By name: run-pipeline description: "Orchestrates a closed-loop marketing/content initiative end to end — each stage a leaf skill — pausing at the execution fork and the human review gate. Use to thread a produce-and-measure loop, or --eval-only to scaffold just the loop workspace (improvement loop, experiment ledger) in .forsvn/loops/[slug]/. Not for one-shot artifacts (use the leaf skill) or debate (use debate-agents)." argument-hint: "[initiative name + asset type, e.g. 'launch hero image' or 'founder outbound sequence'] [--category image|video|design|publish|research] [--eval-only]" allowed-tools: Read Write Edit Grep Glob Bash disable-model-invocation: true metadata: version: "1.2.0" budget: standard estimated-cost: "$0.20-0.90"
Run Pipeline — Closed-Loop Orchestrator
Meta process skill. Threads the staged spine into one resumable loop — sequences leaf skills, stops at every gate, records the loop tree; it does NOT do each stage's work. Orchestrate, don't fuse (CLOSED-LOOP.md §2).
Core Question: "Can a future agent resume this from one loop folder — the live stage, the gate it waits on, what shipped — instead of re-deriving the chain?"
Stage model, gate rules, loop-tree contract:
references/pipeline-spec.md[SPEC]. Shares the schema in_shared/eval-loop-spec.md.
Critical Gates
- Orchestrate, don't fuse. Each stage dispatches its leaf skill (
research-*,brief-*/write-*,produce-*,evaluate-*); this skill owns only sequencing +pipeline.md. - Stop at every gate — never auto-approve. Execute pauses for the execution fork (
_shared/execution-fork.md); brief/execute outputs landdecision_state: pending(architecture §9.2). Guard this here, not in the leaf. - Fork is registry-gated.
list_tools(category)per_shared/tool-registry-spec.md: 0 verified → Brief-only; ≥1 → offer Assisted/Direct. Degrade cleanly. - Close the loop. Evaluate scores the re-ingested real asset (return-leg §6), not the brief. No ingest → open loop; flag it.
- Shared tree, no new store. State lives in
.forsvn/loops/[slug]/. No database.
Quality Gate
Critic (agents/critic-agent.md) PASS/FAIL:
- Stage integrity — all 6 stages named, each maps to a real leaf skill or gate; current-stage pointer present.
- Gate discipline — execute + brief stages marked with their gate; no auto-approve.
- Resumability —
pipeline.mdalone tells a fresh agent the live stage + next action. - Schema integrity —
results.tsv8-col header matches the evaluate-* contract. - Loop closure — ingest stage present so the evaluator scores the shipped asset.
FAIL → revise pipeline.md once, re-score. Eval-only runs are gated by the Eval Loop Critic instead (agents/eval-loop-critic-agent.md, all dimensions ≥ 8/10).
Before Starting
Prior loop folder in .forsvn/loops/[slug]/ → resume (read Current stage), don't re-scaffold. Verified engines (list_tools) decide the fork options. Manifest stale → bun bin/manifest-sync.ts. Full: references/_shared/before-starting-check.md [PLAYBOOK].
Pre-Dispatch
Warm/Cold Start: references/_shared/pre-dispatch-protocol.md. Cold Start needs: initiative name, asset category, primary metric, the measurable surface (eval-only adds mutable/frozen surface + baseline — see the mode reference). Mode: references/_shared/mode-resolver.md; --fast collapses Layer 1 + skips Critic revision on clean PASS, never the 5 Critical Gates.
Session execution profile (single-vs-multi): inherit per references/_shared/execution-policy.md.
Dispatch
Scaffold via bun scripts/scaffold-pipeline.ts "<name>" [--category <cat>] (idempotent). Per cycle: read state → dispatch the current stage's leaf → stop at its gate → on approval, advance Current stage + log the transition. At execute, the fork; at ingest, forsvn-preview attach; at evaluate, the matching evaluate-* appends a results.tsv row. Full sequence + resume rules: references/pipeline-spec.md.
Eval-Only Mode
--eval-only (or an eval-scaffold ask: improvement loop, experiment ledger, loop workspace) scaffolds/resumes the loop workspace without pipeline.md or the execute/ingest stages, via scripts/scaffold-eval-loop.ts. Gates, cold-start bundle, agents, loop-file conventions, upgrade path: references/eval-only-mode.md [PROCEDURE]. Per-cycle scoring still routes to the matching evaluate-* sibling.
Agent Manifest
Agent (agents/*.md) |
Mode | Focus |
|---|---|---|
| Pipeline Architect → Critic | full loop | Stage plan → Quality Gate above |
| Loop Architect + Metric Designer → Scope Guard → Eval Loop Critic | eval-only | Loop contract + measurement plan → boundary + rubric |
Artifact Contract
Loop folder layout (pipeline.md is the one artifact THIS skill owns; eval-only creates the same tree minus it): references/pipeline-spec.md § Shared loop tree. Frontmatter per _shared/artifact-contract-template.md. Cross-stack contract: results.tsv schema changes ripple to _shared/eval-loop-spec.md — evaluate-* siblings consume it.
Anti-Patterns
references/anti-patterns.md [ANTI-PATTERN] — fusing a stage, auto-approving a gate, closing a loop without ingest, metric-less loops, skill-centered folders, hand-edited ledgers.
Worked Example
"Launch hero image" end to end, plus the eval-only entry on the same tree: references/examples/closed-loop-walkthrough.md [EXAMPLE].
Durable Rules (protected)
Completion Status
- DONE — pipeline (or eval-only loop) scaffolded/resumed; current stage + next gate clear.
- DONE_WITH_CONCERNS — a stage lacks a leaf skill, the loop is open (no ingest), or the metric/baseline is weak.
- NEEDS_CONTEXT — missing initiative surface, asset category, or metric.
- BLOCKED — filesystem/scaffold failure or conflicting loop state.
Next Step
Dispatch the Current stage's leaf skill; on the human's gate decision, advance the pointer and log the transition. Eval-only loops: dispatch the cycle's strategy or brief-* skill, then the matching evaluate-* sibling — keep rows promote into learnings.md.