advanced-dispatcher-skill（高级调度员技能）

Advanced Dispatcher

Classify work, dispatch it to the best spawned model, return the result to the main session.

Non-negotiable rules

• Never use Anthropic models unless the current prompt contains --force-claude.
• Reject --use-claude, --force-opus, --no-opus.
• --force-claude is prompt-scoped, not session-scoped.
• Deterministic routing over ad hoc judgment.
• openai-codex/* → long cache retention.
• opencode-go/* → short cache retention.
• Claude never appears in build pipelines; --force-claude is limited to tradeoff proposals.
• For coding work intended to merge, use Trunk-Based Development: one short-lived branch per feature/bug/fix/area.
• Every PR-ready build flow must run robust automated tests before calling the work ready.
• Keep git-visible diffs minimal, atomic, and limited to necessary files.

Routing table

Spanning categories → route by highest-risk deliverable: architecture > math > long-context > formatting.

Tradeoff protocol

Trigger when the user compares approaches, evaluates tradeoffs, or asks "which is better."

Default proposals: opencode-go/glm-5 + openai-codex/gpt-5.3-codex With --force-claude: anthropic/claude-sonnet-4-6 + anthropic/claude-opus-4-6 Judge (always): openai-codex/gpt-5.4

Build pipelines

Trigger on buildq:, build:, or buildx: prefix.

Pre-step: create or verify one short-lived branch for the scoped work.

buildq: (quick, 5 steps)

1. plan → gpt-5.4
2. implement → gpt-5.4
3. test → glm-5
4. simplify → gpt-5.3-codex
5. retest → glm-5

build: (standard, 10 steps)

1. parallel-plan-a → gpt-5.4
2. parallel-plan-b → glm-5
3. judge-plan → gpt-5.4
4. boilerplate → gpt-5.3-codex-spark
5. implement → gpt-5.4
6. test → glm-5
7. simplify → gpt-5.3-codex
8. retest → glm-5
9. review-resolve → gpt-5.4
10. final-test → glm-5

buildx: (strict, 12 steps)

1. parallel-plan-a → gpt-5.4
2. parallel-plan-b → glm-5
3. judge-plan → gpt-5.4
4. boilerplate → gpt-5.3-codex-spark
5. implement → gpt-5.4
6. test → glm-5
7. simplify → gpt-5.3-codex
8. retest → glm-5
9. review-resolve-a → gpt-5.4
10. test-a → glm-5
11. review-resolve-b → kimi-k2.5
12. final-test → glm-5

Exit rule: do not mark the branch PR-ready until automated tests pass and the diff stays atomic.

Judge output contract

Judge-plan must emit:

1. Selected architecture
2. Why it won
3. Project/file structure
4. Implementation order
5. Branch plan (name, scope boundary)
6. Test plan
7. PR/CI test gates
8. Simplification targets
9. Done criteria

For buildx:, also include:

1. Risk list
2. Likely failure modes
3. Review checklist

Simplify contract

Must:

• Remove dead code
• Remove speculative abstractions
• Remove duplication
• Remove over-engineered interfaces
• Prefer fewer files when clarity is preserved

Must not:

• Rewrite architecture
• Add abstractions
• Expand scope

Implementation contract

dispatcher.py must produce deterministic RoutePlan objects, expose route choice + cache retention + rationale + pipeline steps, support tradeoff/buildq/build/buildx, reject empty prompts and legacy flags.

Validation checklist

1. Run test_dispatcher.py — all pass.
2. Smoke test tradeoff, buildq:, build:, buildx:.
3. Confirm Claude unreachable without --force-claude.
4. Confirm Claude never in build pipelines.
5. Confirm cache retention correctness.
6. Confirm judge output includes branch plan and PR/CI test gates.