Release v4.6.40

refactor: security batch 3 — platform service scoping (do not auto-merge)

.env.example

@@ -0,0 +1,43 @@
+# PraisonAI Environment Variables
+# Copy this file to .env and update with your actual values
+
+# === Required ===
+OPENAI_API_KEY=sk-...          # OpenAI key for LLM calls. Get yours at https://platform.openai.com
+
+# === Required for Claw web-search tool ===
+TAVILY_API_KEY=tvly-...        # Get yours at https://app.tavily.com (free tier available)
+
+# === Optional: Claw dashboard security ===
+PRAISONAI_ALLOW_LOCAL_TOOLS=0  # Set to 1 to enable local shell tools (security risk)
+
+# === Optional: Alternative LLM Providers ===
+ANTHROPIC_API_KEY=sk-ant-...   # Anthropic Claude models
+GOOGLE_API_KEY=...             # Google Gemini models  
+GROQ_API_KEY=gsk_...           # Groq (fast inference)
+COHERE_API_KEY=...             # Cohere models
+AZURE_OPENAI_API_KEY=...       # Azure OpenAI
+AZURE_OPENAI_ENDPOINT=...      # Azure OpenAI endpoint
+
+# === Optional: Memory backends ===
+MEM0_API_KEY=...               # Mem0 memory service
+REDIS_URL=redis://localhost:6379  # Redis for state management
+
+# === Optional: Observability ===
+LANGFUSE_SECRET_KEY=sk-lf-...  # Langfuse tracing
+LANGFUSE_PUBLIC_KEY=pk-lf-...  # Langfuse public key
+
+# === Optional: Bot platforms ===
+TELEGRAM_BOT_TOKEN=...         # Telegram bot token
+DISCORD_BOT_TOKEN=...          # Discord bot token  
+SLACK_BOT_TOKEN=xoxb-...       # Slack bot token
+SLACK_APP_TOKEN=xapp-...       # Slack app token
+WHATSAPP_ACCESS_TOKEN=...      # WhatsApp bot access token
+WHATSAPP_PHONE_NUMBER_ID=...   # WhatsApp phone number ID
+
+# === Optional: Database ===
+DATABASE_URL=sqlite:///~/.praison/database.sqlite  # Database connection string
+
+# === Optional: Development ===
+LOGLEVEL=INFO                  # Logging level (DEBUG, INFO, WARNING, ERROR)
+CHAINLIT_AUTH_SECRET=...       # Chainlit authentication secret
+DEBUG=false                    # Enable debug mode

.github/workflows/claude.yml

@@ -181,12 +181,15 @@ jobs:
 
       - name: Fetch PR branch and Setup Remote
         if: github.event.issue.pull_request
+        env:
+          PR_BRANCH: ${{ steps.check_fork.outputs.pr_branch }}
+          IS_FORK: ${{ steps.check_fork.outputs.is_fork }}
         run: |
           # Fetch PR head from base repo and put it into local branch
-          git fetch origin pull/${{ github.event.issue.number }}/head:${{ steps.check_fork.outputs.pr_branch }}
+          git fetch origin pull/${{ github.event.issue.number }}/head:"${PR_BRANCH}"
           
           # If it's a fork, make origin point to local to trick claude-code-action's `git fetch origin <branch>`
-          if [ "${{ steps.check_fork.outputs.is_fork }}" == "true" ]; then
+          if [ "${IS_FORK}" = "true" ]; then
             git remote set-url origin file://$(pwd)
           fi
 

README.md

@@ -23,7 +23,7 @@
 
 </div>
 
-PraisonAI 🦞 — **Hire a 24/7 AI Workforce.** Stop writing boilerplate and start shipping autonomous agents that research, plan, and execute tasks across your apps. From one agent to an entire organization, deployed in 5 lines of code.
+PraisonAI 🦞 — **Hire a 24/7 AI Workforce.** Stop writing boilerplate and start shipping autonomous, self-improving agents that research, plan, and execute tasks across your apps. From one agent to an entire organization, deployed in 5 lines of code.
 
 ```bash
 curl -fsSL https://praison.ai/install.sh | bash
@@ -294,8 +294,35 @@ def search(query: str) -> str:
 
 @tool
 def calculate(expression: str) -> float:
-    """Evaluate a math expression."""
-    return eval(expression)
+    """Safely evaluate a numeric arithmetic expression."""
+    import ast
+    import operator
+    
+    # Define allowed operations
+    _OPS = {
+        ast.Add: operator.add,
+        ast.Sub: operator.sub,
+        ast.Mult: operator.mul,
+        ast.Div: operator.truediv,
+        ast.Pow: operator.pow,
+        ast.USub: operator.neg,
+        ast.UAdd: operator.pos,
+    }
+    
+    def _safe_eval(node):
+        if isinstance(node, ast.Constant) and isinstance(node.value, (int, float)):
+            return node.value
+        elif isinstance(node, ast.BinOp) and type(node.op) in _OPS:
+            return _OPS[type(node.op)](_safe_eval(node.left), _safe_eval(node.right))
+        elif isinstance(node, ast.UnaryOp) and type(node.op) in _OPS:
+            return _OPS[type(node.op)](_safe_eval(node.operand))
+        else:
+            raise ValueError("Unsupported expression")
+    
+    try:
+        return _safe_eval(ast.parse(expression, mode="eval").body)
+    except (ValueError, SyntaxError, TypeError, ZeroDivisionError, OverflowError):
+        raise ValueError("Invalid arithmetic expression")
 
 agent = Agent(
     instructions="You are a helpful assistant",
@@ -304,6 +331,7 @@ agent = Agent(
 agent.start("Search for AI news and calculate 15*4")
 ```
 
+> ⚠️ **Security Note:** Never use `eval()`, `exec()`, or `subprocess` in tool functions that process LLM-generated or user-supplied input. Always validate and sanitize inputs to prevent code injection attacks.
 > 📖 [Full tools docs](https://docs.praison.ai/docs/tools/tools) — BaseTool, tool packages, 100+ built-in tools
 
 ### 5. Persistence (Databases)
@@ -330,6 +358,15 @@ pip install "praisonai[claw]"
 praisonai claw
 ```
 
+#### Required Environment Variables
+
+Copy `.env.example` to `.env` and configure the following variables:
+
+| Variable | Required | Description |
+|----------|----------|-------------|
+| `OPENAI_API_KEY` | Yes | OpenAI API key for all LLM calls |
+| `TAVILY_API_KEY` | Yes (Claw) | Tavily key for the built-in web-search tool. Get one free at https://app.tavily.com |
+
 Open **http://localhost:8082** — the dashboard comes with 13 built-in pages: Chat, Agents, Memory, Knowledge, Channels, Guardrails, Cron, and more. Add messaging channels directly from the UI.
 
 > 📖 [Full Claw docs](https://docs.praison.ai/docs/concepts/claw) — platform tokens, CLI options, Docker, and YAML agent mode

SECURITY_RELEASE.md

@@ -0,0 +1,34 @@
+# Security release checklist (maintainer)
+
+After merging security PRs [#1684](https://github.com/MervinPraison/PraisonAI/pull/1684) and [#1685](https://github.com/MervinPraison/PraisonAI/pull/1685):
+
+## 1. Publish PyPI (requires `PYPI_TOKEN`)
+
+```bash
+cd src/praisonai-agents
+export PYPI_TOKEN=...
+praisonai publish pypi
+
+cd ../praisonai
+python scripts/bump_and_release.py <WRAPPER_VERSION> --agents <AGENTS_VERSION> --wait
+```
+
+## 2. Publish GitHub security advisories
+
+Set versions to match PyPI, then:
+
+```bash
+AGENTS_VERSION=1.6.40 WRAPPER_VERSION=4.6.40 ./scripts/security_publish_advisories.sh
+```
+
+Add reporter `credits` per advisory via `gh api` before publishing if not already set.
+
+## 3. Verify
+
+```bash
+pip index versions praisonaiagents | head -1
+pip index versions praisonai | head -1
+gh api repos/MervinPraison/PraisonAI/security-advisories --jq '.[] | "\(.ghsa_id) \(.state) \(.cve_id // "pending")"'
+```
+
+Resources: https://github.com/MervinPraison/PraisonAI · https://docs.praison.ai · https://praison.ai

SECURITY_TRIAGE.md

@@ -0,0 +1,26 @@
+# Security advisory triage (maintainer)
+
+| GHSA | Status after batch 1+2 | Notes |
+|------|------------------------|-------|
+| GHSA-4mr5-g6f9-cfrh | fixed batch 1 | Sandbox AST |
+| GHSA-5c6w-wwfq-7qqm | fixed batch 1 | spider_tools SSRF |
+| GHSA-9cr9-25q5-8prj | fixed batch 1 | MCP yaml paths |
+| GHSA-86qc-r5v2-v6x6 | fixed batch 1 | call server auth |
+| GHSA-hvhp-v2gc-268q | fixed batch 1 | write_file cwd |
+| GHSA-5cxw-77wg-jrf3 | fixed batch 1 | @url mentions |
+| GHSA-xp85-6wwf-r67c | fixed batch 1 | GHA branch quote |
+| GHSA-3qg8-5g3r-79v5 | fixed batch 1+2 | JWT + issue guard |
+| GHSA-xwq8, 7p8g, w388, g8rr | fixed batch 1 | platform RBAC/IDOR partial |
+| GHSA-943m, 5jx9, 4x6r, 27p4, cp4f | fixed batch 2 | platform IDOR completion |
+| GHSA-vg22-4gmj-prxw | fixed batch 2 | example eval hardening |
+| GHSA-8444-4fhq-fxpq | already-fixed | APIConfig.auth_enabled default True |
+| GHSA-78r8-wwqv-r299 | already-fixed | load_user_module gate |
+| GHSA-gv23-xxxx-xxxx, GHSA-h8q5-cp56-rr65, GHSA-6h6v-xxxx-xxxx, GHSA-h37g-4h4p-9x97 | fixed batch 3 | service-layer workspace_id on get/update/delete |
+| GHSA-c2m8-xxxx-xxxx, GHSA-8g2p-xxxx-xxxx | fixed batch 3 | only owner assigns admin/owner |
+| GHSA-h8q5-cp56-rr65 | fixed batch 3 | bind default 127.0.0.1 (+ PLATFORM_HOST) |
+| GHSA-6xj3-927j-6pqw | not-applicable | Open WebUI path; not in this repo |
+| GHSA-gmjg, 9q28 | published | prior release |
+
+**Code fixed on main; GHSA state still triage until PyPI publish + advisory publish.**
+
+Resources: https://github.com/MervinPraison/PraisonAI · https://docs.praison.ai · https://praison.ai

api.md

@@ -172,6 +172,7 @@ Methods:
 * <code title="class BaseTool">BaseTool.<a href="./src/praisonai-agents/praisonaiagents/tools/base.py">validate</a>() -> bool</code>
 * <code title="class BaseTool">BaseTool.<a href="./src/praisonai-agents/praisonaiagents/tools/base.py">validate_class</a>() -> bool</code>
 * <code title="class FunctionTool">FunctionTool.<a href="./src/praisonai-agents/praisonaiagents/tools/decorator.py">__call__</a>(*args, **kwargs) -> Any</code>
+* <code title="class FunctionTool">FunctionTool.<a href="./src/praisonai-agents/praisonaiagents/tools/decorator.py">check_availability</a>() -> tuple[bool, str]</code>
 * <code title="class FunctionTool">FunctionTool.<a href="./src/praisonai-agents/praisonaiagents/tools/decorator.py">injected_params</a>() -> Dict[str, Any]</code>
 * <code title="class FunctionTool">FunctionTool.<a href="./src/praisonai-agents/praisonaiagents/tools/decorator.py">run</a>(**kwargs) -> Any</code>
 * <code title="class ToolRegistry">ToolRegistry.<a href="./src/praisonai-agents/praisonaiagents/tools/registry.py">clear</a>() -> None</code>
@@ -179,6 +180,7 @@ Methods:
 * <code title="class ToolRegistry">ToolRegistry.<a href="./src/praisonai-agents/praisonaiagents/tools/registry.py">discover_single_file_plugins</a>() -> int</code>
 * <code title="class ToolRegistry">ToolRegistry.<a href="./src/praisonai-agents/praisonaiagents/tools/registry.py">get</a>(name: str) -> Optional[Union[BaseTool, Callable]]</code>
 * <code title="class ToolRegistry">ToolRegistry.<a href="./src/praisonai-agents/praisonaiagents/tools/registry.py">get_all</a>() -> Dict[str, Union[BaseTool, Callable]]</code>
+* <code title="class ToolRegistry">ToolRegistry.<a href="./src/praisonai-agents/praisonaiagents/tools/registry.py">list_available_tools</a>(context: Optional[Dict[str, Any]] = None) -> List[Union[BaseTool, Callable]]</code>
 * <code title="class ToolRegistry">ToolRegistry.<a href="./src/praisonai-agents/praisonaiagents/tools/registry.py">list_base_tools</a>() -> List[BaseTool]</code>
 * <code title="class ToolRegistry">ToolRegistry.<a href="./src/praisonai-agents/praisonaiagents/tools/registry.py">list_tools</a>() -> List[str]</code>
 * <code title="class ToolRegistry">ToolRegistry.<a href="./src/praisonai-agents/praisonaiagents/tools/registry.py">register</a>(tool: Union[BaseTool, Callable], name: Optional[str] = None, overwrite: bool = False) -> None</code>
@@ -206,11 +208,12 @@ Methods:
 * <code title="class Task">Task.<a href="./src/praisonai-agents/praisonaiagents/task/task.py">execute_callback_sync</a>(task_output: TaskOutput) -> None</code>
 * <code title="class Task">Task.<a href="./src/praisonai-agents/praisonaiagents/task/task.py">get_next_task</a>(context: Dict[str, Any]) -> Optional[str]</code>
 * <code title="class Task">Task.<a href="./src/praisonai-agents/praisonaiagents/task/task.py">initialize_memory</a>()</code>
+* <code title="class Task">Task.<a href="./src/praisonai-agents/praisonaiagents/task/task.py">initialize_memory_async</a>()</code>
 * <code title="class Task">Task.<a href="./src/praisonai-agents/praisonaiagents/task/task.py">set_status</a>(new_status: str) -> None</code>
 * <code title="class Task">Task.<a href="./src/praisonai-agents/praisonaiagents/task/task.py">store_in_memory</a>(content: str, agent_name: str = None, task_id: str = None)</code>
 * <code title="class Task">Task.<a href="./src/praisonai-agents/praisonaiagents/task/task.py">to_dict</a>() -> Dict[str, Any]</code>
 * <code title="function">praisonaiagents.<a href="./src/praisonai-agents/praisonaiagents/workflows/workflows.py">loop</a>(step: Any = None, steps: Optional[List[Any]] = None, over: Optional[str] = None, from_csv: Optional[str] = None, from_file: Optional[str] = None, var_name: str = 'item', parallel: bool = False, max_workers: Optional[int] = None, output_variable: Optional[str] = None) -> Loop</code>
-* <code title="function">praisonaiagents.<a href="./src/praisonai-agents/praisonaiagents/workflows/workflows.py">parallel</a>(steps: List) -> Parallel</code>
+* <code title="function">praisonaiagents.<a href="./src/praisonai-agents/praisonaiagents/workflows/workflows.py">parallel</a>(steps: List, max_workers: Optional[int] = None, on_failure: str = 'partial_ok') -> Parallel</code>
 * <code title="function">praisonaiagents.<a href="./src/praisonai-agents/praisonaiagents/workflows/workflows.py">repeat</a>(step: Any, until: Optional[Callable[[WorkflowContext], bool]] = None, max_iterations: int = 10) -> Repeat</code>
 * <code title="function">praisonaiagents.<a href="./src/praisonai-agents/praisonaiagents/workflows/workflows.py">route</a>(routes: Dict[str, List], default: Optional[List] = None) -> Route</code>
 
@@ -467,7 +470,10 @@ Methods:
 * <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">activate_by_name</a>(name: str) -> bool</code>
 * <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">add_skill</a>(skill_path: str) -> Optional[LoadedSkill]</code>
 * <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">clear</a>() -> None</code>
+* <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">create_skill</a>(name: str, content: str, category: str = None) -> dict</code>
+* <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">delete_skill</a>(name: str) -> dict</code>
 * <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">discover</a>(skill_dirs: Optional[List[str]] = None, include_defaults: bool = True) -> int</code>
+* <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">edit_skill</a>(name: str, content: str) -> dict</code>
 * <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">get_allowed_tools</a>(name: str) -> List[str]</code>
 * <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">get_available_skills</a>() -> List[SkillMetadata]</code>
 * <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">get_instructions</a>(name: str) -> Optional[str]</code>
@@ -475,9 +481,12 @@ Methods:
 * <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">get_user_invocable_skills</a>() -> List[LoadedSkill]</code>
 * <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">invoke</a>(name: str, raw_args: str = '', session_id: Optional[str] = None, shell_exec: bool = False) -> Optional[str]</code>
 * <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">load_resources</a>(name: str) -> bool</code>
+* <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">patch_skill</a>(name: str, old_string: str, new_string: str, file_path: str = None, replace_all: bool = False) -> dict</code>
+* <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">remove_skill_file</a>(name: str, file_path: str) -> dict</code>
 * <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">skill_names</a>() -> List[str]</code>
 * <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">skills</a>() -> List[LoadedSkill]</code>
 * <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">to_prompt</a>() -> str</code>
+* <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">write_skill_file</a>(name: str, file_path: str, file_content: str) -> dict</code>
 * <code title="class SkillMetadata">SkillMetadata.<a href="./src/praisonai-agents/praisonaiagents/skills/models.py">from_properties</a>(props: SkillProperties) -> 'SkillMetadata'</code>
 * <code title="class SkillProperties">SkillProperties.<a href="./src/praisonai-agents/praisonaiagents/skills/models.py">to_dict</a>() -> dict</code>
 * <code title="class SkillsConfig">SkillsConfig.<a href="./src/praisonai-agents/praisonaiagents/config/feature_configs.py">to_dict</a>() -> Dict[str, Any]</code>
@@ -704,6 +713,7 @@ Methods:
 * <code title="class AgentTeam">AgentTeam.<a href="./src/praisonai-agents/praisonaiagents/agents/agents.py">astart</a>(content = None, return_dict = False, **kwargs)</code>
 * <code title="class AgentTeam">AgentTeam.<a href="./src/praisonai-agents/praisonaiagents/agents/agents.py">clean_json_output</a>(output: str) -> str</code>
 * <code title="class AgentTeam">AgentTeam.<a href="./src/praisonai-agents/praisonaiagents/agents/agents.py">clear_state</a>() -> None</code>
+* <code title="class AgentTeam">AgentTeam.<a href="./src/praisonai-agents/praisonaiagents/agents/agents.py">close</a>() -> None</code>
 * <code title="class AgentTeam">AgentTeam.<a href="./src/praisonai-agents/praisonaiagents/agents/agents.py">context_manager</a>()</code>
 * <code title="class AgentTeam">AgentTeam.<a href="./src/praisonai-agents/praisonaiagents/agents/agents.py">current_plan</a>()</code>
 * <code title="class AgentTeam">AgentTeam.<a href="./src/praisonai-agents/praisonaiagents/agents/agents.py">default_completion_checker</a>(task, agent_output)</code>
@@ -745,7 +755,7 @@ Methods:
 
 Types:
 ```python
-from praisonai import AgentApp, AgentOS, AnthropicManagedAgent, CloudProvider, DB, Deploy, DeployConfig, DeployType, LocalManagedAgent, LocalManagedConfig, ManagedAgent, ManagedConfig, PraisonAI, __version__
+from praisonai import Agent, AgentApp, AgentOS, AnthropicManagedAgent, CloudProvider, DB, Deploy, DeployConfig, DeployType, HostedAgent, HostedAgentConfig, LocalAgent, LocalAgentConfig, LocalManagedAgent, LocalManagedConfig, ManagedAgent, ManagedConfig, PraisonAI, __version__
 ```
 
 # CLI
@@ -776,6 +786,10 @@ Methods:
 * <code title="cli">praisonai bot discord <a href="./src/praisonai/praisonai/cli/commands/bot.py">--help</a></code>
 * <code title="cli">praisonai bot email <a href="./src/praisonai/praisonai/cli/commands/bot.py">--help</a></code>
 * <code title="cli">praisonai bot install-daemon <a href="./src/praisonai/praisonai/cli/commands/bot.py">--help</a></code>
+* <code title="cli">praisonai bot linear <a href="./src/praisonai/praisonai/cli/commands/bot.py">--help</a></code>
+* <code title="cli">praisonai bot list <a href="./src/praisonai/praisonai/cli/commands/bot.py">--help</a></code>
+* <code title="cli">praisonai bot purge <a href="./src/praisonai/praisonai/cli/commands/bot.py">--help</a></code>
+* <code title="cli">praisonai bot replay <a href="./src/praisonai/praisonai/cli/commands/bot.py">--help</a></code>
 * <code title="cli">praisonai bot slack <a href="./src/praisonai/praisonai/cli/commands/bot.py">--help</a></code>
 * <code title="cli">praisonai bot start <a href="./src/praisonai/praisonai/cli/commands/bot.py">--help</a></code>
 * <code title="cli">praisonai bot telegram <a href="./src/praisonai/praisonai/cli/commands/bot.py">--help</a></code>
@@ -872,6 +886,7 @@ Methods:
 * <code title="cli">praisonai gateway gateway-callback <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
 * <code title="cli">praisonai gateway install <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
 * <code title="cli">praisonai gateway logs <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
+* <code title="cli">praisonai gateway mint-link <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
 * <code title="cli">praisonai gateway send <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
 * <code title="cli">praisonai gateway start <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
 * <code title="cli">praisonai gateway status <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
@@ -944,6 +959,10 @@ Methods:
 * <code title="cli">praisonai package install <a href="./src/praisonai/praisonai/cli/commands/package.py">--help</a></code>
 * <code title="cli">praisonai package list <a href="./src/praisonai/praisonai/cli/commands/package.py">--help</a></code>
 * <code title="cli">praisonai package uninstall <a href="./src/praisonai/praisonai/cli/commands/package.py">--help</a></code>
+* <code title="cli">praisonai pairing approve <a href="./src/praisonai/praisonai/cli/commands/pairing.py">--help</a></code>
+* <code title="cli">praisonai pairing clear <a href="./src/praisonai/praisonai/cli/commands/pairing.py">--help</a></code>
+* <code title="cli">praisonai pairing list <a href="./src/praisonai/praisonai/cli/commands/pairing.py">--help</a></code>
+* <code title="cli">praisonai pairing revoke <a href="./src/praisonai/praisonai/cli/commands/pairing.py">--help</a></code>
 * <code title="cli">praisonai paths show <a href="./src/praisonai/praisonai/cli/commands/paths.py">--help</a></code>
 * <code title="cli">praisonai plugins create <a href="./src/praisonai/praisonai/cli/commands/plugins.py">--help</a></code>
 * <code title="cli">praisonai plugins disable <a href="./src/praisonai/praisonai/cli/commands/plugins.py">--help</a></code>

docker/Dockerfile.chat

@@ -16,7 +16,7 @@ RUN mkdir -p /root/.praison
 # Install Python packages (using latest versions)
 RUN pip install --no-cache-dir \
     praisonai_tools \
-    "praisonai>=4.6.22" \
+    "praisonai>=4.6.40" \
     "praisonai[chat]" \
     "embedchain[github,youtube]"
 

docker/Dockerfile.dev

@@ -20,7 +20,7 @@ RUN mkdir -p /root/.praison
 # Install Python packages (using latest versions)
 RUN pip install --no-cache-dir \
     praisonai_tools \
-    "praisonai>=4.6.22" \
+    "praisonai>=4.6.40" \
     "praisonai[ui]" \
     "praisonai[chat]" \
     "praisonai[realtime]" \

docker/Dockerfile.ui

@@ -16,7 +16,7 @@ RUN mkdir -p /root/.praison
 # Install Python packages (using latest versions)
 RUN pip install --no-cache-dir \
     praisonai_tools \
-    "praisonai>=4.6.22" \
+    "praisonai>=4.6.40" \
     "praisonai[ui]" \
     "praisonai[crewai]"
 

examples/python/linear_agent_example.py

@@ -0,0 +1,68 @@
+"""
+Linear Agent Example for PraisonAI.
+
+Demonstrates how to create a Linear bot that responds to issue mentions
+and assignments with full coding capabilities.
+"""
+
+import asyncio
+from praisonaiagents import Agent
+from praisonai.bots import LinearBot
+
+# Create an agent with Linear tools and coding capabilities
+agent = Agent(
+    name="PraisonAI Coder",
+    instructions="""
+    You are an autonomous coding agent integrated with Linear.
+    
+    When you are mentioned or assigned an issue:
+    1. Analyze the issue description and requirements
+    2. Break down complex tasks into smaller steps  
+    3. Use available tools to implement solutions
+    4. Update the issue with progress comments
+    5. Create GitHub pull requests when code changes are needed
+    
+    Focus on being helpful, thorough, and providing clear updates.
+    """,
+    llm="gpt-4o-mini",
+    tools=[
+        "linear_search_issues",
+        "linear_get_issue", 
+        "linear_add_comment",
+        "linear_update_issue",
+        "linear_list_teams",
+        "linear_list_issue_states",
+        "github_create_branch",
+        "github_commit_and_push", 
+        "github_create_pull_request",
+        "read_file",
+        "write_file",
+        "execute_command"
+    ],
+    memory=True,
+    web_search=True,
+    auto_approve_tools=True,
+)
+
+# Create Linear bot
+bot = LinearBot(
+    token="your-linear-oauth-token",  # or set LINEAR_OAUTH_TOKEN env var
+    signing_secret="your-webhook-secret",  # or set LINEAR_WEBHOOK_SECRET env var
+    agent=agent,
+    webhook_port=8080,
+)
+
+async def main():
+    """Start the Linear bot."""
+    print("Starting Linear bot...")
+    print("Set up your Linear webhook to point to: http://your-host:8080/webhook")
+    print("Configure AgentSession and Comment events in Linear webhook settings")
+    
+    try:
+        await bot.start()
+    except KeyboardInterrupt:
+        print("\nStopping bot...")
+        await bot.stop()
+
+if __name__ == "__main__":
+    asyncio.run(main())

examples/python/managed-agents/provider/all_providers.py

@@ -1,7 +1,7 @@
 """All compute providers — comprehensive test across Local, Docker, E2B, and Modal.
 
-This example mirrors the Anthropic app.py but uses the local provider with
-various compute backends instead of Anthropic's managed infrastructure.
+This example demonstrates local agent loops with various compute backends for tool sandboxing.
+Uses the new LocalAgent class to clearly indicate local execution with optional cloud compute.
 
 Requires:
   - Docker running locally
@@ -9,7 +9,8 @@ Requires:
   - modal CLI configured (modal token set)
 """
 import asyncio
-from praisonai import Agent, ManagedAgent, LocalManagedConfig
+from praisonai import Agent
+from praisonai.integrations import LocalAgent, LocalAgentConfig
 
 
 async def test_provider(name, compute, extra_provision_kwargs=None):
@@ -18,10 +19,9 @@ async def test_provider(name, compute, extra_provision_kwargs=None):
     print(f"  PROVIDER: {name}")
     print(f"{'='*60}")
 
-    managed = ManagedAgent(
-        provider="local",
+    managed = LocalAgent(
         compute=compute,
-        config=LocalManagedConfig(
+        config=LocalAgentConfig(
             model="gpt-4o-mini",
             system="You are a helpful assistant. Be concise.",
             name=f"{name}Agent",

examples/python/managed-agents/provider/all_runtimes.py

@@ -0,0 +1,47 @@
+"""Run every managed-runtime example and report a summary.
+
+Executes the runtime_*.py files in this directory as subprocesses with a
+30s per-example timeout. Exits non-zero if any example fails.
+"""
+import os
+import subprocess
+import sys
+import time
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+EXAMPLES = [
+    # ── LLM providers (where the LLM reasoning happens) ──
+    "runtime_anthropic.py",   # full loop in Anthropic cloud
+    "runtime_openai.py",      # LLM in OpenAI cloud
+    "runtime_gemini.py",      # LLM in Google cloud
+    "runtime_ollama.py",      # LLM on local/self-hosted Ollama
+    # ── Cloud/compute providers (where tool execution happens) ──
+    "runtime_e2b.py",         # tools sandboxed in E2B cloud VM
+    "runtime_modal.py",       # tools sandboxed in Modal serverless cloud
+    "runtime_fly.py",         # tools sandboxed on Fly.io Machines
+    "runtime_daytona.py",     # tools sandboxed in Daytona cloud workspace
+    "runtime_docker.py",      # tools sandboxed in local Docker container
+]
+
+results = []
+for name in EXAMPLES:
+    path = os.path.join(HERE, name)
+    start = time.time()
+    proc = subprocess.run(
+        [sys.executable, path],
+        capture_output=True, text=True, timeout=120,
+        env=os.environ.copy(),  # propagate API keys and PYTHONPATH
+    )
+    dur = time.time() - start
+    ok = proc.returncode == 0
+    results.append((name, ok, dur, proc.stdout.strip().splitlines()[-1] if proc.stdout.strip() else ""))
+    print(f"[{'OK' if ok else 'FAIL'}] {name:30s} {dur:.2f}s")
+    if not ok:
+        print(proc.stdout)
+        print(proc.stderr)
+
+print("\nSummary:")
+for name, ok, dur, tail in results:
+    print(f"  {'✓' if ok else '✗'} {name:30s} {dur:5.2f}s  {tail[:60]}")
+
+sys.exit(0 if all(r[1] for r in results) else 1)

examples/python/managed-agents/provider/local_advanced.py

@@ -23,12 +23,39 @@ print(f"[2] Updated to v{managed.agent_version}")
 result = agent.start("What is 123 + 456?", stream=True)
 
 # ── 3. Custom tool ──
+def _safe_calc(expr: str) -> str:
+    import ast
+    allowed = set("0123456789+-*/.() ")
+    if not all(c in allowed for c in expr):
+        return "error"
+    try:
+        tree = ast.parse(expr, mode="eval")
+        for node in ast.walk(tree):
+            if not isinstance(
+                node,
+                (
+                    ast.Expression,
+                    ast.BinOp,
+                    ast.UnaryOp,
+                    ast.Constant,
+                    ast.Add,
+                    ast.Sub,
+                    ast.Mult,
+                    ast.Div,
+                    ast.USub,
+                    ast.UAdd,
+                ),
+            ):
+                return "error"
+        val = eval(compile(tree, "<expr>", "eval"), {"__builtins__": {}})
+        return str(val)
+    except Exception:
+        return "error"
+
+
 def handle_calculator(tool_name, tool_input):
     expr = tool_input.get("expression", "0")
-    try:
-        val = eval(expr, {"__builtins__": {}})
-    except Exception:
-        val = "error"
+    val = _safe_calc(str(expr))
     print(f"  [Calculator: {expr} = {val}]")
     return str(val)
 

examples/python/managed-agents/provider/local_basic.py

@@ -1,13 +1,14 @@
-"""Local provider — basic managed agent with gpt-4o-mini.
+"""Local agent — basic local execution with gpt-4o-mini.
 
 No external infrastructure needed. Runs the agent loop locally.
+Uses the new canonical LocalAgent class for clarity.
 """
-from praisonai import Agent, ManagedAgent, LocalManagedConfig
+from praisonai import Agent
+from praisonai.integrations import LocalAgent, LocalAgentConfig
 
-# Create a local managed agent (auto-detects local when no ANTHROPIC_API_KEY)
-managed = ManagedAgent(
-    provider="local",
-    config=LocalManagedConfig(
+# Create a local agent (runs locally, no managed runtime)
+managed = LocalAgent(
+    config=LocalAgentConfig(
         model="gpt-4o-mini",
         system="You are a helpful assistant. Be concise.",
         name="LocalAgent",

examples/python/managed-agents/provider/runtime_anthropic.py

@@ -0,0 +1,47 @@
+"""Anthropic managed runtime — full agent loop runs in Anthropic's cloud.
+
+The ENTIRE loop (LLM calls, tool execution, memory, sessions) runs remotely
+on Anthropic's managed infrastructure. This is the only provider today where
+"the whole thing runs in that cloud".
+
+Requires: ANTHROPIC_API_KEY environment variable set.
+"""
+import os
+
+if not os.getenv("ANTHROPIC_API_KEY"):
+    print("[skip] ANTHROPIC_API_KEY not set — set it to run this example.")
+    raise SystemExit(0)
+
+try:
+    import anthropic  # noqa: F401  (required by AnthropicManagedAgent)
+except ImportError:
+    print("[skip] 'anthropic' SDK not installed — pip install 'anthropic>=0.94.0'")
+    raise SystemExit(0) from None
+
+try:
+    from praisonai import Agent, ManagedAgent, ManagedConfig
+except ImportError as e:
+    print(f"[skip] praisonai wrapper not importable ({e}).")
+    raise SystemExit(0) from None
+
+managed = ManagedAgent(
+    provider="anthropic",
+    config=ManagedConfig(
+        model="claude-3-5-sonnet-latest",
+        system="You are a concise coding assistant. One-sentence answers.",
+        name="AnthropicRuntimeAgent",
+    ),
+)
+agent = Agent(name="anthropic-runtime", backend=managed)
+
+# 1. Agent work entirely in the cloud
+print("[1] Agent:", agent.start("Write a Python one-liner that sums 1..10. Just the code."))
+
+# 2. Multi-turn reuses the same cloud session
+print("[2] Agent:", agent.start("Now change it to factorial of 5."))
+
+# 3. Cloud-side session metadata
+info = managed.retrieve_session()
+print(f"[3] session={managed.session_id}  status={info.get('status')}  "
+      f"in={info.get('usage', {}).get('input_tokens', 0)}  "
+      f"out={info.get('usage', {}).get('output_tokens', 0)}")

examples/python/managed-agents/provider/runtime_daytona.py

@@ -0,0 +1,52 @@
+"""Agent runtime with tools sandboxed in a Daytona cloud workspace.
+
+The agent's LLM loop runs locally; tool / code execution runs in a Daytona-
+managed workspace (disposable cloud dev environment, pre-configured images).
+
+Requires:
+  - ``OPENAI_API_KEY`` for the LLM
+  - ``DAYTONA_API_KEY`` for the Daytona API
+  - ``pip install daytona-sdk``
+"""
+import asyncio
+import os
+
+if not os.getenv("OPENAI_API_KEY"):
+    print("[skip] OPENAI_API_KEY not set.")
+    raise SystemExit(0)
+if not os.getenv("DAYTONA_API_KEY"):
+    print("[skip] DAYTONA_API_KEY not set.")
+    raise SystemExit(0)
+
+try:
+    from praisonai import Agent
+    from praisonai.integrations import SandboxedAgent, SandboxedAgentConfig
+except ImportError as e:
+    print(f"[skip] praisonai wrapper not importable ({e}).")
+    raise SystemExit(0) from None
+
+
+async def main():
+    sandboxed = SandboxedAgent(
+        compute="daytona",  # tool execution runs in a Daytona workspace
+        config=SandboxedAgentConfig(
+            model="gpt-4o-mini",
+            system="You are a concise coding assistant.",
+            name="DaytonaRuntimeAgent",
+        ),
+    )
+    agent = Agent(name="daytona-runtime", backend=sandboxed)
+
+    info = await sandboxed.provision_compute(idle_timeout_s=120)
+    print(f"[1] Daytona workspace: {info.instance_id} ({info.status})")
+
+    result = await sandboxed.execute_in_compute("python3 -c 'print(len([1,2,3,4,5]) ** 3)'")
+    print(f"[2] Cloud compute result: {result['stdout'].strip()}")
+
+    print("[3] Agent:", agent.start("What is 5 cubed? Just the number."))
+
+    await sandboxed.shutdown_compute()
+    print("[4] Daytona workspace shut down.")
+
+
+asyncio.run(main())

examples/python/managed-agents/provider/runtime_docker.py

@@ -0,0 +1,59 @@
+"""Agent runtime with tools sandboxed in a local Docker container.
+
+The agent's LLM loop runs locally; tool / code execution is isolated in a
+Docker container. Not a cloud provider, but the same pattern — useful for
+local development and CI.
+
+Requires:
+  - ``OPENAI_API_KEY`` for the LLM
+  - Docker daemon running (``docker info`` must succeed)
+"""
+import asyncio
+import os
+import subprocess
+
+if not os.getenv("OPENAI_API_KEY"):
+    print("[skip] OPENAI_API_KEY not set.")
+    raise SystemExit(0)
+
+try:
+    subprocess.run(
+        ["docker", "info"],
+        check=True, capture_output=True, timeout=5,
+    )
+except (subprocess.CalledProcessError, subprocess.TimeoutExpired, FileNotFoundError):
+    print("[skip] Docker daemon not available — start Docker Desktop.")
+    raise SystemExit(0) from None
+
+try:
+    from praisonai import Agent
+    from praisonai.integrations import SandboxedAgent, SandboxedAgentConfig
+except ImportError as e:
+    print(f"[skip] praisonai wrapper not importable ({e}).")
+    raise SystemExit(0) from None
+
+
+async def main():
+    sandboxed = SandboxedAgent(
+        compute="docker",  # tool execution runs in a local Docker container
+        config=SandboxedAgentConfig(
+            model="gpt-4o-mini",
+            system="You are a concise coding assistant.",
+            name="DockerRuntimeAgent",
+        ),
+    )
+    agent = Agent(name="docker-runtime", backend=sandboxed)
+
+    info = await sandboxed.provision_compute(image="python:3.12-slim")
+    print(f"[1] Docker container: {info.instance_id} ({info.status})")
+
+    result = await sandboxed.execute_in_compute("python3 -c 'print(sum(i*i for i in range(1, 11)))'")
+    print(f"[2] Container compute result: {result['stdout'].strip()}")
+
+    print("[3] Agent:", agent.start("What is the sum of squares from 1 to 10? Just the number."))
+
+    await sandboxed.shutdown_compute()
+    print("[4] Docker container shut down.")
+
+
+asyncio.run(main())

examples/python/managed-agents/provider/runtime_e2b.py

@@ -0,0 +1,56 @@
+"""Agent runtime with tools sandboxed in an E2B cloud VM.
+
+The agent's LLM loop runs locally; any tool / code execution is sandboxed
+in an ephemeral E2B cloud instance. Uses ``SandboxedAgent`` — the honest
+name for the "local loop + cloud tools" pattern.
+
+Requires:
+  - ``OPENAI_API_KEY`` for the LLM
+  - ``E2B_API_KEY``    for the cloud sandbox
+  - ``pip install e2b``
+"""
+import asyncio
+import os
+
+if not os.getenv("OPENAI_API_KEY"):
+    print("[skip] OPENAI_API_KEY not set.")
+    raise SystemExit(0)
+if not os.getenv("E2B_API_KEY"):
+    print("[skip] E2B_API_KEY not set — set it to run in the E2B cloud sandbox.")
+    raise SystemExit(0)
+
+try:
+    from praisonai import Agent
+    from praisonai.integrations import SandboxedAgent, SandboxedAgentConfig
+except ImportError as e:
+    print(f"[skip] praisonai wrapper not importable ({e}).")
+    raise SystemExit(0) from None
+
+
+async def main():
+    sandboxed = SandboxedAgent(
+        compute="e2b",  # tool execution runs in E2B cloud
+        config=SandboxedAgentConfig(
+            model="gpt-4o-mini",
+            system="You are a concise coding assistant.",
+            name="E2BRuntimeAgent",
+        ),
+    )
+    agent = Agent(name="e2b-runtime", backend=sandboxed)
+
+    # 1. Provision the cloud sandbox
+    info = await sandboxed.provision_compute(idle_timeout_s=120)
+    print(f"[1] E2B sandbox: {info.instance_id} ({info.status})")
+
+    # 2. Execute code IN the cloud sandbox
+    result = await sandboxed.execute_in_compute("python3 -c 'print(sum(range(1, 11)))'")
+    print(f"[2] Cloud compute result: {result['stdout'].strip()}")
+
+    # 3. Agent loop (LLM local, any tool calls auto-sandboxed in E2B)
+    print("[3] Agent:", agent.start("What is 13 * 17? Just the number."))
+
+    await sandboxed.shutdown_compute()
+    print("[4] E2B sandbox shut down.")
+
+
+asyncio.run(main())

examples/python/managed-agents/provider/runtime_fly.py

@@ -0,0 +1,57 @@
+"""Agent runtime with tools sandboxed on Fly.io Machines.
+
+The agent's LLM loop runs locally; tool / code execution runs in a Fly.io
+Machine (fast-booting micro-VM with per-second billing and global regions).
+
+Requires:
+  - ``OPENAI_API_KEY`` for the LLM
+  - ``FLY_API_TOKEN``  for the Fly.io Machines API
+  - ``FLY_APP_NAME``   (optional) existing Fly app; auto-created otherwise
+"""
+import asyncio
+import os
+
+if not os.getenv("OPENAI_API_KEY"):
+    print("[skip] OPENAI_API_KEY not set.")
+    raise SystemExit(0)
+if not os.getenv("FLY_API_TOKEN"):
+    print("[skip] FLY_API_TOKEN not set — get one via `fly auth token`.")
+    raise SystemExit(0)
+
+try:
+    from praisonai import Agent
+    from praisonai.integrations import SandboxedAgent, SandboxedAgentConfig
+except ImportError as e:
+    print(f"[skip] praisonai wrapper not importable ({e}).")
+    raise SystemExit(0) from None
+
+
+async def main():
+    sandboxed = SandboxedAgent(
+        compute="flyio",  # tool execution runs on Fly.io Machines
+        config=SandboxedAgentConfig(
+            model="gpt-4o-mini",
+            system="You are a concise coding assistant.",
+            name="FlyioRuntimeAgent",
+        ),
+    )
+    agent = Agent(name="fly-runtime", backend=sandboxed)
+
+    info = await sandboxed.provision_compute(
+        image="python:3.12-slim",
+        cpu=1,
+        memory_mb=512,
+        idle_timeout_s=120,
+    )
+    print(f"[1] Fly Machine: {info.instance_id} ({info.status})")
+
+    result = await sandboxed.execute_in_compute("python3 -c 'print(2 ** 20)'")
+    print(f"[2] Cloud compute result: {result['stdout'].strip()}")
+
+    print("[3] Agent:", agent.start("What is 2 raised to the 20th power? Just the number."))
+
+    await sandboxed.shutdown_compute()
+    print("[4] Fly Machine shut down.")
+
+
+asyncio.run(main())

examples/python/managed-agents/provider/runtime_gemini.py

@@ -0,0 +1,40 @@
+"""Gemini managed runtime — agent loop local, LLM calls go to Google's cloud.
+
+``ManagedAgent(provider="gemini")`` wires a local managed loop to Gemini's
+hosted models via litellm.
+
+Requires: GEMINI_API_KEY (or GOOGLE_API_KEY) environment variable set.
+"""
+import os
+
+if not (os.getenv("GEMINI_API_KEY") or os.getenv("GOOGLE_API_KEY")):
+    print("[skip] GEMINI_API_KEY / GOOGLE_API_KEY not set — set one to run this example.")
+    raise SystemExit(0)
+
+try:
+    from praisonai import Agent, ManagedAgent, LocalManagedConfig
+except ImportError as e:
+    print(f"[skip] praisonai wrapper not importable ({e}).")
+    raise SystemExit(0) from None
+
+managed = ManagedAgent(
+    provider="gemini",
+    config=LocalManagedConfig(
+        # Use explicit litellm prefix to guarantee routing to Google
+        model="gemini/gemini-2.0-flash-exp",
+        system="You are a concise assistant. One-sentence answers.",
+        name="GeminiRuntimeAgent",
+    ),
+)
+agent = Agent(name="gemini-runtime", backend=managed)
+
+# 1. Agent work (LLM call goes to Google cloud)
+print("[1] Agent:", agent.start("What does REST stand for? One sentence."))
+
+# 2. Multi-turn reuses the session
+print("[2] Agent:", agent.start("How is it different from GraphQL?"))
+
+info = managed.retrieve_session()
+print(f"[3] session={managed.session_id}  "
+      f"in={info.get('usage', {}).get('input_tokens', 0)}  "
+      f"out={info.get('usage', {}).get('output_tokens', 0)}")

examples/python/managed-agents/provider/runtime_hosted_anthropic.py

@@ -0,0 +1,63 @@
+"""Anthropic hosted runtime — entire agent runs on Anthropic's managed infrastructure.
+
+Uses the new canonical HostedAgent class which clearly communicates that the entire
+agent loop runs in Anthropic's cloud, not locally.
+"""
+import os
+import sys
+
+# Skip guards - exit cleanly if prerequisites not met
+if not (os.getenv("ANTHROPIC_API_KEY") or os.getenv("CLAUDE_API_KEY")):
+    print("[skip] ANTHROPIC_API_KEY or CLAUDE_API_KEY not set")
+    sys.exit(0)
+
+try:
+    import anthropic
+except ImportError:
+    print("[skip] anthropic SDK not installed")
+    sys.exit(0)
+
+# Heavy imports only after skip-guards pass
+from praisonai import Agent
+from praisonai.integrations import HostedAgent, HostedAgentConfig
+
+# Create a hosted agent running entirely on Anthropic's managed runtime
+hosted = HostedAgent(
+    provider="anthropic",
+    config=HostedAgentConfig(
+        model="claude-3-5-sonnet-latest",
+        system="You are a helpful coding assistant. Be concise.",
+        name="AnthropicHostedAgent",
+        tools=[{"type": "agent_toolset_20260401"}],
+    ),
+)
+
+agent = Agent(name="anthropic-hosted", backend=hosted)
+
+# 1. Basic execution - runs entirely in Anthropic's cloud
+print("[1] Hosted execution on Anthropic infrastructure...")
