Implement Full Spec

Implement Full Spec

This skill turns a parent ticket (audit, RFC, multi-finding report, etc.) with N actionable subtasks into a clean stack of N pull requests — and then keeps driving that stack to merge-ready by addressing every comment and review on every PR until reviewers are satisfied. It is the durable version of the workflow proven out on a real multi-finding security audit (24 P1/P2/P3 findings → 24 stacked PRs → multi-round bot review-response cycle).

Dependencies — check these first

This skill orchestrates two skills that ship as separate plugins. Before Phase B, confirm both are installed and invocable:

• dev-team (plugin: dev-team) — the per-subtask Dev → QA → Reviewer → commit loop that Phase B runs once per subtask.
• pr-autopilot (plugin: pr-autopilot) — owns the single-PR review mechanics that Phase C delegates to.

If either /dev-team or /pr-autopilot is unavailable, stop and tell the operator to install the missing plugin(s) from the same marketplace — e.g. /plugin install dev-team and /plugin install pr-autopilot — then re-run. Do not reimplement those loops inline.

When to use this skill

Triggers when the operator points at a parent containing multiple actionable subtasks AND wants each subtask to land as its own PR. Common phrasings:

• "Work through all the P1/P2/P3 subtasks of this ClickUp/Linear/Jira ticket and create a PR for each."
• "Fix all the findings in this audit. One PR per finding, stacked."
• "Use agent teams to fix all the subtasks of [parent]."
• "Address all the comments and bot reviews on PRs #N..#M."
• "Go through these stacked PRs and respond to every review."
• "Drive these to merge-ready. Re-request review when each one is updated."

It is the right skill when all three are true:

1. The work is naturally a sequence of small, independently-reviewable units (≥3, typically 5–30).
2. Subtasks share enough surface that stacked branches are cheaper than 30 branches off main.
3. The user wants the orchestration to outlast a single attention span — fold in clarifications, hit review feedback, request re-review.

It is the wrong skill when:

• The parent has one big issue, not many subtasks → just open one PR.
• The subtasks are wholly independent (different services, different repos) → open N independent PRs off main.
• The user wants a synchronous pair-programming flow → just work alongside them, no orchestration framework needed.

Three phases

This skill has three distinct phases. They share state (the worktree set, the orchestrator's conversation memory, the PRs on the forge) but are independently invoked.

| Phase | What happens | Trigger | |---|---|---| | A. Plan | Read parent, filter actionable subtasks, design the stack order, pin execution parameters. | First user message with a parent ticket URL. | | B. Execute | For each subtask in order: worktree → run the dev-team loop (Dev → QA → Reviewer → commit) → push → gh pr create --base → mark ticket complete. | After Phase A is approved. | | C. Respond to feedback | Sweep every PR for inline-thread + top-level-review + general-comment feedback. Classify, fix, commit, push, reply with Addressed in , cascade-rebase downstream PRs, then ping @ re-review. Multi-round until reviewers stop flagging things. | After Phase B is open on the forge, OR triggered fresh ("address all the comments on these PRs"). |

Phase A and B are well-covered by references/phase-a-plan.md and references/phase-b-execute-subtask.md. Phase C is in references/phase-c-review-response.md. Common gotchas across phases are in references/gotchas.md. The Phase C agent skeleton and the Phase A kickoff prompt live in references/prompt-templates.md.

The per-subtask Dev → QA → Reviewer loop is its own skill. Phase B does not implement the loop inline — it calls the dev-team skill once per subtask to drive that subtask's spec to a committed, reviewed result, then layers on the stacking, PR, and ticket machinery that's specific to this skill. That keeps the loop reusable: you can fire it off directly on a single change without the full-spec scaffolding. Everything about how that loop runs lives in dev-team — this skill delegates to it and never reimplements those mechanics (see "The per-subtask loop, cycle cap, and role collapsing" below).

For full execution, read the three phase files in order. For a partial invocation (e.g., the user asks only "respond to comments on the PRs we already opened"), read the relevant phase file directly.

Top-level workflow

operator: "Fix all P1/P2/P3 subtasks of parent ticket X and stack the PRs"
   │
   ▼
PHASE A — Plan
   • Read parent ticket. Filter actionable subtasks by priority/label/criteria.
   • **INTERVIEW the operator on PR strategy — three modes:**
       - (A) Single PR: one commit per subtask, all on one branch, one PR.
       - (B) Stacked PRs: N branches stacked, N PRs (the default for
             dependency-ordered work).
       - (C) Independent PRs with selective stacking: one PR per subtask,
             off main when independent, stacked only on real dependency.
   • Decide subtask order (priority + dependency-ordered within tier).
   • Pin remaining parameters (cycle cap, concurrency, worktrees, etc.).
   • Surface the plan; get user approval before touching code.
   │
   ▼
PHASE B — Execute (per subtask, serially in A/B; serial-or-parallel in C)
   For each subtask in order:
   • Set up worktree per the chosen mode (one shared for A; per-subtask
     for B/C, branched off the planned parent).
   • Run the `dev-team` skill on this subtask's spec:
       Dev → QA → Reviewer/code-review → commit, bounded by a cycle cap.
       It returns an APPROVED & COMMITTED SHA (or escalates). The loop
       commits but does NOT push — that's this phase's job.
   • Push the returned commit.
   • Open PR — varies by mode:
       - Mode A: defer; open one PR at the end of the run.
       - Mode B: `gh pr create --base ` after each subtask.
       - Mode C: `gh pr create --base main` for independent;
                 `--base ` for dependent.
   • Mark ticket complete with SHA + branch + PR URL (or "PR pending" in A).
   • Advance.
   │
   ▼
PHASE C — Respond to feedback (after PRs are open OR on demand)
   • For every PR in order:
       - Fetch ALL THREE comment sources (this is the trap):
           * reviewThreads (inline)
           * pulls/{n}/reviews (top-level bodies)
           * issues/{n}/comments (general)
       - Verify counts so silent-empty fetches are caught.
       - Classify each per the pr-autopilot / pr-review-mechanics decision matrix.
       - Fix actionable items; commit (new commit by default).
       - Push.
       - Reply on each thread/comment with: Addressed in []().
       - Resolve inline threads via GraphQL resolveReviewThread mutation.
       - Cascade rebase — varies by mode:
           * Mode A: skip (only one PR).
           * Mode B: rebase ALL downstream PRs onto the new tip; force-push.
           * Mode C: rebase only PRs in this PR's dependency chain.
       - Comment "@ re-review" so they re-evaluate.
   • Multi-round: bots often respond to re-review with NEW items. Repeat.

