/taskwarrior:task-bulk-ops

Operate on many taskwarrior tasks in one pass — queue cleanup, a PR-merge sweep, a project wrap-up — without hitting the four foot-guns that make the obvious for id in 1 2 3; do task $id done; done silently do the wrong thing. Every foot-gun fails invisibly: the loop reports success, the count looks right, and the wrong tasks (or no tasks) were touched.

When to Use This Skill

Context

• Task CLI available: !task --version
• Pending task count: !task status:pending count

Parameters

Parse $ARGUMENTS as an optional taskwarrior filter that narrows the set to operate on (e.g. project:foo, +bulk status:pending, +gh). When empty, ask the user which set to target before mutating anything — never bulk-mutate an unfiltered queue by default.

Execution

Execute bulk operations under these four rules. They are not optional style preferences — skipping any one silently corrupts the batch.

Rule 1 — Address tasks by UUID, never numeric ID

Taskwarrior numeric IDs are a display index over pending tasks: the moment any task completes, every higher ID shifts down by one. A numeric ID is stale across any gap between reading and acting — a loop iteration, a later turn, a new session, or an unrelated task done in another session.

Resolve the immutable UUID at read time, then use only UUIDs for every later op:

# Loop over a set — capture UUIDs up front (they never shift)
UUIDS=$(task status:pending project:foo export | jq -r '.[] | .uuid')
for u in $UUIDS; do
  task "$u" done </dev/null            # (Rule 2: </dev/null)
done

# Acting later? Resolve the UUID the moment you know you'll act, not the number.
PROMOTE=$(task _get 169.uuid)          # 169 is valid RIGHT NOW, at read time
# ...any number of turns / unrelated closes later...
task "$PROMOTE" done </dev/null        # still the right task

If you only have a numeric ID and a gap has passed, re-derive it (re-run the export/list, re-match on description) before acting — do not trust the cached number. The same applies to annotate / modify / delete: any state-changing op iterated over a set references UUIDs.

The +LATEST _get trap. _get is a DOM accessor taking an <id>.<attribute> reference — given a filter (a tag like +LATEST) it silently returns empty (exit 0) and captures nothing. To grab the UUID of the task you just added, use the +LATEST-aware accessor or export | jq:

task +LATEST uuids                          # → UUID of the just-added task
task +LATEST export | jq -r '.[0].uuid'     # always exit-0, parallel-safe

Do not reach for the _get DOM accessor with a tag filter to grab that UUID — task +LATEST _get uuid silently returns empty (exit 0). _get needs an <id>.<attribute> reference: task _get <id>.uuid works for a known numeric id, but the just-added task's UUID comes from task +LATEST uuids.

Rule 2 — Redirect stdin and disable confirmation on every mutation

task done (and other mutating subcommands) read from stdin. Inside a shell for loop the loop's input is stdin, so task done eats the remaining iterations and the loop exits early after one or two — reporting "processed 15" while closing 1. Two safe forms:

# Form A — redirect stdin per inner command
for u in $UUIDS; do
  task "$u" annotate "swept in cleanup" </dev/null
  task "$u" rc.confirmation=no done </dev/null
done

# Form B — xargs runs each command in its own subshell, no stdin link (preferred one-liner)
echo "$UUIDS" | xargs -I {} sh -c 'task rc.confirmation=no {} done'

Always pass rc.confirmation=no to batch task done — without it taskwarrior may prompt "blocked by N other tasks, complete anyway?" and hang.

task config has the same confirmation hazard — and it's worse. task config <name> <value> prompts to confirm by default; non-interactively (a script, this Bash tool, a hook, chezmoi) the un-answered prompt makes it exit 0 without writing the value. The failure is invisible. The sharpest bite is declaring UDAs:

# WRONG — exits 0 but the UDA is NOT written; next `task add foo ghid:1417`
# appends the literal "ghid:1417" to the description instead of setting the UDA
task config uda.ghid.type numeric

# RIGHT — rc.confirmation=no makes the write happen; </dev/null guards stdin
task rc.confirmation=no config uda.ghid.type numeric </dev/null
task rc.confirmation=no config uda.ghid.label "GH Issue" </dev/null

Verify the write landed rather than trusting the exit code:

task _udas | grep -qx ghid && echo "declared" || echo "MISSING — config did not persist"

Rule 3 — Find tasks with export | jq, and mind the filter caveats

Prefer task <filter> export | jq over task <filter> list: export emits [] and exits 0 on an empty result (parallel-safe), while list exits 1 — which silently cancels sibling calls in a parallel Bash batch.

• Empty project: as the first filter errors. task project: status:pending list fails with Unable to find report. Put the value-less filter later, or sidestep the CLI with jq:

  task status:pending export | jq -r '.[] | select(.project == null) | .uuid'
  

• Escape + in test() when matching a tag-style marker that lives in the description text (not a real tag):

  task status:pending export | jq -r '.[] | select(.description | test("\\+upstream_issue")) | .uuid'
  

• Real taskwarrior tags (set via +tag at creation, listed in .tags) filter through the CLI: task +upstream_issue status:pending export. Use underscores or camelCase in tag names, never hyphens — taskwarrior parses a hyphen mid-tag as an exclude filter, so the tag silently never lands.

Rule 4 — Annotate before done, not after

Once a task is completed its id becomes 0 and task <id> no longer addresses it (only the UUID does). Add annotations before closing so the annotate/done workflow stays uniform and the annotation lands on the still-pending task that task <id> info shows.

for u in $UUIDS; do
  task "$u" annotate "closed in PR-merge sweep" </dev/null   # annotate FIRST
  task "$u" rc.confirmation=no done </dev/null                # then close
done

Report

After the pass, re-query and report what actually changed — never trust the loop count. For a close sweep, confirm the pending count dropped by the expected amount and no stray tasks were touched:

task status:pending count

Agentic Optimizations

Quick Reference

Related

• /taskwarrior:task-done — close one task with a landing commit (single-task path)
• /taskwarrior:task-reconcile — close tasks whose linked GitHub issue/PR closed (bulk, but GitHub-driven)
• /taskwarrior:task-status — read-only consolidated queue report
• /taskwarrior:task-add — file a single linked task (the +LATEST uuids capture pattern originates here)
• .claude/rules/parallel-safe-queries.md — the export | jq (exit-0-on-empty) idiom