+result = agent.start("What is the capital of France? One word.", stream=True)
+print(f"    Result: {result}")
+
+# 2. Agent metadata from Anthropic's API
+print(f"\n[2] Agent ID: {hosted.agent_id}")
+print(f"    Version: {hosted.agent_version}")
+print(f"    Env ID:  {hosted.environment_id}")
+print(f"    Session: {hosted.session_id}")
+
+# 3. Multi-turn (same session keeps context in Anthropic's cloud)
+print("\n[3] Multi-turn conversation...")
+result = agent.start("What country is that city in?", stream=True)
+print(f"    Result: {result}")
+
+# 4. Usage tracking from Anthropic's usage API
+info = hosted.retrieve_session()
+print(f"\n[4] Usage: in={info['usage']['input_tokens']}, out={info['usage']['output_tokens']}")
+
+# 5. List all sessions for this agent
+sessions = hosted.list_sessions()
+print(f"\n[5] Sessions: {len(sessions)}")
+for s in sessions[:3]:  # Show first 3
+    print(f"    {s['id']} | {s['status']}")
+
+print("\nDone! Agent loop ran entirely on Anthropic's managed infrastructure.")

examples/python/managed-agents/provider/runtime_local_gemini.py

@@ -0,0 +1,61 @@
+"""Local agent loop with Gemini LLM — runs locally, not in a managed runtime.
+
+Uses the new canonical LocalAgent class which clearly communicates that only the
+agent loop runs locally. The LLM calls go to Google's Gemini API, but there's no managed runtime involved.
+"""
+import os
+import sys
+
+# Skip guards - exit cleanly if prerequisites not met
+if not (os.getenv("GEMINI_API_KEY") or os.getenv("GOOGLE_API_KEY")):
+    print("[skip] GEMINI_API_KEY or GOOGLE_API_KEY not set")
+    sys.exit(0)
+
+try:
+    import google.generativeai
+except ImportError:
+    print("[skip] google-generativeai SDK not installed")
+    sys.exit(0)
+
+# Heavy imports only after skip-guards pass
+from praisonai import Agent
+from praisonai.integrations import LocalAgent, LocalAgentConfig
+
+# Create a local agent using Gemini LLM
+local = LocalAgent(
+    config=LocalAgentConfig(
+        model="gemini/gemini-2.0-flash",  # Use Gemini with litellm routing prefix
+        system="You are a helpful coding assistant. Be concise.",
+        name="LocalGeminiAgent",
+        tools=["execute_command", "read_file", "write_file"],
+    ),
+)
+
+agent = Agent(name="local-gemini", backend=local)
+
+# 1. Basic execution - agent loop runs locally, LLM calls go to Gemini
+print("[1] Local execution with Gemini LLM...")
+result = agent.start("What is the capital of France? One word.", stream=True)
+print(f"    Result: {result}")
+
+# 2. Agent metadata (locally generated UUIDs)
+print(f"\n[2] Agent ID: {local.agent_id}")
+print(f"    Version: {local.agent_version}")
+print(f"    Env ID:  {local.environment_id}")
+print(f"    Session: {local.session_id}")
+
+# 3. Multi-turn conversation (session state maintained locally)
+print("\n[3] Multi-turn conversation...")
+result = agent.start("What country is that city in?", stream=True)
+print(f"    Result: {result}")
+
+# 4. Usage tracking (accumulated locally)
+info = local.retrieve_session()
+print(f"\n[4] Usage: in={info['usage']['input_tokens']}, out={info['usage']['output_tokens']}")
+
+# 5. Tool execution (runs in local subprocess)
+print("\n[5] Tool execution example...")
+result = agent.start("Create a file called hello_gemini.txt with 'Hello from Gemini agent!'", stream=True)
+print(f"    Tool result: {result}")
+
+print("\nDone! Agent loop ran locally with Gemini LLM calls.")

examples/python/managed-agents/provider/runtime_local_ollama.py

@@ -0,0 +1,61 @@
+"""Local agent loop with Ollama LLM — runs locally, not in a managed runtime.
+
+Uses the new canonical LocalAgent class which clearly communicates that only the
+agent loop runs locally. The LLM calls go to a local Ollama instance, no managed runtime involved.
+"""
+import os
+import sys
+import requests
+
+# Skip guards - exit cleanly if prerequisites not met  
+# Check if Ollama daemon is running
+try:
+    response = requests.get("http://localhost:11434/api/tags", timeout=2)
+    if response.status_code != 200:
+        raise Exception("Ollama not responding")
+except Exception:
+    print("[skip] Ollama daemon not running at localhost:11434")
+    sys.exit(0)
+
+# Heavy imports only after skip-guards pass
+from praisonai import Agent
+from praisonai.integrations import LocalAgent, LocalAgentConfig
+
+# Create a local agent using Ollama LLM
+local = LocalAgent(
+    config=LocalAgentConfig(
+        model="ollama/llama3.2",  # Use Ollama with litellm routing prefix
+        system="You are a helpful coding assistant. Be concise.",
+        name="LocalOllamaAgent",
+        tools=["execute_command", "read_file", "write_file"],
+    ),
+)
+
+agent = Agent(name="local-ollama", backend=local)
+
+# 1. Basic execution - agent loop runs locally, LLM calls go to local Ollama
+print("[1] Local execution with Ollama LLM...")
+result = agent.start("What is the capital of France? One word.", stream=True)
+print(f"    Result: {result}")
+
+# 2. Agent metadata (locally generated UUIDs)
+print(f"\n[2] Agent ID: {local.agent_id}")
+print(f"    Version: {local.agent_version}")
+print(f"    Env ID:  {local.environment_id}")
+print(f"    Session: {local.session_id}")
+
+# 3. Multi-turn conversation (session state maintained locally)
+print("\n[3] Multi-turn conversation...")
+result = agent.start("What country is that city in?", stream=True)
+print(f"    Result: {result}")
+
+# 4. Usage tracking (accumulated locally)
+info = local.retrieve_session()
+print(f"\n[4] Usage: in={info['usage']['input_tokens']}, out={info['usage']['output_tokens']}")
+
+# 5. Tool execution (runs in local subprocess)
+print("\n[5] Tool execution example...")
+result = agent.start("Create a file called hello_ollama.txt with 'Hello from Ollama agent!'", stream=True)
+print(f"    Tool result: {result}")
+
+print("\nDone! Agent loop ran locally with Ollama LLM calls.")

examples/python/managed-agents/provider/runtime_local_openai.py

@@ -0,0 +1,61 @@
+"""Local agent loop with OpenAI LLM — runs locally, not in a managed runtime.
+
+Uses the new canonical LocalAgent class which clearly communicates that only the
+agent loop runs locally. The LLM calls go to OpenAI, but there's no managed runtime involved.
+"""
+import os
+import sys
+
+# Skip guards - exit cleanly if prerequisites not met
+if not os.getenv("OPENAI_API_KEY"):
+    print("[skip] OPENAI_API_KEY not set")
+    sys.exit(0)
+
+try:
+    import openai
+except ImportError:
+    print("[skip] openai SDK not installed")
+    sys.exit(0)
+
+# Heavy imports only after skip-guards pass
+from praisonai import Agent
+from praisonai.integrations import LocalAgent, LocalAgentConfig
+
+# Create a local agent using OpenAI's LLM
+local = LocalAgent(
+    config=LocalAgentConfig(
+        model="gpt-4o-mini",
+        system="You are a helpful coding assistant. Be concise.",
+        name="LocalOpenAIAgent",
+        tools=["execute_command", "read_file", "write_file"],
+    ),
+)
+
+agent = Agent(name="local-openai", backend=local)
+
+# 1. Basic execution - agent loop runs locally, LLM calls go to OpenAI
+print("[1] Local execution with OpenAI LLM...")
+result = agent.start("What is the capital of France? One word.", stream=True)
+print(f"    Result: {result}")
+
+# 2. Agent metadata (locally generated UUIDs)
+print(f"\n[2] Agent ID: {local.agent_id}")
+print(f"    Version: {local.agent_version}")
+print(f"    Env ID:  {local.environment_id}")
+print(f"    Session: {local.session_id}")
+
+# 3. Multi-turn conversation (session state maintained locally)
+print("\n[3] Multi-turn conversation...")
+result = agent.start("What country is that city in?", stream=True)
+print(f"    Result: {result}")
+
+# 4. Usage tracking (accumulated locally)
+info = local.retrieve_session()
+print(f"\n[4] Usage: in={info['usage']['input_tokens']}, out={info['usage']['output_tokens']}")
+
+# 5. Tool execution (runs in local subprocess)
+print("\n[5] Tool execution example...")
+result = agent.start("Create a file called hello.txt with 'Hello from local agent!'", stream=True)
+print(f"    Tool result: {result}")
+
+print("\nDone! Agent loop ran locally with OpenAI LLM calls.")

examples/python/managed-agents/provider/runtime_modal.py

@@ -0,0 +1,61 @@
+"""Agent runtime with tools sandboxed in Modal cloud functions.
+
+The agent's LLM loop runs locally; tool / code execution is sandboxed in
+Modal's serverless cloud. Modal gives fast cold-starts and per-second billing.
+
+Requires:
+  - ``OPENAI_API_KEY`` for the LLM
+  - Modal CLI configured: ``modal token set --token-id ... --token-secret ...``
+  - ``pip install modal``
+"""
+import asyncio
+import os
+
+if not os.getenv("OPENAI_API_KEY"):
+    print("[skip] OPENAI_API_KEY not set.")
+    raise SystemExit(0)
+
+try:
+    import modal  # noqa: F401
+except ImportError:
+    print("[skip] 'modal' package not installed — pip install modal")
+    raise SystemExit(0) from None
+
+# Modal auth: presence of ~/.modal.toml or MODAL_TOKEN_ID + MODAL_TOKEN_SECRET
+if not (os.path.exists(os.path.expanduser("~/.modal.toml"))
+        or (os.getenv("MODAL_TOKEN_ID") and os.getenv("MODAL_TOKEN_SECRET"))):
+    print("[skip] Modal not configured — run `modal token set ...`")
+    raise SystemExit(0)
+
+try:
+    from praisonai import Agent
+    from praisonai.integrations import SandboxedAgent, SandboxedAgentConfig
+except ImportError as e:
+    print(f"[skip] praisonai wrapper not importable ({e}).")
+    raise SystemExit(0) from None
+
+
+async def main():
+    sandboxed = SandboxedAgent(
+        compute="modal",  # tool execution runs in Modal's cloud
+        config=SandboxedAgentConfig(
+            model="gpt-4o-mini",
+            system="You are a concise coding assistant.",
+            name="ModalRuntimeAgent",
+        ),
+    )
+    agent = Agent(name="modal-runtime", backend=sandboxed)
+
+    info = await sandboxed.provision_compute(idle_timeout_s=120)
+    print(f"[1] Modal instance: {info.instance_id} ({info.status})")
+
+    result = await sandboxed.execute_in_compute("python3 -c 'import math; print(math.factorial(10))'")
+    print(f"[2] Cloud compute result: {result['stdout'].strip()}")
+
+    print("[3] Agent:", agent.start("What is the 10th Fibonacci number? Just the number."))
+
+    await sandboxed.shutdown_compute()
+    print("[4] Modal instance shut down.")
+
+
+asyncio.run(main())

examples/python/managed-agents/provider/runtime_ollama.py

@@ -0,0 +1,45 @@
+"""Ollama managed runtime — agent runs against a local/self-hosted LLM.
+
+``ManagedAgent(provider="ollama")`` wires the managed loop to an Ollama
+server. Model runs locally (or on a self-hosted box) instead of in a public
+cloud — useful when you want the "managed agent" developer experience with
+no external API dependency.
+
+Requires: An Ollama server reachable (default ``http://localhost:11434``)
+with at least one model pulled (e.g., ``ollama pull llama3.2``).
+"""
+import os
+import urllib.error
+import urllib.request
+
+base = os.getenv("OLLAMA_HOST", "http://localhost:11434")
+try:
+    urllib.request.urlopen(f"{base}/api/tags", timeout=2).read()
+except (urllib.error.URLError, OSError) as e:
+    print(f"[skip] Ollama not reachable at {base} ({e}). Start it with `ollama serve`.")
+    raise SystemExit(0)
+
+try:
+    from praisonai import Agent, ManagedAgent, LocalManagedConfig
+except ImportError as e:
+    print(f"[skip] praisonai wrapper not importable ({e}).")
+    raise SystemExit(0) from None
+
+managed = ManagedAgent(
+    provider="ollama",
+    config=LocalManagedConfig(
+        model="llama3.2",
+        system="You are a concise assistant. One-sentence answers.",
+        name="OllamaRuntimeAgent",
+    ),
+)
+agent = Agent(name="ollama-runtime", backend=managed)
+
+# 1. Agent work (LLM call goes to your Ollama server)
+print("[1] Agent:", agent.start("What is the capital of Japan? One word."))
+
+# 2. Multi-turn reuses the session
+print("[2] Agent:", agent.start("Name its most famous mountain."))
+
+info = managed.retrieve_session()
+print(f"[3] session={managed.session_id}")

examples/python/managed-agents/provider/runtime_openai.py

@@ -0,0 +1,41 @@
+"""OpenAI managed runtime — agent loop local, LLM calls go to OpenAI cloud.
+
+``ManagedAgent(provider="openai")`` wires a local managed loop to OpenAI's
+hosted models. The LLM reasoning (the heavy-lifting) runs in OpenAI's cloud;
+the agent orchestration (tool routing, session state) runs in this process.
+
+Requires: OPENAI_API_KEY environment variable set.
+"""
+import os
+
+if not os.getenv("OPENAI_API_KEY"):
+    print("[skip] OPENAI_API_KEY not set — set it to run this example.")
+    raise SystemExit(0)
+
+try:
+    from praisonai import Agent, ManagedAgent, LocalManagedConfig
+except ImportError as e:
+    print(f"[skip] praisonai wrapper not importable ({e}).")
+    raise SystemExit(0) from None
+
+managed = ManagedAgent(
+    provider="openai",
+    config=LocalManagedConfig(
+        model="gpt-4o-mini",
+        system="You are a concise assistant. One-sentence answers.",
+        name="OpenAIRuntimeAgent",
+    ),
+)
+agent = Agent(name="openai-runtime", backend=managed)
+
+# 1. Agent work (LLM call goes to OpenAI cloud)
+print("[1] Agent:", agent.start("What is the time complexity of binary search? One sentence."))
+
+# 2. Multi-turn reuses the session
+print("[2] Agent:", agent.start("And for a linear scan?"))
+
+# 3. Local session metadata (cloud usage accumulated)
+info = managed.retrieve_session()
+print(f"[3] session={managed.session_id}  "
+      f"in={info.get('usage', {}).get('input_tokens', 0)}  "
+      f"out={info.get('usage', {}).get('output_tokens', 0)}")

examples/scientific_writing/agents.yaml

@@ -0,0 +1,40 @@
+# Scientific paper workflow — YAML equivalent of multi_agent_paper.py
+# Run with:
+#   praisonai agents run --file examples/scientific_writing/agents.yaml
+framework: praisonai
+topic: "Climate change effects on coral reef biodiversity"
+
+roles:
+  literature_reviewer:
+    role: "Literature Reviewer"
+    goal: "Survey recent academic work on {topic} and produce an APA review."
+    backstory: "Expert in academic literature surveying and citation hygiene."
+    llm: "gpt-4o-mini"
+    tasks:
+      review_literature:
+        description: "Produce a concise literature review with 5-8 APA citations."
+        expected_output: "APA-formatted literature review."
+
+  methodology_designer:
+    role: "Methodology Designer"
+    goal: "Design a reproducible research methodology for {topic}."
+    backstory: "Specialises in rigorous experimental design."
+    llm: "gpt-4o-mini"
+    tasks:
+      design_methodology:
+        description: "Write a reproducible methods section in LaTeX."
+        expected_output: "LaTeX methods section."
+
+  scientific_writer:
+    role: "Scientific Writer"
+    goal: "Assemble the final LaTeX paper on {topic}."
+    backstory: >-
+      Specialised in LaTeX-formatted academic papers, fine-tuned on
+      scientific literature (CAJAL-4B checkpoint).
+    llm: "huggingface/Agnuxo/CAJAL-4B-P2PCLAW"
+    tasks:
+      write_paper:
+        description: >-
+          Combine the literature review and methodology into a full
+          scientific paper on {topic}. Use LaTeX formatting.
+        expected_output: "Complete LaTeX paper."

examples/scientific_writing/multi_agent_paper.py

@@ -0,0 +1,73 @@
+"""Multi-agent scientific paper workflow (Researcher -> Methodologist -> Writer).
+
+Demonstrates the agent-centric pattern from AGENTS.md: compose specialised
+Agents with ``AgentTeam`` + sequential ``Task``s instead of creating custom
+Agent subclasses.  The writer uses the CAJAL-4B HuggingFace model; the
+reviewer and methodologist use general-purpose LLMs.
+
+Usage:
+    python examples/scientific_writing/multi_agent_paper.py
+"""
+from praisonaiagents import Agent, AgentTeam, Task
+
+from scientific_writer import (  # noqa: E402 — example imports a sibling file
+    build_scientific_writer,
+    format_citation,
+    format_latex_section,
+)
+
+
+literature_reviewer = Agent(
+    name="Literature Reviewer",
+    instructions=(
+        "You are an expert at surveying academic literature. "
+        "Produce a concise literature review with 5-8 key citations in APA."
+    ),
+    tools=[format_citation],
+)
+
+methodology_designer = Agent(
+    name="Methodology Designer",
+    instructions=(
+        "You design rigorous research methodologies. "
+        "Output a clear, reproducible methods section."
+    ),
+    tools=[format_latex_section],
+)
+
+scientific_writer = build_scientific_writer()
+
+
+def run(topic: str) -> str:
+    """Run the 3-agent workflow for a given research ``topic``."""
+    t1 = Task(
+        name="review_literature",
+        description=f"Review literature on: {topic}",
+        agent=literature_reviewer,
+        expected_output="APA literature review.",
+    )
+    t2 = Task(
+        name="design_methodology",
+        description=f"Design research methodology for: {topic}",
+        agent=methodology_designer,
+        expected_output="LaTeX methods section.",
+    )
+    t3 = Task(
+        name="write_paper",
+        description=(
+            f"Combine the literature review and methodology into a full "
+            f"scientific paper on: {topic}. Use LaTeX formatting."
+        ),
+        agent=scientific_writer,
+        expected_output="Complete LaTeX paper.",
+    )
+    team = AgentTeam(
+        agents=[literature_reviewer, methodology_designer, scientific_writer],
+        tasks=[t1, t2, t3],
+    )
+    return team.start()
+
+
+if __name__ == "__main__":
+    result = run("Transformer architectures for protein folding prediction")
+    print(result)

examples/scientific_writing/scientific_writer.py

@@ -0,0 +1,85 @@
+"""Scientific Paper Writer — single-agent example using the CAJAL model.
+
+CAJAL (``Agnuxo/CAJAL-4B-P2PCLAW``) is a 4B-parameter HuggingFace model
+specialised for LaTeX-formatted academic writing.  This example shows how
+to use it with a regular ``Agent`` and ``@tool`` functions — no new Agent
+subclass is needed.  Run 3 ways: Python (this file), CLI, or YAML.
+
+Requirements:
+    pip install "praisonaiagents[llm]"          # for litellm / HuggingFace
+    export HUGGINGFACE_API_KEY=...              # or run a local inference server
+
+Usage:
+    python examples/scientific_writing/scientific_writer.py
+"""
+from praisonaiagents import Agent, tool
+
+
+@tool
+def format_latex_section(title: str, content: str) -> str:
+    """Wrap prose content in a LaTeX ``\\section{}`` block.
+
+    Args:
+        title: Section heading (e.g. ``"Introduction"``).
+        content: The body text.
+
+    Returns:
+        LaTeX-formatted section string.
+    """
+    return f"\\section{{{title}}}\n{content}\n"
+
+
+@tool
+def format_citation(authors: str, year: int, title: str, venue: str) -> str:
+    """Render a single APA-style citation line.
+
+    Args:
+        authors: ``"Smith, J. & Jones, A."``
+        year: Publication year.
+        title: Paper title.
+        venue: Journal / conference name.
+
+    Returns:
+        One APA-formatted citation string.
+    """
+    return f"{authors} ({year}). {title}. *{venue}*."
+
+
+SCIENTIFIC_INSTRUCTIONS = """You are a specialised scientific paper writer.
+
+Produce high-quality academic content with:
+- Clear structure (Abstract, Introduction, Methodology, Results, Discussion,
+  Conclusion, References)
+- Rigorous methodology and precise language
+- Proper APA citations and LaTeX formatting
+- Honest limitations and future work sections
+
+Use ``format_latex_section`` to wrap each section in LaTeX, and
+``format_citation`` to build the References list.
+"""
+
+
+def build_scientific_writer(
+    model: str = "huggingface/Agnuxo/CAJAL-4B-P2PCLAW",
+) -> Agent:
+    """Construct a scientific-writing Agent.
+
+    Override ``model`` for any HuggingFace / local / API model — the default
+    targets the CAJAL-4B checkpoint for LaTeX-style academic output.
+    """
+    return Agent(
+        name="Scientific Writer",
+        instructions=SCIENTIFIC_INSTRUCTIONS,
+        llm=model,
+        tools=[format_latex_section, format_citation],
+    )
+
+
+if __name__ == "__main__":
+    agent = build_scientific_writer()
+    paper = agent.start(
+        "Write a short scientific paper (Abstract + Introduction + Conclusion) "
+        "on 'Climate change effects on coral reef biodiversity'. "
+        "Use LaTeX formatting and include at least 3 APA citations."
+    )
+    print(paper)

examples/serve/mcp_http_server.py

@@ -44,12 +44,33 @@ def main():
     
     def calculate(expression: str) -> str:
         """Calculate a math expression safely."""
-        try:
-            # Safe eval for basic math
-            allowed = set("0123456789+-*/.() ")
-            if all(c in allowed for c in expression):
-                return str(eval(expression))
+        import ast
+
+        allowed = set("0123456789+-*/.() ")
+        if not all(c in allowed for c in expression):
             return "Error: Invalid expression"
+        try:
+            tree = ast.parse(expression, mode="eval")
+            for node in ast.walk(tree):
+                if not isinstance(
+                    node,
+                    (
+                        ast.Expression,
+                        ast.BinOp,
+                        ast.UnaryOp,
+                        ast.Constant,
+                        ast.Add,
+                        ast.Sub,
+                        ast.Mult,
+                        ast.Div,
+                        ast.USub,
+                        ast.UAdd,
+                    ),
+                ):
+                    return "Error: Invalid expression"
+            return str(
+                eval(compile(tree, "<expr>", "eval"), {"__builtins__": {}})
+            )
         except Exception as e:
             return f"Error: {e}"
     

examples/yaml/cli_backend.yaml

@@ -0,0 +1,12 @@
+framework: praisonai
+topic: coding
+roles:
+  coder:
+    role: Code refactorer
+    goal: Refactor Python modules for better maintainability
+    backstory: Senior engineer with expertise in clean code principles
+    cli_backend: claude-code
+    tasks:
+      refactor:
+        description: Refactor utils.py to improve code quality
+        expected_output: Refactored code with improved structure and documentation

examples/yaml/linear-bot.yaml

@@ -0,0 +1,35 @@
+platform: linear
+token: ${LINEAR_OAUTH_TOKEN}
+signing_secret: ${LINEAR_WEBHOOK_SECRET}
+webhook_port: 8080
+
+agent:
+  name: "PraisonAI Coder"
+  instructions: |
+    You are an autonomous coding agent integrated with Linear.
+    
+    When you are mentioned or assigned an issue:
+    1. Analyze the issue description and requirements
+    2. Break down complex tasks into smaller steps
+    3. Use available tools to implement solutions
+    4. Update the issue with progress comments
+    5. Create GitHub pull requests when code changes are needed
+    
+    Focus on being helpful, thorough, and providing clear updates.
+  llm: gpt-4o-mini
+  tools:
+    - linear_search_issues
+    - linear_get_issue
+    - linear_add_comment
+    - linear_update_issue
+    - linear_list_teams
+    - linear_list_issue_states
+    - github_create_branch
+    - github_commit_and_push
+    - github_create_pull_request
+    - read_file
+    - write_file
+    - execute_command
+  memory: true
+  web_search: true
+  auto_approve_tools: true

scripts/security_publish_advisories.sh

@@ -0,0 +1,57 @@
+#!/usr/bin/env bash
+# Publish security advisories after PyPI release.
+# Usage: AGENTS_VERSION=1.6.40 WRAPPER_VERSION=4.6.40 ./scripts/security_publish_advisories.sh
+set -euo pipefail
+
+REPO="MervinPraison/PraisonAI"
+AGENTS_VERSION="${AGENTS_VERSION:?Set AGENTS_VERSION (e.g. 1.6.40)}"
+WRAPPER_VERSION="${WRAPPER_VERSION:?Set WRAPPER_VERSION (e.g. 4.6.40)}"
+LAST_AGENTS="${LAST_AGENTS:-1.6.39}"
+LAST_WRAPPER="${LAST_WRAPPER:-4.6.39}"
+
+# Batch 1 + 2 GHSA IDs (update credits login per advisory via gh api first if needed)
+AGENTS_GHSAS=(
+  GHSA-4mr5-g6f9-cfrh
+  GHSA-5c6w-wwfq-7qqm
+  GHSA-5cxw-77wg-jrf3
+)
+
+WRAPPER_GHSAS=(
+  GHSA-9cr9-25q5-8prj
+  GHSA-86qc-r5v2-v6x6
+  GHSA-hvhp-v2gc-268q
+  GHSA-xp85-6wwf-r67c
+  GHSA-8444-4fhq-fxpq
+  GHSA-78r8-wwqv-r299
+  GHSA-vg22-4gmj-prxw
+  GHSA-6xj3-927j-6pqw
+)
+
+patch_advisory() {
+  local ghsa="$1" package="$2" last="$3" patched="$4"
+  gh api "repos/${REPO}/security-advisories/${ghsa}" --method PATCH --input - <<EOF
+{
+  "state": "draft",
+  "vulnerabilities": [
+    {
+      "package": {"ecosystem": "pip", "name": "${package}"},
+      "vulnerable_version_range": "<= ${last}",
+      "patched_versions": ">= ${patched}"
+    }
+  ]
+}
+EOF
+  gh api "repos/${REPO}/security-advisories/${ghsa}" --method PATCH -f state=published
+  gh api "repos/${REPO}/security-advisories/${ghsa}/cve" --method POST || true
+  echo "Published ${ghsa}"
+}
+
+for ghsa in "${AGENTS_GHSAS[@]}"; do
+  patch_advisory "$ghsa" "praisonaiagents" "$LAST_AGENTS" "$AGENTS_VERSION"
+done
+
+for ghsa in "${WRAPPER_GHSAS[@]}"; do
+  patch_advisory "$ghsa" "praisonai" "$LAST_WRAPPER" "$WRAPPER_VERSION"
+done
+
+echo "Done. Verify: gh api repos/${REPO}/security-advisories --jq '.[] | select(.state==\"published\") | .ghsa_id'"

src/praisonai-agents/praisonaiagents/__init__.py

@@ -202,6 +202,13 @@ _LAZY_IMPORTS = {
     'FlowDisplay': ('praisonaiagents.flow_display', 'FlowDisplay'),
     'track_workflow': ('praisonaiagents.flow_display', 'track_workflow'),
     
+    # CLI Backend Protocol (Protocol-driven external CLI integration)
+    'CliBackendProtocol': ('praisonaiagents.cli_backend.protocols', 'CliBackendProtocol'),
+    'CliBackendConfig': ('praisonaiagents.cli_backend.protocols', 'CliBackendConfig'),
+    'CliSessionBinding': ('praisonaiagents.cli_backend.protocols', 'CliSessionBinding'),
+    'CliBackendResult': ('praisonaiagents.cli_backend.protocols', 'CliBackendResult'),
+    'CliBackendDelta': ('praisonaiagents.cli_backend.protocols', 'CliBackendDelta'),
+    
     # Agent classes
     'Agent': ('praisonaiagents.agent.agent', 'Agent'),
     'BudgetExceededError': ('praisonaiagents.errors', 'BudgetExceededError'),

src/praisonai-agents/praisonaiagents/agent/__init__.py

@@ -33,6 +33,10 @@ def __getattr__(name):
         from .heartbeat import HeartbeatConfig
         _lazy_cache[name] = HeartbeatConfig
         return HeartbeatConfig
+    if name == 'InterruptController':
+        from .interrupt import InterruptController
+        _lazy_cache[name] = InterruptController
+        return InterruptController
     
     # Specialized agents - lazy loaded (import rich)
     if name == 'ImageAgent':
@@ -194,6 +198,7 @@ __all__ = [
     'BudgetExceededError',
     'Heartbeat',
     'HeartbeatConfig',
+    'InterruptController',
     'ImageAgent',
     'VideoAgent',
     'VideoConfig',

src/praisonai-agents/praisonaiagents/agent/agent.py

@@ -179,13 +179,15 @@ class ServerRegistry:
         self._registered_agents = {}  # Dict of port -> Dict of path -> agent_id  
         self._shared_apps = {}  # Dict of port -> FastAPI app
     
+    # Class-level lock for thread-safe singleton creation
+    _instance_lock = threading.Lock()
+    
     @staticmethod
     def get_default_instance():
         """Get default global registry for backward compatibility."""
         if not hasattr(ServerRegistry, '_default_instance'):
-            import threading
-            # Double-checked locking pattern for thread safety
-            with threading.Lock():
+            # Double-checked locking pattern with shared class-level lock
+            with ServerRegistry._instance_lock:
                 if not hasattr(ServerRegistry, '_default_instance'):
                     ServerRegistry._default_instance = ServerRegistry()
         return ServerRegistry._default_instance
@@ -243,6 +245,7 @@ if TYPE_CHECKING:
     from ..context.models import ContextConfig
     from ..context.manager import ContextManager
     from ..knowledge.knowledge import Knowledge
+    from .interrupt import InterruptController
     from ..agent.autonomy import AutonomyConfig
     from ..task.task import Task
     from .handoff import Handoff, HandoffConfig, HandoffResult
@@ -516,6 +519,7 @@ class Agent(UnifiedExecutionMixin, ToolExecutionMixin, ChatHandlerMixin, Session
         model: Optional[Union[str, Any]] = None,  # Alias for llm=
         base_url: Optional[str] = None,  # Kept separate (connection/auth)
         api_key: Optional[str] = None,  # Kept separate (connection/auth)
+        auth: Optional[str] = None,  # Subscription auth provider: "claude-code", "codex", etc.
         # Tools
         tools: Optional[List[Any]] = None,
         allow_delegation: bool = False,  # Deprecated: use handoffs= instead
@@ -529,7 +533,7 @@ class Agent(UnifiedExecutionMixin, ToolExecutionMixin, ChatHandlerMixin, Session
         # CONSOLIDATED FEATURE PARAMS (agent-centric API)
         # Each follows: False=disabled, True=defaults, Config=custom
         # ============================================================
-        memory: Optional[Union[bool, str, 'MemoryConfig', 'MemoryManager']] = None,
+        memory: Optional[Union[bool, str, 'MemoryConfig', Any]] = None,
         knowledge: Optional[Union[bool, str, List[str], 'KnowledgeConfig', 'Knowledge']] = None,
         planning: Optional[Union[bool, str, 'PlanningConfig']] = False,
         reflection: Optional[Union[bool, str, 'ReflectionConfig']] = None,
@@ -549,6 +553,8 @@ class Agent(UnifiedExecutionMixin, ToolExecutionMixin, ChatHandlerMixin, Session
         parallel_tool_calls: bool = False,  # Gap 2: Enable parallel execution of batched LLM tool calls
         learn: Optional[Union[bool, str, Dict[str, Any], 'LearnConfig']] = None,  # Continuous learning (peer to memory)
         backend: Optional[Any] = None,  # External managed agent backend (e.g., ManagedAgentIntegration)
+        cli_backend: Optional[Union[str, Any]] = None,  # CLI backend for delegating turns (e.g., "claude-code")
+        interrupt_controller: Optional['InterruptController'] = None,  # G2: Cooperative cancellation
     ):
         """Initialize an Agent instance.
 
@@ -573,7 +579,7 @@ class Agent(UnifiedExecutionMixin, ToolExecutionMixin, ChatHandlerMixin, Session
             memory: Memory system configuration. Accepts:
                 - bool: True enables defaults, False disables
                 - MemoryConfig: Custom configuration
-                - MemoryManager: Pre-configured instance
+                - Any: Pre-configured memory instance
             knowledge: Knowledge sources. Accepts:
                 - bool: True enables defaults
                 - List[str]: File paths, URLs, or text content
@@ -645,6 +651,13 @@ class Agent(UnifiedExecutionMixin, ToolExecutionMixin, ChatHandlerMixin, Session
                 - None: Use local execution (default)
                 When provided, agent can delegate execution to managed infrastructure
                 for long-running tasks or when local resources are constrained.
+            cli_backend: CLI backend for delegating full turns to external CLI tools. Accepts:
+                - str: Backend ID ("claude-code", "codex-cli", "gemini-cli")
+                - CliBackendProtocol: Custom CLI backend instance
+                - None: Use standard LLM execution (default)
+                When provided, agent delegates entire conversation turns to the CLI tool
+                instead of using the built-in LLM. Enables session continuity and 
+                tool integration through external AI coding assistants.
 
         Raises:
             ValueError: If all of name, role, goal, backstory, and instructions are None.
@@ -1456,6 +1469,8 @@ class Agent(UnifiedExecutionMixin, ToolExecutionMixin, ChatHandlerMixin, Session
         self.instructions = instructions
         # Gap 2: Store parallel tool calls setting for ToolCallExecutor selection
         self.parallel_tool_calls = parallel_tool_calls
+        # G2: Store interrupt controller for cooperative cancellation
+        self.interrupt_controller = interrupt_controller
         # Check for model name in environment variable if not provided
         self._using_custom_llm = False
         # Flag to track if final result has been displayed to prevent duplicates
@@ -1507,6 +1522,8 @@ class Agent(UnifiedExecutionMixin, ToolExecutionMixin, ChatHandlerMixin, Session
                     llm_config['base_url'] = base_url
                     if api_key:
                         llm_config['api_key'] = api_key
+                    if auth:
+                        llm_config['auth'] = auth
                     llm_config['metrics'] = metrics
                     self.llm_instance = LLM(**llm_config)
                     self.llm = llm.get('model', Agent._get_default_model())
@@ -1517,6 +1534,7 @@ class Agent(UnifiedExecutionMixin, ToolExecutionMixin, ChatHandlerMixin, Session
                         model=model_name,
                         base_url=base_url,
                         api_key=api_key,
+                        auth=auth,
                         metrics=metrics,
                         web_search=web_search,
                         web_fetch=web_fetch,
@@ -1538,6 +1556,10 @@ class Agent(UnifiedExecutionMixin, ToolExecutionMixin, ChatHandlerMixin, Session
                 if api_key and 'api_key' not in llm:
                     llm = llm.copy()
                     llm['api_key'] = api_key
+                # Add auth if provided and not in dict
+                if auth and 'auth' not in llm:
+                    llm = llm.copy()
+                    llm['auth'] = auth
                 # Add metrics parameter
                 llm = llm.copy()
                 llm['metrics'] = metrics
@@ -1557,6 +1579,8 @@ class Agent(UnifiedExecutionMixin, ToolExecutionMixin, ChatHandlerMixin, Session
                 llm_params = {'model': llm}
                 if api_key:
                     llm_params['api_key'] = api_key
+                if auth:
+                    llm_params['auth'] = auth
                 llm_params['metrics'] = metrics
                 llm_params['web_search'] = web_search
                 llm_params['web_fetch'] = web_fetch
@@ -1691,6 +1715,9 @@ Your Goal: {self.goal}
         self.prompt_caching = prompt_caching
         self.claude_memory = claude_memory
         
+        # Initialize closure flag for GC cleanup
+        self._closed = False
+        
         # Session management
         self.auto_save = auto_save  # Session name for auto-saving
         
@@ -1754,6 +1781,31 @@ Your Goal: {self.goal}
             self._approval_backend = None
             self._approve_all_tools = False
             self._approval_timeout = 0
+            # No explicit approval kwarg — honour PRAISONAI_TOOL_SAFETY.
+            # Default preset "default" blocks only destructive ops
+            # (delete_*, execute_command, execute_code, kill_process, move/copy)
+            # while leaving read / create / edit tools fully auto-approved.
+            # Users who want the pre-4.6.27 "trust everything" behaviour
+            # export PRAISONAI_TOOL_SAFETY=off. This adds zero Agent kwargs.
+            _raw_safety_env = os.environ.get("PRAISONAI_TOOL_SAFETY")
+            _safety_env = (_raw_safety_env or "").strip().lower()
+            if _safety_env not in ("off", "full", "none", "0", "false"):
+                from ..approval.registry import PERMISSION_PRESETS
+                _resolved_safety_env = _safety_env or "default"
+                _preset_deny = PERMISSION_PRESETS.get(_resolved_safety_env)
+                if _preset_deny is None and _safety_env:
+                    # Unknown env value - fall back to safe default and log warning
+                    # (logging is already imported at module level; a local `import logging`
+                    # here would shadow it and cause UnboundLocalError on earlier uses
+                    # inside __init__ such as the custom-LLM tools branch.)
+                    logging.getLogger(__name__).warning(
+                        "Unknown PRAISONAI_TOOL_SAFETY value %r; falling back to 'default' preset.",
+                        _raw_safety_env,
+                    )
+                    _resolved_safety_env = "default"
+                    _preset_deny = PERMISSION_PRESETS.get(_resolved_safety_env)
+                if _preset_deny is not None:
+                    self._perm_deny = _preset_deny
         elif isinstance(approval, ApprovalConfig):
             self._approval_backend = approval.backend
             self._approve_all_tools = approval.all_tools
@@ -1779,6 +1831,7 @@ Your Goal: {self.goal}
             self._approval_backend = AutoApproveBackend()
         # Pending approvals for async (non-blocking) mode
         self._pending_approvals = {}
+        self._approvals_lock = asyncio.Lock()
         
         # P8/G11: Tool timeout - prevent slow tools from blocking
         self._tool_timeout = tool_timeout
@@ -1884,6 +1937,11 @@ Your Goal: {self.goal}
 
         # Backend - external managed agent backend for hybrid execution
         self.backend = backend
+        
+        # CLI Backend - external CLI backend for delegating full turns
+        self._cli_backend = None
+        if cli_backend is not None:
+            self._cli_backend = self._resolve_cli_backend(cli_backend)
 
         # Telemetry - lazy initialized via property for performance
         self.__telemetry = None
@@ -1908,13 +1966,29 @@ Your Goal: {self.goal}
     
     @chat_history.setter
     def chat_history(self, value):
-        """Set chat history (updates the underlying async-safe state)."""
-        self.__chat_history_state.value = value
+        """Set chat history (updates the underlying async-safe state with lock)."""
+        with self.__chat_history_state.lock():
+            self.__chat_history_state.value = value
     
     @property
     def _history_lock(self):
         """Get appropriate lock for chat history based on execution context."""
         return self.__chat_history_state
+    
+    def _append_to_chat_history(self, message: dict):
+        """Thread-safe append to chat history using proper locking."""
+        with self._history_lock.lock():
+            self._history_lock.value.append(message)
+    
+    def _truncate_chat_history(self, length: int):
+        """Thread-safe truncation of chat history using proper locking."""
+        with self._history_lock.lock():
+            self._history_lock.value[:] = self._history_lock.value[:length]
+    
+    def _replace_chat_history(self, new_history: List[Dict[str, Any]]):
+        """Thread-safe replacement of entire chat history using proper locking."""
+        with self._history_lock.lock():
+            self._history_lock.value[:] = new_history
 
     @property
     def _cache_lock(self):
@@ -2255,9 +2329,11 @@ Summary:"""
         Uses lazy imports from praisonaiagents.tools to avoid performance impact
         when skills are not used.
 
-        Honours ``PRAISONAI_DISABLE_SKILL_TOOLS=1`` so hosts that do not want
-        the subprocess-backed ``run_skill_script`` tool auto-injected can opt
-        out without touching code.
+        G-E fix: run_skill_script is now safer by default:
+        - PRAISONAI_DISABLE_SKILL_TOOLS=1: disables all skill tools (explicit deny wins)
+        - PRAISONAI_ENABLE_SKILL_TOOLS=1: enables run_skill_script by default
+        - Any loaded skill with 'run_skill_script' in allowed-tools: enables it
+        - Otherwise: only read_file is added (safer default)
         """
         import os as _os
         if _os.environ.get("PRAISONAI_DISABLE_SKILL_TOOLS") in ("1", "true", "True"):
@@ -2272,7 +2348,7 @@ Summary:"""
             elif hasattr(tool, 'name'):
                 tool_names.add(tool.name)
         
-        # Add read_file if not present
+        # Add read_file if not present (low risk, always enabled)
         if 'read_file' not in tool_names:
             try:
                 from ..tools import read_file
@@ -2281,8 +2357,26 @@ Summary:"""
             except ImportError:
                 logging.warning("Could not import read_file tool for skills")
         
-        # Add run_skill_script from skill_tools module
-        if 'run_skill_script' not in tool_names:
+        # G-E fix: run_skill_script safer by default
+        # Only add if explicitly enabled OR any skill declares it in allowed-tools
+        should_add_script_tool = False
+        
+        # Check explicit environment enable
+        if _os.environ.get("PRAISONAI_ENABLE_SKILL_TOOLS") in ("1", "true", "True"):
+            should_add_script_tool = True
+            logging.debug("run_skill_script enabled via PRAISONAI_ENABLE_SKILL_TOOLS")
+        
+        # Check if any loaded skill declares it in allowed-tools
+        if self._skill_manager and not should_add_script_tool:
+            for skill in self._skill_manager.skills:
+                allowed_tools = self._skill_manager.get_allowed_tools(skill.properties.name)
+                if "run_skill_script" in allowed_tools:
+                    should_add_script_tool = True
+                    logging.debug(f"run_skill_script enabled by skill '{skill.properties.name}' allowed-tools")
+                    break
+        
+        # Add run_skill_script if conditions are met
+        if should_add_script_tool and 'run_skill_script' not in tool_names:
             try:
                 from ..tools.skill_tools import create_skill_tools
                 # Create skill tools with current working directory
@@ -2813,6 +2907,19 @@ Summary:"""
                         started_at=started_at,
                     )
                 
+                # G2: Check for interrupt request (cooperative cancellation) - sync version
+                if self.interrupt_controller and self.interrupt_controller.is_set():
+                    reason = self.interrupt_controller.reason or "unknown"
+                    return AutonomyResult(
+                        success=False,
+                        output=f"Task interrupted: {reason}",
+                        completion_reason="interrupted",
+                        iterations=iterations,
+                        stage=stage,
+                        actions=actions_taken,
+                        duration_seconds=time_module.time() - start_time,
+                        started_at=started_at,
+                    )
                 
                 # Execute one turn using the agent's chat method
                 # Always use the original prompt (prompt re-injection)
@@ -3210,6 +3317,20 @@ Summary:"""
                         started_at=started_at,
                     )
                 
+                # G2: Check for interrupt request (cooperative cancellation)
+                if self.interrupt_controller and self.interrupt_controller.is_set():
+                    reason = self.interrupt_controller.reason or "unknown"
+                    return AutonomyResult(
+                        success=False,
+                        output=f"Task interrupted: {reason}",
+                        completion_reason="interrupted",
+                        iterations=iterations,
+                        stage=stage,
+                        actions=actions_taken,
+                        duration_seconds=time_module.time() - start_time,
+                        started_at=started_at,
+                    )
+                
                 
                 # Execute one turn using the agent's async chat method
                 # Always use the original prompt (prompt re-injection)
@@ -4582,9 +4703,141 @@ Answer:"""
     # -------------------------------------------------------------------------
     
     
-    @contextlib.contextmanager
+    # -------------------------------------------------------------------------
+    #                       CLI Backend Management
+    # -------------------------------------------------------------------------
+    
+    def _resolve_cli_backend(self, cli_backend):
+        """Validate and return CLI backend protocol instance.
+        
+        Args:
+            cli_backend: CliBackendProtocol instance or callable that returns one
+            
+        Returns:
+            CliBackendProtocol instance
+            
+        Note:
+            String backend IDs must be resolved to instances in the wrapper layer
+            before passing to Agent. This maintains proper dependency direction
+            per AGENTS.md (core SDK should not import from wrapper).
+        """
+        # Import protocols
+        try:
+            from ..cli_backend import CliBackendProtocol
+        except ImportError:
+            raise ImportError(
+                "CLI backend features requested but protocols not available. "
+                "This should not happen in a properly installed praisonaiagents package."
+            )
+        
+        # If already a protocol instance, return as-is
+        if hasattr(cli_backend, 'execute') and hasattr(cli_backend, 'stream'):
+            return cli_backend
+            
+        # If callable (factory function), call it to get instance
+        if callable(cli_backend):
+            instance = cli_backend()
+            if hasattr(instance, 'execute') and hasattr(instance, 'stream'):
+                return instance
+            raise TypeError(f"CLI backend factory returned invalid type: {type(instance)}. Expected CliBackendProtocol.")
+        
+        # String IDs are no longer supported at core level - must be resolved in wrapper
+        if isinstance(cli_backend, str):
+            raise TypeError(
+                f"String CLI backend IDs ('{cli_backend}') must be resolved to instances "
+                "in the wrapper layer. Use: praisonai.Agent(cli_backend='claude-code') "
+                "or manually resolve: agent = praisonaiagents.Agent(cli_backend=resolve_cli_backend('claude-code'))"
+            )
+        
+        raise TypeError(f"Invalid cli_backend type: {type(cli_backend)}. Expected CliBackendProtocol instance or factory callable.")
