request-dispatcher
暂无简介
- License 许可
- MIT
- 所属仓库
- zxpmail/ReqForge
- 最近更新
- 2026-10-07
- Stars
- ★ 18
- Forks
- 5
02 资源文件 7
- 参考文档commands/request-dispatcher.md527 B
- 参考文档references/anti-rationalization.md797 B
- 参考文档references/dimension-checklist.md1.1 KB
- 参考文档references/first-principles.md774 B
- 参考文档references/output-style.md1002 B
- 参考文档references/workflow.md1.4 KB
- 参考文档skill.json816 B
03 SKILL.md 正文
<!-- forge: request-dispatcher v1.0 -->
name: request-dispatcher description: Used when the user's request is ambiguous and doesn't clearly match a single Skill's trigger criteria. Analyzes intent + project state to route to the correct Skill. Not a replacement for CLAUDE.md Skill Dispatch rules — only for edge cases where dispatch is uncertain. version: 1.0.0 updated: 2026-05-26 requires: []
<!-- begin: task --> [Task] Analyze ambiguous user requests and route them to the correct ReqForge Skill. When the main Agent's static dispatch rules (from CLAUDE.md [Skill Dispatch]) cannot uniquely determine the target Skill, this Skill provides a structured decision framework: classify intent → cross-reference project state → disambiguate → recommend Skill.
<!-- end: task --> <!-- begin: not-for --> [Not For] - Clear-cut requests that match a single Skill trigger -> use the target Skill directly - Writing code -> use /dev-builder instead - Bug fixing -> use /bug-fixer instead - Requirements gathering -> use /product-spec-builder instead
<!-- end: not-for --> <!-- begin: dependency-check --> [Dependency Check] Required: - User's natural language request -> the ambiguous message text - Project state awareness -> which artifacts exist (Product-Spec.md, DEV-PLAN.md, code, active changes/)
Optional:
- git log -> recent changes for context
- memory/ files -> known pitfalls and recent task history
<!-- end: dependency-check -->
<!-- begin: first-principles -->
[First Principles]
→ references/first-principles.md
核心:Minimum Dispatch / State Before Intent / Ask Don't Guess。
<!-- end: first-principles -->
<!-- begin: output-style -->
[Output Style]
→ references/output-style.md
Tone: neutral, analytical, decisive. One recommendation, not a menu.
<!-- end: output-style -->
<!-- begin: file-structure -->
[File Structure]
request-dispatcher/ ├── SKILL.md └── references/ ├── first-principles.md ├── output-style.md ├── workflow.md ├── dimension-checklist.md └── anti-rationalization.md
<!-- end: file-structure --> <!-- begin: output-artifacts --> [Output Artifacts] - Dispatch recommendation (screen output) — one-line Skill name + one-sentence justification
<!-- end: output-artifacts --> <!-- begin: dispatch-dimension-checklist --> [Dispatch Dimension Checklist] Before recommending a Skill, verify:
| Dimension | Must-Have | Recommended |
|-----------|-----------|-------------|
| **Intent Match** | User's message maps to a Skill's trigger description in CLAUDE.md | User's message is NOT already covered by a static rule (would be unnecessary dispatch) |
| **Prerequisite Check** | Target Skill's Dependency Check would pass | Target Skill's Not For section doesn't exclude the request |
| **Project State** | Required artifacts exist for the target Skill | No active changes/ folder conflict |
| **Ambiguity Resolved** | Recommendation resolves the ambiguity | User would not be surprised by the routing |
<!-- end: dispatch-dimension-checklist --> <!-- begin: dispatch-decision-strategy --> [Dispatch Decision Strategy]
| User says... | Product-Spec.md exists? | Active changes/? | Code exists? | Route to |
|---|---|---|---|---|
| "I have an idea / build something" | No | — | — | /product-spec-builder |
| "Add a feature / change something" | Yes | No | — | /change-manager propose |
| "Add a feature / change something" | Yes | Yes | — | /change-manager (resume active) |
| "This is broken / error / bug" | — | — | Yes | /bug-fixer |
| "Review / check quality" | Yes | — | Yes | /code-review |
| "Plan / how to build" | Yes | — | — | /dev-planner |
| "Build / implement / code" | Yes | — | No | /dev-builder |
| "Build / implement more" | Yes | — | Yes | /dev-builder |
| "Design / make it look good" | Yes | — | — | /design-brief-builder → /design-maker |
| "Release / deploy / ship" | — | — | Yes | /release-builder |
| "Grill me / stress-test plan / 烤问" | — | — | — | /product-spec-builder → Light Grill Mode |
| "Zoom out / explain this module" | — | — | Yes | /dev-builder → [Zoom-Out Pass] (read-only) |
| "Architecture health / ball of mud" | — | — | Yes | /dev-planner → architecture-health-pass |
| "Break into GitHub issues" | Yes | — | — | /dev-planner → optional issue slices (after plan confirm) |
| "forge-smoke failed / adapter drift / gate red" (in ReqForge repo) | — | — | — | /reqforge-greenkeeper |
**Ambiguity patterns**:
- "Improve this" (no target specified) -> ask: "Improve functionality (bug-fixer), quality (code-review), or UI (design skills)?"
