Document stderr regex branch unreachability and add turn.failed fixture-driven tests (#4122)

## Summary

Add a clarifying comment in `_exit_classification.py` documenting that
the stderr regex branch
(`_CODEX_CONTEXT_EXHAUSTION_PATTERN.search(result.stderr)`) is
unreachable for Codex because `turn.failed` events arrive via stdout
NDJSON, not stderr. Extend the `_turn_failed_ndjson` test helper with an
optional `error_code` kwarg and add five new test methods exercising all
`error.code` / `error.message` variants of `turn.failed` classification
plus a stderr boundary test.

## Implementation Plan

Plan file:
`/home/talon/projects/autoskillit-runs/impl-20260627-014523-646887/.autoskillit/temp/make-plan/t5_p6_a14_wp1_stderr_regex_unreachable_plan_2026-06-27_015000.md`

Closes #4050

🤖 Generated with [Claude Code](https://claude.com/claude-code) via
AutoSkillit
<!-- autoskillit:pipeline-signature
steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr
-->

## Token Usage Summary

| Step | Model | count | uncached | output | cache_read | peak_ctx |
turns | cache_write | time |

|------|-------|-------|----------|--------|------------|----------|-------|-------------|------|
| plan* | opus[1m] | 1 | 3.5k | 7.6k | 672.2k | 68.1k | 26 | 54.1k | 4m
17s |
| verify* | sonnet | 1 | 70 | 7.1k | 390.3k | 70.7k | 23 | 52.2k | 4m 9s
|
| implement* | MiniMax-M3 | 1 | 62.7k | 6.2k | 1.3M | 0 | 68 | 0 | 2m
13s |
| audit_impl* | sonnet | 1 | 1.6k | 6.3k | 159.7k | 38.4k | 14 | 27.7k |
2m 56s |
| prepare_pr* | MiniMax-M3 | 1 | 40.1k | 1.7k | 110.1k | 0 | 11 | 0 |
38s |
| compose_pr* | MiniMax-M3 | 1 | 36.5k | 1.5k | 249.1k | 0 | 14 | 0 |
55s |
| **Total** | | | 144.5k | 30.2k | 2.9M | 70.7k | | 134.0k | 15m 10s |

\* *Step used a non-Anthropic provider; caching behavior may differ.*

## Token Efficiency

| Step | LoC Changed | cache_read/LoC | cache_write/LoC | output/LoC |
|------|-------------|----------------|-----------------|------------|
| plan | 0 | — | — | — |
| verify | 0 | — | — | — |
| implement | 49 | 27125.6 | 0.0 | 126.1 |
| audit_impl | 0 | — | — | — |
| prepare_pr | 0 | — | — | — |
| compose_pr | 0 | — | — | — |
| **Total** | **49** | 59398.6 | 2735.1 | 617.3 |

## Model Usage Breakdown

| Model | steps | uncached | output | cache_read | cache_write | time |
|-------|-------|----------|--------|------------|-------------|------|
| opus[1m] | 1 | 3.5k | 7.6k | 672.2k | 54.1k | 4m 17s |
| sonnet | 2 | 1.7k | 13.4k | 550.0k | 79.9k | 7m 6s |
| MiniMax-M3 | 3 | 139.4k | 9.3k | 1.7M | 0 | 3m 46s |

---------

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: Trecek

src/autoskillit/execution/session/_exit_classification.py

@@ -115,6 +115,7 @@ def classify_infra_exit(
     if capabilities is None or capabilities.supports_context_exhaustion_detection:
         if session._is_context_exhausted():
             return InfraExitCategory.CONTEXT_EXHAUSTED
+        # Codex turn.failed arrives on stdout (NDJSON); this stderr branch is non-Codex fallback.
         if _CODEX_CONTEXT_EXHAUSTION_PATTERN.search(result.stderr):
             return InfraExitCategory.CONTEXT_EXHAUSTED
     # Rate limit detection — must precede all API_ERROR checks (including

tests/execution/test_exit_classification.py

@@ -56,8 +56,11 @@ def _sr(
     )
 
 
-def _turn_failed_ndjson(error_message: str) -> str:
-    return json.dumps({"type": "turn.failed", "error": {"message": error_message}})
+def _turn_failed_ndjson(error_message: str, *, error_code: str = "") -> str:
+    error: dict[str, str] = {"message": error_message}
+    if error_code:
+        error["code"] = error_code
+    return json.dumps({"type": "turn.failed", "error": error})
 
 
 class TestClassifyInfraExit:
@@ -412,6 +415,49 @@ def test_turn_failed_rate_limit_exceeded_jsonl_context_exhausted_false(self) ->
         adapted = _adapt_agent_result(agent_result)
         assert adapted.jsonl_context_exhausted is False
 
+    def test_turn_failed_error_code_only_classify_returns_context_exhausted(self) -> None:
+        ndjson = _turn_failed_ndjson(
+            "Agent terminated unexpectedly", error_code="context_length_exceeded"
+        )
+        agent_result = CodexResultParser().parse_stdout(ndjson)
+        adapted = _adapt_agent_result(agent_result)
+        result = _sr(returncode=1)
+        assert classify_infra_exit(adapted, result) == InfraExitCategory.CONTEXT_EXHAUSTED
+
+    def test_turn_failed_error_code_and_message_classify_returns_context_exhausted(self) -> None:
+        ndjson = _turn_failed_ndjson(
+            "The model's context length has been exceeded",
+            error_code="context_length_exceeded",
+        )
+        agent_result = CodexResultParser().parse_stdout(ndjson)
+        adapted = _adapt_agent_result(agent_result)
+        result = _sr(returncode=1)
+        assert classify_infra_exit(adapted, result) == InfraExitCategory.CONTEXT_EXHAUSTED
+
+    def test_turn_failed_message_only_classify_returns_context_exhausted(self) -> None:
+        ndjson = _turn_failed_ndjson("context_length_exceeded", error_code="")
+        agent_result = CodexResultParser().parse_stdout(ndjson)
+        adapted = _adapt_agent_result(agent_result)
+        result = _sr(returncode=1)
+        assert classify_infra_exit(adapted, result) == InfraExitCategory.CONTEXT_EXHAUSTED
+
+    def test_turn_failed_non_exhaustion_error_code_not_context_exhausted(self) -> None:
+        ndjson = _turn_failed_ndjson("Internal server error", error_code="server_error")
+        agent_result = CodexResultParser().parse_stdout(ndjson)
+        adapted = _adapt_agent_result(agent_result)
+        result = _sr(returncode=1)
+        assert classify_infra_exit(adapted, result) == InfraExitCategory.API_ERROR
+
+    def test_stderr_non_exhaustion_not_context_exhausted(self) -> None:
+        session = ClaudeSessionResult(
+            subtype="success",
+            is_error=False,
+            result="",
+            session_id="s1",
+        )
+        result = _sr(returncode=1, stderr="some transient error")
+        assert classify_infra_exit(session, result) == InfraExitCategory.COMPLETED
+
 
 class TestRateLimitClassification:
     def test_rate_limited_text_classified_as_rate_limited(self) -> None: