Use when the user describes a multi-step feature or refactor and you need to decide whether to handle it solo, dispatch a parallel `/fan-out`, or run a `/worktree-team` planner→generator→reviewer pipeline. Outputs a routing recommendation with the reasoning behind it.
インストール
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
Use when the user describes a multi-step feature or refactor and you need to decide whether to handle it solo, dispatch a parallel `/fan-out`, or run a `/worktree-team` planner→generator→reviewer pipeline. Outputs a routing recommendation with the reasoning behind it.
when_to_use
Reach for this before starting any task that may touch 5+ files, has independent sub-tasks worth parallelizing, or carries enough risk to warrant separated planning and review. After the planner phase, dispatch also scales the reviewer pool adaptively (N≥3 independent files → N parallel reviewers, capped at 5). Do NOT use for executing the dispatched pattern — use `/fan-out` for parallel batches or `/worktree-team` for full pipelines instead. Do NOT spawn a pool for mutually coupled files where review must consider all files together — use a single reviewer.
disable-model-invocation
true
allowed-tools
["Read"]
counterexamples
["Executing the dispatched pattern — use /fan-out or /worktree-team after the route is picked.","A one-line fix or small bug touching ≤2 files — execute directly without routing overhead.","Exploratory questions where no concrete task exists to route.","Non-independent file work (mutual coupling means review must consider all files together — use single reviewer)."]
contract
{"required_outputs":["Routing recommendation block (Route / Reason / Agent(s) / Estimated scope / Risk level).","Reviewer pool decision block when running the Pipeline route (pool_size, files, decision_reason)."],"budget":"1 model turn","permission_scope":"Read-only on task description","completion_conditions":["Exactly one route classification emitted (Single Agent | Fan-Out | Pipeline | TDD-Loop).","One-sentence reason and risk level included.","Pool decision logged to .claude/handoffs.jsonl when pipeline route is chosen."],"output_paths":["stdout"]}
scheduling
user describes a task whose scope, file count, or operation type warrants a routing decision before execution; or a planner output arrives and reviewer pool size must be decided
structural
["Read the task description and infer file count, operation type, interdependence","Apply the dispatch decision matrix (single-agent vs fan-out vs pipeline vs tdd-loop)","Emit a route classification with reason and risk level",{"When pipeline route":"parse planner output for explicitly enumerated independent files; decide pool size; log decision; dispatch reviewers"},"Hand off to the chosen pattern's invocation skill"]
logical
a single route classification (single-agent | fan-out | pipeline | tdd-loop) is emitted with a one-line reason and a risk-level tag; reviewer pool size matches planner's declared file count (capped at 5)
/dispatch — Task Decomposition Router
Decision Matrix
Analyze the task against these criteria:
Signal
Single Agent
Fan-Out
Pipeline (P/G/R)
Files touched
1-3
4-15 (similar ops)
4-15 (different ops)
Task type
Bug fix, small feature
Batch migration, bulk refactor
New feature, architecture change
Interdependence
High (changes depend on each other)
Low (same operation, different files)
Medium (phases depend on prior phase)
Risk
Low
Low-Medium
Medium-High
Routing Rules
Route 1: Single Agent (do it yourself)
Task touches ≤ 3 files
Changes are interdependent
No ambiguity in requirements
Action: Execute directly, no subagents needed
Route 2: Fan-Out (parallel batch)
Same operation applied to multiple files
Changes are independent of each other
Action: Use /fan-out to dispatch parallel subagents
Sweet spot: 3-5 parallel agents. More than that is hard to review.
Route 3: Planner → Generator → Reviewer Pipeline
New feature or architectural change
Multiple phases with different concerns
Higher risk warrants review before completion
Action: Dispatch agents sequentially:
Planner (Write/Edit scoped to .claude/plans/): Explore codebase, identify patterns, propose approach, and write the plan to .claude/plans/s<N>-<slug>.md. Must include a ## Contract section with testable criteria and verification method, plus ### Tasks / #### T<n> headings for multi-task sprints.
Generator (read-write): Invoke /contract to confirm criteria, then implement based on planner's output.
Reviewer (read-only): Check contract compliance first, then challenge the implementation.
Adaptive Reviewer Pool
After the planner subagent finishes, scan its output for explicitly enumerated independent files before dispatching reviewers. "Independent" is the planner's call to declare — dispatch counts what the planner lists, it does not infer independence from file content.
Parsing planner output for independent files
Look for any of these markers in the planner's output:
Numbered list with file paths — lines matching ^\d+\. .*(\.py|\.sh|\.md|\.json|\.ts|\.js).
Explicit FILES: block — lines following a FILES: or ## Files heading, each starting with - or a path.
Markdown task list — lines matching ^- \[ \] .*path.* that name distinct file paths.
Count only paths from the planner's explicit enumeration. If the planner describes work in prose without a list, treat file count as 0 and fall through to single-reviewer behavior.
Pool size decision
N = count of explicitly enumerated independent files (as listed by planner)
N = min(N, 5) ← hard cap; context overhead at 6+ reviewers outweighs coverage gain
if N >= 3:
dispatch N reviewer subagents in parallel (one Agent tool call per file, with target_file: <path>)
dispatch 1 aggregator reviewer (sees all N individual findings, no target_file)
else:
dispatch 1 reviewer (current behavior, no target_file needed)
The cap is enforced in the pool decision, not just documented. If the planner enumerates 8 files, dispatch 5 reviewers + 1 aggregator covering files 1–5; note the cap in the handoff log.