+    
+    async def _chat_via_cli_backend(self, prompt: str, **kwargs) -> Optional[str]:
+        """Chat implementation using CLI backend delegation.
+        
+        Args:
+            prompt: User prompt
+            **kwargs: Additional chat parameters (passed through as metadata)
+            
+        Returns:
+            CLI backend response content
+        """
+        if not self._cli_backend:
+            raise RuntimeError("CLI backend not configured")
+        
+        try:
+            # Import backend types
+            from ..cli_backend import CliSessionBinding
+            
+            # Build session binding for state management
+            session_id = getattr(self, '_session_id', None) or f"agent-{self.agent_id}"
+            session_binding = CliSessionBinding(session_id=session_id)
+            
+            # Get system prompt from agent configuration
+            system_prompt = None
+            if hasattr(self, 'system_prompt') and self.system_prompt:
+                system_prompt = self.system_prompt
+            elif self.backstory or self.role or self.goal:
+                # Build system prompt from agent identity
+                parts = []
+                if self.backstory:
+                    parts.append(self.backstory)
+                if self.role:
+                    parts.append(f"Your Role: {self.role}")
+                if self.goal:
+                    parts.append(f"Your Goal: {self.goal}")
+                system_prompt = "\n".join(parts)
+            
+            # Extract images from attachments (if any)
+            images = kwargs.get('attachments')
+            if images:
+                # Filter for image files
+                image_extensions = {'.png', '.jpg', '.jpeg', '.gif', '.bmp', '.webp'}
+                images = [
+                    img for img in images 
+                    if any(img.lower().endswith(ext) for ext in image_extensions)
+                ]
+                if not images:
+                    images = None
+            
+            # Execute CLI backend
+            result = await self._cli_backend.execute(
+                prompt=prompt,
+                session=session_binding,
+                images=images,
+                system_prompt=system_prompt
+            )
+            
+            # Check for CLI backend errors
+            if result is None:
+                raise RuntimeError(
+                    f"CLI backend returned no result for agent={self.display_name!r}, "
+                    f"session_id={session_id!r}"
+                )
+            if getattr(result, "error", None):
+                raise RuntimeError(
+                    f"CLI backend failed for agent={self.display_name!r}, "
+                    f"session_id={session_id!r}: {result.error}"
+                )
+            
+            # Update chat history with the exchange
+            if hasattr(self, '_append_to_chat_history'):
+                self._append_to_chat_history({
+                    "role": "user", 
+                    "content": prompt
+                })
+                if result.content:
+                    self._append_to_chat_history({
+                        "role": "assistant", 
+                        "content": result.content
+                    })
+            
+            return result.content if result else None
+            
+        except Exception as e:
+            raise RuntimeError(
+                f"CLI backend execution failed for agent={self.display_name!r}: {e}"
+            ) from e
     
-
     # -------------------------------------------------------------------------
     #                       Resource Lifecycle Management
     # -------------------------------------------------------------------------
@@ -4609,10 +4862,10 @@ Answer:"""
                 if hasattr(self.llm_instance, 'aclose'):
                     # Try async close first
                     try:
-                        import asyncio
                         if asyncio.iscoroutinefunction(self.llm_instance.aclose):
-                            # We're in sync context, so use asyncio.run() for the cleanup
-                            asyncio.run(self.llm_instance.aclose())
+                            # Use async bridge to safely close from any context
+                            from ..utils.async_bridge import run_coroutine_from_any_context
+                            run_coroutine_from_any_context(self.llm_instance.aclose())
                         else:
                             self.llm_instance.aclose()
                     except Exception:
@@ -4659,6 +4912,19 @@ Answer:"""
         except Exception as e:
             logger.warning(f"Task cleanup failed: {e}")
 
+        # ThreadPoolExecutor cleanup
+        try:
+            if hasattr(self, '_tool_executor') and self._tool_executor:
+                # Use cancel_futures only if supported (Python 3.9+)
+                import sys
+                if sys.version_info >= (3, 9):
+                    self._tool_executor.shutdown(wait=False, cancel_futures=True)
+                else:
+                    self._tool_executor.shutdown(wait=False)
+                delattr(self, '_tool_executor')
+        except Exception as e:
+            logger.warning(f"ThreadPoolExecutor cleanup failed: {e}")
+
         # Always set closed flag
         self._closed = True
     
@@ -4693,6 +4959,23 @@ Answer:"""
                         await task
                     except asyncio.CancelledError:
                         pass
+
+            # ThreadPoolExecutor cleanup (async-safe)
+            if hasattr(self, '_tool_executor') and self._tool_executor:
+                import sys
+                loop = asyncio.get_running_loop()
+                # Use run_in_executor to avoid blocking the event loop
+                if sys.version_info >= (3, 9):
+                    await loop.run_in_executor(
+                        None, 
+                        lambda: self._tool_executor.shutdown(wait=False, cancel_futures=True)
+                    )
+                else:
+                    await loop.run_in_executor(
+                        None, 
+                        lambda: self._tool_executor.shutdown(wait=False)
+                    )
+                delattr(self, '_tool_executor')
             
             self._closed = True
             
@@ -4732,8 +5015,21 @@ Answer:"""
         await self.aclose()
     
     def __del__(self):
-        """Destructor safely does nothing to avoid GC pollution in test loops."""
-        pass
+        """Lightweight cleanup that only closes resources if not already closed."""
+        if not getattr(self, '_closed', False):
+            # Only close connections, skip anything that could fail during GC
+            try:
+                memory = getattr(self, "_memory_instance", None)
+                if memory and hasattr(memory, 'close_connections'):
+                    memory.close_connections()
+            except Exception as exc:  # noqa: BLE001 - finalizers must not raise
+                import contextlib
+                with contextlib.suppress(Exception):
+                    import logging
+                    logger = logging.getLogger(__name__)
+                    logger.debug("Agent GC memory cleanup failed: %s", exc)
+            finally:
+                self._closed = True
         
     @property
     def is_closed(self) -> bool:

src/praisonai-agents/praisonaiagents/agent/async_safety.py

@@ -8,6 +8,7 @@ import asyncio
 import threading
 from typing import Any, Optional, Union
 from contextlib import contextmanager, asynccontextmanager
+from weakref import WeakKeyDictionary
 
 
 class DualLock:
@@ -28,15 +29,16 @@ class DualLock:
             pass
             
         # In async context  
-        async with lock.async():
+        async with lock.async_lock():
             # Uses asyncio.Lock
             pass
         ```
     """
     
     def __init__(self):
-        """Initialize with unified thread-safe locking."""
-        self._thread_lock = threading.Lock()  # Single canonical lock for all contexts
+        """Initialize with separate threading and asyncio locks."""
+        self._thread_lock = threading.RLock()  # Re-entrant lock to handle nested acquisitions
+        self._async_locks = WeakKeyDictionary()  # Per-event-loop async locks
     
     @contextmanager
     def sync(self):
@@ -44,15 +46,27 @@ class DualLock:
         with self._thread_lock:
             yield
             
+    def _get_async_lock(self):
+        """Get or create an asyncio.Lock for the current event loop."""
+        try:
+            loop = asyncio.get_running_loop()
+        except RuntimeError:
+            raise RuntimeError("async_lock() must be called from an async context")
+        
+        if loop not in self._async_locks:
+            self._async_locks[loop] = asyncio.Lock()
+        return self._async_locks[loop]
+
     @asynccontextmanager
     async def async_lock(self):
-        """Acquire lock in asynchronous context using threading.Lock via asyncio.to_thread()."""
-        # Use asyncio.to_thread to acquire the thread lock without blocking the event loop
-        await asyncio.to_thread(self._thread_lock.acquire)
-        try:
+        """Acquire lock in asynchronous context using asyncio.Lock.
+        
+        Uses a per-event-loop asyncio.Lock to ensure proper async coordination
+        without blocking the event loop or violating thread ownership semantics.
+        """
+        async_lock = self._get_async_lock()
+        async with async_lock:
             yield
-        finally:
-            self._thread_lock.release()
             
     def is_async_context(self) -> bool:
         """Check if we're currently in an async context."""
@@ -114,12 +128,14 @@ class AsyncSafeState:
         
     async def __aenter__(self):
         """Support for asynchronous context manager protocol."""
-        await asyncio.to_thread(self._lock._thread_lock.acquire)
+        async_lock = self._lock._get_async_lock()
+        await async_lock.acquire()
         return self.value
         
     async def __aexit__(self, exc_type, exc_val, exc_tb):
         """Support for asynchronous context manager protocol."""
-        self._lock._thread_lock.release()
+        async_lock = self._lock._get_async_lock()
+        async_lock.release()
         return None
             
     def get(self) -> Any:

src/praisonai-agents/praisonaiagents/agent/chat_mixin.py

@@ -11,8 +11,6 @@ import time
 import json
 import logging
 from praisonaiagents._logging import get_logger
-import asyncio
-import threading
 from ..errors import BudgetExceededError
 
 # Fallback helpers to avoid circular imports
@@ -45,9 +43,7 @@ logger = logging.getLogger(__name__)
 
 
 
-import traceback
-from typing import List, Optional, Any, Dict, Union, Callable, Generator, TYPE_CHECKING
-from collections import OrderedDict
+from typing import List, Optional, Any, Dict, Union, Generator, TYPE_CHECKING
 
 if TYPE_CHECKING:
     pass
@@ -233,8 +229,21 @@ Your Goal: {self.goal}"""
         """
         try:
             from ..llm.model_capabilities import supports_structured_outputs
+        except ImportError:
+            return False  # Module genuinely not available — acceptable
+
+        try:
             return supports_structured_outputs(self.llm)
-        except Exception:
+        except Exception as e:
+            logging.warning(
+                "Structured output capability check failed for agent %s (model=%r); "
+                "falling back to prompt-based schema formatting. Check model capability "
+                "configuration and optional provider dependencies: %s",
+                getattr(self, "name", "<unknown>"),
+                self.llm,
+                e,
+                exc_info=True,
+            )
             return False
 
     def _build_messages(self, prompt, temperature=1.0, output_json=None, output_pydantic=None, tools=None, use_native_format=False):
@@ -473,8 +482,11 @@ Your Goal: {self.goal}"""
                     if tool_calls:
                         names = [getattr(tc.function, "name", "?") for tc in tool_calls]
                         return f"[tool_calls: {', '.join(names)}]"
-        except (AttributeError, IndexError, TypeError):
-            pass
+        except (AttributeError, IndexError, TypeError) as e:
+            logging.warning(
+                f"Failed to extract LLM response content (falling back to str): {e}"
+            )
+            # Fallback to str(response) is still fine, but now it's visible
         return str(response)
 
     def _process_stream_response(self, messages, temperature, start_time, formatted_tools=None, reasoning_steps=False):
@@ -495,7 +507,7 @@ Your Goal: {self.goal}"""
             emit_events=True
         )
 
-    def _chat_completion(self, messages, temperature=1.0, tools=None, stream=True, reasoning_steps=False, task_name=None, task_description=None, task_id=None, response_format=None):
+    def _chat_completion(self, messages, temperature=1.0, tools=None, stream=True, reasoning_steps=False, task_name=None, task_description=None, task_id=None, response_format=None, _retry_depth=0):
         start_time = time.time()
 
         # --- Context compaction (opt-in via ExecutionConfig.context_compaction) ---
@@ -510,8 +522,10 @@ Your Goal: {self.goal}"""
                 if _compactor.needs_compaction(messages):
                     try:
                         self._hook_runner.execute_sync(_HookEvent.BEFORE_COMPACTION, None)
-                    except Exception:
-                        pass
+                    except Exception as e:
+                        logging.warning(f"BEFORE_COMPACTION hook failed: {e}")
+                        if getattr(self, '_strict_hooks', False):
+                            raise
                     compacted_msgs, _cr = _compactor.compact(messages)
                     messages[:] = compacted_msgs  # in-place update so callers see the change
                     logging.info(
@@ -520,9 +534,13 @@ Your Goal: {self.goal}"""
                     )
                     try:
                         self._hook_runner.execute_sync(_HookEvent.AFTER_COMPACTION, None)
-                    except Exception:
-                        pass
+                    except Exception as e:
+                        logging.warning(f"AFTER_COMPACTION hook failed: {e}")
+                        if getattr(self, '_strict_hooks', False):
+                            raise
             except Exception as _ce:
+                if getattr(self, '_strict_hooks', False):
+                    raise
                 logging.debug(f"[compaction] skipped (non-fatal): {_ce}")
 
         # Trigger BEFORE_LLM hook
@@ -538,7 +556,11 @@ Your Goal: {self.goal}"""
             temperature=temperature
         )
         self._hook_runner.execute_sync(HookEvent.BEFORE_LLM, before_llm_input)
-        
+        # C7 — honour any BEFORE_LLM hook that mutated the message stream
+        # (e.g. PII redactor). The runner applies modified_input in-place on
+        # before_llm_input.messages; adopt that value for the actual LLM call.
+        messages = before_llm_input.messages
+
         logging.debug(f"{self.name} sending messages to LLM: {messages}")
         
         # Emit LLM request trace event (zero overhead when not set)
@@ -645,6 +667,7 @@ Your Goal: {self.goal}"""
         except BudgetExceededError:
             raise
         except Exception as e:
+            from ..errors import LLMError
             error_str = str(e).lower()
             
             # Check if this is a context overflow error
@@ -677,11 +700,22 @@ Your Goal: {self.goal}"""
                         f"{estimate_messages_tokens(truncated_messages)} tokens"
                     )
                     
-                    # Retry with truncated messages (recursive call with truncated context)
-                    return self._chat_completion(
-                        truncated_messages, temperature, tools, stream, 
-                        reasoning_steps, task_name, task_description, task_id, response_format
-                    )
+                    # Retry with truncated messages (recursive call with depth limit)
+                    if _retry_depth < 2:
+                        return self._chat_completion(
+                            truncated_messages, temperature, tools, stream, 
+                            reasoning_steps, task_name, task_description, task_id, response_format, 
+                            _retry_depth=_retry_depth + 1
+                        )
+                    else:
+                        logging.error(f"[{self.name}] Context overflow retry limit exceeded")
+                        raise LLMError(
+                            f"Context overflow could not be resolved after {_retry_depth} attempts", 
+                            model_name=model_name, agent_id=self.name, is_retryable=False
+                        ) from e
+                except LLMError:
+                    # Re-raise LLMError (including depth limit errors) without swallowing
+                    raise
                 except Exception as recovery_error:
                     logging.error(f"[{self.name}] Overflow recovery failed: {recovery_error}")
             
@@ -693,8 +727,47 @@ Your Goal: {self.goal}"""
                 finish_reason="error",
                 response_content=str(e),  # Include error for context replay
             )
-            _get_display_functions()['display_error'](f"Error in chat completion: {e}")
-            return None
+            
+            # Classify and raise structured error with improved transient failure detection
+            model_name = self.llm if isinstance(self.llm, str) else "unknown"
+            session_id = getattr(self, '_session_id', 'unknown')
+            
+            # Check for retryable errors (rate limits, transient network issues, provider errors)
+            retryable_indicators = [
+                "rate limit", "429", "too many requests",
+                "timeout", "connection reset", "connection error", "socket error",
+                "500", "502", "503", "504", "service unavailable", "internal server error",
+                "dns", "network", "connection refused"
+            ]
+            
+            # Check for non-retryable errors (auth issues)
+            auth_indicators = ["401", "403", "authentication", "unauthorized", "invalid_api_key"]
+            
+            if any(phrase in error_str.lower() for phrase in retryable_indicators):
+                is_retryable = True
+            elif any(phrase in error_str.lower() for phrase in auth_indicators):
+                is_retryable = False
+            else:
+                # Default to retryable for unknown errors to be more resilient
+                is_retryable = True
+            
+            # Create LLMError with contextual metadata
+            error = LLMError(
+                str(e), 
+                model_name=model_name, 
+                agent_id=self.name, 
+                is_retryable=is_retryable,
+                session_id=session_id
+            )
+            
+            # Call error hook if available for error interception
+            if hasattr(self, 'on_error') and self.on_error:
+                try:
+                    self.on_error(error)
+                except Exception as hook_error:
+                    logging.debug(f"Error in on_error hook: {hook_error}")
+            
+            raise error from e
 
     def _execute_unified_chat_completion(
         self, 
@@ -1042,30 +1115,41 @@ Your Goal: {self.goal}"""
         rendered = mgr.invoke(name, raw_args=args)
         if rendered is None:
             return prompt
-        # G6: Best-effort pre-approve any tools declared under
+        # G-A fix: Best-effort pre-approve any tools declared under
         # `allowed-tools` in the skill frontmatter. Non-fatal on error.
         try:
             tool_names = mgr.get_allowed_tools(name)
             if tool_names:
-                from ..approval import get_approval_registry, AutoApproveBackend
+                from ..approval import get_approval_registry
 
                 registry = get_approval_registry()
-                agent_name = getattr(self, "name", None)
-                for _tn in tool_names:
-                    try:
-                        registry.set_backend(
-                            AutoApproveBackend(),
-                            agent_name=agent_name,
-                            tool_name=_tn,
-                        )
-                    except TypeError:
-                        # Older registry may not accept tool_name kwarg
-                        registry.set_backend(AutoApproveBackend(), agent_name=agent_name)
-        except Exception:  # pragma: no cover - approval is optional
-            pass
+                agent_name = getattr(self, "display_name", getattr(self, "name", None))
+                if agent_name:  # Only approve if we have a stable agent identifier
+                    for _tn in tool_names:
+                        try:
+                            registry.auto_approve_tool(_tn, agent_name=agent_name)
+                        except Exception as exc:  # pragma: no cover - approval is optional
+                            logging.debug(
+                                "Failed to auto-approve skill tool '%s' for skill '%s' on agent '%s': %s. "
+                                "The skill will continue, but this tool may still require explicit approval.",
+                                _tn,
+                                name,
+                                agent_name,
+                                exc,
+                                exc_info=True,
+                            )
+        except Exception as exc:  # pragma: no cover - approval is optional
+            logging.debug(
+                "Failed to resolve allowed tools for skill '%s' on agent '%s': %s. "
+                "The skill will continue without pre-approving tools.",
+                name,
+                getattr(self, "name", None),
+                exc,
+                exc_info=True,
+            )
         return rendered
 
-    def chat(self, prompt: str, temperature: float = 1.0, tools: Optional[List[Any]] = None, output_json: Optional[Any] = None, output_pydantic: Optional[Any] = None, reasoning_steps: bool = False, stream: Optional[bool] = None, task_name: Optional[str] = None, task_description: Optional[str] = None, task_id: Optional[str] = None, config: Optional[Dict[str, Any]] = None, force_retrieval: bool = False, skip_retrieval: bool = False, attachments: Optional[List[str]] = None, tool_choice: Optional[str] = None) -> Optional[str]:
+    def chat(self, prompt: str, temperature: float = 1.0, tools: Optional[List[Any]] = None, output_json: Optional[Any] = None, output_pydantic: Optional[Any] = None, reasoning_steps: bool = False, stream: Optional[bool] = None, task_name: Optional[str] = None, task_description: Optional[str] = None, task_id: Optional[str] = None, config: Optional[Dict[str, Any]] = None, force_retrieval: bool = False, skip_retrieval: bool = False, attachments: Optional[List[str]] = None, tool_choice: Optional[str] = None, seed: Optional[int] = None, cancel_token: Optional[Any] = None) -> Optional[str]:
         """
         Chat with the agent.
         
@@ -1109,11 +1193,17 @@ Your Goal: {self.goal}"""
         _trace_emitter.agent_start(self.name, {"role": self.role, "goal": self.goal})
         
         try:
-            return self._chat_impl(prompt, temperature, tools, output_json, output_pydantic, reasoning_steps, stream, task_name, task_description, task_id, config, force_retrieval, skip_retrieval, attachments, _trace_emitter, tool_choice)
+            # C2 — cooperative cancellation: abort early if a pre-set token is given
+            _cancel = cancel_token if cancel_token is not None else getattr(self, "interrupt_controller", None)
+            if _cancel is not None and getattr(_cancel, "is_set", lambda: False)():
+                reason = getattr(_cancel, "reason", None) or "cancelled before LLM call"
+                raise InterruptedError(f"Agent chat cancelled: {reason}")
+
+            return self._chat_impl(prompt, temperature, tools, output_json, output_pydantic, reasoning_steps, stream, task_name, task_description, task_id, config, force_retrieval, skip_retrieval, attachments, _trace_emitter, tool_choice, seed=seed, cancel_token=_cancel)
         finally:
             _trace_emitter.agent_end(self.name)
 
-    def _chat_impl(self, prompt, temperature, tools, output_json, output_pydantic, reasoning_steps, stream, task_name, task_description, task_id, config, force_retrieval, skip_retrieval, attachments, _trace_emitter, tool_choice=None):
+    def _chat_impl(self, prompt, temperature, tools, output_json, output_pydantic, reasoning_steps, stream, task_name, task_description, task_id, config, force_retrieval, skip_retrieval, attachments, _trace_emitter, tool_choice=None, seed=None, cancel_token=None):
         """Internal chat implementation (extracted for trace wrapping)."""
         # Apply rate limiter if configured (before any LLM call)
         if self._rate_limiter is not None:
@@ -1334,7 +1424,16 @@ Your Goal: {self.goal}"""
                     effective_tool_choice = tool_choice or getattr(self, '_yaml_tool_choice', None)
                     if effective_tool_choice:
                         llm_kwargs['tool_choice'] = effective_tool_choice
-                    
+
+                    # C1 — per-call seed overrides llm_instance.seed for determinism
+                    if seed is not None:
+                        llm_kwargs['seed'] = seed
+
+                    # C2 — last-chance cancel check before handing to the LLM
+                    if cancel_token is not None and getattr(cancel_token, 'is_set', lambda: False)():
+                        reason = getattr(cancel_token, 'reason', None) or 'cancelled'
+                        raise InterruptedError(f"Agent chat cancelled: {reason}")
+
                     response_text = self.llm_instance.get_response(**llm_kwargs)
 
                     self._add_to_chat_history("assistant", response_text)
@@ -1355,11 +1454,11 @@ Your Goal: {self.goal}"""
                     except Exception as e:
                         logging.error(f"Agent {self.name}: Guardrail validation failed for custom LLM: {e}")
                         # Rollback chat history on guardrail failure
-                        self.chat_history = self.chat_history[:chat_history_length]
+                        self._truncate_chat_history(chat_history_length)
                         return None
                 except Exception as e:
                     # Rollback chat history if LLM call fails
-                    self.chat_history = self.chat_history[:chat_history_length]
+                    self._truncate_chat_history(chat_history_length)
                     _get_display_functions()['display_error'](f"Error in LLM chat: {e}")
                     return None
             except Exception as e:
@@ -1400,7 +1499,7 @@ Your Goal: {self.goal}"""
                     self.chat_history[-1].get("role") == "user" and 
                     self.chat_history[-1].get("content") == normalized_content):
                 # Add user message to chat history BEFORE LLM call so handoffs can access it
-                self.chat_history.append({"role": "user", "content": normalized_content})
+                self._append_to_chat_history({"role": "user", "content": normalized_content})
                 # Persist user message to DB (OpenAI path)
                 self._persist_message("user", normalized_content)
 
@@ -1444,7 +1543,7 @@ Your Goal: {self.goal}"""
                         response = self._chat_completion(messages, temperature=temperature, tools=tools if tools else None, reasoning_steps=reasoning_steps, stream=stream, task_name=task_name, task_description=task_description, task_id=task_id, response_format=response_format)
                         if not response:
                             # Rollback chat history on response failure
-                            self.chat_history = self.chat_history[:chat_history_length]
+                            self._truncate_chat_history(chat_history_length)
                             return None
 
                         # Handle None content (can happen with tool calls or empty responses)
@@ -1455,7 +1554,7 @@ Your Goal: {self.goal}"""
                         if output_json or output_pydantic:
                             # Add to chat history and return raw response
                             # User message already added before LLM call via _build_messages
-                            self.chat_history.append({"role": "assistant", "content": response_text})
+                            self._append_to_chat_history({"role": "assistant", "content": response_text})
                             # Persist assistant message to DB
                             self._persist_message("assistant", response_text)
                             # Apply guardrail validation even for JSON output
@@ -1467,12 +1566,12 @@ Your Goal: {self.goal}"""
                             except Exception as e:
                                 logging.error(f"Agent {self.name}: Guardrail validation failed for JSON output: {e}")
                                 # Rollback chat history on guardrail failure
-                                self.chat_history = self.chat_history[:chat_history_length]
+                                self._truncate_chat_history(chat_history_length)
                                 return None
 
                         if not self.self_reflect:
                             # User message already added before LLM call via _build_messages
-                            self.chat_history.append({"role": "assistant", "content": response_text})
+                            self._append_to_chat_history({"role": "assistant", "content": response_text})
                             # Persist assistant message to DB (non-reflect path)
                             self._persist_message("assistant", response_text)
                             if self.verbose:
@@ -1488,7 +1587,7 @@ Your Goal: {self.goal}"""
                                 except Exception as e:
                                     logging.error(f"Agent {self.name}: Guardrail validation failed for reasoning content: {e}")
                                     # Rollback chat history on guardrail failure
-                                    self.chat_history = self.chat_history[:chat_history_length]
+                                    self._truncate_chat_history(chat_history_length)
                                     return None
                             # Apply guardrail to regular response
                             try:
@@ -1499,7 +1598,7 @@ Your Goal: {self.goal}"""
                             except Exception as e:
                                 logging.error(f"Agent {self.name}: Guardrail validation failed: {e}")
                                 # Rollback chat history on guardrail failure
-                                self.chat_history = self.chat_history[:chat_history_length]
+                                self._truncate_chat_history(chat_history_length)
                                 return None
 
                         reflection_prompt = f"""
@@ -1558,7 +1657,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                 if self.verbose:
                                     _get_display_functions()['display_self_reflection']("Agent marked the response as satisfactory after meeting minimum reflections", console=self.console)
                                 # User message already added before LLM call via _build_messages
-                                self.chat_history.append({"role": "assistant", "content": response_text})
+                                self._append_to_chat_history({"role": "assistant", "content": response_text})
                                 # Apply guardrail validation after satisfactory reflection
                                 try:
                                     validated_response = self._apply_guardrail_with_retry(response_text, original_prompt, temperature, tools, task_name, task_description, task_id)
@@ -1569,7 +1668,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                 except Exception as e:
                                     logging.error(f"Agent {self.name}: Guardrail validation failed after reflection: {e}")
                                     # Rollback chat history on guardrail failure
-                                    self.chat_history = self.chat_history[:chat_history_length]
+                                    self._truncate_chat_history(chat_history_length)
                                     self._end_run(None, "error", {"error": str(e)})
                                     return None
 
@@ -1578,7 +1677,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                 if self.verbose:
                                     _get_display_functions()['display_self_reflection']("Maximum reflection count reached, returning current response", console=self.console)
                                 # User message already added before LLM call via _build_messages
-                                self.chat_history.append({"role": "assistant", "content": response_text})
+                                self._append_to_chat_history({"role": "assistant", "content": response_text})
                                 # Apply guardrail validation after max reflections
                                 try:
                                     validated_response = self._apply_guardrail_with_retry(response_text, original_prompt, temperature, tools, task_name, task_description, task_id)
@@ -1588,7 +1687,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                 except Exception as e:
                                     logging.error(f"Agent {self.name}: Guardrail validation failed after max reflections: {e}")
                                     # Rollback chat history on guardrail failure
-                                    self.chat_history = self.chat_history[:chat_history_length]
+                                    self._truncate_chat_history(chat_history_length)
                                     return None
                             
                             # If not satisfactory and not at max reflections, continue with regeneration
@@ -1615,7 +1714,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                 # Catch any exceptions that escape the while loop
                 _get_display_functions()['display_error'](f"Unexpected error in chat: {e}", console=self.console)
                 # Rollback chat history
-                self.chat_history = self.chat_history[:chat_history_length]
+                self._truncate_chat_history(chat_history_length)
                 return None
 
     def clean_json_output(self, output: str) -> str:
@@ -1630,7 +1729,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
             cleaned = cleaned[:-3].strip()
         return cleaned  
 
-    async def achat(self, prompt: str, temperature: float = 1.0, tools: Optional[List[Any]] = None, output_json: Optional[Any] = None, output_pydantic: Optional[Any] = None, reasoning_steps: bool = False, stream: Optional[bool] = None, task_name: Optional[str] = None, task_description: Optional[str] = None, task_id: Optional[str] = None, config: Optional[Dict[str, Any]] = None, force_retrieval: bool = False, skip_retrieval: bool = False, attachments: Optional[List[str]] = None, tool_choice: Optional[str] = None):
+    async def achat(self, prompt: str, temperature: float = 1.0, tools: Optional[List[Any]] = None, output_json: Optional[Any] = None, output_pydantic: Optional[Any] = None, reasoning_steps: bool = False, stream: Optional[bool] = None, task_name: Optional[str] = None, task_description: Optional[str] = None, task_id: Optional[str] = None, config: Optional[Dict[str, Any]] = None, force_retrieval: bool = False, skip_retrieval: bool = False, attachments: Optional[List[str]] = None, tool_choice: Optional[str] = None, seed: Optional[int] = None, cancel_token: Optional[Any] = None):
         """Async version of chat method with self-reflection support.
         
         Args:
@@ -1653,13 +1752,20 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                 reasoning_steps=reasoning_steps, stream=stream,
                 task_name=task_name, task_description=task_description, task_id=task_id,
                 config=config, force_retrieval=force_retrieval, skip_retrieval=skip_retrieval,
-                attachments=attachments, _trace_emitter=_trace_emitter, tool_choice=tool_choice
+                attachments=attachments, _trace_emitter=_trace_emitter, tool_choice=tool_choice,
+                seed=seed, cancel_token=cancel_token
             )
         finally:
             _trace_emitter.agent_end(self.name)
 
-    async def _achat_impl(self, prompt, temperature, tools, output_json, output_pydantic, reasoning_steps, stream, task_name, task_description, task_id, config, force_retrieval, skip_retrieval, attachments, _trace_emitter, tool_choice=None):
+    async def _achat_impl(self, prompt, temperature, tools, output_json, output_pydantic, reasoning_steps, stream, task_name, task_description, task_id, config, force_retrieval, skip_retrieval, attachments, _trace_emitter, tool_choice=None, seed=None, cancel_token=None):
         """Internal async chat implementation (extracted for trace wrapping)."""
+        # C2 — cooperative cancellation: abort early if a pre-set token is given
+        _cancel = cancel_token if cancel_token is not None else getattr(self, "interrupt_controller", None)
+        if _cancel is not None and getattr(_cancel, "is_set", lambda: False)():
+            reason = getattr(_cancel, "reason", None) or "cancelled before LLM call"
+            raise InterruptedError(f"Agent chat cancelled: {reason}")
+        
         # Use agent's stream setting if not explicitly provided
         if stream is None:
             stream = self.stream
@@ -1747,7 +1853,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                         self.chat_history[-1].get("role") == "user" and 
                         self.chat_history[-1].get("content") == normalized_content):
                     # Add user message to chat history BEFORE LLM call so handoffs can access it
-                    self.chat_history.append({"role": "user", "content": normalized_content})
+                    self._append_to_chat_history({"role": "user", "content": normalized_content})
 
                 # --- Context compaction (async custom LLM path) ---
                 _exec_cfg = getattr(self, 'execution', None)
@@ -1760,49 +1866,67 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                         if _cw.needs_compaction(self.chat_history):
                             try:
                                 await self._hook_runner.execute(_HE.BEFORE_COMPACTION, None)
-                            except Exception:
-                                pass
+                            except Exception as e:
+                                logging.warning(f"BEFORE_COMPACTION hook failed: {e}")
+                                if getattr(self, '_strict_hooks', False):
+                                    raise
                             _ch, _cr = _cw.compact(self.chat_history)
-                            self.chat_history[:] = _ch
+                            self._replace_chat_history(_ch)
                             logging.info(
                                 f"[compaction] {self.name}: {_cr.original_tokens}→{_cr.compacted_tokens} tokens "
                                 f"({_cr.messages_removed} messages removed)"
                             )
                             try:
                                 await self._hook_runner.execute(_HE.AFTER_COMPACTION, None)
-                            except Exception:
-                                pass
+                            except Exception as e:
+                                logging.warning(f"AFTER_COMPACTION hook failed: {e}")
+                                if getattr(self, '_strict_hooks', False):
+                                    raise
                     except Exception as _ce:
+                        if getattr(self, '_strict_hooks', False):
+                            raise
                         logging.debug(f"[compaction] skipped (non-fatal): {_ce}")
 
                 try:
-                    response_text = await self.llm_instance.get_response_async(
-                        prompt=prompt,
-                        system_prompt=self._build_system_prompt(tools),
-                        chat_history=self.chat_history,
-                        temperature=temperature,
-                        tools=tools,
-                        output_json=output_json,
-                        output_pydantic=output_pydantic,
-                        verbose=self.verbose,
-                        markdown=self.markdown,
-                        reflection=self.self_reflect,
-                        max_reflect=self.max_reflect,
-                        min_reflect=self.min_reflect,
-                        console=self.console,
-                        agent_name=self.name,
-                        agent_role=self.role,
-                        agent_tools=[t.__name__ if hasattr(t, '__name__') else str(t) for t in (tools if tools is not None else self.tools)],
-                        task_name=task_name,
-                        task_description=task_description,
-                        task_id=task_id,
-                        execute_tool_fn=self.execute_tool_async,
-                        parallel_tool_calls=getattr(getattr(self, "execution", None), "parallel_tool_calls", False),
-                        reasoning_steps=reasoning_steps,
-                        stream=stream
-                    )
+                    # C1 — per-call seed forwarding (async path)  
+                    llm_kwargs = {
+                        'prompt': prompt,
+                        'system_prompt': self._build_system_prompt(tools),
+                        'chat_history': self.chat_history,
+                        'temperature': temperature,
+                        'tools': tools,
+                        'output_json': output_json,
+                        'output_pydantic': output_pydantic,
+                        'verbose': self.verbose,
+                        'markdown': self.markdown,
+                        'reflection': self.self_reflect,
+                        'max_reflect': self.max_reflect,
+                        'min_reflect': self.min_reflect,
+                        'console': self.console,
+                        'agent_name': self.name,
+                        'agent_role': self.role,
+                        'agent_tools': [t.__name__ if hasattr(t, '__name__') else str(t) for t in (tools if tools is not None else self.tools)],
+                        'task_name': task_name,
+                        'task_description': task_description,
+                        'task_id': task_id,
+                        'execute_tool_fn': self.execute_tool_async,
+                        'parallel_tool_calls': getattr(getattr(self, "execution", None), "parallel_tool_calls", False),
+                        'reasoning_steps': reasoning_steps,
+                        'stream': stream
+                    }
+                    
+                    # C1 — per-call seed overrides llm_instance.seed for determinism  
+                    if seed is not None:
+                        llm_kwargs['seed'] = seed
+                    
+                    # C2 — last-chance cancel check before handing to the LLM
+                    if _cancel is not None and getattr(_cancel, 'is_set', lambda: False)():
+                        reason = getattr(_cancel, 'reason', None) or 'cancelled'
+                        raise InterruptedError(f"Agent chat cancelled: {reason}")
+                    
+                    response_text = await self.llm_instance.get_response_async(**llm_kwargs)
 
-                    self.chat_history.append({"role": "assistant", "content": response_text})
+                    self._append_to_chat_history({"role": "assistant", "content": response_text})
 
                     if get_logger().getEffectiveLevel() == logging.DEBUG:
                         total_time = time.time() - start_time
@@ -1817,11 +1941,11 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                     except Exception as e:
                         logging.error(f"Agent {self.name}: Guardrail validation failed for custom LLM: {e}")
                         # Rollback chat history on guardrail failure
-                        self.chat_history = self.chat_history[:chat_history_length]
+                        self._truncate_chat_history(chat_history_length)
                         return None
                 except Exception as e:
                     # Rollback chat history if LLM call fails
-                    self.chat_history = self.chat_history[:chat_history_length]
+                    self._truncate_chat_history(chat_history_length)
                     _get_display_functions()['display_error'](f"Error in LLM chat: {e}")
                     if get_logger().getEffectiveLevel() == logging.DEBUG:
                         total_time = time.time() - start_time
@@ -1846,7 +1970,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                     self.chat_history[-1].get("role") == "user" and 
                     self.chat_history[-1].get("content") == normalized_content):
                 # Add user message to chat history BEFORE LLM call so handoffs can access it
-                self.chat_history.append({"role": "user", "content": normalized_content})
+                self._append_to_chat_history({"role": "user", "content": normalized_content})
 
             # --- Context compaction (async standard OpenAI path) ---
             _exec_cfg2 = getattr(self, 'execution', None)
@@ -1996,8 +2120,8 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                         if self.verbose:
                                             _get_display_functions()['display_self_reflection'](f"Agent {self.name}: Self-reflection with structured output is not supported for custom LLM providers. Skipping reflection.", console=self.console)
                                         # Return the original response without reflection
-                                        self.chat_history.append({"role": "user", "content": original_prompt})
-                                        self.chat_history.append({"role": "assistant", "content": response_text})
+                                        self._append_to_chat_history({"role": "user", "content": original_prompt})
+                                        self._append_to_chat_history({"role": "assistant", "content": response_text})
                                         if get_logger().getEffectiveLevel() == logging.DEBUG:
                                             total_time = time.time() - start_time
                                             logging.debug(f"Agent.achat completed in {total_time:.2f} seconds")
@@ -2063,7 +2187,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                         except Exception as e:
                             logging.error(f"Agent {self.name}: Guardrail validation failed for OpenAI client: {e}")
                             # Rollback chat history on guardrail failure
-                            self.chat_history = self.chat_history[:chat_history_length]
+                            self._truncate_chat_history(chat_history_length)
                             return None
                 except Exception as e:
                     _get_display_functions()['display_error'](f"Error in chat completion: {e}")
@@ -2125,13 +2249,9 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                     except Exception as _hook_err:
                         logging.debug(f"BEFORE_TOOL hook error (non-fatal): {_hook_err}")
 
-                    # Check if the tool is async
-                    if asyncio.iscoroutinefunction(tool):
-                        result = await tool(**arguments)
-                    else:
-                        # Run sync function in executor to avoid blocking
-                        loop = asyncio.get_running_loop()
-                        result = await loop.run_in_executor(None, lambda: tool(**arguments))
+                    # Route through safety pipeline instead of direct execution
+                    # Pass the tools list to honor task-scoped tools
+                    result = await self.execute_tool_async(function_name, arguments, tools_override=tools)
 
                     # --- AFTER_TOOL hook ---
                     try:
@@ -2306,7 +2426,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                 if not (self.chat_history and 
                         self.chat_history[-1].get("role") == "user" and 
                         self.chat_history[-1].get("content") == normalized_content):
-                    self.chat_history.append({"role": "user", "content": normalized_content})
+                    self._append_to_chat_history({"role": "user", "content": normalized_content})
                 
                 try:
                     # Use the new streaming generator from LLM class
@@ -2335,11 +2455,11 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                     
                     # Add complete response to chat history
                     if response_content:
-                        self.chat_history.append({"role": "assistant", "content": response_content})
+                        self._append_to_chat_history({"role": "assistant", "content": response_content})
                         
                 except Exception as e:
                     # Rollback chat history on error
-                    self.chat_history = self.chat_history[:chat_history_length]
+                    self._truncate_chat_history(chat_history_length)
                     logging.error(f"Custom LLM streaming error: {e}")
                     raise
                     
@@ -2382,7 +2502,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                 if not (self.chat_history and 
                         self.chat_history[-1].get("role") == "user" and 
                         self.chat_history[-1].get("content") == normalized_content):
-                    self.chat_history.append({"role": "user", "content": normalized_content})
+                    self._append_to_chat_history({"role": "user", "content": normalized_content})
                 
                 try:
                     # Check if OpenAI client is available
@@ -2494,7 +2614,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                     "function": tc['function']
                                 } for tc in tool_calls_data if tc['id']
                             ]
-                        self.chat_history.append(assistant_message)
+                        self._append_to_chat_history(assistant_message)
                         
                         # Execute tool calls and add results to chat history
                         for tool_call in tool_calls_data:
@@ -2513,7 +2633,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                         tool_call_id=tool_call.get('id')
                                     )
                                     # Add tool result to chat history
-                                    self.chat_history.append({
+                                    self._append_to_chat_history({
                                         "role": "tool",
                                         "tool_call_id": tool_call['id'],
                                         "content": str(tool_result)
@@ -2521,7 +2641,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                 except Exception as tool_error:
                                     logging.error(f"Tool execution error in streaming: {tool_error}")
                                     # Add error result to chat history
-                                    self.chat_history.append({
+                                    self._append_to_chat_history({
                                         "role": "tool", 
                                         "tool_call_id": tool_call['id'],
                                         "content": f"Error: {str(tool_error)}"
@@ -2529,11 +2649,11 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                     else:
                         # Add complete response to chat history (text-only response)
                         if response_text:
-                            self.chat_history.append({"role": "assistant", "content": response_text})
+                            self._append_to_chat_history({"role": "assistant", "content": response_text})
                         
                 except Exception as e:
                     # Rollback chat history on error
-                    self.chat_history = self.chat_history[:chat_history_length]
+                    self._truncate_chat_history(chat_history_length)
                     logging.error(f"OpenAI streaming error: {e}")
                     # Fall back to simulated streaming
                     response = self.chat(prompt, **kwargs)

src/praisonai-agents/praisonaiagents/agent/code_agent.py

@@ -164,7 +164,8 @@ Follow best practices and coding standards."""
             messages=messages,
             **{k: v for k, v in kwargs.items() if k != "model"}
         )
-        
+        if not response.choices or response.choices[0].message is None:
+            raise ValueError("LLM returned empty or filtered response")
         code = response.choices[0].message.content
         self._log("Code generated successfully", "green")
         return code
@@ -204,7 +205,8 @@ Follow best practices and coding standards."""
             messages=messages,
             **{k: v for k, v in kwargs.items() if k != "model"}
         )
-        
+        if not response.choices or response.choices[0].message is None:
+            raise ValueError("LLM returned empty or filtered response")
         code = response.choices[0].message.content
         self._log("Code generated successfully", "green")
         return code
@@ -365,6 +367,8 @@ Provide clear, actionable feedback."""
             **{k: v for k, v in kwargs.items() if k != "model"}
         )
         
+        if not response.choices or response.choices[0].message is None:
+            raise ValueError("LLM returned empty or filtered response")
         review = response.choices[0].message.content
         self._log("Review completed", "green")
         return review
@@ -405,6 +409,8 @@ Include:
             **{k: v for k, v in kwargs.items() if k != "model"}
         )
         
+        if not response.choices or response.choices[0].message is None:
+            raise ValueError("LLM returned empty or filtered response")
         explanation = response.choices[0].message.content
         self._log("Explanation completed", "green")
         return explanation
@@ -451,6 +457,8 @@ Only output the refactored code."""
             **{k: v for k, v in kwargs.items() if k != "model"}
         )
         
+        if not response.choices or response.choices[0].message is None:
+            raise ValueError("LLM returned empty or filtered response")
         refactored = response.choices[0].message.content
         self._log("Refactoring completed", "green")
         return refactored
@@ -495,6 +503,8 @@ Ensure the fix:
             **{k: v for k, v in kwargs.items() if k != "model"}
         )
         
+        if not response.choices or response.choices[0].message is None:
+            raise ValueError("LLM returned empty or filtered response")
         fixed = response.choices[0].message.content
         self._log("Fix completed", "green")
         return fixed

src/praisonai-agents/praisonaiagents/agent/execution_mixin.py

@@ -5,12 +5,10 @@ Contains all methods for run/start/launch lifecycle, planning,
 and server endpoints. Extracted from agent.py for maintainability.
 """
 
-import os
 import time
 import json
 import logging
 import inspect
-from praisonaiagents._logging import get_logger
 
 import asyncio
 import threading
@@ -126,7 +124,7 @@ class ExecutionMixin:
             config=HandoffConfig(context_policy=ContextPolicy.NONE),
         )
 
-    async def arun(self, prompt: str, **kwargs):
+    async def arun(self, prompt: str, **kwargs) -> Optional[str]:
         """Async version of run() - silent, non-streaming, returns structured result.
         
         Production-friendly async execution. Does not stream or display output.
