fix(codex): auth-aware default model — fixes codex empty builds (#82 Gap 3)

[AbirAbbas]

Summary

Root-causes and fixes the codex empty-builds observation in #82 (Gap 3). It is not a structured-output problem — it's a model/auth mismatch, confirmed live.

SWE-AF defaulted every codex role to gpt-5.3-codex. The -codex models are only available with OpenAI API-key auth — under ChatGPT-account auth they return:

HTTP 400 invalid_request_error:
"The 'gpt-5.3-codex' model is not supported when using Codex with a ChatGPT account."

…in ~3 seconds. The coder turned that fast error into files_changed: [] → "Coder agent failed" → the foundation issue failed → cascade → empty build. That's the reporter's exact symptom (~9–11s, codex-only; claude_code fine, because it never touched codex).

Fix

1. Auth-aware codex default model (schemas.py, fast/schemas.py)
Resolve the codex base model by auth mode instead of a constant:

• API-key auth (SWE_CODEX_AUTH_MODE=api_key, or auto with OPENAI_API_KEY set) → gpt-5.3-codex (unchanged).
• ChatGPT-account auth (chatgpt, or auto with no key) → gpt-5.5 (a ChatGPT-compatible model).
  Explicit overrides (models.default / per-role / SWE_DEFAULT_MODEL) still win.

2. Codex model/auth errors are fatal + surfaced (fatal_error.py, execution_agents.py)
A model/auth 400 is non-retryable, so:

• check_fatal_harness_error now matches "not supported when using Codex with a ChatGPT account" and "requires a newer version of Codex" → raises FatalHarnessError with the real message, short-circuiting the retry cap.
• run_coder includes the underlying error_message in the CoderResult summary when no result parses, so an empty result always carries why (the reporter saw a bare "Coder agent failed" with no reason).

Validation Contract

• ChatGPT-auth codex default = gpt-5.5; API-key-auth codex default = gpt-5.3-codex; explicit overrides win; claude_code/open_code unaffected.
• A codex "model not supported on ChatGPT account" / "needs newer CLI" error is fatal (no silent retry) and its message reaches the build output.

Test Plan — verified live (codex CLI 0.142.4, ChatGPT auth)

• gpt-5.3-codex → 400 "not supported … ChatGPT account" in 3s; resolved gpt-5.5 wrote a real hello.py in 8s.
• resolve_runtime_models / fast_resolve_models return gpt-5.5 under chatgpt env, gpt-5.3-codex under api_key env (new unit tests).
• New fatal-pattern tests for both codex error strings.
• Full make check on py3.12: 1010 passed, 1 skipped.

Scope

SWE-AF-only; independent of #84 (Gaps 1 & 2) and of the agentfield SDK — no cross-repo dependency. Tip for ChatGPT-plan users: gpt-5.5 runs at high reasoning effort (slower per call); set models.coder or SWE_DEFAULT_MODEL to override if you want lower latency.

Refs #82 (Gap 3).

🤖 Generated with Claude Code