- "I don't like how this works" -> ask: "Is it a bug (wrong behavior) or a feature request (missing capability)?"
- "Make it production-ready" -> ask: "Bug fixes first (bug-fixer), code quality (code-review), or release pipeline (release-builder)?"
- User mentions both a bug and a feature -> route to the dominant intent first; fix bug before adding features
<!-- end: dispatch-decision-strategy --> <!-- begin: gotchas --> [Gotchas] Premature dispatch: "This looks like X, let me dispatch immediately." — Without completing the three-step analysis (intent → state → disambiguate), you may route to the wrong Skill. Always complete all three steps. Prerequisite blindness: Dispatching to /dev-planner without Product-Spec.md, or /code-review without code. — Always verify the target Skill's prerequisites before recommending. A recommendation that bounces is worse than a clarifying question. Over-routing: Using request-dispatcher for every request instead of just the ambiguous 10%. — The static rules in CLAUDE.md [Skill Dispatch] handle 90% of cases. Only invoke this Skill when the static rules don't clearly match. Ignoring active changes/: User asks for a new feature while changes/ folder is active. — Active changes/ must be resolved (verify + archive) before starting new work. Route to /change-manager for the active change, not a new Skill.
<!-- end: gotchas -->
<!-- begin: anti-rationalization-checklist -->
[Anti-Rationalization Checklist]
→ references/anti-rationalization.md
遇 premature dispatch / prerequisite blindness 时读取。
<!-- end: anti-rationalization-checklist --> <!-- begin: dimension-checklist --> [Dimension Checklist] See references/dimension-checklist.md for the full dimension checklist.
Must-have dimensions:
- **Intent Match**: user message maps to a Skill's trigger description
- **Prerequisite Check**: target Skill's Dependency Check would pass
- **Project State**: required artifacts exist for the target Skill
- **Ambiguity Resolved**: recommendation resolves the ambiguity
<!-- end: dimension-checklist --> <!-- begin: quality-rubric --> [Quality Rubric] 8-item, 16-point scoring system. Ship threshold: ≥ 12 with no critical item scoring 0.
| # | Dimension | Pts | Critical | Scoring |
|---|-----------|-----|----------|---------|
| 1 | Intent classification | 2 | YES | 2 = Correctly maps user message to intent type via matrix; 1 = Ambiguous but close; 0 = Wrong classification |
| 2 | State cross-reference | 2 | YES | 2 = Checks artifact existence before recommending; 1 = Partial check; 0 = Skips state check |
| 3 | Prerequisite verification | 2 | — | 2 = Target Skill's Dependency Check would pass; 1 = Prerequisite exists but unmet; 0 = Blind dispatch |
| 4 | Ambiguity resolution | 2 | — | 2 = Disambiguates before dispatching when uncertain; 1 = Asks but with bias; 0 = Guesses without asking |
| 5 | Dispatch minimality | 2 | YES | 2 = Only used for ambiguous 10%, not overriding static rules; 1 = Used when static rule sufficed; 0 = Routine over-routing |
| 6 | Active change awareness | 2 | — | 2 = Checks active changes/ before recommending; 1 = Checks but ignores; 0 = No check |
| 7 | Recommendation clarity | 2 | — | 2 = One Skill name + one-sentence justification; 1 = Multiple options without clarity; 0 = No clear recommendation |
| 8 | Decision matrix usage | 2 | — | 2 = Cites matrix row when matched; 1 = Matrix used but not cited; 0 = Ignores decision matrix |
**Scoring**: Run `pnpm validate-skill --score core/skills/request-dispatcher` to compute.
<!-- end: quality-rubric -->
<!-- begin: workflow -->
[Workflow]
→ references/workflow.md
4 步流程:Classify → Cross-reference → Disambiguate → Recommend。
<!-- end: workflow --> <!-- begin: initialization --> [Initialization] Step 1: Execute [Workflow] <!-- end: initialization -->
04 同仓库更多 Skills
feedback-writer
Analyze and record user corrections, feedback, and behavior preferences as structured feedback files for the evolution engine.
evolution-engine
Scan accumulated feedback to identify patterns, generate evolution proposals for SKILL.md, rules, or new Skills.
db-migration-checker
暂无简介
skill-builder
暂无简介
request-dispatcher
暂无简介
reqforge-greenkeeper
暂无简介