kaxil commented on issue #74741: URL: https://github.com/apache/airflow/issues/74741#issuecomment-6228178573
Tested common.ai 1.0.0rc1 from the PyPI wheel on Airflow 3.3.2 (Python 3.12), with live runs against Claude Sonnet 5.5, TypeSafe Jev and Modal sandboxes. Also ran the Airflow 2.11.2 path, and Airflow 3.4.0b1 for the 3.4-only features. Works as described in the PRs: - #74385, #74704, #74314, #73873, #74720, #74313 (durable execution): a retry replays completed model steps and tool calls instead of running them again (a side-effecting tool ran once across two tries). A tool whose result couldn't be recorded is named and runs again. Answers the output validator rejected are asked again, not replayed. Usage with `audio_seconds` loads on every resume. - #74396, #73706 (usage and spend across retries): attempt 2 hits `UsageLimitExceeded` early because attempt 1's requests count, and the failed attempt's `usage` XCom holds its own usage. - #74308, #74707 (retry policies): with `include_traceback=True` the chained traceback reaches the classifier; `ClassifierRetryPolicy` on Jev picks retry or fail per category and escalates a below-bar answer to `LLMRetryPolicy`. - #73994 (prompt caching): a repeated task reads from the cache, and `cache_prompt=False` turns it off. - #73962, #73898, #73587, #73901, #74404 (framework adapters): the Anthropic tool runner, Strands and Google ADK examples run unmodified, with masked tool results and `common_ai.tool_calls` counted per framework. A Strands task resumes from its checkpoint on retry and re-runs only the failed tool. - #73532 (managed-agent contract): exercised with a stub implementation of the hook contract, including failover and refusals. The Bedrock and Vertex hooks themselves were not exercised. - #73586, #74686, #73904, #74450, #74351, #74272, #72202 (tool approval and HITL), under a real scheduler and API server: approve and reject both behave, multi-step approval works on 3.4.0b1 (and fails cleanly on 3.3 as documented), the approved typed answer reaches a downstream task, an unreadable review XCom fails after 10 reads, and review by map index and by loop pass touches only that slot. - #73899, #73925, #70433, #73900, #74380, #74447, #74449, #73957, #74448, #73897, #73993, #73938, #73984, #74312, #72002, #74626 (toolsets): checked live with an agent where the model is involved. Secrets in tool output reach the model as `***` from toolsets, capabilities and retry messages alike. #74662 was checked offline (the batch submission itself was stubbed). - #73418, #74302, #73990, #73559 (Modal): the `modal` connection authenticates with no ambient credentials, and an agent ran commands, exported a file to object storage, and attached to a sandbox another task provisioned. (The Modal provider was 0.1.0rc1 in this environment.) - #74705, #73896, #74679 (lifecycle): production lifecycle, and no experimental warnings on the graduated features. The old usage in each 1.0.0 migration note fails as described and the new usage works; two notes overstate details (below). Found, with fixes in open PRs for 1.0.1. None of them blocks this release: - #74381: `AgentSkillsToolset` itself works, but a fresh `[skills]` install resolves pydantic-ai-skills 2.1.0, which removed `SkillsToolset`, so every run fails. 0.10.0 is affected the same way. #74899 caps it below 2. - #74334: the Model tab never loads on 3.4.0b1, because it scopes on `operator_names` and core expects `operator_name`. Only affects 3.4, which is in beta. #74898. - `durable=True` with duplicate toolset ids, or a tool that defers, fails with pydantic-ai errors instead of the documented ones. #74902, which also makes the changelog's "unique `id=`" note true. - The changelog's `get_schema` note says to check `truncated` first, but a full result has no `truncated` key. #74904. - Stale rows on the supported services page (#72939) and other docs fixes: #74904. Example Dags that never reach their feature against a real model: #74901. The pydantic-ai first-run banner in task logs when Airflow is started from a coding agent: #74900. +1 for common.ai 1.0.0rc1. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