The cap of 5 is a reviewer-pool ceiling, not a limit on the work itself. When the independent work set is genuinely large (a codebase-wide migration or audit), the execution can hand off to a native dynamic workflow (the ultracode keyword) that orchestrates many more subagents under the runtime — while /dispatch keeps its distinct role: the routing decision, reviewer-pool sizing, and handoff logging that the native feature does not do. Available on Claude Code builds that support dynamic workflows.
Dispatcher prompt for a pooled reviewer
You are reviewer-<N> in a parallel review pool. Review only the file assigned to you.
target_file: <path>
Scope your review to <path> and its direct callers or tests that import it.
Do not re-review files assigned to sibling reviewers.
Emit your findings in the standard reviewer format (Verdict, per-check evidence, findings).
Aggregator reviewer role
The aggregator receives all individual reviewer findings (not the original code) and has a distinct job from per-file reviewers:
Merge findings across files, deduplicating identical issues reported by multiple reviewers.
Surface inter-file inconsistencies — a type declared in a.py but used incorrectly in b.py that no single reviewer saw because each scoped to one file.
Escalate any finding that a single reviewer marked low-severity but which, in context of other findings, rises to high-severity.
Emit a unified verdict (ACCEPT / REJECT / NEEDS DISCUSSION) with a summary of merged high/medium/low counts.
You are the aggregator reviewer. You have received findings from <N> parallel per-file reviewers.
Your job: merge, deduplicate, surface inter-file inconsistencies, and emit a single unified verdict.
You do NOT re-read source files. You reason over the reviewer findings only.
Logging the pool decision
After deciding pool size, append one JSONL entry to .claude/handoffs.jsonl:
Paper reference: arXiv:2605.18747 §4.1.3 (SoA, MAGIS) — agent pool size should scale with task complexity; fixed-size pools leave coverage gaps on large file sets.
REVIEWER POOL:
Pool size: <1 or N> reviewer(s) [+ 1 aggregator if N≥3]
Files: <list or "single-reviewer mode">
Decision reason: <single-file|pool-of-N|capped-from-M>
Reviewer prompt template
When dispatching a reviewer subagent, structure the prompt so the response starts with a 2-line verdict, then evidence, then findings. This survives output truncation: the verdict is always readable even if the tail is cut.
If the script prints output, prepend that output verbatim to the subagent's prompt — preserving the [contract] header line. This gives the subagent the freshly-written contract block that contract-reread.sh produced at SubagentStart time, rather than a potentially compacted in-context copy.
If the script prints nothing (missing or empty active-contract.md), proceed without contract injection. This is normal for non-pipeline workflows where no active plan exists.
Why: contract-reread.sh fires on every SubagentStart and writes the current plan's ## Contract section to .claude/state/active-contract.md with a fresh mtime. Reading it at dispatch time — rather than recalling from context — guarantees the subagent sees a post-compaction-safe contract per Sprint Contract Protocol (HARNESS_SPEC.md). File-based handoff survives context boundaries; in-context memory does not.
Route picked from file count alone. Five similar files reading "fan-out" can still be sequential if each step depends on the previous result. Before locking the route, check interdependence, not just count.
Pipeline chosen for a one-shot bug fix. The planner→generator→reviewer round-trip costs ~3× tokens; a 1–3 file fix doesn't earn the overhead. Prefer Single Agent unless the change spans concerns or carries deploy risk.
Fan-out with shared mutable state. Two parallel subagents editing the same file race; both succeed, second overwrites first. The dispatch decision should refuse fan-out when the file list overlaps.
LLM fallback non-termination. When route-prompt-llm.sh is in play and disagrees with the shell verdict, the router can flap. Cap with WORKFLOW_ROUTER_MODE=shell for deterministic dispatch when investigating.
Pool spawned for coupled files. If the planner lists 4 files but they share mutable state or a type that crosses all of them, per-file reviewers will miss the cross-cutting inconsistency. The aggregator catches some of this, but the planner should mark files as non-independent when coupling is tight.
Planner prose instead of list. If the planner describes files in prose without a numbered/task list, the file count is treated as 0 and single-reviewer mode applies. Ask the planner to re-emit a FILES: block if pooling is desired.
Rebuttals
Common rationalizations for shortcutting the routing decision, with rebuttals:
Excuse
Rebuttal
"It's obviously a single-file change — single-agent."
"Obvious" is the most common failure pre-condition. Single-file edits with cross-cutting type or test impact still benefit from pipeline review. The classification cost is one inference; the cost of a wrong route is the whole task.
"Fan-out is overkill for 3 files."
Fan-out's value is isolation, not parallelism. Three independent files in one context window contaminate each other; three subagents do not. File count is a weak proxy for the right route.
"Skip TDD just this once — the requirement is clear."
"Clear" requirements are exactly when TDD is cheapest — the test writes itself. Skipping it forfeits the artifact that proves the requirement was met.
"I'll just do it without recording the route."
An undocumented routing decision is unreviewable. The one-line classification with reason is the audit trail; without it, a wrong route looks identical to a right one in retrospect.
"Pipeline is too heavy for a refactor."
Pipeline overhead is fixed; refactor risk scales with surface area. Renaming three callers across two files is exactly when the planner→generator→reviewer separation pays off.