The per-subtask loop, cycle cap, and role collapsing

The per-subtask loop — its cycle cap, role-collapsing decision table, agent-team execution model, and commit hard rules — lives in the dev-team skill. Phase B calls it once per subtask; read its SKILL.md for the mechanics and the operational cap value. The one knob this skill surfaces is the cycle cap, which Phase A pins as an execution parameter (its default mirrors dev-team's); otherwise this skill delegates the loop and doesn't reimplement it.

Two stack-specific consequences worth stating here:

• The cycle cap blocks the stack. When a subtask exhausts the cycle cap without an APPROVED commit, downstream subtasks can't proceed — they need this branch as their base. Escalate to the user (the loop hands you the diff + failing output + last feedback); do not silently retry past the cap.
• Each subtask gets its own dev team; never share one across subtasks. Keep your parent-level progress tracking out of any shared TaskList the teammates can see (it leaks and they auto-claim the wrong work) — track it in conversation memory. See the dev-team skill's "Team guardrails" for the rest.

Hard rules (do not relax without explicit user override)

The loop's own commit-and-workspace invariants — no --no-verify, no git add -A, no amend by default, don't touch the operator's primary checkout, verbatim spec, mental-revert clause, IN/OUT-of-scope partition — are owned by the dev-team skill and inherited by every per-subtask agent. The rules below are the ones specific to driving a stack of PRs, which the loop knows nothing about:

• Amend is the stack's call, not the loop's. The loop never amends. The one place amend is correct: a bot explicitly flags the original commit's subject/body (e.g., "subject 75 chars > 72") — fix that with git rebase -i HEAD~N + reword, then force-push and cascade the stack.
• Stack-aware rebases after every push: git rebase --onto . See references/gotchas.md for why the naive `` form breaks once upstream has been force-pushed.
• Three comment sources in Phase C. Inline review threads + top-level review bodies + issue-level comments. Missing the top-level-review fetch is a documented failure mode; references/gotchas.md has the GraphQL/REST snippets that get all three.
• Reply format: Addressed in [\\](). Inline threads get a resolveReviewThread GraphQL mutation; top-level reviews and general comments don't have a resolve op, just the reply.
• Re-review request: after pushing changes that address review feedback, post @ re-review so the bot reruns. Do this for every PR in every round.

Adapt the plan in flight

The Phase A plan is a starting hypothesis, not a contract. Three things will almost certainly change in flight:

1. Role collapsing. A run might start full-fat 3-agent and discover ticket #5 is trivial enough to single-shot. Switch when the evidence supports it — the dev-team skill makes this decision per subtask.
2. Cycle cap behavior. The loop escalates when its cycle cap is exhausted. When the user has explicitly said "don't stop until done," push the cap with judgment — but report any cap hit even when continuing past it.
3. Conflict resolution patterns during cascade-rebase. Architectural changes (auth token format, API shapes) tend to mutate across the stack. The same conflict shape recurs; resolve it consistently. references/gotchas.md has a worked example.

If a documented step is producing pain instead of value, change it and write the change down. The plan file at ~/.claude/plans/.md is the persistence path across context resets — update it; don't pretend the old version still applies.

Pointers

• dev-team — the per-subtask Dev → QA → Reviewer/code-review → commit loop, cycle cap, role collapsing, commit hard rules, and the Dev/QA/Reviewer prompt skeletons. Phase B calls this once per subtask.
• pr-autopilot skill (separate plugin) — owns the single-PR review mechanics (three-source fetch, classify, reply/resolve, per-bot re-review triggers) in its pr-review-mechanics.md reference. Phase C is the stacked-PR superset: it delegates each individual PR's comment mechanics to /pr-autopilot --single-pass, then layers the cascade-rebase + per-round re-review in references/phase-c-review-response.md on top. Requires the pr-autopilot plugin (see "Dependencies — check these first" above).
• references/phase-a-plan.md — survey the parent, filter, design the stack, present the plan
• references/phase-b-execute-subtask.md — the per-subtask flow: worktree → loop → push → PR → ticket, in mechanical detail
• references/phase-c-review-response.md — three-source fetch + classify + fix + reply + cascade-rebase + re-review ping
• references/prompt-templates.md — Phase C agent skeleton + Phase A kickoff prompt; copy-paste
• references/gotchas.md — cross-cutting traps: jq+control-chars, three-source fetch, rebase --onto, force-push-with-lease, amend-vs-new-commit, evolving-API conflicts

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: donnfelker
• Source: donnfelker/loop-skills
• License: MIT

Install and usage instructions live in the source repository linked above.