9e1f77f52a
## Summary Make Claude and Codex agents easier to configure and test behind AgentOS. Ordinary settings no longer require untyped `_kwargs` dictionaries, and response printers return the final run and raise on failure so cookbook failures are visible. - Add typed native options and named tools, permissions, MCP and configuration fields. Named non-None settings take precedence; caller-owned configuration is copied. Legacy `_kwargs` aliases warn. - Group related constructor parameters and make built-in adapters keyword-only. Expose read-only `agent.sdk`; retain the Python `framework` compatibility alias and legacy session reads. API/session metadata emits only `sdk`. - Give sync/async response printers a shared result/error contract, tool-event deduplication and persistence warnings. Correct public streaming and async types, including optional final `RunOutput`. - Improve Claude/Codex cookbooks in their existing framework folders: streaming printers, native SDK comparisons, tool fixtures, structured output, sessions, media and AgentOS HTTP/SSE examples. Include a reproducible reliability kit and separately pinned historical results. - Integrate current main, including media, retries, metrics, compaction, structured output, and #10958/#10962 session-busy/replay changes. Preserve media on successful, failed and cancelled runs and both upstream/DX regression coverage. ## Type of change - [x] Bug fix - [x] New feature - [x] Breaking change - [x] Improvement - [ ] Model update - [x] Other: Cookbook and test coverage --- ## Checklist - [x] Code complies with style guidelines - [x] Ran format/validation scripts (`./scripts/format.sh` and `./scripts/validate.sh`) - [x] Self-review completed - [x] Documentation updated (comments, docstrings) - [x] Examples and guides: Relevant cookbook examples have been included or updated (if applicable) - [ ] Tested in clean environment - [x] Tests added/updated (if applicable) ### Duplicate and AI-Generated PR Check - [x] I have searched existing open pull requests and confirmed that no other PR already addresses this issue - [ ] If a similar PR exists, I have explained below why this PR is a better approach - [x] Check if this PR was entirely AI-generated (by Copilot, Claude Code, Cursor, etc.) --- ## Additional Notes ### Migration Use named constructor arguments and select the adapter class instead of passing `framework=`. Use `options`, `thread_options` and `turn_options` instead of their deprecated `_kwargs` names. Printers return the terminal `RunOutput` and raise on errors/cancellation by default; use `raise_on_error=False` to opt out. Unsupported separate media and native input objects fail explicitly. Import paths and transcript namespaces remain unchanged. Agno owns session selection; native lifecycle overrides that conflict with it are rejected. ### Validation — 11 October 2026 Integrated main `dbdca9ac6d9de6604383e6f4f04653a8244eb3c7` into DX head `c88e47d54de487f4d14b95c684a4c3e5889e8d0f`. From the normal checkout in the existing `.venvs/claude-dx-validation` environment: ```bash source .venvs/claude-dx-validation/bin/activate python -m pytest libs/agno/tests/unit/agents \ libs/agno/tests/unit/os/test_external_agent_background_stream.py \ libs/agno/tests/unit/os/test_schemas.py \ libs/agno/tests/unit/os/test_db_replay_fallback.py \ libs/agno/tests/unit/os/test_queue_worker.py \ libs/agno/tests/unit/run/test_queue_store.py \ libs/agno/tests/unit/os/test_ws_replay_floor.py -q -o addopts='' ./scripts/format.sh ./scripts/validate.sh ``` 620 tests pass, including public typing, media/schema combinations, retries, busy-session handling, replay and queue contracts. Full formatting and validation pass (1115 Agno / 21 agnoctl mypy files). No new live SDK calls were made for this merge refresh. Earlier live/provider and eight-hour paced-soak results are historical, with pinned revisions, first failures and limitations retained in the framework and reliability TEST_LOG files; they are not certification of this combined revision. The session-busy check is not an atomic distributed claim. Applications must serialize same-session submissions across replicas; the cookbooks state that limit. Native Claude resume and Codex bounded-history fallback remain distinct. Retries can repeat tool effects. Sandbox execution and final live release acceptance remain separate work. ### Final review follow-up — 11 October 2026 Final revision: `1cc091cc4c`. Found and fixed an interaction between native options and media: attachments now use the effective SDK workspace (including Claude native/legacy options and Codex thread/turn overrides) and absolute paths for relative working directories. Cleanup and recorded upload roots use the same workspace. Fifteen targeted cases failed before the fix; all 20 cases pass after. Expanded local validation: ```bash python -m pytest libs/agno/tests/unit/agents libs/agno/tests/unit/os \ libs/agno/tests/unit/run -q -o addopts='' python -m pytest libs/agno/tests/unit/os/test_public_json_bounds.py -q -o addopts='' ``` 4,484 unique tests pass across these runs; 23 cases skip. The broad run passed 4,467 cases; 17 HTTP cases initially hit the execution sandbox's loopback-bind restriction, then passed when the full 29-case HTTP file was rerun with loopback access. An initial optional Telegram import failure was resolved in the isolated validation environment. Required full format/validation passes. No new live provider calls or soak; final-commit CI is a separate merge gate. --------- Co-authored-by: kausmeows <[email protected]> Co-authored-by: Claude Fable 5.1 <[email protected]> Co-authored-by: Yash Pratap Solanky <[email protected]>
Framework integrations
Run a coding harness through Agno, inspect its tools, then serve it through AgentOS. Start with one provider: Claude or Codex. Both review the same small shipping project.
The harness owns its agent loop, tool execution and native context management. Agno adapts its output to runs and events. AgentOS adds the HTTP API and run storage used in this first exercise.
This directory also contains LangGraph, DSPy, Antigravity, and the mixed-framework quickstart. Claude and Codex each keep their starting and advanced examples in one folder.
