Repository navigation
Tags
Tags
fix: harden autonomous continuation and cut 0.13.2 Add work.stop handling, tighten autonomous continuation guards, persist stop state consistently, validate restored autonomous ownership via pid liveness, and add focused tests for work.stop and agent_end behavior. Also bump version to 0.13.2 and update changelog.
feat(crew): auto-review tasks after worker completion Calls reviewImplementation sequentially for each succeeded task between worker results and wave reporting. SHIP keeps it done, NEEDS_WORK resets via resetTask for retry with feedback, MAJOR_RETHINK blocks. Gated by existing review.enabled and review.maxIterations config. Adds review_count to Task, task.review feed event type, and tests.
feat(crew): dynamic skill loading for workers Workers acquire domain knowledge on demand via three-tier skill discovery (user, extension, project). Planner tags tasks with relevant skills; workers see recommendations and load via read(). Zero token cost when unconfigured. Also fixes: artifact dir creation, multiline feed sanitization, config model override.
fix: split provider/model into separate CLI flags for worker spawn When spawning workers with a model like 'zai/glm-5', pi's model resolver matched it as a literal model ID under vercel-ai-gateway instead of interpreting 'zai' as the provider. Added pushModelArgs helper that splits on the first slash into --provider and --model flags, used by both task workers and lobby workers.
feat: v0.12.0 — overlay redesign, advisory deps, legacy routing removal Unified crew dashboard overlay with dynamic width, live worker progress, task management, chat input, and feed notifications. Advisory dependency mode makes all todo tasks schedulable. Lobby workers, task splitting, task revision, prompt-based planning, and worker coordination system. Removed legacy key-based routing (join/claim/swarm/etc parameters) in favor of the action-based API. Fixed overlay rendering corruption from multi-line bash args, SIGKILL escalation, and numerous worker lifecycle bugs. Default planning passes reduced from 3 to 1. 281 tests across 23 suites.
PreviousNext