fix: update broken Full Flow docs link in praisonai package README

- Fix duplicate broken link in src/praisonai/README.md
- Change from /docs/concepts/flow (404) to /docs/concepts/agentflow (200)
- Ensures PyPI package readers get working documentation link
- Addresses issue identified by Qodo code review

Co-authored-by: Mervin Praison <[email protected]>

.env.example

@@ -0,0 +1,51 @@
+# 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: Tool Security ===
+# ALLOWED_TOOLS=search,send_message       # Whitelist specific tools to prevent name collisions
+                                        # Useful in multi-environment setups with overlapping tool names
+                                        # Leave unset to use all tools (with collision warnings)
+                                        # Do NOT set to empty string: that's treated as configuration error
+                                        # In CI, unknown tool names in the whitelist cause strict startup failure
+# HERMES_ONLY_TOOLS=search,send_message  # (Deprecated) Backward compatibility alias for ALLOWED_TOOLS
+
+# === 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
 

.github/workflows/integration-cross-repo.yml

@@ -0,0 +1,158 @@
+name: Cross-Repo Integration Tests
+
+on:
+  push:
+    branches: [ main ]
+    paths:
+      - 'src/praisonai/praisonai/integration/**'
+      - 'src/praisonai/tests/integration/test_aiui_*'
+      - 'src/praisonai/pyproject.toml'
+  pull_request:
+    branches: [ main ]
+    paths:
+      - 'src/praisonai/praisonai/integration/**'
+      - 'src/praisonai/tests/integration/test_aiui_*'
+      - 'src/praisonai/pyproject.toml'
+  workflow_dispatch:
+
+jobs:
+  cross-repo-integration:
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    
+    steps:
+    - name: Checkout PraisonAI
+      uses: actions/checkout@v4
+      with:
+        path: PraisonAI
+        ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+        
+    - name: Checkout PraisonAIUI
+      uses: actions/checkout@v4
+      with:
+        repository: MervinPraison/PraisonAIUI
+        path: PraisonAIUI
+        ref: main
+        
+    - name: Set up Python
+      uses: actions/setup-python@v4
+      with:
+        python-version: '3.11'
+        
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        # Install PraisonAI packages in development mode
+        cd PraisonAI/src/praisonai-agents
+        pip install -e .
+        cd ../praisonai
+        pip install -e ".[ui,dev]"
+        # Install PraisonAIUI in development mode
+        cd ../../../PraisonAIUI
+        pip install -e .
+        
+    - name: Verify versions
+      run: |
+        python -c "import praisonaiagents; print(f'praisonaiagents: {praisonaiagents.__version__}')"
+        python -c "import praisonai; print(f'praisonai: {praisonai.__version__}')"
+        python -c "import praisonaiui; print(f'praisonaiui: {praisonaiui.__version__}')"
+        
+    - name: Run PraisonAI integration tests
+      run: |
+        cd PraisonAI/src/praisonai
+        python -m pytest tests/integration/test_aiui_* -v --timeout=300
+        
+    - name: Run PraisonAIUI integration tests
+      run: |
+        cd PraisonAIUI
+        if [ -f tests/integration/test_agentic_roundtrip.py ]; then
+          python -m pytest tests/integration/test_agentic_roundtrip.py -v --timeout=300
+        fi
+        if [ -f tests/test_feature_sdk_backends.py ]; then
+          python -m pytest tests/test_feature_sdk_backends.py -v --timeout=300
+        fi
+        
+    - name: Test Pattern C CLI
+      run: |
+        cd PraisonAI/src/praisonai
+        # Test that ui-gateway command exists
+        python -m praisonai serve --help | grep ui-gateway
+        
+    - name: Test public API exports
+      run: |
+        cd PraisonAI/src/praisonai
+        python -c "from praisonai import run_integrated_gateway; print('✓ run_integrated_gateway imported')"
+        python -c "from praisonai import configure_host; print('✓ configure_host imported')"
+        
+    - name: Integration smoke test
+      env:
+        PRAISONAI_TEST_MODE: "1"
+        PYTHONPATH: ${{ github.workspace }}/PraisonAI/src/praisonai:${{ github.workspace }}/PraisonAI/src/praisonai-agents:${{ github.workspace }}/PraisonAIUI
+      run: |
+        cd PraisonAI/src/praisonai
+        python -c "
+        import sys
+        sys.path.insert(0, '.')
+        from praisonai.integration.host_app import configure_host
+        from praisonai import run_integrated_gateway
+        print('✓ Integration imports successful')
+        
+        # Test configure_host with new parameters
+        try:
+            configure_host(
+                style='dashboard', 
+                context_paths=['AGENTS.md'],
+                title='Test App'
+            )
+            print('✓ configure_host with new parameters works')
+        except Exception as e:
+            print(f'× configure_host failed: {e}')
+            sys.exit(1)
+        "
+        
+  # Optional job with API key for agentic tests (if secret is available)
+  agentic-integration:
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    if: ${{ vars.RUN_AGENTIC_TESTS == 'true' }}
+    needs: cross-repo-integration
+    
+    steps:
+    - name: Checkout PraisonAI
+      uses: actions/checkout@v4
+      with:
+        path: PraisonAI
+        ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+        
+    - name: Checkout PraisonAIUI
+      uses: actions/checkout@v4
+      with:
+        repository: MervinPraison/PraisonAIUI
+        path: PraisonAIUI
+        ref: main
+        
+    - name: Set up Python
+      uses: actions/setup-python@v4
+      with:
+        python-version: '3.11'
+        
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        cd PraisonAI/src/praisonai-agents
+        pip install -e .
+        cd ../praisonai
+        pip install -e ".[ui,dev]"
+        cd ../../../PraisonAIUI
+        pip install -e .
+        
+    - name: Run agentic integration test
+      env:
+        OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+      run: |
+        cd PraisonAI/src/praisonai
+        if [ -n "$OPENAI_API_KEY" ]; then
+          python -m pytest tests/integration/test_aiui_host_agentic.py -v --timeout=600
+        else
+          echo "OPENAI_API_KEY not available, skipping agentic tests"
+        fi

.github/workflows/praisonai-issue-triage.yml

@@ -29,9 +29,6 @@ jobs:
         uses: actions/checkout@v4
         with:
           persist-credentials: false
-        with:
-          persist-credentials: false
-        with:
           fetch-depth: 0
           
       - name: Generate GitHub App Token
@@ -63,10 +60,12 @@ jobs:
           GH_TOKEN: ${{ steps.generate_token.outputs.token }}
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
           PRAISONAI_AUTO_APPROVE: "true"
+          ISSUE_NUMBER_INPUT: ${{ inputs.issue_number }}
+          ISSUE_NUMBER_EVENT: ${{ github.event.issue.number }}
         run: |
           if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
-            export ISSUE_NUMBER="${{ inputs.issue_number }}"
+            export ISSUE_NUMBER="$ISSUE_NUMBER_INPUT"
           else
-            export ISSUE_NUMBER="${{ github.event.issue.number }}"
+            export ISSUE_NUMBER="$ISSUE_NUMBER_EVENT"
           fi
           praisonai github triage --issue $ISSUE_NUMBER --agent-file .github/praisonai-issue-triage.yaml

.github/workflows/praisonai-pr-review.yml

@@ -30,22 +30,23 @@ jobs:
     steps:
       - name: Determine checkout ref
         id: dest
+        env:
+          ISSUE_NUMBER: ${{ github.event.issue.number }}
+          PR_NUMBER_INPUT: ${{ inputs.pr_number }}
+          PR_HEAD_SHA: ${{ github.event.pull_request.head.sha }}
         run: |
           if [ "${{ github.event_name }}" = "issue_comment" ]; then
-            echo "ref=refs/pull/${{ github.event.issue.number }}/head" >> "$GITHUB_OUTPUT"
+            echo "ref=refs/pull/$ISSUE_NUMBER/head" >> "$GITHUB_OUTPUT"
           elif [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
-            echo "ref=refs/pull/${{ inputs.pr_number }}/head" >> "$GITHUB_OUTPUT"
+            echo "ref=refs/pull/$PR_NUMBER_INPUT/head" >> "$GITHUB_OUTPUT"
           else
-            echo "ref=${{ github.event.pull_request.head.sha }}" >> "$GITHUB_OUTPUT"
+            echo "ref=$PR_HEAD_SHA" >> "$GITHUB_OUTPUT"
           fi
 
       - name: Checkout Repository
         uses: actions/checkout@v4
         with:
           persist-credentials: false
-        with:
-          persist-credentials: false
-        with:
           ref: ${{ steps.dest.outputs.ref }}
           fetch-depth: 0
           

.github/workflows/test-windows.yml

@@ -26,8 +26,8 @@ jobs:
     
     steps:
     - uses: actions/checkout@v4
-        with:
-          persist-credentials: false
+      with:
+        persist-credentials: false
     - name: Set up Python 3.11
       uses: actions/setup-python@v5
       with:
@@ -176,6 +176,18 @@ jobs:
       run: praisonai queue list
       continue-on-error: true
 
+    # ==================== PACKAGING TESTS ====================
+    - name: Test packaging/daemon entry points
+      run: |
+        echo "Testing both entry point methods..."
+        python -m praisonai --version
+        praisonai --version
+
+    - name: Test doctor packaging command
+      run: |
+        echo "Testing doctor packaging diagnostics..."
+        praisonai doctor packaging --json
+
     # ==================== REAL API TESTS ====================
     - name: Real API test - Agent run
       if: env.OPENAI_API_KEY != ''

MIGRATION_NOTES.md

@@ -1,52 +0,0 @@
-# Migration Notes for auto_approve_tools Behavioral Change
-
-## Breaking Change: auto_approve_tools Default Changed to True
-
-As of this release, the default value for `auto_approve_tools` in `BotConfig` has been changed from `False` to `True`.
-
-### Impact
-
-This affects all bot adapters that use the fallback pattern `config or BotConfig(...)`:
-- **TelegramBot**
-- **SlackBot** 
-- **DiscordBot**
-- **WhatsAppBot**
-- **EmailBot**
-- **AgentMailBot**
-
-### Behavior Change
-
-**Before:**
-```python
-# Tool calls required manual confirmation by default
-bot = TelegramBot(token="...", agent=agent)  # auto_approve_tools=False
-```
-
-**After:**
-```python
-# Tool calls are now auto-approved by default
-bot = TelegramBot(token="...", agent=agent)  # auto_approve_tools=True
-```
-
-### Rationale
-
-Chat bots cannot show CLI approval prompts to users, making manual approval impractical in messaging environments. This change enables bots to work seamlessly out-of-the-box while maintaining safety through:
-
-1. **Safe tool defaults**: Only non-destructive tools (web search, memory, scheduling) are auto-injected
-2. **Destructive tool filtering**: Tools like `execute_command` are excluded from auto-injection
-3. **Explicit opt-in for risky tools**: Destructive tools require explicit configuration
-
-### Migration
-
-If your bot adapter previously relied on manual tool approval, explicitly set `auto_approve_tools=False`:
-
-```python
-# Maintain old behavior (manual approval required)
-config = BotConfig(auto_approve_tools=False)
-bot = TelegramBot(token="...", agent=agent, config=config)
-```
-
-### Affected Packages
-
-- `praisonaiagents.bots.config.BotConfig` (default changed)
-- All bot adapters in `praisonai.bots.*` (when using default config)

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
@@ -345,7 +382,7 @@ praisonai flow
 
 Open **http://localhost:7861** — use the **Agent** and **Agent Team** components to create sequential or parallel workflows. Connect Chat Input → Agent Team → Chat Output for instant multi-agent pipelines.
 
-> 📖 [Full Flow docs](https://docs.praison.ai/docs/concepts/flow) — visual agent building, component reference, and deployment
+> 📖 [Full Flow docs](https://docs.praison.ai/docs/concepts/agentflow) — visual agent building, component reference, and deployment
 
 ### 8. PraisonAI UI 🤖 (Clean Chat)
 

api.md

@@ -40,6 +40,7 @@ Methods:
 * <code title="class Agent">Agent.<a href="./src/praisonai-agents/praisonaiagents/agent/agent.py">chat_with_context</a>(message: str, context: 'ContextPack', **kwargs) -> str</code>
 * <code title="class Agent">Agent.<a href="./src/praisonai-agents/praisonaiagents/agent/agent.py">checkpoints</a>() -> Optional[bool]</code>
 * <code title="class Agent">Agent.<a href="./src/praisonai-agents/praisonaiagents/agent/agent.py">checkpoints</a>(value: Optional[bool]) -> None</code>
+* <code title="class Agent">Agent.<a href="./src/praisonai-agents/praisonaiagents/agent/agent.py">clone_for_channel</a>() -> 'Agent'</code>
 * <code title="class Agent">Agent.<a href="./src/praisonai-agents/praisonaiagents/agent/agent.py">close</a>() -> None</code>
 * <code title="class Agent">Agent.<a href="./src/praisonai-agents/praisonaiagents/agent/agent.py">console</a>() -> Optional[Any]</code>
 * <code title="class Agent">Agent.<a href="./src/praisonai-agents/praisonaiagents/agent/agent.py">context_manager</a>() -> Optional[Any]</code>
@@ -171,6 +172,7 @@ Methods:
 * <code title="class BaseTool">BaseTool.<a href="./src/praisonai-agents/praisonaiagents/tools/base.py">safe_run</a>(**kwargs) -> ToolResult</code>
 * <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 BaseTool">BaseTool.<a href="./src/praisonai-agents/praisonaiagents/tools/base.py">validate_schema_roundtrip</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>
@@ -183,6 +185,7 @@ Methods:
 * <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">list_tools_with_allowed_filter</a>(context: Optional[Dict[str, Any]] = None) -> 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>
 * <code title="class ToolRegistry">ToolRegistry.<a href="./src/praisonai-agents/praisonaiagents/tools/registry.py">unregister</a>(name: str) -> bool</code>
 * <code title="class Tools">Tools.<a href="./src/praisonai-agents/praisonaiagents/tools/tools.py">internet_search</a>(*args, **kwargs)</code>
@@ -213,7 +216,7 @@ Methods:
 * <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, max_workers: Optional[int] = None) -> 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>
 
@@ -476,8 +479,10 @@ Methods:
 * <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_available_skills_by_state</a>(state: SkillState) -> List[LoadedSkill]</code>
 * <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">get_instructions</a>(name: str) -> Optional[str]</code>
 * <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">get_skill</a>(name: str) -> Optional[LoadedSkill]</code>
+* <code title="class SkillManager">SkillManager.<a href="./src/praisonai-agents/praisonaiagents/skills/manager.py">get_skills_diagnostics</a>() -> Dict[str, ValidationResult]</code>
 * <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>
@@ -486,6 +491,7 @@ Methods:
 * <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">validate_skill_capabilities</a>(skill_name: str, force_refresh: bool = False) -> ValidationResult</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>
@@ -601,6 +607,13 @@ Methods:
 * <code title="class ContextManager">ContextManager.<a href="./src/praisonai-agents/praisonaiagents/context/manager.py">set_tool_budget</a>(tool_name: str, max_tokens: int, protected: bool = False) -> None</code>
 * <code title="class ContextManager">ContextManager.<a href="./src/praisonai-agents/praisonaiagents/context/manager.py">truncate_tool_output</a>(tool_name: str, output: str) -> str</code>
 
+# UI
+
+Types:
+```python
+from praisonaiagents import A2UI
+```
+
 # Config
 
 Types:
@@ -755,7 +768,7 @@ Methods:
 
 Types:
 ```python
-from praisonai import Agent, 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__, arun, run, run_integrated_gateway
 ```
 
 # CLI
@@ -786,6 +799,10 @@ 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>
@@ -853,8 +870,10 @@ Methods:
 * <code title="cli">praisonai doctor env <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
 * <code title="cli">praisonai doctor mcp <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
 * <code title="cli">praisonai doctor network <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
+* <code title="cli">praisonai doctor packaging <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
 * <code title="cli">praisonai doctor performance <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
 * <code title="cli">praisonai doctor selftest <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
+* <code title="cli">praisonai doctor skills <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
 * <code title="cli">praisonai doctor tools <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
 * <code title="cli">praisonai doctor troubleshoot <a href="./src/praisonai/praisonai/cli/commands/doctor.py">--help</a></code>
 * <code title="cli">praisonai endpoints list <a href="./src/praisonai/praisonai/cli/commands/endpoints.py">--help</a></code>
@@ -886,6 +905,7 @@ Methods:
 * <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>
+* <code title="cli">praisonai gateway stop <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
 * <code title="cli">praisonai gateway uninstall <a href="./src/praisonai/praisonai/cli/commands/gateway.py">--help</a></code>
 * <code title="cli">praisonai github triage <a href="./src/praisonai/praisonai/cli/commands/github.py">--help</a></code>
 * <code title="cli">praisonai hooks add <a href="./src/praisonai/praisonai/cli/commands/hooks.py">--help</a></code>
@@ -1048,6 +1068,7 @@ Methods:
 * <code title="cli">praisonai serve status <a href="./src/praisonai/praisonai/cli/commands/serve.py">--help</a></code>
 * <code title="cli">praisonai serve stop <a href="./src/praisonai/praisonai/cli/commands/serve.py">--help</a></code>
 * <code title="cli">praisonai serve ui <a href="./src/praisonai/praisonai/cli/commands/serve.py">--help</a></code>
+* <code title="cli">praisonai serve ui-gateway <a href="./src/praisonai/praisonai/cli/commands/serve.py">--help</a></code>
 * <code title="cli">praisonai serve unified <a href="./src/praisonai/praisonai/cli/commands/serve.py">--help</a></code>
 * <code title="cli">praisonai session delete <a href="./src/praisonai/praisonai/cli/commands/session.py">--help</a></code>
 * <code title="cli">praisonai session export <a href="./src/praisonai/praisonai/cli/commands/session.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>=4.6.30" \
+    "praisonai>=4.6.48" \
     "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>=4.6.30" \
+    "praisonai>=4.6.48" \
     "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>=4.6.30" \
+    "praisonai>=4.6.48" \
     "praisonai[ui]" \
     "praisonai[crewai]"
 

examples/python/a2ui/send_surface.py

@@ -0,0 +1,27 @@
+#!/usr/bin/env python3
+"""Minimal A2UI example — agent + send_a2ui_messages tool.
+
+Install:
+    pip install praisonaiagents[a2ui]
+
+Run:
+    python send_surface.py
+
+For live rendering, use PraisonAIUI with AG-UI (/agui) — see integrate-a2ui-frontend docs.
+"""
+
+from praisonaiagents import Agent
+from praisonaiagents.tools.a2ui_tools import send_a2ui_messages
+
+agent = Agent(
+    name="ui-assistant",
+    instructions=(
+        "When asked for UI, call send_a2ui_messages with A2UI v0.9 JSON "
+        "(createSurface on surface id 'main')."
+    ),
+    tools=[send_a2ui_messages],
+)
+
+if __name__ == "__main__":
+    result = agent.start("Show a simple welcome card on surface main")
+    print(result)