Start here
Use a source checkout containing this cookbook. A package version alone does not identify which unreleased adapter changes are present.
From the repository root:
uv venv .venvs/harnesses --python 3.12
source .venvs/harnesses/bin/activate
uv pip install -e 'libs/agno[os,sqlite]'
Then install and authenticate the SDK in the provider's README. You only need the provider you choose. These scripts make real model calls. Use the same interpreter for installation, examples and tests.
Follow the same four steps with either provider
| Step | Claude | Codex | What to inspect |
|---|---|---|---|
| 1. First run | claude_basic.py | codex_basic.py | Answer, run ID and terminal status |
| 2. Compare the SDK | claude_native_sdk.py | codex_native_sdk.py | Same prompt/model through the native SDK |
| 3. Inspect code | claude_tools.py | codex_tools.py | Streamed text, visible tools and actual file results |
| 4. Serve it | claude_agentos.py | codex_agentos.py | HTTP/SSE and persisted runs |
Start with python cookbook/frameworks/claude-agent-sdk/claude_basic.py or
python cookbook/frameworks/codex/codex_basic.py. Every script is independent.
The tools exercise should explain fees of 8, 0 and 0 dollars for the three
orders. No editing of the sample project is required.
The basic and native SDK examples contain the policy in the prompt. They
establish a round trip, not file access. The tools example must actually read
the fixture. The basic examples use print_response(), which returns a RunOutput and
raises on errors or cancellation by default. The native SDK and Codex tools examples also check
terminal results explicitly; Claude tools use the same response printer. Failed runs exit unsuccessfully.
What this first set establishes
| Area | Current exercise and boundary |
|---|---|
| Text and tools | Live standalone streaming and HTTP streaming/non-streaming; SDK-specific details are translated, not exposed losslessly |
| Product history | The AgentOS examples save Agno runs/tool results in a local SQLite file |
| Claude native state | With this SQLite adapter, native entries are mirrored through the Claude SDK session store; successful mirroring and cross-process recovery need their own tests |
| Codex native state | Agno stores the thread ID; native thread files remain in Codex's local storage |
| Durability | This set does not prove disconnect survival, durable queue acceptance, retries or crash recovery |
| Advanced SDK features | Compaction, subagents, skills, plugins, approvals and checkpoint branching are separate exercises |
Product history, native conversation state and workspace files are different things. Saving a run does not establish recovery of all three. In the current adapters, a missing native session can fall back to a new conversation seeded with Agno history; that is not proof of native continuity.
Support limitations live here and in the provider READMEs. Fresh observations live in TEST_LOG.md; use TEST_PROMPT.md to reproduce them.
Runtime boundaries
The server examples bind to 127.0.0.1, have no authentication and use SQLite
for local development. Production requires explicit authorization, durable
storage and an execution/coordination topology appropriate to the product.
Neither a working directory nor a tool allowlist provides tenant isolation.
Both examples use bounded prompts. Native scripts time out after 120 seconds, and the live test suite bounds subprocess/HTTP waits. Those are test limits, not billing guarantees.
PORT changes the local server port. HARNESS_STATE_DIR changes the SQLite
directory; otherwise it is tmp/harnesses/<provider>/ under your launch directory.
Stop the server before removing its disposable local state. That does not
remove SDK-native session files.
Existing examples and the next exercises
The Claude guide and Codex guide include sessions, MCP tools, background execution and other advanced examples. Their historical test logs do not imply verification by the starting-example suite.
Next: native session recovery and storage failures; long conversations and compaction; background/queue/disconnect/retry/multi-replica exercises; step replay; skills/plugins/registry; interfaces and session-owned sandboxes. These are follow-up topics, not guarantees made by this first set.
Adapter API changes for 3.2
Configure built-in adapters with keywords, for example ClaudeAgent(name="Reviewer", model="claude-sonnet-5-5"). Positional constructor arguments are no longer accepted.
Related fields are grouped in the constructor signature.
This applies to Claude, Codex, LangGraph, DSPy and Antigravity. Python 3.9 remains
supported by the adapter layer; each native SDK has its own Python requirements.
The integration is identified by read-only agent.sdk, selected by the agent class.
For example, Claude uses "claude-agent-sdk" and Codex uses "codex". Neither sdk
nor framework is a constructor option. agent.framework remains a compatibility
alias for reads; calls such as ClaudeAgent(framework="claude-agent-sdk") must
remove that argument and select the corresponding adapter class instead.
API metadata and new session metadata include both keys; existing transcript
storage namespaces and database columns keep their original names and values.
print_response() and aprint_response() return the final RunOutput, including
run ID, status and tools. Errors and cancellations are displayed explicitly and
raise by default. To inspect an unsuccessful result without raising:
result = agent.print_response("Review the project", stream=True, raise_on_error=False)
print(result.status)
The async equivalent is result = await agent.aprint_response(..., stream=True). Interactive
printing rejects background=True; call arun(background=True) directly for
background execution. Foreground streams yield event objects; background streams
yield SSE strings. With yield_run_output=True, either stream appends a terminal
RunOutput.
Separate images, audio, videos and files inputs are rejected by the current
adapters before work starts. Empty collections are accepted. This does not restrict
file reads through the harness's native tools; it prevents uploaded inputs from
being silently discarded. Media support needs an explicit adapter implementation.
Codex now accepts native client_options plus typed thread_options and
turn_options. The old thread_kwargs / turn_kwargs names are deprecated.
Explicit named settings take precedence over either form, including empty values
and False; callers that relied on conflicting raw overrides must update them.
Use mcp_servers for native MCP configuration. Codex input must be a string;
unsupported native input objects fail explicitly. See the
Codex configuration guide.