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:

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)". subagents have no transcript memory.
4. Constrain: explicitly include the tools_denied list as "Hard rules: do NOT use these tools" in the briefing. Subagents can't read the schema, you have to inline the rules.
5. Dispatch via the Agent tool with subagent_type = the recommended one.
6. Verify the diff on return. Subagent wrote files? Read them. Don't trust the summary blind. that's the "trust the summary, verify the diff" delegation rule.

Briefing template (universal)

Every brief answers 6 questions. Domain scaffolds in the JSON add 3-4 domain-tuned lines on top.

1. Goal? (one sentence)
2. Why? (why this task exists. the subagent has no transcript)
3. What's known / ruled out? (don't re-discover)
4. Scope? (specific files/paths/questions, NOT "investigate X area")
5. Out of scope? (denylist)
6. Success criteria? (output format, length budget, pass test)

Hard rules (inherit from delegation discipline)

• Max depth = 2. The parent (this session) can spawn ONE child via the Agent tool. The child must NOT spawn further Agent calls. When briefing, NEVER write "spawn N more agents" or "delegate to other agents" in the prompt.
• Memory isolation. Don't paste full memory dumps. Paste the SPECIFIC fact/file path the subagent needs.
• No MCP write tools in subagents. The parent owns external writes (chat, spreadsheet, knowledge graph). Always.
• Verify diffs on return. Read changed files, don't blindly accept summary text.
• Skip dispatch for <3min tasks: overhead exceeds value.

Anti-patterns (from agent-phase-domains.json)

• ❌ Pasting the full voice constitution / user model into a draft subagent. Paste the specific voice rules + anti-AI-slop subset relevant to the piece.
• ❌ Letting a research subagent edit files. They report findings, the parent writes the reference page.
• ❌ Blending build + verify into one subagent. Build, return the diff, dispatch verify separately if needed.
• ❌ Allowing an outbound-message tool in a draft/build/ops subagent. They write to outputs/, the parent decides if it goes external.

Example dispatches

research domain. repo doc-mine

subagent_type: Explore
brief includes:
  - Goal: "Deep-mine github.com/X/Y for 5 specific patterns"
  - Stack context: paste the relevant 5-10 lines of your existing infra
  - Specific files to read: explicit URLs/paths
  - Output format: structured markdown with file path + line range cites
  - Tools denied: Edit, Write, npm/pip install, subagent spawning

build domain. new script

subagent_type: general-purpose
brief includes:
  - Goal: "Build scripts/your_new_tool.py that does X"  # e.g. an example placeholder path
  - Files to read first (existing patterns to mirror): list
  - Smoke-test command: single bash invocation
  - Out-of-scope: don't touch skills/, don't add MCP wiring
  - Tools denied: external-write MCP tools, schedule_create

verify domain. output QA

subagent_type: general-purpose
brief includes:
  - Goal: "Smoke-test outputs/raw/agent/<file>.md against the rubric"
  - Pass criteria: explicit checks (frontmatter present, no slop phrases, links resolve)
  - Output: pass/fail + line cites for each failure
  - Tools denied: Edit, Write, all MCP writes (verify never mutates)

ops domain. status-progression routine

subagent_type: general-purpose
brief includes:
  - Goal: "Run a status-progression check on active items"
  - Data sources: explicit knowledge-graph paths + spreadsheet ranges
  - Output destination: outputs/raw/agent/status-<date>.md
  - CONFIRM-gate flags: any terminal-state candidates require the parent's confirmation BEFORE a knowledge-graph edit
  - Tools denied: outbound message, spreadsheet update, knowledge-graph create (parent writes external)

Logging

Log the dispatch + verification outcome to logs/session.jsonl:

{"ts":"<ISO>","persona":"default","category":"agent-dispatch","event":"dispatched","domain":"research","subagent_type":"Explore","verified":true}

Verified = the parent read returned files / confirmed claims / accepted the summary. If verified=false, also log the discrepancy.

Cross-references

• memory/Infra/agent-phase-domains.json: canonical schema (this skill reads it)
• docs/agent-protocols.md: the broader agent protocol (3 surfaces: persona handoff / subagent briefing / cross-bot handoff)
• The system prompt's # Session Self-Regulation → Delegation threshold + Delegation discipline (3 hard rules)

Adoption

Opt-in. Use for: repo audits, multi-step research, ops routines with a pattern (status-progression, follow-up). Skip for: trivial inline work, multi-turn conversation, single-file ops.