[ENG-3419] Honor the selected reasoning effort in cloud turns

[tino097]

https://linear.app/mindsdb/issue/ENG-3419

What changed

• Apply selected reasoning effort in cloud Anton sessions: an optional reasoning_effort on TurnRequestV1 sets the planning model's effort only; coding and history summarization keep their configured effort
• Fall back to the default effort when the request carries a non-string value: a malformed value no longer fails the turn
• Read the requested effort once and treat a blank value as unset: " " keeps the configured effort
• Resend once at the configured effort when the provider refuses the selected one: a fallback provider around the planning role, removed again below
• Document the effort fallback provider's methods: docstrings only, removed again below
• Replace the effort fallback provider with the planning effort setting: the selected effort is applied through settings.planning_reasoning_effort next to the model assignment, so the client factory and its callers are unchanged from staging
• Assert the selected effort on the wire instead of a private builder: the tests drive real clients over a mock transport and check the reasoning_effort each role's request carried
• Drop the leftover docstring on the client factory: no code change
• Build the streamed answer in the session tests and document the effort helpers: tests only

Notes for the reviewer

• Effort is planning only on purpose. The router shares the coding provider when none is configured, so changing coding effort would also change summarization.
• An unadvertised level never fails the turn, and the pod does not judge it. cowork-server drops a level the model does not advertise when the org catalog is cached. Otherwise the gateway clamps an off-ladder level to the nearest rung or the alias default and drops one it cannot carry; it never answers 400. Every cloud job carries the gateway llm block, so no other provider sees the value.
• The fallback provider is gone. The gateway never refuses a level, so the resend could not fire on the cloud path; the net diff is the contract field, the one setting assignment, and tests.
• Cross repo contract. The field is optional; an older pod ignores the key.

Ships with

• mindsdb/scratchpad-controller#87 and [ENG-3419] Send the selected reasoning effort to cloud turns cowork-server#671
• Deploy order: this one (pod image), then the controller, then cowork-server. Roll back cowork-server first; older hops ignore the field.

Checks

Check	Observed result
uv run pytest (full suite, at the wire test commit)	5043 passed, 91 skipped, 1 failed: the live fetch of example.com, which passed when rerun alone
uv run pytest tests/test_cloud_turn_session.py tests/test_cloud_turn_contract.py tests/test_cloud_turn_process.py tests/test_cloud_turn_usage_notes.py tests/test_client.py tests/test_bad_request_reaches_the_classifier.py at the head	243 passed

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-13T15:09:00.097172Z	6fb32e1	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6fb32e1c19

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Close delegated streams when the wrapper is closed

When a selected-effort planning stream is cancelled or its consumer stops early, LLMClient._guarded_stream closes this wrapper, but the bare async for does not close the underlying provider async generator. Consequently, its HTTP response can remain open—and potentially continue consuming server resources—until async-generator finalization. Wrap both delegated streams in contextlib.aclosing, as the other provider delegation paths do.

Useful? React with 👍 / 👎.