@@ -309,8 +307,8 @@ class ExecutionMixin:
         # ── Chat history & session linkage ──
         # Record prompt+response in chat_history so SessionStore/auto_save works
         if result is not None:
-            self.chat_history.append({"role": "user", "content": prompt})
-            self.chat_history.append({"role": "assistant", "content": str(result)})
+            self._append_to_chat_history({"role": "user", "content": prompt})
+            self._append_to_chat_history({"role": "assistant", "content": str(result)})
             # Link managed session ID into SessionStore gateway_session_id
             if hasattr(self.backend, 'managed_session_id'):
                 msid = self.backend.managed_session_id
@@ -319,8 +317,14 @@ class ExecutionMixin:
                         sid = getattr(self, 'auto_save', None) or getattr(self, '_session_id', None)
                         if sid and hasattr(self._session_store, 'set_gateway_info'):
                             self._session_store.set_gateway_info(sid, gateway_session_id=msid)
-                    except Exception:
-                        pass  # Best-effort linkage
+                    except Exception as e:
+                        # Log session linkage failures for debugging
+                        logger.warning(
+                            "Session gateway linkage failed: %s",
+                            e,
+                            extra={"session_id": sid, "managed_session_id": msid},
+                            exc_info=True,
+                        )
             self._auto_save_session()
         
         return result
@@ -942,7 +946,7 @@ Write the complete compiled report:"""
         task_id = getattr(task, 'id', None)
         return await self.achat(prompt, task_name=task_name, task_description=task_description, task_id=task_id)
 
-    async def execute_tool_async(self, function_name: str, arguments: Dict[str, Any], tool_call_id: Optional[str] = None) -> Any:
+    async def execute_tool_async(self, function_name: str, arguments: Dict[str, Any], tool_call_id: Optional[str] = None, tools_override: Optional[List] = None) -> Any:
         """Async version of execute_tool"""
         try:
             logging.info(f"Executing async tool: {function_name} with arguments: {arguments}")
@@ -958,9 +962,10 @@ Write the complete compiled report:"""
                 logging.error(error_msg)
                 return {"error": error_msg, "approval_error": True}
             
-            # Try to find the function in the agent's tools list first
+            # Try to find the function in the override tools list first, then agent's tools list
             func = None
-            for tool in self.tools:
+            tools_to_search = tools_override if tools_override is not None else self.tools
+            for tool in tools_to_search:
                 if (callable(tool) and getattr(tool, '__name__', '') == function_name):
                     func = tool
                     break

src/praisonai-agents/praisonaiagents/agent/handoff.py

@@ -175,6 +175,7 @@ class Handoff:
     # Class-level semaphore for concurrency control
     _semaphore: Optional[asyncio.Semaphore] = None
     _sync_semaphore: Optional[threading.Semaphore] = None
+    _semaphore_lock: threading.Lock = threading.Lock()  # Lock for semaphore initialization
     
     def __init__(
         self,
@@ -461,9 +462,11 @@ class Handoff:
         Returns:
             HandoffResult with response or error
         """
-        # Initialize semaphore if needed
+        # Initialize semaphore if needed (thread-safe)
         if self.config.max_concurrent > 0 and Handoff._semaphore is None:
-            Handoff._semaphore = asyncio.Semaphore(self.config.max_concurrent)
+            with Handoff._semaphore_lock:  # Thread-safe initialization
+                if Handoff._semaphore is None:  # Double-check after acquiring lock
+                    Handoff._semaphore = asyncio.Semaphore(self.config.max_concurrent)
         
         start_time = time.time()
         kwargs = context or {}

src/praisonai-agents/praisonaiagents/agent/interrupt.py

@@ -0,0 +1,105 @@
+"""
+Interrupt Controller - Cooperative cancellation for agent runs.
+
+Provides thread-safe, cooperative cancellation mechanism for long-running agent
+operations. Follows protocol-driven design with zero overhead when not used.
+"""
+
+import threading
+from typing import Optional, Protocol
+from dataclasses import dataclass, field
+
+__all__ = ["InterruptControllerProtocol", "InterruptController"]
+
+
+class InterruptControllerProtocol(Protocol):
+    """Protocol for interrupt controller extension point."""
+    
+    def request(self, reason: str = "user") -> None:
+        """Request cancellation with optional reason."""
+        ...
+    
+    def clear(self) -> None:
+        """Clear interrupt state."""
+        ...
+    
+    def is_set(self) -> bool:
+        """Check if interrupt was requested."""
+        ...
+    
+    @property
+    def reason(self) -> Optional[str]:
+        """Get interrupt reason if set."""
+        ...
+    
+    def check(self) -> None:
+        """Check for interrupt and raise if set."""
+        ...
+
+
+@dataclass
+class InterruptController:
+    """Thread-safe cooperative cancellation for agent runs.
+    
+    Provides a lightweight mechanism for requesting cancellation of agent
+    operations. Uses threading.Event for thread safety and cooperative
+    checking patterns.
+    
+    Examples:
+        Basic usage:
+        >>> controller = InterruptController()
+        >>> # In another thread:
+        >>> controller.request("user_cancel")
+        >>> # In agent loop:
+        >>> if controller.is_set():
+        >>>     return f"Cancelled: {controller.reason}"
+    """
+    
+    _flag: threading.Event = field(default_factory=threading.Event, init=False, repr=False)
+    _reason: Optional[str] = field(default=None, init=False)
+    _lock: threading.Lock = field(default_factory=threading.Lock, init=False, repr=False)
+
+    def request(self, reason: str = "user") -> None:
+        """Request cancellation with a reason.
+        
+        Args:
+            reason: Human-readable reason for cancellation
+        """
+        with self._lock:
+            if not self._flag.is_set():
+                self._reason = reason
+                self._flag.set()
+
+    def clear(self) -> None:
+        """Clear the cancellation request."""
+        with self._lock:
+            self._reason = None
+            self._flag.clear()
+
+    def is_set(self) -> bool:
+        """Check if cancellation has been requested.
+        
+        Returns:
+            True if cancellation was requested
+        """
+        return self._flag.is_set()
+
+    @property
+    def reason(self) -> Optional[str]:
+        """Get the reason for cancellation.
+        
+        Returns:
+            Reason string if cancelled, None otherwise
+        """
+        with self._lock:
+            return self._reason
+
+    def check(self) -> None:
+        """Check for cancellation and raise if requested.
+        
+        Raises:
+            InterruptedError: If cancellation was requested
+        """
+        if self.is_set():
+            reason = self.reason or "unknown"
+            raise InterruptedError(f"Operation cancelled: {reason}")

src/praisonai-agents/praisonaiagents/agent/memory_mixin.py

@@ -9,7 +9,6 @@ for maintainability.
 import os
 import logging
 from typing import Optional
-from praisonaiagents._logging import get_logger
 
 # Fallback helpers to avoid circular imports
 def _get_console():
@@ -42,7 +41,7 @@ logger = logging.getLogger(__name__)
 
 
 import contextlib
-from typing import List, Optional, Any, Dict, Generator, TYPE_CHECKING
+from typing import Optional, Dict, Generator, TYPE_CHECKING
 from collections import OrderedDict
 
 if TYPE_CHECKING:
@@ -82,7 +81,7 @@ class MemoryMixin:
         with self._history_lock:
             self.chat_history.append({"role": role, "content": content})
 
-    def _add_to_chat_history_if_not_duplicate(self, role, content):
+    def _add_to_chat_history_if_not_duplicate(self, role, content) -> bool:
         """Thread-safe method to add messages to chat history only if not duplicate.
         
         Atomically checks for duplicate and adds message under the same lock to prevent TOCTOU races.

src/praisonai-agents/praisonaiagents/agent/tool_execution.py

@@ -17,6 +17,8 @@ import concurrent.futures
 from typing import List, Optional, Any, Dict, Union, TYPE_CHECKING
 from ..errors import ToolExecutionError
 
+logger = logging.getLogger(__name__)
+
 if TYPE_CHECKING:
     pass
 
@@ -192,6 +194,21 @@ class ToolExecutionMixin:
                 if res.output and res.output.modified_data:
                     arguments.update(res.output.modified_data)
 
+            # C4 — optional tool-argument validation via ToolValidatorProtocol.
+            # Zero overhead when not set. Users wire via `agent._tool_validator = MyValidator()`.
+            _validator = getattr(self, '_tool_validator', None)
+            if _validator is not None:
+                try:
+                    _vres = _validator.validate_args(function_name, arguments)
+                    if _vres is not None and not getattr(_vres, 'valid', True):
+                        _errs = "; ".join(getattr(_vres, 'errors', []) or ["validation failed"])
+                        logging.warning(
+                            f"Tool {function_name} args rejected by validator: {_errs}"
+                        )
+                        return f"Tool arguments rejected: {_errs}"
+                except Exception as _ve:  # noqa: BLE001 — never break tool exec on validator bug
+                    logging.debug(f"Tool validator raised; skipping validation: {_ve}")
+
             # P8/G11: Apply tool timeout if configured
             tool_timeout = getattr(self, '_tool_timeout', None)
             if tool_timeout and tool_timeout > 0:
@@ -200,7 +217,7 @@ class ToolExecutionMixin:
                 
                 def execute_with_context():
                     with with_injection_context(state):
-                        return self._execute_tool_impl(function_name, arguments)
+                        return self._execute_tool_with_circuit_breaker(function_name, arguments)
                 
                 # Use reusable executor to prevent resource leaks
                 if not hasattr(self, '_tool_executor'):
@@ -217,7 +234,7 @@ class ToolExecutionMixin:
                     result = {"error": f"Tool timed out after {tool_timeout}s", "timeout": True}
             else:
                 with with_injection_context(state):
-                    result = self._execute_tool_impl(function_name, arguments)
+                    result = self._execute_tool_with_circuit_breaker(function_name, arguments)
             
             # Apply tool output truncation to prevent context overflow
             # Uses context manager budget if enabled, otherwise applies default limit
@@ -346,9 +363,10 @@ class ToolExecutionMixin:
         try:
             nudge = self._maybe_emit_nudge(prompt if isinstance(prompt, str) else str(prompt))
             if nudge and hasattr(self, "chat_history") and isinstance(self.chat_history, list):
-                self.chat_history.append({"role": "system", "content": nudge.strip()})
-        except Exception:
-            pass
+                self._append_to_chat_history({"role": "system", "content": nudge.strip()})
+        except Exception as e:
+            # Log learning nudge failures for debugging
+            logger.warning("Learning nudge generation failed: %s", e, exc_info=True)
 
         return response
 
@@ -381,9 +399,10 @@ class ToolExecutionMixin:
         try:
             nudge = self._maybe_emit_nudge(prompt if isinstance(prompt, str) else str(prompt))
             if nudge and hasattr(self, "chat_history") and isinstance(self.chat_history, list):
-                self.chat_history.append({"role": "system", "content": nudge.strip()})
-        except Exception:
-            pass
+                self._append_to_chat_history({"role": "system", "content": nudge.strip()})
+        except Exception as e:
+            # Log learning nudge failures for debugging
+            logger.warning("Learning nudge generation failed: %s", e, exc_info=True)
 
         return response
 
@@ -590,6 +609,73 @@ class ToolExecutionMixin:
             logging.info(f"Using modified arguments: {arguments}")
         return None, arguments
 
+    def _execute_tool_with_circuit_breaker(self, function_name, arguments):
+        """Execute tool with circuit breaker protection.
+        
+        Args:
+            function_name: Name of the tool to execute
+            arguments: Arguments for the tool
+            
+        Returns:
+            Tool execution result or circuit breaker error
+        """
+        # Import circuit breaker components first (lazy import for performance)
+        try:
+            from ..tools.circuit_breaker import get_circuit_breaker, CircuitBreakerConfig, CircuitBreakerException
+        except ImportError:
+            # Circuit breaker not available - fallback to direct execution
+            logging.debug("Circuit breaker not available, falling back to direct tool execution")
+            return self._execute_tool_impl(function_name, arguments)
+
+        try:
+            
+            # Get or create circuit breaker for this tool
+            breaker_name = f"tool_{function_name}"
+            config = CircuitBreakerConfig(
+                failure_threshold=5,        # Open after 5 failures
+                recovery_timeout=60.0,      # Wait 60s before trying half-open
+                timeout=30.0,               # Tool call timeout
+                graceful_degradation=True   # Return error instead of raising exception
+            )
+            breaker = get_circuit_breaker(breaker_name, config)
+            
+            # Execute tool through circuit breaker with failure detection wrapper
+            def _tool_wrapper():
+                result = self._execute_tool_impl(function_name, arguments)
+                # Convert error dicts to exceptions so circuit breaker can detect failures
+                # Don't treat approval/permission denials as circuit breaker failures
+                if isinstance(result, dict) and result.get("error") and \
+                   not result.get("approval_denied") and \
+                   not result.get("permission_denied") and \
+                   not result.get("approval_error"):
+                    # Create a sentinel exception to register failure with circuit breaker
+                    class _ToolFailure(Exception):
+                        def __init__(self, error_dict):
+                            self.error_dict = error_dict
+                            super().__init__(error_dict.get("error", "Tool execution failed"))
+                    raise _ToolFailure(result)
+                return result
+            
+            try:
+                return breaker.call(_tool_wrapper)
+            except Exception as e:
+                # Check if this is our sentinel exception
+                if hasattr(e, 'error_dict'):
+                    return e.error_dict  # Return the original error dict
+                else:
+                    raise  # Re-raise other exceptions
+            
+        except CircuitBreakerException as e:
+            # Circuit breaker is open - return error dict instead of raising
+            logging.warning(f"Tool '{function_name}' circuit breaker open: {e}")
+            return {
+                "error": f"Tool '{function_name}' circuit breaker open - too many recent failures",
+                "circuit_open": True,
+                "agent_name": getattr(self, "name", None),
+                "session_id": getattr(self, "_session_id", None),
+                "remediation": "Wait for recovery_timeout (60s) or investigate recent tool failures.",
+            }
+
     def _execute_tool_impl(self, function_name, arguments):
         """Internal tool execution implementation."""
 
@@ -687,11 +773,9 @@ class ToolExecutionMixin:
                 pass
         
         if func is None:
-            # If not found in tools or registry, try globals and main
-            func = globals().get(function_name)
-            if not func:
-                import __main__
-                func = getattr(__main__, function_name, None)
+            # Tool not found in declared tools or registry — do not fall back to
+            # globals() or __main__ as that allows undeclared callables to execute.
+            pass
 
         if func:
             try:
@@ -756,12 +840,18 @@ class ToolExecutionMixin:
 
         tracking_id = str(uuid.uuid4())[:8]
         task = asyncio.ensure_future(backend.request_approval(request))
-        self._pending_approvals[tracking_id] = {
-            "task": task,
-            "function_name": function_name,
-            "arguments": arguments,
-            "request": request,
-        }
+        
+        # Make defensive copy to prevent TOCTOU mutations
+        import copy
+        frozen_args = copy.deepcopy(arguments)
+        
+        async with self._approvals_lock:
+            self._pending_approvals[tracking_id] = {
+                "task": task,
+                "function_name": function_name,
+                "arguments": frozen_args,
+                "request": request,
+            }
         logging.info(f"Approval request submitted: {tracking_id} for {function_name}")
         return tracking_id
 
@@ -773,40 +863,63 @@ class ToolExecutionMixin:
         and their results included.
         """
         results = {}
-        completed_ids = []
+        # Collect completed items under the lock, then execute tools outside the lock
+        # to avoid holding it during potentially slow async tool execution.
+        approved_items = []
+        denied_items = []
+        error_items = []
 
-        for tid, info in self._pending_approvals.items():
-            task = info["task"]
-            if task.done():
-                completed_ids.append(tid)
-                try:
-                    decision = task.result()
-                    if decision.approved:
-                        # Auto-execute the approved tool
-                        tool_result = await self.execute_tool_async(
-                            info["function_name"], info["arguments"],
-                        )
-                        results[tid] = {
-                            "status": "approved_and_executed",
-                            "tool_name": info["function_name"],
-                            "decision": decision,
-                            "result": tool_result,
-                        }
-                    else:
-                        results[tid] = {
-                            "status": "denied",
-                            "tool_name": info["function_name"],
-                            "decision": decision,
-                        }
-                except Exception as e:
-                    results[tid] = {
-                        "status": "error",
-                        "tool_name": info["function_name"],
-                        "error": str(e),
-                    }
+        async with self._approvals_lock:
+            completed_ids = []
+            for tid, info in list(self._pending_approvals.items()):
+                task = info["task"]
+                if task.done():
+                    completed_ids.append(tid)
+                    try:
+                        decision = task.result()
+                        if decision.approved:
+                            approved_items.append((tid, info, decision))
+                        else:
+                            denied_items.append((tid, info, decision))
+                    except Exception as e:
+                        error_items.append((tid, info, e))
 
-        for tid in completed_ids:
-            del self._pending_approvals[tid]
+            # Remove completed entries while still holding the lock
+            for tid in completed_ids:
+                del self._pending_approvals[tid]
+
+        # Execute approved tools outside the lock to avoid long lock hold
+        for tid, info, decision in approved_items:
+            try:
+                tool_result = await self.execute_tool_async(
+                    info["function_name"], info["arguments"],
+                )
+                results[tid] = {
+                    "status": "approved_and_executed",
+                    "tool_name": info["function_name"],
+                    "decision": decision,
+                    "result": tool_result,
+                }
+            except Exception as e:
+                results[tid] = {
+                    "status": "error",
+                    "tool_name": info["function_name"],
+                    "error": str(e),
+                }
+
+        for tid, info, decision in denied_items:
+            results[tid] = {
+                "status": "denied",
+                "tool_name": info["function_name"],
+                "decision": decision,
+            }
+
+        for tid, info, exc in error_items:
+            results[tid] = {
+                "status": "error",
+                "tool_name": info["function_name"],
+                "error": str(exc),
+            }
 
         return results
 

src/praisonai-agents/praisonaiagents/agent/unified_execution_mixin.py

@@ -78,9 +78,28 @@ class UnifiedExecutionMixin:
         _trace_emitter.agent_start(self.name, {"role": self.role, "goal": self.goal})
         
         try:
+            # CLI Backend routing - delegate entire turn if configured
+            if hasattr(self, '_cli_backend') and self._cli_backend is not None:
+                return await self._chat_via_cli_backend(
+                    prompt=prompt,
+                    temperature=temperature,
+                    tools=tools,
+                    output_json=output_json,
+                    output_pydantic=output_pydantic,
+                    reasoning_steps=reasoning_steps,
+                    stream=stream,
+                    task_name=task_name,
+                    task_description=task_description,
+                    task_id=task_id,
+                    config=config,
+                    force_retrieval=force_retrieval,
+                    skip_retrieval=skip_retrieval,
+                    attachments=attachments,
+                    tool_choice=tool_choice
+                )
             # Apply rate limiter if configured (before any LLM call)
             if hasattr(self, '_rate_limiter') and self._rate_limiter is not None:
-                self._rate_limiter.acquire()
+                await self._rate_limiter.acquire_async()
             
             # Process ephemeral attachments (DRY - builds multimodal prompt)
             # IMPORTANT: Original text 'prompt' is stored in history, attachments are NOT

src/praisonai-agents/praisonaiagents/agents/agents.py

@@ -126,7 +126,24 @@ class _AgentServerRegistry:
         
         thread = threading.Thread(target=run_server, daemon=True)
         thread.start()
-        ready_event.wait(timeout=5.0)  # Deterministic wait instead of sleep(0.5)
+        
+        # Check for configurable timeout via environment variable
+        try:
+            timeout = float(os.environ.get("PRAISONAI_SERVER_READY_TIMEOUT", "5.0"))
+        except ValueError:
+            logger.warning("Invalid PRAISONAI_SERVER_READY_TIMEOUT value. Using default 5.0s.")
+            timeout = 5.0
+        became_ready = ready_event.wait(timeout=timeout)
+        
+        if not became_ready:
+            logger.warning(
+                "Agent server on port %s did not become ready within %.1fs. "
+                "Proceeding, but some features may not work correctly. "
+                "Check server logs for startup errors.",
+                port,
+                timeout,
+            )
+        
         return True
 
 
@@ -1010,20 +1027,74 @@ class AgentTeam:
             logger.info(f"Task with ID {task_id} is already completed")
             return
 
+        # Use per-task max_retries if available
+        task_max = getattr(task, "max_retries", self.max_retries)
         retries = 0
-        while task.status != "completed" and retries < self.max_retries:
+        while task.status != "completed" and retries < task_max:
             logger.debug(f"Attempt {retries+1} for task {task_id}")
             if task.status in ["not started", "in progress"]:
                 task_output = await self.aexecute_task(task_id)
                 if task_output and self.completion_checker(task, task_output.raw):
+                    # Run guardrail validation BEFORE marking task complete
+                    if task._guardrail_fn:
+                        try:
+                            guardrail_result = task._process_guardrail(task_output)
+                            if not guardrail_result.success:
+                                if task.retry_count >= task.max_retries:
+                                    raise Exception(
+                                        f"Task failed guardrail validation after {task.max_retries} retries. "
+                                        f"Last error: {guardrail_result.error}"
+                                    )
+                                
+                                task.retry_count += 1
+                                task.status = "in progress"  # Keep task in progress for retry
+                                logger.warning(f"Task {task_id}: Guardrail validation failed (retry {task.retry_count}/{task.max_retries}): {guardrail_result.error}")
+                                retries += 1
+                                continue  # Actually retry the task
+                            
+                            # If guardrail passed and returned a modified result
+                            if guardrail_result.result is not None:
+                                if isinstance(guardrail_result.result, str):
+                                    # Update the task output with the modified result
+                                    task_output.raw = guardrail_result.result
+                                    # Clear structured fields to avoid stale cache
+                                    if hasattr(task_output, 'json_dict'):
+                                        task_output.json_dict = None
+                                    if hasattr(task_output, 'pydantic'):
+                                        task_output.pydantic = None
+                                    task.result = task_output
+                                elif hasattr(guardrail_result.result, 'raw'):
+                                    # Replace with the new task output
+                                    task_output = guardrail_result.result
+                                    task.result = task_output
+                            
+                            logger.info(f"Task {task_id}: Guardrail validation passed")
+                        except Exception as e:
+                            logger.error(f"Task {task_id}: Error in guardrail processing: {e}")
+                            # Handle guardrail failure with retry logic
+                            if task.retry_count >= task.max_retries:
+                                raise Exception(
+                                    f"Task failed due to guardrail processing error after {task.max_retries} retries. "
+                                    f"Last error: {e}"
+                                ) from e
+                            task.retry_count += 1
+                            task.status = "in progress"
+                            logger.warning(f"Task {task_id}: Guardrail processing error (retry {task.retry_count}/{task.max_retries}): {e}")
+                            retries += 1
+                            continue  # Retry the task
+                    
                     task.status = "completed"
                     # Run execute_callback for memory operations
                     try:
-                        # Use the new sync wrapper to avoid pending coroutine issues
-                        task.execute_callback_sync(task_output)
+                        await task.execute_callback(task_output)
                     except Exception as e:
                         logger.error(f"Error executing memory callback for task {task_id}: {e}")
                         logger.exception(e)
+                        # Respect task failure policies - re-raise if configured
+                        if hasattr(task, 'fail_on_callback_error') and task.fail_on_callback_error:
+                            raise
+                        if hasattr(task, 'fail_on_memory_error') and task.fail_on_memory_error:
+                            raise
                     
                     # Run task callback if exists
                     if task.callback:
@@ -1046,7 +1117,11 @@ class AgentTeam:
                     task.status = "in progress"
                     if self.verbose >= 1:
                         logger.info(f"Task {task_id} not completed, retrying")
-                    await asyncio.sleep(1)
+                    # Use task's retry policy instead of hardcoded sleep (with cap)
+                    delay = getattr(task, 'retry_delay', 1)
+                    max_delay = getattr(task, 'max_retry_delay', 300)  # 5 min cap
+                    actual_delay = min(delay * (2 ** retries), max_delay)
+                    await asyncio.sleep(actual_delay)
                     retries += 1
             else:
                 if task.status == "failed":
@@ -1056,8 +1131,9 @@ class AgentTeam:
                     logger.info("Invalid Task status")
                     break
 
-        if retries == self.max_retries and task.status != "completed":
-            logger.info(f"Task {task_id} failed after {self.max_retries} retries.")
+        if retries == task_max and task.status != "completed":
+            task.status = "failed"  # Set failed status to match sync behavior
+            logger.info(f"Task {task_id} failed after {task_max} retries.")
 
     async def arun_all_tasks(self):
         """Async version of run_all_tasks method"""
@@ -1239,12 +1315,61 @@ class AgentTeam:
         if self.variables and not getattr(task, 'variables', None):
             task.variables = self.variables
         
+        # Use per-task max_retries if available
+        task_max = getattr(task, "max_retries", self.max_retries)
         retries = 0
-        while task.status != "completed" and retries < self.max_retries:
+        while task.status != "completed" and retries < task_max:
             logger.debug(f"Attempt {retries+1} for task {task_id}")
             if task.status in ["not started", "in progress"]:
                 task_output = self.execute_task(task_id)
                 if task_output and self.completion_checker(task, task_output.raw):
+                    # Add guardrail validation (matches arun_task logic)
+                    if task._guardrail_fn:
+                        try:
+                            guardrail_result = task._process_guardrail(task_output)
+                            if not guardrail_result.success:
+                                if task.retry_count >= task.max_retries:
+                                    raise Exception(
+                                        f"Task failed guardrail validation after {task.max_retries} retries. "
+                                        f"Last error: {guardrail_result.error}"
+                                    )
+                                task.retry_count += 1
+                                task.status = "in progress"  # Keep task in progress for retry
+                                logger.warning(f"Task {task_id}: Guardrail validation failed (retry {task.retry_count}/{task.max_retries}): {guardrail_result.error}")
+                                retries += 1
+                                continue  # Retry the task
+                            
+                            # If guardrail passed and returned a modified result
+                            if guardrail_result.result is not None:
+                                if isinstance(guardrail_result.result, str):
+                                    # Update the task output with the modified result
+                                    task_output.raw = guardrail_result.result
+                                    # Clear structured fields to avoid stale cache
+                                    if hasattr(task_output, 'json_dict'):
+                                        task_output.json_dict = None
+                                    if hasattr(task_output, 'pydantic'):
+                                        task_output.pydantic = None
+                                    task.result = task_output
+                                elif hasattr(guardrail_result.result, 'raw'):
+                                    # Replace with the new task output
+                                    task_output = guardrail_result.result
+                                    task.result = task_output
+                            
+                            logger.info(f"Task {task_id}: Guardrail validation passed")
+                        except Exception as e:
+                            logger.error(f"Task {task_id}: Error in guardrail processing: {e}")
+                            # Handle guardrail failure with retry logic
+                            if task.retry_count >= task.max_retries:
+                                raise Exception(
+                                    f"Task failed due to guardrail processing error after {task.max_retries} retries. "
+                                    f"Last error: {e}"
+                                ) from e
+                            task.retry_count += 1
+                            task.status = "in progress"
+                            logger.warning(f"Task {task_id}: Guardrail processing error (retry {task.retry_count}/{task.max_retries}): {e}")
+                            retries += 1
+                            continue  # Retry the task
+                    
                     task.status = "completed"
                     # Run execute_callback for memory operations
                     try:
@@ -1283,7 +1408,11 @@ class AgentTeam:
                     task.status = "in progress"
                     if self.verbose >= 1:
                         logger.info(f"Task {task_id} not completed, retrying")
-                    time.sleep(1)
+                    # Use task's retry policy instead of hardcoded sleep (with cap)
+                    delay = getattr(task, 'retry_delay', 1)
+                    max_delay = getattr(task, 'max_retry_delay', 300)  # 5 min cap
+                    actual_delay = min(delay * (2 ** retries), max_delay)
+                    time.sleep(actual_delay)
                     retries += 1
             else:
                 if task.status == "failed":
@@ -1293,8 +1422,9 @@ class AgentTeam:
                     logger.info("Invalid Task status")
                     break
 
-        if retries == self.max_retries and task.status != "completed":
-            logger.info(f"Task {task_id} failed after {self.max_retries} retries.")
+        if retries == task_max and task.status != "completed":
+            task.status = "failed"  # Set failed status
+            logger.info(f"Task {task_id} failed after {task_max} retries.")
 
     def run_all_tasks(self):
         """Synchronous version of run_all_tasks method"""
@@ -1417,56 +1547,103 @@ class AgentTeam:
             ))
             console.print()
             
-            # Execute tasks with verbose output
+            # Use callbacks for verbose display while maintaining process orchestration
             total_agents = len(self.agents)
             workflow_start_time = time_module.time()
+            task_times = {}
+            task_idx = {}
             
-            for idx, (task_id, task) in enumerate(self.tasks.items(), 1):
-                agent = task.agent
-                agent_name = agent.display_name if agent else "Unknown"
-                agent_model = getattr(agent, 'llm', 'gpt-4o-mini') if agent else "unknown"
-                
-                # Show agent task panel with model info
-                task_desc = task.description[:100] + "..." if len(task.description) > 100 else task.description
-                panel_content = f"[bold {PRAISON_COLORS['task_text']}]📋 Task:[/] {task_desc}\n"
-                panel_content += f"[dim]🤖 Model: {agent_model}[/dim]"
-                console.print(Panel.fit(
-                    panel_content,
-                    title=f"[bold]Agent [{idx}/{total_agents}]: {agent_name}[/]",
-                    border_style=PRAISON_COLORS["task"]
-                ))
-                
-                # Execute with timing and status
-                start_time = time_module.time()
-                
-                # Show working spinner
-                with console.status(
-                    f"[bold yellow]Working...[/]  {agent_name} generating response...",
-                    spinner="dots",
-                    spinner_style="yellow"
-                ):
-                    # Run the task
-                    if self.planning:
-                        self._run_with_planning()
-                        break  # Planning mode handles all tasks
-                    else:
-                        self.run_task(task_id)
-                
-                elapsed = time_module.time() - start_time
-                
-                # Show response panel - FULL response, no truncation
-                result = self.get_task_result(task_id)
-                if result:
-                    response_text = str(result.raw)
-                    # No truncation - show full response in verbose mode
-                    from rich.markdown import Markdown
-                    console.print(Panel(
-                        Markdown(response_text),
-                        title=f"[bold]Agent [{idx}/{total_agents}] Complete ({elapsed:.1f}s)[/]",
-                        border_style=PRAISON_COLORS["response"],
-                        padding=(1, 2)
+            # Set up index mapping for tasks
+            for idx, task_id in enumerate(self.tasks.keys(), 1):
+                task_idx[task_id] = idx
+            
+            # Define callbacks to display progress while using proper orchestration
+            def verbose_task_start_callback(task, task_id):
+                try:
+                    agent = task.agent
+                    agent_name = agent.display_name if agent else "Unknown"
+                    agent_model = getattr(agent, 'llm', 'gpt-4o-mini') if agent else "unknown"
+                    idx = task_idx.get(task_id, 0)
+                    
+                    # Show agent task panel with model info
+                    task_desc = task.description[:100] + "..." if len(task.description) > 100 else task.description
+                    panel_content = f"[bold {PRAISON_COLORS['task_text']}]📋 Task:[/] {task_desc}\n"
+                    panel_content += f"[dim]🤖 Model: {agent_model}[/dim]"
+                    console.print(Panel.fit(
+                        panel_content,
+                        title=f"[bold]Agent [{idx}/{total_agents}]: {agent_name}[/]",
+                        border_style=PRAISON_COLORS["task"]
                     ))
-                console.print()
+                    
+                    # Store start time for this task
+                    task_times[task_id] = time_module.time()
+                    
+                    # Show working spinner
+                    console.print(f"[bold yellow]Working...[/]  {agent_name} generating response...")
+                except Exception as e:
+                    logging.debug(f"Error in verbose task start callback: {e}")
+            
+            def verbose_task_complete_callback(task, task_output):
+                try:
+                    task_id = getattr(task, 'id', 'unknown')
+                    start_time = task_times.get(task_id, time_module.time())
+                    elapsed = time_module.time() - start_time
+                    idx = task_idx.get(task_id, 0)
+                    
+                    # Show response panel - FULL response, no truncation
+                    if task_output:
+                        response_text = str(task_output.raw)
+                        # No truncation - show full response in verbose mode
+                        from rich.markdown import Markdown
+                        console.print(Panel(
+                            Markdown(response_text),
+                            title=f"[bold]Agent [{idx}/{total_agents}] Complete ({elapsed:.1f}s)[/]",
+                            border_style=PRAISON_COLORS["response"],
+                            padding=(1, 2)
+                        ))
+                    console.print()
+                except Exception as e:
+                    logging.debug(f"Error in verbose task complete callback: {e}")
+            
+            # Set callbacks for verbose display (compose with existing callbacks)
+            original_on_task_start = self.on_task_start
+            original_on_task_complete = self.on_task_complete
+            
+            def composed_on_task_start(task):
+                try:
+                    verbose_task_start_callback(task)
+                except Exception as e:
+                    logging.debug(f"Error in verbose task start callback: {e}")
+                if original_on_task_start:
+                    try:
+                        original_on_task_start(task)
+                    except Exception as e:
+                        logging.debug(f"Error in original task start callback: {e}")
+            
+            def composed_on_task_complete(task):
+                try:
+                    verbose_task_complete_callback(task)
+                except Exception as e:
+                    logging.debug(f"Error in verbose task complete callback: {e}")
+                if original_on_task_complete:
+                    try:
+                        original_on_task_complete(task)
+                    except Exception as e:
+                        logging.debug(f"Error in original task complete callback: {e}")
+            
+            self.on_task_start = composed_on_task_start
+            self.on_task_complete = composed_on_task_complete
+            
+            # Use proper process orchestration with verbose callbacks
+            try:
+                if self.planning:
+                    self._run_with_planning()
+                else:
+                    self.run_all_tasks()
+            finally:
+                # Restore original callbacks
+                self.on_task_start = original_on_task_start
+                self.on_task_complete = original_on_task_complete
             
             # Workflow summary panel
             total_elapsed = time_module.time() - workflow_start_time
@@ -2519,6 +2696,86 @@ class AgentTeam:
         self._plan_tasks = self.tasks.copy()
         # Keep plan tasks for results but note original tasks are preserved in _plan_tasks
 
+    # Resource Lifecycle Management
+    def close(self) -> None:
+        """Close all agent resources and cleanup connections.
+        
+        This method ensures proper cleanup of:
+        - Agent resources (connections, file handles)
+        - Shared memory connections (SQLite, ChromaDB, MongoDB)
+        - Context manager resources
+        """
+        from .._logging import get_logger
+        logger = get_logger(__name__)
+        
+        # Close all agents
+        for agent in self.agents:
+            try:
+                if hasattr(agent, 'close') and callable(agent.close):
+                    agent.close()
+                    logger.debug(f"Closed resources for agent: {agent.name}")
+            except Exception as e:
+                logger.warning(f"Agent {getattr(agent, 'name', 'unknown')} cleanup failed: {e}")
+        
+        # Close shared memory resources
+        if hasattr(self, 'shared_memory') and self.shared_memory:
+            try:
+                if hasattr(self.shared_memory, 'close') and callable(self.shared_memory.close):
+                    self.shared_memory.close()
+                    logger.debug("Closed shared memory resources")
+            except Exception as e:
+                logger.warning(f"Shared memory cleanup failed: {e}")
+        
+        # Close context manager if initialized
+        if hasattr(self, '_context_manager') and self._context_manager:
+            try:
+                if hasattr(self._context_manager, 'close') and callable(self._context_manager.close):
+                    self._context_manager.close()
+                    logger.debug("Closed context manager resources")
+            except Exception as e:
+                logger.warning(f"Context manager cleanup failed: {e}")
+
+    def __enter__(self):
+        """Context manager entry point for resource management."""
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        """Context manager exit point - ensures resource cleanup."""
+        self.close()
+
+    async def __aenter__(self):
+        """Async context manager entry point for resource management."""
+        return self
+
+    async def __aexit__(self, exc_type, exc_val, exc_tb):
+        """Async context manager exit point - ensures async resource cleanup."""
+        from .._logging import get_logger
+        logger = get_logger(__name__)
+        
+        # Close all agents async if they support it
+        for agent in self.agents:
+            try:
+                if hasattr(agent, 'aclose') and callable(agent.aclose):
+                    await agent.aclose()
+                    logger.debug(f"Async closed resources for agent: {agent.name}")
+                elif hasattr(agent, 'close') and callable(agent.close):
+                    agent.close()
+                    logger.debug(f"Closed resources for agent: {agent.name}")
+            except Exception as e:
+                logger.warning(f"Agent {getattr(agent, 'name', 'unknown')} async cleanup failed: {e}")
+        
+        # Close shared memory resources (async if supported)
+        if hasattr(self, 'shared_memory') and self.shared_memory:
+            try:
+                if hasattr(self.shared_memory, 'aclose') and callable(self.shared_memory.aclose):
+                    await self.shared_memory.aclose()
+                    logger.debug("Async closed shared memory resources")
+                elif hasattr(self.shared_memory, 'close') and callable(self.shared_memory.close):
+                    self.shared_memory.close()
+                    logger.debug("Closed shared memory resources")
+            except Exception as e:
+                logger.warning(f"Shared memory async cleanup failed: {e}")
+
 # Backward compatibility aliases (silent - no deprecation warnings)
 # AgentTeam is the primary name (v1.0+)
 # AgentManager, Agents, PraisonAIAgents are silent aliases

src/praisonai-agents/praisonaiagents/agents/delegator.py

@@ -9,6 +9,7 @@ import logging
 from praisonaiagents._logging import get_logger
 import asyncio
 import warnings
+import threading
 from dataclasses import dataclass, field
 from typing import Any, Dict, List, Optional, Callable
 from enum import Enum
@@ -147,6 +148,7 @@ class SubagentDelegator:
         # Note: asyncio.Semaphore() is created lazily via _get_semaphore() to support
         # Python 3.9 where Semaphore() requires an event loop at creation time.
         self.__semaphore: Optional[asyncio.Semaphore] = None
+        self.__semaphore_lock: threading.Lock = threading.Lock()  # Thread-safe initialization
     
     def _get_semaphore(self) -> asyncio.Semaphore:
         """Get the semaphore, creating it lazily if needed.
@@ -155,7 +157,9 @@ class SubagentDelegator:
         where asyncio.Semaphore() calls get_event_loop() at creation time.
         """
         if self.__semaphore is None:
-            self.__semaphore = asyncio.Semaphore(self.config.max_concurrent_subagents)
+            with self.__semaphore_lock:  # Thread-safe initialization
+                if self.__semaphore is None:  # Double-check after acquiring lock
+                    self.__semaphore = asyncio.Semaphore(self.config.max_concurrent_subagents)
         return self.__semaphore
     
     def get_available_agents(self) -> List[str]:

src/praisonai-agents/praisonaiagents/approval/__init__.py

@@ -183,14 +183,21 @@ def require_approval(risk_level: RiskLevel = "high"):
                 mark_approved(tool_name)
                 return func(*args, **kwargs)
             try:
-                try:
-                    asyncio.get_running_loop()
-                    raise RuntimeError("Use sync fallback in async context")
-                except RuntimeError:
-                    decision = asyncio.run(request_approval(tool_name, kwargs))
+                from ..utils.async_bridge import is_async_context, run_coroutine_from_any_context
+                registry = get_approval_registry()
+                if is_async_context():
+                    # We're in an async context - cannot safely use console I/O
+                    raise RuntimeError(
+                        f"Tool '{tool_name}' requires approval but cannot use console I/O from async context. "
+                        f"Configure a non-console approval backend or call from sync context."
+                    )
+                else:
+                    # Safe to run async approval using registry
+                    decision = run_coroutine_from_any_context(request_approval(tool_name, kwargs))
             except Exception as e:
-                logging.warning(f"Async approval failed, using sync fallback: {e}")
-                decision = console_approval_callback(tool_name, kwargs, risk_level)
+                logging.warning(f"Approval request failed: {e}", exc_info=True)
+                # Fail closed - do not fall back to console
+                raise PermissionError(f"Approval request failed for {tool_name}: {e}") from e
             if not decision.approved:
                 raise PermissionError(f"Execution of {tool_name} denied: {decision.reason}")
             mark_approved(tool_name)

src/praisonai-agents/praisonaiagents/approval/registry.py

@@ -47,16 +47,31 @@ DEFAULT_DANGEROUS_TOOLS: Dict[str, str] = {
 }
 
 # Permission presets — resolved to deny frozensets at Agent.__init__ time.
-# Usage: Agent(approval="safe")
+# Usage: Agent(approval="safe") — or set PRAISONAI_TOOL_SAFETY=<preset>
+# which applies as the default when no ``approval=`` kwarg is passed.
+#
+# ``default`` is the baseline we apply when nothing is configured: it
+# only blocks operations the LLM should never execute unattended —
+# destructive file ops (delete/move/copy) and arbitrary shell/code
+# execution. Read, create and edit stay allowed because those are
+# what 99% of useful agent workflows need. Users who want the old
+# ``trust the LLM with everything`` behaviour pass ``approval="full"``
+# or set ``PRAISONAI_TOOL_SAFETY=off``; users who want stricter
+# controls can opt into ``approval="safe"`` or ``"read_only"``.
 PERMISSION_PRESETS = {
+    # "default" — blocks delete + shell/code exec. Allows read/create/edit.
+    "default": frozenset({
+        "execute_command", "kill_process", "execute_code", "acp_execute_command",
+        "delete_file", "move_file", "copy_file", "acp_delete_file",
+    }),
     # "safe" — blocks all dangerous tools (file writes, shell exec, etc.)
     "safe": frozenset(DEFAULT_DANGEROUS_TOOLS.keys()),
-    # "read_only" — blocks dangerous tools + write operations
-    "read_only": frozenset(DEFAULT_DANGEROUS_TOOLS.keys()) | frozenset({
-        "write_file", "copy_file", "move_file",
-    }),
-    # "full" — no restrictions
+    # "read_only" — alias of "safe" (blocks all dangerous tools)
+    "read_only": frozenset(DEFAULT_DANGEROUS_TOOLS.keys()),
+    # "full" — no restrictions (trust the LLM). Equivalent to "off" env.
     "full": frozenset(),
+    # "off" — alias of "full" for the env-var off-switch.
+    "off": frozenset(),
 }
 
 class ApprovalRegistry:
@@ -76,6 +91,9 @@ class ApprovalRegistry:
         self._required_tools: Set[str] = set()
         self._risk_levels: Dict[str, str] = {}
 
+        # Per-agent, per-tool auto-approval (G-A fix)
+        self._agent_tool_auto_approve: Dict[tuple[str, str], bool] = {}
+
         # Context variables (per-coroutine / per-thread)
         self._approved_context: contextvars.ContextVar[Set[str]] = contextvars.ContextVar(
             "approved_context", default=set()
@@ -139,6 +157,20 @@ class ApprovalRegistry:
     def get_risk_level(self, tool_name: str) -> Optional[str]:
         return self._risk_levels.get(tool_name)
 
+    # ── Per-tool auto-approval (G-A fix) ─────────────────────────────────
+
+    def auto_approve_tool(self, tool_name: str, agent_name: str) -> None:
+        """Pre-approve a single tool for a specific agent."""
+        if not agent_name:
+            raise ValueError("Skill auto-approval requires a stable agent/session scope")
+        self._agent_tool_auto_approve[(agent_name, tool_name)] = True
+
+    def is_auto_approved(self, tool_name: str, agent_name: str) -> bool:
+        """Check if a tool is auto-approved for a specific agent."""
+        if not agent_name:
+            return False
+        return self._agent_tool_auto_approve.get((agent_name, tool_name), False)
+
     # ── Context helpers ──────────────────────────────────────────────────
 
     def mark_approved(self, tool_name: str) -> None:
@@ -189,6 +221,11 @@ class ApprovalRegistry:
         if self.is_already_approved(tool_name):
             return ApprovalDecision(approved=True, reason="Already approved in context")
 
+        # Check per-tool auto-approval (G-A fix)
+        if self.is_auto_approved(tool_name, agent_name):
+            self.mark_approved(tool_name)
+            return ApprovalDecision(approved=True, reason="Auto-approved (skill)", approver="skill")
+
         # Env auto-approve
         if self.is_env_auto_approve():
             self.mark_approved(tool_name)
@@ -237,6 +274,11 @@ class ApprovalRegistry:
         if self.is_already_approved(tool_name):
             return ApprovalDecision(approved=True, reason="Already approved in context")
 
+        # Check per-tool auto-approval (G-A fix)
+        if self.is_auto_approved(tool_name, agent_name):
+            self.mark_approved(tool_name)
+            return ApprovalDecision(approved=True, reason="Auto-approved (skill)", approver="skill")
+
         if self.is_env_auto_approve():
             self.mark_approved(tool_name)
             return ApprovalDecision(approved=True, reason="Auto-approved (env)", approver="env")

src/praisonai-agents/praisonaiagents/auth/__init__.py

@@ -14,6 +14,13 @@ Quick start::
 
         async def authorize(self, identity, resource, action):
             return "owner" in identity.roles
+
+Subscription auth (lazy-loaded)::
+
+    from praisonaiagents.auth import resolve_subscription_credentials
+    
+    creds = resolve_subscription_credentials("claude-code")
+    print(creds.api_key, creds.headers)
 """
 
 from .protocols import (
@@ -25,7 +32,36 @@ from .protocols import (
 
 __all__ = [
     "AuthIdentity",
-    "AuthBackendProtocol",
+    "AuthBackendProtocol", 
     "WorkspaceContextProtocol",
     "AuthConfig",
+    # Subscription auth (lazy-loaded)
+    "SubscriptionAuthProtocol",
+    "SubscriptionCredentials",
+    "register_subscription_provider",
+    "list_subscription_providers",
+    "resolve_subscription_credentials",
+    "get_subscription_provider",
 ]
+
+
+def __getattr__(name):
+    """Lazy-load subscription auth to avoid import overhead."""
+    if name in (
+        "SubscriptionAuthProtocol", 
+        "SubscriptionCredentials",
+        "register_subscription_provider",
+        "list_subscription_providers", 
+        "resolve_subscription_credentials",
+        "get_subscription_provider",
+    ):
+        from .subscription import (
+            SubscriptionAuthProtocol,
+            SubscriptionCredentials,
+            register_subscription_provider,
+            list_subscription_providers,
+            resolve_subscription_credentials,
+            get_subscription_provider,
+        )
+        return locals()[name]
+    raise AttributeError(f"module '{__name__}' has no attribute '{name}'")

src/praisonai-agents/praisonaiagents/auth/subscription/__init__.py

@@ -0,0 +1,35 @@
+"""Subscription / OAuth auth for LLM providers.
+
+Opt-in, zero-overhead. Lazy-loaded via __getattr__ so importing
+praisonaiagents stays fast.
+"""
+from __future__ import annotations
+
+__all__ = [
+    "SubscriptionAuthProtocol",
+    "SubscriptionCredentials",
+    "register_subscription_provider",
+    "list_subscription_providers",
+    "resolve_subscription_credentials",
+    "get_subscription_provider",
+]
+
+
+def __getattr__(name):
+    if name in ("SubscriptionAuthProtocol", "SubscriptionCredentials"):
+        from .protocols import SubscriptionAuthProtocol, SubscriptionCredentials
+        return locals()[name]
+    if name in (
+        "register_subscription_provider",
+        "list_subscription_providers",
+        "resolve_subscription_credentials",
+        "get_subscription_provider",
+    ):
+        from .registry import (
+            register_subscription_provider,
+            list_subscription_providers,
+            resolve_subscription_credentials,
+            get_subscription_provider,
+        )
+        return locals()[name]
+    raise AttributeError(name)

src/praisonai-agents/praisonaiagents/auth/subscription/claude_code.py

@@ -0,0 +1,202 @@
+"""Claude Code OAuth — read ~/.claude/.credentials.json or macOS Keychain."""