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/MULTI_CHANNEL_GUIDE.md

@@ -0,0 +1,222 @@
+# Multi-Channel Bot Deployment Guide
+
+This guide explains how to set up multiple bots on the same platform (e.g., multiple Telegram bots) for different roles using PraisonAI, similar to Hermes-style AI workforces.
+
+## Overview
+
+PraisonAI supports running multiple bots on the same platform, each routing to different agents with specialized knowledge and capabilities. This is useful for:
+
+- **Role-based bots**: `@company_cfo_bot`, `@company_ops_bot`, `@company_content_bot`
+- **Department-specific agents**: Each with different knowledge brains and tool access
+- **Workflow separation**: Different bots handle different types of requests
+
+## Quick Start with Onboard Wizard
+
+The easiest way to set up multiple channels is using the onboard wizard:
+
+```bash
+praisonai onboard
+```
+
+The wizard now supports:
+1. **Multi-channel setup**: Add multiple bots for the same platform
+2. **Role-based naming**: Automatically generates `TELEGRAM_<ROLE>_BOT_TOKEN` environment variables
+3. **Agent routing**: Each channel routes to a different agent
+
+### Onboard Wizard Flow
+
+1. **Choose platform**: Start with your first platform (e.g., telegram)
+2. **Add more channels**: Wizard asks "Add another bot channel?"
+3. **Specify role**: Enter role name (e.g., "cfo", "ops", "content")
+4. **Configure tokens**: Each channel gets its own environment variable
+5. **Test connections**: Wizard tests each bot separately
+
+## Environment Variable Convention
+
+For multiple bots on the same platform, use this naming convention:
+
+| Channel | Environment Variable | Bot Purpose |
+|---------|---------------------|-------------|
+| `telegram_cfo` | `TELEGRAM_CFO_BOT_TOKEN` | CFO-related tasks |
+| `telegram_ops` | `TELEGRAM_OPS_BOT_TOKEN` | Operations tasks |
+| `telegram_content` | `TELEGRAM_CONTENT_BOT_TOKEN` | Content creation |
+| `telegram_support` | `TELEGRAM_SUPPORT_BOT_TOKEN` | Customer support |
+
+### Setting Up Tokens
+
+1. **Create bots in @BotFather**:
+   ```text
+   /newbot
+   Bot name: Company CFO Bot
+   Username: @company_cfo_bot
+   -> Copy token
+   
+   /newbot
+   Bot name: Company Ops Bot
+   Username: @company_ops_bot
+   -> Copy token
+   ```
+
+2. **Set environment variables**:
+   ```bash
+   export TELEGRAM_CFO_BOT_TOKEN="123456789:ABC..."
+   export TELEGRAM_OPS_BOT_TOKEN="987654321:DEF..."
+   export TELEGRAM_CONTENT_BOT_TOKEN="456789123:GHI..."
+   ```
+
+3. **Or use `~/.praisonai/.env`**:
+   ```env
+   TELEGRAM_CFO_BOT_TOKEN=123456789:ABC...
+   TELEGRAM_OPS_BOT_TOKEN=987654321:DEF...
+   TELEGRAM_CONTENT_BOT_TOKEN=456789123:GHI...
+   TELEGRAM_ALLOWED_USERS=123456789,987654321
+   ```
+
+## Example Configuration
+
+See `multi-telegram-hermes-workforce.yaml` for a complete example.
+
+### Key Features
+
+1. **Multiple Channels**:
+   ```yaml
+   channels:
+     telegram_cfo:
+       platform: telegram
+       token: ${TELEGRAM_CFO_BOT_TOKEN}
+       routes:
+         default: cfo
+     
+     telegram_ops:
+       platform: telegram
+       token: ${TELEGRAM_OPS_BOT_TOKEN}
+       routes:
+         default: ops
+   ```
+
+2. **Specialized Agents**:
+   ```yaml
+   agents:
+     cfo:
+       instructions: "You are a CFO agent. Help with financial analysis..."
+       model: gpt-4o-mini
+     
+     ops:
+       instructions: "You are an ops agent. Help with system monitoring..."
+       model: gpt-4o-mini
+   ```
+
+## Validation with Doctor
+
+Use the doctor command to validate your multi-channel setup:
+
+```bash
+praisonai doctor
+```
+
+The doctor checks for:
+- ✅ **Unique tokens**: Each channel uses a different bot token
+- ✅ **Naming convention**: Follows `PLATFORM_ROLE_BOT_TOKEN` pattern  
+- ✅ **Missing tokens**: All referenced environment variables are set
+- ❌ **Duplicate tokens**: Same token used by multiple channels (fails)
+
+## Deployment
+
+### Start Gateway
+
+```bash
+# Start the multi-bot gateway
+praisonai gateway start --config multi-telegram-hermes-workforce.yaml
+
+# Or install as daemon
+praisonai onboard  # Wizard installs daemon automatically
+praisonai gateway status
+```
+
+### Verify Setup
+
+1. **Check health**: `curl http://127.0.0.1:8765/health`
+2. **View dashboard**: `praisonai claw` → http://127.0.0.1:8082
+3. **Test bots**: Send messages to each bot to verify routing
+
+## Common Patterns
+
+### Basic Multi-Role Setup
+```yaml
+# Single platform, multiple roles
+channels:
+  telegram_support:
+    token: ${TELEGRAM_SUPPORT_BOT_TOKEN}
+    routes:
+      default: support_agent
+  
+  telegram_sales:
+    token: ${TELEGRAM_SALES_BOT_TOKEN}
+    routes:
+      default: sales_agent
+```
+
+### Cross-Platform Workforce
+```yaml
+# Multiple platforms, multiple roles
+channels:
+  telegram_cfo:
+    platform: telegram
+    token: ${TELEGRAM_CFO_BOT_TOKEN}
+    routes:
+      default: cfo
+  
+  discord_ops:
+    platform: discord
+    token: ${DISCORD_OPS_BOT_TOKEN}
+    routes:
+      default: ops
+  
+  slack_content:
+    platform: slack
+    token: ${SLACK_CONTENT_BOT_TOKEN}
+    routes:
+      default: content
+```
+
+## Troubleshooting
+
+### Common Issues
+
+1. **"Token is used by multiple channels"**
+   - Each bot needs its own unique token from @BotFather
+   - Check environment variables for duplicates
+
+2. **"Channel skipped - no token"**
+   - Set the environment variable: `export TELEGRAM_CFO_BOT_TOKEN="..."`
+   - Or update `~/.praisonai/.env`
+
+3. **"Bot not responding"**
+   - Verify token is correct: `praisonai doctor`
+   - Check bot is not already running elsewhere
+   - Ensure allowlist includes your user ID
+
+### Debugging Commands
+
+```bash
+# Check configuration
+praisonai doctor
+
+# View logs
+praisonai gateway logs
+
+# Test connection
+curl -H "Authorization: Bearer $GATEWAY_AUTH_TOKEN" \
+     http://127.0.0.1:8765/info
+```
+
+## Migration from Single-Bot Setup
+
+If you currently have a single bot setup, you can add more channels:
+
+1. **Run onboard wizard**: `praisonai onboard`
+2. **Add new channels**: Wizard will preserve existing setup
+3. **Update configuration**: New channels added to existing `bot.yaml`
+4. **Restart gateway**: `praisonai gateway restart`
+
+The wizard preserves backward compatibility with single-bot configurations.

examples/yaml/a2ui-canvas.yaml

@@ -0,0 +1,19 @@
+# A2UI canvas agent — YAML example
+#
+# Usage:
+#   pip install praisonaiagents[a2ui]
+#   praisonai agent run --file a2ui-canvas.yaml --prompt "Show a welcome card"
+#
+# For live A2UI rendering, use PraisonAIUI (AG-UI / surfaces):
+#   pip install praisonaiui
+#   See examples/python/a2ui/send_surface.py and PraisonAIUI example 29.
+
+agent:
+  name: "ui-assistant"
+  instructions: |
+    You are a helpful assistant. When the user asks for UI elements
+    (buttons, cards, forms), call send_a2ui_messages with valid A2UI v0.9 JSON.
+    Use surface id "main" unless told otherwise.
+  llm: gpt-4o-mini
+  tools:
+    - send_a2ui_messages

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

examples/yaml/multi-telegram-hermes-workforce.yaml

@@ -0,0 +1,66 @@
+# Multi-Telegram Channel Hermes Workforce Example
+# Generated by `praisonai onboard` with multi-channel support
+# Start with: praisonai gateway start --config multi-telegram-hermes-workforce.yaml
+
+gateway:
+  host: "127.0.0.1"
+  port: 8765
+
+# Legacy single-agent block for backward compatibility
+agent:
+  name: "assistant"
+  instructions: "You are a helpful AI assistant. You can search the web, manage schedules, and remember past conversations. Use your tools proactively."
+  model: "gpt-4o-mini"
+
+# Multi-agent workforce with different specialized roles
+agents:
+  cfo:
+    instructions: "You are a CFO agent. You are a helpful AI assistant. Use your tools proactively to help with CFO-related tasks like financial analysis, budget planning, and cost optimization."
+    model: gpt-4o-mini
+    memory: false
+
+  ops:
+    instructions: "You are an ops agent. You are a helpful AI assistant. Use your tools proactively to help with ops-related tasks like system monitoring, deployment, and infrastructure management."
+    model: gpt-4o-mini
+    memory: false
+
+  content:
+    instructions: "You are a content agent. You are a helpful AI assistant. Use your tools proactively to help with content-related tasks like writing, editing, and content strategy."
+    model: gpt-4o-mini
+    memory: false
+
+# Multiple Telegram channels - each with unique tokens and routing
+channels:
+  telegram_cfo:
+    platform: telegram
+    token: ${TELEGRAM_CFO_BOT_TOKEN}
+    group_policy: mention_only
+    allowed_users: ${TELEGRAM_ALLOWED_USERS}
+    # auto_approve_tools: true
+                               # Chat bots cannot show CLI approval prompts; safe tools auto-approved by default.
+    routes:
+      default: cfo
+
+  telegram_ops:
+    platform: telegram
+    token: ${TELEGRAM_OPS_BOT_TOKEN}
+    group_policy: mention_only
+    allowed_users: ${TELEGRAM_ALLOWED_USERS}
+    # auto_approve_tools: true
+                               # Chat bots cannot show CLI approval prompts; safe tools auto-approved by default.
+    routes:
+      default: ops
+
+  telegram_content:
+    platform: telegram
+    token: ${TELEGRAM_CONTENT_BOT_TOKEN}
+    group_policy: mention_only
+    allowed_users: ${TELEGRAM_ALLOWED_USERS}
+    # auto_approve_tools: true
+                               # Chat bots cannot show CLI approval prompts; safe tools auto-approved by default.
+    routes:
+      default: content
+
+# Legacy routing block
+routing:
+  default: assistant

src/praisonai-agents/praisonaiagents/__init__.py

@@ -343,6 +343,7 @@ _LAZY_IMPORTS = {
     # UI (optional)
     'AGUI': ('praisonaiagents.ui.agui', 'AGUI'),
     'A2A': ('praisonaiagents.ui.a2a', 'A2A'),
+    'A2UI': ('praisonaiagents.ui.a2ui', 'A2UI'),
     
     # Feature configs
     'MemoryConfig': ('praisonaiagents.config.feature_configs', 'MemoryConfig'),
@@ -814,6 +815,6 @@ def __dir__():
 # - Workflow, Task, Route, Parallel, Loop, Repeat, Pipeline, etc.
 # - MCP, FlowDisplay, track_workflow, FastContext, etc.
 # - Plan, PlanStep, TodoList, PlanningAgent, ApprovalCallback, etc.
-# - RAG, RAGConfig, RAGResult, AGUI, A2A, etc.
+# - RAG, RAGConfig, RAGResult, AGUI, A2A, A2UI, etc.
 # - All telemetry, display, and utility functions
 # ============================================================================

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

@@ -21,6 +21,7 @@ from .chat_handler import ChatHandlerMixin
 from .session_manager import SessionManagerMixin
 from .async_safety import AsyncSafeState
 from .unified_execution_mixin import UnifiedExecutionMixin
+from .sandbox_mixin import SandboxMixin
 
 # Module-level logger for thread safety errors and debugging
 logger = get_logger(__name__)
@@ -255,7 +256,7 @@ if TYPE_CHECKING:
 # Import structured error from central errors module
 from ..errors import BudgetExceededError
 
-class Agent(UnifiedExecutionMixin, ToolExecutionMixin, ChatHandlerMixin, SessionManagerMixin, ChatMixin, ExecutionMixin, MemoryMixin, AsyncMemoryMixin):
+class Agent(SandboxMixin, UnifiedExecutionMixin, ToolExecutionMixin, ChatHandlerMixin, SessionManagerMixin, ChatMixin, ExecutionMixin, MemoryMixin, AsyncMemoryMixin):
     # Class-level counter for generating unique display names for nameless agents
     _agent_counter = 0
     _agent_counter_lock = threading.Lock()
@@ -519,6 +520,7 @@ 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
@@ -554,6 +556,7 @@ class Agent(UnifiedExecutionMixin, ToolExecutionMixin, ChatHandlerMixin, Session
         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
+        sandbox: Optional[Union[bool, 'SandboxConfig']] = None,  # Sandbox for safe code execution
     ):
         """Initialize an Agent instance.
 
@@ -1521,6 +1524,8 @@ 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())
@@ -1531,6 +1536,7 @@ 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,
@@ -1552,6 +1558,10 @@ 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
@@ -1571,6 +1581,8 @@ 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
@@ -1785,7 +1797,9 @@ Your Goal: {self.goal}
                 _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
-                    import logging
+                    # (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,
@@ -1819,6 +1833,7 @@ 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
@@ -1933,6 +1948,132 @@ Your Goal: {self.goal}
         # Telemetry - lazy initialized via property for performance
         self.__telemetry = None
         self.__telemetry_initialized = False
+        
+        # Sandbox configuration - initialize SandboxMixin
+        super().__init__(sandbox=sandbox)
+
+    def __deepcopy__(self, memo: dict) -> "Agent":
+        """Custom deepcopy that creates fresh threading primitives.
+
+        threading.RLock (self.__cache_lock) and threading.Lock (self._cost_lock)
+        cannot be deep-copied on CPython < 3.13.  This hook deep-copies every
+        other attribute normally and replaces the locks with new instances so
+        that copy.deepcopy(agent) works in any Python version.
+        """
+        import copy
+        cls = self.__class__
+        result = cls.__new__(cls)
+        memo[id(self)] = result
+        for k, v in self.__dict__.items():
+            if k in ("_Agent__cache_lock",):
+                object.__setattr__(result, k, threading.RLock())
+            elif k == "_cost_lock":
+                object.__setattr__(result, k, threading.Lock())
+            else:
+                object.__setattr__(result, k, copy.deepcopy(v, memo))
+        return result
+
+    def clone_for_channel(self) -> "Agent":
+        """Return a fully independent copy of this agent for a gateway channel.
+        
+        This method safely clones an agent for use in multi-channel scenarios
+        by avoiding deepcopy issues with tools/handoffs and creating fresh
+        instances of mutable objects like interrupt_controller.
+        
+        Returns:
+            Agent: A new Agent instance with copied configuration but fresh
+                  locks, interrupt controller, and shallow-copied tools.
+        """
+        import copy
+        from ..tools.base import BaseTool
+        
+        # Build kwargs for new Agent instance, copying core attributes
+        clone_kwargs = {
+            # Core identity
+            'name': self.name,
+            'role': self.role, 
+            'goal': self.goal,
+            'backstory': self.backstory,
+            'instructions': self.instructions,
+            
+            # LLM configuration 
+            'llm': self.llm,
+            'base_url': getattr(self, 'base_url', None),
+            'api_key': getattr(self, 'api_key', None),
+            'auth': getattr(self, 'auth', None),
+            
+            # Shallow copy tools to avoid deepcopy issues with nested objects
+            'tools': list(self.tools) if self.tools else None,
+            
+            # Skip handoffs entirely - they shouldn't be shared across channels
+            # and can contain nested Agent instances that cause RLock issues
+            'handoffs': None,
+            
+            # Feature configurations - check for actual stored config objects
+            'memory': getattr(self, '_memory_config', None),
+            'knowledge': getattr(self, '_knowledge_config', None), 
+            'planning': getattr(self, '_planning_config', None),
+            'reflection': getattr(self, '_reflection_config', None),
+            'guardrails': getattr(self, '_guardrails_config', None),
+            'web': getattr(self, '_web_config', None),
+            'context': getattr(self, '_context_config', None),
+            'autonomy': getattr(self, '_autonomy_config', None),
+            'output': getattr(self, '_output_config', None),
+            'execution': getattr(self, '_execution_config', None),
+            'templates': getattr(self, '_template_config', None),
+            'caching': getattr(self, '_caching_config', None),
+            'hooks': getattr(self, '_hooks_config', None),
+            'skills': getattr(self, '_skills_config', None),
+            'approval': getattr(self, '_approval_config', None),
+            'learn': getattr(self, '_learn_config', None),
+            
+            # Tool configuration
+            'tool_timeout': getattr(self, '_tool_timeout', None),
+            'parallel_tool_calls': getattr(self, 'parallel_tool_calls', False),
+            
+            # CLI backend
+            'cli_backend': getattr(self, '_cli_backend', None),
+            
+            # Create fresh interrupt controller to avoid shared state
+            'interrupt_controller': None,  # Let new instance create its own
+            
+            # Sandbox config
+            'sandbox': getattr(self, '_sandbox_config', None),
+        }
+        
+        # Handle deprecated parameters for backward compatibility
+        import warnings
+        
+        # Check for deprecated standalone attributes and emit warnings
+        deprecated_attrs = [
+            ('allow_code_execution', 'execution'),
+            ('code_execution_mode', 'execution'), 
+            ('auto_save', 'memory'),
+            ('rate_limiter', 'execution'),
+            ('allow_delegation', 'handoffs'),
+            ('verification_hooks', 'autonomy')
+        ]
+        
+        for old_attr, new_param in deprecated_attrs:
+            if hasattr(self, old_attr):
+                value = getattr(self, old_attr)
+                if value is not None and value != False:  # Skip None/False defaults
+                    warnings.warn(
+                        f"Deprecated attribute '{old_attr}' found in agent clone. "
+                        f"Use '{new_param}=' parameter instead.",
+                        DeprecationWarning,
+                        stacklevel=2
+                    )
+                    # For basic compatibility, include in clone_kwargs if not already set
+                    if old_attr == 'allow_code_execution' and clone_kwargs['execution'] is None:
+                        from ..config.execution import ExecutionConfig
+                        clone_kwargs['execution'] = ExecutionConfig(code_execution=value)
+                    elif old_attr == 'auto_save' and clone_kwargs['memory'] is None:
+                        from ..config.memory import MemoryConfig  
+                        clone_kwargs['memory'] = MemoryConfig(auto_save=value)
+        
+        # Create new Agent instance
+        return self.__class__(**{k: v for k, v in clone_kwargs.items() if v is not None})
 
     @property
     def _telemetry(self):
@@ -4598,60 +4739,57 @@ Answer:"""
                 error=f"Agent guardrail validation error: {str(e)}"
             )
 
-    def _apply_guardrail_with_retry(self, response_text, prompt, temperature=1.0, tools=None, task_name=None, task_description=None, task_id=None):
-        """Apply guardrail validation with retry logic.
-        
-        Args:
-            response_text: The response to validate
-            prompt: Original prompt for regeneration if needed
-            temperature: Temperature for regeneration
-            tools: Tools for regeneration
-            
-        Returns:
-            str: The validated response text or None if validation fails after retries
-        """
+    def _validate_with_guardrail(self, response_text):
+        """Validate response with guardrail. Returns (success, result, error)."""
         if not self._guardrail_fn:
-            return response_text
+            return True, response_text, None
             
         from ..main import TaskOutput
         
+        task_output = TaskOutput(
+            description="Agent response output",
+            raw=response_text,
+            agent=self.name
+        )
+        
+        guardrail_result = self._process_guardrail(task_output)
+        
+        if guardrail_result.success:
+            # Return the potentially modified result
+            if guardrail_result.result and hasattr(guardrail_result.result, 'raw'):
+                return True, guardrail_result.result.raw, None
+            elif guardrail_result.result:
+                return True, str(guardrail_result.result), None
+            else:
+                return True, response_text, None
+        else:
+            return False, None, guardrail_result.error
+
+    def _apply_guardrail_with_retry(self, response_text, prompt, temperature=1.0, tools=None, task_name=None, task_description=None, task_id=None):
+        """Apply guardrail validation with retry logic (sync version)."""
         retry_count = 0
         current_response = response_text
         
         while retry_count <= self.max_guardrail_retries:
-            # Create TaskOutput object
-            task_output = TaskOutput(
-                description="Agent response output",
-                raw=current_response,
-                agent=self.name
-            )
+            success, result, error = self._validate_with_guardrail(current_response)
             
-            # Process guardrail
-            guardrail_result = self._process_guardrail(task_output)
-            
-            if guardrail_result.success:
+            if success:
                 logging.info(f"Agent {self.name}: Guardrail validation passed")
-                # Return the potentially modified result
-                if guardrail_result.result and hasattr(guardrail_result.result, 'raw'):
-                    return guardrail_result.result.raw
-                elif guardrail_result.result:
-                    return str(guardrail_result.result)
-                else:
-                    return current_response
+                return result
             
             # Guardrail failed
             if retry_count >= self.max_guardrail_retries:
                 raise Exception(
                     f"Agent {self.name} response failed guardrail validation after {self.max_guardrail_retries} retries. "
-                    f"Last error: {guardrail_result.error}"
+                    f"Last error: {error}"
                 )
             
             retry_count += 1
-            logging.warning(f"Agent {self.name}: Guardrail validation failed (retry {retry_count}/{self.max_guardrail_retries}): {guardrail_result.error}")
+            logging.warning(f"Agent {self.name}: Guardrail validation failed (retry {retry_count}/{self.max_guardrail_retries}): {error}")
             
