/loop-task

name: loop-task
description: Runs non-trivial implementation/fix work that needs a verifiable output through a rubric + independent verifier grading cycle. Invoke before starting code/docs work. Not for simple questions, exploration, or one-line fixes.
argument-hint: "[optional #id / issue-key] [task description]"

Loop-Task — rubric-based self-correction cycle

If `.loop/config.json` does not exist, stop this skill and point the user to `/loopcraft:loop-init` first.

If the argument starts with a `#N` (GitHub) or an issue key (e.g. `ABC-123`), treat that token as an optional **write-back target** and the rest as the task description (used in §6). No such token → no write-back — a normal loop-task.

0. Pick the rubric

• From the `rubrics` array in `config.json`, use the rubric whose glob matches the files you're working on

(`.loop/rubrics/<name>.md`). If nothing matches, or it's ambiguous, ask the user which rubric to use.

• Read the rubric file and note its criteria count and gates (frontmatter `gates`, else the config-wide gates).

1. Consult

Find and read notes for the categories/tags relevant to the task from `.loop/memory/INDEX.md`. Treat `verified: false` notes as hypotheses only.

2. Create the marker (Stop-gate integration)

mkdir -p .loop/state && printf '%s | rubric=%s | started=%s\n' "<task one-liner>" "<rubric name>" "$(date +%F)" > .loop/state/current-task

While this marker exists, ending the session is blocked by the stop-gate — you can't walk away before grading finishes.

3. Do the work (maker)

Record a baseline before starting: `BASE=$(git rev-parse HEAD)`. Do the work, but don't commit yet (grade against the working tree).

4. Verifier grading

Call `subagent_type: "loopcraft:verifier"` with the Agent tool (a fresh subagent, **no fork**). Put **only the following** in the delegation prompt — do not mix in your reasoning, your conversation, or excuses:

1. The full rubric (the file content verbatim) 2. The output: the full `git diff "$BASE"` (if it's large, a list of changed file paths + their absolute paths) + paths of any new files 3. The gate output, if you already ran it

Read the `Result:` line from the verifier's Verdict.

5. Handle the ruling

• **FAIL** → fix using only the "FAIL summary" as grounds, then go back to 4. The retry cap is the rubric frontmatter `max_retries` if present, else the config `maxRetries` (default 3).

On exceeding it, **escalate**: show the user the last Verdict and ask for a decision. If you stop, record the escalation reason in .loop/memory/STATE.md and delete the marker.

• **If the Verdict's "Unscorable criteria" line has entries** → it means the rubric needs revision, separate from the task itself.

Include it in the completion report and propose editing `.loop/rubrics/<name>.md` (spec §6).

• **PASS** → go to 6.

6. Gate → commit → cleanup

1. Run the gates (rubric frontmatter `gates` first, else the config-wide gates). Green is required. If a gate fails, treat it exactly like a verifier FAIL — fix and return to §4 for re-grading (sharing the retry count), and keep the marker. 2. Commit — add a trailer at the end of the message:

   Loop-Verified: <pass>/<total>

3. **Write-back (opt-in)** — if the argument carried a `#id`/issue-key **and** `config.backlog.report` is configured, run it once (best-effort) to comment the verdict on that item: `LOOP_ITEM_ID=<id> LOOP_EVENT=verified LOOP_WRITEBACK=comment LOOP_VERDICT=<pass>/<total> LOOP_COMMIT=<sha> LOOP_BRANCH=<branch> <config.backlog.report>`. Don't pass `title`/`body`. If it fails, don't fail the task — note "report failed" in the report. Comment only (draft-pr is loop-run's). No id, or no `report` configured → skip. 4. Save the full verdict to `.loop/journal/$(date +%F)-<task-slug>.md` (gitignored). 5. Delete the marker: `rm -f .loop/state/current-task` 6. If there were failures/findings in this cycle, follow up with `/loopcraft:distill`.

Include in the completion report

Verdict summary (N/M), retry count, commit hash, write-back result (if an id was given), unscorable criteria (if any), distilled notes (if any).