+from __future__ import annotations
+
+import json
+import os
+import re
+import subprocess
+import sys
+import time
+import urllib.parse
+import urllib.request
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+from .protocols import AuthError, SubscriptionAuthProtocol, SubscriptionCredentials
+from .registry import register_subscription_provider
+
+_OAUTH_CLIENT_ID = "9d1c250a-e61b-44d9-88ed-5944d1962f5e"
+_TOKEN_ENDPOINTS = (
+    "https://platform.claude.com/v1/oauth/token",
+    "https://console.anthropic.com/v1/oauth/token",
+)
+_CLAUDE_CODE_VERSION_FALLBACK = "2.1.74"
+
+
+def _detect_claude_code_version() -> str:
+    """Detect the installed Claude Code version, fall back to a static constant.
+    
+    Anthropic's OAuth infrastructure validates the user-agent version and may
+    reject requests with a version that's too old. Detecting dynamically means
+    users who keep Claude Code updated never hit stale-version 400s.
+    """
+    cached = getattr(_detect_claude_code_version, "_cache", None)
+    if cached:
+        return cached
+    try:
+        out = subprocess.run(
+            ["claude", "--version"], capture_output=True, text=True, timeout=2, check=False
+        )
+        match = re.match(r"(\d+\.\d+\.\d+)", out.stdout.strip())
+        version = match.group(1) if match else _CLAUDE_CODE_VERSION_FALLBACK
+    except Exception:
+        version = _CLAUDE_CODE_VERSION_FALLBACK
+    _detect_claude_code_version._cache = version
+    return version
+
+
+_CLI_USER_AGENT_FALLBACK = f"claude-cli/{_CLAUDE_CODE_VERSION_FALLBACK} (external, cli)"
+
+
+def _read_keychain_credentials() -> Optional[Dict[str, Any]]:
+    """Read 'Claude Code-credentials' entry from macOS Keychain.
+
+    Returns dict with accessToken/refreshToken/expiresAt or None.
+    """
+    if sys.platform != "darwin":
+        return None
+    import subprocess
+    try:
+        out = subprocess.run(
+            ["security", "find-generic-password", "-s", "Claude Code-credentials", "-w"],
+            capture_output=True, text=True, timeout=5, check=False,
+        )
+        if out.returncode != 0 or not out.stdout.strip():
+            return None
+        data = json.loads(out.stdout)
+        oauth = data.get("claudeAiOauth") or {}
+        if oauth.get("accessToken"):
+            return {
+                "accessToken":  oauth["accessToken"],
+                "refreshToken": oauth.get("refreshToken", ""),
+                "expiresAt":    int(oauth.get("expiresAt", 0)),
+                "source":       "claude-code-keychain",
+            }
+    except Exception:
+        return None
+    return None
+
+
+def _read_file_credentials() -> Optional[Dict[str, Any]]:
+    """Read ~/.claude/.credentials.json."""
+    path = Path.home() / ".claude" / ".credentials.json"
+    if not path.exists():
+        return None
+    try:
+        data = json.loads(path.read_text(encoding="utf-8"))
+    except Exception:
+        return None
+    oauth = data.get("claudeAiOauth") or {}
+    if not oauth.get("accessToken"):
+        return None
+    return {
+        "accessToken":  oauth["accessToken"],
+        "refreshToken": oauth.get("refreshToken", ""),
+        "expiresAt":    int(oauth.get("expiresAt", 0)),
+        "source":       "claude-code-file",
+    }
+
+
+def _is_expiring(expires_at_ms: Optional[int], skew_ms: int = 60_000) -> bool:
+    if not expires_at_ms:
+        return False
+    return int(time.time() * 1000) >= (expires_at_ms - skew_ms)
+
+
+def _refresh(refresh_token: str) -> Dict[str, Any]:
+    """Pure refresh — does not touch local files."""
+    if not refresh_token:
+        raise AuthError("no refresh_token; please re-run 'claude /login'")
+    body = urllib.parse.urlencode({
+        "grant_type": "refresh_token",
+        "refresh_token": refresh_token,
+        "client_id": _OAUTH_CLIENT_ID,
+    }).encode()
+    last = None
+    for endpoint in _TOKEN_ENDPOINTS:
+        req = urllib.request.Request(
+            endpoint, data=body, method="POST",
+            headers={
+                "Content-Type": "application/x-www-form-urlencoded",
+                "User-Agent":   f"claude-cli/{_detect_claude_code_version()} (external, cli)",
+            },
+        )
+        try:
+            with urllib.request.urlopen(req, timeout=10) as resp:
+                payload = json.loads(resp.read().decode())
+            return {
+                "accessToken":  payload["access_token"],
+                "refreshToken": payload.get("refresh_token", refresh_token),
+                "expiresAt":    int(time.time() * 1000) + int(payload.get("expires_in", 3600)) * 1000,
+            }
+        except Exception as exc:
+            last = exc
+    raise AuthError(f"Anthropic refresh failed: {last}")
+
+
+class ClaudeCodeAuth:
+    """Claude Code OAuth subscription auth."""
+
+    name = "claude-code"
+
+    def resolve_credentials(self) -> SubscriptionCredentials:
+        # 1. ANTHROPIC_TOKEN / CLAUDE_CODE_OAUTH_TOKEN env (Hermes-style)
+        for env_key in ("ANTHROPIC_TOKEN", "CLAUDE_CODE_OAUTH_TOKEN"):
+            tok = os.environ.get(env_key, "").strip()
+            if tok:
+                return SubscriptionCredentials(
+                    api_key=tok,
+                    base_url="https://api.anthropic.com",
+                    headers=self.headers_for("https://api.anthropic.com", ""),
+                    auth_scheme="bearer",
+                    source=f"env:{env_key}",
+                )
+
+        # 2. Keychain or file
+        creds = _read_keychain_credentials() or _read_file_credentials()
+        if not creds:
+            raise AuthError(
+                "No Claude Code credentials found. Install Claude Code and run "
+                "'claude /login', or set ANTHROPIC_TOKEN."
+            )
+        if _is_expiring(creds["expiresAt"]) and creds.get("refreshToken"):
+            creds.update(_refresh(creds["refreshToken"]))
+        return SubscriptionCredentials(
+            api_key=creds["accessToken"],
+            base_url="https://api.anthropic.com",
+            headers=self.headers_for("https://api.anthropic.com", ""),
+            auth_scheme="bearer",
+            expires_at_ms=creds.get("expiresAt"),
+            source=creds["source"],
+        )
+
+    def refresh(self) -> SubscriptionCredentials:
+        creds = _read_keychain_credentials() or _read_file_credentials()
+        if not creds or not creds.get("refreshToken"):
+            raise AuthError("cannot refresh: no refresh token available")
+        new = _refresh(creds["refreshToken"])
+        return SubscriptionCredentials(
+            api_key=new["accessToken"],
+            base_url="https://api.anthropic.com",
+            headers=self.headers_for("https://api.anthropic.com", ""),
+            auth_scheme="bearer",
+            expires_at_ms=new["expiresAt"],
+            source="claude-code-refreshed",
+        )
+
+    def headers_for(self, base_url: str, model: str) -> Dict[str, str]:
+        # CRITICAL: without these headers Anthropic returns 500s on OAuth tokens.
+        # Match Hermes' OAuth header set exactly. litellm auto-adds oauth-2025-04-20.
+        return {
+            "anthropic-beta": ",".join([
+                "interleaved-thinking-2025-05-14",
+                "fine-grained-tool-streaming-2025-05-14", 
+                "context-1m-2025-08-07",
+                "claude-code-20250219",
+            ]),
+            "user-agent":     f"claude-cli/{_detect_claude_code_version()} (external, cli)",
+            "x-app":          "cli",
+        }
+
+
+register_subscription_provider("claude-code", lambda: ClaudeCodeAuth())

src/praisonai-agents/praisonaiagents/auth/subscription/codex.py

@@ -0,0 +1,69 @@
+"""Codex / ChatGPT subscription — read ~/.codex/auth.json (or CODEX_HOME)."""
+from __future__ import annotations
+
+import json
+import os
+import time
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+from .protocols import AuthError, SubscriptionCredentials
+from .registry import register_subscription_provider
+
+
+_CODEX_BASE_URL = "https://chatgpt.com/backend-api/codex"
+
+
+def _read_codex_tokens() -> Optional[Dict[str, Any]]:
+    home = os.environ.get("CODEX_HOME", "").strip() or str(Path.home() / ".codex")
+    auth_path = Path(home).expanduser() / "auth.json"
+    if not auth_path.is_file():
+        return None
+    try:
+        payload = json.loads(auth_path.read_text(encoding="utf-8"))
+    except Exception:
+        return None
+    tokens = payload.get("tokens") or {}
+    if not tokens.get("access_token") or not tokens.get("refresh_token"):
+        return None
+    return tokens
+
+
+def _is_jwt_expiring(jwt: str, skew_seconds: int = 60) -> bool:
+    """Decode JWT exp claim without verifying signature."""
+    import base64
+    try:
+        _hdr, payload_b64, _sig = jwt.split(".")
+        pad = "=" * (-len(payload_b64) % 4)
+        payload = json.loads(base64.urlsafe_b64decode(payload_b64 + pad))
+        exp = int(payload.get("exp", 0))
+        return exp and (time.time() + skew_seconds) >= exp
+    except Exception:
+        return False  # treat unknown as not-expiring; refresh on actual 401
+
+
+class CodexAuth:
+    name = "codex"
+
+    def resolve_credentials(self) -> SubscriptionCredentials:
+        # EXPERIMENTAL: Codex uses Responses API, not Chat Completions
+        # OpenAI chat-completions requests will 404 against this endpoint
+        raise AuthError(
+            "Codex auth is experimental and requires custom Responses API transport. "
+            "Use a standard OpenAI API key or wait for proper Codex support in a future release."
+        )
+
+    def refresh(self) -> SubscriptionCredentials:
+        # OpenAI Codex tokens are refreshed by the Codex CLI itself; we rely
+        # on the CLI being run periodically. A direct refresh would require
+        # replicating their device-flow client — out of scope for v1.
+        return self.resolve_credentials()
+
+    def headers_for(self, base_url: str, model: str) -> Dict[str, str]:
+        return {
+            "openai-beta": "responses=v1",
+            "user-agent":  "codex-cli/0.1.0 (external)",
+        }
+
+
+register_subscription_provider("codex", lambda: CodexAuth())

src/praisonai-agents/praisonaiagents/auth/subscription/gemini_cli.py

@@ -0,0 +1,51 @@
+"""Gemini CLI OAuth — read ~/.gemini/oauth_creds.json."""
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+from .protocols import AuthError, SubscriptionCredentials
+from .registry import register_subscription_provider
+
+
+_GEMINI_BASE_URL = "https://cloudcode-pa.googleapis.com/v1internal"
+
+
+def _read_gemini_tokens() -> Optional[Dict[str, Any]]:
+    """Read Gemini CLI OAuth credentials."""
+    auth_path = Path.home() / ".gemini" / "oauth_creds.json"
+    if not auth_path.is_file():
+        return None
+    try:
+        payload = json.loads(auth_path.read_text(encoding="utf-8"))
+    except Exception:
+        return None
+    if not payload.get("access_token"):
+        return None
+    return payload
+
+
+class GeminiCliAuth:
+    name = "gemini-cli"
+
+    def resolve_credentials(self) -> SubscriptionCredentials:
+        # EXPERIMENTAL: Gemini CLI uses Code Assist API, not OpenAI chat format
+        # Standard OpenAI chat-completions requests will not work with cloudcode-pa endpoints
+        raise AuthError(
+            "Gemini CLI auth is experimental and requires custom Code Assist transport. "
+            "Use standard Gemini API keys or wait for proper Code Assist support in a future release."
+        )
+
+    def refresh(self) -> SubscriptionCredentials:
+        # Gemini CLI handles refresh internally
+        return self.resolve_credentials()
+
+    def headers_for(self, base_url: str, model: str) -> Dict[str, str]:
+        return {
+            "user-agent": "gemini-cli/1.0.0 (external)",
+        }
+
+
+register_subscription_provider("gemini-cli", lambda: GeminiCliAuth())

src/praisonai-agents/praisonaiagents/auth/subscription/protocols.py

@@ -0,0 +1,41 @@
+"""Subscription-auth protocol — what every provider must implement."""
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any, Dict, Optional, Protocol, runtime_checkable
+
+
+@dataclass
+class SubscriptionCredentials:
+    """Resolved credentials ready to inject into an LLM client."""
+    api_key: str                            # access_token (or "" for SDK-default chains like AWS)
+    base_url: str = ""                       # provider endpoint
+    headers: Dict[str, str] = field(default_factory=dict)  # extra headers (UA, x-app, betas)
+    auth_scheme: str = "bearer"              # "bearer" | "x-api-key" | "sdk-default"
+    expires_at_ms: Optional[int] = None     # epoch ms; None = no expiry tracked
+    source: str = ""                         # "claude-code-keychain", "codex-cli-file", etc.
+
+
+@runtime_checkable
+class SubscriptionAuthProtocol(Protocol):
+    """Discover, refresh, and inject subscription credentials."""
+
+    def resolve_credentials(self) -> SubscriptionCredentials:
+        """Return current valid credentials, refreshing if expiring soon.
+
+        Raises:
+            AuthError: if no usable credentials can be resolved.
+        """
+        ...
+
+    def refresh(self) -> SubscriptionCredentials:
+        """Force a refresh and return the new credentials."""
+        ...
+
+    def headers_for(self, base_url: str, model: str) -> Dict[str, str]:
+        """Provider-specific headers (UA, beta flags) keyed off endpoint/model."""
+        ...
+
+
+class AuthError(RuntimeError):
+    """Raised when subscription credentials cannot be resolved."""

src/praisonai-agents/praisonaiagents/auth/subscription/qwen_cli.py

@@ -0,0 +1,58 @@
+"""Qwen CLI OAuth — read ~/.qwen/oauth_creds.json."""
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+from .protocols import AuthError, SubscriptionCredentials
+from .registry import register_subscription_provider
+
+
+_QWEN_BASE_URL = "https://portal.qwen.ai/v1"
+
+
+def _read_qwen_tokens() -> Optional[Dict[str, Any]]:
+    """Read Qwen CLI OAuth credentials."""
+    auth_path = Path.home() / ".qwen" / "oauth_creds.json"
+    if not auth_path.is_file():
+        return None
+    try:
+        payload = json.loads(auth_path.read_text(encoding="utf-8"))
+    except Exception:
+        return None
+    if not payload.get("access_token"):
+        return None
+    return payload
+
+
+class QwenCliAuth:
+    name = "qwen-cli"
+
+    def resolve_credentials(self) -> SubscriptionCredentials:
+        tokens = _read_qwen_tokens()
+        if not tokens:
+            raise AuthError(
+                "No Qwen CLI credentials at ~/.qwen/oauth_creds.json. "
+                "Install Qwen CLI and run 'qwen login'."
+            )
+        return SubscriptionCredentials(
+            api_key=tokens["access_token"],
+            base_url=os.environ.get("PRAISONAI_QWEN_BASE_URL", "").strip() or _QWEN_BASE_URL,
+            headers=self.headers_for(_QWEN_BASE_URL, ""),
+            auth_scheme="bearer",
+            source="qwen-cli-file",
+        )
+
+    def refresh(self) -> SubscriptionCredentials:
+        # Qwen CLI handles refresh internally
+        return self.resolve_credentials()
+
+    def headers_for(self, base_url: str, model: str) -> Dict[str, str]:
+        return {
+            "user-agent": "qwen-cli/1.0.0 (external)",
+        }
+
+
+register_subscription_provider("qwen-cli", lambda: QwenCliAuth())

src/praisonai-agents/praisonaiagents/auth/subscription/registry.py

@@ -0,0 +1,64 @@
+"""Thread-safe provider registry for subscription auth."""
+from __future__ import annotations
+
+import threading
+from typing import Callable, Dict, List, Optional
+
+from .protocols import AuthError, SubscriptionAuthProtocol, SubscriptionCredentials
+
+_REGISTRY: Dict[str, Callable[[], SubscriptionAuthProtocol]] = {}
+_LOCK = threading.RLock()
+_BUILTIN_REGISTERED = False
+
+
+def register_subscription_provider(
+    provider_id: str,
+    factory: Callable[[], SubscriptionAuthProtocol],
+) -> None:
+    """Register a new subscription-auth provider.
+
+    Args:
+        provider_id: short id e.g. "claude-code", "codex", "gemini-cli".
+        factory: zero-arg callable returning a SubscriptionAuthProtocol impl.
+    """
+    with _LOCK:
+        _REGISTRY[provider_id] = factory
+
+
+def list_subscription_providers() -> List[str]:
+    _ensure_builtins()
+    with _LOCK:
+        return sorted(_REGISTRY)
+
+
+def resolve_subscription_credentials(provider_id: str) -> SubscriptionCredentials:
+    """Look up provider by id and resolve its credentials."""
+    _ensure_builtins()
+    with _LOCK:
+        factory = _REGISTRY.get(provider_id)
+    if factory is None:
+        raise AuthError(
+            f"Unknown subscription provider {provider_id!r}. "
+            f"Registered: {list_subscription_providers()}"
+        )
+    return factory().resolve_credentials()
+
+
+def get_subscription_provider(provider_id: str) -> Optional[SubscriptionAuthProtocol]:
+    """Return a new provider instance for *provider_id*, or None if not registered."""
+    _ensure_builtins()
+    with _LOCK:
+        factory = _REGISTRY.get(provider_id)
+    return factory() if factory is not None else None
+
+
+def _ensure_builtins() -> None:
+    """Lazy register built-in providers on first use."""
+    global _BUILTIN_REGISTERED
+    with _LOCK:
+        if _BUILTIN_REGISTERED:
+            return
+        # Import modules inside the lock to prevent race conditions
+        from . import claude_code, codex, gemini_cli, qwen_cli   # noqa: F401
+        # Each module registers itself at import time.
+        _BUILTIN_REGISTERED = True

src/praisonai-agents/praisonaiagents/background/runner.py

@@ -6,6 +6,7 @@ Manages background task execution.
 
 import asyncio
 import logging
+import threading
 from praisonaiagents._logging import get_logger
 from typing import Optional, List, Dict, Any, Callable, Union
 from datetime import datetime
@@ -53,13 +54,16 @@ class BackgroundRunner:
         self._tasks: Dict[str, BackgroundTask] = {}
         # Lazily create asyncio.Semaphore() to avoid Python 3.9 event loop issues
         self._semaphore: Optional[asyncio.Semaphore] = None
+        self._semaphore_lock: threading.Lock = threading.Lock()  # Thread-safe initialization
         self._running = False
         self._cleanup_task: Optional[asyncio.Task] = None
     
     def _get_semaphore(self) -> asyncio.Semaphore:
         """Get or create the semaphore lazily (Python 3.9 compatible)."""
         if self._semaphore is None:
-            self._semaphore = asyncio.Semaphore(self.config.max_concurrent_tasks)
+            with self._semaphore_lock:  # Thread-safe initialization
+                if self._semaphore is None:  # Double-check after acquiring lock
+                    self._semaphore = asyncio.Semaphore(self.config.max_concurrent_tasks)
         return self._semaphore
     
     @property
@@ -403,7 +407,7 @@ def _get_bg_loop() -> asyncio.AbstractEventLoop:
     """Return a running event loop on a background daemon thread.
 
     Created once on first call; reused thereafter.  The thread is a
-    daemon so it won't prevent process exit.
+    daemon so it won't prevent process exit, but includes graceful shutdown.
     """
     global _bg_loop
     if _bg_loop is not None and _bg_loop.is_running():
@@ -420,3 +424,31 @@ def _get_bg_loop() -> asyncio.AbstractEventLoop:
         t.start()
 
     return _bg_loop
+
+
+def _shutdown_bg_loop():
+    """Gracefully shutdown the background event loop."""
+    global _bg_loop
+    if _bg_loop is None or not _bg_loop.is_running():
+        return
+
+    async def _cancel_pending():
+        current = asyncio.current_task()
+        pending = [t for t in asyncio.all_tasks() if t is not current]
+        for t in pending:
+            t.cancel()
+        if pending:
+            await asyncio.gather(*pending, return_exceptions=True)
+
+    try:
+        fut = asyncio.run_coroutine_threadsafe(_cancel_pending(), _bg_loop)
+        fut.result(timeout=2)
+    except Exception:
+        pass
+    finally:
+        _bg_loop.call_soon_threadsafe(_bg_loop.stop)
+
+
+# Register cleanup on process exit
+import atexit
+atexit.register(_shutdown_bg_loop)

src/praisonai-agents/praisonaiagents/bots/config.py

@@ -5,7 +5,9 @@ Provides configuration dataclasses for bot settings.
 """
 
 from dataclasses import dataclass, field
-from typing import Any, Dict, List, Optional
+from typing import Any, Dict, List, Literal, Optional
+
+from .pairing_types import UnknownUserPolicy
 
 
 @dataclass
@@ -48,16 +50,32 @@ class BotConfig:
     # Group message policy
     group_policy: str = "mention_only"  # respond_all, mention_only, command_only
     
-    # Default tools enabled for all bots
+    # Default safe tools (auto-injected for bots with no tools configured)
+    # With workspace scoping, file operations are now safe by construction
     default_tools: List[str] = field(default_factory=lambda: [
-        "execute_command", "search_web", "web_crawl",
-        "schedule_add", "schedule_list", "schedule_remove",
+        # Web (existing)
+        "search_web", "web_crawl",
+        # Memory / learning (existing)
         "store_memory", "search_memory",
         "store_learning", "search_learning",
+        # Scheduling (existing)
+        "schedule_add", "schedule_list", "schedule_remove",
+        # Clarify tool (new from main)
+        "clarify",
+        # Files — NEW (workspace-scoped, safe by construction)
+        "read_file", "write_file", "edit_file", "list_files", "search_files",
+        # Planning — NEW
+        "todo_add", "todo_list", "todo_update",
+        # Skills (self-improving) — NEW
+        "skills_list", "skill_view", "skill_manage",
+        # Delegation & session tools are intentionally NOT auto-injected yet;
+        # their reference implementations are placeholders. Users can opt in
+        # via BotConfig(default_tools=[..., "delegate_task", "session_search"]).
     ])
     
     # Auto-approve tool calls (useful for trusted environments)
-    auto_approve_tools: bool = False  # If True, skip confirmation for tool execution
+    # Default to True for chat bots - they can't show CLI approval prompts
+    auto_approve_tools: bool = True  # If True, skip confirmation for safe tool execution
     
     # Inbound message debounce (ms). Coalesces rapid messages from same user.
     # 0 = disabled (default). Recommended: 1000-2000 for chat bots.
@@ -73,6 +91,24 @@ class BotConfig:
     # When set, stale sessions older than this are auto-reaped.
     session_ttl: int = 0
     
+    # Workspace settings for file operation containment and security
+    workspace_dir: Optional[str] = None  # default: ~/.praisonai/workspaces/<scope>/<session_key>
+    workspace_access: str = "rw"  # "rw" (read-write) | "ro" (read-only) | "none" (copy-on-write sandbox)
+    workspace_scope: str = "session"  # "shared" | "session" | "user" | "agent"
+    
+    # Unknown user policy: "deny" (default), "pair", or "allow"
+    unknown_user_policy: UnknownUserPolicy = "deny"
+    
+    # Owner user ID for pairing approvals (platform-specific format)
+    owner_user_id: Optional[str] = None
+
+    def __post_init__(self) -> None:
+        if self.unknown_user_policy not in {"deny", "pair", "allow"}:
+            raise ValueError(
+                f"unknown_user_policy must be one of: deny, pair, allow. Got: {self.unknown_user_policy}"
+            )
+    
+    
     metadata: Dict[str, Any] = field(default_factory=dict)
     
     def to_dict(self) -> Dict[str, Any]:
@@ -100,6 +136,11 @@ class BotConfig:
             "ack_emoji": self.ack_emoji,
             "done_emoji": self.done_emoji,
             "session_ttl": self.session_ttl,
+            "workspace_dir": self.workspace_dir,
+            "workspace_access": self.workspace_access,
+            "workspace_scope": self.workspace_scope,
+            "unknown_user_policy": self.unknown_user_policy,
+            "owner_user_id": "***" if self.owner_user_id else None,
             "metadata": self.metadata,
         }
     

src/praisonai-agents/praisonaiagents/bots/pairing_types.py

@@ -0,0 +1,37 @@
+"""
+Pairing types and protocols for bot user approval.
+
+Contains lightweight types and protocols for the pairing approval system.
+Heavy implementations are in the wrapper package.
+"""
+
+from dataclasses import dataclass
+from typing import Literal, Optional
+
+# Policy for handling unknown users
+UnknownUserPolicy = Literal["deny", "pair", "allow"]
+
+
+@dataclass
+class PairingReply:
+    """Reply data for pairing approval requests."""
+    
+    user_name: str
+    user_id: str
+    channel: str
+    code: str
+    message: str = ""
+    
+    def __post_init__(self):
+        if not self.message:
+            self.message = f"User {self.user_name} wants to chat. Approve access?"
+
+
+@dataclass
+class PairingApprovalResult:
+    """Result of a pairing approval action."""
+    
+    success: bool
+    message: str
+    user_id: Optional[str] = None
+    channel: Optional[str] = None

src/praisonai-agents/praisonaiagents/bus/bus.py

@@ -9,6 +9,7 @@ import logging
 from praisonaiagents._logging import get_logger
 import threading
 from typing import Any, Callable, Dict, List, Optional, Set, Union
+from ..utils.async_bridge import run_coroutine_from_any_context
 from dataclasses import dataclass, field
 from .event import Event, EventType
 
@@ -209,13 +210,13 @@ class EventBus:
         for sub in subscribers:
             try:
                 if sub.is_async:
-                    # Schedule async callback
+                    # Schedule async callback safely
                     try:
                         loop = asyncio.get_running_loop()
                         loop.create_task(sub.callback(event))
                     except RuntimeError:
-                        # No running loop, run synchronously
-                        asyncio.run(sub.callback(event))
+                        # No running loop, use safe async bridge
+                        run_coroutine_from_any_context(sub.callback(event))
                 else:
                     sub.callback(event)
             except Exception as e:

src/praisonai-agents/praisonaiagents/cli_backend/__init__.py

@@ -0,0 +1,37 @@
+"""CLI Backend Protocol for PraisonAI Agents.
+
+Provides protocol-driven core for integrating external CLI tools as agent backends.
+
+This module follows AGENTS.md protocol-driven design:
+- Core SDK contains only protocols and dataclasses
+- Heavy implementations live in wrapper (praisonai package)
+- Zero optional dependencies at import time
+"""
+
+def __getattr__(name: str):
+    """Lazy loading for CLI backend components."""
+    if name == "CliBackendProtocol":
+        from .protocols import CliBackendProtocol
+        return CliBackendProtocol
+    elif name == "CliBackendConfig":
+        from .protocols import CliBackendConfig
+        return CliBackendConfig
+    elif name == "CliSessionBinding":
+        from .protocols import CliSessionBinding
+        return CliSessionBinding
+    elif name == "CliBackendResult":
+        from .protocols import CliBackendResult
+        return CliBackendResult
+    elif name == "CliBackendDelta":
+        from .protocols import CliBackendDelta
+        return CliBackendDelta
+    else:
+        raise AttributeError(f"module '{__name__}' has no attribute '{name}'")
+
+__all__ = [
+    "CliBackendProtocol",
+    "CliBackendConfig", 
+    "CliSessionBinding",
+    "CliBackendResult",
+    "CliBackendDelta"
+]

src/praisonai-agents/praisonaiagents/cli_backend/protocols.py

@@ -0,0 +1,135 @@
+"""CLI Backend protocols and data structures.
+
+Protocol-driven design following AGENTS.md:
+- Lightweight protocols only (no implementations)
+- Dataclasses for configuration
+- Async-first with proper typing
+"""
+
+from dataclasses import dataclass, field
+from typing import Protocol, AsyncIterator, Optional, List, Dict, Any, runtime_checkable
+
+
+@dataclass
+class CliBackendConfig:
+    """Declarative CLI backend configuration.
+    
+    Mirrors OpenClaw's CliBackendConfig for cross-ecosystem compatibility.
+    """
+    # Core command
+    command: str  # e.g., "claude", "codex", "gemini"
+    args: List[str] = field(default_factory=list)
+    
+    # Session management
+    resume_args: Optional[List[str]] = None  # supports "{session_id}" placeholder
+    session_arg: Optional[str] = None  # e.g., "--session-id"
+    session_mode: str = "none"  # "always" | "existing" | "none"
+    session_id_fields: List[str] = field(default_factory=list)
+    
+    # Input/output format
+    output: str = "text"  # "text" | "json" | "jsonl"
+    input: str = "arg"  # "arg" | "stdin"
+    max_prompt_arg_chars: Optional[int] = None
+    
+    # Model configuration
+    model_arg: Optional[str] = None  # e.g., "--model"
+    model_aliases: Dict[str, str] = field(default_factory=dict)
+    
+    # System prompt handling
+    system_prompt_arg: Optional[str] = None  # e.g., "--append-system-prompt"
+    system_prompt_when: str = "always"  # "first" | "always" | "never"
+    system_prompt_mode: str = "append"  # "append" | "replace"
+    
+    # Image support
+    image_arg: Optional[str] = None  # e.g., "--image"
+    image_mode: str = "repeat"  # "repeat" | "list"
+    
+    # Environment
+    clear_env: List[str] = field(default_factory=list)  # Env vars to clear
+    env: Dict[str, str] = field(default_factory=dict)  # Env vars to set
+    
+    # Advanced features
+    live_session: Optional[str] = None  # e.g., "claude-stdio"
+    bundle_mcp: bool = False
+    bundle_mcp_mode: Optional[str] = None  # e.g., "claude-config-file"
+    serialize: bool = False  # Queue operations to avoid conflicts
+    
+    # Timeouts
+    no_output_timeout_ms: Optional[int] = None
+    timeout_ms: int = 300_000
+
+
+@dataclass
+class CliSessionBinding:
+    """Session binding for CLI backend state tracking."""
+    session_id: Optional[str] = None
+    auth_profile_id: Optional[str] = None
+    system_prompt_hash: Optional[str] = None
+    mcp_config_hash: Optional[str] = None
+    is_resume: bool = False  # Whether this is resuming an existing session
+
+
+@dataclass
+class CliBackendResult:
+    """Result from CLI backend execution."""
+    content: str
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    session_id: Optional[str] = None
+    error: Optional[str] = None
+
+
+@dataclass
+class CliBackendDelta:
+    """Streaming delta from CLI backend."""
+    type: str  # "text" | "tool_call" | "thinking" | "error"
+    content: str = ""
+    metadata: Dict[str, Any] = field(default_factory=dict)
+
+
+@runtime_checkable
+class CliBackendProtocol(Protocol):
+    """Protocol for CLI backend implementations.
+    
+    Any object implementing these methods can serve as an Agent backend.
+    Follows typing.Protocol pattern from AGENTS.md.
+    """
+    config: CliBackendConfig
+    
+    async def execute(
+        self, 
+        prompt: str, 
+        *,
+        session: Optional[CliSessionBinding] = None,
+        images: Optional[List[str]] = None,
+        system_prompt: Optional[str] = None,
+        **kwargs
+    ) -> CliBackendResult:
+        """Execute a single prompt and return result.
+        
+        Args:
+            prompt: User prompt/query
+            session: Session binding for state management
+            images: List of image paths for multimodal prompts
+            system_prompt: System prompt override
+            **kwargs: Additional backend-specific options
+            
+        Returns:
+            CliBackendResult with response content and metadata
+        """
+        ...
+    
+    async def stream(
+        self, 
+        prompt: str, 
+        **kwargs
+    ) -> AsyncIterator[CliBackendDelta]:
+        """Stream response deltas from CLI backend.
+        
+        Args:
+            prompt: User prompt/query
+            **kwargs: Additional options (session, images, etc.)
+            
+        Yields:
+            CliBackendDelta objects with incremental response content
+        """
+        ...

src/praisonai-agents/praisonaiagents/context/session_tracker.py

@@ -7,7 +7,7 @@ Inspired by Agno's SessionContextStore pattern.
 
 from dataclasses import dataclass, field, asdict
 from typing import Any, Dict, List, Optional
-from datetime import datetime
+from datetime import datetime, timezone
 import json
 import logging
 from praisonaiagents._logging import get_logger
@@ -97,8 +97,8 @@ class SessionContextTracker:
         
         self._state = SessionState(
             session_id=self.session_id,
-            created_at=datetime.utcnow().isoformat() + "Z",
-            updated_at=datetime.utcnow().isoformat() + "Z",
+            created_at=datetime.now(timezone.utc).isoformat(),
+            updated_at=datetime.now(timezone.utc).isoformat(),
         )
     
     def _generate_session_id(self) -> str:
@@ -163,7 +163,7 @@ class SessionContextTracker:
     
     def _mark_updated(self) -> None:
         """Update the updated_at timestamp."""
-        self._state.updated_at = datetime.utcnow().isoformat() + "Z"
+        self._state.updated_at = datetime.now(timezone.utc).isoformat()
         self._state.turn_count += 1
     
     def to_context_string(self) -> str:
@@ -257,7 +257,7 @@ Current messages take precedence if there's any conflict with this summary.
         self._state = SessionState(
             session_id=self.session_id,
             created_at=self._state.created_at,
-            updated_at=datetime.utcnow().isoformat() + "Z",
+            updated_at=datetime.now(timezone.utc).isoformat(),
         )
     
     def to_dict(self) -> Dict[str, Any]:

src/praisonai-agents/praisonaiagents/eval/grader.py

@@ -12,7 +12,7 @@ import os
 import logging
 from praisonaiagents._logging import get_logger
 from dataclasses import dataclass, field, asdict
-from datetime import datetime
+from datetime import datetime, timezone
 from typing import Any, Dict, List, Optional
 
 logger = get_logger(__name__)
@@ -39,7 +39,7 @@ class GradeResult:
     input_text: str = ""
     output: str = ""
     expected_output: Optional[str] = None
-    timestamp: str = field(default_factory=lambda: datetime.utcnow().isoformat())
+    timestamp: str = field(default_factory=lambda: datetime.now(timezone.utc).isoformat())
     
     def to_dict(self) -> Dict[str, Any]:
         """Convert to dictionary."""
@@ -55,7 +55,7 @@ class GradeResult:
             input_text=data.get("input_text", ""),
             output=data.get("output", ""),
             expected_output=data.get("expected_output"),
-            timestamp=data.get("timestamp", datetime.utcnow().isoformat()),
+            timestamp=data.get("timestamp", datetime.now(timezone.utc).isoformat()),
         )
 
 class BaseLLMGrader:

src/praisonai-agents/praisonaiagents/gateway/__init__.py

@@ -18,8 +18,25 @@ from .protocols import (
     GatewayEvent,
     GatewayMessage,
     EventType,
+    # Push protocols and dataclasses
+    PushChannelProtocol,
+    PresenceProtocol,
+    DeliveryGuaranteeProtocol,
+    ChannelInfo,
+    PresenceInfo,
+)
+from .config import (
+    GatewayConfig,
+    SessionConfig,
+    ChannelRouteConfig,
+    MultiChannelGatewayConfig,
+    # Push config
+    PushConfig,
+    RedisConfig,
+    PresenceConfig,
+    DeliveryConfig,
+    PollingConfig,
 )
-from .config import GatewayConfig, SessionConfig, ChannelRouteConfig, MultiChannelGatewayConfig
 
 # Lazy loading cache
 _lazy_cache = {}
@@ -68,11 +85,22 @@ __all__ = [
     "GatewayEvent",
     "GatewayMessage",
     "EventType",
+    # Push protocols (always available)
+    "PushChannelProtocol",
+    "PresenceProtocol",
+    "DeliveryGuaranteeProtocol",
+    "ChannelInfo",
+    "PresenceInfo",
     # Config (always available)
     "GatewayConfig",
     "SessionConfig",
     "ChannelRouteConfig",
     "MultiChannelGatewayConfig",
+    "PushConfig",
+    "RedisConfig",
+    "PresenceConfig",
+    "DeliveryConfig",
+    "PollingConfig",
     # Implementations (lazy loaded from praisonai wrapper)
     "WebSocketGateway",
     "GatewaySession",

src/praisonai-agents/praisonaiagents/gateway/config.py

@@ -37,6 +37,158 @@ class SessionConfig:
         }
 
 
+# ---------------------------------------------------------------------------
+# Push notification configuration
+# ---------------------------------------------------------------------------
+
+@dataclass
+class RedisConfig:
+    """Redis connection configuration for push service scaling.
+    
+    Attributes:
+        url: Full Redis URL (takes precedence over host/port)
+        host: Redis host
+        port: Redis port
+        db: Redis database number
+        password: Redis password
+        prefix: Key prefix namespace
+        max_connections: Connection pool size
+    """
+    
+    url: Optional[str] = None
+    host: str = "localhost"
+    port: int = 6379
+    db: int = 0
+    password: Optional[str] = None
+    prefix: str = "praison:push:"
+    max_connections: int = 20
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary (hides sensitive data)."""
+        return {
+            "url": "***" if self.url else None,
+            "host": self.host,
+            "port": self.port,
+            "db": self.db,
+            "password": "***" if self.password else None,
+            "prefix": self.prefix,
+            "max_connections": self.max_connections,
+        }
+
+
+@dataclass
+class PresenceConfig:
+    """Configuration for presence tracking.
+    
+    Attributes:
+        enabled: Toggle presence tracking
+        heartbeat_interval: Expected heartbeat frequency from clients (seconds)
+        offline_timeout: Mark offline after this many seconds without heartbeat
+        broadcast_changes: Broadcast presence changes to subscribed channels
+    """
+    
+    enabled: bool = True
+    heartbeat_interval: int = 15
+    offline_timeout: int = 45
+    broadcast_changes: bool = True
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary."""
+        return {
+            "enabled": self.enabled,
+            "heartbeat_interval": self.heartbeat_interval,
+            "offline_timeout": self.offline_timeout,
+            "broadcast_changes": self.broadcast_changes,
+        }
+
+
+@dataclass
+class DeliveryConfig:
+    """Configuration for at-least-once delivery guarantees.
+    
+    Attributes:
+        enabled: Toggle delivery guarantees
+        ack_timeout: Seconds to wait for ACK before retrying
+        max_retries: Maximum retry attempts
+        retry_backoff: Exponential backoff multiplier
+        message_ttl: How long to retain unacknowledged messages (seconds)
+        store_backend: Message store backend ("memory" or "redis")
+    """
+    
+    enabled: bool = True
+    ack_timeout: int = 30
+    max_retries: int = 3
+    retry_backoff: float = 2.0
+    message_ttl: int = 86400
+    store_backend: str = "memory"
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary."""
+        return {
+            "enabled": self.enabled,
+            "ack_timeout": self.ack_timeout,
+            "max_retries": self.max_retries,
+            "retry_backoff": self.retry_backoff,
+            "message_ttl": self.message_ttl,
+            "store_backend": self.store_backend,
+        }
+
+
+@dataclass
+class PollingConfig:
+    """Configuration for HTTP long-polling fallback.
+    
+    Attributes:
+        enabled: Toggle polling fallback
+        long_poll_timeout: Long-poll hang duration (seconds)
+        max_batch_size: Max messages per poll response
+    """
+    
+    enabled: bool = True
+    long_poll_timeout: int = 30
+    max_batch_size: int = 100
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary."""
+        return {
+            "enabled": self.enabled,
+            "long_poll_timeout": self.long_poll_timeout,
+            "max_batch_size": self.max_batch_size,
+        }
+
+
+@dataclass
+class PushConfig:
+    """Top-level configuration for the push notification service.
+    
+    All push capabilities are opt-in. When enabled=False (default),
+    zero overhead is added to the gateway.
+    
+    Attributes:
+        enabled: Feature toggle (push is opt-in)
+        redis: Redis config for cross-server scaling (None = local-only)
+        presence: Presence tracking settings
+        delivery: Delivery guarantee settings
+        polling: Polling fallback settings
+    """
+    
+    enabled: bool = False
+    redis: Optional[RedisConfig] = None
+    presence: PresenceConfig = field(default_factory=PresenceConfig)
+    delivery: DeliveryConfig = field(default_factory=DeliveryConfig)
+    polling: PollingConfig = field(default_factory=PollingConfig)
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary."""
+        return {
+            "enabled": self.enabled,
+            "redis": self.redis.to_dict() if self.redis else None,
+            "presence": self.presence.to_dict(),
+            "delivery": self.delivery.to_dict(),
+            "polling": self.polling.to_dict(),
+        }
+
+
 @dataclass
 class GatewayConfig:
     """Configuration for the gateway server.
@@ -45,6 +197,7 @@ class GatewayConfig:
         host: Host to bind to
         port: Port to listen on
         cors_origins: Allowed CORS origins
+        allowed_origins: Allowed origins for WebSocket connections (CSWSH defense)
         auth_token: Optional authentication token
         max_connections: Maximum concurrent connections
         max_sessions_per_agent: Maximum sessions per agent (0 = unlimited)
@@ -53,11 +206,14 @@ class GatewayConfig:
         reconnect_timeout: Time to wait for reconnection before closing session
         ssl_cert: Path to SSL certificate (for HTTPS/WSS)
         ssl_key: Path to SSL key
+        push: Push notification service configuration
     """
     
     host: str = "127.0.0.1"
     port: int = 8765
+    bind_host: Optional[str] = None
     cors_origins: List[str] = field(default_factory=lambda: [])
+    allowed_origins: List[str] = field(default_factory=lambda: [])
     auth_token: Optional[str] = None
     max_connections: int = 1000
     max_sessions_per_agent: int = 0  # 0 = unlimited
@@ -66,6 +222,12 @@ class GatewayConfig:
     reconnect_timeout: int = 60
     ssl_cert: Optional[str] = None
     ssl_key: Optional[str] = None
+    push: PushConfig = field(default_factory=PushConfig)
+
+    def __post_init__(self) -> None:
+        """Post-initialization to set bind_host from host if not specified."""
+        if self.bind_host is None:
+            self.bind_host = self.host
     
     def to_dict(self) -> Dict[str, Any]:
         """Convert to dictionary (hides sensitive data)."""
@@ -73,6 +235,7 @@ class GatewayConfig:
             "host": self.host,
             "port": self.port,
             "cors_origins": self.cors_origins,
+            "allowed_origins": self.allowed_origins,
             "auth_token": "***" if self.auth_token else None,
             "max_connections": self.max_connections,
             "max_sessions_per_agent": self.max_sessions_per_agent,
@@ -80,6 +243,7 @@ class GatewayConfig:
             "heartbeat_interval": self.heartbeat_interval,
             "reconnect_timeout": self.reconnect_timeout,
             "ssl_enabled": bool(self.ssl_cert and self.ssl_key),
+            "push": self.push.to_dict(),
         }
     
     @property
@@ -205,6 +369,7 @@ class MultiChannelGatewayConfig:
             host=gw_data.get("host", "127.0.0.1"),
             port=gw_data.get("port", 8765),
             cors_origins=gw_data.get("cors_origins", []),
+            allowed_origins=gw_data.get("allowed_origins", []),
             auth_token=gw_data.get("auth_token"),
             max_connections=gw_data.get("max_connections", 1000),
         )

src/praisonai-agents/praisonaiagents/gateway/protocols.py

@@ -20,6 +20,7 @@ from typing import (
     Callable,
     Dict,
     List,
+    Literal,
     Optional,
     Protocol,
     Union,
@@ -27,6 +28,7 @@ from typing import (
 )
 
 if TYPE_CHECKING:
+    from praisonai.gateway.pairing import PairedChannel
     from ..agent import Agent
 
 
@@ -62,6 +64,26 @@ class EventType(str, Enum):
     HEALTH = "health"
     ERROR = "error"
     BROADCAST = "broadcast"
+    
+    # Push channel events
+    CHANNEL_SUBSCRIBE = "channel_subscribe"
+    CHANNEL_UNSUBSCRIBE = "channel_unsubscribe"
+    CHANNEL_MESSAGE = "channel_message"
+    CHANNEL_CREATED = "channel_created"
+    CHANNEL_DELETED = "channel_deleted"
+    
+    # Presence events
+    PRESENCE_JOIN = "presence_join"
+    PRESENCE_LEAVE = "presence_leave"
+    PRESENCE_UPDATE = "presence_update"
+    
+    # Delivery events
+    MESSAGE_NACK = "message_nack"
+    DELIVERY_RETRY = "delivery_retry"
+    
+    # Polling events
+    POLL_REQUEST = "poll_request"
+    POLL_RESPONSE = "poll_response"
 
 
 @dataclass
@@ -407,3 +429,478 @@ class GatewayProtocol(Protocol):
             - clients: Number of connected clients
         """
         ...
+
+
+# ---------------------------------------------------------------------------
+# Push notification dataclasses
+# ---------------------------------------------------------------------------
+
+@dataclass
+class ChannelInfo:
+    """Metadata about a push channel/topic.
+    
+    Attributes:
+        name: Channel name (unique identifier)
+        created_at: Channel creation timestamp
+        metadata: Arbitrary channel metadata
+        subscriber_count: Current number of subscribers
+    """
+    
+    name: str
+    created_at: float = field(default_factory=time.time)
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    subscriber_count: int = 0
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary for serialization."""
