feat(stt): mémoire court-terme de session côté serveur (5a)

Serveur : SessionStore (historique borné + TTL, en mémoire), /v1/ask accepte session_id
et injecte l'historique dans l'appel LLM, /v1/reset l'efface. Dockerfile en 1 worker
(cohérence mémoire process).

Client : session_id généré par run (uuid), envoyé à chaque requête ; commande /reset
en mode texte.

Testé en process (TestClient) : historique croît 0→2→4, reset→0, sessions isolées,
sans session_id = sans état.

Mémoire long-terme Qdrant (5b) + réparation Qdrant (5c) à suivre.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013FmcxGsyXZXogiAHQLjnZT

stt/server/stt_server/app.py

@@ -14,13 +14,16 @@ from pydantic import BaseModel
 from stt_server import __version__
 from stt_server.brain import ask as brain_ask
 from stt_server.config import settings
+from stt_server.memory import SessionStore
 
 app = FastAPI(title="STT-server", version=__version__)
+sessions = SessionStore()
 
 
 class AskRequest(BaseModel):
     text: str
-    model: str | None = None   # alias LiteLLM ; défaut serveur si absent
+    model: str | None = None      # alias LiteLLM ; défaut serveur si absent
+    session_id: str | None = None  # mémoire court-terme : fil de conversation
 
 
 class AskReply(BaseModel):
@@ -38,6 +41,13 @@ async def v1_models() -> dict:
     return {"default": settings.model, "available": settings.allowed_models}
 
 
+@app.post("/v1/reset")
+async def v1_reset(req: AskRequest) -> dict:
+    if req.session_id:
+        sessions.reset(req.session_id)
+    return {"status": "reset"}
+
+
 @app.post("/v1/ask", response_model=AskReply)
 async def v1_ask(req: AskRequest) -> AskReply:
     text = req.text.strip()
@@ -49,10 +59,14 @@ async def v1_ask(req: AskRequest) -> AskReply:
             status_code=400,
             detail=f"modèle '{model}' non autorisé ; dispo : {settings.allowed_models}",
         )
+    history = sessions.history(req.session_id) if req.session_id else None
     try:
-        reply = await brain_ask(text, model)
+        reply = await brain_ask(text, model, history)
     except httpx.HTTPError as e:
         raise HTTPException(status_code=502, detail=f"upstream LiteLLM : {e}") from e
+    if req.session_id:
+        sessions.add(req.session_id, "user", text)
+        sessions.add(req.session_id, "assistant", reply)
     return AskReply(reply=reply, model=model)
 
 

stt/server/stt_server/brain.py

@@ -12,13 +12,14 @@ import httpx
 from stt_server.config import settings
 
 
-async def ask(text: str, model: str | None = None) -> str:
+async def ask(text: str, model: str | None = None, history: list[dict] | None = None) -> str:
+    messages = [{"role": "system", "content": settings.system_prompt}]
+    if history:
+        messages += history
+    messages.append({"role": "user", "content": text})
     payload = {
         "model": model or settings.model,
-        "messages": [
-            {"role": "system", "content": settings.system_prompt},
-            {"role": "user", "content": text},
-        ],
+        "messages": messages,
         "max_tokens": settings.max_tokens,
         "temperature": settings.temperature,
     }

stt/server/stt_server/memory.py

@@ -0,0 +1,38 @@
+"""Mémoire court-terme côté serveur : historique de conversation par session.
+
+En mémoire (process), borné et avec TTL. Suffisant pour un usage mono-utilisateur ;
+le Deployment tourne en 1 worker pour que les sessions soient cohérentes.
+La mémoire long-terme sémantique (Qdrant) viendra en phase 5b.
+"""
+
+from __future__ import annotations
+
+import time
+from collections import deque
+
+
+class SessionStore:
+    def __init__(self, max_turns: int = 10, ttl_sec: int = 3600):
+        self.max_turns = max_turns          # paires user/assistant conservées
+        self.ttl = ttl_sec
+        self._data: dict[str, dict] = {}     # sid -> {"turns": deque, "ts": float}
+
+    def _gc(self) -> None:
+        now = time.time()
+        for sid in [s for s, v in self._data.items() if now - v["ts"] > self.ttl]:
+            del self._data[sid]
+
+    def history(self, sid: str) -> list[dict]:
+        self._gc()
+        sess = self._data.get(sid)
+        return list(sess["turns"]) if sess else []
+
+    def add(self, sid: str, role: str, content: str) -> None:
+        sess = self._data.setdefault(
+            sid, {"turns": deque(maxlen=self.max_turns * 2), "ts": time.time()}
+        )
+        sess["turns"].append({"role": role, "content": content})
+        sess["ts"] = time.time()
+
+    def reset(self, sid: str) -> None:
+        self._data.pop(sid, None)