AgentSkillsCN

capability-discovery

通过动态插件枚举与能力映射,结合委托优先级与显式路由公告,实现高效协作。

SKILL.md
--- frontmatter
name: capability-discovery
description: Dynamic plugin enumeration and capability mapping with delegation priority and explicit routing announcements
triggers:
  - plugin discovery
  - capability mapping
  - agent routing
  - plugin enumeration
  - delegation

Capability Discovery Skill

This skill enables dynamic discovery and mapping of installed plugins to abstract capability categories, allowing the workflow to route tasks to appropriate agents without hardcoding plugin names. It includes explicit delegation announcements so users understand routing decisions.

Core Principles

  1. Technology Agnostic: Map capabilities, not specific plugins
  2. Dynamic Discovery: Scan at runtime, not compile time
  3. Graceful Fallback: Use internal default agents when no plugin matches
  4. Project-Aware Routing: Prefer plugins matching detected project type
  5. Transparent Delegation: Always announce why a particular agent was chosen

Capability Categories

CapabilityPurposePhaseDefault Agent
requirements-gatheringStructured requirements elicitationDISCUSSdefaults/interviewer
brainstormingIdeation and option explorationDISCUSSdefaults/interviewer
codebase-analysisUnderstanding existing codeDISCUSS, PLANdefaults/researcher
tdd-implementationTest-driven developmentEXECUTEdefaults/executor
code-implementationGeneral code writingEXECUTEdefaults/executor
infrastructureDevOps/infra tasksEXECUTEdefaults/executor
code-reviewQuality validationVERIFYvalidator
requirements-validationRequirements complianceVERIFYvalidator

Discovery Process

Step 1: Enumerate Installed Plugins

markdown
1. Locate Plugins/ directory
2. List all subdirectories (each is a plugin)
3. For each plugin:
   a. Read .claude-plugin/plugin.json for metadata
   b. Scan agents/*.md for agent definitions
   c. Scan commands/*.md for command definitions
4. Build plugin registry

Step 2: Extract Descriptions

For each agent/command file, extract the description from YAML frontmatter:

yaml
---
name: implementer
description: TDD implementation specialist for .NET. Use to write minimal code that makes tests pass.
model: opus
tools: [Read, Write, Edit, Bash, Grep, Glob]
---

Step 3: Map to Capabilities (Keyword Scoring)

Apply keyword matching to descriptions:

markdown
For each plugin agent/command:
  1. Tokenize description into words (lowercase)
  2. Match against capability keywords
  3. Calculate score:
     - Exact keyword match: +10 points
     - Partial match (word contains keyword): +3 points
     - Keyword in first sentence: +5 bonus
     - Multiple keywords from same category: +2 each after first
  4. Assign to highest-scoring capability
  5. Record match confidence (High: 25+, Medium: 15-24, Low: <15)

Step 4: Detect Project Type

Scan project root for technology indicators:

File PatternProject TypeKeywords
*.csproj, *.slndotnet.NET, dotnet, C#, csharp
package.json, *.js, *.tsnodeNode.js, JavaScript, TypeScript
*.py, requirements.txtpythonPython, pip, pytest
go.mod, *.gogoGo, golang
Cargo.toml, *.rsrustRust, Cargo
pom.xml, build.gradlejavaJava, Maven, Gradle

Step 5: Cache Results in FLOW.md

Store capability mappings in the Capabilities Cache section of FLOW.md.

Delegation Priority System

When routing a task to an agent, follow this priority order:

Priority 1: Project-Type Match

If project is detected (e.g., dotnet), prefer plugins with matching technology keywords in their description.

markdown
Detected: dotnet project
Capability needed: tdd-implementation

Matches:
- dotnet-tdd:implementer (description: "TDD for .NET") → Score: 35 + project bonus
- node-tdd:implementer (description: "TDD for Node.js") → Score: 30

Winner: dotnet-tdd:implementer (project type match)

Priority 2: Exact Keyword Match

Prefer agents with exact keyword matches over partial matches.

Priority 3: Agent Over Command

When both an agent and a command match a capability, prefer the agent (more capable).

Priority 4: Internal Default

When no plugin matches the capability, use the appropriate default agent.

Delegation Announcement Pattern

CRITICAL: Always announce delegation decisions to the user with reasoning.

When Plugin Found

markdown
**Delegating [capability]** → [plugin-name]:[agent-name]

Matched via keyword scoring:
- Keywords: [matched keywords]
- Score: [score] (High confidence)
- Project type: [type] (matched)

Spawning agent...

When Using Default

markdown
**Using built-in agent** for [capability] → flow-workflow:defaults/[agent]

No installed plugin matched this capability.
- Searched: [N] plugins
- Keywords tried: [keywords]

Consider installing: [suggested plugin type]

Quick Delegation (For Status/Logs)

markdown
→ [capability]: [plugin:agent] (keyword match)
→ [capability]: defaults/[agent] (no plugin match)

Routing Logic

When Phase Needs Capability

markdown
1. Determine capability needed for current task
2. Look up FLOW.md capability cache
3. If cache stale (>24h) or missing, re-scan plugins
4. If multiple matches for capability:
   a. Filter by project type preference
   b. Prefer exact keyword matches
   c. Prefer agents over commands
5. If single match:
   a. Announce delegation with reasoning
   b. Return agent reference
6. If no match:
   a. Announce using default
   b. Return default agent reference

Capability → Phase Mapping

PhasePrimary CapabilitySecondary Capabilities
DISCUSSrequirements-gatheringbrainstorming, codebase-analysis
PLAN(internal coordinator)codebase-analysis
EXECUTEcode-implementationtdd-implementation, infrastructure
VERIFYcode-reviewrequirements-validation

Default Agents

When no plugin matches a capability:

CapabilityDefault AgentLocation
requirements-gatheringinterviewerflow-workflow:defaults/interviewer
brainstorminginterviewerflow-workflow:defaults/interviewer
codebase-analysisresearcherflow-workflow:defaults/researcher
tdd-implementationexecutorflow-workflow:defaults/executor
code-implementationexecutorflow-workflow:defaults/executor
infrastructureexecutorflow-workflow:defaults/executor
code-reviewvalidatorflow-workflow:validator
requirements-validationvalidatorflow-workflow:validator

Re-Discovery Triggers

Re-scan plugins when:

  1. /flow-workflow:start is run on uninitialized project
  2. FLOW.md capability cache is older than 24 hours
  3. Requested capability returns no match (try fresh scan)
  4. User explicitly requests refresh via /flow-workflow:status --refresh

Capability Gap Logging

When no plugin matches a capability, log in FLOW.md:

markdown
## Capability Warnings

**[TIMESTAMP]**: No plugin found for capability 'infrastructure'
- Searched: 5 plugins
- Keywords tried: infra, devops, pipeline, deploy
- Using: flow-workflow:defaults/executor
- Suggestion: Consider installing a DevOps plugin

Integration Points

  • State Management: Cache capability map in FLOW.md
  • Coordinator Agent: Performs discovery on initialization, uses for routing
  • Smart Continuation: Uses cached capabilities for agent selection

See categories.md for detailed capability category definitions.