             # Regenerate response for retry
             try:
-                retry_prompt = f"{prompt}\n\nNote: Previous response failed validation due to: {guardrail_result.error}. Please provide an improved response."
+                retry_prompt = f"{prompt}\n\nNote: Previous response failed validation due to: {error}. Please provide an improved response."
                 response = self._chat_completion([{"role": "user", "content": retry_prompt}], temperature, tools, task_name=task_name, task_description=task_description, task_id=task_id)
                 if response and response.choices:
                     content = response.choices[0].message.content
@@ -4660,10 +4798,46 @@ Answer:"""
                     raise Exception("Failed to generate retry response")
             except Exception as e:
                 logging.error(f"Agent {self.name}: Error during guardrail retry: {e}")
-                # If we can't regenerate, fail the guardrail
+                raise Exception(f"Agent {self.name} guardrail retry failed: {e}")
+        
+        return current_response
+
+    async def _aapply_guardrail_with_retry(self, response_text, prompt, temperature=1.0, tools=None, task_name=None, task_description=None, task_id=None):
+        """Apply guardrail validation with retry logic (async version)."""
+        retry_count = 0
+        current_response = response_text
+        
+        while retry_count <= self.max_guardrail_retries:
+            success, result, error = self._validate_with_guardrail(current_response)
+            
+            if success:
+                logging.info(f"Agent {self.name}: Guardrail validation passed")
+                return result
+            
+            # Guardrail failed
+            if retry_count >= self.max_guardrail_retries:
                 raise Exception(
-                    f"Agent {self.name} guardrail retry failed: {e}"
+                    f"Agent {self.name} response failed guardrail validation after {self.max_guardrail_retries} retries. "
+                    f"Last error: {error}"
                 )
+            
+            retry_count += 1
+            logging.warning(f"Agent {self.name}: Guardrail validation failed (retry {retry_count}/{self.max_guardrail_retries}): {error}")
+            
+            # Regenerate response for retry (async version)
+            try:
+                retry_prompt = f"{prompt}\n\nNote: Previous response failed validation due to: {error}. Please provide an improved response."
+                response = await self._execute_unified_achat_completion([{"role": "user", "content": retry_prompt}], temperature, tools, task_name=task_name, task_description=task_description, task_id=task_id)
+                if response and hasattr(response, 'choices') and response.choices:
+                    content = response.choices[0].message.content
+                    current_response = content.strip() if content else ""
+                elif isinstance(response, str):
+                    current_response = response.strip()
+                else:
+                    raise Exception("Failed to generate retry response")
+            except Exception as e:
+                logging.error(f"Agent {self.name}: Error during guardrail retry: {e}")
+                raise Exception(f"Agent {self.name} guardrail retry failed: {e}")
         
         return current_response
     
@@ -4899,6 +5073,19 @@ 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
     
@@ -4933,6 +5120,23 @@ 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
             

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
@@ -511,7 +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) ---
@@ -560,7 +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)
@@ -667,6 +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
@@ -699,11 +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}")
             
@@ -715,8 +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,
+                context={"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, 
@@ -851,7 +902,7 @@ Your Goal: {self.goal}"""
                 max_tokens=getattr(self, 'max_tokens', None),
                 stream=stream,
                 response_format=response_format,
-                execute_tool_fn=getattr(self, 'execute_tool', None),
+                execute_tool_fn=getattr(self, 'execute_tool_async', None),
                 console=self.console if (self.verbose or stream) else None,
                 display_fn=self._display_generating if self.verbose else None,
                 stream_callback=stream_callback,
@@ -884,6 +935,151 @@ Your Goal: {self.goal}"""
             logging.error(f"Unified async chat completion failed: {e}")
             raise
 
+    async def _finalize_unified_achat_response(
+        self,
+        response,
+        original_prompt,
+        temperature,
+        tools,
+        output_json,
+        output_pydantic,
+        reasoning_steps,
+        stream,
+        messages,
+        chat_history_length,
+        start_time,
+        task_name=None,
+        task_description=None,
+        task_id=None,
+    ):
+        """Post-process unified achat completion (parity with sync _chat_impl)."""
+        if not response:
+            self._truncate_chat_history(chat_history_length)
+            return None
+
+        if hasattr(response, "choices") and response.choices:
+            msg = response.choices[0].message
+            content = getattr(msg, "content", None)
+            response_text = content.strip() if content else ""
+            if (
+                reasoning_steps
+                and getattr(msg, "reasoning_content", None)
+            ):
+                response_text = msg.reasoning_content.strip()
+        else:
+            extracted = self._extract_llm_response_content(response)
+            response_text = (extracted or "").strip() if isinstance(extracted, str) else str(extracted or "")
+
+        if output_json or output_pydantic:
+            self._append_to_chat_history({"role": "assistant", "content": response_text})
+            self._persist_message("assistant", response_text)
+            try:
+                validated_response = self._apply_guardrail_with_retry(
+                    response_text, original_prompt, temperature, tools,
+                    task_name, task_description, task_id,
+                )
+                self._execute_callback_and_display(
+                    original_prompt, validated_response, time.time() - start_time,
+                    task_name, task_description, task_id,
+                )
+                return await self._atrigger_after_agent_hook(
+                    original_prompt, validated_response, start_time,
+                )
+            except Exception as e:
+                logging.error(f"Agent {self.name}: Guardrail validation failed for JSON output: {e}")
+                self._truncate_chat_history(chat_history_length)
+                return None
+
+        if not self.self_reflect:
+            self._append_to_chat_history({"role": "assistant", "content": response_text})
+            self._persist_message("assistant", response_text)
+            try:
+                validated_response = self._apply_guardrail_with_retry(
+                    response_text, original_prompt, temperature, tools,
+                    task_name, task_description, task_id,
+                )
+                self._execute_callback_and_display(
+                    original_prompt, validated_response, time.time() - start_time,
+                    task_name, task_description, task_id,
+                )
+                return await self._atrigger_after_agent_hook(
+                    original_prompt, validated_response, start_time,
+                )
+            except Exception as e:
+                logging.error(f"Agent {self.name}: Guardrail validation failed: {e}")
+                self._truncate_chat_history(chat_history_length)
+                return None
+
+        # Self-reflection (legacy OpenAI path)
+        reflection_count = 0
+        while True:
+            reflection_prompt = f"""
+Reflect on your previous response: '{response_text}'.
+{self.reflect_prompt if self.reflect_prompt else "Identify any flaws, improvements, or actions."}
+Provide a "satisfactory" status ('yes' or 'no').
+Output MUST be JSON with 'reflection' and 'satisfactory'.
+            """
+            reflection_messages = messages + [
+                {"role": "assistant", "content": response_text},
+                {"role": "user", "content": reflection_prompt},
+            ]
+            try:
+                if self._openai_client is None:
+                    self._append_to_chat_history({"role": "assistant", "content": response_text})
+                    return await self._atrigger_after_agent_hook(
+                        original_prompt, response_text, start_time,
+                    )
+                reflection_response = await self._openai_client.async_client.beta.chat.completions.parse(
+                    model=self.reflect_llm if self.reflect_llm else self.llm,
+                    messages=reflection_messages,
+                    temperature=temperature,
+                    response_format=_get_display_functions()['ReflectionOutput'],
+                )
+                reflection_output = reflection_response.choices[0].message.parsed
+                if reflection_output.satisfactory == "yes" and reflection_count >= self.min_reflect - 1:
+                    break
+                if reflection_count >= self.max_reflect - 1:
+                    break
+                regenerate_messages = reflection_messages + [
+                    {
+                        "role": "assistant",
+                        "content": (
+                            f"Self Reflection: {reflection_output.reflection} "
+                            f"Satisfactory?: {reflection_output.satisfactory}"
+                        ),
+                    },
+                    {"role": "user", "content": "Now regenerate your response using the reflection you made"},
+                ]
+                new_response = await self._openai_client.async_client.chat.completions.create(
+                    model=self.llm,
+                    messages=regenerate_messages,
+                    temperature=temperature,
+                )
+                response_text = new_response.choices[0].message.content
+                reflection_count += 1
+            except Exception as e:
+                logging.error("Reflection parsing failed.", exc_info=True)
+                reflection_count += 1
+                if reflection_count >= self.max_reflect:
+                    break
+
+        try:
+            validated_response = self._apply_guardrail_with_retry(
+                response_text, original_prompt, temperature, tools,
+                task_name, task_description, task_id,
+            )
+            self._execute_callback_and_display(
+                original_prompt, validated_response, time.time() - start_time,
+                task_name, task_description, task_id,
+            )
+            return await self._atrigger_after_agent_hook(
+                original_prompt, validated_response, start_time,
+            )
+        except Exception as e:
+            logging.error(f"Agent {self.name}: Guardrail validation failed for OpenAI client: {e}")
+            self._truncate_chat_history(chat_history_length)
+            return None
+
     def _execute_callback_and_display(self, prompt: str, response: str, generation_time: float, task_name=None, task_description=None, task_id=None):
         """Helper method to execute callbacks and display interaction.
         
@@ -1098,7 +1294,7 @@ Your Goal: {self.goal}"""
             )
         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.
         
@@ -1142,11 +1338,17 @@ 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:
@@ -1367,7 +1569,16 @@ 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)
@@ -1663,7 +1874,7 @@ 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:
@@ -1686,13 +1897,20 @@ 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
@@ -1815,31 +2033,43 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                         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._append_to_chat_history({"role": "assistant", "content": response_text})
 
@@ -1849,7 +2079,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                     
                     # Apply guardrail validation for custom LLM response
                     try:
-                        validated_response = self._apply_guardrail_with_retry(response_text, prompt, temperature, tools, task_name, task_description, task_id)
+                        validated_response = await self._aapply_guardrail_with_retry(response_text, prompt, temperature, tools, task_name, task_description, task_id)
                         # Execute callback after validation
                         self._execute_callback_and_display(normalized_content, validated_response, time.time() - start_time, task_name, task_description, task_id)
                         return await self._atrigger_after_agent_hook(prompt, validated_response, start_time)
@@ -1957,8 +2187,216 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                             task_id=task_id,
                             response_format=response_format
                         )
-                        # Continue to handle structured outputs and other processing
-                        # Don't return immediately - fall through to existing logic
+                        
+                        # Process response - mirror sync _chat_impl behavior (lines ~1544-1602)
+                        if not response:
+                            # Rollback chat history on response failure
+                            self._truncate_chat_history(chat_history_length)
+                            return None
+
+                        # Extract response content using the same method as sync
+                        response_text = self._extract_llm_response_content(response)
+                        if isinstance(response_text, str):
+                            response_text = response_text.strip()
+
+                        # Handle output_json or output_pydantic if specified
+                        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._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
+                            try:
+                                validated_response = await self._aapply_guardrail_with_retry(response_text, original_prompt, temperature, tools, task_name, task_description, task_id)
+                                # Execute callback after validation
+                                self._execute_callback_and_display(original_prompt, validated_response, time.time() - start_time, task_name, task_description, task_id)
+                                return await self._atrigger_after_agent_hook(original_prompt, validated_response, start_time)
+                            except Exception as e:
+                                logging.error(f"Agent {self.name}: Guardrail validation failed for JSON output: {e}")
+                                # Rollback chat history on guardrail failure
+                                self._truncate_chat_history(chat_history_length)
+                                return None
+
+                        # For regular responses (no self-reflection)
+                        if not self.self_reflect:
+                            # User message already added before LLM call via _build_messages
+                            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:
+                                logging.debug(f"Agent {self.name} final response: {response_text}")
+                            # Return only reasoning content if reasoning_steps is True
+                            if reasoning_steps and hasattr(response, 'choices') and response.choices and hasattr(response.choices[0].message, 'reasoning_content') and response.choices[0].message.reasoning_content:
+                                # Apply guardrail to reasoning content
+                                try:
+                                    validated_reasoning = await self._aapply_guardrail_with_retry(response.choices[0].message.reasoning_content, original_prompt, temperature, tools, task_name, task_description, task_id)
+                                    # Execute callback after validation
+                                    self._execute_callback_and_display(original_prompt, validated_reasoning, time.time() - start_time, task_name, task_description, task_id)
+                                    return await self._atrigger_after_agent_hook(original_prompt, validated_reasoning, start_time)
+                                except Exception as e:
+                                    logging.error(f"Agent {self.name}: Guardrail validation failed for reasoning content: {e}")
+                                    # Rollback chat history on guardrail failure
+                                    self._truncate_chat_history(chat_history_length)
+                                    return None
+                            else:
+                                # Apply guardrail to regular response content
+                                try:
+                                    validated_response = await self._aapply_guardrail_with_retry(response_text, original_prompt, temperature, tools, task_name, task_description, task_id)
+                                    # Execute callback after validation
+                                    self._execute_callback_and_display(original_prompt, validated_response, time.time() - start_time, task_name, task_description, task_id)
+                                    return await self._atrigger_after_agent_hook(original_prompt, validated_response, start_time)
+                                except Exception as e:
+                                    logging.error(f"Agent {self.name}: Guardrail validation failed: {e}")
+                                    # Rollback chat history on guardrail failure
+                                    self._truncate_chat_history(chat_history_length)
+                                    return None
+                        
+                        # If self-reflection is enabled, implement reflection logic
+                        if self.self_reflect:
+                            # Implement async self-reflection similar to sync path
+                            reflection_prompt = f"""
+Reflect on your previous response: '{response_text}'.
+{self.reflect_prompt if self.reflect_prompt else "Identify any flaws, improvements, or actions."}
+Provide a "satisfactory" status ('yes' or 'no').
+Output MUST be JSON with 'reflection' and 'satisfactory'.
+                            """
+                            
+                            # Add reflection prompt to messages
+                            reflection_messages = messages + [
+                                {"role": "assistant", "content": response_text},
+                                {"role": "user", "content": reflection_prompt}
+                            ]
+                            
+                            reflection_count = 0
+                            
+                            while True:
+                                try:
+                                    # Check if OpenAI client is available for self-reflection
+                                    if self._openai_client is None:
+                                        # For custom LLMs, self-reflection with structured output is not supported
+                                        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._append_to_chat_history({"role": "assistant", "content": response_text})
+                                        # Persist assistant message to DB
+                                        self._persist_message("assistant", response_text)
+                                        try:
+                                            validated_response = await self._aapply_guardrail_with_retry(response_text, original_prompt, temperature, tools, task_name, task_description, task_id)
+                                            # Execute callback after validation
+                                            self._execute_callback_and_display(original_prompt, validated_response, time.time() - start_time, task_name, task_description, task_id)
+                                            return await self._atrigger_after_agent_hook(original_prompt, validated_response, start_time)
+                                        except Exception as e:
+                                            logging.error(f"Agent {self.name}: Guardrail validation failed: {e}")
+                                            # Rollback chat history on guardrail failure
+                                            self._truncate_chat_history(chat_history_length)
+                                            return None
+                                    
+                                    reflection_response = await self._openai_client.async_client.beta.chat.completions.parse(
+                                        model=self.reflect_llm if self.reflect_llm else self.llm,
+                                        messages=reflection_messages,
+                                        temperature=temperature,
+                                        response_format=_get_display_functions()['ReflectionOutput']
+                                    )
+                                    
+                                    reflection_output = reflection_response.choices[0].message.parsed
+                                    
+                                    if self.verbose:
+                                        _get_display_functions()['display_self_reflection'](f"Agent {self.name} self reflection (using {self.reflect_llm if self.reflect_llm else self.llm}): reflection='{reflection_output.reflection}' satisfactory='{reflection_output.satisfactory}'", console=self.console)
+                                    
+                                    # Only consider satisfactory after minimum reflections
+                                    if reflection_output.satisfactory == "yes" and reflection_count >= self.min_reflect - 1:
+                                        if self.verbose:
+                                            _get_display_functions()['display_self_reflection']("Agent marked the response as satisfactory after meeting minimum reflections", console=self.console)
+                                        # Add to chat history and return
+                                        self._append_to_chat_history({"role": "assistant", "content": response_text})
+                                        # Persist assistant message to DB
+                                        self._persist_message("assistant", response_text)
+                                        try:
+                                            validated_response = await self._aapply_guardrail_with_retry(response_text, original_prompt, temperature, tools, task_name, task_description, task_id)
+                                            # Execute callback after validation
+                                            self._execute_callback_and_display(original_prompt, validated_response, time.time() - start_time, task_name, task_description, task_id)
+                                            return await self._atrigger_after_agent_hook(original_prompt, validated_response, start_time)
+                                        except Exception as e:
+                                            logging.error(f"Agent {self.name}: Guardrail validation failed after reflection: {e}")
+                                            # Rollback chat history on guardrail failure
+                                            self._truncate_chat_history(chat_history_length)
+                                            return None
+                                    
+                                    # Check if we've hit max reflections
+                                    if reflection_count >= self.max_reflect - 1:
+                                        if self.verbose:
+                                            _get_display_functions()['display_self_reflection']("Maximum reflection count reached, returning current response", console=self.console)
+                                        # Add to chat history and return
+                                        self._append_to_chat_history({"role": "assistant", "content": response_text})
+                                        # Persist assistant message to DB
+                                        self._persist_message("assistant", response_text)
+                                        try:
+                                            validated_response = await self._aapply_guardrail_with_retry(response_text, original_prompt, temperature, tools, task_name, task_description, task_id)
+                                            # Execute callback after validation
+                                            self._execute_callback_and_display(original_prompt, validated_response, time.time() - start_time, task_name, task_description, task_id)
+                                            return await self._atrigger_after_agent_hook(original_prompt, validated_response, start_time)
+                                        except Exception as e:
+                                            logging.error(f"Agent {self.name}: Guardrail validation failed after max reflections: {e}")
+                                            # Rollback chat history on guardrail failure
+                                            self._truncate_chat_history(chat_history_length)
+                                            return None
+                                    
+                                    # Regenerate response based on reflection
+                                    regenerate_messages = reflection_messages + [
+                                        {"role": "assistant", "content": f"Self Reflection: {reflection_output.reflection} Satisfactory?: {reflection_output.satisfactory}"},
+                                        {"role": "user", "content": "Now regenerate your response using the reflection you made"}
+                                    ]
+                                    
+                                    new_response = await self._execute_unified_achat_completion(
+                                        messages=regenerate_messages,
+                                        temperature=temperature,
+                                        tools=formatted_tools,
+                                        stream=stream,
+                                        reasoning_steps=reasoning_steps,
+                                        task_name=task_name,
+                                        task_description=task_description,
+                                        task_id=task_id,
+                                        response_format=response_format
+                                    )
+                                    
+                                    if new_response:
+                                        new_response_text = self._extract_llm_response_content(new_response)
+                                        if isinstance(new_response_text, str):
+                                            response_text = new_response_text.strip()
+                                        # Update reflection_messages to include the new response for next iteration
+                                        reflection_messages = regenerate_messages + [
+                                            {"role": "assistant", "content": response_text}
+                                        ]
+                                    
+                                    reflection_count += 1
+                                    
+                                except Exception as e:
+                                    if self.verbose:
+                                        _get_display_functions()['display_error'](f"Error in parsing self-reflection json {e}. Retrying", console=self.console)
+                                    logging.error("Reflection parsing failed.", exc_info=True)
+                                    reflection_count += 1
+                                    if reflection_count >= self.max_reflect:
+                                        # Return original response after max reflection attempts
+                                        self._append_to_chat_history({"role": "assistant", "content": response_text})
+                                        # Persist assistant message to DB
+                                        self._persist_message("assistant", response_text)
+                                        try:
+                                            validated_response = await self._aapply_guardrail_with_retry(response_text, original_prompt, temperature, tools, task_name, task_description, task_id)
+                                            # Execute callback after validation
+                                            self._execute_callback_and_display(original_prompt, validated_response, time.time() - start_time, task_name, task_description, task_id)
+                                            return await self._atrigger_after_agent_hook(original_prompt, validated_response, start_time)
+                                        except Exception as guard_e:
+                                            logging.error(f"Agent {self.name}: Guardrail validation failed after reflection error: {guard_e}")
+                                            # Rollback chat history on guardrail failure
+                                            self._truncate_chat_history(chat_history_length)
+                                            return None
+                                    continue
+                        
+                        # This should never be reached due to the returns above
+                        # But adding as safety fallback
+                        logging.warning(f"Agent {self.name}: Unexpected code path reached in unified dispatch")
+                        return None
                     else:
                         # LEGACY: Check if OpenAI client is available
                         if self._openai_client is None:
@@ -2095,7 +2533,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                         
                         # Apply guardrail validation for OpenAI client response
                         try:
-                            validated_response = self._apply_guardrail_with_retry(response_text, original_prompt, temperature, tools, task_name, task_description, task_id)
+                            validated_response = await self._aapply_guardrail_with_retry(response_text, original_prompt, temperature, tools, task_name, task_description, task_id)
                             # Execute callback after validation
                             self._execute_callback_and_display(original_prompt, validated_response, time.time() - start_time, task_name, task_description, task_id)
                             return await self._atrigger_after_agent_hook(original_prompt, validated_response, start_time)
@@ -2164,13 +2602,9 @@ 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:

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.
@@ -948,7 +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}")
@@ -964,9 +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/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.
@@ -404,9 +403,33 @@ class MemoryMixin:
                     self._session_store.add_user_message(self._session_id, content)
                 elif role == "assistant":
                     self._session_store.add_assistant_message(self._session_id, content)
+                self._persist_session_stats()
             except Exception as e:
                 logging.warning(f"Failed to persist message to session store: {e}")
 
