Skip to content

[aw-failures] [aw-fix] P1: Daily Cache Strategy Analyzer — Codex requests decommissioned model gpt-5-codex-alpha-2025-11-07 (404, not retrie [Content truncated due to length] #46301

Description

@github-actions

Pin the Codex request model to a supported name — the engine is resolving to the decommissioned alpha snapshot gpt-5-codex-alpha-2025-11-07, which 404s at the proxy and hard-fails the run before any agent work.

Parent: #46170

Problem statement

The Daily Cache Strategy Analyzer Codex run dies in Execute Codex CLI with a fatal 404 Not Found: Model not found gpt-5-codex-alpha-2025-11-07 from (apiproxy/redacted) The codex-harness classifies it isInvalidModelError=true and explicitly does **not** retry (invalid/unsupported model configuration — not retrying`), so the job exits 1 with no agent work. This is a deterministic config failure that will recur on every scheduled run until the model is fixed.

Affected workflows & run IDs

  • Daily Cache Strategy Analyzer (.github/workflows/daily-cache-strategy-analyzer.lock.yml) — representative run §29604657433 (2026-07-17T18:39Z).
  • No comparator success for this workflow in the 6h window; nearest prior failure is the same signature.

Probable root cause

The workflow's experiment spec (GH_AW_EXPERIMENT_SPEC) offers model_size variants gpt-5.4 / gpt-5.4-mini, wired to GH_AW_INFO_MODEL telemetry. But the actual Codex CLI request (engine codex v0.144.4) resolves to the hardcoded/default gpt-5-codex-alpha-2025-11-07 snapshot, which the api-proxy no longer serves. The experiment model_size value is not being propagated into the Codex request model — it only reaches telemetry, so the request falls back to the decommissioned alpha default.

Proposed remediation

  • Propagate the resolved model_size (gpt-5.4 / gpt-5.4-mini) into the Codex CLI request model (--model / codex config), not just GH_AW_INFO_MODEL, so no request ever falls back to gpt-5-codex-alpha-*.
  • Add a compile-time / startup validation that rejects unknown or alpha-snapshot Codex model names and fails fast with a clear message pointing at frontmatter.
  • Recompile affected Codex workflows against the current supported model set and remove any references to gpt-5-codex-alpha-2025-11-07.

Success criteria / verification

  • Execute Codex CLI for Daily Cache Strategy Analyzer no longer emits Model not found gpt-5-codex-alpha-2025-11-07.
  • Next scheduled run reaches the agent with a supported model and non-zero token usage.
  • No isInvalidModelError=true in codex-harness across 5 consecutive runs.

Related to #46170
Related to #46170

Generated by 🔍 [aw] Failure Investigator (6h) · 93 AIC · ⌖ 41.5 AIC · ⊞ 5.2K ·

  • expires on Jul 24, 2026, 11:20 AM UTC-08:00

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions