Natural Language Intent Routing

Comienza revisando los workflows existentes del dispatcher y su gestión de los slash commands; después, examina el paso de clasificación propuesto mediante run-gemini-cli. Define cómo se enrutan las solicitudes de review, triage, invoke y las ambiguas, y verifica que los ejemplos en lenguaje natural funcionen mientras los slash commands existentes sigan gestionándose adecuadamente.

Problem

The current dispatcher pattern relies on slash commands (@gemini-cli /review, @gemini-cli /triage) limits the natural interaction experience:

• Rigid syntax: Users must remember exact slash commands (/review, /triage)
• Poor discoverability: No way to know what commands are available except by reading workflows
• Unnatural interaction: Doesn't leverage Gemini's natural language understanding capabilities
• Limited flexibility: Can't handle variations or synonyms (e.g., "please review this" vs "/review")
• User friction: Have to context-switch to remember command syntax

Solution

Intent routing involves:

• Parse natural language requests instead of requiring slash commands
• Automatically detect intent from user requests
• Route to appropriate workflows based on semantic understanding
• Provide fallback handling for ambiguous requests

Here are some examples:

• Review intent: "review", "analyze", "check", "examine", "audit"
• Triage intent: "categorize", "label", "classify", "organize", "tag"
• Implement intent: "fix", "update", "change", "modify", "improve"
• General query intent: "explain", "help", "what", "how", "why"

Example

This is an example routing step:

# Example enhanced dispatcher logic
- name: 'Classify intent'
  if: |-
    ${{ !startsWith(github.event.comment.body, '@gemini-cli /') }}
  env:
    REQUEST_BODY: '${{ github.event.comment.body || github.event.review.body || github.event.issue.body }}'
  uses: 'google-github-actions/run-gemini-cli@main'
  with:
    prompt: |-
      Classify this request into one of: review, triage, invoke
      Request: "${REQUEST_BODY}"
      Return only the classification.

This would be the outcome:

❌ Before: "@gemini-cli /review focus on security"
✅ After: "@gemini-cli please review this PR for security issues"

❌ Before: "@gemini-cli /triage" 
✅ After: "@gemini-cli can you help categorize this issue?"