task-release

/taskwarrior:task-release

Release an active claim without closing the task. Pairs with /taskwarrior:task-claim: claim picks the task up, release puts it back down.

When to Use This Skill

Use this skill when...	Use a sibling skill instead when...
Pausing a task mid-flight (lunch break, context switch, end of session)	The work has landed in a commit — use /taskwarrior:task-done
Handing off to another agent — leaving an annotation describing where you stopped	Starting work on a fresh task — use /taskwarrior:task-claim
Aborting a started task because scope was wrong	Filing a brand-new task to capture the work — use /taskwarrior:task-add

Context

• Task CLI available: !task --version
• Git repo detected: !find . -maxdepth 1 -name '.git' -print -quit

Git probes (git rev-parse --show-toplevel, git branch --show-current) write to stderr in a no-git cwd, and stderr from a Context backtick aborts the skill before its body runs. Git toplevel and current-branch resolution are done in the body via the Bash tool (where 2>/dev/null and exit-code handling are tolerated).

Parameters

Parse $ARGUMENTS:

• $0 — task ID (required).
• Remaining positional words — concatenate as the freeform state message for the handoff annotation. If empty, prompt for one before annotating (an empty handoff annotation defeats the purpose).
• --clear-identity — clear agent / host / branch / worktree UDAs in addition to pid. Use when the next agent should not inherit any context (e.g. cross-host handoff). Default behaviour preserves these so the queue still shows who last touched the task.
• --no-coworker-marker — skip the /git:coworker-check --release step. By default, releasing the task also drops the git-side session marker so destructive ops in this clone are no longer guarded.

Execution

Execute this release workflow:

Step 1: Load the task

task "$TASKID" export | jq '.[0] | {id, uuid, description, status, start, agent, pid, host, branch, worktree, annotations}'

Use export | jq — never info / list. Capture the uuid field as $TASK_UUID — every mutating call from Step 3 onward addresses the task by $TASK_UUID, never $TASKID. Step 2's AskUserQuestion pause is a real gap during which another agent's concurrent task done can renumber $TASKID (see .claude/rules/task-id-stability.md). Decide:

Condition	Action
status is not pending	Abort. Released tasks must be pending; if it is already closed, the release is a no-op.
start is unset (not +ACTIVE)	Warn but proceed — annotation + UDA drain still useful as a "I looked at this" signal.
start is set	Standard path. Continue.

Step 2: Resolve state message

If the user did not pass a message, ask them for one with a single AskUserQuestion. The annotation is the handoff payload — without it the next agent has only timestamps. Keep the prompt focused: "Where did you stop? (one sentence)".

Step 3: Annotate the handoff

Annotate before stopping — if the stop fails (e.g. concurrent modify), the annotation is still captured:

task "$TASK_UUID" annotate "released by $AGENT — state: $MESSAGE"

$AGENT is read from the task's existing agent UDA (the claim stamped it); fall back to claude-${CLAUDE_SESSION_ID:0:8} if unset.

Step 4: Stop the active state

task "$TASK_UUID" stop

This clears the built-in start attribute and removes the +ACTIVE virtual tag. The task is back in the dispatch pool for /taskwarrior:task-coordinate.

Step 5: Drain identity UDAs

Always clear pid (the process is going away):

task "$TASK_UUID" modify pid:

When --clear-identity is passed, also clear the rest:

task "$TASK_UUID" modify agent: host: branch: worktree:

The empty-value field: syntax removes a UDA in taskwarrior. Default behaviour keeps agent / host / branch / worktree so task-status "Recently touched" can attribute the work.

Step 6: Drop the coworker-check marker

Unless --no-coworker-marker was passed:

Use SlashCommand to invoke `/git:coworker-check --release`.

Fallback if SlashCommand is unavailable:

bash "$(git rev-parse --show-toplevel)/git-plugin/skills/git-coworker-check/scripts/release-session.sh" --project-dir "$(pwd)"

Step 7: Report

Print:

• Task ID, description, current status (still pending)
• Time spent active (from start timestamp to now — read before stopping)
• Annotation appended
• Identity drain: pid only / full drain
• Coworker-check status: marker released / skipped
• Suggested next step:
  • /taskwarrior:task-coordinate to dispatch the now-unclaimed work onto the next wave
  • /taskwarrior:task-claim <id> to re-claim later

Agentic Optimizations

Context	Command
Read existing claim before release	task _get "$TASKID".agent (single-field, exit 0 even when unset)
Time-spent computation	task "$TASKID" export | jq '.[0].start' then compute against now
Bulk release of stale claims	Pair with task +ACTIVE start.before:now-4h export | jq '.[].id'
Skip empty-result failures	Always export | jq, never task active list

Quick Reference

Step	Command
Annotate	task UUID annotate "released by AGENT — state: MSG"
Stop active	task UUID stop
Clear pid	task UUID modify pid:
Full identity drain	task UUID modify agent: host: branch: worktree: pid:
Drop git marker	/git:coworker-check --release

Related

• /taskwarrior:task-claim — pair: pick up an unclaimed task
• /taskwarrior:task-done — close after landing (also auto-stops + drains identity)
• /taskwarrior:task-coordinate — finds the next dispatch candidate
• /git:coworker-check — sister marker that this skill drops on --release
• .claude/rules/agent-coworker-detection.md — combined-signal rationale
• .claude/rules/task-id-stability.md — why this skill resolves $TASK_UUID once and mutates by UUID