[aw-failures] PR Sous Chef 100% red — Copilot main agent requests `gpt-5.5`, SDK returns 400 "not accessible via /chat/completio [Content truncated due to length]

[github-actions]

Fix PR Sous Chef's Copilot model routing — gpt-5.5 must go through the OpenAI Responses API, not /chat/completions, or every scheduled run dies before the agent does any work.

In-flight fix already open: #42421 ("route gpt-5.5 through OpenAI Responses API wireApi=responses"). This issue tracks the failure + a secondary harness bug it exposes.

Problem statement

Both PR Sous Chef runs in the last 6h failed identically: the Copilot SDK driver rejects the configured model on its first request with 400 model "gpt-5.5" is not accessible via the /chat/completions endpoint. The agent produces 0 turns and empty output (agent_output.json: {"items":[],"errors":[]}) — no safe outputs, no PR triage. This is a hard red, distinct from the false-red tracked in #41636.

Affected workflows and run IDs

• PR Sous Chef (.github/workflows/pr-sous-chef.lock.yml)
  • §28425976809 — representative
  • §28422468107 — comparator (same signature, all 4 retries)
• Successful baseline for diff: run 28406960482 (cohort match) ran 28 turns; failed run ran 0 — classic immediate-400 regression.

Probable root cause

gpt-5.5 is a Responses-API-only model; routing it through the chat-completions wire API returns a deterministic 400. PR Sous Chef's resolved model/engine config still targets /chat/completions.

Secondary harness bug (actionable separately): the harness misclassifies this terminal error — isModelNotSupportedError=false, isHTTP400ResponseError=false, every flag false, failureClass=partial_execution — so it burns all 4 retries on an identical, non-retryable 400 before exiting 1. It should detect not accessible via the /chat/completions endpoint as terminal and fail fast (or auto-route to /responses).

Evidence (agent-stdio.log)

[copilot-sdk-driver] [sdk-driver] error: 400 model "gpt-5.5" is not accessible via the /chat/completions endpoint
[copilot-harness] attempt 1 failed: exitCode=1 failureClass=partial_execution isModelNotSupportedError=false isHTTP400ResponseError=false ... retriesRemaining=3
... attempts 2,3,4 identical ... retriesRemaining=0
agent_output.json: {"items":[],"errors":[]}

Proposed remediation

1. Land fix: route gpt-5.5 through OpenAI Responses API (wireApi=responses) #42421 so gpt-5.5 resolves through the Responses API (wireApi=responses) for the Copilot engine; verify PR Sous Chef's lock file picks it up after recompile.
2. Add a harness terminal-error matcher for not accessible via the /chat/completions endpoint → set isModelNotSupportedError/endpoint-mismatch true and stop retrying.

Success criteria / verification

• Next scheduled PR Sous Chef run reaches >0 turns and emits safe outputs (or completes clean no-op).
• No gpt-5.5 ... /chat/completions 400 in agent-stdio.log.
• A simulated endpoint-mismatch 400 fails fast (1 attempt, not 4) in harness retry tests.

Existing issue correlation

• Not [aw-failures] PR Code Quality Reviewer red — Copilot general-purpose subagent requests tier-unsupported model → SDK 400 `model [Content truncated due to length] #42095 (that is a subagent requesting a tier-unsupported model on PR Code Quality Reviewer, isModelNotSupportedError=true, no retry). Different agent, error class, and retry behavior.
• Not [aw-failures] Copilot CLI false-red — runs marked failure (exit 1) after safe-outputs succeed, via "numerous permission denied" [Content truncated due to length] #41636 (Copilot false-red after safe-outputs succeed). Here safe-outputs never ran.

References: §28425976809, §28422468107

Generated by 🔍 [aw] Failure Investigator (6h) · 143.3 AIC · ⌖ 32.3 AIC · ⊞ 5.2K · ◷

• expires on Jul 7, 2026, 12:13 AM UTC-08:00