Title generation: pin to Llama-3.1-8B-Instruct via Cerebras, plain-text output (#53)

* Hardcode title generation to gpt-oss-20b via Groq, strip markdown

The chat tab headline renders the returned string as plain text, so any
markdown characters in the LLM's response render as literal garbage.
Pins the title endpoint to openai/gpt-oss-20b:groq on the HF router
(bypassing the shared resolver, which routes openai/ prefixes to the
native OpenAI API), tightens the system prompt to forbid markdown, and
strips common formatting characters on the way out as a safety net.

* Fix title gen: double-prefix model id, enable reasoning, widen budget

Verified end-to-end against the HF router:

- LiteLLM strips its provider prefix before dispatch, so passing
'openai/gpt-oss-20b:groq' sent 'gpt-oss-20b:groq' to the router and
got back 'model does not exist'. Double-prefixing to
'openai/openai/gpt-oss-20b:groq' leaves the correct HF id on the wire.
- gpt-oss-20b is a reasoning model; with max_tokens=20 it used the
whole budget on reasoning and returned empty content. Setting
reasoning_effort=low (~5 reasoning tokens, ~310 ms) and widening
max_tokens to 60 yields clean titles reliably.

* Switch title gen from gpt-oss-20b to Llama-3.1-8B-Instruct via Cerebras

Non-reasoning model, so drop the reasoning_effort knob and shrink
max_tokens back to 20 (6-word title fits easily). Verified end-to-end
against the HF router: ~400 ms, clean titles, no empty-content
surprises.

backend/routes/agent.py

@@ -157,39 +157,55 @@ async def set_model(body: dict, user: dict = Depends(get_current_user)) -> dict:
     return {"model": model_id}
 
 
+_TITLE_STRIP_CHARS = str.maketrans("", "", "`*_~#[]()")
+
+
 @router.post("/title")
 async def generate_title(
     request: SubmitRequest, user: dict = Depends(get_current_user)
 ) -> dict:
-    """Generate a short title for a chat session based on the first user message."""
-    model = session_manager.config.model_name
-    llm_params = _resolve_llm_params(model, reasoning_effort="high")
+    """Generate a short title for a chat session based on the first user message.
+
+    Always uses Llama-3.1-8B-Instruct via Cerebras on the HF router. The tab
+    headline renders as plain text, so the model is told to avoid markdown
+    and any stray formatting characters are stripped before returning.
+    """
+    api_key = (
+        os.environ.get("INFERENCE_TOKEN")
+        or (user.get("hf_token") if isinstance(user, dict) else None)
+        or os.environ.get("HF_TOKEN")
+    )
     try:
         response = await acompletion(
+            # Double openai/ prefix: LiteLLM strips the first as its provider
+            # prefix, leaving the HF model id on the wire for the router.
+            model="openai/meta-llama/Llama-3.1-8B-Instruct:cerebras",
+            api_base="https://router.huggingface.co/v1",
+            api_key=api_key,
             messages=[
                 {
                     "role": "system",
                     "content": (
                         "Generate a very short title (max 6 words) for a chat conversation "
                         "that starts with the following user message. "
-                        "Reply with ONLY the title, no quotes, no punctuation at the end."
+                        "Reply with ONLY the title in plain text. "
+                        "Do NOT use markdown, backticks, asterisks, quotes, brackets, or any "
+                        "formatting characters. No punctuation at the end."
                     ),
                 },
                 {"role": "user", "content": request.text[:500]},
             ],
             max_tokens=20,
             temperature=0.3,
-            timeout=8,
-            **llm_params,
+            timeout=10,
         )
         title = response.choices[0].message.content.strip().strip('"').strip("'")
-        # Safety: cap at 50 chars
+        title = title.translate(_TITLE_STRIP_CHARS).strip()
         if len(title) > 50:
             title = title[:50].rstrip() + "…"
         return {"title": title}
     except Exception as e:
         logger.warning(f"Title generation failed: {e}")
-        # Fallback: truncate the message
         fallback = request.text.strip()
         title = fallback[:40].rstrip() + "…" if len(fallback) > 40 else fallback
         return {"title": title}