What problem does it solve?
Coordinating multi-task implementation work across AI agent sessions is error-prone: tasks get skipped, workers are duplicated after lost responses, and completion claims are accepted without proof. This Skill turns a spec's task graph into a disciplined delegation loop where every task runs in its own bounded worker session and only the task file's on-disk status counts as done.
Core Features & Use Cases
- Graph-ordered task queue: Reads
.compozy/tasks/<slug>/_tasks.md and each task_NN.md frontmatter to queue pending or in-progress tasks in dependency order.
- Bounded worker spawning: Creates one TTL-limited, parent-bound CompozyOS session per task with idempotency keys, runtime selection by task type, and reconciliation logic that prevents duplicate workers after ambiguous spawn responses.
- Proof-based completion: Dispatches a structured briefing, waits on the blocking prompt, and accepts only
status: completed in the task file frontmatter, with one corrective prompt before marking the task blocked.
- Use Case: Given a spec slug with eight pending tasks, run this Skill to sequentially spawn eight worker sessions, collect per-task JSONL event logs, and produce a structured report mapping each task id to its worker session id.
Quick Start
Orchestrate the tasks for the spec slug 'auth-refactor' under .compozy/tasks/ using the code_implementer agent across worker sessions.
