Serialize local model calls and skip concurrent context discovery

LM Studio and Ollama run one model on one GPU — concurrent requests
cause crashes. Two fixes:

1. Per-upstream semaphore (concurrency=1) in _route_agent_chat for
   lm-studio/ollama providers. All agent-routed calls to the same
   base URL queue instead of hitting the GPU simultaneously.

2. skip_discovery=True when routing to a local model. Context discovery
   would fire a second LM Studio call alongside the main inference.
   Novel words are still registered in SOAS (low saliency) but the
   LLM confirmation step waits. Configure write_model_id or a separate
   agent model pointing at a cloud/remote model to re-enable live
   context discovery.

3. _LLM_CONCURRENCY 2 → 1 in write_queue for the same reason.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

plugins/festinger/festinger/main.py

@@ -669,6 +669,25 @@ def _openai_sse_from_response(raw: dict) -> bytes:
     return "".join(parts).encode()
 
 
+# ---------------------------------------------------------------------------
+# Per-upstream concurrency control for local models
+# ---------------------------------------------------------------------------
+
+# Local inference servers (LM Studio, Ollama) typically run a single model
+# on one GPU and queue or crash under concurrent requests.  We serialize all
+# agent-routed calls that share the same base URL through a semaphore.
+# Cloud providers (claude, openai) are excluded — they handle concurrency fine.
+_local_upstream_semaphores: dict[str, asyncio.Semaphore] = {}
+
+
+def _get_upstream_semaphore(base_url: str) -> asyncio.Semaphore:
+    """Return (creating if needed) a semaphore that limits concurrency to 1 for this upstream."""
+    key = base_url.rstrip("/").lower()
+    if key not in _local_upstream_semaphores:
+        _local_upstream_semaphores[key] = asyncio.Semaphore(1)
+    return _local_upstream_semaphores[key]
+
+
 async def _get_agent_routing_model(pool, agent_name: str) -> ModelConfig | None:
     """
     Look up the agent's configured model from the agent_models table.
@@ -711,10 +730,19 @@ async def _route_agent_chat(
 
     claude   → Anthropic Messages API (translated in both directions)
     openai   → OpenAI-compatible endpoint (base_url + model swap)
-    lm-studio→ same as openai
+    lm-studio→ same as openai, serialized through a per-upstream semaphore
     """
-    # Run recollection injection (same pipeline as standard path)
+    is_local = agent_model.provider in ("lm-studio", "ollama")
-    body = await process_prompt(body, "/v1/chat/completions", pool, cfg, request_headers)
+
+    # Local single-GPU models can't handle concurrent requests.
+    # Skip context discovery so we don't fire a second LM Studio call
+    # at the same time as the main inference.
+    # (Configure write_model_id or a per-agent model for a cloud/separate
+    # model if you want memory building alongside local inference.)
+    body = await process_prompt(
+        body, "/v1/chat/completions", pool, cfg, request_headers,
+        skip_discovery=is_local,
+    )
 
     sess = session_key(agent_model.model_name, body.get("messages", []))
 
@@ -738,6 +766,8 @@ async def _route_agent_chat(
             "authorization": f"Bearer {agent_model.api_key or 'lm-studio'}",
             "content-type": "application/json",
         }
+        # One request at a time to local inference servers.
+        _local_sem = _get_upstream_semaphore(oai_upstream) if is_local else None
 
     async def _call(current_body: dict) -> tuple[str, dict]:
         """Call the agent's upstream; always return (text, openai_format_dict)."""
@@ -748,7 +778,11 @@ async def _route_agent_chat(
         else:
             b = dict(current_body)
             b["model"] = agent_model.model_name
-            text, raw_o = await call_openai(b, oai_upstream, oai_headers)
+            if _local_sem is not None:
+                async with _local_sem:
+                    text, raw_o = await call_openai(b, oai_upstream, oai_headers)
+            else:
+                text, raw_o = await call_openai(b, oai_upstream, oai_headers)
             return text, raw_o
 
     text, raw = await _call(body)
@@ -806,10 +840,22 @@ def _extract_agent_name(body: dict, headers: dict) -> tuple[str, dict]:
     return identity, clean_body
 
 
-async def process_prompt(body: dict, path: str, pool, cfg: dict, request_headers: dict | None = None) -> dict:
+async def process_prompt(
+    body: dict,
+    path: str,
+    pool,
+    cfg: dict,
+    request_headers: dict | None = None,
+    skip_discovery: bool = False,
+) -> dict:
     """
     Run the saliency + recollection pipeline over the prompt.
     Returns a (possibly modified) body dict with the recollection block injected.
+
+    skip_discovery: when True, novel-word candidates are still registered in
+    SOAS (low saliency) but the LLM context-discover call is suppressed.
+    Used when the agent is routed to a local single-GPU model that cannot
+    handle simultaneous requests.
     """
     read_threshold = float(await get_config(pool, "saliency_read_threshold", "0.5"))
     conf_floor = float(await get_config(pool, "recollection_confidence_floor", "0.6"))
@@ -880,7 +926,7 @@ async def process_prompt(body: dict, path: str, pool, cfg: dict, request_headers
         # Do NOT add to salient_for_read — no zero-hit recollection until confirmed.
 
     # 3. Enqueue for LLM-driven discovery if there are candidates to evaluate.
-    if novel_candidates and len(user_text) >= 20:
+    if novel_candidates and len(user_text) >= 20 and not skip_discovery:
         await enqueue_context_discover(
             user_text, novel_candidates,
             agent_name=agent_name,

plugins/festinger/festinger/write_queue.py

@@ -58,7 +58,9 @@ _running: bool = False
 
 # Limit concurrent LM Studio / cloud-LLM calls so we don't flood a slow
 # local model server.  Cue processing (pure DB ops) is never rate-limited.
-_LLM_CONCURRENCY = 2
+# Local models (LM Studio, Ollama) can typically only handle one request at a
+# time.  Keep this at 1 to avoid overwhelming a single-GPU inference server.
+_LLM_CONCURRENCY = 1
 _llm_semaphore: asyncio.Semaphore | None = None