sc-mcp（SC MCP 搜索）

MCP Orchestration Skill

Central orchestration hub for PAL MCP and Rube MCP capabilities. Use this skill for complex workflows requiring multi-model reasoning, external service integration, or both.

Quick Start

# PAL-powered analysis
/sc:mcp analyze --pal consensus --question "Should we use microservices?"

# Rube-powered automation
/sc:mcp automate --rube --apps slack,github --workflow "notify on PR"

# Combined orchestration
/sc:mcp orchestrate --pal thinkdeep --rube --full-validation

PAL MCP Integration

Available Tools

PAL Workflows

Consensus Decision Making

Use consensus for:
- Architectural decisions (2-3 models)
- Security validations (security-focused models)
- Technology choices (diverse perspectives)
- Complex trade-off analysis

Recommended model combinations:

• Architectural: gpt-5.2 (for), gemini-3-pro (against), deepseek (neutral)
• Security: gpt-5.2 (security focus), gemini-3-pro (attack surface)
• Performance: gpt-5.2 (optimization), deepseek (efficiency)

Debug Investigation

Use debug for:
- Complex bugs with unclear causes
- Performance issues
- Race conditions
- Memory leaks
- Integration problems

Debug confidence levels: exploring -> low -> medium -> high -> very_high -> almost_certain -> certain

Code Review

Use codereview for:
- Pre-merge validation
- Security audits
- Performance reviews
- Architecture compliance

Review types: full, security, performance, quick

Rube MCP Integration

Available Tools

Rube Workflows

External Integration Flow

1. SEARCH_TOOLS - Find relevant tools for use case
2. GET_SCHEMAS - Get input requirements (if schemaRef returned)
3. MANAGE_CONNECTIONS - Verify/create auth
4. MULTI_EXECUTE - Execute tools
5. CREATE_RECIPE - Save for reuse (optional)

Bulk Processing Flow

1. SEARCH_TOOLS - Find data source/destination tools
2. REMOTE_WORKBENCH - Process with Python helpers:
   - run_composio_tool() - Execute Composio tools
   - invoke_llm() - AI processing
   - upload_local_file() - Export results
   - proxy_execute() - Direct API calls

Supported Apps (500+)

Communication: Slack, Discord, Teams, Gmail, Outlook, WhatsApp, Telegram Development: GitHub, GitLab, Jira, Linear, Asana, Vercel Productivity: Google Workspace, Notion, Airtable, Trello Data: Snowflake, BigQuery, Datadog, Amplitude AI: OpenAI, Anthropic, Replicate

Combined Orchestration Patterns

Pattern 1: Research + Decide + Execute

1. PAL thinkdeep - Investigate problem deeply
2. PAL consensus - Get multi-model decision
3. Rube SEARCH_TOOLS - Find execution tools
4. Rube MULTI_EXECUTE - Implement decision

Pattern 2: Review + Validate + Notify

1. PAL codereview - Review code changes
2. PAL precommit - Validate git changes
3. Rube MULTI_EXECUTE - Send notifications (Slack, email)
4. Rube CREATE_RECIPE - Save for CI/CD

Pattern 3: Debug + Fix + Verify

1. PAL debug - Root cause analysis
2. Implement fix locally
3. PAL codereview - Validate fix
4. Rube MULTI_EXECUTE - Update tickets, notify team

Pattern 4: Plan + Consensus + Automate

1. PAL planner - Create implementation plan
2. PAL consensus - Validate approach with multiple models
3. Rube MULTI_EXECUTE - Execute across apps
4. Rube MULTI_EXECUTE - Execute across apps
5. Rube CREATE_RECIPE - Save as reusable workflow

Flags

Behavioral Flow

1. Analyze - Understand what MCP capabilities are needed
2. Discover - Use RUBE_SEARCH_TOOLS for external needs, listmodels for PAL
3. Plan - Create execution plan (PAL planner or RUBE_CREATE_PLAN)
4. Validate - Use consensus for critical decisions
5. Execute - Run PAL analysis and/or Rube tools
6. Persist - Save recipes, store memory for continuity
7. Report - Present findings with tool attribution

Memory & State Management

PAL Continuation

Use continuation_id to maintain context across PAL tool calls:

# First call returns continuation_id
result = mcp__pal__thinkdeep(...)
continuation_id = result["continuation_id"]

# Subsequent calls reuse it
result = mcp__pal__thinkdeep(..., continuation_id=continuation_id)

Rube Session & Memory

Use session_id and memory for Rube continuity:

# First search generates session_id
result = mcp__rube__RUBE_SEARCH_TOOLS(..., session={"generate_id": True})
session_id = result["session_id"]

# Subsequent calls reuse session and build memory
result = mcp__rube__RUBE_MULTI_EXECUTE_TOOL(
    ...,
    session_id=session_id,
    memory={"slack": ["Channel general is C123"]}
)

Examples

Multi-Model Architecture Review

/sc:mcp analyze --pal consensus --models "gpt-5.2,gemini-3-pro,deepseek" \
  --question "Is event sourcing appropriate for this use case?"

Automated PR Workflow

/sc:mcp automate --rube --apps github,slack \
  --workflow "On PR merge, post summary to #releases"
  --save-recipe --schedule "0 9 * * 1-5"

Full Investigation Pipeline

/sc:mcp orchestrate --pal debug --rube \
  --issue "Memory leak in production" \
  --notify slack,jira --full-validation

Guardrails

• Always search tools before executing unknown integrations
• Use consensus for decisions with >$1000 impact
• Validate schemas before multi-execute
• Store memory for frequently used IDs
• Check connection status before automation
• Use thinking_mode=high for complex PAL analysis

Error Handling

Resources

• PAL MCP: codereview, debug, consensus, thinkdeep, precommit, planner, chat, challenge, apilookup
• Rube MCP: 500+ app integrations via Composio
• Trait: mcp-pal-enabled - Apply PAL to any agent
• Trait: mcp-rube-enabled - Apply Rube to any agent