Feature Request: Unified agent orchestration + function calling on a single listener

[mahmoudibrahim98]

Problem

Currently, agent orchestration (type: agent) and function calling (type: prompt with prompt_targets) operate on separate listeners with separate ports. There is no way to have a single entry point that:

1. First checks if the user request matches a prompt_target (function calling → direct API call)
2. Falls back to agent orchestration if no function matches

We tried several approaches on v0.4.11:

1. use_agent_orchestrator: true in overrides

The WASM source code (stream_context.rs) has the logic to fall back from function calling to agent orchestration when use_agent_orchestrator is set. However:f

• The Python config generator (config_generator.py line 450-458) requires exactly one endpoint when use_agent_orchestrator: true, and counts agent URLs as endpoints, so any config with agents + a function calling endpoint fails
• The Rust brightstaff binary has the same validation
• The WASM filter (prompt_gateway.wasm) also validates this at runtime

2. Separate listeners on different ports

This works — we can run type: prompt on port 10000 and type: agent on port 8001 in the same container. But the caller (our gateway) must decide which port to send to, re-introducing client-side routing logic that Plano should handle.

3. Patching the config generator

We patched the Python endpoint count check (elif len(endpoints) > 1: → elif False:), which let the config generate. But the envoy template creates two listeners on the same port (one for the prompt WASM filter, one for the agent routing), causing duplicate address errors.

Desired behavior

A single listener that combines both capabilities:

listeners:
  - type: unified  # or type: prompt with agent support
    name: platform
    port: 8001
    router: plano_orchestrator_v1
    prompt_targets:
      - name: fetch_application_fields
        description: Fetch fields for a certain database
        parameters:
          - name: applicationId
            type: str
            required: true
        endpoint:
          name: server
          path: /api/v1/meta/queries/appfields
          http_method: GET
    agents:
      - id: consultant_agent
        description: Knowledge questions
      - id: version_api_creator_agent
        description: Create a Version API (multi-turn workflow)
      - id: general_chat_agent
        description: General conversation
        default: true

Flow:

1. User sends "Fetch CUSTOMER fields" → Arch-Function matches fetch_application_fields → direct API call → fast response
2. User sends "Explain FUNDS.TRANSFER validation rules" → Arch-Function finds no match → falls back to Plano-Orchestrator → routes to consultant_agent
3. User sends "Hello" → No function match → No agent match → routes to general_chat_agent (default)

Environment

• Plano v0.4.11 (katanemo/plano:0.4.11)
• Cloud orchestrator at archfc.katanemo.dev
• Note: v0.4.12 doesn't work with the cloud (provider_interface: plano not recognized by cloud)