+    def _persist_session_stats(self):
+        """Flush agent cost/token stats into session JSON metadata."""
+        store = getattr(self, "_session_store", None)
+        session_id = getattr(self, "_session_id", None)
+        if store is None or session_id is None:
+            return
+        if not hasattr(store, "update_session_metadata"):
+            return
+        try:
+            summary = getattr(self, "cost_summary", {}) or {}
+            llm = getattr(self, "llm", None)
+            model = str(llm) if llm is not None else None
+            store.update_session_metadata(
+                session_id,
+                model=model,
+                total_tokens=summary.get("tokens_in", 0) + summary.get("tokens_out", 0),
+                cost=summary.get("cost", 0),
+                source=getattr(self, "_session_source", "chat"),
+                agent_id=getattr(self, "agent_id", None) or getattr(self, "_agent_id", None),
+            )
+        except Exception as e:
+            logging.debug(f"Session stats persist skipped: {e}")
+
     def session_id(self) -> Optional[str]:
         """Get the current session ID."""
         return self._session_id

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

@@ -0,0 +1,231 @@
+"""
+Sandbox execution mixin for Agent.
+
+Provides safe code execution capabilities to agents via the sandbox framework.
+"""
+
+import logging
+from typing import Optional, Dict, Any, Union
+
+from ..sandbox import (
+    SandboxConfig, 
+    SandboxManager, 
+    SandboxResult,
+    check_code_safety,
+    format_warnings,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class SandboxMixin:
+    """Mixin that adds sandbox execution capabilities to Agent.
+    
+    Provides safe code execution in isolated environments with security pre-checks.
+    """
+    
+    def __init__(self, *args, **kwargs):
+        # Extract sandbox config from kwargs
+        self.sandbox_config: Optional[SandboxConfig] = kwargs.pop('sandbox', None)
+        
+        # Convert boolean True to default config
+        if self.sandbox_config is True:
+            self.sandbox_config = SandboxConfig.subprocess()
+        
+        super().__init__(*args, **kwargs)
+        
+        self._sandbox_manager: Optional[SandboxManager] = None
+    
+    @property
+    def has_sandbox(self) -> bool:
+        """Check if sandbox is configured for this agent."""
+        return self.sandbox_config is not None
+    
+    def get_sandbox_manager(self) -> Optional[SandboxManager]:
+        """Get or create sandbox manager."""
+        if not self.has_sandbox:
+            return None
+        
+        if self._sandbox_manager is None:
+            self._sandbox_manager = SandboxManager(self.sandbox_config)
+        
+        return self._sandbox_manager
+    
+    async def execute_code(
+        self,
+        code: str,
+        language: str = "python",
+        check_security: bool = True,
+        **kwargs
+    ) -> SandboxResult:
+        """Execute code safely in configured sandbox.
+        
+        Args:
+            code: Code to execute
+            language: Programming language (python, bash, etc.)
+            check_security: Whether to run security pre-checks
+            **kwargs: Additional arguments passed to sandbox.execute()
+            
+        Returns:
+            SandboxResult with execution details
+            
+        Raises:
+            RuntimeError: If no sandbox is configured
+        """
+        if not self.has_sandbox:
+            raise RuntimeError(
+                "No sandbox configured. Set sandbox=True or provide SandboxConfig to Agent()"
+            )
+        
+        # Security pre-check if enabled
+        if check_security:
+            warnings = check_code_safety(code, language)
+            if warnings:
+                warning_text = format_warnings(warnings)
+                if self.verbose:
+                    logger.warning(f"Security warnings for code execution:\n{warning_text}")
+                
+                # Store warnings in result metadata
+                kwargs.setdefault('metadata', {})['security_warnings'] = warnings
+        
+        manager = self.get_sandbox_manager()
+        return await manager.run_code(code, language=language, **kwargs)
+    
+    def execute_code_sync(
+        self,
+        code: str,
+        language: str = "python", 
+        check_security: bool = True,
+        **kwargs
+    ) -> SandboxResult:
+        """Synchronous wrapper for execute_code.
+        
+        Args:
+            code: Code to execute
+            language: Programming language
+            check_security: Whether to run security pre-checks
+            **kwargs: Additional arguments
+            
+        Returns:
+            SandboxResult with execution details
+        """
+        from ..approval.utils import run_coroutine_safely
+        return run_coroutine_safely(self.execute_code(code, language, check_security, **kwargs))
+    
+    async def run_shell_command(
+        self,
+        command: Union[str, list],
+        check_security: bool = True,
+        **kwargs
+    ) -> SandboxResult:
+        """Run a shell command in the sandbox.
+        
+        Args:
+            command: Shell command to run
+            check_security: Whether to run security pre-checks
+            **kwargs: Additional arguments
+            
+        Returns:
+            SandboxResult with execution details
+        """
+        if not self.has_sandbox:
+            raise RuntimeError("No sandbox configured")
+        
+        # Convert command to string for security checking
+        if isinstance(command, list):
+            command_str = " ".join(command)
+        else:
+            command_str = command
+        
+        # Security pre-check
+        if check_security:
+            warnings = check_code_safety(command_str, "bash")
+            if warnings:
+                warning_text = format_warnings(warnings)
+                if self.verbose:
+                    logger.warning(f"Security warnings for command:\n{warning_text}")
+        
+        async with self.get_sandbox_manager() as sandbox:
+            return await sandbox.run_command(command, **kwargs)
+    
+    def get_sandbox_status(self) -> Dict[str, Any]:
+        """Get sandbox status information.
+        
+        Returns:
+            Dictionary with sandbox status details
+        """
+        if not self.has_sandbox:
+            return {"configured": False}
+        
+        status = {
+            "configured": True,
+            "config": self.sandbox_config.to_dict() if self.sandbox_config else None,
+        }
+        
+        manager = self.get_sandbox_manager()
+        if manager:
+            # Get available sandbox types
+            available_types = manager.get_available_types()
+            status["available_types"] = available_types
+            status["current_type"] = self.sandbox_config.sandbox_type
+        
+        return status
+    
+    async def sandbox_cleanup(self):
+        """Clean up sandbox resources."""
+        if self._sandbox_manager:
+            # The context manager handles cleanup, but we can force it here
+            try:
+                await self._sandbox_manager.__aexit__(None, None, None)
+            except Exception as e:
+                logger.warning(f"Error during sandbox cleanup: {e}")
+            finally:
+                self._sandbox_manager = None
+    
+    def _get_code_execution_tools(self):
+        """Get sandbox-based code execution tools for this agent.
+        
+        Returns tools that use the configured sandbox for safe execution.
+        
+        Returns:
+            List of tool functions
+        """
+        if not self.has_sandbox:
+            return []
+        
+        from ..tools import tool
+        
+        @tool
+        def execute_python_code(code: str) -> str:
+            """Execute Python code safely in sandbox.
+            
+            Args:
+                code: Python code to execute
+                
+            Returns:
+                Execution output
+            """
+            result = self.execute_code_sync(code, language="python")
+            if result.success:
+                return result.stdout or result.output
+            else:
+                return f"Error: {result.error or result.stderr}"
+        
+        @tool  
+        def execute_shell_command(command: str) -> str:
+            """Execute shell command safely in sandbox.
+            
+            Args:
+                command: Shell command to execute
+                
+            Returns:
+                Command output
+            """
+            from ..approval.utils import run_coroutine_safely
+            result = run_coroutine_safely(self.run_shell_command(command))
+            if result.success:
+                return result.stdout or result.output
+            else:
+                return f"Error: {result.error or result.stderr}"
+        
+        return [execute_python_code, execute_shell_command]

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

@@ -26,6 +26,21 @@ if TYPE_CHECKING:
 class ToolExecutionMixin:
     """Mixin providing toolexecution methods for the Agent class."""
 
+    def _get_existing_stream_emitter(self):
+        """Return an already-initialized stream emitter without creating one."""
+        emitter = getattr(self, "_stream_emitter", None)
+        if emitter is not None:
+            return emitter
+
+        # Support name-mangled private attributes across class renames/inheritance.
+        for cls in type(self).mro():
+            mangled = f"_{cls.__name__}__stream_emitter"
+            if hasattr(self, mangled):
+                emitter = getattr(self, mangled, None)
+                if emitter is not None:
+                    return emitter
+        return None
+
     def _resolve_tool_names(self, tool_names):
         """Resolve tool names to actual tool instances from registry.
         
@@ -160,8 +175,9 @@ class ToolExecutionMixin:
         
         # Emit TOOL_CALL_START to stream_emitter (for AIUI/AG-UI consumers)
         # Zero overhead when no callbacks registered
-        if hasattr(self, '_Agent__stream_emitter') and getattr(self, "_Agent__stream_emitter", None) is not None and getattr(self, "_Agent__stream_emitter", None).has_callbacks:
-            getattr(self, "_Agent__stream_emitter", None).emit(StreamEvent(
+        _stream_emitter = self._get_existing_stream_emitter()
+        if _stream_emitter is not None and _stream_emitter.has_callbacks:
+            _stream_emitter.emit(StreamEvent(
                 type=StreamEventType.TOOL_CALL_START,
                 timestamp=_tool_start_perf,
                 tool_call={
@@ -194,6 +210,21 @@ 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:
@@ -202,7 +233,7 @@ 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'):
@@ -219,7 +250,7 @@ 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
@@ -260,10 +291,10 @@ class ToolExecutionMixin:
             
             # Emit TOOL_CALL_RESULT to stream_emitter (for AIUI/AG-UI consumers)
             # Zero overhead when no callbacks registered
-            if hasattr(self, '_Agent__stream_emitter') and getattr(self, "_Agent__stream_emitter", None) is not None and getattr(self, "_Agent__stream_emitter", None).has_callbacks:
+            if _stream_emitter is not None and _stream_emitter.has_callbacks:
                 # Truncate result for stream event (keep it reasonable for UI display)
                 result_summary = str(result)[:500] if result else None
-                getattr(self, "_Agent__stream_emitter", None).emit(StreamEvent(
+                _stream_emitter.emit(StreamEvent(
                     type=StreamEventType.TOOL_CALL_RESULT,
                     timestamp=_time.perf_counter(),
                     tool_call={
@@ -275,6 +306,18 @@ class ToolExecutionMixin:
                     agent_id=self.name,
                     metadata={"duration_ms": _duration_ms},
                 ))
+                _stream_emitter.emit(StreamEvent(
+                    type=StreamEventType.TOOL_CALL_END,
+                    timestamp=_time.perf_counter(),
+                    tool_call={
+                        "name": function_name,
+                        "arguments": arguments,
+                        "result": result_summary,
+                        "id": tool_call_id,
+                    },
+                    agent_id=self.name,
+                    metadata={"duration_ms": _duration_ms},
+                ))
             
             # Trigger AFTER_TOOL hook
             from ..hooks import HookEvent, AfterToolInput
@@ -594,6 +637,73 @@ 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."""
 
@@ -691,11 +801,9 @@ 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:
@@ -760,12 +868,18 @@ 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
 
@@ -777,44 +891,66 @@ 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
 
     def pending_approval_count(self) -> int:
         """Number of approval requests still waiting."""
         return len(self._pending_approvals)
-

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

@@ -99,7 +99,7 @@ class UnifiedExecutionMixin:
                 )
             # 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

@@ -15,9 +15,9 @@ from enum import Enum
 
 # Import token tracking
 try:
-    from ..telemetry.token_collector import _token_collector
+    from ..telemetry.token_collector import get_token_collector
 except ImportError:
-    _token_collector = None
+    get_token_collector = None
 
 # Import async utility for hot-path usage
 try:
@@ -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
 
 
@@ -1010,8 +1027,10 @@ 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)
@@ -1038,6 +1057,11 @@ class AgentTeam:
                                 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
@@ -1047,7 +1071,17 @@ class AgentTeam:
                             logger.info(f"Task {task_id}: Guardrail validation passed")
                         except Exception as e:
                             logger.error(f"Task {task_id}: Error in guardrail processing: {e}")
-                            # Continue execution even if guardrail fails to avoid breaking the task
+                            # 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
@@ -1056,6 +1090,11 @@ class AgentTeam:
                     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:
@@ -1078,7 +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":
@@ -1088,8 +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"""
@@ -1271,12 +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:
@@ -1315,7 +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":
@@ -1325,8 +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"""
@@ -1449,56 +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, task_id):
+                try:
+                    verbose_task_start_callback(task, task_id)
+                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, task_id)
+                    except Exception as e:
+                        logging.debug(f"Error in original task start callback: {e}")
+            
+            def composed_on_task_complete(task, task_output):
+                try:
+                    verbose_task_complete_callback(task, task_output)
+                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, task_output)
+                    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
@@ -1681,18 +1826,19 @@ class AgentTeam:
 
     def get_token_usage_summary(self) -> Dict[str, Any]:
         """Get a summary of token usage across all agents and tasks."""
-        if not _token_collector:
+        if not get_token_collector:
             return {"error": "Token tracking not available"}
         
-        return _token_collector.get_session_summary()
+        return get_token_collector().get_session_summary()
     
     def get_detailed_token_report(self) -> Dict[str, Any]:
         """Get a detailed token usage report."""
-        if not _token_collector:
+        if not get_token_collector:
             return {"error": "Token tracking not available"}
         
-        summary = _token_collector.get_session_summary()
-        recent = _token_collector.get_recent_interactions(limit=20)
+        collector = get_token_collector()
+        summary = collector.get_session_summary()
+        recent = collector.get_recent_interactions(limit=20)
         
         # Calculate cost estimates (example rates)
         cost_per_1k_input = 0.0005  # $0.0005 per 1K input tokens
@@ -1716,11 +1862,11 @@ class AgentTeam:
     
     def display_token_usage(self):
         """Display token usage in a formatted table."""
-        if not _token_collector:
+        if not get_token_collector:
             print("Token tracking not available")
             return
         
-        summary = _token_collector.get_session_summary()
+        summary = get_token_collector().get_session_summary()
         
         print("\n" + "="*50)
         print("TOKEN USAGE SUMMARY")

src/praisonai-agents/praisonaiagents/allowed_tools_filter.py

@@ -0,0 +1,257 @@
+"""
+Tool whitelist filter module with diagnostics.
+
+This module provides a canonical implementation of the ALLOWED_TOOLS filter
+that prevents tool name collisions in multi-environment agent systems by 
+whitelisting only specified tools.
+
+The filter is designed to solve the "tool shadowing" problem where multiple
+modules register tools with overlapping or ambiguous names, causing agents
+to invoke the wrong implementation.
+
+Environment Variables:
+    ALLOWED_TOOLS: Primary variable (comma-separated tool names)
+    HERMES_ONLY_TOOLS: Backward compatibility alias for ALLOWED_TOOLS
+
+Usage:
+    from praisonaiagents.allowed_tools_filter import AllowedToolsFilter
+    
+    # Create filter from environment variable  
+    filter = AllowedToolsFilter()
+    
+    # Filter tools
+    available_tools = {"search", "send_message", "extract_pdf"}
+    filtered_tools = filter.filter_tools(available_tools)
+    
+    # Get diagnostics
+    filter.log_diagnostics()
+"""
+
+import os
+import logging
+from typing import Any, Dict, List, Optional, Set, Union
+
+logger = logging.getLogger(__name__)
+
+
+class AllowedToolsFilter:
+    """
+    Canonical tool whitelist filter with diagnostics and backward compatibility.
+    
+    This filter implements the semantics for tool filtering:
+    - ALLOWED_TOOLS unset (or HERMES_ONLY_TOOLS): all tools visible (with warning about collisions)
+    - ALLOWED_TOOLS empty string: error - must specify tools or unset
+    - ALLOWED_TOOLS with values: only whitelisted tools visible
+    - Unknown tools in list: warn and strip in dev, strict fail in CI
+    
+    Backward Compatibility:
+    - Supports both ALLOWED_TOOLS and HERMES_ONLY_TOOLS
+    - ALLOWED_TOOLS takes precedence if both are set
+    """
+    
+    def __init__(self, env_var_name: str = "ALLOWED_TOOLS"):
+        """
+        Initialize the filter with backward compatibility support.
+        
+        Args:
+            env_var_name: Primary environment variable name (default: "ALLOWED_TOOLS")
+        """
+        # Support both new (ALLOWED_TOOLS) and legacy (HERMES_ONLY_TOOLS) naming
+        self.primary_var = env_var_name
+        self.legacy_var = "HERMES_ONLY_TOOLS"
+        
+        # ALLOWED_TOOLS takes precedence over HERMES_ONLY_TOOLS for backward compatibility
+        self.env_value = os.environ.get(self.primary_var) or os.environ.get(self.legacy_var)
+        self.env_var_name = self.primary_var if os.environ.get(self.primary_var) else self.legacy_var
+        
+        self.is_ci = os.environ.get("CI", "").lower() in ("true", "1", "yes")
+        self._whitelist: Optional[Set[str]] = self._parse_whitelist()
+        self._diagnostics: Dict[str, Any] = {}
+        
+    def _parse_whitelist(self) -> Optional[Set[str]]:
+        """
+        Parse the environment variable value into a whitelist set.
+        
+        Returns:
+            None if unset, Set of tool names if set, raises ValueError for empty string
+        """
+        if self.env_value is None:
+            return None
+        
+        # Empty string is an error - must be explicit
+        if self.env_value.strip() == "":
+            raise ValueError(
+                f"{self.env_var_name} cannot be empty. Either unset it for all tools "
+                "or provide a comma-separated list of tool names."
+            )
+        
+        # Parse comma-separated values, strip whitespace
+        tools = {name.strip() for name in self.env_value.split(",") if name.strip()}
+        
+        if not tools:
+            raise ValueError(
+                f"{self.env_var_name} contains no valid tool names. "
+                "Provide comma-separated tool names or unset the variable."
+            )
+        
+        return tools
+    
+    def filter_tools(self, available_tools: Union[Set[str], List[str], Dict[str, Any]]) -> Set[str]:
+        """
+        Filter tools based on ALLOWED_TOOLS whitelist.
+        
+        Args:
+            available_tools: Available tools as set, list, or dict with tool names as keys
+            
+        Returns:
+            Filtered set of tool names
+            
+        Raises:
+            ValueError: If unknown tools in CI mode or configuration errors
+        """
+        # Convert input to set of tool names
+        if isinstance(available_tools, dict):
+            tool_names = set(available_tools.keys())
+        else:
+            tool_names = set(available_tools)
+        
+        # Store for diagnostics
+        self._diagnostics["registered_before_filter"] = sorted(tool_names)
+        
+        # If no whitelist, return all tools (with warning)
+        if self._whitelist is None:
+            logger.warning(
+                "Tool whitelist is unset. All %d tools are visible. "
+                "Consider using ALLOWED_TOOLS to prevent tool name collisions.",
+                len(tool_names)
+            )
+            self._diagnostics["registered_after_filter"] = sorted(tool_names)
+            self._diagnostics["dropped_tools"] = []
+            self._diagnostics["unknown_tools"] = []
+            return tool_names
+        
+        # Find intersection and unknown tools
+        available_whitelisted = tool_names.intersection(self._whitelist)
+        unknown_tools = self._whitelist.difference(tool_names)
+        dropped_tools = tool_names.difference(self._whitelist)
+        
+        # Store for diagnostics
+        self._diagnostics["registered_after_filter"] = sorted(available_whitelisted)
+        self._diagnostics["dropped_tools"] = sorted(dropped_tools)
+        self._diagnostics["unknown_tools"] = sorted(unknown_tools)
+        
+        # Handle unknown tools based on environment
+        if unknown_tools:
+            unknown_list = sorted(unknown_tools)
+            if self.is_ci:
+                raise ValueError(
+                    f"Unknown tools in {self.env_var_name}: {unknown_list}. "
+                    "All specified tools must be available in CI mode."
+                )
+            else:
+                logger.warning(
+                    "Unknown tools in %s will be ignored: %s",
+                    self.env_var_name,
+                    unknown_list
+                )
+        
+        if dropped_tools:
+            logger.info(
+                "Dropped %d tools not in %s whitelist: %s",
+                len(dropped_tools),
+                self.env_var_name,
+                sorted(dropped_tools)
+            )
+        
+        logger.info(
+            "%s filtered tools: %d available, %d after filter",
+            self.env_var_name,
+            len(tool_names),
+            len(available_whitelisted)
+        )
+        
+        return available_whitelisted
+    
+    def log_diagnostics(self) -> None:
+        """Log startup diagnostics section as specified in the issue."""
+        if not self._diagnostics:
+            logger.warning("No diagnostics available. Call filter_tools() first.")
+            return
+        
+        logger.info("=" * 50)
+        logger.info("ALLOWED_TOOLS FILTER DIAGNOSTICS")
+        logger.info("=" * 50)
+        logger.info("%s=%s", self.env_var_name, self.env_value or "<unset>")
+        logger.info("RegisteredBeforeFilter=%s", self._diagnostics.get("registered_before_filter", []))
+        logger.info("RegisteredAfterFilter=%s", self._diagnostics.get("registered_after_filter", []))
+        logger.info("DroppedTools=%s", self._diagnostics.get("dropped_tools", []))
+        logger.info("UnknownTools=%s", self._diagnostics.get("unknown_tools", []))
+        logger.info("IsCI=%s", self.is_ci)
+        logger.info("=" * 50)
+    
+    def get_diagnostics(self) -> Dict[str, Any]:
+        """
+        Get diagnostics data as a dictionary.
+        
+        Returns:
+            Dictionary with diagnostics information
+        """
+        return {
+            "env_var_name": self.env_var_name,
+            "env_value": self.env_value,
+            "is_ci": self.is_ci,
+            "whitelist": list(self._whitelist) if self._whitelist else None,
+            **self._diagnostics
+        }
+    
+    def is_enabled(self) -> bool:
+        """
+        Check if filtering is enabled.
+        
+        Returns:
+            True if tool whitelist is set and filtering is active
+        """
+        return self._whitelist is not None
+    
+    def get_whitelist(self) -> Optional[Set[str]]:
+        """
+        Get the current whitelist.
+        
+        Returns:
+            Set of whitelisted tool names, or None if not set
+        """
+        return self._whitelist.copy() if self._whitelist else None
+
+
+def filter_tools_with_allowed_tools(
+    available_tools: Union[Set[str], List[str], Dict[str, Any]],
+    env_var_name: str = "ALLOWED_TOOLS",
+    log_diagnostics: bool = True
+) -> Set[str]:
+    """
+    Convenience function to filter tools with ALLOWED_TOOLS.
+    
+    Args:
+        available_tools: Available tools to filter
+        env_var_name: Environment variable name (default: "ALLOWED_TOOLS")
+        log_diagnostics: Whether to log diagnostics (default: True)
+        
+    Returns:
+        Filtered set of tool names
+    """
+    filter_instance = AllowedToolsFilter(env_var_name)
+    filtered = filter_instance.filter_tools(available_tools)
+    
+    if log_diagnostics:
+        filter_instance.log_diagnostics()
+    
+    return filtered
+
+
+# Backward compatibility aliases
+filter_tools_with_hermes = filter_tools_with_allowed_tools
+hermes_filter = filter_tools_with_allowed_tools
+apply_hermes_filter = filter_tools_with_allowed_tools
+
+# Legacy class alias for backward compatibility
+HermesToolFilter = AllowedToolsFilter

