fix(chat-path): classify signal exits before auth fallback (#906)

[oleksandr-korin]

Summary

• Add the missing _classify_signal_exit call to the chat path (claude_code.py:450), mirroring the existing pattern in headless_executor.py:683 that was added for Issue bug: SIGKILL/timeout terminations of claude subprocess misclassified as authentication failure #516
• A SIGKILL at 0 turns (OOM, watchdog, operator cancel) no longer gets misclassified as "Subscription token may be expired"
• Regression test pins the call-ordering invariant in both files

Background

Issue #516 introduced _classify_signal_exit to fix the SIGKILL → "token expired" misclassification in the headless path (/api/task, scheduled tasks). The fix never landed on the parallel chat path (/api/chat, interactive chats), so the same misclassification persists there.

Two-codepath analysis from the issue's diagnosis:

Codepath	File	Called _classify_signal_exit first?
Scheduled tasks (/api/task)	headless_executor.py:683	✅ Yes (Issue #516)
Interactive chat (/api/chat)	claude_code.py:453	❌ No — bug

The misclassification has two compounding effects beyond bad operator-facing text:

1. Operator misdirection — operators chase token regeneration when the actual cause is OOM / timeout / external kill (the issue cites a 72-hour wrong-fix cycle)
2. SUB-003 auto-switch state corruption — the auto-switch trigger pattern matcher reads the error string and rotates subscriptions on phantom auth failures, burning the 2-hour skip-list slot

Changes

docker/base-image/agent_server/services/claude_code.py (+12)

• Add _classify_signal_exit to the imports from error_classifier
• Insert the signal-exit check before _diagnose_exit_failure in the chat handler's if return_code != 0 block. Pattern matches headless_executor.py:683-687 exactly.

tests/unit/test_signal_exit_classification.py (+52)

• Add a structural regression test parametrized over both files (claude_code.py, headless_executor.py)
• Test asserts _classify_signal_exit( appears before _diagnose_exit_failure( in each file's source. Order reversal in either file would re-introduce the misclassification — verified by manually flipping the order and confirming the test fails.

Out of scope (deferred to follow-ups)

• Fix 2 (gate auto-switch on observed wire signal) — adding a structured error_class field to executions and refactoring the SUB-003 matcher to read it instead of the freeform error text. Larger refactor; filed for follow-up.
• Fix 3 (cgroup OOM event reading) — upgrading SIGKILL classification to "container OOM observed" when memory.events.oom_kill incremented. Optional per the issue body.

Deployment

This change touches docker/base-image/agent_server/. Per CLAUDE.md note #7, base image rebuild + agent restart is required for the fix to take effect on running agents:

./scripts/deploy/build-base-image.sh
# Then restart agents (e.g. via /agent stop+start in UI, or fleet-wide restart)

Until the rebuild + restart, agents already running will continue to emit the old misclassified error.

Test plan

• Unit tests: python3.11 -m pytest tests/unit/test_signal_exit_classification.py --confcutdir=tests/unit -q — 23 passed (21 existing + 2 new parametrized cases)
• Regression test verified by temporarily reversing the call ordering in claude_code.py — test fails as expected, then restores
• Manual verification post-deploy: trigger a SIGKILL on an agent's chat session (e.g. via cgroup OOM with small memory limit + memory-allocating skill) and observe the recorded error is "Execution terminated by SIGKILL after 0 tool calls / 0 turns…" instead of "Subscription token may be expired…"
• Verify SUB-003 auto-switch no longer fires on the SIGKILL repro (subscription should remain healthy / unrotated)

Fixes #906

🤖 Generated with Claude Code