propose-workflow

Propose a reviewable workflow

Use when the user wants to design or build a workflow and review it before it runs — "let's design a workflow for X", "build me a workflow but let me see it first", "what shape would you use for X". You emit a blueprint (the shape: pattern, phases, per-node role/model/effort/instruction, symbolic fan-out, edges), the panel draws it as a graph, and the user approves / rejects / edits. Only on approve do you author and run it.

The judgment worth reviewing is which algorithm you chose — fan-out map→reduce for "audit this repo" vs a judge panel for "design an API". Emit the shape; let the human check it.

1 — Propose, or just run?

Mode is the user's intent, not a setting — infer it:

• "Just run a workflow" / "go" / "do X across the files" → skip this gate. Author and run it the normal way — the declare → lock → trust → agentstack workflow run pipeline in §4. Same pipeline, gate auto-skips.
• "Design / build / let me review / show me the shape first" → propose a blueprint (below), then stop and wait. Do not author, do not run.

When unsure which the user meant, propose — a review gate is cheap to approve and expensive to skip.

2 — Emit the blueprint, then stop

Pick the best-fitting pattern for the task and name it. One node per role/step; give each a model/effort/instruction; edges carry a kind. Emit it in a fenced block whose language tag is exactly agentstack-blueprint — that tag is how the panel intercepts and renders it.

```agentstack-blueprint
{
  "workflow": "repo-audit",
  "pattern": "map-reduce",
  "goal": "Find and rank bugs across the changed files",
  "nodes": [
    { "id": "map", "phase": "Find", "role": "reviewer",
      "model": "gpt-5.5", "effort": "low",
      "instruction": "Scan ONE changed file for correctness bugs",
      "fanout": "1 per changed file" },
    { "id": "reduce", "phase": "Rank", "role": "synthesizer",
      "model": "opus", "effort": "high",
      "instruction": "Dedupe and rank all findings by severity",
      "fanout": null }
  ],
  "edges": [ { "from": "map", "to": "reduce", "kind": "fan-in" } ]
}
```