+        return {
+            "name": self.name,
+            "created_at": self.created_at,
+            "metadata": self.metadata,
+            "subscriber_count": self.subscriber_count,
+        }
+    
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> "ChannelInfo":
+        """Create from dictionary."""
+        return cls(
+            name=data.get("name", ""),
+            created_at=data.get("created_at", time.time()),
+            metadata=data.get("metadata", {}),
+            subscriber_count=data.get("subscriber_count", 0),
+        )
+
+
+@dataclass
+class PresenceInfo:
+    """Presence information for a connected client.
+    
+    Attributes:
+        client_id: Client identifier
+        status: Presence status ("online", "idle", "offline")
+        last_seen: Last heartbeat timestamp
+        metadata: Client-provided metadata (e.g., display name)
+        channels: Channels this client is subscribed to
+    """
+    
+    client_id: str
+    status: str = "online"
+    last_seen: float = field(default_factory=time.time)
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    channels: List[str] = field(default_factory=list)
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary for serialization."""
+        return {
+            "client_id": self.client_id,
+            "status": self.status,
+            "last_seen": self.last_seen,
+            "metadata": self.metadata,
+            "channels": self.channels,
+        }
+    
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> "PresenceInfo":
+        """Create from dictionary."""
+        return cls(
+            client_id=data.get("client_id", ""),
+            status=data.get("status", "online"),
+            last_seen=data.get("last_seen", time.time()),
+            metadata=data.get("metadata", {}),
+            channels=data.get("channels", []),
+        )
+
+
+# ---------------------------------------------------------------------------
+# Push notification protocols
+# ---------------------------------------------------------------------------
+
+@runtime_checkable
+class PushChannelProtocol(Protocol):
+    """Protocol for channel/topic-based push messaging.
+    
+    Channels group clients by interest. Messages published to a channel
+    are delivered to all subscribed clients.
+    """
+    
+    def add_channel(
+        self, channel_name: str, metadata: Optional[Dict[str, Any]] = None,
+    ) -> bool:
+        """Create a named channel.
+        
+        Args:
+            channel_name: Unique channel name
+            metadata: Optional channel metadata
+            
+        Returns:
+            True if created, False if already exists
+        """
+        ...
+    
+    def remove_channel(self, channel_name: str) -> bool:
+        """Delete a channel and unsubscribe all clients.
+        
+        Returns:
+            True if removed, False if not found
+        """
+        ...
+    
+    def get_channel(self, channel_name: str) -> Optional[ChannelInfo]:
+        """Get channel metadata."""
+        ...
+    
+    def list_channels(self) -> List[str]:
+        """List all active channel names."""
+        ...
+    
+    def subscribe_client(self, client_id: str, channel_name: str) -> bool:
+        """Subscribe a client to a channel.
+        
+        Returns:
+            True if subscribed, False if already subscribed or channel missing
+        """
+        ...
+    
+    def unsubscribe_client(self, client_id: str, channel_name: str) -> bool:
+        """Unsubscribe a client from a channel.
+        
+        Returns:
+            True if unsubscribed, False if not subscribed
+        """
+        ...
+    
+    def get_subscribers(self, channel_name: str) -> List[str]:
+        """List client IDs subscribed to a channel."""
+        ...
+    
+    def get_client_channels(self, client_id: str) -> List[str]:
+        """List channels a client is subscribed to."""
+        ...
+    
+    async def publish_to_channel(
+        self,
+        channel_name: str,
+        event: GatewayEvent,
+        exclude: Optional[List[str]] = None,
+    ) -> int:
+        """Publish an event to all subscribers of a channel.
+        
+        Args:
+            channel_name: Target channel
+            event: The event to deliver
+            exclude: Optional client IDs to skip
+            
+        Returns:
+            Number of clients the event was sent to
+        """
+        ...
+
+
+@runtime_checkable
+class PresenceProtocol(Protocol):
+    """Protocol for tracking client presence (online/idle/offline)."""
+    
+    async def track_presence(
+        self,
+        client_id: str,
+        status: str = "online",
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """Set or update a client's presence status.
+        
+        Args:
+            client_id: Client identifier
+            status: Presence status ("online", "idle", "offline")
+            metadata: Optional client metadata
+        """
+        ...
+    
+    async def remove_presence(self, client_id: str) -> None:
+        """Remove a client's presence (on disconnect)."""
+        ...
+    
+    def get_presence(self, client_id: str) -> Optional[PresenceInfo]:
+        """Get a single client's presence info."""
+        ...
+    
+    def get_all_presence(
+        self, channel_name: Optional[str] = None,
+    ) -> List[PresenceInfo]:
+        """Get presence info, optionally filtered by channel.
+        
+        Args:
+            channel_name: If provided, only return presence for channel members
+        """
+        ...
+    
+    def get_online_count(self, channel_name: Optional[str] = None) -> int:
+        """Count online clients, optionally filtered by channel."""
+        ...
+
+
+@runtime_checkable
+class DeliveryGuaranteeProtocol(Protocol):
+    """Protocol for at-least-once message delivery.
+    
+    Messages are stored, tracked, and retried until acknowledged.
+    """
+    
+    async def store_message(self, event: GatewayEvent) -> str:
+        """Persist a message to the store.
+        
+        Returns:
+            The event_id of the stored message
+        """
+        ...
+    
+    async def acknowledge(self, client_id: str, event_id: str) -> bool:
+        """Mark a message as acknowledged by a client.
+        
+        Returns:
+            True if found and acknowledged, False if not found
+        """
+        ...
+    
+    async def nack(self, client_id: str, event_id: str) -> None:
+        """Negative acknowledge - request redelivery."""
+        ...
+    
+    async def get_unacknowledged(
+        self, client_id: str, limit: int = 100,
+    ) -> List[GatewayEvent]:
+        """Get pending unacknowledged messages for a client."""
+        ...
+    
+    async def retry_unacknowledged(self, client_id: str) -> int:
+        """Redeliver all unacknowledged messages to a client.
+        
+        Returns:
+            Number of messages redelivered
+        """
+        ...
+    
+    async def purge_acknowledged(self, max_age_seconds: int = 86400) -> int:
+        """Remove old acknowledged messages from the store.
+        
+        Returns:
+            Number of messages purged
+        """
+        ...
+
+
+# ---------------------------------------------------------------------------
+# Auth Mode protocols and helpers (bind-aware authentication posture)
+# ---------------------------------------------------------------------------
+
+AuthMode = Literal["local", "token", "password", "trusted-proxy"]
+"""Authentication mode for gateway/UI components.
+
+- "local": Permissive mode for loopback interfaces (127.0.0.1, localhost, ::1)
+- "token": Token-based authentication required (default for external interfaces)  
+- "password": Username/password authentication
+- "trusted-proxy": Authentication handled by upstream proxy
+"""
+
+
+def is_loopback(host: str) -> bool:
+    """Check if a host is a loopback interface.
+    
+    Args:
+        host: Host/IP address to check
+        
+    Returns:
+        True if the host is a loopback address
+        
+    Examples:
+        >>> is_loopback("127.0.0.1")
+        True
+        >>> is_loopback("localhost") 
+        True
+        >>> is_loopback("::1")
+        True
+        >>> is_loopback("0.0.0.0")
+        False
+        >>> is_loopback("192.168.1.1")
+        False
+    """
+    import ipaddress
+    
+    # Handle localhost specially
+    if host in ("localhost", "0:0:0:0:0:0:0:1"):
+        return True
+    
+    try:
+        ip = ipaddress.ip_address(host)
+        return ip.is_loopback
+    except ValueError:
+        # Not a valid IP address (e.g., domain name)
+        return False
+
+
+def resolve_auth_mode(bind_host: str, configured: Optional[AuthMode] = None) -> AuthMode:
+    """Resolve authentication mode based on bind host and explicit configuration.
+    
+    Args:
+        bind_host: Host/IP that the service is bound to
+        configured: Explicitly configured auth mode (takes precedence)
+        
+    Returns:
+        The resolved authentication mode
+        
+    Examples:
+        >>> resolve_auth_mode("127.0.0.1")
+        'local'
+        >>> resolve_auth_mode("0.0.0.0")  
+        'token'
+        >>> resolve_auth_mode("127.0.0.1", "token")
+        'token'
+    """
+    if configured is not None:
+        return configured
+    
+    return "local" if is_loopback(bind_host) else "token"
+
+
+# ---------------------------------------------------------------------------
+# Auth, Pairing, and Session Binding Protocols (Issue #1588 Gap 3)
+# ---------------------------------------------------------------------------
+
+@runtime_checkable
+class AuthProtocol(Protocol):
+    """Protocol for authentication implementations.
+    
+    Enables structural typing for different authentication strategies
+    (token-based, local loopback, trusted proxy, etc.).
+    """
+    
+    def check(self, request: Any) -> Dict[str, Any]:
+        """Check authentication for a request.
+        
+        Args:
+            request: The request object to authenticate
+            
+        Returns:
+            Authentication decision with metadata:
+            - success: bool - whether authentication succeeded
+            - user_id: Optional[str] - authenticated user ID
+            - role: Optional[str] - user role/permissions
+            - metadata: Dict[str, Any] - additional auth context
+        """
+        ...
+
+
+@runtime_checkable  
+class PairingProtocol(Protocol):
+    """Protocol for channel pairing implementations.
+    
+    Manages the authorization of external channels (Telegram, Slack, UI)
+    to communicate with the gateway through signed codes.
+    """
+    
+    def generate_code(
+        self, 
+        channel_type: str = "unknown", 
+        channel_id: Optional[str] = None
+    ) -> str:
+        """Generate a new pairing code for a channel.
+        
+        Args:
+            channel_type: Type of channel (e.g., "telegram", "slack", "ui")
+            channel_id: Optional channel identifier
+            
+        Returns:
+            The generated pairing code
+        """
+        ...
+    
+    def approve(
+        self, 
+        channel_type: str,
+        code: str,
+        user_id: str = "",
+        user_name: str = ""
+    ) -> bool:
+        """Approve a pairing code, authorizing the channel.
+        
+        Args:
+            channel_type: Type of channel
+            code: The pairing code to approve
+            user_id: User identifier (optional, defaults to empty string)
+            user_name: Human-readable username (optional, defaults to empty string)
+            
+        Returns:
+            True if approval successful, False if code invalid/expired
+        """
+        ...
+    
+    def is_paired(self, channel_id: str, channel_type: str) -> bool:
+        """Check if a channel is authorized.
+        
+        Args:
+            channel_id: Channel identifier
+            channel_type: Type of channel
+            
+        Returns:
+            True if channel is paired/authorized
+        """
+        ...
+    
+    def list_paired(self) -> List["PairedChannel"]:
+        """List all authorized channels.
+        
+        Returns:
+            List of paired channel information
+        """
+        ...
+    
+    def revoke(self, channel_id: str, channel_type: str) -> bool:
+        """Revoke authorization for a channel.
+        
+        Args:
+            channel_id: Channel identifier
+            channel_type: Type of channel
+            
+        Returns:
+            True if revocation successful, False if not found
+        """
+        ...
+    
+    def list_pending(self, channel_type: Optional[str] = None) -> List[Dict[str, any]]:
+        """List pending pairing requests.
+        
+        Args:
+            channel_type: Optional filter by channel type
+            
+        Returns:
+            List of pending requests with channel, code, user info, and age
+        """
+        ...
+
+
+@runtime_checkable
+class SessionBindingProtocol(Protocol):
+    """Protocol for session binding implementations.
+    
+    Manages the association between sessions and authenticated principals
+    (users, agents, etc.) for state tracking and authorization.
+    """
+    
+    def bind(self, session_id: str, principal: Dict[str, Any]) -> None:
+        """Bind a session to an authenticated principal.
+        
+        Args:
+            session_id: Unique session identifier
+            principal: Principal information (user_id, roles, metadata, etc.)
+        """
+        ...
+    
+    def lookup(self, session_id: str) -> Optional[Dict[str, Any]]:
+        """Look up the principal bound to a session.
+        
+        Args:
+            session_id: Session identifier to look up
+            
+        Returns:
+            Principal information if found, None otherwise
+        """
+        ...

src/praisonai-agents/praisonaiagents/hooks/runner.py

@@ -146,12 +146,12 @@ class HookRunner:
             loop = None
 
         if loop is not None and loop.is_running():
-            # Already inside a running async loop — cannot use run_until_complete.
-            # Schedule as fire-and-forget task to avoid leaking unawaited coroutines.
-            # Callers that need results in async context should use
-            # ``await runner.execute(...)`` directly.
-            loop.create_task(self.execute(event, input_data, target))
-            return []
+            # Cannot execute sync in running event loop - would block the loop
+            # Guide users to use the async API instead
+            raise RuntimeError(
+                "execute_sync() cannot be called from within a running event loop. "
+                "Use 'await runner.execute(event, input_data, target)' instead in async contexts."
+            )
 
         # No running loop — safe to create one
         return asyncio.run(self.execute(event, input_data, target))

src/praisonai-agents/praisonaiagents/knowledge/adapters/mem0_adapter.py

@@ -71,7 +71,8 @@ class Mem0Adapter:
             from mem0 import Memory
         except ImportError:
             raise ImportError(
-                "mem0 is not installed. Install with: pip install mem0ai"
+                "mem0ai is not installed. "
+                "Run: pip install 'praisonaiagents[memory]'"
             )
         
         if self._config:

src/praisonai-agents/praisonaiagents/llm/__init__.py

@@ -9,8 +9,11 @@ import os
 # Ensure litellm telemetry is disabled before any imports
 os.environ["LITELLM_TELEMETRY"] = "False"
 
+import threading
+
 # Module-level cache for lazy-loaded classes
 _lazy_cache = {}
+_cache_lock = threading.Lock()
 
 
 def __getattr__(name):
@@ -18,139 +21,165 @@ def __getattr__(name):
     if name in _lazy_cache:
         return _lazy_cache[name]
     
-    if name == "LLM":
-        from .llm import LLM
-        _lazy_cache[name] = LLM
-        return LLM
-    elif name == "LLMContextLengthExceededException":
-        from .llm import LLMContextLengthExceededException
-        _lazy_cache[name] = LLMContextLengthExceededException
-        return LLMContextLengthExceededException
-    elif name == "OpenAIClient":
-        from .openai_client import OpenAIClient
-        _lazy_cache[name] = OpenAIClient
-        return OpenAIClient
-    elif name == "get_openai_client":
-        from .openai_client import get_openai_client
-        _lazy_cache[name] = get_openai_client
-        return get_openai_client
-    elif name == "ChatCompletionMessage":
-        from .openai_client import ChatCompletionMessage
-        _lazy_cache[name] = ChatCompletionMessage
-        return ChatCompletionMessage
-    elif name == "Choice":
-        from .openai_client import Choice
-        _lazy_cache[name] = Choice
-        return Choice
-    elif name == "CompletionTokensDetails":
-        from .openai_client import CompletionTokensDetails
-        _lazy_cache[name] = CompletionTokensDetails
-        return CompletionTokensDetails
-    elif name == "PromptTokensDetails":
-        from .openai_client import PromptTokensDetails
-        _lazy_cache[name] = PromptTokensDetails
-        return PromptTokensDetails
-    elif name == "CompletionUsage":
-        from .openai_client import CompletionUsage
-        _lazy_cache[name] = CompletionUsage
-        return CompletionUsage
-    elif name == "ChatCompletion":
-        from .openai_client import ChatCompletion
-        _lazy_cache[name] = ChatCompletion
-        return ChatCompletion
-    elif name == "ToolCall":
-        from .openai_client import ToolCall
-        _lazy_cache[name] = ToolCall
-        return ToolCall
-    elif name == "process_stream_chunks":
-        from .openai_client import process_stream_chunks
-        _lazy_cache[name] = process_stream_chunks
-        return process_stream_chunks
-    elif name == "supports_structured_outputs":
-        from .model_capabilities import supports_structured_outputs
-        _lazy_cache[name] = supports_structured_outputs
-        return supports_structured_outputs
-    elif name == "supports_streaming_with_tools":
-        from .model_capabilities import supports_streaming_with_tools
-        _lazy_cache[name] = supports_streaming_with_tools
-        return supports_streaming_with_tools
-    elif name == "ModelRouter":
-        from .model_router import ModelRouter
-        _lazy_cache[name] = ModelRouter
-        return ModelRouter
-    elif name == "ModelProfile":
-        from .model_router import ModelProfile
-        _lazy_cache[name] = ModelProfile
-        return ModelProfile
-    elif name == "TaskComplexity":
-        from .model_router import TaskComplexity
-        _lazy_cache[name] = TaskComplexity
-        return TaskComplexity
-    elif name == "create_routing_agent":
-        from .model_router import create_routing_agent
-        _lazy_cache[name] = create_routing_agent
-        return create_routing_agent
-    elif name == "RateLimiter":
-        from .rate_limiter import RateLimiter
-        _lazy_cache[name] = RateLimiter
-        return RateLimiter
-    elif name == "TokenUsage":
-        from .llm import TokenUsage
-        _lazy_cache[name] = TokenUsage
-        return TokenUsage
-    elif name == "LLMProviderProtocol":
-        from .protocols import LLMProviderProtocol
-        _lazy_cache[name] = LLMProviderProtocol
-        return LLMProviderProtocol
-    elif name == "ModelCapabilitiesProtocol":
-        from .protocols import ModelCapabilitiesProtocol
-        _lazy_cache[name] = ModelCapabilitiesProtocol
-        return ModelCapabilitiesProtocol
-    elif name == "LLMRateLimiterProtocol":
-        from .protocols import LLMRateLimiterProtocol
-        _lazy_cache[name] = LLMRateLimiterProtocol
-        return LLMRateLimiterProtocol
-    elif name == "LLMFailoverProtocol":
-        from .protocols import LLMFailoverProtocol
-        _lazy_cache[name] = LLMFailoverProtocol
-        return LLMFailoverProtocol
-    elif name == "LLMProviderError":
-        from .protocols import LLMProviderError
-        _lazy_cache[name] = LLMProviderError
-        return LLMProviderError
-    elif name == "RateLimitError":
-        from .protocols import RateLimitError
-        _lazy_cache[name] = RateLimitError
-        return RateLimitError
-    elif name == "ModelNotAvailableError":
-        from .protocols import ModelNotAvailableError
-        _lazy_cache[name] = ModelNotAvailableError
-        return ModelNotAvailableError
-    elif name == "ContextLengthExceededError":
-        from .protocols import ContextLengthExceededError
-        _lazy_cache[name] = ContextLengthExceededError
-        return ContextLengthExceededError
-    elif name == "UnifiedLLMProtocol":
-        from .protocols import UnifiedLLMProtocol
-        _lazy_cache[name] = UnifiedLLMProtocol
-        return UnifiedLLMProtocol
-    elif name == "LiteLLMAdapter":
-        from .unified_adapters import LiteLLMAdapter
-        _lazy_cache[name] = LiteLLMAdapter
-        return LiteLLMAdapter
-    elif name == "OpenAIAdapter":
-        from .unified_adapters import OpenAIAdapter
-        _lazy_cache[name] = OpenAIAdapter
-        return OpenAIAdapter
-    elif name == "UnifiedLLMDispatcher":
-        from .unified_adapters import UnifiedLLMDispatcher
-        _lazy_cache[name] = UnifiedLLMDispatcher
-        return UnifiedLLMDispatcher
-    elif name == "create_llm_dispatcher":
-        from .unified_adapters import create_llm_dispatcher
-        _lazy_cache[name] = create_llm_dispatcher
-        return create_llm_dispatcher
-    
+    with _cache_lock:
+        # Double-check after acquiring lock
+        if name in _lazy_cache:
+            return _lazy_cache[name]
+
+        if name == "LLM":
+            from .llm import LLM
+            _lazy_cache[name] = LLM
+            return LLM
+        elif name == "LLMContextLengthExceededException":
+            from .llm import LLMContextLengthExceededException
+            _lazy_cache[name] = LLMContextLengthExceededException
+            return LLMContextLengthExceededException
+
+        elif name == "OpenAIClient":
+            from .openai_client import OpenAIClient
+            _lazy_cache[name] = OpenAIClient
+            return OpenAIClient
+        elif name == "get_openai_client":
+            from .openai_client import get_openai_client
+            _lazy_cache[name] = get_openai_client
+            return get_openai_client
+        elif name == "ChatCompletionMessage":
+            from .openai_client import ChatCompletionMessage
+            _lazy_cache[name] = ChatCompletionMessage
+            return ChatCompletionMessage
+        elif name == "Choice":
+            from .openai_client import Choice
+            _lazy_cache[name] = Choice
+            return Choice
+        elif name == "CompletionTokensDetails":
+            from .openai_client import CompletionTokensDetails
+            _lazy_cache[name] = CompletionTokensDetails
+            return CompletionTokensDetails
+        elif name == "PromptTokensDetails":
+            from .openai_client import PromptTokensDetails
+            _lazy_cache[name] = PromptTokensDetails
+            return PromptTokensDetails
+        elif name == "CompletionUsage":
+            from .openai_client import CompletionUsage
+            _lazy_cache[name] = CompletionUsage
+            return CompletionUsage
+        elif name == "ChatCompletion":
+            from .openai_client import ChatCompletion
+            _lazy_cache[name] = ChatCompletion
+            return ChatCompletion
+        elif name == "ToolCall":
+            from .openai_client import ToolCall
+            _lazy_cache[name] = ToolCall
+            return ToolCall
+        elif name == "process_stream_chunks":
+            from .openai_client import process_stream_chunks
+            _lazy_cache[name] = process_stream_chunks
+            return process_stream_chunks
+        elif name == "supports_structured_outputs":
+            from .model_capabilities import supports_structured_outputs
+            _lazy_cache[name] = supports_structured_outputs
+            return supports_structured_outputs
+        elif name == "supports_streaming_with_tools":
+            from .model_capabilities import supports_streaming_with_tools
+            _lazy_cache[name] = supports_streaming_with_tools
+            return supports_streaming_with_tools
+        elif name == "ModelRouter":
+            from .model_router import ModelRouter
+            _lazy_cache[name] = ModelRouter
+            return ModelRouter
+        elif name == "ModelProfile":
+            from .model_router import ModelProfile
+            _lazy_cache[name] = ModelProfile
+            return ModelProfile
+        elif name == "TaskComplexity":
+            from .model_router import TaskComplexity
+            _lazy_cache[name] = TaskComplexity
+            return TaskComplexity
+        elif name == "create_routing_agent":
+            from .model_router import create_routing_agent
+            _lazy_cache[name] = create_routing_agent
+            return create_routing_agent
+        elif name == "RateLimiter":
+            from .rate_limiter import RateLimiter
+            _lazy_cache[name] = RateLimiter
+            return RateLimiter
+        elif name == "TokenUsage":
+            from .llm import TokenUsage
+            _lazy_cache[name] = TokenUsage
+            return TokenUsage
+        elif name == "LLMProviderProtocol":
+            from .protocols import LLMProviderProtocol
+            _lazy_cache[name] = LLMProviderProtocol
+            return LLMProviderProtocol
+        elif name == "ModelCapabilitiesProtocol":
+            from .protocols import ModelCapabilitiesProtocol
+            _lazy_cache[name] = ModelCapabilitiesProtocol
+            return ModelCapabilitiesProtocol
+        elif name == "LLMRateLimiterProtocol":
+            from .protocols import LLMRateLimiterProtocol
+            _lazy_cache[name] = LLMRateLimiterProtocol
+            return LLMRateLimiterProtocol
+        elif name == "LLMFailoverProtocol":
+            from .protocols import LLMFailoverProtocol
+            _lazy_cache[name] = LLMFailoverProtocol
+            return LLMFailoverProtocol
+        elif name == "LLMProviderError":
+            from .protocols import LLMProviderError
+            _lazy_cache[name] = LLMProviderError
+            return LLMProviderError
+        elif name == "RateLimitError":
+            from .protocols import RateLimitError
+            _lazy_cache[name] = RateLimitError
+            return RateLimitError
+        elif name == "ModelNotAvailableError":
+            from .protocols import ModelNotAvailableError
+            _lazy_cache[name] = ModelNotAvailableError
+            return ModelNotAvailableError
+        elif name == "ContextLengthExceededError":
+            from .protocols import ContextLengthExceededError
+            _lazy_cache[name] = ContextLengthExceededError
+            return ContextLengthExceededError
+        elif name == "UnifiedLLMProtocol":
+            from .protocols import UnifiedLLMProtocol
+            _lazy_cache[name] = UnifiedLLMProtocol
+            return UnifiedLLMProtocol
+        elif name == "LiteLLMAdapter":
+            from .unified_adapters import LiteLLMAdapter
+            _lazy_cache[name] = LiteLLMAdapter
+            return LiteLLMAdapter
+        elif name == "OpenAIAdapter":
+            from .unified_adapters import OpenAIAdapter
+            _lazy_cache[name] = OpenAIAdapter
+            return OpenAIAdapter
+        elif name == "UnifiedLLMDispatcher":
+            from .unified_adapters import UnifiedLLMDispatcher
+            _lazy_cache[name] = UnifiedLLMDispatcher
+            return UnifiedLLMDispatcher
+        elif name == "create_llm_dispatcher":
+            from .unified_adapters import create_llm_dispatcher
+            _lazy_cache[name] = create_llm_dispatcher
+            return create_llm_dispatcher
+        elif name == "sanitize_messages":
+            from .sanitize import sanitize_messages
+            _lazy_cache[name] = sanitize_messages
+            return sanitize_messages
+        elif name == "strip_surrogates":
+            from .sanitize import strip_surrogates
+            _lazy_cache[name] = strip_surrogates
+            return strip_surrogates
+        elif name == "sanitize_text":
+            from .sanitize import sanitize_text
+            _lazy_cache[name] = sanitize_text
+            return sanitize_text
+        elif name == "ErrorCategory":
+            from .error_classifier import ErrorCategory
+            _lazy_cache[name] = ErrorCategory
+            return ErrorCategory
+        elif name == "classify_error":
+            from .error_classifier import classify_error
+            _lazy_cache[name] = classify_error
+            return classify_error
+
     raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
 
 
@@ -190,5 +219,12 @@ __all__ = [
     "LLMProviderError",
     "RateLimitError",
     "ModelNotAvailableError",
-    "ContextLengthExceededError"
+    "ContextLengthExceededError",
+    # Sanitization (G6)
+    "sanitize_messages",
+    "strip_surrogates", 
+    "sanitize_text",
+    # Error Classification (G5)
+    "ErrorCategory",
+    "classify_error"
 ]

src/praisonai-agents/praisonaiagents/llm/error_classifier.py

@@ -0,0 +1,252 @@
+"""
+Error Classification - Multi-category error classifier for intelligent retry logic.
+
+Extends the existing single-category rate limit detection with comprehensive
+error classification for better handling of different failure modes.
+"""
+
+import re
+import random
+from enum import Enum
+from typing import Dict, Tuple, List, Optional
+
+__all__ = [
+    "ErrorCategory",
+    "classify_error", 
+    "should_retry", 
+    "get_retry_delay",
+    "extract_retry_after",
+    "get_error_context",
+]
+
+
+class ErrorCategory(str, Enum):
+    """Categories of LLM provider errors for intelligent handling."""
+    RATE_LIMIT = "rate_limit"           # Too many requests, temporary
+    CONTEXT_LIMIT = "context_limit"     # Input too long, need compression
+    AUTH = "auth"                       # Authentication/authorization failure  
+    INVALID_REQUEST = "invalid_request" # Malformed request, permanent
+    TRANSIENT = "transient"            # Network/server issues, temporary
+    PERMANENT = "permanent"            # Unrecoverable error
+
+
+# Error patterns for classification (case-insensitive)
+_ERROR_PATTERNS: Dict[ErrorCategory, List[str]] = {
+    ErrorCategory.RATE_LIMIT: [
+        r"rate.?limit",
+        r"429",
+        r"too.?many.?request",
+        r"resource.?exhausted", 
+        r"quota.?exceeded",
+        r"tokens.?per.?minute",
+        r"requests.?per.?minute",
+        r"concurrent.?requests",
+    ],
+    
+    ErrorCategory.CONTEXT_LIMIT: [
+        r"context.?length",
+        r"maximum.?context", 
+        r"token.?limit",
+        r"input.?too.?long",
+        r"sequence.?too.?long",
+        r"context.?window",
+        r"413",  # Request Entity Too Large
+        r"payload.?too.?large",
+    ],
+    
+    ErrorCategory.AUTH: [
+        r"authenticat",
+        r"authoriz",
+        r"401", 
+        r"403",
+        r"invalid.?api.?key",
+        r"api.?key.*invalid",
+        r"permission.?denied", 
+        r"access.?denied",
+        r"forbidden",
+        r"unauthorized",
+        r"invalid.?token",
+        r"expired.*token",
+    ],
+    
+    ErrorCategory.INVALID_REQUEST: [
+        r"invalid.*request",
+        r"bad.?request",
+        r"400",
+        r"malformed",
+        r"invalid.*parameter",
+        r"unsupported.*model",
+        r"model.*not.*found",
+        r"validation.*error",
+        r"schema.*error",
+    ],
+    
+    ErrorCategory.TRANSIENT: [
+        r"timeout",
+        r"timed.?out",
+        r"500", r"502", r"503", r"504",
+        r"internal.?server.?error",
+        r"bad.?gateway", 
+        r"service.?unavailable",
+        r"gateway.?timeout",
+        r"connection.*error",
+        r"network.*error",
+        r"temporary.*unavailable",
+        r"server.?overload",
+        r"retry.?after",
+    ],
+}
+
+
+def classify_error(error: Exception) -> ErrorCategory:
+    """Classify an error into a category for intelligent handling.
+    
+    Args:
+        error: Exception to classify
+        
+    Returns:
+        ErrorCategory indicating how the error should be handled
+        
+    Examples:
+        >>> classify_error(Exception("Rate limit exceeded"))
+        ErrorCategory.RATE_LIMIT
+        >>> classify_error(Exception("Context length 8192 exceeded"))  
+        ErrorCategory.CONTEXT_LIMIT
+        >>> classify_error(Exception("Invalid API key"))
+        ErrorCategory.AUTH
+    """
+    error_text = f"{type(error).__name__} {error}".lower()
+    
+    # Check each category's patterns
+    for category, patterns in _ERROR_PATTERNS.items():
+        for pattern in patterns:
+            if re.search(pattern, error_text, re.IGNORECASE):
+                return category
+    
+    # Default to permanent for unknown errors
+    return ErrorCategory.PERMANENT
+
+
+def should_retry(category: ErrorCategory) -> bool:
+    """Determine if an error category should be retried.
+    
+    Args:
+        category: Error category from classify_error()
+        
+    Returns:
+        True if the error type should be retried
+    """
+    return category in {
+        ErrorCategory.RATE_LIMIT,
+        ErrorCategory.CONTEXT_LIMIT,  # Could retry with compression
+        ErrorCategory.TRANSIENT,
+    }
+
+
+def get_retry_delay(category: ErrorCategory, attempt: int = 1, base_delay: float = 1.0) -> float:
+    """Get the appropriate delay before retrying based on error category.
+    
+    Uses full jitter to prevent thundering herd problems in multi-agent setups
+    where multiple agents hit rate limits simultaneously.
+    
+    Args:
+        category: Error category
+        attempt: Current attempt number (1-based)
+        base_delay: Base delay in seconds
+        
+    Returns:
+        Delay in seconds, or 0 if should not retry
+        
+    Examples:
+        >>> # With jitter, these will return random values in range:
+        >>> get_retry_delay(ErrorCategory.RATE_LIMIT, attempt=1)  # 0.0 to 3.0
+        >>> get_retry_delay(ErrorCategory.TRANSIENT, attempt=3)   # 0.0 to 8.0
+        >>> get_retry_delay(ErrorCategory.AUTH, attempt=1)        # Always 0
+    """
+    attempt = max(1, attempt)
+
+    if not should_retry(category):
+        return 0
+    
+    if category == ErrorCategory.RATE_LIMIT:
+        # Exponential backoff with equal jitter for rate limits (minimum floor to prevent instant retries)
+        max_delay = min(base_delay * (3 ** attempt), 60.0)
+        return base_delay + random.uniform(0, max_delay - base_delay)
+    
+    elif category == ErrorCategory.CONTEXT_LIMIT:
+        # Short delay for context limits (no jitter needed - not a contention issue)
+        return base_delay * 0.5
+    
+    elif category == ErrorCategory.TRANSIENT:
+        # Exponential backoff with equal jitter for transient errors (minimum floor to prevent instant retries)
+        max_delay = min(base_delay * (2 ** attempt), 30.0)
+        return base_delay + random.uniform(0, max_delay - base_delay)
+    
+    return 0
+
+
+def extract_retry_after(error: Exception) -> Optional[float]:
+    """Extract Retry-After header value from rate limit errors.
+    
+    Args:
+        error: Exception potentially containing Retry-After info
+        
+    Returns:
+        Delay in seconds if found, None otherwise
+    """
+    error_str = str(error)
+    
+    # Look for common Retry-After patterns
+    patterns = [
+        r"retry.?after[:\s]+(\d+)",
+        r"retry[:\s]+(\d+)",
+        r"wait[:\s]+(\d+)",
+        r"(\d+).*second",
+    ]
+    
+    for pattern in patterns:
+        match = re.search(pattern, error_str, re.IGNORECASE)
+        if match:
+            try:
+                delay = float(match.group(1))
+                return min(delay, 300.0)  # Cap at 5 minutes
+            except (ValueError, IndexError):
+                continue
+    
+    return None
+
+
+def get_error_context(error: Exception) -> Dict[str, str]:
+    """Extract structured context from an error for logging/debugging.
+    
+    Args:
+        error: Exception to analyze
+        
+    Returns:
+        Dictionary with error context information
+    """
+    category = classify_error(error)
+    
+    context = {
+        "error_type": type(error).__name__,
+        "category": category.value,
+        "should_retry": str(should_retry(category)),
+        "message": str(error)[:500],  # Truncate long messages
+    }
+    
+    # Add category-specific context
+    if category == ErrorCategory.RATE_LIMIT:
+        retry_after = extract_retry_after(error)
+        if retry_after:
+            context["retry_after"] = str(retry_after)
+    
+    elif category == ErrorCategory.CONTEXT_LIMIT:
+        context["suggestion"] = "Try reducing input size or enabling compression"
+        
+    elif category == ErrorCategory.AUTH:
+        context["suggestion"] = "Check API key configuration"
+    
+    elif category == ErrorCategory.INVALID_REQUEST:
+        context["suggestion"] = "Review request parameters"
+        
+    return context

src/praisonai-agents/praisonaiagents/llm/llm.py

@@ -6,11 +6,17 @@ import re
 import inspect
 import asyncio
 from dataclasses import dataclass
-from typing import Any, Dict, List, Optional, Union, Literal, Callable, TYPE_CHECKING
+from typing import Any, Dict, List, Optional, Union, Literal, Callable, TYPE_CHECKING, Protocol
 
 if TYPE_CHECKING:
     from rich.console import Console
     from rich.live import Live
+    
+class FailoverManagerProtocol(Protocol):
+    """Protocol for failover manager implementations."""
+    def get_next_profile(self) -> Optional["AuthProfile"]: ...
+    def mark_failure(self, profile: "AuthProfile", error: str, is_rate_limit: bool = False) -> None: ...
+    def mark_success(self, profile: "AuthProfile") -> None: ...
 from pydantic import BaseModel
 import time
 import json
@@ -354,6 +360,8 @@ Respond with ONLY a valid JSON tool call in this format:
         web_fetch: Optional[Union[bool, Dict[str, Any]]] = None,
         prompt_caching: Optional[bool] = None,
         claude_memory: Optional[Union[bool, Any]] = None,
+        failover_manager: Optional[FailoverManagerProtocol] = None,
+        auth: Optional[str] = None,           # NEW: "claude-code", "codex", "gemini-cli", "qwen-cli"
         **extra_settings
     ):
         # Configure logging only once at the class level
@@ -429,6 +437,18 @@ Respond with ONLY a valid JSON tool call in this format:
         self._rate_limiter = extra_settings.get('rate_limiter', None)
         self._max_retries = extra_settings.get('max_retries', 3)
         self._retry_delay = extra_settings.get('retry_delay', 60)  # Default 60 seconds
+        
+        # Subscription auth
+        self._auth_provider_id = auth
+        self._cached_subscription_creds = None  # SubscriptionCredentials | None
+        
+        # Failover management
+        self._failover_manager = failover_manager
+        self._current_profile = None  # Track current auth profile for failover
+        if self._failover_manager:
+            self._current_profile = self._failover_manager.get_next_profile()
+            if self._current_profile:
+                self._switch_to_profile(self._current_profile)
 
         # Cache for formatted tools and messages
         self._formatted_tools_cache = {}
@@ -650,8 +670,87 @@ Respond with ONLY a valid JSON tool call in this format:
 
         return any(indicator in error_str or indicator in error_type for indicator in indicators)
 
+    def _classify_error_and_should_retry(self, error: Exception, attempt: int = 1) -> tuple[str, bool, float]:
+        """Classify error and determine retry strategy using G5 error classifier.
+        
+        Args:
+            error: Exception to classify
+            attempt: Current attempt number (1-based)
+            
+        Returns:
+            Tuple of (category, should_retry, retry_delay)
+        """
+        try:
+            from .error_classifier import classify_error, should_retry, get_retry_delay, extract_retry_after
+            
+            category = classify_error(error)
+            can_retry = should_retry(category)
+            
+            if not can_retry:
+                return category.value, False, 0.0
+            
+            # For rate limits, try to extract specific retry-after first
+            if category.value == "rate_limit":
+                retry_after = extract_retry_after(error)
+                if retry_after:
+                    return category.value, True, retry_after
+            
+            # Use category-specific delay calculation with proper attempt
+            delay = get_retry_delay(category, attempt=attempt, base_delay=self._retry_delay)
+            return category.value, True, delay
+            
+        except ImportError:
+            # Fallback to legacy rate limit detection
+            is_rate_limit = self._is_rate_limit_error(error)
+            delay = self._parse_retry_delay(str(error)) if is_rate_limit else 0.0
+            return "rate_limit" if is_rate_limit else "unknown", is_rate_limit, delay
+
+    def _switch_to_profile(self, profile: "AuthProfile") -> None:
+        """Switch to a new auth profile for failover.
+        
+        Args:
+            profile: AuthProfile to switch to
+        """
+        if profile.api_key:
+            self.api_key = profile.api_key
+        if profile.base_url:
+            self.base_url = profile.base_url
+        if profile.model and profile.model != self.model:
+            # Only log if model actually changes
+            logging.info(f"Failover: switching from {self.model} to {profile.model}")
+            self.model = profile.model
+
+    def _resolve_subscription_creds(self):
+        """Lazy resolve + cache subscription credentials, checking expiration."""
+        if not self._auth_provider_id:
+            return None
+        
+        # Check if cached credentials are expired
+        if (self._cached_subscription_creds and 
+            self._cached_subscription_creds.expires_at_ms and
+            self._cached_subscription_creds.expires_at_ms <= int(time.time() * 1000)):
+            self._cached_subscription_creds = None
+            
+        if self._cached_subscription_creds is None:
+            from ..auth import resolve_subscription_credentials
+            self._cached_subscription_creds = resolve_subscription_credentials(
+                self._auth_provider_id,
+            )
+        return self._cached_subscription_creds
+
+    def _refresh_subscription_creds(self):
+        """Force refresh of subscription credentials."""
+        if not self._auth_provider_id:
+            return None
+        from ..auth.subscription.registry import get_subscription_provider
+        provider = get_subscription_provider(self._auth_provider_id)
+        if provider is None:
+            return None
+        self._cached_subscription_creds = provider.refresh()
+        return self._cached_subscription_creds
+
     def _call_with_retry(self, func, *args, **kwargs):
-        """Call a function with automatic retry on rate limit errors.
+        """Call a function with automatic retry on rate limit errors and failover support.
 
         Args:
             func: The function to call (e.g., litellm.completion)
@@ -672,20 +771,63 @@ Respond with ONLY a valid JSON tool call in this format:
                 if self._rate_limiter is not None:
                     self._rate_limiter.acquire()
 
-                return func(*args, **kwargs)
+                result = func(*args, **kwargs)
+                
+                # Mark success if failover is configured
+                if self._failover_manager and self._current_profile:
+                    self._failover_manager.mark_success(self._current_profile)
+                    
+                return result
 
             except Exception as e:
-                if not self._is_rate_limit_error(e):
-                    raise
-
+                category, can_retry, retry_delay = self._classify_error_and_should_retry(e, attempt + 1)
+                
                 last_error = e
                 error_str = str(e)
 
-                if attempt < self._max_retries:
-                    retry_delay = self._parse_retry_delay(error_str)
+                # Check for auth errors and try refreshing subscription credentials
+                if category == "auth" and self._auth_provider_id and attempt == 0:
+                    try:
+                        logging.info("Authentication error detected - attempting credential refresh")
+                        refreshed_creds = self._refresh_subscription_creds()
+                        if refreshed_creds:
+                            # Update parameters with refreshed credentials (don't clear cache)
+                            kwargs = self._build_completion_params(**kwargs)
+                            # Retry immediately with refreshed credentials
+                            can_retry = True
+                            retry_delay = 0.0
+                            logging.info("Subscription credentials refreshed, retrying...")
+                    except Exception as refresh_error:
+                        logging.warning(f"Failed to refresh subscription credentials: {refresh_error}")
 
+                # Failover: mark failure and try next profile (do this before early exit)
+                if self._failover_manager and self._current_profile:
+                    is_rate_limit = (category == "rate_limit")
+                    self._failover_manager.mark_failure(
+                        self._current_profile, error_str, is_rate_limit=is_rate_limit
+                    )
+                    next_profile = self._failover_manager.get_next_profile()
+                    if next_profile and next_profile != self._current_profile:
+                        self._switch_to_profile(next_profile)
+                        self._current_profile = next_profile
+                        # Update the kwargs with new profile values for the next retry
+                        if "api_key" in kwargs:
+                            kwargs["api_key"] = self.api_key
+                        if "base_url" in kwargs:
+                            kwargs["base_url"] = self.base_url
+                        if "model" in kwargs:
+                            kwargs["model"] = self.model
+                        # Enable retry for profile switch even if originally non-retryable
+                        can_retry = True
+                        retry_delay = 0.0
+                        logging.info(f"Failover: switched to profile '{next_profile.name}'")
+                
+                if not can_retry:
+                    raise
+
+                if attempt < self._max_retries:
                     logging.warning(
-                        f"Rate limit hit (attempt {attempt + 1}/{self._max_retries + 1}), "
+                        f"{category} error hit (attempt {attempt + 1}/{self._max_retries + 1}), "
                         f"waiting {retry_delay:.1f}s before retry..."
                     )
 
@@ -712,7 +854,7 @@ Respond with ONLY a valid JSON tool call in this format:
         raise last_error
 
     async def _call_with_retry_async(self, func, *args, **kwargs):
-        """Async version of _call_with_retry.
+        """Async version of _call_with_retry with failover support.
 
         Args:
             func: The async function to call
@@ -733,20 +875,63 @@ Respond with ONLY a valid JSON tool call in this format:
                 if self._rate_limiter is not None:
                     await self._rate_limiter.acquire_async()
 
-                return await func(*args, **kwargs)
+                result = await func(*args, **kwargs)
+                
+                # Mark success if failover is configured
+                if self._failover_manager and self._current_profile:
+                    self._failover_manager.mark_success(self._current_profile)
+                    
+                return result
 
             except Exception as e:
-                if not self._is_rate_limit_error(e):
-                    raise
-
+                category, can_retry, retry_delay = self._classify_error_and_should_retry(e, attempt + 1)
+                
                 last_error = e
                 error_str = str(e)
 
-                if attempt < self._max_retries:
-                    retry_delay = self._parse_retry_delay(error_str)
+                # Check for auth errors and try refreshing subscription credentials
+                if category == "auth" and self._auth_provider_id and attempt == 0:
+                    try:
+                        logging.info("Authentication error detected - attempting credential refresh")
+                        refreshed_creds = self._refresh_subscription_creds()
+                        if refreshed_creds:
+                            # Update parameters with refreshed credentials (don't clear cache)
+                            kwargs = self._build_completion_params(**kwargs)
+                            # Retry immediately with refreshed credentials
+                            can_retry = True
+                            retry_delay = 0.0
+                            logging.info("Subscription credentials refreshed, retrying...")
+                    except Exception as refresh_error:
+                        logging.warning(f"Failed to refresh subscription credentials: {refresh_error}")
 
+                # Failover: mark failure and try next profile (do this before early exit)
+                if self._failover_manager and self._current_profile:
+                    is_rate_limit = (category == "rate_limit")
+                    self._failover_manager.mark_failure(
+                        self._current_profile, error_str, is_rate_limit=is_rate_limit
+                    )
+                    next_profile = self._failover_manager.get_next_profile()
+                    if next_profile and next_profile != self._current_profile:
+                        self._switch_to_profile(next_profile)
+                        self._current_profile = next_profile
+                        # Update the kwargs with new profile values for the next retry
+                        if "api_key" in kwargs:
+                            kwargs["api_key"] = self.api_key
+                        if "base_url" in kwargs:
+                            kwargs["base_url"] = self.base_url
+                        if "model" in kwargs:
+                            kwargs["model"] = self.model
+                        # Enable retry for profile switch even if originally non-retryable
+                        can_retry = True
+                        retry_delay = 0.0
+                        logging.info(f"Failover: switched to profile '{next_profile.name}'")
+                
+                if not can_retry:
+                    raise
+
+                if attempt < self._max_retries:
                     logging.warning(
-                        f"Rate limit hit (attempt {attempt + 1}/{self._max_retries + 1}), "
+                        f"{category} error hit (attempt {attempt + 1}/{self._max_retries + 1}), "
                         f"waiting {retry_delay:.1f}s before retry..."
