| name | message-routing |
| description | Use when classifying user messages at process gates. Carried by orchestrator agents to determine how to handle user responses. |
Message Routing
Classify user messages at gates to determine the correct response. Used by orchestrator agents during supervised mode when a gate pauses for user input.
When to Use
- A phase has completed and the orchestrator is presenting output at a gate
- The user has responded to a gate prompt
- You need to determine: does this response mean "proceed" or something else?
Classifications
1. Approval
Signals: "looks good", "continue", "approved", "go", "next", "yes", "lgtm", thumbs up, or any clear indication of satisfaction.
Behavior: Proceed to the next phase. No further action needed at this gate.
2. Feedback
Signals: "this is too promotional", "the tone is wrong", "needs more sources", or any critique of the output quality.
Behavior:
- Fix in session: route the feedback to the appropriate agent to revise the output. Re-present the gate after revision.
- Queue for permanent improvement: if feedback is enabled, write a structured signal to the workspace feedback inbox (
.pas/workspace/{process}/{slug}/feedback/). Use the self-evaluation signal format:
- PPU if the feedback expresses a persistent preference
- OQI if the feedback identifies an output quality issue
- Include
Target: pointing to the skill or agent that produced the output
3. Question
Signals: "why this angle?", "what sources did you use?", "can you explain...", or any request for information about the output.
Behavior: Answer the question using context from the phase output and agent work. Then re-present the gate. Do not proceed until the user gives an explicit approval or instruction.
4. Instruction
Signals: "also include the SEC filing", "add a quote from...", "change the headline to...", or any directive to modify or extend the output.
Behavior: Incorporate the instruction into the current phase output. Route to the appropriate agent if needed. Re-present the gate after the change is made.
Ambiguous Messages
If the user's response does not clearly fit one classification:
- Default to Question (safest: ask for clarification rather than proceeding incorrectly)
- Ask: "I want to make sure I understand. Are you asking me to [interpretation], or would you like to proceed?"
- Never assume approval from an ambiguous message
Multiple Classifications
A single message can contain multiple classifications. Handle them in order:
- Answer any questions first
- Incorporate any instructions
- Fix any feedback
- Only then check for approval
Example: "Why did you use that source? Also add the SEC filing. Otherwise looks good." = Question + Instruction + Approval. Answer the question, add the SEC filing, then proceed.
