Implement Plan

Fetches a structured plan from a GitHub issue, steps through each unchecked task in dependency order, commits the result, runs a PR review, and opens a PR linked back to the plan issue. Composes create-commit and review-pr rather than reimplementing their logic.

Instructions

Side effects: implements code changes, creates a git commit, may open a GitHub PR. Task checkboxes on the plan issue are left unchecked throughout, allowing the skill to be re-run or retried.

1. Accept input and fetch the plan issue

Accept the plan as:

• A bare integer (e.g. 126) — a GitHub issue number; infer the repo from git remote get-url origin
• A #N reference (e.g. #126)
• A full GitHub issue URL (e.g. https://github.com/owner/repo/issues/126)

If no plan is specified, ask: "Which plan issue should I implement? Provide an issue number, #N reference, or full URL."

Fetch the issue body using the GitHub tools available in the environment. If the issue cannot be fetched, report the error and stop.

2. Parse and present unchecked tasks

Extract all unchecked checkboxes (- [ ]) from the issue body, preserving their order and any phase groupings. Checked tasks (- [x]) are skipped — they represent already-completed work.

Respect dependency annotations (e.g. _(depends on: Task A)_) to surface tasks in an order that satisfies all declared dependencies. When tasks are marked _(parallel with: Task X)_, implement them sequentially but note they can be batched in future runs.

Present the task list to the user before starting:

• State the plan issue number and title
• List all unchecked tasks grouped by phase (if phases exist)
• Call out any tasks blocked by incomplete dependencies

3. Implement each task in order

For each unchecked task (in dependency-resolved order):

1. State which task you are beginning (e.g. "Starting: Create .agents/skills/implement-plan/SKILL.md").
2. Implement the task using the tools available (Read, Write, Edit, Bash, Glob, Grep, GitHub tools, etc.).
3. Report completion: "Done: <task description>."
4. If a task cannot be completed (missing dependency, ambiguity, external blocker), stop and ask the user for guidance before continuing.

Defer all lint, test, and typecheck runs until the dedicated final-validation step (step 6). In this step, do not run lint, tests, or typecheck after individual tasks — intermediate states are not the validation target, and per-task runs accumulate redundant work across the loop. All such commands are deferred to step 6, where they execute exactly once against the final post-review state. The only exception is running a command to confirm a task's concrete output (e.g. a generator script the task explicitly requires); even then, do not invoke repo-wide lint or test suites.

Do not check off task checkboxes on the plan issue — leave them unchecked so the skill can be retried or re-run without losing progress markers.

After all tasks are implemented, write a brief summary of what was done, then immediately proceed to step 4 — do not stop or wait for user input.

4. Commit via create-commit

Immediately after all tasks are complete, invoke the create-commit skill (/commit) — do not pause, summarize and stop, or ask for confirmation. Proceed directly.

Wait for the commit to succeed, then proceed immediately to step 5. If create-commit reports nothing to commit, note this and proceed immediately to step 5 — do not stop.

5. Review via review-pr

Immediately after the commit in step 4 completes, invoke the review-pr skill on the current branch against the base branch (default: main) — do not pause or ask for confirmation:

/review-pr <current-branch>

review-pr will automatically fix all actionable findings and commit them. Do not run lint, tests, or typecheck during review-pr fix iteration — lint and tests are deferred to step 6, where they run exactly once on the final post-review state. Validating against each intermediate fix-and-recommit cycle is redundant, since only the final state ships. When review-pr completes, proceed immediately to step 6 — do not stop or wait for user input.

6. Run lint and tests on final state

Immediately after review-pr completes, run the repository's lint, typecheck, and test commands exactly once against the final post-review state — do not pause or ask for confirmation. This is the single validation gate; per-task and per-review runs were deferred here specifically to avoid redundant validation of intermediate states.

Command discovery (in priority order — use the first source that defines each command):

1. AGENTS.md / CLAUDE.md — look for an explicit "Testing", "Linting", or "Validation" section naming the commands to run
2. package.json scripts — prefer lint, typecheck, test (in that order); fall back to check or ci if they exist
3. Makefile — prefer make lint, make typecheck, make test
4. Language defaults — ruff check . && pytest (Python), cargo clippy && cargo test (Rust), go vet ./... && go test ./... (Go), etc.

