From dee3cf126e916d0993508db5831b93000f8b3d61 Mon Sep 17 00:00:00 2001
From: Mickael Farina <farina.mickael@gmail.com>
Date: Wed, 16 Sep 2026 17:03:02 +0200
Subject: [PATCH] =?UTF-8?q?fix(chat):=20persona=20models=20speak=20only=20?=
 =?UTF-8?q?as=20themselves=20=E2=80=94=20no=20CODEC=20prompt,=20no=20scaff?=
 =?UTF-8?q?old?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Fréd / M Corpus / Mwen (config extra_models[].system_prompt*) are personas the
user talks TO. The server built their persona prompt correctly, then glued the
frontend's generic "You are CODEC Deep Chat — a J.A.R.V.I.S.-class AI…" system
message onto it and (Think on) appended the FINAL-ANSWER reasoning scaffold. On
a 4B that produced two conflicting identities: it answered in the generic voice,
confabulated an attachment it never read, and emitted a stray <Final Answer>
tag the UI stripper (### FINAL ANSWER only) never removed.

For a persona model: REPLACE the client system message with the persona prompt
instead of concatenating, and skip _REASON_SCAFFOLD. Also strip <Final Answer>
tag variants in the UI as defense-in-depth. Operator persona (everyday model)
unchanged — it still gets the CODEC prompt + scaffold + skills.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---
 codec_chat.html |  4 ++--
 routes/chat.py  | 25 +++++++++++++++++++++++--
 2 files changed, 25 insertions(+), 4 deletions(-)

diff --git a/codec_chat.html b/codec_chat.html
index 746ff6c..cc53e37 100644
--- a/codec_chat.html
+++ b/codec_chat.html
@@ -888,7 +888,7 @@ <h1><a href="/" style="color:inherit;text-decoration:none">CODEC</a></h1>
     // (also stripping an UNCLOSED <thinking> tail — a cut/degenerate stream
     // must never dump raw reasoning into the bubble) > honest glitch notice
     // with the reasoning kept in the reveal panel.
-    var answer=(pf.answer&&pf.answer.trim())||raw.replace(/<thinking>[\s\S]*?<\/thinking>/i,'').replace(/<thinking>[\s\S]*$/i,'').replace(/###\s*FINAL ANSWER:?/i,'').trim();
+    var answer=(pf.answer&&pf.answer.trim())||raw.replace(/<thinking>[\s\S]*?<\/thinking>/i,'').replace(/<thinking>[\s\S]*$/i,'').replace(/###\s*FINAL ANSWER:?/i,'').replace(/<\/?final\s*answer>/gi,'').trim();
     if(!answer&&pf.think&&pf.think.trim()){answer='*My reasoning ran on without reaching a final answer — a local-model glitch, not a real reply. The train of thought is below; please ask again.*'}
     if(answer){
       div.remove();var md=addMessage('assistant',answer);
@@ -1703,7 +1703,7 @@ <h1><a href="/" style="color:inherit;text-decoration:none">CODEC</a></h1>
       var pf2=parseScaffold(raw);
       // Same fallback order as the primary handler: never dump an unclosed
       // <thinking> tail into the bubble; keep reasoning in the reveal panel.
-      var answer2=(pf2.answer&&pf2.answer.trim())||raw.replace(/<thinking>[\s\S]*?<\/thinking>/i,'').replace(/<thinking>[\s\S]*$/i,'').replace(/###\s*FINAL ANSWER:?/i,'').trim();
+      var answer2=(pf2.answer&&pf2.answer.trim())||raw.replace(/<thinking>[\s\S]*?<\/thinking>/i,'').replace(/<thinking>[\s\S]*$/i,'').replace(/###\s*FINAL ANSWER:?/i,'').replace(/<\/?final\s*answer>/gi,'').trim();
       if(!answer2&&pf2.think&&pf2.think.trim()){answer2='*My reasoning ran on without reaching a final answer — a local-model glitch, not a real reply. The train of thought is below; please ask again.*'}
       if(answer2){div.remove();var md2=addMessage('assistant',answer2);if(thinkingEnabled&&pf2.think&&pf2.think.trim()&&md2){var fp2=makeToT(md2,'Train of thought',false);totSet(fp2,pf2.think.trim())}chatHist.push({role:'assistant',content:answer2});saveMessages([{role:'assistant',content:answer2}])}
       else{bubble.innerHTML='<em style="color:var(--text-dim)">No response</em>'}
diff --git a/routes/chat.py b/routes/chat.py
index 9544429..91db249 100644
--- a/routes/chat.py
+++ b/routes/chat.py
@@ -1110,16 +1110,37 @@ async def _skill_stream():
                 f"Treat it as the current working directory for this conversation."
             )
 
+        # A persona model (Fréd, M Corpus, Mwen — config extra_models[].system_prompt*)
+        # OWNS the turn: it must speak only as itself. The frontend still ships its
+        # generic "You are CODEC Deep Chat — a J.A.R.V.I.S.-class AI…" system
+        # message; gluing that onto the persona gave the 4B two conflicting
+        # identities (it answered in the generic voice and confabulated an
+        # attachment it never read). So for a persona, REPLACE the client system
+        # message with the persona prompt instead of concatenating.
+        _persona_active = False
+        try:
+            import codec_models as _cmp
+            _persona_active = bool(_cmp.model_extras().get("system_prompt"))
+        except Exception as _e:
+            log.debug("persona check failed: %s", _e)
+
         # Prepend system message (or replace existing one)
         if messages and messages[0].get("role") == "system":
-            messages[0]["content"] = sys_prompt + "\n\n" + messages[0]["content"]
+            messages[0]["content"] = (
+                sys_prompt if _persona_active
+                else sys_prompt + "\n\n" + messages[0]["content"]
+            )
         else:
             messages.insert(0, {"role": "system", "content": sys_prompt})
 
         # Think mode: append the reasoning scaffold LAST so it wins over the base
         # prompt's "answer directly"/emoji rules. Frontend sends reason_scaffold
         # = Think-toggle state; it parses the resulting <thinking>/### FINAL ANSWER.
-        if body.get("reason_scaffold") and messages and messages[0].get("role") == "system":
+        # Skipped for persona models: a small fine-tune renders the format as its
+        # own <Final Answer> variant the UI stripper misses, and it fights the
+        # persona's voice.
+        if (body.get("reason_scaffold") and not _persona_active
+                and messages and messages[0].get("role") == "system"):
             messages[0]["content"] += _REASON_SCAFFOLD
 
         stream_mode = body.get("stream", False)