M Plan Roll

m_plan_roll

One-shot, hands-off pipeline: plan large → implement → deploy → verify, no questions. A thin wrapper over m_plan (Phases 0–1) and m_plan_implement (execution) that swaps the interactive parts for autonomous defaults and maximum coverage.

This trades interactivity for autonomy. It still cannot fake verification and still obeys the hard safety gates below — those are the only things that can halt or block the run.

Inputs

• $ARGUMENTS: the task. If empty, ask once for the task, then stop asking.

How it differs from /m_plan

m_plan asks blockers and right-sizes. m_plan_roll does neither — it assumes the most thorough path and decides everything itself.

Phase 1 — Plan (autonomous, large)

Run m_plan Phases 0–1 with these overrides:

• No AskUserQuestion, no blocker batch. For every blocker in m_plan's BLOCKERS.md, pick the safest thorough default yourself. Record each non-obvious decision under a ## Assumptions heading in 01_architecture.md so the run stays auditable.
• Force size = large → write all 9 artifacts (01–09), maximum coverage (08 included).
• Gitignore: append .m_plan/ to .gitignore by default (no prompt); note it in the run log.
• No Approve pause. Do not stop for Approve? (yes / edit / stop) — go straight to Phase 2.
• Still obey m_plan's artifact rules: a per-step check in every 05 step, build/test commands live only in 07, fill every <…> placeholder, no secrets / real tokens / prod hostnames.

Phase 2 — Implement (autonomous)

Immediately invoke m_plan_implement and run its Phase B to completion, with these overrides:

• Auto-answer its prompts: Resume the plan just written; Ready? = yes; no slug-selection question (use the new plan).
• Deploy runs automatically once every non-deploy V-* is [x]/[~] (treat it as if the user typed deploy) — against the real declared target only, per 06 or the discovered deploy path.
• If a non-deploy V-* is itself [!] blocked (e.g., the target or a required check is unreachable), the auto-deploy precondition isn't met — mark the dependent V-DEPLOY-* / V-SMOKE-* [!] too, finish everything else, and report PARTIAL. Don't force a deploy and don't hang waiting.
• Walk steps, flip V-*, log deviations immediately, run the final V-* sweep, then emit the three-section report.

Hard safety gates (still apply — never bypassed)

Autonomy means don't ask for preferences. It does not mean skip safety:

• Never fake verification. 09_verification.md still decides done. Local build/unit/dev-server checks never satisfy V-DEPLOY-* / V-SMOKE-* / V-E2E-* (the deploy reality invariant).
• Can't access a required check (no real deploy path, or the browser MCP / credentials / URL for a check are missing): do not stop the whole run and do not fake it — mark that V-* [!] blocked, log a deviation, and keep going with everything else. Surface it in the final report.
• Destructive actions on production / shared services (restarts, migrations, replacing containers, data deletion, reboots): apply the repo/user destructive-action policy. If that policy requires confirmation, this is the one place the run pauses — stop at that specific action and report it rather than doing something irreversible unprompted.
• No secrets / real tokens / prod hostnames in any artifact or commit.

Phase 3 — Report

Emit m_plan_implement's three-section report (✅ Done / 🔀 Worked around or changed / ❌ Not done), then append:

• Assumptions made: every autonomous decision from Phase 1 (copy from 01's ## Assumptions).
• Halted on: any destructive-action gate that paused the run (or "none").
• Final: DONE | PARTIAL | BLOCKED. If any V-* is [ ]/[!], it is not DONE — never close by lying.

Files in this skill

• SKILL.md — this file (orchestrator; execution semantics live in m_plan / m_plan_implement)
• evals/evals.json — test cases