Close a taskwarrior task with landing commit annotation and optional GitHub issue/PR close. Use when finishing a coordination task or marking a work order complete.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Close a taskwarrior task with landing commit annotation and optional GitHub issue/PR close. Use when finishing a coordination task or marking a work order complete.
Git probes (git remote, git rev-parse --short HEAD,
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. HEAD
commit, current branch, and remote-presence checks happen in the body
(Steps 2 and 5) via the Bash tool where 2>/dev/null and exit-code
handling are tolerated.
Parameters
Parse $ARGUMENTS:
$0 — task ID (required)
$1 — commit hash (optional; defaults to HEAD)
--no-gh — skip GitHub close/comment even when remote is present
--no-tracker — skip blueprint tracker drain
--drain-identity — also clear agent / pid / host / branch / UDAs after closing (default keeps them as audit trail)
worktree
--no-coworker-marker — skip the /git:coworker-check --release step
Execution
Execute this workflow:
Step 1: Load the task
task "$TASKID"export | jq '.[0]'
Never use task $TASKID info or task $TASKID list — both can exit 1 and
cancel parallel siblings. export | jq returns valid JSON even when the
task is already closed (treat empty as "no such open task" and abort).
Capture: uuid, bpid, bpdoc, ghid, ghpr, tags, description, plus
identity UDAs agent, pid, host, branch, worktree, and start
(if the task was claimed via /taskwarrior:task-claim).
Capture uuid as $TASK_UUID — every mutating call from Step 2 onward
(including the post-close Step 6 query) addresses the task by $TASK_UUID,
never $TASKID (see .claude/rules/task-id-stability.md).
Step 2: Resolve commit hash
If $1 is unset, read git rev-parse --short HEAD. Confirm the HEAD commit
actually touches work this task covers before annotating — a stale HEAD is
a common footgun.
Annotation first, then done — if close fails (e.g. dependencies), the
annotation is still captured.
Taskwarrior auto-stops a +ACTIVE task on done, so an explicit
task stop is not needed. The task transition removes +ACTIVE and
records the duration. If you want to drain the identity UDAs (so the
closed task does not retain the stamp), do so separately after the
close — see Step 4b below.
Step 4b: Drain identity UDAs (optional)
After the task is closed, optionally clear the identity stamp left by
the original claim:
Default behaviour is to leave these set on closed tasks — the audit
trail of "who claimed and landed this" is useful in task-status
recently-completed reports. Drain them only when the user explicitly
asks (e.g. compliance / privacy hygiene), or when handing the queue
file off to another team.
Step 4: Drain the blueprint tracker
If bpdoc is set and points to a valid path, read the file and advance
its status marker in place. Typical patterns by tracker format:
Work-order markdown: check off the relevant bullet, add landed-in line
PRP: flip the "Implementation" status field
Use Edit with a narrow old_string / new_string pair — do not rewrite
the file. When bpdoc references a shared tracker (manifest.json, global
feature-tracker), fall back to reporting "manual tracker update required"
rather than concurrent-write the shared file.
Step 5: Close linked GitHub items (optional)
When GitHub mode is active and --no-gh was not passed:
ghid set → offer gh issue close "$GHID" --comment "Closed by $COMMIT_SHORT (branch $BRANCH)"
Always confirm before mutating GitHub state — the user may want to close
the issue as part of the PR merge rather than ahead of time.
Step 5b: Drop the coworker-check marker
If the task was +ACTIVE (claimed via /taskwarrior:task-claim), the
matching git-side session marker should be released so destructive ops
in this clone are no longer guarded:
Use SlashCommand to invoke `/git:coworker-check --release`.
Skip when the user has more work in flight on this branch — releasing
the marker lifts the cross-agent guard.
Step 6: Report
Print:
Task closed: id + description + new status
Commit annotated: $COMMIT_SHORT
Tracker drained: path + diff summary, or "skipped"
Unblocked siblings: any tasks whose depends: pointed at this one now
free to start (query via task depends:"$TASK_UUID" export | jq). Once
completed, this task's numeric id becomes 0 and may be reassigned to
an unrelated task — a query keyed on the stale $TASKID here would check
dependencies against the wrong task. depends: accepts UUIDs, so
$TASK_UUID is the only safe form for this read.
Agentic Optimizations
Context
Command
Load task
task "$TASKID" export | jq '.[0]' (capture .uuid as $TASK_UUID)
Annotate + close
Two separate calls (hook-friendly), both by $TASK_UUID
Blocked children
task depends:"$TASK_UUID" export | jq '.[]'
Skip empty-result failures
Always export | jq
Quick Reference
Step
Command
Load
task TASKID export | jq (capture uuid)
Annotate
task UUID annotate "msg"
Close
task UUID done
Check unblocked siblings
task depends:UUID export | jq '.[]'
GitHub close
gh issue close N --comment "msg"
PR comment
gh pr comment N --body "msg"
Bulk-close patterns
Closing many tasks in one pass has two silent foot-guns. The obvious shape — for id in 1 2 3; do task $id done; done — reports success while doing the wrong thing.
Foot-gun 1: numeric IDs renumber after every task done
Numeric IDs are a display index over pending tasks. The moment one closes, every higher ID shifts down by one. A loop over numeric IDs closes the original first task, then keeps targeting wrong tasks as IDs slide underneath the iterator. No error surfaces.
Fix: use UUIDs (immutable). The same applies to any iterated state-changing op — annotate, modify, delete.
# WRONG — IDs shift mid-loopforidin 35 36 37 38 39; do task "$id"done; done# Correct — capture immutable UUIDs first
UUIDS=$(task status:pending project:myrepo export | jq -r '.[].uuid')
for u in$UUIDS; do task "$u"done </dev/null; done# stdin redirect — see below
Foot-gun 2: task done consumes loop stdin
task done reads from stdin (for confirmation prompts). In a shell for loop, the loop's input is also stdin — so task done eats subsequent iterations and the loop exits early, usually after one or two passes, with no error.
Symptom: a loop over 15 UUIDs reports "processed 15" but only 1 task closed.
# Fix A — redirect stdin per inner commandfor u in$UUIDS; do
task "$u" rc.confirmation=no done </dev/null
done# Fix B — xargs (preferred; each invocation runs in its own subshell with no# stdin link to the source loop)echo"$UUIDS" | xargs -I {} sh -c 'task rc.confirmation=no {} done'
Always pass rc.confirmation=no for batch closes
Without it, taskwarrior may prompt "this task is blocked by N other tasks, complete anyway? (yes/no)" per task, hanging the loop. rc.confirmation=no makes batch closes deterministic.
Annotate before done, not after
Once completed, a task's id becomes 0 and task <id> no longer addresses it — only the UUID does. Annotate first to keep the UUID-or-ID workflow uniform.
Finding tasks for the bulk close
Use export | jq (never list — it exits 1 on empty filters and cancels parallel siblings; see .claude/rules/parallel-safe-queries.md):
task status:pending project:myrepo +pr_ready export | jq -r '.[].uuid'# Tasks with no project (CLI filter quirk — empty value as first filter# errors; sidestep through jq)
task status:pending export | jq -r '.[] | select(.project == null) | .uuid'# Substring matches on description (for markers that aren't real tags)
task status:pending export \
| jq -r '.[] | select(.description | test("\\[triage\\]")) | .uuid'
