teamharness-task-execution

# Task Execution Use this skill when executing assigned work as a Worker or remote member. Claim only tasks assigned to you. Read the task spec, keep deliverables in the task directory, ask focused blocker questions, and submit a structured result. Do not change project-level state or project result files. Do not use this skill or taskflow for readiness checks, direct questions, or explicit requests to reply with specific text. Answer those directly in the current room. ## Task Directory Your assigned task lives under: ```text shared/tasks/{task-id}/ ``` Your Leader owns: ```text shared/tasks/{task-id}/meta.json shared/tasks/{task-id}/spec.md ``` You own: ```text shared/tasks/{task-id}/workspace/ shared/tasks/{task-id}/progress/ shared/tasks/{task-id}/result.md ``` Do not edit: ```text shared/projects/{project-id}/meta.json shared/projects/{project-id}/plan.md shared/projects/{project-id}/result.md ``` If a task spec asks you to write or submit `shared/projects/...`, report that boundary conflict to the Leader. Put Worker-owned deliverables under `shared/tasks/{task-id}/...`; your Leader owns project-level reports. ## Acknowledge When you are mentioned with `TASK_ASSIGNED`, first pull or inspect the task directory if needed, then acknowledge with `taskflow`: ```json { "role": "worker", "action": "ack_task", "payload": { "taskId": "demo-project-001-01" } } ``` If `ack_task` fails because the task is missing or assigned elsewhere, stop and report the blocker to the Leader in the current room. Do not invent task files. Read: ```text shared/tasks/{task-id}/spec.md ``` before doing the work. ## Execute Keep all deliverables under the task directory. If you need private notes, use: ```text shared/tasks/{task-id}/workspace/ ``` Use meaningful filenames for deliverables. Prefer names that describe the artifact's purpose, such as `analysis.md`, `trace-summary.log`, or `patch-notes.md`; avoid vague names like `output.md` when there are multiple files. For report-style tasks whose expected output is `shared/tasks/{task-id}/result.md`, write the full report content to that file before calling `submit_task`. `submit_task` records structured status in task metadata and does not create or rewrite `result.md`. For long-running tasks, report progress periodically with the `report_progress` action (a short `note`, ≤ 200 chars). It records a progress entry in the task's audit history without changing task state or notifying the room — use it to make multi-stage work visible, not as a status update. If blocked, submit a `BLOCKED` result instead of silently waiting. ## Submit Submit with `taskflow`: ```json { "role": "worker", "action": "submit_task", "payload": { "taskId": "demo-project-001-01", "status": "SUCCESS", "summary": "Completed the assigned work.", "deliverables": [ "shared/tasks/demo-project-001-01/workspace/output.md" ] } } ``` Use one of: - `SUCCESS` - `SUCCESS_WITH_NOTES` - `PARTIAL` - `REVISION_NEEDED` - `BLOCKED` - `INTERRUPTED` Use `INTERRUPTED` when execution stopped before you could finish. If your Leader accepts either `INTERRUPTED` or `BLOCKED`, TeamHarness records the task and plan node as `blocked` and resolves the continuation with `resolution: blocked`. Your first persisted submission records `submission_id`, UTC `submitted_at`, `result_digest`, and a pending `continuation` marker in TaskMeta. Treat `submission_id` as an opaque fence: compare it for equality, but do not parse it or assume a UUID format. The digest covers your trimmed status, whitespace- collapsed summary, and validated deliverable paths in their persisted order; it does not cover notes or the rendered `result.md` text. If shared-storage sync is interrupted, retry with exactly the same status, summary, and ordered deliverables. Your retry reuses the original submission, timestamp, digest, and continuation `delivery_id`, and repairs missing project or task projections. If you change any digest input, the retry conflicts with the submitted task and you must wait for a Leader decision or a new task. A pending continuation is durable state for a future Controller; it does not mean that a Matrix wake was sent or that your Leader has already resumed the task. You cannot accept, reject, cancel, or resolve your own submission. Those are trusted-Leader-only decisions, and putting `role: leader` in a payload cannot override your Worker runtime identity. After submitting, use the returned `submissionId` only as an opaque value in reports or exact retries. For a legacy task already marked `submitted` without a submission identity, retry only when TaskMeta has `submitted_at`, and send the complete original status, summary, and ordered deliverables. CoPaw adopts it only if that result exactly matches the persisted result; otherwise it fails closed. Do not invent an identity or try to resolve the legacy task yourself. Submitting ends the task. Do not keep editing the old task after submission unless the Leader assigns a new task. If you already wrote a detailed `shared/tasks/{task-id}/result.md`, include that path in `deliverables`; do not replace the report with only a short summary. Do not include `shared/projects/...` paths in `deliverables`. When Matrix room context is available, `submit_task` automatically publishes `shared/tasks/{task-id}/result.md` and eligible deliverable files under the current task directory as Matrix `m.file` events. Check `publishedArtifacts` in the tool result for published, skipped, or failed artifact status. Do not upload task artifacts manually with `message` just to make them appear in the room file panel. ## Completion Message `submit_task` automatically publishes the completion event to the Task room: the first line is the contract below, and the event @mentions the Leader and the human members of the team (the task initiator), so the requester is routed with the same salience the leader is. After `submit_task` returns `ok: true`, do not send another completion line. The event first line carries one token per result status (code-generated): ```text @leader-user:matrix.local TASK_COMPLETED: demo-project-001-01 - Result: shared/tasks/demo-project-001-01/result.md @leader-user:matrix.local TASK_REVISION_NEEDED: demo-project-001-01 - <summary> @leader-user:matrix.local TASK_BLOCKED: demo-project-001-01 - <short blocker summary> @leader-user:matrix.local TASK_INTERRUPTED: demo-project-001-01 - <summary> ``` If the task spec gives an exact completion line, the code event keeps that line format; a short human-readable summary message may still follow in the room but never replaces the event. Do not use `NO_REPLY` after successful submission. While the task is still in flight and you need a human decision (approval / decision / escalation) instead of guessing, call `taskflow` with `action: request_attention` (payload: `kind`, `question`). Do not rely on the human noticing an ambient room message. Once the decision is made, close the loop by calling again with the same `kind` plus `resolved: true` — closing needs an existing record of that kind (it is rejected without one) and reuses the original event (no new ping).