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

@@ -407,7 +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():
@@ -424,3 +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/config/feature_configs.py

@@ -42,6 +42,7 @@ class MemoryBackend(str, Enum):
     FILE = "file"
     SQLITE = "sqlite"
     REDIS = "redis"
+    VALKEY = "valkey"
     POSTGRES = "postgres"
     MEM0 = "mem0"
     MONGODB = "mongodb"

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/protocols.py

@@ -28,6 +28,7 @@ from typing import (
 )
 
 if TYPE_CHECKING:
+    from praisonai.gateway.pairing import PairedChannel
     from ..agent import Agent
 
 
@@ -758,3 +759,148 @@ def resolve_auth_mode(bind_host: str, configured: Optional[AuthMode] = 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/query_protocol.py

@@ -0,0 +1,13 @@
+"""Hooks query protocol for UI/API consumers."""
+
+from __future__ import annotations
+
+from typing import Dict, List, Protocol, runtime_checkable
+
+
+@runtime_checkable
+class HooksQueryProtocol(Protocol):
+    """Read path for registered hooks."""
+
+    def list_hooks_for_api(self) -> List[Dict]:
+        ...

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

@@ -439,3 +439,13 @@ def has_hook(event: Union[str, HookEvent]) -> bool:
             )
     
     return get_default_registry().has_hooks(event)
+
+
+def list_hooks_for_api() -> List[Dict]:
+    """Flat hook list suitable for REST/UI consumers."""
+    registry = get_default_registry()
+    flat: List[Dict] = []
+    for event, hooks in registry.list_hooks().items():
+        for h in hooks:
+            flat.append({"event": event, "source": "sdk", **h})
+    return flat

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/hooks/types.py

@@ -68,6 +68,14 @@ class HookEvent(str, Enum):
     SCHEDULE_REMOVE = "schedule_remove"
     SCHEDULE_TRIGGER = "schedule_trigger"
 
+    # Kanban task lifecycle (wrapper dispatcher + tools emit these)
+    KANBAN_TASK_CREATED = "kanban_task_created"
+    KANBAN_TASK_CLAIMED = "kanban_task_claimed"
+    KANBAN_TASK_MOVED = "kanban_task_moved"
+    KANBAN_TASK_DONE = "kanban_task_done"
+    KANBAN_TASK_BLOCKED = "kanban_task_blocked"
+    KANBAN_TASK_FAILED = "kanban_task_failed"
+
     # Claude Code parity events
     USER_PROMPT_SUBMIT = "user_prompt_submit"  # When user submits a prompt
     NOTIFICATION = "notification"              # When notification is sent
@@ -101,6 +109,34 @@ class HookInput:
         }
 
 
+@dataclass
+class KanbanHookInput(HookInput):
+    """Hook input for kanban task lifecycle events.
+    
+    Used by wrapper dispatcher and tools when emitting kanban events.
+    Observability adapters can subscribe to these via the hook registry.
+    """
+    task_id: str = ""
+    board: str = "default"
+    status: str = ""
+    assignee: Optional[str] = field(default=None)
+    from_status: Optional[str] = field(default=None)
+    to_status: Optional[str] = field(default=None)
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary for JSON serialization."""
+        result = super().to_dict()
+        result.update({
+            "task_id": self.task_id,
+            "board": self.board,
+            "status": self.status,
+            "assignee": self.assignee,
+            "from_status": self.from_status,
+            "to_status": self.to_status,
+        })
+        return result
+
+
 @dataclass
 class HookOutput:
     """Base hook output - common fields for all events."""

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

@@ -0,0 +1,23 @@
+"""
+Kanban protocols and types for PraisonAI Agents.
+
+This module provides the protocol contracts for kanban functionality,
+allowing the wrapper (praisonai) and PraisonAIUI to share a stable
+interface without coupling the core to SQLite or other heavy implementations.
+"""
+
+from praisonaiagents.kanban.protocols import (
+    KanbanStoreProtocol,
+    KanbanTaskProtocol,
+    KanbanCommentingProtocol,
+    KanbanLinkingProtocol,
+    VALID_KANBAN_STATUSES,
+)
+
+__all__ = [
+    "KanbanStoreProtocol",
+    "KanbanTaskProtocol",
+    "KanbanCommentingProtocol", 
+    "KanbanLinkingProtocol",
+    "VALID_KANBAN_STATUSES",
+]

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

@@ -0,0 +1,221 @@
+"""
+Kanban protocols for PraisonAI Agents.
+
+Defines the protocol contracts that the wrapper (praisonai) and PraisonAIUI
+can use to implement kanban functionality with a stable interface.
+
+This follows AGENTS.md §3.2 Protocol-First Design:
+- Protocols define WHAT (interface contract)
+- Implementations define HOW (concrete behavior)
+- Core SDK has protocols only, wrapper has heavy implementations
+"""
+
+from typing import Protocol, TypedDict, runtime_checkable
+
+
+# Valid kanban statuses matching PraisonAIUI columns + archived
+VALID_KANBAN_STATUSES: frozenset[str] = frozenset([
+    "triage",
+    "todo", 
+    "ready",
+    "running",
+    "blocked",
+    "review",
+    "done",
+    "archived"
+])
+
+
+class KanbanTaskProtocol(TypedDict, total=False):
+    """Typed dict shape for kanban task fields.
+    
+    Defines the expected structure of task objects returned by
+    KanbanStoreProtocol implementations.
+    """
+    id: str
+    title: str
+    body: str
+    status: str
+    assignee: str | None
+    priority: str | None
+    tenant: str | None
+    board: str
+    created_at: float
+    updated_at: float
+
+
+@runtime_checkable
+class KanbanStoreProtocol(Protocol):
+    """Protocol contract for kanban store implementations.
+    
+    This protocol defines the core interface that PraisonAIUI expects
+    for injected kanban stores. Wrapper implementations must
+    implement all methods to be compatible.
+    
+    Duck typing contract matches InjectedKanbanStore in PraisonAIUI.
+    """
+    
+    def get_board(
+        self,
+        *,
+        board: str = "default",
+        tenant: str | None = None,
+        include_archived: bool = False,
+    ) -> dict:
+        """Get board data with tasks grouped by status.
+        
+        Args:
+            board: Board name (default: "default")
+            tenant: Optional tenant filter
+            include_archived: Whether to include archived tasks
+            
+        Returns:
+            Board data with tasks organized by status columns
+        """
+        ...
+
+    def get_task(self, task_id: str) -> dict | None:
+        """Get a single task by ID.
+        
+        Args:
+            task_id: Task identifier
+            
+        Returns:
+            Task data or None if not found
+        """
+        ...
+
+    def create_task(self, data: dict) -> dict:
+        """Create a new task.
+        
+        Args:
+            data: Task creation data (title, body, status, etc.)
+            
+        Returns:
+            Created task data with assigned ID
+        """
+        ...
+
+    def update_task(self, task_id: str, data: dict) -> dict | None:
+        """Update an existing task.
+        
+        Args:
+            task_id: Task identifier
+            data: Fields to update
+            
+        Returns:
+            Updated task data or None if not found
+        """
+        ...
+
+    def move_task(self, task_id: str, status: str) -> dict | None:
+        """Move a task to a different status.
+        
+        Args:
+            task_id: Task identifier
+            status: New status (must be in VALID_KANBAN_STATUSES)
+            
+        Returns:
+            Updated task data or None if not found
+        """
+        ...
+
+    def bulk_update(self, task_ids: list[str], status: str) -> dict:
+        """Update multiple tasks to the same status.
+        
+        Args:
+            task_ids: List of task identifiers
+            status: New status for all tasks
+            
+        Returns:
+            Results summary with success/failure counts
+        """
+        ...
+
+    def delete_task(self, task_id: str) -> bool:
+        """Delete a task.
+        
+        Args:
+            task_id: Task identifier
+            
+        Returns:
+            True if deleted, False if not found
+        """
+        ...
+
+    def list_events(self, since: float = 0.0, board: str = "default") -> list[dict]:
+        """List kanban events since a timestamp.
+        
+        Args:
+            since: Unix timestamp to filter events from
+            board: Board name to filter events
+            
+        Returns:
+            List of event data
+        """
+        ...
+
+    def health(self) -> dict:
+        """Check store health status.
+        
+        Returns:
+            Health status information
+        """
+        ...
+
+
+@runtime_checkable
+class KanbanCommentingProtocol(Protocol):
+    """Extension protocol for kanban task commenting functionality.
+    
+    This protocol is implemented separately from KanbanStoreProtocol to allow
+    stores to optionally support commenting without breaking isinstance checks
+    on the core protocol.
+    """
+    
+    def add_comment(self, task_id: str, text: str, author: str | None = None) -> dict | None:
+        """Add a comment to a task.
+        
+        Args:
+            task_id: Task identifier
+            text: Comment text
+            author: Optional comment author
+            
+        Returns:
+            Comment data or None if task not found
+        """
+        ...
+
+
+@runtime_checkable 
+class KanbanLinkingProtocol(Protocol):
+    """Extension protocol for kanban task linking functionality.
+    
+    This protocol is implemented separately from KanbanStoreProtocol to allow
+    stores to optionally support task relationships without breaking isinstance
+    checks on the core protocol.
+    """
+    
+    def link_tasks(self, parent_id: str, child_id: str) -> bool:
+        """Link two tasks in parent-child relationship.
+        
+        Args:
+            parent_id: Parent task identifier
+            child_id: Child task identifier
+            
+        Returns:
+            True if linked successfully, False otherwise
+        """
+        ...
+
+    def unlink_tasks(self, parent_id: str, child_id: str) -> bool:
+        """Unlink parent-child task relationship.
+        
+        Args:
+            parent_id: Parent task identifier
+            child_id: Child task identifier
+            
+        Returns:
+            True if unlinked successfully, False otherwise
+        """
+        ...

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

@@ -174,7 +174,14 @@ class ChromaKnowledgeAdapter:
             
             # Add where filter if any conditions specified
             if where_filter:
-                query_kwargs["where"] = where_filter
+                # ChromaDB validation expects exactly one condition in the where clause
+                # When multiple conditions exist, use $and operator to combine them
+                if len(where_filter) == 1:
+                    query_kwargs["where"] = where_filter
+                else:
+                    # Convert multiple conditions to ChromaDB's $and format
+                    and_conditions = [{k: v} for k, v in where_filter.items()]
+                    query_kwargs["where"] = {"$and": and_conditions}
             
             response = self.collection.query(**query_kwargs)
             
@@ -197,6 +204,10 @@ class ChromaKnowledgeAdapter:
             return SearchResult(results=items)
             
         except Exception as e:
+            # Log ChromaDB query errors for better debugging
+            import logging
+            logger = logging.getLogger(__name__)
+            logger.warning(f"ChromaDB search failed: {e}")
             return SearchResult(results=[])
     
     def add(self, content: Any, *, user_id: Optional[str] = None, agent_id: Optional[str] = None,

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/error_classifier.py

@@ -6,6 +6,7 @@ error classification for better handling of different failure modes.
 """
 
 import re
+import random
 from enum import Enum
 from typing import Dict, Tuple, List, Optional
 
@@ -145,6 +146,9 @@ def should_retry(category: ErrorCategory) -> bool:
 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)
@@ -154,12 +158,10 @@ def get_retry_delay(category: ErrorCategory, attempt: int = 1, base_delay: float
         Delay in seconds, or 0 if should not retry
         
     Examples:
-        >>> get_retry_delay(ErrorCategory.RATE_LIMIT, attempt=1)
-        3.0
-        >>> get_retry_delay(ErrorCategory.TRANSIENT, attempt=3)
-        8.0
-        >>> get_retry_delay(ErrorCategory.AUTH, attempt=1)
-        0
+        >>> # 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)
 
@@ -167,16 +169,18 @@ def get_retry_delay(category: ErrorCategory, attempt: int = 1, base_delay: float
         return 0
     
     if category == ErrorCategory.RATE_LIMIT:
-        # Longer delay for rate limits to avoid hitting limits again
-        return min(base_delay * (3 ** attempt), 60.0)
+        # 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 (compression should be tried)
+        # Short delay for context limits (no jitter needed - not a contention issue)
         return base_delay * 0.5
     
     elif category == ErrorCategory.TRANSIENT:
-        # Exponential backoff for transient errors
-        return min(base_delay * (2 ** attempt), 30.0)
+        # 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
 

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
@@ -67,10 +73,10 @@ def _get_live():
 
 # Import token tracking
 try:
-    from ..telemetry.token_collector import TokenMetrics, _token_collector
+    from ..telemetry.token_collector import TokenMetrics, get_token_collector
 except ImportError:
     TokenMetrics = None
-    _token_collector = None
+    get_token_collector = None
 
 # Logging is already configured in _logging.py via __init__.py
 
@@ -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 = {}
@@ -685,8 +705,52 @@ Respond with ONLY a valid JSON tool call in this format:
             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)
@@ -707,16 +771,60 @@ 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:
                 category, can_retry, retry_delay = self._classify_error_and_should_retry(e, attempt + 1)
-                if not can_retry:
-                    raise
-
+                
                 last_error = e
                 error_str = str(e)
 
+                # 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"{category} error hit (attempt {attempt + 1}/{self._max_retries + 1}), "
@@ -746,7 +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
@@ -767,16 +875,60 @@ 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:
                 category, can_retry, retry_delay = self._classify_error_and_should_retry(e, attempt + 1)
-                if not can_retry:
-                    raise
-
+                
                 last_error = e
                 error_str = str(e)
 
+                # 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"{category} error hit (attempt {attempt + 1}/{self._max_retries + 1}), "
@@ -3401,7 +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,
@@ -3417,7 +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()
@@ -3436,7 +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,
@@ -4358,7 +4519,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
 
     def _track_token_usage(self, response: Dict[str, Any], model: str) -> Optional[TokenMetrics]:
         """Extract and track token usage from LLM response."""
-        if not TokenMetrics or not _token_collector:
+        if not TokenMetrics or not get_token_collector:
             return None
         
         # Note: metrics check moved to call sites for performance
@@ -4390,7 +4551,7 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
             # Track in global collector
             # Extract provider from model string (e.g., 'openai' from 'openai/gpt-4o-mini')
             provider = model.split('/')[0] if '/' in model else 'litellm'
-            _token_collector.track_tokens(
+            get_token_collector().track_tokens(
                 model=model,
                 agent=self.current_agent_name,
                 metrics=metrics,
@@ -4496,6 +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

@@ -69,12 +69,14 @@ class RateLimiter:
     _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)
@@ -93,6 +95,7 @@ class RateLimiter:
         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:
@@ -105,6 +108,7 @@ class RateLimiter:
         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:
@@ -184,15 +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.
@@ -208,17 +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.
@@ -274,23 +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).
@@ -315,14 +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/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/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:
@@ -1960,23 +2014,30 @@ 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 MongoDB client if it exists
+        # 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()

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

@@ -181,6 +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"""
@@ -828,6 +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 = []
@@ -854,22 +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
@@ -883,10 +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:
@@ -940,6 +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:
@@ -1068,6 +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.")
@@ -1464,7 +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):
@@ -1497,6 +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 = []
@@ -1523,7 +1603,7 @@ Provide a JSON with the structure:
    "agent_name": "<string>",
    "action": "<execute or stop>"
 }}
-"""
+""" + error_context
 
             try:
                 logging.info("Requesting manager instructions...")
@@ -1547,10 +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/rag/context.py

@@ -6,7 +6,13 @@ No heavy imports - only stdlib.
 """
 
 import hashlib
-from typing import Any, Dict, List, Optional, Set, Tuple
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Set, Tuple, Union
+
+if TYPE_CHECKING:
+    from ..knowledge.models import SearchResultItem
+
+
+ResultItem = Union[Dict[str, Any], "SearchResultItem"]
 
 
 def _estimate_tokens(text: str) -> int:
@@ -19,6 +25,39 @@ def _estimate_tokens(text: str) -> int:
     return len(text) // 4 + 1
 
 
+def _extract_value(item: ResultItem, key: str, default: Any = None) -> Any:
+    """
+    Extract value from either dict or object (e.g., SearchResultItem).
+    
+    Args:
+        item: Dictionary or object with attributes
+        key: Key/attribute name to extract
+        default: Default value if key not found
+        
+    Returns:
+        Extracted value or default
+    """
+    if isinstance(item, dict):
+        return item.get(key, default)
+    else:
+        return getattr(item, key, default)
+
+
+def _extract_metadata_value(
+    item: ResultItem,
+    metadata: Dict[str, Any],
+    key: str,
+    default: Any = "",
+) -> Any:
+    """
+    Extract metadata value with fallback to top-level item attribute.
+    """
+    value = metadata.get(key)
+    if value is None:
+        return _extract_value(item, key, default)
+    return value
+
+
 def _chunk_hash(text: str, source: Optional[str] = None) -> str:
     """
     Generate stable hash for chunk deduplication.
@@ -35,31 +74,38 @@ def _chunk_hash(text: str, source: Optional[str] = None) -> str:
 
 
 def deduplicate_chunks(
-    results: List[Dict[str, Any]],
+    results: List[ResultItem],
     similarity_threshold: float = 0.9,
-) -> List[Dict[str, Any]]:
+) -> List[ResultItem]:
     """
     Deduplicate chunks by content hash.
     
+    Handles both dict format and SearchResultItem objects.
+    
     Args:
-        results: List of search results with 'text' or 'memory' key
+        results: List of search results (dicts or SearchResultItem objects)
         similarity_threshold: Not used currently (hash-based dedup)
         
     Returns:
         Deduplicated list of results
     """
     seen_hashes: Set[str] = set()
-    unique_results: List[Dict[str, Any]] = []
+    unique_results: List[ResultItem] = []
     
     for result in results:
         # Skip None results
         if result is None:
             continue
-        # Handle different result formats
-        text = result.get("text") or result.get("memory", "")
+        
+        # Handle different result formats (dict or SearchResultItem)
+        text = _extract_value(result, "text") or _extract_value(result, "memory", "")
+        
         # CRITICAL: Handle metadata=None from mem0 - ensure always dict
-        metadata = result.get("metadata") or {}
-        source = metadata.get("source", "")
+        metadata = _extract_value(result, "metadata") or {}
+        if not isinstance(metadata, dict):
+            metadata = {}
+        
+        source = _extract_metadata_value(result, metadata, "source", "")
         
         chunk_id = _chunk_hash(text, source)
         
@@ -106,17 +152,19 @@ def truncate_context(
 
 
 def build_context(
-    results: List[Dict[str, Any]],
+    results: List[ResultItem],
     max_tokens: int = 4000,
     deduplicate: bool = True,
     separator: str = "\n\n---\n\n",
     include_source: bool = True,
-) -> Tuple[str, List[Dict[str, Any]]]:
+) -> Tuple[str, List[ResultItem]]:
     """
     Build context string from retrieval results.
     