If none of these are present and no convention is documented, skip this step and note it in the PR body.

Execution order: lint → typecheck → tests. Stop at the first failure; a failing lint run typically masks real type or test errors.

Fix-and-recommit loop (capped at 2 fix attempts after the initial failure):

1. If a command fails, inspect the output and fix the root cause — do not suppress warnings or disable rules to make them pass.
2. Invoke create-commit to commit the fix.
3. Re-run the failed command (and any later commands in the order above).
4. Repeat up to 2 fix attempts before escalating.

Escalation: if failures persist after 2 fix attempts, or if a failure is clearly outside the scope of this plan (e.g. a pre-existing broken test unrelated to the changed files), stop and report to the user:

• which command failed
• the relevant error excerpt
• what was tried
• whether the agent believes the failure is in-scope or pre-existing

When all commands pass (or the step is skipped per the rules above), proceed immediately to step 7 — do not stop or wait for user input.

7. Open a PR linked to the plan issue

Immediately after step 6 completes, derive the PR title from the plan issue title and open the PR — do not pause or ask for confirmation.

Keep the title concise (under 70 characters) and imperative (e.g. "Add implement-plan skill to orchestrate plan execution").

Create the PR using the GitHub tools available in the environment, linking back to the plan issue in the body:

## Summary
<bullet points summarizing what was implemented>

## Plan
Closes #<plan-issue-number>

## Test plan
<bulleted checklist of how to verify the changes>

Report the PR URL to the user.

Examples

Invocation with issue number:

/implement-plan 126

Invocation with #N reference:

/implement-plan #126

Invocation with URL:

/implement-plan https://github.com/geemus/hax-yax/issues/126

Invocation (interactive):

/implement-plan

The agent will ask which plan issue to implement.

Sample task presentation:

Plan: #126 — Add implement-plan skill to orchestrate plan execution

Phase 1 — Design (2 tasks):
  - [ ] Define task-parsing contract
  - [ ] Confirm checkpointing behavior

Phase 2 — Implementation (5 tasks):
  - [ ] Create .agents/skills/implement-plan/SKILL.md  _(depends on: Phase 1)_
  - [ ] Write Step 1 instructions  _(depends on: Phase 1)_
  - [ ] Write Step 2 instructions  _(depends on: Step 1)_
  - [ ] Write Step 3 instructions  _(depends on: Step 2)_
  - [ ] Write Step 4 instructions  _(depends on: Step 3)_

Phase 3 — Housekeeping (1 task):
  - [ ] Run upsert-agents-md  _(depends on: Phase 2)_

Ready to implement these 8 tasks. Proceed?

Sample blocking-finding fix:

review-pr found 1 blocking issue:
  issue [blocking]: `SKILL.md` frontmatter is missing the required `name` field

Fixing automatically…
[edits file]
Committing fix via create-commit…
[commits]
Proceeding to step 6 (lint and tests on final state).

Sample lint/test failure recovery:

Step 6 — running lint and tests on final post-review state.

Discovered commands from package.json scripts: `npm run lint`, `npm test`.

Attempt 1:
  $ npm run lint
  ✗ src/plan.ts:42  no-unused-vars: 'tmp' is defined but never used

Fix attempt 1/2: remove unused `tmp` binding in src/plan.ts:42.
Committing fix via create-commit…
[commits]

Attempt 2:
  $ npm run lint
  ✓ passed
  $ npm test
  ✓ 47 passed

All validation commands pass. Proceeding to step 7 (open PR).

Sample full flow (happy path):

Step 3 — implementing 4 tasks, deferring all lint/test runs to step 6.
[task 1 done] [task 2 done] [task 3 done] [task 4 done]

Step 4 — /commit
Step 5 — /review-pr claude/implement-plan-126 (no blocking findings)
Step 6 — lint + typecheck + tests on final state (all pass)
Step 7 — opening PR