spec-orchestrator

# Spec Orchestrator

Coordinates the full lifecycle of specification development: writing, multi-agent review, revision, and approval. Simulates diverse team perspectives during review and iterates until the spec is approved or escalated to the user.

## Arguments

- `topic-or-spec` (positional): A topic to write a spec for, or path to existing spec
- `--skip-review`: Write spec without review cycle (just invoke spec-writer)
- `--reviewers=<p1,p2>`: Use only specific personas (comma-separated names)
- `--max-iterations=<n>`: Override max revision cycles
- `--require-approval`: Always require human approval (never auto-approve)

## Configuration

Check for `.claude/spec-workflow/config.yaml`:

```yaml
review:
  maxIterations: 3              # Cycles before escalating to human
  autoApproveThreshold: 0.8     # Consensus for auto-approval (0.0-1.0)
  selection:
    strategy: "auto"            # "auto" | "all" | "explicit"
    minimum: 2                  # Minimum reviewers
    maximum: 4                  # Maximum reviewers
```

### Philosophy Injection

If `.claude/spec-workflow/philosophy/review-criteria.md` exists, inject it into all reviewer prompts to guide their focus.

### Custom Personas

If `.claude/spec-workflow/personas/` exists and contains `.md` files:
- Load each file as a reviewer persona
- Parse frontmatter for `name` and `triggers`
- Use these instead of built-in personas

Otherwise, use built-in personas from references/personas.md.

---

## Modes

1. **Full loop**: Write spec → review → revise → re-review until approved
2. **Review only**: Review an existing spec and provide feedback

---

## Flow

```
Input (idea or spec)
     │
     ▼
┌─────────────────────┐
│ B: Is spec written? │
└─────────────────────┘
     │
     ├─No──► Invoke spec-writer (autonomous mode)
     │              │
     │              ▼
     └─Yes─► C: Does spec need agent review?
                   │
                   ├─No (trivial)──► Human review
                   │