+    Handles both dict format and SearchResultItem objects.
+    
     Args:
-        results: List of search results
+        results: List of search results (dicts or SearchResultItem objects)
         max_tokens: Maximum tokens for context
         deduplicate: Whether to deduplicate chunks
         separator: Separator between chunks
@@ -133,7 +181,7 @@ def build_context(
         results = deduplicate_chunks(results)
     
     context_parts: List[str] = []
-    used_results: List[Dict[str, Any]] = []
+    used_results: List[ResultItem] = []
     current_tokens = 0
     separator_tokens = _estimate_tokens(separator)
     
@@ -141,17 +189,21 @@ def build_context(
         # Skip None results
         if result is None:
             continue
-        # Handle different result formats
-        text = result.get("text") or result.get("memory", "")
+        
+        # Handle different result formats (dict or SearchResultItem)
+        text = _extract_value(result, "text") or _extract_value(result, "memory", "")
         if not text:
             continue
         
         # Build chunk text with optional source
         if include_source:
             # CRITICAL: Handle metadata=None from mem0 - ensure always dict
-            metadata = result.get("metadata") or {}
-            source = metadata.get("source", "")
-            filename = metadata.get("filename", "")
+            metadata = _extract_value(result, "metadata") or {}
+            if not isinstance(metadata, dict):
+                metadata = {}
+            
+            source = _extract_metadata_value(result, metadata, "source", "")
+            filename = _extract_metadata_value(result, metadata, "filename", "")
             source_label = filename or source or f"Source {i + 1}"
             chunk_text = f"[{source_label}]\n{text}"
         else:
@@ -194,11 +246,14 @@ class DefaultContextBuilder:
     
     def build(
         self,
-        results: List[Dict[str, Any]],
+        results: List[ResultItem],
         max_tokens: int = 4000,
         deduplicate: bool = True,
     ) -> str:
-        """Build context string from results."""
+        """Build context string from results.
+        
+        Handles both dict format and SearchResultItem objects.
+        """
         context, _ = build_context(
             results=results,
             max_tokens=max_tokens,

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

@@ -15,6 +15,8 @@ from .protocols import (
     ResourceLimits,
 )
 from .config import SandboxConfig, SecurityPolicy
+from .manager import SandboxManager
+from .security import SecurityWarning, check_code_safety, format_warnings, get_security_summary
 
 __all__ = [
     "SandboxProtocol",
@@ -22,5 +24,10 @@ __all__ = [
     "SandboxStatus",
     "ResourceLimits",
     "SandboxConfig",
-    "SecurityPolicy",
+    "SecurityPolicy", 
+    "SandboxManager",
+    "SecurityWarning",
+    "check_code_safety",
+    "format_warnings", 
+    "get_security_summary",
 ]

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

@@ -0,0 +1,302 @@
+"""
+SandboxManager for PraisonAI Agents.
+
+Factory and context manager for sandbox backends.
+Core SDK component that routes to appropriate sandbox implementations
+in the praisonai wrapper package.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import Dict, Any, Optional, TYPE_CHECKING
+
+from .config import SandboxConfig
+from .protocols import SandboxProtocol, SandboxResult
+
+if TYPE_CHECKING:
+    from typing import AsyncContextManager
+
+logger = logging.getLogger(__name__)
+
+
+class SandboxManager:
+    """Factory and context manager for all sandbox backends.
+    
+    Routes to appropriate sandbox implementations based on configuration.
+    Lightweight manager in core SDK - heavy implementations in wrapper.
+    
+    Example:
+        from praisonaiagents.sandbox import SandboxManager, SandboxConfig
+        
+        # Simple usage
+        config = SandboxConfig.docker("python:3.11-slim")
+        manager = SandboxManager(config)
+        result = await manager.run_code("print('Hello, World!')")
+        
+        # Context manager usage  
+        async with SandboxManager(config) as sandbox:
+            result = await sandbox.execute("print('Hello, World!')")
+    """
+    
+    def __init__(self, config: Optional[SandboxConfig] = None):
+        """Initialize the sandbox manager.
+        
+        Args:
+            config: Sandbox configuration. Defaults to subprocess sandbox.
+        """
+        self.config = config or SandboxConfig.subprocess()
+        self._sandbox: Optional[SandboxProtocol] = None
+        
+    async def __aenter__(self) -> SandboxProtocol:
+        """Async context manager entry."""
+        self._sandbox = await self._create_sandbox()
+        return self._sandbox
+        
+    async def __aexit__(self, exc_type, exc_val, exc_tb):
+        """Async context manager exit with cleanup."""
+        if self._sandbox:
+            try:
+                await self._sandbox.stop()
+                await self._sandbox.cleanup()
+            except Exception as e:
+                logger.warning(f"Error during sandbox cleanup: {e}")
+            finally:
+                self._sandbox = None
+    
+    async def run_code(
+        self, 
+        code: str,
+        language: str = "python",
+        **kwargs
+    ) -> SandboxResult:
+        """Convenience method: create sandbox, run code, cleanup.
+        
+        Args:
+            code: Code to execute
+            language: Programming language
+            **kwargs: Additional arguments passed to execute()
+            
+        Returns:
+            SandboxResult with execution details
+        """
+        async with self as sandbox:
+            return await sandbox.execute(code, language=language, **kwargs)
+    
+    async def _create_sandbox(self) -> SandboxProtocol:
+        """Create appropriate sandbox backend based on config.
+        
+        Routes to implementations in praisonai wrapper package.
+        
+        Returns:
+            Configured sandbox instance
+            
+        Raises:
+            ValueError: For unknown sandbox types
+            ImportError: If required dependencies not available
+        """
+        sandbox_type = self.config.sandbox_type.lower()
+        
+        if sandbox_type == "docker":
+            return await self._create_docker_sandbox()
+        elif sandbox_type in ("subprocess", "local"):
+            return await self._create_subprocess_sandbox()  
+        elif sandbox_type == "e2b":
+            return await self._create_e2b_sandbox()
+        elif sandbox_type == "sandlock":
+            return await self._create_sandlock_sandbox()
+        elif sandbox_type == "ssh":
+            return await self._create_ssh_sandbox()
+        elif sandbox_type == "modal":
+            return await self._create_modal_sandbox()
+        elif sandbox_type == "daytona":
+            return await self._create_daytona_sandbox()
+        else:
+            raise ValueError(
+                f"Unknown sandbox type: {sandbox_type!r}. "
+                f"Supported: 'docker', 'subprocess', 'e2b', 'sandlock', 'ssh', 'modal', 'daytona'"
+            )
+    
+    async def _create_docker_sandbox(self) -> SandboxProtocol:
+        """Create Docker sandbox."""
+        try:
+            # Lazy import from wrapper package
+            from praisonai.sandbox import DockerSandbox
+        except ImportError as e:
+            raise ImportError(
+                "Docker sandbox not available. Install with: "
+                "pip install praisonaiagents[sandbox-docker]"
+            ) from e
+        
+        sandbox = DockerSandbox(
+            image=self.config.image,
+            config=self.config,
+        )
+        
+        if not sandbox.is_available:
+            raise RuntimeError(
+                "Docker is not available. Please install Docker and ensure it's running."
+            )
+        
+        await sandbox.start()
+        return sandbox
+    
+    async def _create_subprocess_sandbox(self) -> SandboxProtocol:
+        """Create subprocess sandbox."""
+        try:
+            from praisonai.sandbox import SubprocessSandbox
+        except ImportError as e:
+            raise ImportError(
+                "Subprocess sandbox not available. This should not happen - "
+                "please check your installation."
+            ) from e
+        
+        sandbox = SubprocessSandbox(config=self.config)
+        await sandbox.start()
+        return sandbox
+    
+    async def _create_e2b_sandbox(self) -> SandboxProtocol:
+        """Create E2B sandbox."""
+        try:
+            from praisonai.sandbox import E2BSandbox
+        except ImportError as e:
+            raise ImportError(
+                "E2B sandbox not available. Install with: "
+                "pip install praisonaiagents[sandbox] e2b-code-interpreter"
+            ) from e
+        
+        sandbox = E2BSandbox(config=self.config)
+        
+        if not sandbox.is_available:
+            raise RuntimeError(
+                "E2B is not available. Please set E2B_API_KEY environment variable."
+            )
+        
+        await sandbox.start()
+        return sandbox
+    
+    async def _create_sandlock_sandbox(self) -> SandboxProtocol:
+        """Create Sandlock sandbox.""" 
+        try:
+            from praisonai.sandbox import SandlockSandbox
+        except ImportError as e:
+            raise ImportError(
+                "Sandlock sandbox not available. Install sandlock."
+            ) from e
+        
+        sandbox = SandlockSandbox(config=self.config)
+        await sandbox.start()
+        return sandbox
+        
+    async def _create_ssh_sandbox(self) -> SandboxProtocol:
+        """Create SSH sandbox."""
+        try:
+            from praisonai.sandbox import SSHSandbox
+        except ImportError as e:
+            raise ImportError(
+                "SSH sandbox not available. Install with: "
+                "pip install paramiko"
+            ) from e
+        
+        sandbox = SSHSandbox(config=self.config)
+        await sandbox.start()
+        return sandbox
+        
+    async def _create_modal_sandbox(self) -> SandboxProtocol:
+        """Create Modal sandbox."""
+        try:
+            from praisonai.sandbox import ModalSandbox
+        except ImportError as e:
+            raise ImportError(
+                "Modal sandbox not available. Install with: "
+                "pip install modal"
+            ) from e
+        
+        sandbox = ModalSandbox(config=self.config)
+        await sandbox.start()
+        return sandbox
+        
+    async def _create_daytona_sandbox(self) -> SandboxProtocol:
+        """Create Daytona sandbox."""
+        try:
+            from praisonai.sandbox import DaytonaSandbox
+        except ImportError as e:
+            raise ImportError(
+                "Daytona sandbox not available. Install daytona client."
+            ) from e
+        
+        sandbox = DaytonaSandbox(config=self.config)
+        await sandbox.start()
+        return sandbox
+
+    def get_available_types(self) -> Dict[str, Dict[str, Any]]:
+        """Get available sandbox types and their status.
+        
+        Returns:
+            Dictionary mapping sandbox types to availability info
+        """
+        types = {}
+        
+        # Check Docker
+        try:
+            from praisonai.sandbox import DockerSandbox
+            docker = DockerSandbox()
+            types["docker"] = {
+                "available": docker.is_available,
+                "description": "Isolated Docker containers",
+                "requires": ["docker"],
+            }
+        except ImportError:
+            types["docker"] = {
+                "available": False,
+                "description": "Isolated Docker containers", 
+                "requires": ["docker", "praisonaiagents[sandbox-docker]"],
+            }
+        
+        # Check subprocess (always available)
+        types["subprocess"] = {
+            "available": True,
+            "description": "Local subprocess (limited isolation)",
+            "requires": [],
+        }
+        
+        # Check E2B
+        try:
+            from praisonai.sandbox import E2BSandbox
+            e2b = E2BSandbox()
+            types["e2b"] = {
+                "available": e2b.is_available,
+                "description": "E2B cloud sandboxes",
+                "requires": ["e2b-code-interpreter", "E2B_API_KEY"],
+            }
+        except ImportError:
+            types["e2b"] = {
+                "available": False,
+                "description": "E2B cloud sandboxes",
+                "requires": ["e2b-code-interpreter", "E2B_API_KEY"],
+            }
+        
+        # Add other types with try/except blocks
+        for sandbox_type, module_name, description, requirements in [
+            ("sandlock", "SandlockSandbox", "OS-native sandboxing", ["sandlock"]),
+            ("ssh", "SSHSandbox", "Remote SSH execution", ["paramiko"]),
+            ("modal", "ModalSandbox", "Modal cloud compute", ["modal"]),
+            ("daytona", "DaytonaSandbox", "Daytona workspaces", ["daytona"]),
+        ]:
+            try:
+                from praisonai import sandbox
+                sandbox_class = getattr(sandbox, module_name)
+                # Check if class exists without instantiating (some need required args)
+                types[sandbox_type] = {
+                    "available": False,  # Can't easily check without required args
+                    "description": description,
+                    "requires": requirements,
+                }
+            except (ImportError, AttributeError):
+                types[sandbox_type] = {
+                    "available": False,
+                    "description": description,
+                    "requires": requirements,
+                }
+        
+        return types

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

@@ -0,0 +1,345 @@
+"""
+Security pre-checks for sandbox code execution.
+
+Provides pre-execution analysis of code for potentially dangerous patterns.
+Note: These are warnings only - the sandbox provides the real isolation.
+"""
+
+import re
+import ast
+import logging
+from typing import List, Optional, Dict, Any
+from dataclasses import dataclass
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass 
+class SecurityWarning:
+    """A security warning from code analysis."""
+    
+    pattern: str
+    message: str
+    severity: str  # "low", "medium", "high", "critical"
+    line_number: Optional[int] = None
+    context: Optional[str] = None
+
+
+# Dangerous patterns to detect in code
+DANGEROUS_PATTERNS = [
+    # System/OS access
+    (r"import\s+os\s*;?\s*os\.system", "Direct system command execution via os.system", "high"),
+    (r"subprocess\.call.*shell\s*=\s*True", "Shell injection risk with subprocess", "high"),
+    (r"subprocess\.run.*shell\s*=\s*True", "Shell injection risk with subprocess", "high"),
+    (r"subprocess\.Popen.*shell\s*=\s*True", "Shell injection risk with subprocess", "high"),
+    (r"__import__\(['\"]os['\"]\)", "Dynamic import of os module", "medium"),
+    (r"eval\s*\(", "Dynamic code execution with eval()", "critical"),
+    (r"exec\s*\(", "Dynamic code execution with exec()", "critical"),
+    (r"compile\s*\(.*exec", "Code compilation for execution", "high"),
+    
+    # File system access
+    (r"open\s*\(.*['\"]\/etc", "Access to system configuration files", "medium"),
+    (r"open\s*\(.*['\"]\/root", "Access to root directory", "high"),
+    (r"open\s*\(.*['\"]\/home", "Access to user directories", "medium"),
+    (r"shutil\.rmtree", "Recursive directory deletion", "high"),
+    (r"os\.remove\s*\(", "File deletion", "medium"),
+    (r"os\.unlink\s*\(", "File deletion", "medium"),
+    (r"pathlib\.Path.*unlink", "File deletion via pathlib", "medium"),
+    
+    # Network access
+    (r"socket\.socket", "Direct socket creation", "medium"),
+    (r"urllib\.request", "HTTP requests via urllib", "low"),
+    (r"requests\.", "HTTP requests via requests library", "low"),
+    (r"http\.client", "HTTP client usage", "low"),
+    
+    # Process manipulation
+    (r"os\.fork\s*\(", "Process forking", "high"),
+    (r"multiprocessing\.", "Multiprocessing usage", "medium"),
+    (r"threading\.Thread", "Thread creation", "low"),
+    
+    # Potentially dangerous builtins
+    (r"getattr\s*\(.*__", "Dynamic attribute access to dunder methods", "medium"),
+    (r"setattr\s*\(.*__", "Dynamic attribute setting of dunder methods", "medium"),
+    (r"hasattr\s*\(.*__", "Dynamic attribute checking of dunder methods", "low"),
+    (r"globals\s*\(\)", "Access to global namespace", "medium"),
+    (r"locals\s*\(\)", "Access to local namespace", "low"),
+    (r"vars\s*\(", "Variable inspection", "low"),
+    (r"dir\s*\(", "Object inspection", "low"),
+    
+    # Code modification
+    (r"sys\.modules", "Module manipulation", "high"),
+    (r"importlib\.", "Dynamic import manipulation", "medium"),
+    (r"__builtins__", "Access to builtin functions", "high"),
+    
+    # Memory/performance
+    (r"while\s+True:", "Infinite loop (potential DoS)", "medium"),
+    (r"for.*in.*range\s*\(\s*[0-9]{6,}", "Large range iteration", "low"),
+    (r"list\s*\(\s*range\s*\(\s*[0-9]{6,}", "Large list creation", "low"),
+    (r"\*\s*[0-9]{6,}", "Large data multiplication", "low"),
+]
+
+
+def check_code_safety(code: str, language: str = "python") -> List[SecurityWarning]:
+    """Analyze code for potentially dangerous patterns.
+    
+    Args:
+        code: Code to analyze
+        language: Programming language (currently only 'python' supported)
+        
+    Returns:
+        List of security warnings found
+        
+    Note:
+        These are warnings only. The sandbox provides real isolation.
+        This is a best-effort static analysis for user awareness.
+    """
+    warnings = []
+    
+    if language == "python":
+        warnings.extend(_check_python_patterns(code))
+        warnings.extend(_check_python_ast(code))
+    elif language == "bash":
+        warnings.extend(_check_bash_patterns(code))
+    else:
+        # Fall back to basic pattern matching
+        warnings.extend(_check_generic_patterns(code))
+    
+    return warnings
+
+
+def _check_python_patterns(code: str) -> List[SecurityWarning]:
+    """Check Python code using regex patterns."""
+    warnings = []
+    lines = code.split('\n')
+    
+    for i, line in enumerate(lines, 1):
+        for pattern, message, severity in DANGEROUS_PATTERNS:
+            if re.search(pattern, line, re.IGNORECASE):
+                warnings.append(SecurityWarning(
+                    pattern=pattern,
+                    message=message,
+                    severity=severity,
+                    line_number=i,
+                    context=line.strip(),
+                ))
+    
+    return warnings
+
+
+def _check_python_ast(code: str) -> List[SecurityWarning]:
+    """Check Python code using AST analysis."""
+    warnings = []
+    
+    try:
+        tree = ast.parse(code)
+        visitor = DangerousASTVisitor()
+        visitor.visit(tree)
+        warnings.extend(visitor.warnings)
+    except SyntaxError:
+        # Invalid Python syntax - let the sandbox handle it
+        pass
+    except Exception as e:
+        logger.debug(f"AST analysis failed: {e}")
+    
+    return warnings
+
+
+def _check_bash_patterns(code: str) -> List[SecurityWarning]:
+    """Check bash/shell code for dangerous patterns."""
+    warnings = []
+    lines = code.split('\n')
+    
+    bash_patterns = [
+        (r"rm\s+-rf\s+\/", "Recursive deletion of root directory", "critical"),
+        (r"rm\s+-rf\s+\*", "Recursive deletion with wildcard", "high"),
+        (r">\s*\/dev\/sd[a-z]", "Direct disk device access", "critical"),
+        (r"dd\s+if=.*of=", "Low-level disk operations", "high"),
+        (r"mkfs\.", "Filesystem creation", "critical"),
+        (r"fdisk", "Disk partitioning", "critical"),
+        (r"shutdown", "System shutdown", "high"),
+        (r"reboot", "System reboot", "high"),
+        (r"halt", "System halt", "high"),
+        (r"init\s+[0-6]", "System runlevel change", "high"),
+        (r"chmod\s+777", "Overly permissive file permissions", "medium"),
+        (r"curl.*\|\s*bash", "Piped execution from network", "high"),
+        (r"wget.*\|\s*bash", "Piped execution from network", "high"),
+        (r"\$\(.*\)", "Command substitution", "medium"),
+        (r"`.*`", "Command substitution (backticks)", "medium"),
+    ]
+    
+    for i, line in enumerate(lines, 1):
+        for pattern, message, severity in bash_patterns:
+            if re.search(pattern, line, re.IGNORECASE):
+                warnings.append(SecurityWarning(
+                    pattern=pattern,
+                    message=message,
+                    severity=severity,
+                    line_number=i,
+                    context=line.strip(),
+                ))
+    
+    return warnings
+
+
+def _check_generic_patterns(code: str) -> List[SecurityWarning]:
+    """Check code using generic patterns."""
+    warnings = []
+    lines = code.split('\n')
+    
+    generic_patterns = [
+        (r"eval\s*\(", "Dynamic code execution", "critical"),
+        (r"exec\s*\(", "Dynamic code execution", "critical"),
+        (r"system\s*\(", "System command execution", "high"),
+        (r"shell\s*\(", "Shell command execution", "high"),
+    ]
+    
+    for i, line in enumerate(lines, 1):
+        for pattern, message, severity in generic_patterns:
+            if re.search(pattern, line, re.IGNORECASE):
+                warnings.append(SecurityWarning(
+                    pattern=pattern,
+                    message=message,
+                    severity=severity,
+                    line_number=i,
+                    context=line.strip(),
+                ))
+    
+    return warnings
+
+
+class DangerousASTVisitor(ast.NodeVisitor):
+    """AST visitor to detect dangerous Python constructs."""
+    
+    def __init__(self):
+        self.warnings = []
+    
+    def visit_Call(self, node):
+        """Visit function calls."""
+        # Check for dangerous function calls
+        if isinstance(node.func, ast.Name):
+            func_name = node.func.id
+            if func_name in ("eval", "exec", "compile"):
+                self.warnings.append(SecurityWarning(
+                    pattern=f"{func_name}()",
+                    message=f"Dynamic code execution with {func_name}()",
+                    severity="critical",
+                    line_number=node.lineno,
+                ))
+        
+        # Check for attribute calls like os.system
+        elif isinstance(node.func, ast.Attribute):
+            if isinstance(node.func.value, ast.Name):
+                module = node.func.value.id
+                func = node.func.attr
+                
+                if module == "os" and func == "system":
+                    self.warnings.append(SecurityWarning(
+                        pattern="os.system()",
+                        message="Direct system command execution",
+                        severity="high",
+                        line_number=node.lineno,
+                    ))
+                elif module == "subprocess" and func in ("call", "run", "Popen"):
+                    # Check if shell=True is used
+                    for keyword in node.keywords:
+                        if keyword.arg == "shell" and isinstance(keyword.value, ast.Constant):
+                            if keyword.value.value is True:
+                                self.warnings.append(SecurityWarning(
+                                    pattern=f"subprocess.{func}(..., shell=True)",
+                                    message="Shell injection risk with subprocess",
+                                    severity="high",
+                                    line_number=node.lineno,
+                                ))
+        
+        self.generic_visit(node)
+    
+    def visit_Import(self, node):
+        """Visit import statements."""
+        for alias in node.names:
+            if alias.name in ("os", "subprocess", "sys"):
+                self.warnings.append(SecurityWarning(
+                    pattern=f"import {alias.name}",
+                    message=f"Import of potentially dangerous module: {alias.name}",
+                    severity="low",
+                    line_number=node.lineno,
+                ))
+        
+        self.generic_visit(node)
+    
+    def visit_ImportFrom(self, node):
+        """Visit from ... import statements."""
+        if node.module in ("os", "subprocess", "sys"):
+            for alias in node.names:
+                if alias.name in ("system", "call", "run", "Popen", "modules"):
+                    self.warnings.append(SecurityWarning(
+                        pattern=f"from {node.module} import {alias.name}",
+                        message=f"Import of dangerous function: {alias.name}",
+                        severity="medium",
+                        line_number=node.lineno,
+                    ))
+        
+        self.generic_visit(node)
+
+
+def format_warnings(warnings: List[SecurityWarning]) -> str:
+    """Format security warnings for display.
+    
+    Args:
+        warnings: List of security warnings
+        
+    Returns:
+        Formatted string of warnings
+    """
+    if not warnings:
+        return "No security issues detected."
+    
+    # Group by severity
+    by_severity = {"critical": [], "high": [], "medium": [], "low": []}
+    for warning in warnings:
+        by_severity[warning.severity].append(warning)
+    
+    lines = []
+    lines.append(f"Security analysis found {len(warnings)} potential issue(s):")
+    lines.append("")
+    
+    for severity in ["critical", "high", "medium", "low"]:
+        if by_severity[severity]:
+            lines.append(f"{severity.upper()} RISK:")
+            for warning in by_severity[severity]:
+                line_info = f" (line {warning.line_number})" if warning.line_number else ""
+                context_info = f"\n  Context: {warning.context}" if warning.context else ""
+                lines.append(f"  - {warning.message}{line_info}{context_info}")
+            lines.append("")
+    
+    lines.append("Note: These are warnings only. The sandbox provides real isolation.")
+    return "\n".join(lines)
+
+
+def get_security_summary(warnings: List[SecurityWarning]) -> Dict[str, Any]:
+    """Get security analysis summary.
+    
+    Args:
+        warnings: List of security warnings
+        
+    Returns:
+        Dictionary with summary statistics
+    """
+    by_severity = {"critical": 0, "high": 0, "medium": 0, "low": 0}
+    for warning in warnings:
+        by_severity[warning.severity] += 1
+    
+    max_severity = "low"
+    if by_severity["critical"] > 0:
+        max_severity = "critical"
+    elif by_severity["high"] > 0:
+        max_severity = "high"  
+    elif by_severity["medium"] > 0:
+        max_severity = "medium"
+    
+    return {
+        "total_warnings": len(warnings),
+        "by_severity": by_severity,
+        "max_severity": max_severity,
+        "is_safe": len(warnings) == 0,
+        "has_critical": by_severity["critical"] > 0,
+    }

