fix(chat): stabilize system prompt, sequence memory extraction, and send stable session id to preserve KV cache (#3360)

* fix(chat): stabilize system prompt, sequence memory extraction, send stable session id to preserve KV cache

Fixes #2927. As diagnosed in the issue, three things in Odysseus's request
pattern actively destroyed local backends' (llama.cpp / LM Studio) KV-cache
continuity, forcing a full prompt re-evaluation (15-30s+) on every turn:

1. Dynamic content folded into the system prompt every turn. Both the chat
   preface (ChatProcessor.build_context_preface) and the agent system prompt
   (_build_system_prompt) injected current_datetime_prompt() — text that
   changes every minute — directly into system-role messages, which llm_core
   then concatenates into the single system message sent as the cached
   prefix. Any byte difference there invalidates the entire cache. Moved this
   to a new current_datetime_context_message() helper that returns a
   standalone user-role message, inserted near the end of the array (right
   before the latest user turn) instead of mixed into the system prompt. The
   static system prefix (preset prompt + safety policy + agent base prompt)
   now stays byte-identical across turns of the same session.

2. Memory/skill extraction side-requests competed with the main completion.
   run_post_response_tasks fired extract_and_store / maybe_extract_skill via
   asyncio.create_task — fire-and-forget coroutines that could overlap the
   next turn's main request and steal llama.cpp's limited processing slots,
   evicting the cached checkpoint. They're now queued through a new
   _run_extraction_jobs_sequentially helper that waits for the session's
   stream to go idle and runs the jobs strictly one at a time.

3. No stable session identifier was sent to local backends, so llama.cpp
   assigned a new processing slot via LRU every turn ("session_id=<empty>
   server-selected (LCP/LRU)"), losing slot affinity. Added
   _apply_local_cache_affinity() in llm_core, which sets session_id and
   cache_prompt: true on outgoing payloads — gated to self-hosted
   OpenAI-compatible endpoints only (never api.openai.com or other cloud
   providers, which reject unrecognized request fields with a 400). Threaded
   session_id through stream_llm / llm_call_async / stream_agent_loop from
   the existing Odysseus session id.

Tests in tests/test_kv_cache_invalidation_2927.py exercise the real payload-
assembly and scheduling code paths: byte-identical system prefix across two
turns of the same session (with a regression check that genuinely changed
instructions DO still change it), the dynamic time block landing as a
user-role message, extraction jobs waiting for the stream to go idle and
running sequentially, and the outgoing payload carrying a stable session_id
(same across turns of one session, different across sessions) only for
self-hosted endpoints. Updated tests/test_user_time.py for the new message
placement.

* fix(tests): accept owner= kwarg in normalize_model_id monkeypatch

The upstream normalize_model_id signature now takes an owner= keyword
argument, and chat_helpers.py passes owner=getattr(sess, "owner", None)
at the call site. Update the test stub lambda to **kwargs so it handles
the new argument without breaking, and update chat_helpers.py to forward
the owner parameter consistently.

---------

Co-authored-by: Alexandre Teixeira <111787685+alteixeira20@users.noreply.github.com>

src/llm_core.py

@@ -455,6 +455,43 @@ def _detect_provider(url: str) -> str:
     return "openai"
 
 
+def _is_self_hosted_openai_compatible(url: str) -> bool:
+    """True for custom/local OpenAI-compatible servers (llama.cpp, LM Studio,
+    vLLM, text-generation-webui, etc.) as opposed to api.openai.com itself.
+
+    Used to gate llama.cpp-server-specific payload extras (``session_id``,
+    ``cache_prompt``) — sending unrecognized top-level fields to OpenAI's
+    actual API returns a 400 ("Unrecognized request argument"), but
+    self-hosted servers generally ignore unknown fields and many (notably
+    llama.cpp's server) use them for KV-cache slot affinity (issue #2927).
+    """
+    return _detect_provider(url) == "openai" and not _host_match(url, "openai.com")
+
+
+def _apply_local_cache_affinity(payload: Dict, url: str, session_id: Optional[str]) -> None:
+    """Add llama.cpp-server slot-affinity hints to an outgoing payload, in place.
+
+    As diagnosed in issue #2927, llama.cpp assigns requests to processing
+    slots via LRU when no stable identifier is present ("session_id=<empty>
+    server-selected (LCP/LRU)"), which means consecutive turns of the same
+    chat can land on different slots and lose their cached prefix entirely.
+    Sending a stable ``session_id`` (derived from the Odysseus session) lets
+    the server keep routing the same conversation to the same slot, and
+    ``cache_prompt: true`` asks it to retain/reuse the prefix it already has.
+
+    Both fields are llama.cpp / LM Studio extensions to the OpenAI schema; we
+    only set them for self-hosted OpenAI-compatible endpoints (never
+    api.openai.com or other cloud providers, which reject unrecognized
+    top-level request fields).
+    """
+    if not session_id:
+        return
+    if not _is_self_hosted_openai_compatible(url):
+        return
+    payload.setdefault("session_id", str(session_id))
+    payload.setdefault("cache_prompt", True)
+
+
 def _provider_headers(provider: str, headers: Optional[Dict] = None) -> Dict[str, str]:
     h = {"Content-Type": "application/json"}
     if isinstance(headers, dict):
@@ -1269,7 +1306,8 @@ async def llm_call_async(
     headers: Optional[Dict] = None,
     timeout: int = LLMConfig.STREAM_TIMEOUT,
     max_retries: int = LLMConfig.MAX_RETRIES,
-    prompt_type: Optional[str] = None
+    prompt_type: Optional[str] = None,
+    session_id: Optional[str] = None,
 ) -> str:
     """Asynchronous LLM call using httpx with connection pooling, timeout, retry logic, and performance logging."""
     provider = _detect_provider(url)
@@ -1369,6 +1407,7 @@ async def llm_call_async(
         # Suppress thinking for qwen3/gemma4 on Ollama /v1 — same as stream_llm.
         if _is_ollama_openai_compat_url(url) and _supports_thinking(model):
             payload["think"] = False
+        _apply_local_cache_affinity(payload, url, session_id)
 
     if _is_host_dead(target_url):
         raise HTTPException(503, f"Upstream {_host_key(target_url)} marked unreachable (cooldown active)")
@@ -1426,7 +1465,7 @@ async def llm_call_async(
 async def stream_llm(url: str, model: str, messages: List[Dict], temperature: float = LLMConfig.DEFAULT_TEMPERATURE,
                      max_tokens: int = LLMConfig.DEFAULT_MAX_TOKENS, headers: Optional[Dict] = None,
                      timeout: int = LLMConfig.STREAM_TIMEOUT, prompt_type: Optional[str] = None,
-                     tools: Optional[List[Dict]] = None):
+                     tools: Optional[List[Dict]] = None, session_id: Optional[str] = None):
     """Stream LLM responses with improved error handling.
 
     Yields SSE chunks:
@@ -1491,6 +1530,7 @@ async def stream_llm(url: str, model: str, messages: List[Dict], temperature: fl
         # <think> blocks. Ollama /v1 accepts "think": false as a top-level param.
         if _is_ollama_openai_compat_url(url) and _supports_thinking(model):
             payload["think"] = False
+        _apply_local_cache_affinity(payload, url, session_id)
         h = _provider_headers(provider, headers)
         if provider == "copilot":
             from src.copilot import apply_request_headers