                     )
 
@@ -3368,7 +3553,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                         
                         # Continue conversation after tool execution - get follow-up response
                         try:
-                            follow_up_response = litellm.completion(
+                            follow_up_response = self._completion_with_retry(
                                 **self._build_completion_params(
                                     messages=messages,
                                     tools=formatted_tools,
@@ -3384,7 +3569,16 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                     # Yield the follow-up response after tool execution
                                     yield follow_up_content
                         except Exception as e:
-                            logging.error(f"Follow-up response failed: {e}")
+                            import time
+                            error_ref = f"followup-{int(time.time() * 1000)}"
+                            logging.error(
+                                "Follow-up response failed after retries (ref=%s, model=%s): %s",
+                                error_ref, getattr(self, 'model', 'unknown'), e
+                            )
+                            yield (
+                                f"\n\n[Error: Failed to generate final response after tool execution "
+                                f"(ref: {error_ref}). Please retry. If it continues, try reducing prompt size.]"
+                            )
                             
                 except Exception as e:
                     error_msg = str(e).lower()
@@ -3403,7 +3597,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
             if not use_streaming:
                 # Fall back to non-streaming and yield the complete response
                 try:
-                    response = litellm.completion(
+                    response = self._completion_with_retry(
                         **self._build_completion_params(
                             messages=messages,
                             tools=formatted_tools,
@@ -4463,6 +4657,18 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
             filtered_extra_settings = {k: v for k, v in self.extra_settings.items() if k != 'metrics'}
             params.update(filtered_extra_settings)
         
+        # Inject subscription credentials if auth provider is set
+        creds = self._resolve_subscription_creds()
+        if creds:
+            # Use litellm's native OAuth detection (auto-detects sk-ant-oat-* and switches to Bearer)
+            params["api_key"] = creds.api_key
+            if creds.base_url:
+                params["base_url"] = creds.base_url
+            if creds.headers:
+                extra_headers = dict(params.get("extra_headers") or {})
+                extra_headers.update(creds.headers)
+                params["extra_headers"] = extra_headers
+        
         # Override with any provided parameters
         params.update(override_params)
         

src/praisonai-agents/praisonaiagents/llm/openai_client.py

@@ -907,6 +907,8 @@ class OpenAIClient:
                     final_response = self._responses_to_chat_completion(raw)
                     
                     # Display the response text if display_fn provided
+                    if not final_response.choices or final_response.choices[0].message is None:
+                        raise ValueError("LLM returned empty or filtered response")
                     response_text = final_response.choices[0].message.content or ""
                     if display_fn and response_text:
                         from rich.markdown import Markdown
@@ -1167,6 +1169,8 @@ class OpenAIClient:
                     final_response = self._responses_to_chat_completion(raw)
                     
                     # Display the response text if display_fn provided
+                    if not final_response.choices or final_response.choices[0].message is None:
+                        raise ValueError("LLM returned empty or filtered response")
                     response_text = final_response.choices[0].message.content or ""
                     if display_fn and response_text:
                         from rich.markdown import Markdown
@@ -1638,8 +1642,10 @@ class OpenAIClient:
             )
             
             # Check for tool calls
+            if not final_response.choices or final_response.choices[0].message is None:
+                raise ValueError("LLM returned empty or filtered response")
             tool_calls = getattr(final_response.choices[0].message, 'tool_calls', None)
-            
+
             # Emit llm_content for intermediate narrative display
             # (gpt-4.1+ models produce text alongside tool calls)
             response_content = getattr(final_response.choices[0].message, 'content', None)
@@ -1835,10 +1841,12 @@ class OpenAIClient:
             
             if not final_response:
                 return None
-            
+
             # Check for tool calls
+            if not final_response.choices or final_response.choices[0].message is None:
+                raise ValueError("LLM returned empty or filtered response")
             tool_calls = getattr(final_response.choices[0].message, 'tool_calls', None)
-            
+
             # Emit llm_content for intermediate narrative display
             # (gpt-4.1+ models produce text alongside tool calls)
             response_content = getattr(final_response.choices[0].message, 'content', None)
@@ -2037,6 +2045,8 @@ class OpenAIClient:
                     return
                 
                 # Check for tool calls
+                if not final_response.choices or final_response.choices[0].message is None:
+                    raise ValueError("LLM returned empty or filtered response")
                 tool_calls = getattr(final_response.choices[0].message, 'tool_calls', None)
                 
                 if tool_calls and execute_tool_fn:

src/praisonai-agents/praisonaiagents/llm/rate_limiter.py

@@ -24,6 +24,7 @@ Usage:
 import time
 import asyncio
 import logging
+import threading
 from praisonaiagents._logging import get_logger
 from dataclasses import dataclass, field
 from typing import Optional, Callable
@@ -67,11 +68,15 @@ class RateLimiter:
     _tokens: float = field(default=None, init=False, repr=False)
     _last_update: float = field(default=None, init=False, repr=False)
     _lock: asyncio.Lock = field(default=None, init=False, repr=False)
+    _lock_init: "threading.Lock" = field(default=None, init=False, repr=False)
+    _sync_lock: "threading.Lock" = field(default=None, init=False, repr=False)
 
     # Internal state for token-based limiting
     _api_tokens: float = field(default=None, init=False, repr=False)
     _api_tokens_last_update: float = field(default=None, init=False, repr=False)
     _api_tokens_lock: asyncio.Lock = field(default=None, init=False, repr=False)
+    _api_tokens_lock_init: "threading.Lock" = field(default=None, init=False, repr=False)
+    _sync_api_tokens_lock: "threading.Lock" = field(default=None, init=False, repr=False)
 
     # Injectable functions for testing
     _get_time: Callable[[], float] = field(default=None, init=False, repr=False)
@@ -88,6 +93,9 @@ class RateLimiter:
         self._last_update = time.monotonic()
         # Lazily create asyncio.Lock() to avoid Python 3.9 event loop issues
         self._lock = None
+        if self._lock_init is None:
+            self._lock_init = threading.Lock()  # Threading lock for async lock initialization
+        self._sync_lock = threading.Lock()
 
         # Token-based limiting
         if self.tokens_per_minute is not None:
@@ -98,6 +106,9 @@ class RateLimiter:
         self._api_tokens_last_update = time.monotonic()
         # Lazily create asyncio.Lock() to avoid Python 3.9 event loop issues
         self._api_tokens_lock = None
+        if self._api_tokens_lock_init is None:
+            self._api_tokens_lock_init = threading.Lock()  # Threading lock for async lock initialization
+        self._sync_api_tokens_lock = threading.Lock()
 
         # Default implementations (can be overridden for testing)
         if self._get_time is None:
@@ -110,13 +121,17 @@ class RateLimiter:
     def _get_lock(self) -> asyncio.Lock:
         """Get or create the request lock lazily (Python 3.9 compatible)."""
         if self._lock is None:
-            self._lock = asyncio.Lock()
+            with self._lock_init:  # Thread-safe initialization
+                if self._lock is None:  # Double-check after acquiring lock
+                    self._lock = asyncio.Lock()
         return self._lock
 
     def _get_api_tokens_lock(self) -> asyncio.Lock:
         """Get or create the API tokens lock lazily (Python 3.9 compatible)."""
         if self._api_tokens_lock is None:
-            self._api_tokens_lock = asyncio.Lock()
+            with self._api_tokens_lock_init:  # Thread-safe initialization
+                if self._api_tokens_lock is None:  # Double-check after acquiring lock
+                    self._api_tokens_lock = asyncio.Lock()
         return self._api_tokens_lock
 
     @property
@@ -173,15 +188,15 @@ class RateLimiter:
         if self.requests_per_minute is None:
             return
 
-        self._refill()
-
-        wait = self._wait_time()
-        if wait > 0:
+        while True:
+            with self._sync_lock:
+                self._refill()
+                if self._tokens >= 1.0:
+                    self._tokens -= 1.0
+                    return
+                wait = self._wait_time()
             logger.debug(f"Rate limit: waiting {wait:.2f}s before next request")
             self._sleep(wait)
-            self._refill()
-
-        self._tokens -= 1.0
 
     def acquire_tokens(self, num_tokens: int) -> None:
         """Acquire API tokens, blocking if necessary.
@@ -197,17 +212,17 @@ class RateLimiter:
         if self.tokens_per_minute is None:
             return
 
-        self._refill_api_tokens()
-
-        wait = self._wait_time_for_tokens(num_tokens)
-        if wait > 0:
+        while True:
+            with self._sync_api_tokens_lock:
+                self._refill_api_tokens()
+                if self._api_tokens >= num_tokens:
+                    self._api_tokens -= num_tokens
+                    return
+                wait = self._wait_time_for_tokens(num_tokens)
             # Do NOT cap internal wait time - this ensures correct rate limiting
             # The wait time is calculated based on token refill rate and must be honored
             logger.debug(f"Token limit: waiting {wait:.2f}s for {num_tokens} tokens")
             self._sleep(wait)
-            self._refill_api_tokens()
-
-        self._api_tokens -= num_tokens
 
     async def acquire_async(self) -> None:
         """Acquire a request token asynchronously, waiting if necessary.
@@ -263,23 +278,24 @@ class RateLimiter:
         if self.requests_per_minute is None:
             return True
 
-        self._refill()
-
-        if self._tokens >= 1.0:
-            self._tokens -= 1.0
-            return True
+        with self._sync_lock:
+            self._refill()
+            if self._tokens >= 1.0:
+                self._tokens -= 1.0
+                return True
         return False
 
     def reset(self) -> None:
         """Reset the rate limiter to initial state."""
-        if self.requests_per_minute is not None:
-            self._tokens = float(self.burst)
-        self._last_update = self._get_time()
-
-        if self.tokens_per_minute is not None:
-            token_burst = max(self.burst, self.tokens_per_minute // 60)
-            self._api_tokens = float(token_burst)
-        self._api_tokens_last_update = self._get_time()
+        with self._sync_lock:
+            if self.requests_per_minute is not None:
+                self._tokens = float(self.burst)
+            self._last_update = self._get_time()
+        with self._sync_api_tokens_lock:
+            if self.tokens_per_minute is not None:
+                token_burst = max(self.burst, self.tokens_per_minute // 60)
+                self._api_tokens = float(token_burst)
+            self._api_tokens_last_update = self._get_time()
 
     def wait_for_retry(self, retry_delay: float) -> None:
         """Wait for a specified retry delay (for handling 429 errors).
@@ -304,14 +320,16 @@ class RateLimiter:
     @property
     def available_tokens(self) -> float:
         """Get current available request tokens (for monitoring)."""
-        self._refill()
-        return self._tokens
+        with self._sync_lock:
+            self._refill()
+            return self._tokens
 
     @property
     def available_api_tokens(self) -> float:
         """Get current available API tokens (for monitoring)."""
-        self._refill_api_tokens()
-        return self._api_tokens
+        with self._sync_api_tokens_lock:
+            self._refill_api_tokens()
+            return self._api_tokens
 
     def __repr__(self) -> str:
         parts = []

src/praisonai-agents/praisonaiagents/llm/sanitize.py

@@ -0,0 +1,145 @@
+"""
+Message Sanitization - Clean surrogate/non-ASCII characters before LLM calls.
+
+Prevents Unicode encoding issues that can cause silent failures with some
+providers when processing emoji, non-Latin text, or corrupted Unicode.
+"""
+
+import re
+from typing import List, Dict, Any, Union
+
+__all__ = ["sanitize_messages", "strip_surrogates"]
+
+
+def strip_surrogates(text: str) -> str:
+    """Remove surrogate pairs and malformed Unicode from text.
+    
+    Surrogate pairs (U+D800-U+DFFF) are used in UTF-16 encoding but are
+    invalid in UTF-8/Unicode strings. They can appear from:
+    - Incorrect Unicode conversion
+    - Corrupted text data  
+    - Invalid emoji/character sequences
+    
+    Args:
+        text: Input text potentially containing surrogates
+        
+    Returns:
+        Text with surrogates removed or replaced
+        
+    Examples:
+        >>> strip_surrogates("Hello \\uD83D World")  # Missing low surrogate
+        'Hello  World'
+        >>> strip_surrogates("Valid text 🌍")
+        'Valid text 🌍'
+    """
+    if not text:
+        return text
+    
+    try:
+        # Method 1: Encode with surrogatepass, decode with ignore
+        # This converts surrogates to UTF-16, then back to UTF-8 safely
+        return text.encode('utf-16', 'surrogatepass').decode('utf-16', 'ignore')
+    except (UnicodeError, LookupError):
+        # Fallback: Remove surrogate code points directly
+        return re.sub(r'[\uD800-\uDFFF]', '', text)
+
+
+def _sanitize_value_recursive(value: Any) -> tuple[Any, bool]:
+    """Recursively sanitize any nested string values in a data structure.
+    
+    Args:
+        value: Value to sanitize (can be dict, list, str, or other)
+        
+    Returns:
+        Tuple of (sanitized_value, changed_flag)
+    """
+    if isinstance(value, str):
+        sanitized = strip_surrogates(value)
+        return sanitized, sanitized != value
+    
+    elif isinstance(value, list):
+        changed = False
+        sanitized_items = []
+        for item in value:
+            sanitized_item, item_changed = _sanitize_value_recursive(item)
+            sanitized_items.append(sanitized_item)
+            changed = changed or item_changed
+        return sanitized_items, changed
+    
+    elif isinstance(value, dict):
+        changed = False
+        sanitized_dict = {}
+        for key, nested_value in value.items():
+            # Sanitize the key if it's a string
+            if isinstance(key, str):
+                sanitized_key = strip_surrogates(key)
+                if sanitized_key != key:
+                    changed = True
+                    key = sanitized_key
+            
+            # Recursively sanitize the value
+            sanitized_value, value_changed = _sanitize_value_recursive(nested_value)
+            sanitized_dict[key] = sanitized_value
+            changed = changed or value_changed
+        return sanitized_dict, changed
+    
+    else:
+        # Return other types unchanged
+        return value, False
+
+
+def sanitize_messages(messages: List[Dict[str, Any]]) -> bool:
+    """Sanitize message content in-place, removing problematic Unicode.
+    
+    Processes all string content in message dictionaries, including:
+    - message.content (string or list)  
+    - message.name
+    - Any nested string values (including tool_calls[].function.arguments)
+    
+    Args:
+        messages: List of message dicts to sanitize in-place
+        
+    Returns:
+        True if any changes were made, False otherwise
+        
+    Examples:
+        >>> messages = [{"content": "Hello \\uD83D World", "role": "user"}]
+        >>> changed = sanitize_messages(messages)
+        >>> assert changed == True
+        >>> assert messages[0]["content"] == "Hello  World"
+    """
+    if not messages:
+        return False
+    
+    changed = False
+    
+    for message in messages:
+        if not isinstance(message, dict):
+            continue
+            
+        # Recursively sanitize the entire message structure
+        sanitized_message, message_changed = _sanitize_value_recursive(message)
+        if message_changed:
+            message.clear()
+            message.update(sanitized_message)
+            changed = True
+    
+    return changed
+
+
+def sanitize_text(text: Union[str, None]) -> Union[str, None]:
+    """Sanitize a single text string.
+    
+    Convenience function for sanitizing individual strings.
+    
+    Args:
+        text: Text to sanitize, or None
+        
+    Returns:
+        Sanitized text, or None if input was None
+    """
+    if text is None:
+        return None
+    if not isinstance(text, str):
+        return text
+    return strip_surrogates(text)

src/praisonai-agents/praisonaiagents/main.py

@@ -676,6 +676,8 @@ class TaskOutput(BaseModel):
     agent: str
     output_format: Literal["RAW", "JSON", "Pydantic"] = "RAW"
     token_metrics: Optional['TokenMetrics'] = None  # Add token metrics field
+    callback_error: Optional[str] = None
+    non_fatal_errors: Optional[list[str]] = None
 
     def json(self) -> Optional[str]:
         if self.output_format == "JSON" and self.json_dict:
@@ -696,4 +698,4 @@ class TaskOutput(BaseModel):
         elif self.json_dict:
             return json.dumps(self.json_dict)
         else:
-            return self.raw 
+            return self.raw 

src/praisonai-agents/praisonaiagents/managed/__init__.py

@@ -25,6 +25,7 @@ from .protocols import (
     ComputeConfig,
     InstanceInfo,
     InstanceStatus,
+    ManagedRuntimeProtocol,
 )
 
 __all__ = [
@@ -41,6 +42,7 @@ __all__ = [
     "ComputeConfig",
     "InstanceInfo",
     "InstanceStatus",
+    "ManagedRuntimeProtocol",
     "ManagedBackendProtocol",
 ]
 

src/praisonai-agents/praisonaiagents/managed/protocols.py

@@ -23,6 +23,7 @@ from dataclasses import dataclass, field
 from enum import Enum
 from typing import (
     Any,
+    AsyncIterator,
     Dict,
     List,
     Optional,
@@ -216,3 +217,140 @@ class ComputeProviderProtocol(Protocol):
             List of InstanceInfo for running instances.
         """
         ...
+
+
+@runtime_checkable
+class ManagedRuntimeProtocol(Protocol):
+    """Protocol for hosted agent runtime — the agent LOOP runs on remote infra.
+    
+    Separates from ComputeProviderProtocol which only sandboxes tool execution.
+    Here, the entire agent loop (LLM calls, tool calling, event streaming, 
+    memory, session management) runs remotely.
+    
+    This is the protocol that Anthropic Managed Agents, E2B Managed Runtime,
+    Modal Managed Functions, etc. should implement.
+    
+    Example lifecycle::
+    
+        runtime = AnthropicManagedAgent(config=cfg)
+        agent_id = await runtime.create_agent(config)
+        env_id = await runtime.create_environment(config)
+        session_id = await runtime.create_session(agent_id, env_id)
+        
+        # Send user message and stream agent responses
+        await runtime.send_event(session_id, {"type": "user.message", ...})
+        async for event in runtime.stream_events(session_id):
+            if event["type"] == "agent.message":
+                print(event["content"])
+    """
+
+    async def create_agent(self, config: Dict[str, Any]) -> str:
+        """Create agent definition on remote infrastructure.
+        
+        Args:
+            config: Agent configuration (model, system, tools, etc.)
+            
+        Returns:
+            Agent ID assigned by the remote provider
+        """
+        ...
+
+    async def create_environment(self, config: Dict[str, Any]) -> str:
+        """Create environment (container/sandbox) on remote infrastructure.
+        
+        Args:
+            config: Environment configuration (packages, networking, etc.)
+            
+        Returns:
+            Environment ID assigned by the remote provider
+        """
+        ...
+
+    async def create_session(
+        self, 
+        agent_id: str, 
+        environment_id: str, 
+        **kwargs
+    ) -> str:
+        """Create running session (agent-in-environment) on remote infrastructure.
+        
+        Args:
+            agent_id: Agent ID from create_agent()
+            environment_id: Environment ID from create_environment()
+            **kwargs: Additional session config (vault_ids, resources, etc.)
+            
+        Returns:
+            Session ID assigned by the remote provider
+        """
+        ...
+
+    async def send_event(self, session_id: str, event: Dict[str, Any]) -> None:
+        """Send event to running session.
+        
+        Args:
+            session_id: Session ID from create_session()
+            event: Event dict (e.g. {"type": "user.message", "content": [...]})
+        """
+        ...
+
+    def stream_events(self, session_id: str) -> AsyncIterator[Dict[str, Any]]:
+        """Stream events from running session.
+
+        Implementations should be async generators: ``async def`` with ``yield``.
+        Declared without ``async def`` here so the protocol signature matches the
+        async-generator return type (``AsyncIterator``) rather than a coroutine
+        that resolves to one, letting callers write ``async for e in runtime.stream_events(sid)``.
+
+        Args:
+            session_id: Session ID from create_session()
+
+        Yields:
+            Event dicts (agent.message, agent.tool_use, session.status_idle, etc.)
+        """
+        ...
+
+    async def interrupt(self, session_id: str) -> None:
+        """Send interrupt signal to running session.
+        
+        Args:
+            session_id: Session ID to interrupt
+        """
+        ...
+
+    async def retrieve_session(self, session_id: str) -> Dict[str, Any]:
+        """Retrieve session metadata and status.
+        
+        Args:
+            session_id: Session ID to retrieve
+            
+        Returns:
+            Session info dict with id, status, title, usage, etc.
+        """
+        ...
+
+    async def list_sessions(self, **filters) -> List[Dict[str, Any]]:
+        """List sessions with optional filtering.
+        
+        Args:
+            **filters: Filter criteria (agent_id, status, limit, etc.)
+            
+        Returns:
+            List of session info dicts
+        """
+        ...
+
+    async def archive_session(self, session_id: str) -> None:
+        """Archive session (preserve but mark inactive).
+        
+        Args:
+            session_id: Session ID to archive
+        """
+        ...
+
+    async def delete_session(self, session_id: str) -> None:
+        """Delete session permanently.
+        
+        Args:
+            session_id: Session ID to delete
+        """
+        ...

src/praisonai-agents/praisonaiagents/memory/adapters/factories.py

@@ -40,7 +40,8 @@ def create_mem0_memory_adapter(**kwargs) -> MemoryProtocol:
         import mem0
     except ImportError:
         raise ImportError(
-            "mem0 is required for mem0 adapter. Install with: pip install mem0ai"
+            "mem0ai is not installed. "
+            "Run: pip install 'praisonaiagents[memory]'"
         )
     
     return Mem0MemoryAdapter(mem0_config=kwargs)
@@ -304,6 +305,12 @@ class ChromaMemoryAdapter:
             else:
                 sanitized[k] = str(v)
         return sanitized
+    
+    def close(self):
+        """Clean up ChromaDB resources."""
+        # ChromaDB PersistentClient has no explicit close; release references.
+        self.client = None
+        self.collection = None
 
 
 class MongoDBMemoryAdapter:
@@ -357,7 +364,7 @@ class MongoDBMemoryAdapter:
     
     def store_short_term(self, text: str, metadata: Optional[Dict[str, Any]] = None, **kwargs) -> str:
         """Store in MongoDB short-term collection."""
-        from datetime import datetime
+        from datetime import datetime, timezone
         import time
         
         doc_id = str(time.time_ns())
@@ -365,7 +372,7 @@ class MongoDBMemoryAdapter:
             "_id": doc_id,
             "content": text,
             "metadata": metadata or {},
-            "created_at": datetime.now(datetime.timezone.utc),
+            "created_at": datetime.now(timezone.utc),
             "memory_type": "short_term"
         }
         
@@ -389,7 +396,7 @@ class MongoDBMemoryAdapter:
     
     def store_long_term(self, text: str, metadata: Optional[Dict[str, Any]] = None, **kwargs) -> str:
         """Store in MongoDB long-term collection."""
-        from datetime import datetime
+        from datetime import datetime, timezone
         import time
         
         doc_id = str(time.time_ns())
@@ -397,7 +404,7 @@ class MongoDBMemoryAdapter:
             "_id": doc_id,
             "content": text,
             "metadata": metadata or {},
-            "created_at": datetime.now(datetime.timezone.utc),
+            "created_at": datetime.now(timezone.utc),
             "memory_type": "long_term"
         }
         
@@ -492,4 +499,15 @@ class MongoDBMemoryAdapter:
             result = embedding(text, model="text-embedding-3-small")
             return result.embeddings[0] if result.embeddings else None
         except Exception:
-            return None
+            return None
+    
+    def close(self):
+        """Clean up MongoDB resources."""
+        if hasattr(self, 'client') and self.client:
+            try:
+                self.client.close()
+            except Exception as e:
+                import logging
+                logging.warning(f"MongoDB cleanup failed: {e}")
+            finally:
+                self.client = None

src/praisonai-agents/praisonaiagents/memory/adapters/in_memory_adapter.py

@@ -23,17 +23,27 @@ class InMemoryAdapter:
 
     def __init__(self, **kwargs) -> None:
         self._data: List[Dict[str, Any]] = []
+        self._max_size: int = kwargs.get("max_size", 10_000)  # Default limit
+        self._next_id: int = 0  # Monotonic counter to prevent ID collisions after eviction
+    
+    def _evict_if_needed(self):
+        """Evict old entries if we exceed max size (FIFO eviction)."""
+        if len(self._data) > self._max_size:
+            # Keep only the most recent entries
+            self._data = self._data[len(self._data) - self._max_size:]
 
     def store_short_term(
         self, text: str, metadata: Optional[Dict[str, Any]] = None, **kwargs
     ) -> str:
         entry: Dict[str, Any] = {
-            "id": str(len(self._data)),
+            "id": str(self._next_id),
             "text": text,
             "type": "short",
             "metadata": metadata,
         }
+        self._next_id += 1
         self._data.append(entry)
+        self._evict_if_needed()
         return entry["id"]
 
     def search_short_term(
@@ -50,12 +60,14 @@ class InMemoryAdapter:
         self, text: str, metadata: Optional[Dict[str, Any]] = None, **kwargs
     ) -> str:
         entry: Dict[str, Any] = {
-            "id": str(len(self._data)),
+            "id": str(self._next_id),
             "text": text,
             "type": "long",
             "metadata": metadata,
         }
+        self._next_id += 1
         self._data.append(entry)
+        self._evict_if_needed()
         return entry["id"]
 
     def search_long_term(

src/praisonai-agents/praisonaiagents/memory/core.py

@@ -416,7 +416,7 @@ class MemoryCoreMixin:
 
     # Async variants to prevent event loop blocking
     async def store_short_term_async(self, content: str, metadata: Optional[Dict] = None, quality_score: Optional[float] = None, 
-                                   user_id: Optional[str] = None, auto_promote: bool = True) -> str:
+                                   user_id: Optional[str] = None, auto_promote: bool = True, **kwargs) -> str:
         """
         Async version of store_short_term to prevent event loop blocking.
         
@@ -426,6 +426,7 @@ class MemoryCoreMixin:
             quality_score: Optional pre-calculated quality score
             user_id: Optional user identifier
             auto_promote: Whether to automatically promote to LTM if quality is high
+            **kwargs: Additional provider-specific arguments forwarded to the adapter
             
         Returns:
             The memory ID of the stored content
@@ -442,8 +443,7 @@ class MemoryCoreMixin:
         raw_metadata = metadata.copy() if metadata else {}
         raw_metadata.update({
             "timestamp": datetime.now().isoformat(),
-            "quality_score": quality_score,
-            "memory_type": "short_term"
+            "quality": quality_score,
         })
         if user_id:
             raw_metadata["user_id"] = user_id
@@ -463,7 +463,7 @@ class MemoryCoreMixin:
         # Only use SQLite fallback if primary storage failed completely
         if not memory_id and hasattr(self, '_sqlite_adapter') and self._sqlite_adapter != getattr(self, 'memory_adapter', None):
             try:
-                memory_id = await asyncio.to_thread(self._store_sqlite_stm, content, clean_metadata, quality_score)
+                memory_id = await asyncio.to_thread(self._sqlite_adapter.store_short_term, content, metadata=clean_metadata, **kwargs)
                 self._log_verbose(f"Stored in SQLite async STM as fallback: {content[:100]}...")
             except Exception as e:
                 logging.error(f"Failed to store in SQLite async STM fallback: {e}")
@@ -472,7 +472,7 @@ class MemoryCoreMixin:
         # Auto-promote to long-term memory if quality is high (async)
         if auto_promote and quality_score >= 7.5:  # High quality threshold
             try:
-                await self.store_long_term_async(content, clean_metadata, quality_score, user_id)
+                await self.store_long_term_async(content, clean_metadata, quality_score, user_id, **kwargs)
                 self._log_verbose(f"Auto-promoted STM content to LTM (score: {quality_score:.2f})")
             except Exception as e:
                 logging.warning(f"Failed to auto-promote to LTM: {e}")
@@ -485,7 +485,7 @@ class MemoryCoreMixin:
         return memory_id or ""
 
     async def store_long_term_async(self, content: str, metadata: Optional[Dict] = None, quality_score: Optional[float] = None,
-                                  user_id: Optional[str] = None) -> str:
+                                  user_id: Optional[str] = None, **kwargs) -> str:
         """
         Async version of store_long_term to prevent event loop blocking.
         
@@ -494,6 +494,7 @@ class MemoryCoreMixin:
             metadata: Optional metadata dictionary
             quality_score: Optional pre-calculated quality score
             user_id: Optional user identifier
+            **kwargs: Additional provider-specific arguments forwarded to the adapter
             
         Returns:
             The memory ID of the stored content
@@ -506,4 +507,4 @@ class MemoryCoreMixin:
             quality_score = self.compute_quality_score(content, metadata)
         
         # Use sync version in thread to avoid blocking event loop
-        return await asyncio.to_thread(self.store_long_term, content, metadata, quality_score, user_id)
+        return await asyncio.to_thread(self.store_long_term, content, metadata, quality_score, user_id, **kwargs)

src/praisonai-agents/praisonaiagents/memory/memory.py

@@ -7,7 +7,7 @@ import threading
 from typing import Any, Dict, List, Optional, Union, Literal
 import logging
 from praisonaiagents._logging import get_logger
-from datetime import datetime
+from datetime import datetime, timezone
 
 # Decomposed memory functionality - imported as mixins for backward compatibility
 from .storage import StorageMixin
@@ -143,6 +143,9 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
         adapter_name = provider_mapping.get(provider, provider)
         
         # Try to get preferred adapter, fallback to available ones
+        adapter = None
+        provider_explicitly_requested = provider != "rag"  # "rag" is default, others are explicit
+        
         try:
             adapter_config = self._get_adapter_config_for_provider(adapter_name)
             adapter = get_memory_adapter(adapter_name, **adapter_config)
@@ -154,6 +157,37 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
             adapter = None
         
         if adapter is None:
+            # If the provider was explicitly requested, try to give a helpful error message
+            if provider_explicitly_requested:
+                try:
+                    # Try the factory function directly to get the specific ImportError message
+                    from .adapters.factories import (
+                        create_mem0_memory_adapter,
+                        create_chroma_memory_adapter,
+                        create_mongodb_memory_adapter,
+                    )
+                    factory_map = {
+                        "mem0": create_mem0_memory_adapter,
+                        "chroma": create_chroma_memory_adapter, 
+                        "mongodb": create_mongodb_memory_adapter,
+                    }
+                    if adapter_name in factory_map:
+                        # Call the factory to get ImportError with installation hint
+                        # or capture successful adapter creation
+                        adapter_config = self._get_adapter_config_for_provider(adapter_name)
+                        adapter = factory_map[adapter_name](**adapter_config)
+                except ImportError:
+                    # Re-raise ImportError with installation instructions
+                    raise
+                # All other exceptions (e.g., ValueError for missing API key) 
+                # propagate naturally, preserving backward compatibility
+            
+            # If factory succeeded, use that adapter
+            if adapter is not None:
+                self.memory_adapter = adapter
+                self.provider = adapter_name
+                return
+                
             # Fallback to first available adapter
             self._log_verbose(f"Provider '{adapter_name}' not available, trying fallbacks")
             # Try each fallback preference individually
@@ -460,6 +494,9 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
     def _init_mongodb(self):
         """Initialize MongoDB client for memory storage."""
         try:
+            # Import MongoClient locally to handle optional dependency
+            from pymongo import MongoClient
+            
             mongo_cfg = self.cfg.get("config", {})
             self.connection_string = mongo_cfg.get("connection_string", "mongodb://localhost:27017/")
             self.database_name = mongo_cfg.get("database", "praisonai")
@@ -635,36 +672,49 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
         ident = str(time.time_ns())
         created_at = time.time()
         
-        # Store in MongoDB if enabled
-        if self.use_mongodb and hasattr(self, "mongo_short_term"):
+        # Protocol-driven storage: Try adapter first if available
+        adapter_success = False
+        if hasattr(self, 'memory_adapter') and self.memory_adapter:
             try:
-                doc = {
-                    "_id": ident,
-                    "content": text,
-                    "metadata": metadata,
-                    "created_at": datetime.utcnow(),
-                    "memory_type": "short_term"
-                }
-                self.mongo_short_term.insert_one(doc)
-                logger.info(f"Successfully stored in MongoDB short-term memory with ID: {ident}")
+                result_id = self.memory_adapter.store_short_term(text, metadata=metadata)
+                logger.info(f"Successfully stored via memory adapter with ID: {result_id}")
+                adapter_success = True
+                ident = result_id  # Use adapter-provided ID
             except Exception as e:
-                logger.error(f"Failed to store in MongoDB short-term memory: {e}")
-                raise
+                logger.warning(f"Failed to store via memory adapter, falling back to direct storage: {e}")
+        
+        # Only use direct storage if adapter failed or doesn't exist
+        if not adapter_success:
+            # Store in MongoDB if enabled
+            if self.use_mongodb and hasattr(self, "mongo_short_term"):
+                try:
+                    doc = {
+                        "_id": ident,
+                        "content": text,
+                        "metadata": metadata,
+                        "created_at": datetime.now(timezone.utc),
+                        "memory_type": "short_term"
+                    }
+                    self.mongo_short_term.insert_one(doc)
+                    logger.info(f"Successfully stored in MongoDB short-term memory with ID: {ident}")
+                except Exception as e:
+                    logger.error(f"Failed to store in MongoDB short-term memory: {e}")
+                    raise
 
-        # Existing SQLite store logic (with write lock for concurrency safety)
-        try:
-            conn = self._get_stm_conn()
-            with self._write_lock:  # Serialize write operations
-                conn.execute(
-                    "INSERT INTO short_mem (id, content, meta, created_at) VALUES (?,?,?,?)",
-                    (ident, text, json.dumps(metadata), created_at)
-                )
-                conn.commit()
-            logger.info(f"Successfully stored in SQLite short-term memory with ID: {ident}")
-        except Exception as e:
-            logger.error(f"Failed to store in SQLite short-term memory: {e}")
-            if not self.use_mongodb:  # Only raise if we're not using MongoDB as fallback
-                raise
+            # Existing SQLite store logic (with write lock for concurrency safety)
+            try:
+                conn = self._get_stm_conn()
+                with self._write_lock:  # Serialize write operations
+                    conn.execute(
+                        "INSERT INTO short_mem (id, content, meta, created_at) VALUES (?,?,?,?)",
+                        (ident, text, json.dumps(metadata), created_at)
+                    )
+                    conn.commit()
+                logger.info(f"Successfully stored in SQLite short-term memory with ID: {ident}")
+            except Exception as e:
+                logger.error(f"Failed to store in SQLite short-term memory: {e}")
+                if not self.use_mongodb:  # Only raise if we're not using MongoDB as fallback
+                    raise
         
         # Emit trace event for memory store
         self._emit_memory_event("store", "short_term", len(text), metadata=metadata)
@@ -868,7 +918,7 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
                     "_id": ident,
                     "content": text,
                     "metadata": metadata,
-                    "created_at": datetime.utcnow(),
+                    "created_at": datetime.now(timezone.utc),
                     "memory_type": "long_term"
                 }
                 
@@ -965,9 +1015,9 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
             # Filter by quality
             filtered = [r for r in results if r.get("metadata", {}).get("quality", 0.0) >= min_quality]
             logger.info(f"Found {len(filtered)} results in Mem0")
-            return filtered
+            found.extend(filtered)
 
-        elif self.use_mongodb and hasattr(self, "mongo_long_term"):
+        if not found and self.use_mongodb and hasattr(self, "mongo_long_term"):
             try:
                 results = []
                 
@@ -1031,13 +1081,13 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
                         })
                 
                 logger.info(f"Found {len(results)} results in MongoDB")
-                return results
+                found.extend(results)
                 
             except Exception as e:
                 self._log_verbose(f"Error searching MongoDB long-term memory: {e}", logging.ERROR)
                 # Fall through to SQLite search
 
-        elif self.use_rag and hasattr(self, "chroma_col"):
+        if not found and self.use_rag and hasattr(self, "chroma_col"):
             try:
                 from praisonaiagents.embedding import embedding as get_embedding
                 result = get_embedding(query, model=self.embedding_model)
@@ -1045,29 +1095,28 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
                 
                 if query_embedding is None:
                     self._log_verbose("Failed to get embedding for query", logging.WARNING)
-                    return []
-                
-                # Search ChromaDB with embedding
-                resp = self.chroma_col.query(
-                    query_embeddings=[query_embedding],
-                    n_results=limit,
-                    include=["documents", "metadatas", "distances"]
-                )
-                
-                results = []
-                if resp["ids"]:
-                    for i in range(len(resp["ids"][0])):
-                        metadata = resp["metadatas"][0][i] if "metadatas" in resp else {}
-                        text = resp["documents"][0][i]
-                        # Add memory record citation
-                        text = f"{text} (Memory record: {resp['ids'][0][i]})"
-                        found.append({
-                            "id": resp["ids"][0][i],
-                            "text": text,
-                            "metadata": metadata,
-                            "score": 1.0 - (resp["distances"][0][i] if "distances" in resp else 0.0)
-                        })
-                logger.info(f"Found {len(found)} results in ChromaDB")
+                else:
+                    # Search ChromaDB with embedding
+                    resp = self.chroma_col.query(
+                        query_embeddings=[query_embedding],
+                        n_results=limit,
+                        include=["documents", "metadatas", "distances"]
+                    )
+                    
+                    results = []
+                    if resp["ids"]:
+                        for i in range(len(resp["ids"][0])):
+                            metadata = resp["metadatas"][0][i] if "metadatas" in resp else {}
+                            text = resp["documents"][0][i]
+                            # Add memory record citation
+                            text = f"{text} (Memory record: {resp['ids'][0][i]})"
+                            found.append({
+                                "id": resp["ids"][0][i],
+                                "text": text,
+                                "metadata": metadata,
+                                "score": 1.0 - (resp["distances"][0][i] if "distances" in resp else 0.0)
+                            })
+                    logger.info(f"Found {len(found)} results in ChromaDB")
 
             except Exception as e:
                 self._log_verbose(f"Error searching ChromaDB: {e}", logging.ERROR)
@@ -1085,7 +1134,7 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
             text = row[1]
             # Add memory record citation if not already present
             if "(Memory record:" not in text:
-                text = f"{text} (Memory record: {text})"
+                text = f"{text} (Memory record: {row[0]})"
             # Only add if not already found by ChromaDB/Mem0
             if not any(f["id"] == row[0] for f in found):
                 found.append({
@@ -1392,14 +1441,14 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
             self.mem0_client.add(text, user_id=user_id, metadata=meta)
         elif self.use_mongodb and hasattr(self, "mongo_users"):
             try:
-                from datetime import datetime
+                from datetime import datetime, timezone
                 ident = str(time.time_ns())
                 doc = {
                     "_id": ident,
                     "user_id": user_id,
                     "content": text,
                     "metadata": meta,
-                    "created_at": datetime.utcnow()
+                    "created_at": datetime.now(timezone.utc)
                 }
                 self.mongo_users.insert_one(doc)
                 self._log_verbose(f"Successfully stored user memory for {user_id}")
@@ -1945,14 +1994,19 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
     def close_connections(self):
         """
         Close database connections.
-        
-        Closes the current thread's connections and attempts to close all known
-        connections from other threads. Each thread should call this method before
-        terminating to ensure proper cleanup.
+
+        Closes only the calling thread's STM/LTM SQLite connections (and the
+        process-wide MongoDB client, if any). Connections owned by other threads
+        are left untouched so concurrent agents are not interfered with — each
+        thread is responsible for calling this before terminating. For full
+        process shutdown, use the adapter's own bulk-close path.
         """
         # Close current thread's connections
         if hasattr(self._local, 'stm_conn') and self._local.stm_conn:
             try:
+                # Remove from registry before closing
+                with self._connection_lock:
+                    self._all_connections.discard(self._local.stm_conn)
                 self._local.stm_conn.close()
                 self._local.stm_conn = None
             except Exception as e:
@@ -1960,21 +2014,37 @@ class Memory(StorageMixin, SearchMixin, MemoryCoreMixin):
         
         if hasattr(self._local, 'ltm_conn') and self._local.ltm_conn:
             try:
+                # Remove from registry before closing
+                with self._connection_lock:
+                    self._all_connections.discard(self._local.ltm_conn)
                 self._local.ltm_conn.close()
                 self._local.ltm_conn = None
             except Exception as e:
                 logger.warning(f"Error closing current thread's LTM connection: {e}")
         
-        # Close all known connections from the registry
-        with self._connection_lock:  # Ensure thread safety during cleanup
-            connections_to_close = list(self._all_connections)
-            for conn in connections_to_close:
-                try:
-                    conn.close()
-                except Exception as e:
-                    logger.debug(f"Error closing registered connection: {e}")
-            # Clear the registry
-            self._all_connections.clear()
+        # NOTE: Only close THIS thread's connections, not ALL connections
+        # Other active threads should manage their own connections
+        # Only clear full registry in __del__ or explicit shutdown
+        
+        # Close memory adapter if it exists (protocol-driven path)
+        if hasattr(self, 'memory_adapter') and self.memory_adapter:
+            try:
+                if hasattr(self.memory_adapter, 'close') and callable(self.memory_adapter.close):
+                    self.memory_adapter.close()
+                    logger.debug("Memory adapter closed successfully")
+            except Exception as e:
+                logger.warning(f"Error closing memory adapter: {e}")
+            finally:
+                self.memory_adapter = None
+        
+        # Close MongoDB client if it exists (legacy path)
+        if hasattr(self, 'mongo_client') and self.mongo_client:
+            try:
+                self.mongo_client.close()
+                self.mongo_client = None
+                logger.debug("MongoDB client closed successfully")
+            except Exception as e:
+                logger.warning(f"Error closing MongoDB client: {e}")
     
     def __enter__(self):
         """Allow Memory to be used as a context manager."""

src/praisonai-agents/praisonaiagents/memory/storage.py

@@ -95,13 +95,23 @@ class StorageMixin:
         # Initialize collections
         try:
             self.stm_collection = self.chroma_client.get_collection("short_term_memory")
-        except:
+        except ValueError:  # ChromaDB raises ValueError for missing collections
             self.stm_collection = self.chroma_client.create_collection("short_term_memory")
+        except Exception as e:
+            from .._logging import get_logger
+            logger = get_logger(__name__)
+            logger.error(f"Failed to initialize short_term_memory collection: {e}")
+            raise
             
         try:
             self.ltm_collection = self.chroma_client.get_collection("long_term_memory")
-        except:
+        except ValueError:  # ChromaDB raises ValueError for missing collections
             self.ltm_collection = self.chroma_client.create_collection("long_term_memory")
+        except Exception as e:
+            from .._logging import get_logger
+            logger = get_logger(__name__)
+            logger.error(f"Failed to initialize long_term_memory collection: {e}")
+            raise
     
     def _init_mongodb(self):
         """Initialize MongoDB for storage."""

src/praisonai-agents/praisonaiagents/output/models.py

@@ -24,6 +24,8 @@ class TaskOutput(BaseModel):
     agent: str
     output_format: Literal["RAW", "JSON", "Pydantic"] = "RAW"
     token_metrics: Optional['TokenMetrics'] = None
+    callback_error: Optional[str] = None
+    non_fatal_errors: Optional[list[str]] = None
 
     def json(self) -> Optional[str]:
         if self.output_format == "JSON" and self.json_dict:

src/praisonai-agents/praisonaiagents/process/process.py

@@ -67,7 +67,9 @@ class Process:
     async def _get_state_lock(self):
         """Get or create the async state lock (must be called from async context)."""
         if self._state_lock is None:
-            self._state_lock = asyncio.Lock()
+            with self._state_lock_init:  # Use existing threading.Lock for synchronization
+                if self._state_lock is None:  # Double-check after acquiring
+                    self._state_lock = asyncio.Lock()
         return self._state_lock
 
     def _create_llm_instance(self):
@@ -179,6 +181,14 @@ class Process:
             logging.error(f"Failed to create subtasks for loop task {loop_task.name}: {e}")
             import traceback
             traceback.print_exc()
+            # Mark task as failed to ensure error is visible
+            loop_task.status = "failed"
+            if hasattr(loop_task, 'result') and loop_task.result is None:
+                from ..output.models import TaskOutput
+                loop_task.result = TaskOutput(
+                    raw=f"Failed to read input file: {e}",
+                    task_id=loop_task.id
+                )
 
     def _build_task_context(self, current_task: Task) -> str:
         """Build context for a task based on its retain_full_context setting"""
@@ -236,6 +246,12 @@ class Process:
                         break  # Only include the most recent one
                         
         return context
+    
+    async def _set_workflow_finished(self, value: bool):
+        """Thread-safe setter for workflow_finished flag."""
+        lock = await self._get_state_lock()
+        async with lock:
+            self.workflow_finished = value
 
     def _find_next_not_started_task(self) -> Optional[Task]:
         """Fallback mechanism to find the next 'not started' task."""
@@ -243,9 +259,10 @@ class Process:
         temp_current_task = None
         
         # Clear previous task context before finding next task
+        # NOTE: Fixed to only clear _execution_context instead of mutating descriptions
         for task in self.tasks.values():
-            if hasattr(task, 'description') and 'Input data from previous tasks:' in task.description:
-                task.description = task.description.split('Input data from previous tasks:')[0].strip()
+            if hasattr(task, '_execution_context'):
+                task._execution_context = None
         
         while fallback_attempts < self.max_retries and not temp_current_task:
             fallback_attempts += 1
@@ -261,18 +278,18 @@ class Process:
                     
                     if not leads_to_task and not task_candidate.next_tasks:
                         continue  # Skip if no valid path exists
-                        
-                    if self.task_retry_counter.get(task_candidate.id, 0) < self.max_retries:
-                        # Atomic increment using thread lock to prevent race conditions
-                        with self._state_lock_init:
-                            current_count = self.task_retry_counter.get(task_candidate.id, 0)
+                    
+                    # Fix TOCTOU bug: Move the entire check-and-increment inside the lock
+                    with self._state_lock_init:
+                        current_count = self.task_retry_counter.get(task_candidate.id, 0)
+                        if current_count < self.max_retries:
                             self.task_retry_counter[task_candidate.id] = current_count + 1
-                        temp_current_task = task_candidate
-                        logging.debug(f"Fallback attempt {fallback_attempts}: Found 'not started' task: {temp_current_task.name}, retry count: {self.task_retry_counter[temp_current_task.id]}")
-                        return temp_current_task # Return the found task immediately
-                    else:
-                        logging.debug(f"Max retries ({self.max_retries}) reached for task {task_candidate.name} in fallback mode, marking as failed.")
