feat(models): persona models same on chat + Telegram, no_think, one resident

codec_bridges.py

@@ -113,7 +113,21 @@ def call_llm(channel, text, llm_cfg, conversation_history=None,
     bridge degradation). `chat_template_kwargs` is filtered out of
     `llm_cfg["kwargs"]` so codec_llm's enable_thinking=False is preserved."""
     import codec_llm
-    if system_prompt_override:
+    # A persona model (Fréd / M Corpus / Mwen / the 8B reformat voice — config
+    # extra_models[].system_prompt*) owns the turn on every surface, so when it
+    # is the active model the bridge speaks as it, not as the channel persona,
+    # and applies its sampling. This is what makes "switch the model in CODEC,
+    # then use it on Telegram" behave the same as the dashboard chat.
+    _persona = {"system_prompt": None, "sampling": {}}
+    try:
+        import codec_models
+        _persona = codec_models.model_extras(llm_cfg.get("model"))
+    except Exception:
+        pass
+
+    if _persona.get("system_prompt"):
+        sys_prompt = _persona["system_prompt"]           # persona wins over channel
+    elif system_prompt_override:
         sys_prompt = system_prompt_override
     else:
         now_str = datetime.now().strftime("%A %B %d, %Y at %H:%M")
@@ -125,6 +139,7 @@ def call_llm(channel, text, llm_cfg, conversation_history=None,
     messages.append({"role": "user", "content": text})
 
     extra = {k: v for k, v in llm_cfg["kwargs"].items() if k != "chat_template_kwargs"}
+    extra.update(_persona.get("sampling") or {})         # per-model sampling
     content = codec_llm.call(
         messages, base_url=llm_cfg["base_url"], model=llm_cfg["model"],
         api_key=llm_cfg["api_key"], max_tokens=1500, temperature=0.7,

codec_models.py

@@ -211,7 +211,7 @@ def model_extras(model_id: Optional[str] = None,
     """
     cfg = config if config is not None else _load_config()
     mid = model_id or get_active(cfg)
-    out: Dict[str, Any] = {"system_prompt": None, "sampling": {}}
+    out: Dict[str, Any] = {"system_prompt": None, "sampling": {}, "no_think": False}
     for em in (cfg.get("extra_models") or []):
         if not isinstance(em, dict) or em.get("id") != mid:
             continue
@@ -229,6 +229,10 @@ def model_extras(model_id: Optional[str] = None,
         if isinstance(sampling, dict):
             out["sampling"] = {k: v for k, v in sampling.items()
                                if isinstance(k, str) and isinstance(v, (int, float, str, bool))}
+        # A reformat/persona fine-tune that should answer directly (no <think>
+        # block) declares `"no_think": true`. Chat forces enable_thinking=False;
+        # the bridge path is already thinking-off via codec_llm.
+        out["no_think"] = bool(em.get("no_think", False))
         break
     return out
 

routes/chat.py

@@ -1147,6 +1147,15 @@ async def _skill_stream():
         # Dashboard chat & Vibe benefit from thinking mode (deeper answers).
         # Frontend can send thinking=false to override for speed.
         thinking = body.get("thinking", True)
+        # A reformat/persona model declaring no_think answers directly — force
+        # thinking off regardless of the Think toggle (a <think> block would eat
+        # the budget and delay the rewrite).
+        try:
+            import codec_models as _cm_nt
+            if _cm_nt.model_extras(model).get("no_think"):
+                thinking = False
+        except Exception as _e:
+            log.debug("no_think lookup failed: %s", _e)
         # Train-of-thought reveal: when the frontend's Thoughts toggle is ON,
         # also stream the model's <think> reasoning as separate SSE `think`
         # events. Off by default → no think frames → identical to before.

tests/test_model_switching.py

@@ -400,8 +400,21 @@ def test_model_extras_reads_persona_and_sampling(cfg, tmp_path):
     x = codec_models.model_extras()
     assert x["system_prompt"].startswith("You are Fred.")
     assert x["sampling"] == {"repetition_penalty": 1.2, "repetition_context_size": 512}
+    assert x["no_think"] is False           # not declared → defaults off
     # a model with no entry gets no override
-    assert codec_models.model_extras("mlx-community/A") == {"system_prompt": None, "sampling": {}}
+    assert codec_models.model_extras("mlx-community/A") == {
+        "system_prompt": None, "sampling": {}, "no_think": False}
+
+
+def test_model_extras_reads_no_think_flag(cfg, tmp_path):
+    """A reformat model declares no_think:true so chat forces thinking off."""
+    import json as _json
+    data = _json.loads(cfg.read_text())
+    data["llm_model"] = "/abs/reformat-8b"
+    data["extra_models"] = [{"id": "/abs/reformat-8b", "label": "Reformat",
+                             "system_prompt": "Rewrite it.", "no_think": True}]
+    cfg.write_text(_json.dumps(data))
+    assert codec_models.model_extras()["no_think"] is True
 
 
 def test_model_extras_survives_missing_prompt_file(cfg):