dispatch

dispatch

Split independent work into ≥2 concurrent agent tracks, or invoke a remote agent-box task/prompt.

Preamble — RAG pre-flight

Before planning dispatch, query prior dispatch runs for this task domain:

graphify query "dispatch <task-domain> agent plan" --budget 300

• If result shows a dispatch plan for the same task within 24h → surface it, ask user to confirm whether to reuse agent assignments or start fresh.
• If no recent match → proceed to Phase 1.

Done when: cached dispatch plan surfaced or no match found (proceed).

---

When to use

• Per-repo scans (multiple repos, no cross-repo deps)
• CI/lint/test audits (independent job analysis)
• Dependency audits (N packages in one sweep)
• Multiple PR review passes (independent review lenses)
• Standards or docs comparison (independent sources)
• Remote agent-box task invocations (see Remote Agent Dispatch below)

Workflow

Step 1: Validate independence

Check that each track does NOT depend on output from another track. Done when: you can name each track's input + output without forward references.

Step 2: Define track specs

For each track, name:

• Input (what it starts with)
• Single responsibility (what it answers)
• Output format (bullet list, counts, verdict, or artifact path)

Done when: each track has ≥1 sentence describing its output, printed for user review; user has not objected within 10s.

Step 2a: Critic gate (after agent plan is drafted)

Dispatch ONE Explore agentType critic — read-only, never edits — with:

"Challenge this dispatch plan: Which agent assignments are wrong for the task type? Which tasks are under-specified (agent will hit NEEDS_CONTEXT)? What dependency between tasks was missed? What should be sequential that was marked parallel?"

• If critic identifies ≥1 wrong assignment or missing dependency → revise plan before dispatch.
• Minor concerns → log in dispatch summary, proceed.

Done when: critic verdict returned; wrong assignments corrected or none found.