agent-dispatch

# agent-dispatch Right subagent, right tools, right brief. Pull from the domain registry. don't freestyle. ## When to invoke About to spawn an Agent AND: - Task is non-trivial (>3min inline, or 3+ files / 5+ commands per the delegation threshold) - Fits one of the 5 domains below Skip dispatch (do inline) if: <3min, depends on active session state, or multi-turn user conversation. ## The 5 domains Read `memory/Infra/agent-phase-domains.json` (canonical schema). One-line cheat sheet: | domain | purpose | subagent_type | writes? | |---|---|---|---| | **research** | doc-mining, repo audits, web research, codebase exploration | `Explore` | ❌ read-only | | **build** | code-writing: skills, scripts, configs | `general-purpose` | ✅ filesystem (scope-limited) | | **draft** | creative content: proposals, briefs, posts, decks | `general-purpose` | ✅ outputs/raw/ only | | **verify** | smoke tests, validation, audits, output QA | `general-purpose` | ❌ read+exec only | | **ops** | routine operations: status progression, follow-up, briefings, triage | `general-purpose` | ⚠️ outputs/ only, NO external writes | ## Dispatch flow 1. **Classify** the task. Pick ONE domain. If it spans 2+, decompose into separate subagent calls (one per domain): don't blend. 2. **Load** the domain entry from `agent-phase-domains.json`. Pull: `recommended_subagent_type`, `tools_denied`, `briefing_scaffold`. 3. **Brief** using the scaffold. Fill EVERY placeholder. Don't skip "stack context (paste-in)". su...