rpa_vision_v3/mcp_rpa_vision.py

#!/usr/bin/env python3
# /// script
# requires-python = ">=3.10"
# dependencies = ["mcp[cli]", "httpx"]
# ///
"""
MCP Server — RPA Vision V3
Expose les endpoints de Léa comme outils sémantiques pour Claude Code.

Usage (claude mcp add) :
  uv run /home/dom/ai/rpa_vision_v3/mcp_rpa_vision.py

Env vars :
  RPA_API_TOKEN   — token Bearer pour le streaming server (5005)
  RPA_STREAM_URL  — défaut http://localhost:5005
  RPA_VWB_URL     — défaut http://localhost:5002
  OLLAMA_BASE_URL — défaut http://localhost:11434
"""

import json
import os

import httpx
from mcp.server.fastmcp import FastMCP

# --- Config ---
STREAM = os.getenv("RPA_STREAM_URL", "http://localhost:5005")
VWB = os.getenv("RPA_VWB_URL", "http://localhost:5002")
OLLAMA = os.getenv("OLLAMA_BASE_URL", "http://localhost:11434")
TOKEN = os.getenv("RPA_API_TOKEN", "")

mcp = FastMCP("rpa-vision")


def _hdrs():
    return {"Authorization": f"Bearer {TOKEN}"}


def _get(url: str, **params) -> dict:
    with httpx.Client(timeout=30) as c:
        r = c.get(url, headers=_hdrs(), params=params)
        r.raise_for_status()
        return r.json()


def _post(url: str, data: dict | None = None) -> dict:
    with httpx.Client(timeout=60) as c:
        r = c.post(url, headers=_hdrs(), json=data or {})
        r.raise_for_status()
        return r.json()


def _vwb(url: str) -> dict:
    with httpx.Client(timeout=15) as c:
        r = c.get(url)
        r.raise_for_status()
        return r.json()


def _j(obj) -> str:
    return json.dumps(obj, ensure_ascii=False, indent=2)


# ── Workflows VWB ──────────────────────────────────────────────────────────────

@mcp.tool()
def list_workflows() -> str:
    """Liste tous les workflows VWB : id, nom, nombre d'étapes."""
    data = _vwb(f"{VWB}/api/workflows/")
    wfs = data if isinstance(data, list) else data.get("workflows", [])
    return _j([
        {"id": w.get("id"), "name": w.get("name"), "steps": len(w.get("steps", []))}
        for w in wfs
    ])


@mcp.tool()
def get_workflow(workflow_id: str) -> str:
    """Retourne le détail complet d'un workflow VWB (étapes, actions, paramètres)."""
    return _j(_vwb(f"{VWB}/api/workflows/{workflow_id}"))


# ── Sessions streaming ─────────────────────────────────────────────────────────

@mcp.tool()
def list_sessions() -> str:
    """Liste les sessions streaming actives (session_id, machine, dernière activité)."""
    return _j(_get(f"{STREAM}/api/v1/traces/stream/sessions"))


@mcp.tool()
def get_session(session_id: str) -> str:
    """Retourne les détails d'une session : état, heartbeat, actions récentes."""
    return _j(_get(f"{STREAM}/api/v1/traces/stream/session/{session_id}"))


@mcp.tool()
def get_streaming_stats() -> str:
    """Statistiques du serveur streaming : sessions actives, replays en cours, charge."""
    return _j(_get(f"{STREAM}/api/v1/traces/stream/stats"))


# ── Replays ────────────────────────────────────────────────────────────────────

@mcp.tool()
def list_replays() -> str:
    """Liste les replays actifs ou récents avec leur état (running/paused/done/error)."""
    return _j(_get(f"{STREAM}/api/v1/traces/stream/replays"))


@mcp.tool()
def get_replay_state(replay_id: str) -> str:
    """Retourne l'état complet d'un replay : statut, action courante, variables runtime, progression."""
    return _j(_get(f"{STREAM}/api/v1/traces/stream/replay/{replay_id}"))


@mcp.tool()
def start_replay(workflow_id: str, session_id: str) -> str:
    """Lance un replay d'un workflow sur une session active. Retourne le replay_id."""
    return _j(_post(f"{STREAM}/api/v1/traces/stream/replay", {
        "workflow_id": workflow_id,
        "session_id": session_id,
    }))


@mcp.tool()
def resume_replay(replay_id: str) -> str:
    """Reprend un replay en état paused_need_help (équivalent bouton Continuer de Léa)."""
    return _j(_post(f"{STREAM}/api/v1/traces/stream/replay/{replay_id}/resume"))


@mcp.tool()
def list_stream_workflows() -> str:
    """Liste les workflows compilés côté streaming server (peut différer du VWB)."""
    return _j(_get(f"{STREAM}/api/v1/traces/stream/workflows"))


# ── Ollama ─────────────────────────────────────────────────────────────────────

@mcp.tool()
def ollama_list_models() -> str:
    """Liste les modèles Ollama disponibles sur cette machine."""
    with httpx.Client(timeout=10) as c:
        r = c.get(f"{OLLAMA}/api/tags")
        r.raise_for_status()
    return _j([m["name"] for m in r.json().get("models", [])])


@mcp.tool()
def ollama_query(prompt: str, model: str = "qwen2.5:7b") -> str:
    """Envoie un prompt à un modèle Ollama local. Utile pour tester t2a_decision ou debug LLM."""
    with httpx.Client(timeout=120) as c:
        r = c.post(f"{OLLAMA}/api/generate", json={
            "model": model,
            "prompt": prompt,
            "stream": False,
        })
        r.raise_for_status()
    return r.json().get("response", "")


@mcp.tool()
def ollama_chat(messages: list, model: str = "qwen2.5:7b") -> str:
    """Chat avec un modèle Ollama. messages = [{"role": "user", "content": "..."}]"""
    with httpx.Client(timeout=120) as c:
        r = c.post(f"{OLLAMA}/api/chat", json={
            "model": model,
            "messages": messages,
            "stream": False,
        })
        r.raise_for_status()
    return r.json().get("message", {}).get("content", "")


if __name__ == "__main__":
    mcp.run()