Classify unsupported Codex custom tools as invalid-model failures - #57578

Fixed in b920635. The detector now parses the raw turn.failed JSONL event and unwraps the escaped nested provider JSON structurally; the regression test uses the exact escaped event shape.

The new detector hard-codes a very narrow JSON text shape for the Codex tools: custom rejection. We already have extractNestedProviderErrorDetails() / isUnsupportedModelToolsError() that parse nested provider payloads structurally, and those tests cover envelopes like metadata.raw; this regex will miss the same error as soon as the SDK emits a slightly different serialization or escapes the JSON differently, which means the workflow falls back to the generic engine-failure path again.

function isUnsupportedModelToolsModelError(output) {
  // parse each JSON event/object line, unwrap nested payloads, then check
  // message === "Invalid value: 'custom'" && param === "tools".
}

modelNotSupportedError:
  MODEL_NOT_SUPPORTED_PATTERN.test(logContent) || isUnsupportedModelToolsModelError(logContent)

Fixed in b920635. The serialized-JSON regex was removed and replaced with bounded structural traversal of error, nested message JSON, and metadata.raw payloads.

This reordered-payload test only exercises MODEL_NOT_SUPPORTED_PATTERN.test(...), so it can still pass if detectErrors(...) stops wiring that match into modelNotSupportedError. The previous case checks end-to-end classification for the canonical ordering, but this one leaves the exact regression you are adding here partially unguarded.

expect(MODEL_NOT_SUPPORTED_PATTERN.test(errorOutput)).toBe(true);
expect(detectErrors(errorOutput).modelNotSupportedError).toBe(true);

Fixed in b920635. The reordered raw payload now asserts detectErrors(...).modelNotSupportedError end to end.

The [^{}]{0,300}? span limit guards against cross-object matches, but 300 characters may be too tight if the API returns an error object with a long details, description, or localized message field between the two required keys. Consider raising the cap (e.g. {0,1000}) or adding an inline comment explaining why 300 was chosen, so future maintainers understand the deliberate trade-off.

@copilot please address this.

Obsolete after b920635: the capped regex span was removed in favor of structural JSON parsing.

actions/setup/js/detect_agent_errors.cjs:160: shrink: the two field-order branches in the new Codex regex are duplicated. Replace them with a small helper that checks both substrings once instead of a giant alternation.

Fixed in b920635. The duplicated regex branches were removed and replaced with a structural helper.

actions/setup/js/detect_agent_errors.test.cjs:163: delete: three new tests cover the same Codex error shape with only field-order and object-boundary differences. One happy-path test plus one reordered-field case is enough.

Addressed in b920635. Coverage now focuses on the raw escaped event, reordered fields, and the required same-object negative guard.

[/diagnosing-bugs] The {0,300} character budget might silently miss real Codex payloads if the API ever adds more fields between message and param (e.g. a long details string). Consider raising the limit to {0,1000} — false negatives (missed classifications) are the risk here.

// Before: [^{}]{0,300}?
// After:  [^{}]{0,1000}?

@copilot please address this.

Obsolete after b920635: structural parsing replaces the fixed character-budget regex.

[/tdd] The test for reversed field order doesn't call detectErrors() to confirm the full classification pipeline, unlike the first test. For completeness, add expect(detectErrors(errorOutput).modelNotSupportedError).toBe(true) so we verify the end-to-end path (not just the raw pattern).

it("matches when the Codex error fields are reordered", () => {
  const errorOutput = `{ ... }`;
  expect(MODEL_NOT_SUPPORTED_PATTERN.test(errorOutput)).toBe(true);
  // Add this:
  expect(detectErrors(errorOutput).modelNotSupportedError).toBe(true);
});

@copilot please address this.

Fixed in b920635. The reordered event now exercises the full detectErrors(...) classification path.