richkuo/rk-skills/fableplan-loop

fableplan-loop

Chain fableplan → work-on-issue-loop into one autonomous run: Fable 5.1 plans the implementation for a GitHub issue (plan posted to the issue), then work-on-issue-loop implements that plan in an isolated worktree, opens a PR that closes the issue, and drives the PR through @claude review to convergence.

It has no validate-issue step, no update-issue edits, and no score gate: fableplan always runs.

Do not skip or reorder the chain. The plan gates implementation. Every step of each skill still runs; only the "wait for the user's reply" moments are replaced by the handoff below.

Input

A GitHub issue is required (work-on-issue-loop targets an issue): a full URL, #<N>, bare <N>, or owner/repo#N. With nothing supplied, use the issue validated or planned in this session; with none, stop and ask. Never pick an issue from a list, and never plan or implement against a paraphrase.

Record the resolved issue as the full reference owner/repo#N per fable-validate-loop step 1 (the repository from the user's reference, else gh repo view --json nameWithOwner -q .nameWithOwner). Every step below passes that full reference.

Optional targetBranch (orchestration form { issue, targetBranch } or a prose "target branch <name>"): passed unchanged to every plan and build step in the chain, so the baseline and the PR base are that branch instead of the repo default. work-on-issue step 1 ("Target") owns its validation.

Steps

0. Pre-plan gate — check the issue is still worth planning

Before dispatching any planning:

• gh issue view <N> -R <owner>/<repo> --json state,title,url,closedByPullRequestsReferences: is the issue still open?
• Check for PRs already addressing it (linked PRs on the issue, or an open PR in <owner>/<repo> whose branch/body references #<N>, via gh pr list -R <owner>/<repo>).

If the issue is closed, or a merged/open PR already addresses it, do not plan — alert the user and ask what to do next (e.g. plan anyway, target the existing PR, or stop). Only proceed on their say-so.

1. Run fableplan — planning phase only

Invoke the fableplan skill for the recorded owner/repo#N (Skill tool, skill: fableplan) and follow fableplan's "Planning-phase-only invocation" section. Instruct fableplan to use the harness suffix fableplan-loop in the posted comment's attribution footer.

Keep the plan's scratchpad file for step 2. On a structurally wrong plan, a fableplan dispatch failure after its internal retry, or a snapshot diff that shows the planning subagent wrote, stop and report per that section; don't hand a broken plan to work-on-issue-loop, and don't implement unplanned.

2. Hand off to work-on-issue-loop

Invoke the work-on-issue-loop skill for the recorded owner/repo#N (Skill tool, skill: work-on-issue-loop). Pass that full reference explicitly, so it cannot resolve a different issue from session context or the current checkout, and instruct it that the implementation must follow the posted plan: point it at the plan's scratchpad file and the posted issue comment (## Implementation plan (<model>)), and tell it deviations follow work-on-issue step 2's plan-deviation policy — the traced code, a newer comment or body edit from a work-on-issue step 0 trusted author, then correctness and safety — and must each be named in the PR body. Don't narrow that policy here.

It runs its full loop: work-on-issue implements in a fresh worktree and opens the PR (Closes #<N>), the loop triggers the first review, and fix-pr-review cycles run. work-on-issue-loop step 2 owns the stop rules and step 4 owns the terminal-state table; this skill states neither. Relay whatever terminal state it reports in step 3.

3. Report

Relay work-on-issue-loop's final summary (PR URL, number of review cycles, final verdict, which model each fix cycle ran on, any follow-on issues it filed), prefixed with one line covering the head of the chain: plan posted (comment URL). Relay every terminal state of work-on-issue-loop step 4 under its own name: Done, Done, with leftovers, Diverging, Blocked on a test, Fixer stopped, the bot-never-responded escalation, and Nothing to drive. Only the two Done states report a finished review; never imply an approved PR exists when it does not.

Cap the whole report at 55 words and 5 sentences, plain simple English in ASD-STE100; apply the Response Style rules in CLAUDE.md/AGENTS.md, written for a reader with no context on this codebase.

Red Flags — STOP