[aw-failures] Claude `effort`/adaptive-thinking 400 kills daily-doc-updater + daily-function-namer (P0)

[github-actions]

Fix the Claude engine to stop sending effort/adaptive-thinking params to models that reject them — this kills every run of 2 scheduled workflows at turn 1, zero work done. Parent: #37282.

Problem

Two Claude-engine scheduled workflows abort on the first API call with 400 invalid_request_error, producing zero turns of useful work. Both set model: "${{ needs.activation.outputs.model_size }}" (an A/B model_size experiment); the small model the experiment selects does not support the effort or adaptive-thinking parameters that the Claude harness unconditionally sends.

Verbatim error (run 27061743674, agent-stdio.log):

[ERROR] API error (attempt 1/11): 400 {"type":"error","error":{"type":"invalid_request_error","message":"This model does not support the effort parameter."}}
[ERROR] API error (attempt 1/11): 400 {"type":"error","error":{"type":"invalid_request_error","message":"adaptive thinking is not supported on this model"}}
result: API Error: 400 This model does not support the effort parameter.  (num_turns=1, exitCode=1)

Affected workflows and run IDs

Workflow	Source	Failing run	Last green
Daily Go Function Namer	.github/workflows/daily-function-namer.md	27061743674 (06-06 12:01Z)	—
Daily Documentation Updater	.github/workflows/daily-doc-updater.md	27060575713 (06-06 11:04Z)	26949194158 (06-04)

Regression window: between 2026-06-04 11:34Z (last green) and 2026-06-05 11:39Z (first failure on Daily Documentation Updater).

audit-diff evidence — base 26949194158 (green) vs 27060575713 (fail)

• Core GitHub API calls consumed -92% (156 → 13): the agent did essentially nothing.
• Claude telemetry domain o205451.ingest.us.sentry.io dropped in the failing run — the agent died before reaching normal operation.
• num_turns=1, no token usage.

Probable root cause

The Claude harness sends effort (reasoning effort) and adaptive-thinking parameters on every request, but the model_size experiment resolves needs.activation.outputs.model_size to a Claude model that rejects both. Both workflows share the identical experiment block, so both break identically. Engine-level bug: parameter support is not gated on the resolved model.

Proposed remediation

1. Gate effort/adaptive-thinking in the Claude engine so they are only sent when the resolved model supports them (preferred — fixes all model_size-experiment workflows at once).
2. OR pin model_size to a model that supports effort, OR drop the effort/thinking config for the small-model variant.
3. Recompile and confirm the next scheduled runs reach the agent stage.

Success criteria / verification

• A scheduled run of Daily Documentation Updater and Daily Go Function Namer completes ≥2 turns without a 400 ... effort parameter / adaptive thinking error.
• No invalid_request_error for effort/adaptive-thinking in agent-stdio.log across model_size-experiment workflows.
  Related to [aw-failures] Agentic Maintenance blocked (P0): 6 workflows use deprecated max-effective-tokens #37282

Generated by 🔍 [aw] Failure Investigator (6h) · 367.6 AIC · ⌖ 13.3 AIC · ◷

• expires on Jun 13, 2026, 1:34 PM UTC