Add MCP skill discovery and on-demand loading

[RaghavSood]

Summary

• Integrates MCP resources protocol (resources/list, resources/read) to discover skill guides from the MCP server
• Skills are listed in the system prompt so the LLM knows what's available, but content is only loaded on-demand via the new get_skill tool
• Skill list and content are cached with the same TTL as MCP tools, with background refresh on stale cache

Stacked on #3 → #2.

Architecture

Startup:
  main.go → mcpClient.ListSkills() → cache skill metadata

Per-request (ProcessMessage):
  System prompt += SkillSummary()     → "## Available Skills\n- **dca-setup**: ...\n- **swap-tokens**: ..."
  Tools list  += GetSkillTool         → only if skills exist

On-demand (decision loop):
  LLM calls get_skill(slug) → executor → mcpClient.ReadSkill(slug)
                                           ↓
                                       resources/read → "skills/{slug}.md"
                                           ↓
                                       markdown content returned to LLM as tool result

Key design decision: The skill list (names + descriptions) is always in the system prompt so the LLM can decide relevance. The skill content (full markdown) is only loaded when the LLM calls get_skill. This keeps the context window lean as the skill library grows.

Changes

File	Change
internal/mcp/client.go	Add MCP resource types, skillCache, ListSkills, ReadSkill, SkillSummary
internal/service/agent/agent.go	Extend MCPToolProvider interface, inject skill summary into prompt, register GetSkillTool
internal/service/agent/tools.go	Add GetSkillTool definition
internal/service/agent/executor.go	Add get_skill dispatch + execGetSkill handler
cmd/server/main.go	Pre-warm skill cache at startup

MCP Protocol Details

Skill discovery uses standard MCP resources protocol:

• resources/list → filter for URIs matching skills/*.md
• resources/read with URI skills/{slug}.md → skill markdown content

Caching:

• Skill list: TTL-based cache (same as tools), background refresh when stale
• Skill content: cached in sync.Map for process lifetime (docs change with deployments)

Test plan

• Build passes: go build ./...
• Startup with MCP server that has skills: verify mcp skills loaded log
• Startup with MCP server without resources support: verify graceful degradation
• Startup without MCP: verify no skill-related logs or errors
• Send message that matches a skill: verify LLM calls get_skill and uses the content
• Send message that doesn't match any skill: verify LLM does NOT call get_skill
• Verify skill content is cached (second load is instant, no resources/read call)

🤖 Generated with Claude Code