implement

name	implement
description	Implement ONE approved Synto implementation plan end-to-end by driving the implement-plan workflow. Takes the plan's path (or slug) under docs/superpowers/plans/, runs preflight, hands it to the batched self-rotating engine (per-unit green-gate + bookmark-B integration + deep end-review), and reports what landed. Use when a written plan is ready to build.
user_invocable	true

implement

The Synto-specific entry point for implementing one approved plan. It resolves the plan path, then hands it to the implement-plan workflow (.claude/workflows/implement-plan.js) — the agnostic batched engine that owns all the orchestration (the self-rotating green-unit loop, Plan-Tasks commit-trailer progress, the completeness gate, per-unit integration onto bookmark B, and the deep end-review).

Read first (on demand): CLAUDE.md (jj, the green gate, no AI footer, experimental branches) and — only if you need the severity bar — .claude/playbook/project-phase.md. You do not pre-read the plan; the workflow's setup step parses it.

What to pass

The user gives you a plan reference. Resolve it to the exact top-level path docs/superpowers/plans/<file>.md (top level only — never drafts/, never completed/, never a subdirectory). If they gave a slug or bare filename, find the matching top-level file. If no such top-level plan exists, stop and say so (an un-promoted draft is not ready — it must be moved to the top level first).

Run it

Call the Workflow tool with the resolved path:

Workflow({ name: 'implement-plan', args: { plan: 'docs/superpowers/plans/<file>.md' } })

• Mode. Default is integrate (per-unit CI that advances bookmark B). Pass args.mode = 'dry-run' to implement + gate per unit without advancing the bookmark or pushing — use this for a rehearsal.
• Push gating is an environment knob, not an arg. In the default LOCAL mode nothing is pushed; the work lands on B locally and the operator's working copy follows it via jj auto-rebase. Only SYNTO_FLOW_INTEGRATE=push in the environment makes the engine also jj git push. Per CLAUDE.md, experimental work stays on an experimental/* line — do not set push mode unless the operator explicitly asked for it.
• Plan-review checklist (optional). If a plan-review pass produced critical/high concerns, pass them deduped as args.planReviewConcerns: ["[high] …", …] — the deep end-review uses them as a focusing checklist. Omit for a standalone plan.

The engine is plan-file-driven: it reads no issue and no board. It runs preflight (base-branch.sh resolves bookmark B; probe.sh checks the jj working copy), creates an isolated jj workspace rooted at B, then a single batched self-rotating implementer burns small green units (≈one plan task each): a fast commit-stage build → jj commit carrying Plan:/Plan-Tasks: trailers (these are machine-readable progress markers, not an AI footer — keep the message otherwise footer-clean) → integrate.sh, which rebases the stack onto the latest B, runs the full green-gate (dotnet build · dotnet test · dotnet format whitespace --verify-no-changes) on the rebased tree, and advances B. It self-rotates near its quality zone; a fresh generation resumes mechanically from the trailers until every task has landed. Then one deep end-review of the cumulative diff fix-forwards the critical/high/medium defects this change introduced, and the plan file is archived to completed/.

Present the result

The workflow returns a single planResult. Lead with the outcome:

• merged === true && reason === 'merged + archived' — success. Report the landed commit ids (planResult.commits) and pushedTasks. Surface planResult.reviewWarnings (unresolved introduced mediums the review proceeded past) and planResult.reviewDeferred (pre-existing / pre-release-deferred findings) as plain informational notes if non-empty — never auto-file them.
• Anything else — report reason + problems and what to do next. The engine resumes mechanically from the trailers, so for most parks a bare retry is safe:
  • Resumable park — no progress (stuck) / generation cap reached: some units may already be on B; a fresh run reads the trailers and continues the un-landed tasks (it will not redo landed work). Skim why it stuck before retrying.
  • Escalation — reason starts with escalated:: the engine judged the plan wrong/infeasible, or integrate.sh hit an unresolvable rebase conflict or a push-rejected remote. A human fixes the plan / resolves the conflict / reconciles the remote first; landed units stay on B.
  • Transient/environment park — reason mentions transient/environment (a NuGet restore blip or a build-host file lock under concurrent workspaces): not a code defect; retry once the environment settles.
  • { aborted: true } — preflight failed (no bookmark B, or the jj working copy has conflicts). Fix the precondition (probe.sh's reason says which) and re-run.

Notes

• All code lands inside the workflow's isolated jj workspace and advances bookmark B; this skill makes no commit in the operator's default working copy and must leave it clean.
• One plan per run, by exact path — no other plan is ever swept.