AgentXchain v2.155.73
v2.155.73 publishes the recovery and release-readiness work accumulated since v2.155.72. The release focuses on governed-run resilience: recovery events are classified for operators, crash resume refuses duplicate live workers, continuous sessions keep their persisted checkpoints aligned, and the recent Claude/DOGFOOD recovery fixes are included in the shipped npm and Homebrew surfaces.
What's New
- Structured recovery classification. Recovery events are enriched as they are emitted, so current and historical governance reports can show recovery class, severity, actor action, operator action, and health score without relying on one-off prose.
- Crash-resume PID liveness guard.
step --resumechecks retained worker PIDs before reactivating active or blocked runs. Alive workers are rejected to prevent duplicate dispatch, while stale crash metadata is cleaned when the worker is gone. - Configurable turn deadlines. Local dispatch timeout now follows
per_turn_minutes, giving governed runs a config-backed deadline instead of a fixed operator assumption. - Restart-safe intake. Intent and event intake data survives process restarts, keeping continuous runs aligned with their governed work queue.
Bug Fixes
- Ghost blocker session consistency. Ghost blocker clearing now writes the corresponding
blocker_clearedsession checkpoint, and failed continuous steps are recovered only when persisted governed state is still active for the same session. - Claude recovery hardening. Retained Claude auth escalations, provider timeouts, Node incompatibility crashes, and refreshed-auth blockers now route through typed recovery paths with fail-closed negative cases.
- DOGFOOD-100 shipped-package helper.
agentxchain-dogfood-claude-smokeremains available as a direct npx bin for credential diagnostics without relying on a local checkout.
Evidence
- npm test -- --test-timeout=60000 -> 7386 tests / 1485 suites / 0 failures / 0 skipped
Tester Re-Run Contract
Verify: npx --yes -p [email protected] -c "agentxchain --version"