Parallel Dispatch

Parallel Dispatch Skill

Philosophy

Parallel dispatch is not "fire and forget" — it is "fire, verify, and integrate." The value of parallelism comes from independent execution, but the risk comes from unverified output. Every agent is treated as an untrusted contractor: given a clear scope, expected to deliver, but verified independently before their work is accepted. The dispatch orchestrator's primary job is not delegation — it is verification.

When to Use

• When 2+ independent tracks can execute concurrently without file conflicts
• For multi-module fixes, refactors, or feature additions across a workspace
• When time pressure demands parallel execution over sequential safety
• For commission-style work where each track has clear boundaries

Workflow

Phase 1: Plan

1. Parse tracks from user input
2. For each track, define:
   • Clear objective and scope
   • Specific file manifest (files the agent may touch)
   • Expected outputs and success criteria
   • Build/test commands to verify
3. Check for file conflicts — if two tracks touch the same file, use isolation: "worktree" or serialize them
4. Create a TodoWrite checklist with one item per track

Phase 2: Dispatch

1. Dispatch agents in parallel — one per track, max 5 concurrent
2. Each agent prompt must include:
   • The full file manifest and success criteria
   • Instruction: "Run all tests before reporting completion"
   • The build/test commands that will be used to verify
3. Model routing: Sonnet for straightforward implementation, Opus for architectural/complex tracks

File Manifest Protocol

Every agent prompt must include an explicit file manifest stating exactly which files the agent may create or modify. This is the zero-overlap guarantee that makes parallel execution safe.

Agent Dispatch Template:

"You own ONLY these files: [explicit list]. DO NOT read, modify, or create any files outside this list. If you need changes to files outside your manifest, report the need in your output — do not make the change."

Ownership rules:

• Agents that CREATE new files never conflict — new-file creation is inherently exclusive
• Agents that MODIFY existing files need exclusive ownership — one agent per file, no exceptions
• One agent per concern, not per file. Assign a complete feature slice (component + test + wiring), not a single file
• Main thread integrates: after all agents land, run go build ./..., go test ./..., wire cross-component imports, then one final build check

Evidence: validated across 20+ parallel dispatches in a single HTMLCraft Studio session — 3 phases, ~20 agents, 81 files committed, zero merge conflicts. The manifest is the mechanism: without explicit ownership, silent collisions are inevitable; with it, parallelism scales to any team size.

Phase 3: Verify

After ALL agents complete, run independent verification for each track:

1. git status to confirm expected files were changed
2. grep for expected patterns in modified files
3. Run the build/test commands specified in Phase 1
4. Check for duplicate imports, wrong formats, or regressions

Phase 4: Resolve

If any track fails verification:

1. Read the agent's actual output
2. Diagnose what went wrong
3. Fix directly or re-dispatch with corrected instructions
4. Re-verify after fix

Phase 5: Report

Output a verified status table: track | agent status | verification status | files changed

Output

The skill produces:

1. A per-track execution plan with file manifests
2. Parallel agent dispatches with scoped prompts
3. A verified status table with independent verification results
4. Commits per track (if requested)

Examples

3-track dispatch with clean results:

/dispatch
Track A: Fix auth middleware in AgenticGateway
Track B: Add unit tests to dojo-cli /init command
Track C: Update STATUS.md across all DojoGenesis repos

> Phase 1: 3 tracks, no file conflicts
> Phase 2: 3 agents dispatched (A: sonnet, B: sonnet, C: sonnet)
> Phase 3: All 3 verified green
> RESULT: 3/3 tracks complete, 12 files changed

Dispatch with one failure:

/dispatch
Track A: Consolidate 2 protocol handlers into 1
Track B: Add JWT validation to bridge.go

> Phase 2: 2 agents dispatched (A: opus, B: sonnet)
> Phase 3: Track A verified green, Track B failed (wrong import path)
> Phase 4: Fixed Track B import, re-verified green
> RESULT: 2/2 tracks complete after 1 remediation

Edge Cases

• File conflict between tracks: Serialize the conflicting tracks or use worktree isolation. Never let two agents write to the same file.
• Agent reports completion but files unchanged: Treat as failure. Re-dispatch with explicit instruction to persist changes.
• Agent fix breaks another track's work: Run integration tests after all tracks complete. Resolve conflicts before committing.
• More than 5 tracks: Batch into waves of 5. Complete and verify wave 1 before dispatching wave 2.
• Track depends on another track's output: Do not parallelize dependent tracks. Execute the dependency first, verify, then dispatch the dependent track.

Anti-Patterns

• Trusting agent self-reports: The most common failure mode. Always verify with git status, grep, and build/test commands.
• No file manifest: Without explicit file boundaries, agents may conflict or touch unrelated code.
• Dispatching without a plan: Jumping straight to dispatch without defining objectives, manifests, and success criteria leads to vague agent output.
• Using Opus for simple tasks: Model routing matters for cost and speed. Straightforward implementation gets Sonnet; architectural decisions get Opus.
• Committing before verification: Every track must pass independent verification before any commit is made.
• Re-dispatching without diagnosing: When a track fails, read the output first. The fix may be a one-line correction, not a full re-dispatch.

Model Enforcement

As of Apr 14, 2026, a PostToolUse hook (agent-model-enforce.sh) warns whenever an agent is dispatched without an explicit model: parameter. The rule is non-negotiable: always specify model: "sonnet" or model: "opus" in every agent call — never inherit the default.

Routing split:

• model: "sonnet" — parsing, bulk transforms, template generation, audits, straightforward implementation (~80% of dispatches)
• model: "opus" — architecture decisions, synthesis of competing constraints, strategic calls (~20% of dispatches)

Inheriting the default is an anti-pattern because the default may change, may differ by environment, and makes cost and quality unpredictable across a large swarm. Explicit model selection is part of the dispatch contract.

Quality Checklist

• All tracks defined with explicit objective, file manifest, and success criteria before dispatch
• File conflict check performed — no two tracks share a writable file
• Each agent prompt includes the file manifest and the build/test command
• Model explicitly specified for every agent (sonnet or opus) — never inherited
• Maximum 5 concurrent agents per wave
• All tracks verified after completion using git status, grep, and build/test — not agent self-reports
• Failed tracks diagnosed (output read) before re-dispatch
• Verified status table produced listing track, agent status, verification status, and files changed

Related Skills

• orchestration-pattern-selector (agent-orchestration): Determines whether parallel dispatch is the right topology for a task. Run it first when the coordination shape is non-obvious.
• audit-sweep-dispatch (agent-orchestration): Uses parallel dispatch as the wave execution mechanism for health audit findings. The wave/triage logic sits on top of this skill's dispatch protocol.
• maestro-orchestration (agent-orchestration): Full orchestration harness — uses parallel dispatch for the execution phase, adds persistent coordination and inter-agent communication on top.
• handoff-protocol (agent-orchestration): Structures the output contract between tracks when one track's output is another track's input across sessions or machines.