What problem does it solve? When developing with an AI pipeline, new requirements keep occurring mid-run, and a fixed batch plan forces you to stop, update, and re-invoke every time. This Skill keeps a task pool open so tasks you add while it runs are picked up automatically, and a single task failure never halts the whole run. ## Core Features & Use Cases - Living task pool: Re-reads docs/batches/<pool-id>/BATCH_PLAN.md every iteration so newly appended pending rows are planned into the topological frontier without re-invoking. - Best-effort execution: A failed task is marked and skipped, a needs-human task is deferred and surfaced in the report's ## Needs your input section, and only genuine hazards (baseline verify_all FAIL, STOP intervention, safety-hook block) halt the stream. - Ambient mode: Invoking with no pool-id turns every chat message into a heartbeat that folds new requirements into the default pool and drains it, gated by a session-scoped .harness/ambient.flag. - Use Case: While building a feature you keep thinking of follow-ups — "also add CSV export", "fix the login redirect" — so you run the stream under /loop, type requirements as they occur, and only watch the per-task verdicts and final STREAM_REPORT.md. ## Quick Start Run /harness-stream with my pool-id (or no argument for ambient mode) and keep draining every task I add to the pool, reporting failures without stopping.
