M plan implement

m_plan_implement

Follow-up to m_plan. Reads existing artifacts, drives them to completion. No re-planning. Default mode: direct walk. /goal mode: opt-in.

Inputs

• $ARGUMENTS: optional task slug. If absent, auto-detect.

Phase A — Load the plan

1. List .m_plan/*/ directories.

   • None → No plans found. Run /m_plan first. Stop.
   • One → use it.
   • Many → list with last-modified, ask via AskUserQuestion (max 4; on overflow, ask user to type the slug; if that tool is unavailable in this harness, ask in chat as a numbered list and wait).

2. Read present files. A plan is valid with at minimum 01_architecture.md, 04_implementation_requirements.md, 05_step_plan.md, 09_verification.md. Other files are optional — treat missing as N/A. Read whichever subset exists in parallel.

3. Extract:

   • Step IDs S01..S<N> + per-step check commands from 05.
   • Verification rows V-* + current state from 09.
   • Build / test commands from 07_test_plan.md if present, else from CLAUDE.md.
   • Smoke + e2e IDs from 08_e2e_plan.md if present.
   • Deploy steps, real target hosts/services, release commands, rollback commands, and post-deploy smoke probes from 06_deploy_plan.md if present.
   • If 06 is missing or does not name a real target, inspect repo-local ops docs (AGENTS.md, CLAUDE.md, README, docs/, deploy scripts, compose/k8s/systemd files) for an unambiguous deploy path. If still ambiguous, mark deploy as blocked. Do not silently replace deploy with localhost checks.

4. Detect resume state. Summarize: Already done: <N>. Open: <M>. Blocked: <K>. Ask Resume / Restart / Stop.

5. Readiness check. If 03_infra_requirements.md exists, list its secrets/access requirements and ask Ready? (yes / no — name the blocker). Any open [!] in 09 — including the V-READY-* deploy/e2e go/no-go rows seeded by /m_plan Phase 0 — must be resolved (re-probe and flip to [x]) or explicitly converted to [~] before proceeding. Do not start the walk with an unresolved deploy/e2e readiness blocker.

6. Deployment reality gate + browser preflight. Before editing code, decide whether this plan has an actual runtime surface and a real deploy target, and whether you can run any browser/UI checks 08 requires. Deploy reality invariant: local build / unit / dev-server checks on this workstation are preflight only — they never satisfy V-DEPLOY-*, V-SMOKE-*, or V-E2E-* rows unless 06 declares this workstation as the target. If this plan deploys anything or has browser checks, read templates/deploy_rules.md now for the full gate + preflight checklist before proceeding.

Phase B — Execute (default: direct walk)

Walk 05_step_plan.md step by step:

• Respect dependency order. After each step's per-step check passes, flip its V-STEP-Sxx to [x] via Edit.
• On failing check: diagnose, retry once. Second failure → [!] blocker, stop, jump to Phase C.
• After all steps: local preflight build/test (delegate noisy runs to the m_code-test-runner agent, keep only the summary) → (if 08 exists) e2e against the intended target when possible — run browser/UI checks through the available browser MCP (Playwright/Chrome, or macOS computer-use) only after the Phase A §7 access preflight passes → real deploy → post-deploy smoke from outside or on the target host.
• Update other V-* rows as their evidence appears in the transcript.

Deviation logging (mandatory). Any time the run drifts from 01–08, append a one-line entry under ## Deviations in 09_verification.md immediately — do not wait for Phase C. A deviation is any of:

• A file touched that is not in 04's "Files to add / modify / delete".
• A different approach taken mid-step (e.g., used a different library than 01 named).
• A smoke / e2e check skipped or replaced.
• A per-step check replaced with a different command than 05/07 declares.
• A local check used where the plan expected deploy, remote smoke, or target-environment verification.

Format: - Sxx / V-XXX-YY — plan said: <one quote>. Did: <…>. Why: <…>.

Deploy walks 06 like steps walk 05, but only against the real target. When every non-deploy V-* is [x]/[~], run 06's "Deploy order" entries in declared order, each followed by its matching smoke check. Stop on first failure; flip the matching V-DEPLOY-XX to [!]. Read templates/deploy_rules.md for the deploy walk + real-environment deploy rules (what counts as observing the running artifact, SSH/CI handling, destructive-action policy, post-deploy smoke) — they all enforce the deploy reality invariant.

Final V- sweep (before declaring DONE).* After everything above, re-read 09_verification.md from disk and, for every still-[ ] row, attempt a deterministic check from current repo / build state: greps, file existence, signature matches, last build/test output. Flip what passes to [x]. Anything that cannot be deterministically verified stays [ ] and becomes a "Not done" item in Phase C.

Opt-in /goal mode — if user said use /goal or the run will span sessions, emit:

/goal Complete every checkbox in .m_plan/<slug>/09_verification.md. For each, run the cited command, paste output, flip [ ]→[x]. Stop after <N_steps * 4 + 20> turns and explain residual blockers.

Rules during execution

• Never edit 01–08 to change scope. Contract from /m_plan.
• Never delete a V-* row. Only flip state.
• [~] skip needs a one-line reason inline. No confirmation required.
• [!] blocker needs description + entry under "Blockers".
• Commit policy follows 06 if present; otherwise one commit per step.
• Deploy from 06 or the discovered real deploy path runs automatically after every non-deploy V-* is [x]/[~]. If no real deploy path can be identified for a deployable runtime, stop with V-DEPLOY-* blocked instead of reporting DONE.

Phase C — Three-section report

Re-read 09_verification.md from disk, then emit the three-section report using the skeleton in templates/report_format.md (all headers mandatory, kept even if empty). If not DONE, name the smallest action that unblocks each open / blocked item.

Rules

• Never regenerates the plan. If 01–08 are wrong → stop, tell user to re-run /m_plan or edit manually.
• Never writes secrets / prod hostnames into artifacts or commits.
• All three report sections are mandatory; empty sections kept (explicit empty = info).
• Trust artifacts: blockers were resolved in /m_plan Phase 0. New blockers mid-flight escalate, never silently bypass.

Files in this skill

• SKILL.md — this file (core algorithm)
• templates/deploy_rules.md — deploy reality gate + browser preflight + real-environment deploy rules (read on deploy)
• templates/report_format.md — Phase C report skeleton (read at report time)
• evals/evals.json — test cases