agentsclimarketplace

Run plan

Skill hungv47/meta-skills/skills/meta/run-plan

Execute an APPROVED plan.md (the A3 plan-artifact at .forsvn/runs/<slug>/plan.md) step by step — auto-advance the non-publish steps within the governor envelope, narrate each applied play, and STOP hard at every publish / spend / irreversible gate, never publishing. Use to run a plan a human already approved. Not for building the plan (that's /forsvn's multi-step planner), a channel launch (run-launch), or a generic produce-and-measure loop (run-pipeline).From its SKILL.md

Install
npx -y skills add hungv47/meta-skills --skill run-plan

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 14 stars14 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

6.3 KB, ~1.4k tokens by cl100k_base, as published. Nobody here has run it

Run Plan — Approved-Plan Executor

Meta process skill. Walks an approved plan.md, auto-advancing only the non-publish steps within the A6 governor envelope and stopping hard at every publish gate. Orchestrate, don't fuse; never auto-publish, never auto-ratify, never run an unapproved plan.

Core question: "Did every non-publish step advance legibly within the envelope, and did the run STOP — publishing nothing — at the first publish gate or governor limit?"

The executor loop, the stricter-gate rule, the governor config, the worked example: run-plan-spec.md [SPEC]. Plan schema + writer: bin/plan.ts. Governor: bin/lib/governor.ts.

Critical Gates

  1. Approval first — refuse otherwise. Read .forsvn/runs/<slug>/plan.md. If status != approved, STOP and report it (proposed → a human must approve before step 1). Never run, never self-approve.
  2. Publish gate is absolute. A step resolved to publish STOPS the run unconditionally — hand to the human, publish nothing, do not advance. Hard-coded in gate_class; never downgradable by .forsvn/config.json.
  3. Stricter gate wins. Resolve each step's gate from the live capability gate_class (the capability index, built by bin/build-capability-index.ts) AND the plan row's gate. If either says publish, it's a publish stop. A missing/unknown gate is treated as review (never auto).
  4. Governor envelope (A6). Before each step call bun bin/lib/governor.ts check …. On a stop (max-steps / checkpoint / domain-jump / budget) STOP and narrate the reason. The governor only tightens — it cannot relax a gate.
  5. Legible + never auto-ratify. Every advance narrates ▸ Step n/N …; every leaf artifact stays decision_state: pending. No black-box run; no auto-acceptance.

Quality Gate — 5 dimensions

Pre-exit self-check — all 5 must hold (the executor's critic):

  • Approval honored — the run only proceeded because status: approved; an unapproved plan was refused.
  • Publish-stop honored — every publish step stopped the run; nothing was published or advanced past it.
  • Envelope honored — each advanced step passed the governor check; a governor stop was narrated with its reason.
  • Narrated — every advanced step printed ▸ Step n/N /<skill> — <play>; every stop printed ■ ….
  • No auto-ratify — each produced artifact is decision_state: pending; the executor approved nothing.

Before Starting

Read the plan: bun bin/plan.ts show <slug>. Refuse unless status: approved. Resume is native — the Current step pointer + the steps table's status column say exactly what ran and what's next; no chat history needed. Session profile: _shared/execution-policy.md. Mode: _shared/mode-resolver.md--fast skips narration verbosity, never a gate or the publish stop.

Dispatch

Per ready step (all depends_on done), in depends_on order:

  1. Resolve the gate — stricter of the plan row gate and the capability's live gate_class (default review, never auto).
  2. Check the envelopebun bin/lib/governor.ts check --step-index <done-count> --running-domain <chain-domain> --next-domain <step-domain> [--next-skill <id> --cost-so-far <usd>]. Exit 2 → STOP, narrate the reason.
  3. If auto/review and the governor proceeds — dispatch the step's leaf skill (it runs its own agents + critic), leave the artifact decision_state: pending, narrate ▸ Step n/N /<skill> — <applied play / why>, then bun bin/plan.ts set-status <slug> <stepId> done and bun bin/plan.ts advance <slug>.
  4. If publish — narrate ■ Step n/N /<skill> — publish gate: human required (nothing published). and STOP. Do not dispatch, advance, or publish.

Continue until a publish gate, a governor stop, or no ready step remains. The executor owns sequencing + gate enforcement only — each step's work is the leaf's (orchestrate, don't fuse).

Chain Position

Deference: build/propose the plan → /forsvn multi-step planner (A3); a single channel launch → run-launch; a generic produce-and-measure loop → run-pipeline; approve the plan → the human, in the review surface. run-plan runs what they approved; it never plans, launches, or approves.

Anti-Patterns

anti-patterns.md [ANTI-PATTERN] — run an unapproved plan · publish or advance past a publish gate · let .forsvn/config.json downgrade a gate · auto-ratify an artifact · black-box a step (no line) · fuse a step's work instead of dispatching its leaf · ignore a governor stop.

Durable Rules (protected)

<!-- SLOW_UPDATE_START --> <!-- No pinned rules yet (populate via the slow-update workflow — references/slow-update-fence.md). --> <!-- SLOW_UPDATE_END -->

Completion Status

  • DONE — the plan ran to a clean terminal: every ready non-publish step advanced + narrated within the envelope, and the run stopped at a publish gate, a governor limit, or plan completion.
  • DONE_WITH_CONCERNS — a leaf returned DONE_WITH_CONCERNS, or a step degraded; the run stopped legibly but a step needs attention.
  • NEEDS_CONTEXT — no plan at .forsvn/runs/<slug>/, or the slug is ambiguous.
  • BLOCKEDstatus != approved (a human must approve first), or a leaf GUARD_FAIL.

Worked Example

A 3-step plan that auto-advances research → copy, then stops at the publish gate — narration + gate resolution visible: run-plan-spec.md § Worked example.

What ships with it: 5 files

27.7 KB alongside SKILL.md

Keep looking

Skills are one crate of 325,949. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.