-                        task_candidate.status = "failed"
+                            temp_current_task = task_candidate
+                            logging.debug(f"Fallback attempt {fallback_attempts}: Found 'not started' task: {temp_current_task.name}, retry count: {self.task_retry_counter[temp_current_task.id]}")
+                            return temp_current_task # Return the found task immediately
+                        else:
+                            logging.debug(f"Max retries ({self.max_retries}) reached for task {task_candidate.name} in fallback mode, marking as failed.")
+                            task_candidate.status = "failed"
             if not temp_current_task:
                 logging.debug(f"Fallback attempt {fallback_attempts}: No 'not started' task found within retry limit.")
         return None # Return None if no task found after all attempts
@@ -384,7 +401,7 @@ class Process:
         # Parse JSON and validate with Pydantic
         return self._parse_manager_instructions(response, ManagerInstructions)
 
-    def _check_all_tasks_completed(self) -> bool:
+    async def _check_all_tasks_completed(self) -> bool:
         """Check if all tasks are completed and handle workflow completion.
         
         Returns:
@@ -392,9 +409,27 @@ class Process:
         """
         if all(task.status == "completed" for task in self.tasks.values()):
             logging.info("All tasks are completed.")
-            self.workflow_finished = True
+            await self._set_workflow_finished(True)
             return True
         return False
+    
+    def _check_all_tasks_completed_sync(self) -> bool:
+        """Synchronous version of _check_all_tasks_completed.
+        
+        Returns:
+            bool: True if all tasks are completed and workflow should exit, False otherwise.
+        """
+        if all(task.status == "completed" for task in self.tasks.values()):
+            logging.info("All tasks are completed.")
+            # Use thread-safe setter for sync context
+            self._set_workflow_finished_sync(True)
+            return True
+        return False
+    
+    def _set_workflow_finished_sync(self, value: bool):
+        """Thread-safe setter for workflow_finished flag in sync context."""
+        with self._state_lock_init:
+            self.workflow_finished = value
 
     async def aworkflow(self) -> AsyncGenerator[str, None]:
         """Async version of workflow method"""
@@ -471,7 +506,7 @@ Tasks by type:
             """)
 
             # ADDED: Check if all tasks are completed and set workflow_finished flag
-            if self._check_all_tasks_completed():
+            if await self._check_all_tasks_completed():
                 break  # Exit immediately to prevent task reset
 
             task_id = current_task.id
@@ -486,13 +521,10 @@ Context tasks: {[t.name for t in current_task.context] if current_task.context e
 Description length: {len(current_task.description)}
             """)
 
-            # Build context and set description for this execution pass only
+            # Build context and store separately instead of mutating description
             context = self._build_task_context(current_task)
-            # Store original description if not already stored
-            if not hasattr(current_task, '_original_description'):
-                current_task._original_description = current_task.description
-            # Set description with context for execution; reset after yield to prevent accumulation
-            current_task.description = current_task._original_description + (context if context else "")
+            # Store context in dedicated field instead of concatenating to description
+            current_task._execution_context = context if context else ""
 
             # Skip execution for loop tasks, only process their subtasks
             if current_task.task_type == "loop":
@@ -563,7 +595,7 @@ Subtask: {st.name}
                             if next_task:
                                 next_task.status = "not started"  # Reset status to allow execution
                                 logging.debug(f"Routing to {next_task.name} based on decision: {decision_str}")
-                                self.workflow_finished = False
+                                await self._set_workflow_finished(False)
                                 current_task = next_task
                                 # Ensure the task is yielded for execution
                                 if current_task.id not in visited_tasks:
@@ -572,7 +604,7 @@ Subtask: {st.name}
                             else:
                                 # End workflow if no valid next task found
                                 logging.info(f"No valid next task found for decision: {decision_str}")
-                                self.workflow_finished = True
+                                await self._set_workflow_finished(True)
                                 current_task = None
                                 break
                 else:
@@ -609,7 +641,7 @@ Subtask: {st.name}
                     for t in self.tasks.values()
                 ):
                     logging.info(f"Task {current_task.name} has no next tasks, ending workflow")
-                    self.workflow_finished = True
+                    await self._set_workflow_finished(True)
                     current_task = None
                     break
 
@@ -672,7 +704,7 @@ Subtask: {st.name}
                         # Handle all forms of exit conditions
                         if not target_tasks or target_tasks == "exit" or (isinstance(target_tasks, list) and (not target_tasks or target_tasks[0] == "exit")):
                             logging.info(f"Workflow exit condition met on decision: {decision_str}")
-                            self.workflow_finished = True
+                            await self._set_workflow_finished(True)
                             current_task = None
                             break
                         else:
@@ -712,7 +744,7 @@ Subtask: {st.name}
                                 
                                 logging.debug(f"Routing to {next_task.name} based on decision: {decision_str}")
                                 # Don't mark workflow as finished when following condition path
-                                self.workflow_finished = False
+                                await self._set_workflow_finished(False)
 
             # If no condition-based routing, use next_tasks
             if not next_task and current_task and current_task.next_tasks:
@@ -726,7 +758,7 @@ Subtask: {st.name}
                         next_task.next_tasks and 
                         next_task.next_tasks[0] in self.tasks and 
                         next_task.name in self.tasks[next_task.next_tasks[0]].previous_tasks):
-                        self.workflow_finished = False
+                        await self._set_workflow_finished(False)
                     logging.debug(f"Following next_tasks to {next_task.name}")
 
             current_task = next_task
@@ -804,6 +836,11 @@ Workflow Finished: {self.workflow_finished} # ADDED: Workflow Finished Status
         completed_count = 0
         total_tasks = len(self.tasks) - 1
         logging.info(f"Need to complete {total_tasks} tasks (excluding manager task)")
+        
+        # Track invalid selection attempts and error context
+        invalid_selection_attempts = 0
+        MAX_INVALID_SELECTIONS = 3
+        error_context = ""
 
         while completed_count < total_tasks:
             tasks_summary = []
@@ -830,22 +867,44 @@ Provide a JSON with the structure:
    "agent_name": "<string>",
    "action": "<execute or stop>"
 }}
-"""
+""" + error_context
 
-            try:
-                logging.info("Requesting manager instructions...")
-                if manager_task.async_execution:
-                    parsed_instructions = await self._get_manager_instructions_with_fallback_async(
-                        manager_task, manager_prompt, ManagerInstructions
-                    )
-                else:
-                    parsed_instructions = self._get_manager_instructions_with_fallback(
-                        manager_task, manager_prompt, ManagerInstructions
-                    )
-                logging.info(f"Manager instructions: {parsed_instructions}")
-            except Exception as e:
-                display_error(f"Manager parse error: {e}")
-                logging.error(f"Manager parse error: {str(e)}", exc_info=True)
+            # Retry logic for manager instruction failures
+            MAX_MANAGER_RETRIES = 3
+            parsed_instructions = None
+            
+            for manager_attempt in range(MAX_MANAGER_RETRIES):
+                try:
+                    logging.info(f"Requesting manager instructions (attempt {manager_attempt + 1}/{MAX_MANAGER_RETRIES})...")
+                    if manager_task.async_execution:
+                        parsed_instructions = await self._get_manager_instructions_with_fallback_async(
+                            manager_task, manager_prompt, ManagerInstructions
+                        )
+                    else:
+                        # Offload sync call to thread to avoid blocking event loop
+                        parsed_instructions = await asyncio.to_thread(
+                            self._get_manager_instructions_with_fallback,
+                            manager_task, manager_prompt, ManagerInstructions
+                        )
+                    logging.info(f"Manager instructions: {parsed_instructions}")
+                    break  # Success, exit retry loop
+                except Exception as e:
+                    if manager_attempt < MAX_MANAGER_RETRIES - 1:
+                        # Not the last attempt, wait with exponential backoff
+                        delay = 2 ** manager_attempt
+                        logging.warning(
+                            f"Manager parse error (attempt {manager_attempt + 1}/{MAX_MANAGER_RETRIES}), "
+                            f"retrying in {delay}s: {e}"
+                        )
+                        # Always use async sleep in async context
+                        await asyncio.sleep(delay)
+                    else:
+                        # Final attempt failed
+                        display_error(f"Manager failed after {MAX_MANAGER_RETRIES} attempts: {e}")
+                        logging.error(f"Manager parse error: {str(e)}", exc_info=True)
+            
+            # If all retries failed, break out of the main loop
+            if parsed_instructions is None:
                 break
 
             selected_task_id = parsed_instructions.task_id
@@ -859,10 +918,29 @@ Provide a JSON with the structure:
                 break
 
             if selected_task_id not in self.tasks:
-                error_msg = f"Manager selected invalid task id {selected_task_id}"
-                display_error(error_msg)
-                logging.error(error_msg)
-                break
+                # Re-prompt the manager with valid task IDs instead of terminating
+                invalid_selection_attempts += 1
+                if invalid_selection_attempts > MAX_INVALID_SELECTIONS:
+                    logging.error(
+                        f"Manager produced {invalid_selection_attempts} invalid task selections; aborting."
+                    )
+                    break
+                
+                valid_task_ids = list(self.tasks.keys())
+                logging.warning(
+                    f"Manager selected invalid task_id={selected_task_id} "
+                    f"(attempt {invalid_selection_attempts}/{MAX_INVALID_SELECTIONS}); valid IDs: {valid_task_ids}"
+                )
+                # Set error context for next iteration (instead of appending to prompt that gets rebuilt)
+                error_context = (
+                    f"\n\n[ERROR] Your previous selection of task_id={selected_task_id} was invalid. "
+                    f"Valid task IDs are: {valid_task_ids}. Please select again from the valid options."
+                )
+                continue  # Re-prompt the manager instead of breaking
+            
+            # Reset on valid selection
+            invalid_selection_attempts = 0
+            error_context = ""
 
             original_agent = self.tasks[selected_task_id].agent.name if self.tasks[selected_task_id].agent else "None"
             for a in self.agents:
@@ -916,6 +994,7 @@ Provide a JSON with the structure:
             stacklevel=3
         )
         current_iter = 0  # Track how many times we've looped
+        workflow_start = time.monotonic()  # For timeout enforcement
         # Build workflow relationships first
         for task in self.tasks.values():
             if task.next_tasks:
@@ -1044,6 +1123,14 @@ Provide a JSON with the structure:
                 logging.info(f"Max iteration limit {self.max_iter} reached, ending workflow.")
                 break
 
+            # Enforce workflow timeout if set
+            if self.workflow_timeout is not None:
+                elapsed = time.monotonic() - workflow_start
+                if elapsed > self.workflow_timeout:
+                    logging.warning(f"Workflow timeout ({self.workflow_timeout}s) exceeded after {elapsed:.1f}s, ending workflow.")
+                    self.workflow_cancelled = True
+                    break
+
             # ADDED: Check workflow finished flag at the start of each cycle
             if self.workflow_finished:
                 logging.info("Workflow finished early as all tasks are completed.")
@@ -1071,7 +1158,7 @@ Tasks by type:
             """)
 
             # ADDED: Check if all tasks are completed and set workflow_finished flag
-            if self._check_all_tasks_completed():
+            if self._check_all_tasks_completed_sync():
                 break  # Exit immediately to prevent task reset
 
 
@@ -1160,13 +1247,10 @@ Context tasks: {[t.name for t in current_task.context] if current_task.context e
 Description length: {len(current_task.description)}
             """)
 
-            # Build context and set description for this execution pass only
+            # Build context and store separately instead of mutating description
             context = self._build_task_context(current_task)
-            # Store original description if not already stored
-            if not hasattr(current_task, '_original_description'):
-                current_task._original_description = current_task.description
-            # Set description with context for execution; reset after yield to prevent accumulation
-            current_task.description = current_task._original_description + (context if context else "")
+            # Store context in dedicated field instead of concatenating to description
+            current_task._execution_context = context if context else ""
 
             # Skip execution for loop tasks, only process their subtasks
             if current_task.task_type == "loop":
@@ -1237,7 +1321,7 @@ Subtask: {st.name}
                             if next_task:
                                 next_task.status = "not started"  # Reset status to allow execution
                                 logging.debug(f"Routing to {next_task.name} based on decision: {decision_str}")
-                                self.workflow_finished = False
+                                self._set_workflow_finished_sync(False)
                                 current_task = next_task
                                 # Ensure the task is yielded for execution
                                 if current_task.id not in visited_tasks:
@@ -1246,7 +1330,7 @@ Subtask: {st.name}
                             else:
                                 # End workflow if no valid next task found
                                 logging.info(f"No valid next task found for decision: {decision_str}")
-                                self.workflow_finished = True
+                                self._set_workflow_finished_sync(True)
                                 current_task = None
                                 break
                 else:
@@ -1283,7 +1367,7 @@ Subtask: {st.name}
                     for t in self.tasks.values()
                 ):
                     logging.info(f"Task {current_task.name} has no next tasks, ending workflow")
-                    self.workflow_finished = True
+                    self._set_workflow_finished_sync(True)
                     current_task = None
                     break
 
@@ -1344,7 +1428,7 @@ Subtask: {st.name}
                         # Handle all forms of exit conditions
                         if not target_tasks or target_tasks == "exit" or (isinstance(target_tasks, list) and (not target_tasks or target_tasks[0] == "exit")):
                             logging.info(f"Workflow exit condition met on decision: {decision_str}")
-                            self.workflow_finished = True
+                            self._set_workflow_finished_sync(True)
                             current_task = None
                             break
                         else:
@@ -1384,7 +1468,7 @@ Subtask: {st.name}
                                 
                                 logging.debug(f"Routing to {next_task.name} based on decision: {decision_str}")
                                 # Don't mark workflow as finished when following condition path
-                                self.workflow_finished = False
+                                self._set_workflow_finished_sync(False)
 
             # If no condition-based routing, use next_tasks
             if not next_task and current_task and current_task.next_tasks:
@@ -1398,7 +1482,7 @@ Subtask: {st.name}
                         next_task.next_tasks and 
                         next_task.next_tasks[0] in self.tasks and 
                         next_task.name in self.tasks[next_task.next_tasks[0]].previous_tasks):
-                        self.workflow_finished = False
+                        self._set_workflow_finished_sync(False)
                     logging.debug(f"Following next_tasks to {next_task.name}")
 
             current_task = next_task
@@ -1443,7 +1527,19 @@ Workflow Finished: {self.workflow_finished} # ADDED: Workflow Finished Status
     def sequential(self):
         """Synchronous version of sequential method"""
         for task_id in self.tasks:
-            if self.tasks[task_id].status != "completed":
+            task = self.tasks[task_id]
+            if task.status == "failed":
+                continue  # Skip failed tasks
+            # Check if any context dependency failed
+            if hasattr(task, 'context') and task.context:
+                deps_failed = any(
+                    self.tasks[dep.id].status == "failed"
+                    for dep in task.context if hasattr(dep, 'id') and dep.id in self.tasks
+                )
+                if deps_failed:
+                    task.status = "failed"
+                    continue
+            if task.status != "completed":
                 yield task_id
 
     def hierarchical(self):
@@ -1476,6 +1572,11 @@ Workflow Finished: {self.workflow_finished} # ADDED: Workflow Finished Status
         completed_count = 0
         total_tasks = len(self.tasks) - 1
         logging.info(f"Need to complete {total_tasks} tasks (excluding manager task)")
+        
+        # Track invalid selection attempts and error context
+        invalid_selection_attempts = 0
+        MAX_INVALID_SELECTIONS = 3
+        error_context = ""
 
         while completed_count < total_tasks:
             tasks_summary = []
@@ -1502,7 +1603,7 @@ Provide a JSON with the structure:
    "agent_name": "<string>",
    "action": "<execute or stop>"
 }}
-"""
+""" + error_context
 
             try:
                 logging.info("Requesting manager instructions...")
@@ -1526,10 +1627,29 @@ Provide a JSON with the structure:
                 break
 
             if selected_task_id not in self.tasks:
-                error_msg = f"Manager selected invalid task id {selected_task_id}"
-                display_error(error_msg)
-                logging.error(error_msg)
-                break
+                # Re-prompt the manager with valid task IDs instead of terminating
+                invalid_selection_attempts += 1
+                if invalid_selection_attempts > MAX_INVALID_SELECTIONS:
+                    logging.error(
+                        f"Manager produced {invalid_selection_attempts} invalid task selections; aborting."
+                    )
+                    break
+                
+                valid_task_ids = list(self.tasks.keys())
+                logging.warning(
+                    f"Manager selected invalid task_id={selected_task_id} "
+                    f"(attempt {invalid_selection_attempts}/{MAX_INVALID_SELECTIONS}); valid IDs: {valid_task_ids}"
+                )
+                # Set error context for next iteration (instead of appending to prompt that gets rebuilt)
+                error_context = (
+                    f"\n\n[ERROR] Your previous selection of task_id={selected_task_id} was invalid. "
+                    f"Valid task IDs are: {valid_task_ids}. Please select again from the valid options."
+                )
+                continue  # Re-prompt the manager instead of breaking
+            
+            # Reset on valid selection
+            invalid_selection_attempts = 0
+            error_context = ""
 
             original_agent = self.tasks[selected_task_id].agent.name if self.tasks[selected_task_id].agent else "None"
             for a in self.agents:

src/praisonai-agents/praisonaiagents/push/__init__.py

@@ -0,0 +1,39 @@
+"""Push client public surface.
+
+Concrete implementations live in the `praisonai` wrapper. This module
+re-exports them lazily so user code can keep importing from
+``praisonaiagents.push``.
+"""
+import importlib
+
+from .models import ChannelMessage
+from .protocols import PushTransportProtocol
+
+_lazy_cache: dict = {}
+
+_WRAPPER_MODULE = "praisonai.push"
+_WRAPPER_NAMES = frozenset({"PushClient", "WebSocketTransport", "PollingTransport"})
+
+def __getattr__(name: str):
+    if name in _lazy_cache:
+        return _lazy_cache[name]
+    if name in _WRAPPER_NAMES:
+        try:
+            obj = getattr(importlib.import_module(_WRAPPER_MODULE), name)
+        except ImportError as e:
+            raise ImportError(
+                f"{name} requires the praisonai wrapper package. "
+                "Install with: pip install praisonai"
+            ) from e
+        _lazy_cache[name] = obj
+        return obj
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+
+
+__all__ = [
+    "ChannelMessage",
+    "PushTransportProtocol",
+    "PushClient",
+    "WebSocketTransport",
+    "PollingTransport",
+]

src/praisonai-agents/praisonaiagents/push/client.py

@@ -0,0 +1,10 @@
+"""Backward compatibility shim for praisonaiagents.push.client.
+
+This module maintains backward compatibility for existing imports like:
+    from praisonaiagents.push.client import PushClient
+
+The actual implementation has been moved to the praisonai wrapper package.
+"""
+from . import PushClient
+
+__all__ = ["PushClient"]

src/praisonai-agents/praisonaiagents/push/models.py

@@ -0,0 +1,50 @@
+"""
+Push notification models for PraisonAI Agents.
+"""
+
+from __future__ import annotations
+
+import time
+import uuid
+from dataclasses import dataclass, field
+from typing import Any, Dict, Optional
+
+
+@dataclass
+class ChannelMessage:
+    """A user-friendly wrapper around a push channel message.
+
+    Attributes:
+        channel: Channel name the message was received on
+        data: Message payload
+        event_id: Unique event identifier
+        timestamp: When the event was created
+        source: Source client ID (who sent it)
+    """
+
+    channel: str
+    data: Dict[str, Any] = field(default_factory=dict)
+    event_id: str = field(default_factory=lambda: str(uuid.uuid4()))
+    timestamp: float = field(default_factory=time.time)
+    source: Optional[str] = None
+
+    @classmethod
+    def from_event_dict(cls, event_dict: Dict[str, Any]) -> "ChannelMessage":
+        """Create from a raw gateway event dictionary."""
+        return cls(
+            channel=event_dict.get("channel", ""),
+            data=event_dict.get("data", {}),
+            event_id=event_dict.get("event_id", str(uuid.uuid4())),
+            timestamp=event_dict.get("timestamp", time.time()),
+            source=event_dict.get("source"),
+        )
+
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary."""
+        return {
+            "channel": self.channel,
+            "data": self.data,
+            "event_id": self.event_id,
+            "timestamp": self.timestamp,
+            "source": self.source,
+        }

src/praisonai-agents/praisonaiagents/push/protocols.py

@@ -0,0 +1,29 @@
+"""Push transport protocol — implementations live in the praisonai wrapper."""
+from __future__ import annotations
+from typing import Any, Dict, Optional, Protocol, runtime_checkable
+
+
+@runtime_checkable
+class PushTransportProtocol(Protocol):
+    """Protocol for push client transports."""
+
+    @property
+    def is_connected(self) -> bool:
+        """Whether the transport is currently connected."""
+        ...
+
+    async def connect(self) -> None:
+        """Establish connection."""
+        ...
+
+    async def disconnect(self) -> None:
+        """Close connection."""
+        ...
+
+    async def send(self, data: Dict[str, Any]) -> None:
+        """Send a JSON message."""
+        ...
+
+    async def receive(self) -> Dict[str, Any]:
+        """Receive a JSON message. Blocks until a message arrives."""
+        ...

src/praisonai-agents/praisonaiagents/push/transports.py

@@ -0,0 +1,10 @@
+"""Backward compatibility shim for praisonaiagents.push.transports.
+
+This module maintains backward compatibility for existing imports like:
+    from praisonaiagents.push.transports import WebSocketTransport, PollingTransport
+
+The actual implementations have been moved to the praisonai wrapper package.
+"""
+from . import WebSocketTransport, PollingTransport
+
+__all__ = ["WebSocketTransport", "PollingTransport"]

src/praisonai-agents/praisonaiagents/rag/__init__.py

@@ -67,21 +67,28 @@ _LAZY_IMPORTS = {
     "deduplicate_chunks": ("praisonaiagents.rag.context", "deduplicate_chunks"),
 }
 
+import threading
+
 _cache = {}
+_cache_lock = threading.Lock()
 
 
 def __getattr__(name: str):
-    """Lazy load RAG components."""
+    """Lazy load RAG components with thread safety."""
     if name in _cache:
         return _cache[name]
     
     if name in _LAZY_IMPORTS:
-        module_path, attr_name = _LAZY_IMPORTS[name]
-        import importlib
-        module = importlib.import_module(module_path)
-        value = getattr(module, attr_name)
-        _cache[name] = value
-        return value
+        with _cache_lock:
+            # Double-check after acquiring lock
+            if name in _cache:
+                return _cache[name]
+            module_path, attr_name = _LAZY_IMPORTS[name]
+            import importlib
+            module = importlib.import_module(module_path)
+            value = getattr(module, attr_name)
+            _cache[name] = value
+            return value
     
     raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
 

src/praisonai-agents/praisonaiagents/session.py

@@ -617,15 +617,17 @@ class Session:
         if self.is_remote:
             return  # No cleanup needed for remote sessions
         
-        # Clear memory
+        # Properly cleanup memory
         if self._memory:
             try:
-                # Clear short-term and long-term memory for this session
-                # Note: This is a basic implementation - specific memory backends
-                # might need more sophisticated cleanup
+                # Call close_connections to properly cleanup memory adapters
+                if hasattr(self._memory, 'close_connections'):
+                    self._memory.close_connections()
+            except Exception as e:
+                import logging
+                logging.warning(f"Memory cleanup failed: {e}")
+            finally:
                 self._memory = None
-            except Exception:
-                pass  # Ignore cleanup errors
         
         # Clear knowledge
         if self._knowledge:

src/praisonai-agents/praisonaiagents/session/__init__.py

@@ -90,7 +90,43 @@ def __getattr__(name: str):
         from .hierarchy import ExtendedSessionData
         _module_cache[name] = ExtendedSessionData
         return ExtendedSessionData
-    
+
+    # W1 — Identity resolver
+    if name == "IdentityResolverProtocol":
+        from .identity import IdentityResolverProtocol
+        _module_cache[name] = IdentityResolverProtocol
+        return IdentityResolverProtocol
+    if name == "IdentityLink":
+        from .identity import IdentityLink
+        _module_cache[name] = IdentityLink
+        return IdentityLink
+    if name == "InMemoryIdentityResolver":
+        from .identity import InMemoryIdentityResolver
+        _module_cache[name] = InMemoryIdentityResolver
+        return InMemoryIdentityResolver
+    if name == "FileIdentityResolver":
+        from .identity import FileIdentityResolver
+        _module_cache[name] = FileIdentityResolver
+        return FileIdentityResolver
+
+    # W1 — Task-local session context
+    if name == "SessionContext":
+        from .context import SessionContext
+        _module_cache[name] = SessionContext
+        return SessionContext
+    if name == "set_session_context":
+        from .context import set_session_context
+        _module_cache[name] = set_session_context
+        return set_session_context
+    if name == "get_session_context":
+        from .context import get_session_context
+        _module_cache[name] = get_session_context
+        return get_session_context
+    if name == "clear_session_context":
+        from .context import clear_session_context
+        _module_cache[name] = clear_session_context
+        return clear_session_context
+
     raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
 
 
@@ -105,4 +141,13 @@ __all__ = [
     "get_hierarchical_session_store",
     "SessionSnapshot",
     "ExtendedSessionData",
+    # W1
+    "IdentityResolverProtocol",
+    "IdentityLink",
+    "InMemoryIdentityResolver",
+    "FileIdentityResolver",
+    "SessionContext",
+    "set_session_context",
+    "get_session_context",
+    "clear_session_context",
 ]

src/praisonai-agents/praisonaiagents/session/context.py

@@ -0,0 +1,112 @@
+"""Task-local session context for concurrent message handlers.
+
+W1 — Replaces process-global ``os.environ`` based session metadata
+with ``contextvars.ContextVar`` so that two messages handled
+concurrently never overwrite each other's platform / chat / user IDs.
+
+Each ``set_session_context`` call returns a token; pass it to
+``clear_session_context`` in a ``finally`` block to restore the
+previous state.
+
+Usage in a bot handler::
+
+    from praisonaiagents.session.context import (
+        set_session_context,
+        clear_session_context,
+        get_session_context,
+    )
+
+    token = set_session_context(
+        platform="telegram",
+        chat_id="100",
+        user_id="alice",
+        unified_user_id="alice-global",
+    )
+    try:
+        # ... agent handles message; tools can call get_session_context()
+        ctx = get_session_context()
+    finally:
+        clear_session_context(token)
+"""
+
+from __future__ import annotations
+
+from contextvars import ContextVar, Token
+from dataclasses import asdict, dataclass
+
+
+@dataclass(frozen=True)
+class SessionContext:
+    """Snapshot of the current message-handling context."""
+
+    platform: str = ""
+    chat_id: str = ""
+    chat_name: str = ""
+    thread_id: str = ""
+    user_id: str = ""
+    user_name: str = ""
+    unified_user_id: str = ""
+
+    def to_dict(self) -> dict:
+        return asdict(self)
+
+    @classmethod
+    def from_dict(cls, d: dict) -> "SessionContext":
+        return cls(
+            platform=d.get("platform", ""),
+            chat_id=d.get("chat_id", ""),
+            chat_name=d.get("chat_name", ""),
+            thread_id=d.get("thread_id", ""),
+            user_id=d.get("user_id", ""),
+            user_name=d.get("user_name", ""),
+            unified_user_id=d.get("unified_user_id", ""),
+        )
+
+
+_CTX: ContextVar[SessionContext] = ContextVar(
+    "praisonai_session_context", default=SessionContext()
+)
+
+
+def set_session_context(
+    platform: str = "",
+    chat_id: str = "",
+    chat_name: str = "",
+    thread_id: str = "",
+    user_id: str = "",
+    user_name: str = "",
+    unified_user_id: str = "",
+) -> Token:
+    """Set the task-local session context. Returns a token for ``clear``."""
+    ctx = SessionContext(
+        platform=platform,
+        chat_id=chat_id,
+        chat_name=chat_name,
+        thread_id=thread_id,
+        user_id=user_id,
+        user_name=user_name,
+        unified_user_id=unified_user_id,
+    )
+    return _CTX.set(ctx)
+
+
+def get_session_context() -> SessionContext:
+    """Return the current task's session context (empty if unset)."""
+    return _CTX.get()
+
+
+def clear_session_context(token: Token) -> None:
+    """Restore the previous session context using the token from ``set``."""
+    try:
+        _CTX.reset(token)
+    except (LookupError, ValueError):
+        # Token from a different context; reset to empty
+        _CTX.set(SessionContext())
+
+
+__all__ = [
+    "SessionContext",
+    "set_session_context",
+    "get_session_context",
+    "clear_session_context",
+]

src/praisonai-agents/praisonaiagents/session/hierarchy.py

@@ -481,6 +481,74 @@ class HierarchicalSessionStore(DefaultSessionStore):
         session.title = title
         return self._save_extended_session(session)
     
+    async def auto_title(self, session_id: str) -> bool:
+        """Generate and set title automatically from first exchange.
+        
+        Args:
+            session_id: Session to generate title for
+            
+        Returns:
+            True if title was generated and set, False otherwise
+        """
+        import asyncio
+        
+        # Load session in thread to avoid blocking event loop
+        session = await asyncio.to_thread(self._load_extended_session, session_id)
+        
+        # Skip if already has a title
+        if session.title and session.title.strip():
+            return False
+            
+        # Need at least one user and one assistant message
+        messages = session.messages
+        if not messages or len(messages) < 2:
+            return False
+            
+        # Find first user message and first assistant response
+        user_msg = None
+        assistant_msg = None
+        
+        for msg in messages:
+            # Handle both SessionMessage dataclass and dict formats
+            if hasattr(msg, 'role'):
+                role = msg.role
+                content = msg.content
+            else:
+                role = msg.get("role")
+                content = msg.get("content", "")
+                
+            if role == "user" and not user_msg:
+                if isinstance(content, str) and content.strip():
+                    user_msg = content
+            elif role == "assistant" and not assistant_msg and user_msg:
+                if isinstance(content, str) and content.strip():
+                    assistant_msg = content
+                    break
+        
+        if not user_msg or not assistant_msg:
+            return False
+            
+        try:
+            # Generate title using title module
+            from .title import generate_title_async
+            title = await generate_title_async(user_msg, assistant_msg)
+            
+            if title and title.strip():
+                # Reload session to avoid overwriting concurrent updates
+                fresh_session = await asyncio.to_thread(self._load_extended_session, session_id)
+                # Only set title if it's still empty
+                if not fresh_session.title or not fresh_session.title.strip():
+                    fresh_session.title = title.strip()
+                    return await asyncio.to_thread(self._save_extended_session, fresh_session)
+                
+        except Exception as e:
+            # Title generation failed - log with context instead of silent failure
+            import logging
+            logger = logging.getLogger(__name__)
+            logger.debug("Auto title generation failed for session %s: %s", session_id, str(e))
+            
+        return False
+    
     def get_extended_session(self, session_id: str) -> ExtendedSessionData:
         """Get extended session data."""
         return self._load_extended_session(session_id)

src/praisonai-agents/praisonaiagents/session/identity.py

@@ -0,0 +1,231 @@
+"""Cross-platform identity resolution for agents.
+
+W1 — Optional, opt-in mapping from ``(platform, platform_user_id)``
+to a single ``unified_user_id`` so the same human is recognised across
+Telegram, Discord, Slack, etc.
+
+By default (no links registered) ``resolve()`` returns the
+platform-prefixed string unchanged — no surprises, no privacy leak.
+
+Linking is performed explicitly, typically after a DM-based pairing
+flow has cryptographically verified ownership of both accounts.
+
+Usage::
+
+    from praisonaiagents.session.identity import InMemoryIdentityResolver
+
+    resolver = InMemoryIdentityResolver()
+    resolver.link("telegram", "12345", "alice")
+    resolver.link("discord", "snowflake-1", "alice")
+    assert resolver.resolve("telegram", "12345") == "alice"
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import tempfile
+import threading
+from dataclasses import dataclass
+from pathlib import Path
+from typing import List, Optional, Protocol, runtime_checkable
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass(frozen=True)
+class IdentityLink:
+    """Immutable mapping from a platform-scoped user ID to a unified ID."""
+
+    platform: str
+    platform_user_id: str
+    unified_user_id: str
+
+
+@runtime_checkable
+class IdentityResolverProtocol(Protocol):
+    """Maps platform-scoped user IDs to a unified identity.
+
+    Implementations MUST be thread-safe. Implementations SHOULD return
+    ``f"{platform}:{platform_user_id}"`` for unlinked users so callers
+    never need to special-case the unlinked path.
+    """
+
+    def resolve(self, platform: str, platform_user_id: str) -> str:
+        """Return ``unified_user_id`` for the given platform user.
+
+        If no link is registered, return ``f"{platform}:{platform_user_id}"``.
+        """
+        ...
+
+    def link(
+        self, platform: str, platform_user_id: str, unified_user_id: str
+    ) -> None:
+        """Register a mapping. Overwrites any existing mapping for the same
+        ``(platform, platform_user_id)`` pair.
+        """
+        ...
+
+    def unlink(self, platform: str, platform_user_id: str) -> None:
+        """Remove a mapping. No-op when no link exists."""
+        ...
+
+    def links_for(self, unified_user_id: str) -> List[IdentityLink]:
+        """Return all links pointing to the given unified ID."""
+        ...
+
+
+class InMemoryIdentityResolver:
+    """Thread-safe in-memory ``IdentityResolverProtocol`` implementation.
+
+    Suitable for tests and single-process deployments. Production
+    deployments with multiple processes should use a persistent
+    implementation backed by SQLite, Redis, or a database.
+    """
+
+    def __init__(self) -> None:
+        self._links: dict[tuple[str, str], str] = {}
+        self._lock = threading.RLock()
+
+    def resolve(self, platform: str, platform_user_id: str) -> str:
+        with self._lock:
+            return self._links.get(
+                (platform, platform_user_id),
+                f"{platform}:{platform_user_id}",
+            )
+
+    def link(
+        self, platform: str, platform_user_id: str, unified_user_id: str
+    ) -> None:
+        with self._lock:
+            self._links[(platform, platform_user_id)] = unified_user_id
+
+    def unlink(self, platform: str, platform_user_id: str) -> None:
+        with self._lock:
+            self._links.pop((platform, platform_user_id), None)
+
+    def links_for(self, unified_user_id: str) -> List[IdentityLink]:
+        with self._lock:
+            return [
+                IdentityLink(p, u, unified_user_id)
+                for (p, u), uid in self._links.items()
+                if uid == unified_user_id
+            ]
+
+
+_DEFAULT_IDENTITY_PATH = Path(
+    os.environ.get(
+        "PRAISONAI_IDENTITY_PATH",
+        os.path.expanduser("~/.praisonai/identity.json"),
+    )
+)
+
+
+class FileIdentityResolver(InMemoryIdentityResolver):
+    """JSON-file-backed ``IdentityResolverProtocol`` implementation.
+
+    Loads links from disk on construction; persists on every mutation
+    via atomic temp-file + ``os.replace``. Thread-safe.
+
+    Default path: ``$PRAISONAI_IDENTITY_PATH`` or ``~/.praisonai/identity.json``.
+
+    Storage format::
+
+        {
+          "links": {
+            "telegram::12345": "alice",
+            "discord::snowflake-1": "alice"
+          }
+        }
+    """
+
+    def __init__(self, path: Optional[Path | str] = None) -> None:
+        super().__init__()
+        self._path: Path = Path(path) if path else _DEFAULT_IDENTITY_PATH
+        # Serialises concurrent disk writes so that two simultaneous
+        # link()/unlink() calls cannot interleave their os.replace() and
+        # lose the second writer's data.
+        self._flush_lock = threading.Lock()
+        self._load()
+
+    @staticmethod
+    def _encode_key(platform: str, platform_user_id: str) -> str:
+        return f"{platform}::{platform_user_id}"
+
+    @staticmethod
+    def _decode_key(encoded: str) -> tuple[str, str]:
+        platform, _, user = encoded.partition("::")
+        return platform, user
+
+    def _load(self) -> None:
+        try:
+            if not self._path.exists():
+                return
+            data = json.loads(self._path.read_text(encoding="utf-8"))
+            links = data.get("links", {}) if isinstance(data, dict) else {}
+            with self._lock:
+                self._links = {
+                    self._decode_key(k): v
+                    for k, v in links.items()
+                    if isinstance(v, str)
+                }
+        except (OSError, json.JSONDecodeError) as e:
+            logger.warning("FileIdentityResolver: failed to load %s: %s", self._path, e)
+
+    def _flush(self) -> None:
+        try:
+            self._path.parent.mkdir(parents=True, exist_ok=True)
+            # Hold _flush_lock for the entire write (snapshot + disk I/O) so
+            # that concurrent calls always write the latest in-memory state and
+            # the last os.replace() wins rather than an older snapshot.
+            with self._flush_lock:
+                with self._lock:
+                    payload = {
+                        "links": {
+                            self._encode_key(p, u): uid
+                            for (p, u), uid in self._links.items()
+                        }
+                    }
+                    # Keep the _lock held across the entire write operation to prevent
+                    # TOCTOU race where a concurrent link() could modify _links between
+                    # payload snapshot and os.replace() completion.
+                    fd, tmp = tempfile.mkstemp(
+                        dir=str(self._path.parent), prefix=".identity-", suffix=".tmp"
+                    )
+                    try:
+                        with os.fdopen(fd, "w", encoding="utf-8") as f:
+                            json.dump(payload, f, indent=2, ensure_ascii=False)
+                            f.flush()
+                            os.fsync(f.fileno())
+                        os.replace(tmp, self._path)
+                        try:
+                            os.chmod(self._path, 0o600)
+                        except OSError:
+                            pass
+                    except BaseException:
+                        try:
+                            os.unlink(tmp)
+                        except OSError:
+                            pass
+                        raise
+        except OSError as e:
+            logger.warning("FileIdentityResolver: failed to flush %s: %s", self._path, e)
+
+    def link(
+        self, platform: str, platform_user_id: str, unified_user_id: str
+    ) -> None:
+        super().link(platform, platform_user_id, unified_user_id)
+        self._flush()
+
+    def unlink(self, platform: str, platform_user_id: str) -> None:
+        super().unlink(platform, platform_user_id)
+        self._flush()
+
+
+__all__ = [
+    "IdentityLink",
+    "IdentityResolverProtocol",
+    "InMemoryIdentityResolver",
+    "FileIdentityResolver",
+]

src/praisonai-agents/praisonaiagents/session/store.py

@@ -5,7 +5,6 @@ JSON-based session persistence with file locking and atomic writes.
 Zero dependencies beyond stdlib.
 """
 
-import fcntl
 import json
 import logging
 from praisonaiagents._logging import get_logger
@@ -14,6 +13,13 @@ import sys
 import tempfile
 import threading
 import time
+
+# fcntl is Unix-only; on Windows, use msvcrt for file locking
+if sys.platform != 'win32':
+    import fcntl
+    _HAS_FCNTL = True
+else:
+    _HAS_FCNTL = False
 from dataclasses import dataclass, field
 from datetime import datetime, timezone
 from typing import Any, Dict, List, Optional
@@ -22,6 +28,9 @@ from ..paths import get_sessions_dir
 
 logger = get_logger(__name__)
 
+# Module-level sentinel to track if we've warned about degraded locking
+_WARNED_NO_FCNTL = False
+
 # Default session directory (uses centralized paths - DRY)
 DEFAULT_SESSION_DIR = str(get_sessions_dir())
 
@@ -150,7 +159,17 @@ class FileLock:
                     msvcrt.locking(self._lock_file.fileno(), msvcrt.LK_NBLCK, 1)
                 else:
                     # Unix locking
-                    fcntl.flock(self._lock_file.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
+                    if _HAS_FCNTL:
+                        fcntl.flock(self._lock_file.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
+                    else:
+                        # Warn once about degraded locking (should never happen with current platform detection)
+                        global _WARNED_NO_FCNTL
+                        if not _WARNED_NO_FCNTL:
+                            logger.warning(
+                                "File locking unavailable on this platform (no fcntl/msvcrt); "
+                                "concurrent writers may corrupt session files."
+                            )
+                            _WARNED_NO_FCNTL = True
                 return True
             except (IOError, OSError, BlockingIOError):
                 if self._lock_file:
@@ -171,7 +190,9 @@ class FileLock:
                     import msvcrt
                     msvcrt.locking(self._lock_file.fileno(), msvcrt.LK_UNLCK, 1)
                 else:
-                    fcntl.flock(self._lock_file.fileno(), fcntl.LOCK_UN)
+                    if _HAS_FCNTL:
+                        fcntl.flock(self._lock_file.fileno(), fcntl.LOCK_UN)
+                    # Note: No warning needed in release() as it mirrors acquire() logic
             except (IOError, OSError):
                 pass
             finally:

src/praisonai-agents/praisonaiagents/session/title.py

@@ -0,0 +1,169 @@
+"""
+Session Title Auto-Generation - Generate descriptive titles for chat sessions.
+
+Automatically creates concise, meaningful titles based on the first user-assistant
+exchange in a conversation. Uses a lightweight model for fast generation.
+"""
+
+import asyncio
+from typing import Optional
+
+__all__ = ["generate_title", "generate_title_async"]
+
+
+def generate_title(
+    user_msg: str, 
+    assistant_msg: str,
+    llm_model: str = "gpt-4o-mini",
+    timeout: float = 10.0,
+    max_length: int = 60
+) -> str:
+    """Generate a session title from the first exchange.
+    
+    Args:
+        user_msg: First user message
+        assistant_msg: First assistant response  
+        llm_model: Model to use for title generation (default: fast, cheap model)
+        timeout: Timeout in seconds for generation
+        max_length: Maximum title length in characters
+        
+    Returns:
+        Generated title string, or fallback based on user message
+        
+    Examples:
+        >>> title = generate_title(
+        ...     "Help me debug this Python code",
+        ...     "I'd be happy to help debug your Python code..."
+        ... )
+        >>> # Returns something like: "Python Code Debugging Help"
+    """
+    # Fallback title from user message if generation fails
+    fallback_title = _create_fallback_title(user_msg, max_length)
+    
+    try:
+        # Check if we're already in an event loop
+        asyncio.get_running_loop()
+    except RuntimeError:
+        # No running loop - safe to use asyncio.run
+        try:
+            return asyncio.run(generate_title_async(
+                user_msg, assistant_msg, llm_model, timeout, max_length
+            ))
+        except Exception:
+            import logging
+            logger = logging.getLogger(__name__)
+            logger.debug("Title generation failed in sync context", exc_info=True)
+            return fallback_title
+    else:
+        # Already in event loop - cannot use asyncio.run
+        import logging
+        logger = logging.getLogger(__name__)
+        logger.warning(
+            "generate_title() called from running event loop; "
+            "use generate_title_async() instead. Returning fallback title."
+        )
+        return fallback_title
+
+
+async def generate_title_async(
+    user_msg: str,
+    assistant_msg: str, 
+    llm_model: str = "gpt-4o-mini",
+    timeout: float = 10.0,
+    max_length: int = 60
+) -> str:
+    """Async version of generate_title.
+    
+    Args:
+        user_msg: First user message
+        assistant_msg: First assistant response
+        llm_model: Model to use for title generation
+        timeout: Timeout in seconds for generation
+        max_length: Maximum title length in characters
+        
+    Returns:
+        Generated title string, or fallback based on user message
+    """
+    fallback_title = _create_fallback_title(user_msg, max_length)
+    
+    try:
+        # Lazy import to avoid circular dependencies
+        from ..llm import LLM
+        
+        # Create prompt for title generation
+        prompt = f"""Generate a concise, descriptive title (3-8 words) for this conversation.
+
+USER: {user_msg[:400]}
+ASSISTANT: {assistant_msg[:400]}
+
+Return ONLY the title text, no quotes, no explanation."""
+        
+        # Use lightweight model for fast, cheap generation
+        llm = LLM(model=llm_model)
+        
+        # Generate with timeout
+        try:
+            response = await asyncio.wait_for(
+                llm.aget_response(prompt=prompt),
+                timeout=timeout
+            )
+            
+            if response and isinstance(response, str):
+                # Clean up the response
+                title = response.strip().strip('"').strip("'")
+                title = title.replace('\n', ' ').replace('\r', '')
+                
+                # Truncate if too long
+                if len(title) > max_length:
+                    title = title[:max_length-3] + "..."
+                
+                # Return if non-empty
+                if title and len(title.strip()) > 0:
+                    return title
+                    
+        except asyncio.TimeoutError:
+            # Title generation timed out
+            pass
+        except Exception:
+            # Any other LLM error
+            pass
+            
+    except ImportError:
+        # LLM module not available
+        pass
+    except Exception:
+        # Unexpected error
+        pass
+    
+    return fallback_title
+
+
+def _create_fallback_title(user_msg: str, max_length: int) -> str:
+    """Create a fallback title from the user message.
+    
+    Args:
+        user_msg: User message to base title on
+        max_length: Maximum length for the title
+        
+    Returns:
+        Simple title based on user message
+    """
+    if not user_msg or not user_msg.strip():
+        return "Chat Session"
+    
+    # Clean up the message
+    clean_msg = user_msg.strip()
+    
+    # Remove common question words and make it more title-like
+    clean_msg = clean_msg.replace('?', '')
+    clean_msg = clean_msg.replace('!', '')
+    
+    # Take first sentence or reasonable chunk
+    if '.' in clean_msg:
+        clean_msg = clean_msg.split('.')[0]
+    
+    # Truncate if too long
+    if len(clean_msg) > max_length:
+        clean_msg = clean_msg[:max_length-3] + "..."
+    
+    return clean_msg if clean_msg else "Chat Session"