src/praisonai-agents/praisonaiagents/scheduler/README.md

@@ -0,0 +1,29 @@
+# Scheduler — runner vs loop
+
+## ScheduleRunner (stateless)
+
+`ScheduleRunner` is **stateless**. Call `get_due_jobs()` whenever you want to check for jobs that should fire. It does not run background threads.
+
+Use when:
+- CLI one-shot tick (`praisonai schedule tick`)
+- Host integration that polls on its own timer
+- Tests
+
+## ScheduleLoop (optional daemon)
+
+`ScheduleLoop` is an **optional background daemon**. You must call `.start()` explicitly; it polls the runner and invokes `on_trigger` callbacks.
+
+Use when:
+- Long-running host process (Pattern B/C via `setup_bridges()`)
+- Standalone scheduler service
+
+```python
+from praisonaiagents.scheduler import FileScheduleStore, ScheduleRunner, ScheduleLoop
+
+store = FileScheduleStore()
+runner = ScheduleRunner(store)
+loop = ScheduleLoop(runner)
+loop.start()  # optional daemon
+```
+
+The wrapper host calls `ensure_schedule_runner()` which starts `ScheduleLoop` when the scheduler optional dependency is available.

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/README.md

@@ -32,7 +32,22 @@ When you provide a `session_id` to an Agent:
 
 ### Default Storage Location
 
-Sessions are stored in: `~/.praison/sessions/{session_id}.json`
+Sessions are stored in: `~/.praisonai/sessions/{session_id}.json`
+
+### Session metadata contract (Issue #48)
+
+List endpoints and UI dashboards may expose these optional root-level fields:
+
+| Field | Type | Description |
+|-------|------|-------------|
+| `model` | string | LLM model used in the session |
+| `total_tokens` | int | Cumulative input+output tokens |
+| `cost` | float | Estimated USD cost |
+| `agent_id` | string | Gateway or registry agent id |
+| `source` | string | Origin: `chat`, `gateway`, `cli`, `api` |
+| `agent_name` | string | Human-readable agent name |
+
+Fields are persisted in session JSON `metadata` and mirrored at root on save. Populated automatically after each assistant turn when using the default session store.
 
