conversational: add no-tools multi-turn Brain endpoint for portal-assistant

The portal-assistant voice gateway needs a Claude that is conversational, free
(on the cluster subscription, no metered API), and safe to sit behind a public
edge. Add POST /v1/conversational: it drives a new no-tools `conversational`
agent with per-conversation --resume so a voice turn keeps context, and is lean
on purpose — no workspace clone, no tools, and crucially NO
--dangerously-skip-permissions (so even a leaked agent can't execute anything).
This is deliberately NOT /v1/chat/completions, which clones the git-crypt infra
repo and runs a Bash-enabled agent per turn (portal-assistant ADR-0002).

The conversational agent replies in the speaker's language (Bulgarian/English),
short and TTS-friendly. Tests cover the argv builder (new vs resume), the happy
path, multi-turn resume across calls, auth, and failure → 503. Full suite green.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app/conversational.py

@@ -0,0 +1,98 @@
+"""Conversational Brain — drives the Claude CLI for the portal-assistant gateway.
+
+A lean, no-tools, multi-turn path (portal-assistant ADR-0002): no workspace clone,
+no tool-enabled agent, and NO --dangerously-skip-permissions. Per-conversation
+continuity comes from the Claude CLI's own --session-id / --resume, so the gateway
+only has to hand us a stable session id per conversation.
+"""
+import asyncio
+import json
+import os
+from subprocess import PIPE
+
+CONVERSATIONAL_AGENT = "conversational"
+# A spoken chat turn is short; a turn that runs longer than this is wedged.
+CONVERSATIONAL_TIMEOUT_SECONDS = int(
+    os.environ.get("CONVERSATIONAL_TIMEOUT_SECONDS", "120")
+)
+
+# Session ids the Claude CLI has already opened in THIS process, so a follow-up
+# turn resumes instead of re-opening. In-memory + single-replica: a pod restart
+# clears this AND the CLI's emptyDir session state together, so they stay in sync.
+_started: set[str] = set()
+
+
+def reset_started() -> None:
+    """Forget all opened sessions (used by tests)."""
+    _started.clear()
+
+
+def conversational_argv(
+    session_id: str, message: str, model: str, resume: bool
+) -> list[str]:
+    """Build the argv for one conversational turn.
+
+    A new conversation opens the session with --session-id; subsequent turns
+    continue it with --resume so Claude keeps its own context. We never pass
+    --dangerously-skip-permissions: the conversational agent has no tools and the
+    endpoint is public-facing, so nothing may be auto-permitted.
+    """
+    argv = [
+        "claude", "-p",
+        "--agent", CONVERSATIONAL_AGENT,
+        "--output-format", "json",
+        "--model", model,
+    ]
+    argv += ["--resume", session_id] if resume else ["--session-id", session_id]
+    argv.append(message)
+    return argv
+
+
+def extract_reply(output_lines: list[str]) -> str:
+    """Pull the final assistant text out of `claude -p --output-format json`.
+
+    The CLI emits one JSON object with the final message under `result`; fall
+    back to the raw text if it isn't parseable so callers always get something.
+    """
+    raw = "".join(output_lines).strip()
+    if not raw:
+        return ""
+    try:
+        parsed = json.loads(raw)
+    except json.JSONDecodeError:
+        return raw
+    if isinstance(parsed, dict):
+        for key in ("result", "content", "text"):
+            value = parsed.get(key)
+            if isinstance(value, str) and value:
+                return value
+    return raw
+
+
+async def run_turn(session_id: str, message: str, model: str) -> dict:
+    """Run one conversational turn and return {exit_code, reply, stderr}.
+
+    Resumes the Claude session if we've opened it before; otherwise opens it.
+    The session is only marked opened on success so a failed first turn can be
+    retried cleanly as a new one.
+    """
+    resume = session_id in _started
+    argv = conversational_argv(session_id, message, model, resume)
+
+    proc = await asyncio.create_subprocess_exec(*argv, stdout=PIPE, stderr=PIPE)
+    assert proc.stdout is not None and proc.stderr is not None
+
+    output_lines: list[str] = []
+    async for line in proc.stdout:
+        output_lines.append(line.decode(errors="replace"))
+    stderr = await proc.stderr.read()
+    await proc.wait()
+
+    if proc.returncode == 0:
+        _started.add(session_id)
+
+    return {
+        "exit_code": proc.returncode,
+        "reply": extract_reply(output_lines),
+        "stderr": stderr.decode(errors="replace"),
+    }

app/main.py

@@ -13,6 +13,8 @@ from fastapi import FastAPI, HTTPException, Header
 from fastapi.responses import JSONResponse
 from pydantic import BaseModel, Field
 
+from app import conversational
+
 app = FastAPI(title="Claude Agent Service")
 
 API_TOKEN = os.environ.get("API_BEARER_TOKEN", "")
@@ -104,6 +106,15 @@ class ChatCompletionsRequest(BaseModel):
     model_config = {"extra": "allow"}
 
 
+class ConversationalRequest(BaseModel):
+    # The portal-assistant gateway owns the conversation; it hands us a stable
+    # session id (for Claude --resume) plus the next user message. Model is
+    # selectable per request, same as the OpenAI-compat path.
+    session_id: str
+    message: str
+    model: str | None = None
+
+
 def verify_token(authorization: str | None):
     # Reject everything when the service is unconfigured. compare_digest("", "")
     # returns True, so without this guard an empty API_TOKEN would happily
@@ -510,3 +521,56 @@ async def chat_completions(
             "total_tokens": 0,
         },
     }
+
+
+@app.post("/v1/conversational")
+async def conversational_turn(
+    request: ConversationalRequest,
+    authorization: str | None = Header(default=None),
+):
+    """Lean, multi-turn conversational Brain for the portal-assistant gateway.
+
+    Drives a no-tools conversational agent with per-conversation --resume — no
+    workspace clone, no tools (see portal-assistant ADR-0002). Returns the
+    assistant's reply text keyed to the caller's session id.
+    """
+    verify_token(authorization)
+
+    model = request.model if request.model is not None else DEFAULT_MODEL
+    if model not in SUPPORTED_MODELS:
+        return JSONResponse(
+            status_code=400,
+            content={"error": "unsupported model", "supported": sorted(SUPPORTED_MODELS)},
+        )
+
+    if not _reserve_queue_slot():
+        return JSONResponse(
+            status_code=503,
+            content={"error": "execution failed", "detail": "queue full"},
+        )
+
+    try:
+        async with _execution_slot():
+            result = await asyncio.wait_for(
+                conversational.run_turn(request.session_id, request.message, model),
+                timeout=conversational.CONVERSATIONAL_TIMEOUT_SECONDS,
+            )
+    except asyncio.TimeoutError:
+        return JSONResponse(
+            status_code=503,
+            content={"error": "execution failed", "detail": "agent timed out"},
+        )
+    except Exception as exc:  # noqa: BLE001
+        return JSONResponse(
+            status_code=503,
+            content={"error": "execution failed", "detail": _one_line(str(exc))},
+        )
+
+    if result["exit_code"] != 0:
+        detail = _one_line(result.get("stderr") or "") or f"exit {result['exit_code']}"
+        return JSONResponse(
+            status_code=503,
+            content={"error": "execution failed", "detail": detail},
+        )
+
+    return {"session_id": request.session_id, "reply": result["reply"]}