kaxil opened a new pull request, #74904:
URL: https://github.com/apache/airflow/pull/74904

   Fixes docs statements in the Common AI provider that don't match what the 
provider does. Each one turned up while testing 1.0.0rc1 (#74741), and each was 
rechecked against `main` before editing.
   
   - **Supported services.** The SandboxToolset row named only Docker 
Sandboxes. It now lists all six shipped backends (Modal, Islo, OpenSandbox, 
OpenShell, Boat, sbx). The managed-agent row named Snowflake Cortex Agents and 
Azure AI Foundry, but no provider implements them. Only `BedrockAgentCoreHook` 
(amazon) and `AgentEngineHook` (google) implement the managed-agent contract on 
`main`. The DataFusion row gains Azure Blob Storage (#73374) and GCS, which 
`common.sql`'s DataFusion object-store providers support.
   - **System One servers** are added to the `pydanticai` connection's 
`external-services` in `provider.yaml`. The supported-services page is 
generated from that list, and `decision_models.rst` documents the `system-one:` 
prefix. `get_provider_info.py` is regenerated.
   - **Changelog, `get_schema` note.** It said "check `truncated` first". A 
full result has no `truncated` key, so `result["truncated"]` raises `KeyError` 
on every table that isn't summarized. It now says `result.get("truncated")`.
   - **Install lines.** The Anthropic, Strands and ADK guides all use 
`SQLToolset`, which imports `sqlglot` from this provider's `sql` extra. 
Following any of them in a fresh environment failed with `ModuleNotFoundError: 
No module named 'sqlglot'`. `DataFusionToolset` has the same import guard, so 
its page now names both extras.
   - **New guide: writing your own toolset** (`toolsets/custom.rst`). Until 
now, subclassing `AirflowToolset` and implementing `execute_tool` (#73938) was 
documented only in docstrings. The guide uses a new example Dag, a small 
toolset over PyPI's JSON API through `HttpHook`. It covers `id`, `get_tools`, 
`execute_tool`, `agent_template_fields`, `replayable`, which exceptions reach 
the model and which fail the task, `run_blocking`, and using the toolset from 
other frameworks. The JSON output and retry prompt shown on the page come from 
a run of the toolset against pypi.org. Like everything not on the stability 
page's stable list, the page is marked experimental.
   
   `README.rst`'s `htt  ps` typo is left alone. The release-prep README 
regeneration rewrites the whole file from a template that has the correct URL.
   
   The durable-execution items from the same testing are fixed in code in a 
separate PR, so the 1.0.0 changelog's durable note stays as it is: the code 
change makes it accurate.
   


-- 
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]

Reply via email to