 ### Session File Format
 
@@ -45,7 +60,13 @@ Sessions are stored in: `~/.praison/sessions/{session_id}.json`
   ],
   "created_at": "2026-01-02T04:00:00+00:00",
   "updated_at": "2026-01-02T04:01:00+00:00",
-  "agent_name": "Assistant"
+  "agent_name": "Assistant",
+  "agent_id": "agent-abc",
+  "source": "chat",
+  "model": "gpt-4o-mini",
+  "total_tokens": 128,
+  "cost": 0.0004,
+  "metadata": {}
 }
 ```
 

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

@@ -144,6 +144,35 @@ class HierarchicalSessionStore(DefaultSessionStore):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self._extended_cache: Dict[str, ExtendedSessionData] = {}
+
+    def _load_session_from_disk(self, session_id: str, filepath: str) -> ExtendedSessionData:
+        """Load extended session JSON from disk (caller must hold FileLock)."""
+        if os.path.exists(filepath):
+            try:
+                with open(filepath, "r", encoding="utf-8") as f:
+                    data = json.load(f)
+                return ExtendedSessionData.from_dict(data)
+            except (json.JSONDecodeError, IOError):
+                pass
+        return ExtendedSessionData(session_id=session_id)
+
+    def _modify_session_locked(
+        self,
+        session_id: str,
+        mutator,
+        *,
+        error_label: str = "modify session",
+    ) -> bool:
+        """Locked read-modify-write preserving extended session fields."""
+        result = super()._modify_session_locked(
+            session_id, mutator, error_label=error_label
+        )
+        if result:
+            with self._lock:
+                cached = self._cache.get(session_id)
+                if isinstance(cached, ExtendedSessionData):
+                    self._extended_cache[session_id] = cached
+        return result
     
     def add_message(
         self,
@@ -157,25 +186,22 @@ class HierarchicalSessionStore(DefaultSessionStore):
         
         Overrides parent to preserve extended session data.
         """
-        from .store import SessionMessage
-        
-        # Load extended session (force reload to get latest)
-        session = self._load_extended_session(session_id, force_reload=True)
-        
+
         message = SessionMessage(
             role=role,
             content=content,
             timestamp=time.time(),
             metadata=metadata or {},
         )
-        
-        session.messages.append(message)
-        
-        # Trim messages if over limit
-        if len(session.messages) > self.max_messages:
-            session.messages = session.messages[-self.max_messages:]
-        
-        return self._save_extended_session(session)
+
+        def _apply(session: SessionData) -> None:
+            session.messages.append(message)
+            if len(session.messages) > self.max_messages:
+                session.messages = session.messages[-self.max_messages :]
+
+        return self._modify_session_locked(
+            session_id, _apply, error_label="add message to session"
+        )
     
     def _load_extended_session(self, session_id: str, force_reload: bool = False) -> ExtendedSessionData:
         """Load extended session from disk."""

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/protocols.py

@@ -129,3 +129,16 @@ class SessionStoreProtocol(Protocol):
             True if the session exists.
         """
         ...
+
+
+@runtime_checkable
+class CheckpointQueryProtocol(Protocol):
+    """Read path for session checkpoints / rollback snapshots."""
+
+    def list_checkpoints(self, session_id: str) -> List[Dict[str, Any]]:
+        """Return checkpoint metadata for a session."""
+        ...
+
+    def get_checkpoint(self, session_id: str, checkpoint_id: str) -> Optional[Dict[str, Any]]:
+        """Return a single checkpoint payload."""
+        ...

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
@@ -14,14 +13,24 @@ 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
+from typing import Any, Callable, Dict, List, Optional
 
 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())
 
@@ -72,7 +81,7 @@ class SessionData:
     
     def to_dict(self) -> Dict[str, Any]:
         """Convert to dictionary for JSON serialization."""
-        return {
+        data = {
             "session_id": self.session_id,
             "messages": [m.to_dict() for m in self.messages],
             "created_at": self.created_at,
@@ -83,6 +92,10 @@ class SessionData:
             "gateway_session_id": self.gateway_session_id,
             "agent_id": self.agent_id,
         }
+        for key in ("model", "llm", "total_tokens", "token_count", "cost", "source"):
+            if key in self.metadata:
+                data[key] = self.metadata[key]
+        return data
     
     @classmethod
     def from_dict(cls, data: Dict[str, Any]) -> "SessionData":
@@ -128,7 +141,8 @@ class FileLock:
         self._lock_path = filepath + ".lock"
     
     def __enter__(self):
-        self.acquire()
+        if not self.acquire():
+            raise IOError(f"Failed to acquire file lock for {self.filepath} after {self.timeout}s")
         return self
     
     def __exit__(self, exc_type, exc_val, exc_tb):
@@ -150,7 +164,17 @@ 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 +195,9 @@ 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:
@@ -265,6 +291,72 @@ class DefaultSessionStore:
             self._cache[session_id] = session
         
         return session
+
+    def _load_session_from_disk(self, session_id: str, filepath: str) -> SessionData:
+        """Load session JSON from disk (caller must hold FileLock)."""
+        if os.path.exists(filepath):
+            try:
+                with open(filepath, "r", encoding="utf-8") as f:
+                    data = json.load(f)
+                return SessionData.from_dict(data)
+            except (json.JSONDecodeError, IOError):
+                pass
+        return SessionData(session_id=session_id)
+
+    def _read_session_fresh(self, session_id: str) -> SessionData:
+        """Reload session from disk and refresh the in-process cache."""
+        filepath = self._get_session_path(session_id)
+        with FileLock(filepath, self.lock_timeout):
+            session = self._load_session_from_disk(session_id, filepath)
+        with self._lock:
+            self._cache[session_id] = session
+        return session
+
+    def _modify_session_locked(
+        self,
+        session_id: str,
+        mutator: Callable[[SessionData], None],
+        *,
+        error_label: str = "modify session",
+    ) -> bool:
+        """Apply mutator after reloading from disk under FileLock."""
+        filepath = self._get_session_path(session_id)
+
+        with FileLock(filepath, self.lock_timeout):
+            session = self._load_session_from_disk(session_id, filepath)
+            mutator(session)
+            session.updated_at = datetime.now(timezone.utc).isoformat()
+
+            if len(session.messages) > self.max_messages:
+                session.messages = session.messages[-self.max_messages:]
+
+            try:
+                dir_path = os.path.dirname(filepath) or "."
+                os.makedirs(dir_path, exist_ok=True)
+                with tempfile.NamedTemporaryFile(
+                    mode="w",
+                    encoding="utf-8",
+                    dir=dir_path,
+                    delete=False,
+                    suffix=".tmp",
+                ) as f:
+                    json.dump(session.to_dict(), f, indent=2, ensure_ascii=False)
+                    temp_path = f.name
+
+                os.replace(temp_path, filepath)
+
+                with self._lock:
+                    self._cache[session_id] = session
+
+                return True
+            except (IOError, OSError) as e:
+                logger.error(f"Failed to {error_label} {session_id}: {e}")
+                try:
+                    if "temp_path" in locals():
+                        os.remove(temp_path)
+                except (IOError, OSError):
+                    pass
+                return False
     
     def _save_session(self, session: SessionData) -> bool:
         """Save session to disk with atomic write."""
@@ -414,13 +506,13 @@ class DefaultSessionStore:
         Returns:
             List of {"role": "user/assistant", "content": "..."} dicts.
         """
-        session = self._load_session(session_id)
+        session = self._read_session_fresh(session_id)
         limit = max_messages or self.max_messages
         return session.get_chat_history(limit)
     
     def get_session(self, session_id: str) -> SessionData:
         """Get full session data."""
-        return self._load_session(session_id)
+        return self._read_session_fresh(session_id)
     
     def set_agent_info(
         self,
@@ -429,27 +521,42 @@ class DefaultSessionStore:
         user_id: Optional[str] = None,
     ) -> bool:
         """Set agent info for a session."""
-        session = self._load_session(session_id)
-        
-        with self._lock:
+
+        def _apply(session: SessionData) -> None:
             if agent_name:
                 session.agent_name = agent_name
             if user_id:
                 session.user_id = user_id
-            self._cache[session_id] = session
-        
-        return self._save_session(session)
+
+        return self._modify_session_locked(
+            session_id, _apply, error_label="set agent info for session"
+        )
     
     def clear_session(self, session_id: str) -> bool:
         """Clear all messages from a session."""
-        session = self._load_session(session_id)
-        
-        with self._lock:
-            session.messages.clear()
-            self._cache[session_id] = session
-        
-        return self._save_session(session)
-    
+        return self._modify_session_locked(
+            session_id,
+            lambda session: session.messages.clear(),
+            error_label="clear session",
+        )
+
+    def update_session_metadata(self, session_id: str, **fields: Any) -> bool:
+        """Merge run stats / metadata fields into a persisted session."""
+        if not fields:
+            return True
+
+        def _apply(session: SessionData) -> None:
+            for key, value in fields.items():
+                if value is None:
+                    continue
+                session.metadata[key] = value
+                if key in ("agent_id", "agent_name", "user_id"):
+                    setattr(session, key, value)
+
+        return self._modify_session_locked(
+            session_id, _apply, error_label="update session metadata"
+        )
+
     def delete_session(self, session_id: str) -> bool:
         """Delete a session completely."""
         filepath = self._get_session_path(session_id)
@@ -479,10 +586,16 @@ class DefaultSessionStore:
                             data = json.load(f)
                         sessions.append({
                             "session_id": data.get("session_id", filename[:-5]),
+                            "id": data.get("session_id", filename[:-5]),
                             "agent_name": data.get("agent_name"),
+                            "agent_id": data.get("agent_id") or (data.get("metadata") or {}).get("agent_id"),
+                            "source": data.get("source") or (data.get("metadata") or {}).get("source"),
                             "created_at": data.get("created_at"),
                             "updated_at": data.get("updated_at"),
                             "message_count": len(data.get("messages", [])),
+                            "model": data.get("model") or data.get("llm") or (data.get("metadata") or {}).get("model"),
+                            "total_tokens": data.get("total_tokens") or data.get("token_count") or (data.get("metadata") or {}).get("total_tokens"),
+                            "cost": data.get("cost") or (data.get("metadata") or {}).get("cost"),
                         })
                     except (json.JSONDecodeError, IOError):
                         continue
@@ -542,7 +655,7 @@ class DefaultSessionStore:
             List of SessionData objects for the specified agent
         """
         session_ids = self.list_sessions_by_agent(agent_name, limit)
-        return [self._load_session(sid) for sid in session_ids]
+        return [self._read_session_fresh(sid) for sid in session_ids]
     
     def get_agent_chat_history(
         self,
@@ -596,16 +709,15 @@ class DefaultSessionStore:
         Returns:
             True if saved successfully
         """
-        session = self._load_session(session_id)
-        
-        with self._lock:
+        def _apply(session: SessionData) -> None:
             if gateway_session_id:
                 session.gateway_session_id = gateway_session_id
             if agent_id:
                 session.agent_id = agent_id
-            self._cache[session_id] = session
-        
-        return self._save_session(session)
+
+        return self._modify_session_locked(
+            session_id, _apply, error_label="set gateway info for session"
+        )
     
     def get_by_gateway_session(self, gateway_session_id: str) -> Optional[SessionData]:
         """Get session data linked to a gateway session.

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

@@ -19,7 +19,9 @@ Usage:
 __all__ = [
     # Models
     "SkillProperties",
-    "SkillMetadata",
+    "SkillMetadata", 
+    "SkillRequirements",
+    "SkillState",
     "ParseError",
     "ValidationError",
     # Parser
@@ -54,13 +56,17 @@ __all__ = [
     "SkillActivatedEvent",
     # Budget
     "SkillPromptBudget",
+    # Capability validation
+    "CapabilityValidator",
+    "EnforcementLevel",
+    "ValidationResult",
 ]
 
 
 def __getattr__(name: str):
     """Lazy load module components to avoid import overhead."""
-    if name in ("SkillProperties", "SkillMetadata", "ParseError", "ValidationError"):
-        from .models import SkillProperties, SkillMetadata, ParseError, ValidationError
+    if name in ("SkillProperties", "SkillMetadata", "SkillRequirements", "SkillState", "ParseError", "ValidationError"):
+        from .models import SkillProperties, SkillMetadata, SkillRequirements, SkillState, ParseError, ValidationError
         return locals()[name]
     
     if name in ("parse_frontmatter", "find_skill_md", "read_properties"):
@@ -130,5 +136,9 @@ def __getattr__(name: str):
             return None
 
         return load_skill
+    
+    if name in ("CapabilityValidator", "EnforcementLevel", "ValidationResult"):
+        from .capability_validator import CapabilityValidator, EnforcementLevel, ValidationResult
+        return locals()[name]
 
     raise AttributeError(f"module {__name__!r} has no attribute {name!r}")

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

@@ -0,0 +1,236 @@
+"""Capability validation for Agent Skills.
+
+This module implements the capability gate enforcement system for skills,
+checking that required tools, servers, and environment variables are available
+before allowing skill activation.
+"""
+
+import logging
+import os
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Set, Tuple
+from enum import Enum
+
+from .models import SkillProperties, SkillRequirements, SkillState
+
+logger = logging.getLogger(__name__)
+
+
+class EnforcementLevel(Enum):
+    """Enforcement level for capability gates."""
+    DISABLED = "disabled"      # No enforcement (existing behavior)
+    TELEMETRY = "telemetry"    # Log only, no blocking
+    WARN = "warn"              # Warning but allow activation
+    STRICT = "strict"          # Hard failure, block activation
+
+
+@dataclass
+class ValidationResult:
+    """Result of capability validation for a skill."""
+    skill_name: str
+    state: SkillState
+    satisfied_tools: List[str]
+    missing_tools: List[str]
+    satisfied_servers: List[str]
+    missing_servers: List[str]
+    satisfied_env_vars: List[str]
+    missing_env_vars: List[str]
+    warnings: List[str]
+    errors: List[str]
+    
+    @property
+    def is_fully_satisfied(self) -> bool:
+        """Check if all requirements are satisfied."""
+        return not any([
+            self.missing_tools,
+            self.missing_servers, 
+            self.missing_env_vars,
+            self.errors
+        ])
+    
+    @property
+    def has_critical_missing(self) -> bool:
+        """Check if critical requirements are missing (tools, servers)."""
+        return bool(self.missing_tools or self.missing_servers)
+    
+    def to_dict(self) -> Dict:
+        """Convert to dictionary for serialization."""
+        return {
+            "skill_name": self.skill_name,
+            "state": self.state.value,
+            "satisfied_tools": self.satisfied_tools,
+            "missing_tools": self.missing_tools,
+            "satisfied_servers": self.satisfied_servers,
+            "missing_servers": self.missing_servers,
+            "satisfied_env_vars": self.satisfied_env_vars,
+            "missing_env_vars": self.missing_env_vars,
+            "warnings": self.warnings,
+            "errors": self.errors,
+            "is_fully_satisfied": self.is_fully_satisfied,
+            "has_critical_missing": self.has_critical_missing
+        }
+
+
+class CapabilityValidator:
+    """Validates skill requirements against available capabilities.
+    
+    This is the core capability gate enforcement engine. It checks skill
+    requirements against the runtime context (available tools, servers, etc.)
+    and determines whether a skill should be activated based on policy.
+    """
+    
+    def __init__(self, enforcement_level: EnforcementLevel = EnforcementLevel.WARN):
+        """Initialize the validator.
+        
+        Args:
+            enforcement_level: How strictly to enforce requirements
+        """
+        self.enforcement_level = enforcement_level
+        self._tool_cache: Optional[Set[str]] = None
+        self._server_cache: Optional[Set[str]] = None
+    
+    def validate_skill(
+        self,
+        skill: SkillProperties,
+        available_tools: Optional[Set[str]] = None,
+        available_servers: Optional[Set[str]] = None
+    ) -> ValidationResult:
+        """Validate a skill's requirements against available capabilities.
+        
+        Args:
+            skill: Skill to validate
+            available_tools: Set of available tool names (auto-detected if None)
+            available_servers: Set of available server names (auto-detected if None)
+            
+        Returns:
+            ValidationResult with validation status and details
+        """
+        if not skill.requirements or skill.requirements.is_empty():
+            # No requirements specified - skill is always available
+            return ValidationResult(
+                skill_name=skill.name,
+                state=SkillState.ACTIVE,
+                satisfied_tools=[],
+                missing_tools=[],
+                satisfied_servers=[],
+                missing_servers=[],
+                satisfied_env_vars=[],
+                missing_env_vars=[],
+                warnings=[],
+                errors=[]
+            )
+        
+        # Get available capabilities
+        if available_tools is None:
+            available_tools = self._get_available_tools()
+        if available_servers is None:
+            available_servers = self._get_available_servers()
+            
+        requirements = skill.requirements
+        
+        # Validate tools
+        satisfied_tools = [tool for tool in requirements.tools if tool in available_tools]
+        missing_tools = [tool for tool in requirements.tools if tool not in available_tools]
+        
+        # Validate servers
+        satisfied_servers = [srv for srv in requirements.servers if srv in available_servers]
+        missing_servers = [srv for srv in requirements.servers if srv not in available_servers]
+        
+        # Validate environment variables
+        satisfied_env_vars = [var for var in requirements.env_vars if os.getenv(var)]
+        missing_env_vars = [var for var in requirements.env_vars if not os.getenv(var)]
+        
+        # Determine warnings and errors
+        warnings = []
+        errors = []
+        
+        if missing_tools:
+            msg = f"Missing required tools: {', '.join(missing_tools)}"
+            if self.enforcement_level in (EnforcementLevel.STRICT,):
+                errors.append(msg)
+            else:
+                warnings.append(msg)
+                
+        if missing_servers:
+            msg = f"Missing required servers: {', '.join(missing_servers)}"
+            if self.enforcement_level in (EnforcementLevel.STRICT,):
+                errors.append(msg)
+            else:
+                warnings.append(msg)
+                
+        if missing_env_vars:
+            msg = f"Missing required environment variables: {', '.join(missing_env_vars)}"
+            if self.enforcement_level == EnforcementLevel.STRICT:
+                errors.append(msg)
+            else:
+                warnings.append(msg)
+        
+        # Determine skill state
+        if errors:
+            state = SkillState.UNAVAILABLE
+        elif warnings:
+            state = SkillState.DEGRADED
+        else:
+            state = SkillState.ACTIVE
+            
+        result = ValidationResult(
+            skill_name=skill.name,
+            state=state,
+            satisfied_tools=satisfied_tools,
+            missing_tools=missing_tools,
+            satisfied_servers=satisfied_servers,
+            missing_servers=missing_servers,
+            satisfied_env_vars=satisfied_env_vars,
+            missing_env_vars=missing_env_vars,
+            warnings=warnings,
+            errors=errors
+        )
+        
+        # Log based on enforcement level
+        self._log_validation_result(result)
+        
+        return result
+    
+    def _get_available_tools(self) -> Set[str]:
+        """Get set of available tool names from registry."""
+        if self._tool_cache is None:
+            try:
+                from ..tools.registry import get_registry
+                registry = get_registry()
+                self._tool_cache = set(registry.list_tools())
+            except ImportError:
+                logger.debug("Tool registry not available")
+                self._tool_cache = set()
+        return self._tool_cache
+    
+    def _get_available_servers(self) -> Set[str]:
+        """Get set of available MCP server names."""
+        if self._server_cache is None:
+            # TODO: Implement MCP server discovery
+            # For now, return empty set - this can be extended later
+            self._server_cache = set()
+        return self._server_cache
+    
+    def _log_validation_result(self, result: ValidationResult) -> None:
+        """Log validation result based on enforcement level."""
+        if self.enforcement_level == EnforcementLevel.DISABLED:
+            return
+            
+        skill_name = result.skill_name
+        
+        if result.state == SkillState.UNAVAILABLE:
+            if self.enforcement_level == EnforcementLevel.STRICT:
+                logger.error(f"Skill '{skill_name}' blocked: {'; '.join(result.errors)}")
+            else:
+                logger.warning(f"Skill '{skill_name}' has critical missing dependencies: {'; '.join(result.errors)}")
+                
+        elif result.state == SkillState.DEGRADED:
+            logger.warning(f"Skill '{skill_name}' partially available: {'; '.join(result.warnings)}")
+            
+        elif result.state == SkillState.ACTIVE:
+            logger.debug(f"Skill '{skill_name}' requirements satisfied")
+            
+    def clear_cache(self) -> None:
+        """Clear cached capability information."""
+        self._tool_cache = None
+        self._server_cache = None

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

@@ -108,7 +108,7 @@ class SkillLoader:
             return False
 
         try:
-            content = skill_md.read_text()
+            content = skill_md.read_text(encoding="utf-8")
             _, body = parse_frontmatter(content)
             skill.instructions = body
             return True
@@ -136,7 +136,7 @@ class SkillLoader:
             for script_file in scripts_dir.iterdir():
                 if script_file.is_file():
                     try:
-                        scripts[script_file.name] = script_file.read_text()
+                        scripts[script_file.name] = script_file.read_text(encoding="utf-8")
                     except Exception:
                         continue
         except PermissionError:
@@ -166,7 +166,7 @@ class SkillLoader:
             for ref_file in refs_dir.iterdir():
                 if ref_file.is_file():
                     try:
-                        refs[ref_file.name] = ref_file.read_text()
+                        refs[ref_file.name] = ref_file.read_text(encoding="utf-8")
                     except Exception:
                         continue
         except PermissionError:

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

@@ -5,12 +5,13 @@ from typing import List, Optional, Dict
 
 logger = logging.getLogger(__name__)
 
-from .models import SkillMetadata
-from .discovery import discover_skills
+from .models import SkillMetadata, SkillState
+from .discovery import discover_skills, get_default_skill_dirs
 from .loader import SkillLoader, LoadedSkill
 from .prompt import generate_skills_xml
 from .substitution import render_skill_body
 from .shell_render import render_shell_blocks
+from .capability_validator import CapabilityValidator, EnforcementLevel, ValidationResult
 
 
 class SkillManager:
@@ -34,11 +35,21 @@ class SkillManager:
         manager.activate(skill)
     """
 
-    def __init__(self):
-        """Initialize the SkillManager."""
+    def __init__(self, enforcement_level: Optional[EnforcementLevel] = None):
+        """Initialize the SkillManager.
+        
+        Args:
+            enforcement_level: Capability enforcement level (defaults to environment-based)
+        """
         self._skills: Dict[str, LoadedSkill] = {}
         self._loader = SkillLoader()
         self._discovered = False
+        self._validation_cache: Dict[str, ValidationResult] = {}
+        
+        # Initialize capability validator
+        if enforcement_level is None:
+            enforcement_level = self._get_default_enforcement_level()
+        self._validator = CapabilityValidator(enforcement_level)
 
     @property
     def skills(self) -> List[LoadedSkill]:
@@ -126,15 +137,29 @@ class SkillManager:
     def get_available_skills(self) -> List[SkillMetadata]:
         """Get metadata for all available skills.
 
-        This is used for system prompt injection. Skills with
-        ``disable-model-invocation: true`` are omitted so the LLM never
-        sees them and cannot auto-trigger them.
+        This is used for system prompt injection. Skills are filtered by:
+        1. disable-model-invocation: true (excluded)
+        2. Capability enforcement (if strict mode, exclude UNAVAILABLE skills)
         """
-        return [
-            skill.metadata
-            for skill in self._skills.values()
-            if not getattr(skill.properties, "disable_model_invocation", False)
-        ]
+        available = []
+        for skill in self._skills.values():
+            # Skip if explicitly disabled for model invocation
+            if getattr(skill.properties, "disable_model_invocation", False):
+                continue
+                
+            # Check capability requirements if enforcement is strict
+            if self._validator.enforcement_level == EnforcementLevel.STRICT:
+                try:
+                    validation = self.validate_skill_capabilities(skill.properties.name)
+                    if validation.state == SkillState.UNAVAILABLE:
+                        continue  # Skip unavailable skills in strict mode
+                except Exception as e:
+                    logger.warning(f"Skipping skill '{skill.properties.name}' due to validation error: {e}")
+                    continue
+                    
+            available.append(skill.metadata)
+            
+        return available
 
     def get_user_invocable_skills(self) -> List[LoadedSkill]:
         """Return skills the *user* may invoke via slash-commands.
@@ -271,8 +296,7 @@ class SkillManager:
                 return {"success": False, "error": "Skill content exceeds maximum size (100KB)"}
             
             # Create skill directory
-            from .discovery import get_default_skill_directories
-            skill_dirs = get_default_skill_directories()
+            skill_dirs = get_default_skill_dirs()
             base_dir = skill_dirs[0] if skill_dirs else "~/.praisonai/skills"
             
             import os
@@ -567,9 +591,90 @@ author: agent
                 logger.debug("Failed to clean up temp file %s", temp_path, exc_info=True)
             raise
 
+    def validate_skill_capabilities(self, skill_name: str, force_refresh: bool = False) -> ValidationResult:
+        """Validate a skill's capability requirements.
+        
+        Args:
+            skill_name: Name of the skill to validate
+            force_refresh: If True, bypass cache and re-validate
+            
+        Returns:
+            ValidationResult with capability status
+            
+        Raises:
+            ValueError: If skill not found
+        """
+        skill = self.get_skill(skill_name)
+        if skill is None:
+            raise ValueError(f"Skill '{skill_name}' not found")
+            
+        # Check cache first
+        if not force_refresh and skill_name in self._validation_cache:
+            return self._validation_cache[skill_name]
+            
+        # Perform validation
+        result = self._validator.validate_skill(skill.properties)
+        self._validation_cache[skill_name] = result
+        return result
+    
+    def get_available_skills_by_state(self, state: SkillState) -> List[LoadedSkill]:
+        """Get skills filtered by their capability validation state.
+        
+        Args:
+            state: Desired skill state
+            
+        Returns:
+            List of skills in the specified state
+        """
+        matching = []
+        for skill in self._skills.values():
+            try:
+                result = self.validate_skill_capabilities(skill.properties.name)
+                if result.state == state:
+                    matching.append(skill)
+            except Exception as e:
+                logger.warning(f"Failed to validate skill '{skill.properties.name}': {e}")
+                continue
+        return matching
+    
+    def get_skills_diagnostics(self) -> Dict[str, ValidationResult]:
+        """Get capability diagnostics for all skills.
+        
+        Returns:
+            Dict mapping skill names to their validation results
+        """
+        diagnostics = {}
+        for skill_name in self.skill_names:
+            try:
+                diagnostics[skill_name] = self.validate_skill_capabilities(skill_name)
+            except Exception as e:
+                logger.error(f"Failed to validate skill '{skill_name}': {e}")
+        return diagnostics
+    
+    def _get_default_enforcement_level(self) -> EnforcementLevel:
+        """Get default enforcement level from environment."""
+        import os
+        level_str = os.getenv('SKILL_CAPABILITY_ENFORCEMENT', 'warn').lower()
+        
+        level_map = {
+            'disabled': EnforcementLevel.DISABLED,
+            'off': EnforcementLevel.DISABLED,
+            'telemetry': EnforcementLevel.TELEMETRY,
+            'log': EnforcementLevel.TELEMETRY,
+            'warn': EnforcementLevel.WARN,
+            'warning': EnforcementLevel.WARN,
+            'strict': EnforcementLevel.STRICT,
+            'hard': EnforcementLevel.STRICT,
+            'fail': EnforcementLevel.STRICT,
+        }
+        
+        return level_map.get(level_str, EnforcementLevel.WARN)
+    
     def clear(self) -> None:
         """Clear all loaded skills."""
         self._skills.clear()
+        self._validation_cache.clear()
+        self._validator.clear_cache()
         self._discovered = False
 
     def __len__(self) -> int:

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

@@ -1,8 +1,9 @@
 """Data models for Agent Skills."""
 
 from dataclasses import dataclass, field
-from typing import Optional
+from typing import Optional, List, Union
 from pathlib import Path
+from enum import Enum
 
 
 class ParseError(Exception):
@@ -15,6 +16,105 @@ class ValidationError(Exception):
     pass
 
 
+class SkillState(Enum):
+    """Skill activation state based on requirement validation."""
+    ACTIVE = "active"          # All requirements satisfied
+    DEGRADED = "degraded"      # Some requirements missing (soft warn)
+    UNAVAILABLE = "unavailable" # Critical requirements missing (hard fail)
+    UNKNOWN = "unknown"        # Requirements not yet validated
+
+
+@dataclass
+class SkillRequirements:
+    """Capability requirements for a skill.
+    
+    This represents the parsed and normalized requirements from skill frontmatter.
+    Supports both Hermes/OpenClaw conventions and PraisonAI-specific extensions.
+    
+    Attributes:
+        servers: Required MCP servers or service endpoints
+        tools: Required tool names (from tool registry)
+        env_vars: Required environment variables (use sparingly)
+        openclaw_hints: OpenClaw-specific metadata (passthrough)
+    """
+    servers: List[str] = field(default_factory=list)
+    tools: List[str] = field(default_factory=list) 
+    env_vars: List[str] = field(default_factory=list)
+    openclaw_hints: dict = field(default_factory=dict)
+    
+    @classmethod
+    def from_frontmatter(cls, metadata: dict) -> "SkillRequirements":
+        """Parse requirements from skill frontmatter.
+        
+        Normalizes various naming conventions:
+        - requires_servers, requires_tools (Hermes/OpenClaw style)
+        - allowed-tools (existing PraisonAI convention) 
+        - openclaw (OpenClaw-specific hints)
+        
+        Args:
+            metadata: Raw frontmatter dict
+            
+        Returns:
+            SkillRequirements instance
+        """
+        servers = []
+        tools = []
+        env_vars = []
+        openclaw_hints = {}
+        
+        # Parse servers
+        if "requires_servers" in metadata:
+            servers = cls._normalize_list(metadata["requires_servers"])
+        elif "requires-servers" in metadata:
+            servers = cls._normalize_list(metadata["requires-servers"])
+        
+        # Parse tools (multiple naming conventions)
+        if "requires_tools" in metadata:
+            tools.extend(cls._normalize_list(metadata["requires_tools"]))
+        elif "requires-tools" in metadata:
+            tools.extend(cls._normalize_list(metadata["requires-tools"]))
+            
+        # Support existing allowed-tools convention for backward compatibility
+        if "allowed-tools" in metadata:
+            tools.extend(cls._normalize_list(metadata["allowed-tools"]))
+            
+        # Parse environment variables
+        if "requires_env" in metadata:
+            env_vars = cls._normalize_list(metadata["requires_env"])
+        elif "requires-env" in metadata:
+            env_vars = cls._normalize_list(metadata["requires-env"])
+            
+        # OpenClaw hints (passthrough)
+        if "openclaw" in metadata:
+            openclaw_hints = metadata["openclaw"] if isinstance(metadata["openclaw"], dict) else {}
+            
+        return cls(
+            servers=servers,
+            tools=tools,
+            env_vars=env_vars, 
+            openclaw_hints=openclaw_hints
+        )
+    
+    @staticmethod
+    def _normalize_list(value: Union[str, List[str]]) -> List[str]:
+        """Normalize string or list to list of strings."""
+        if isinstance(value, str):
+            # Support both comma-separated and space-separated strings
+            return [item.strip() for item in value.replace(',', ' ').split() if item.strip()]
+        elif isinstance(value, (list, tuple)):
+            return [str(item).strip() for item in value if str(item).strip()]
+        else:
+            return []
+    
+    def is_empty(self) -> bool:
+        """Check if no requirements are specified."""
+        return not any([self.servers, self.tools, self.env_vars, self.openclaw_hints])
+    
+    def __bool__(self) -> bool:
+        """Truth value: True if any requirements are specified."""
+        return not self.is_empty()
+
+
 @dataclass
 class SkillProperties:
     """Properties parsed from a skill's SKILL.md frontmatter.
@@ -49,6 +149,8 @@ class SkillProperties:
     hooks: Optional[dict] = None
     paths: Optional[object] = None  # list[str] glob patterns
     shell: Optional[str] = None  # "bash" | "powershell"
+    # Capability requirements (new)
+    requirements: Optional[SkillRequirements] = None
 
     def to_dict(self) -> dict:
         """Convert to dictionary, excluding None values and empty metadata."""

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

@@ -3,7 +3,7 @@
 from pathlib import Path
 from typing import Optional, Tuple
 
-from .models import ParseError, ValidationError, SkillProperties
+from .models import ParseError, ValidationError, SkillProperties, SkillRequirements
 
 
 def find_skill_md(skill_dir: Path) -> Optional[Path]:
@@ -100,7 +100,7 @@ def read_properties(skill_dir: Path) -> SkillProperties:
     if skill_md is None:
         raise ParseError(f"SKILL.md not found in {skill_dir}")
 
-    content = skill_md.read_text()
+    content = skill_md.read_text(encoding="utf-8")
     metadata, _ = parse_frontmatter(content)
 
     if "name" not in metadata:
@@ -143,4 +143,6 @@ def read_properties(skill_dir: Path) -> SkillProperties:
         hooks=metadata.get("hooks") if isinstance(metadata.get("hooks"), dict) else None,
         paths=metadata.get("paths"),
         shell=metadata.get("shell"),
+        # Parse capability requirements
+        requirements=SkillRequirements.from_frontmatter(metadata),
     )

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

@@ -7,7 +7,7 @@ plugin bundle, enterprise policy) without touching the core.
 
 from __future__ import annotations
 
-from typing import Protocol, Iterable, Optional, runtime_checkable
+from typing import Protocol, Iterable, Optional, runtime_checkable, List, Dict, Any
 
 from .models import SkillProperties
 
@@ -45,7 +45,8 @@ class SkillMutatorProtocol(Protocol):
     """Agent-managed skill CRUD operations.
     
     Allows agents to create, edit, and manage their own skills at runtime.
-    Implementations should provide safe-by-default behavior (e.g., propose mode).
+    Implementations should provide safe-by-default behaviour: use ``propose=True``
+    (default) to stage mutations for human approval before writing to disk.
     """
 
     def create(self, name: str, content: str, category: Optional[str] = None,
@@ -163,3 +164,30 @@ class SkillMutatorProtocol(Protocol):
             Status message describing the action taken
         """
         ...
+
+
+@runtime_checkable
+class SkillsCatalogProtocol(Protocol):
+    """Metadata listing for skills — UI catalog pages."""
+
+    def list_skills(self) -> List[Dict[str, Any]]:
+        """Return skill metadata dicts (name, description, location)."""
+        ...
+
+
+def list_skills_for_api() -> List[Dict[str, Any]]:
+    """Default catalog adapter using SkillManager when available."""
+    try:
+        from .manager import SkillManager
+
+        manager = SkillManager()
+        return [
+            {
+                "name": s.name,
+                "description": s.description,
+                "location": getattr(s, "location", ""),
+            }
+            for s in manager.get_available_skills()
+        ]
+    except Exception:
+        return []