Skip to main content

AgentXchain v2.155.67

v2.155.67 closes BUG-112, found during the strict Tusq DOGFOOD-100 run after the session reached counter 97. A Claude QA turn exhausted provider API retries, returned Request timed out, and left no staged result, but the retained state surfaced as generic escalation:retries-exhausted:qa. Continuous startup now treats that evidence as a recoverable provider timeout and reissues the turn through the existing productive-timeout path.

Bug Fixes​

  • BUG-112: retained Claude provider timeout recovery. agentxchain run --continuous now inspects retained retries-exhausted dispatch logs for explicit provider timeout markers such as Request timed out.
  • No broad retry masking. Generic Claude api_retry events alone are not enough to trigger auto-reissue; the terminal log must contain explicit timeout evidence, the turn must have first output, and no staged result may exist.
  • Auth precedence preserved. Retained Claude auth failures still reclassify to dispatch:claude_auth_failed before timeout recovery is considered.

Evidence​

  • node --test cli/test/continuous-run.test.js -> 82 tests / 15 suites / 0 failures
  • node --test --test-name-pattern='BUG-100|BUG-111|BUG-112' cli/test/continuous-run.test.js -> 6 tests / 3 suites / 0 failures
  • npm test -- --test-timeout=60000 -> 7330 tests / 1485 suites / 0 failures / 0 skipped

Tester Re-Run Contract​

Verify: npx --yes -p [email protected] -c "agentxchain --version"

For BUG-112: resume tusq.dev from continuous session cont-7dc5b5df, run run_73ffb608f7c8a510, and QA turn turn_aa9664d36f8cac23 using the shipped package. If the retained dispatch log contains Request timed out, the turn has first output, and no staged result exists, startup must emit auto_retried_productive_timeout, clear escalation:retries-exhausted:qa, and continue the same session without agentxchain unblock, manual staging edits, or operator-side accept-turn.

If the reissued QA turn accepts, append counter value 98 to Tusq turn-counter.jsonl within the DOGFOOD-100 window.

For BUG-111 carry-forward: retained Claude 401/authentication failures must still reclassify to dispatch:claude_auth_failed before timeout recovery.

For BUG-110 carry-forward: fresh Claude local CLI auth failures must still block once as dispatch infrastructure instead of consuming turn retries.

For BUG-109 carry-forward: supplemental auto-checkpoint recovery must still run automatically under --auto-checkpoint.

For BUG-108 carry-forward: terminal blocked steps must surface once and must not re-enter paused-active recovery loops.

For BUG-107 carry-forward: paused active continuous sessions must still preserve the existing session ID and recover only when the governed run is active, unblocked, and dispatchable.

Carry-forward tester checks for shipped-package closure proof:

  • BUG-54 startup watchdog evidence should quote startup_latency_ms, elapsed_since_spawn_ms, and runtimes.<id>.startup_watchdog_ms.
  • BUG-55/87 verification-output evidence should quote undeclared_verification_outputs, verification.produced_files, and git status --short after accept-turn followed by checkpoint-turn. Clean tree means BUG-55 is fixed for your reproduction.
  • BUG-52/53 continuous recovery evidence should quote phase_entered, reconciled_before_dispatch, and session_continuation.