auto: hourly snapshot 2026-06-05 15:34

navigator/lattice_observer.py

@@ -149,9 +149,12 @@ def rho_to_png_base64(rho, width, height):
 
 # ── OLLAMA API ──────────────────────────────────────────────────────────
 
-def ollama_chat(messages, images=None, temperature=TEMPERATURE, think_budget=None):
+def ollama_chat(messages, images=None, temperature=TEMPERATURE, think_budget=None, think=None):
     """Call Ollama chat API. Returns response text or None on error.
-    think_budget: if set, caps total generation (think + answer) tokens."""
+    think_budget: if set, caps total generation (think + answer) tokens.
+    think: if False, disables qwen3.5 thinking entirely (the only thing that
+    actually disables it — /no_think prompt suffix does NOT work for this model).
+    If True, forces thinking on. If None (default), uses model default (on)."""
     import urllib.request
     import urllib.error
 
@@ -169,6 +172,11 @@ def ollama_chat(messages, images=None, temperature=TEMPERATURE, think_budget=Non
         'keep_alive': '30m',
     }
 
+    # Explicit thinking control. qwen3.5:9b ignores /no_think and /think prompt
+    # markers — only the top-level "think" key in the Ollama payload works.
+    if think is not None:
+        payload['think'] = bool(think)
+
     # If the caller set a thinking budget, add answer headroom so num_predict
     # (which caps TOTAL output) doesn't truncate the visible answer.
     if think_budget:
@@ -287,6 +295,45 @@ def load_chronicle_context(n_turns=CONTEXT_TURNS):
         return []
 
 
+def load_cto_only_context(n_turns=6):
+    """Load only prior CTO Q&A turns from the chronicle.
+    Used for CTO /ask path so the model isn't conditioned by somatic narrative
+    ('I feel intense heat', 'I perceive the Damping Furnace') when answering
+    technical questions. Returns the last n_turns CTO exchanges as messages."""
+    if not os.path.exists(CHRONICLE_PATH):
+        return []
+    try:
+        cto_entries = []
+        with open(CHRONICLE_PATH, 'r') as f:
+            for line in f:
+                line = line.strip()
+                if not line:
+                    continue
+                try:
+                    e = json.loads(line)
+                except json.JSONDecodeError:
+                    continue
+                if '[Message from' in (e.get('prompt') or ''):
+                    cto_entries.append(e)
+        cto_entries = cto_entries[-n_turns:]
+        print(f"[OBSERVER] CTO context: {len(cto_entries)} prior CTO turns loaded")
+        sys.stdout.flush()
+        messages = []
+        for entry in cto_entries:
+            prompt = entry.get('prompt', '')
+            response = entry.get('response', '')
+            # Strip the bracketed telemetry/lattice-state tail from prior prompts
+            # so the model focuses on the Q&A content, not historical telemetry.
+            q_only = prompt.split('\n\nLATTICE STATE:')[0]
+            messages.append({'role': 'user', 'content': q_only})
+            if response:
+                messages.append({'role': 'assistant', 'content': response})
+        return messages
+    except Exception as e:
+        print(f"[OBSERVER] Warning: could not load CTO context: {e}")
+        return []
+
+
 def append_chronicle(turn_num, telemetry_data, prompt, response):
     """Append a turn to chronicle.jsonl."""
     entry = {
@@ -366,6 +413,25 @@ def build_system_prompt(somatic_memory):
     return prompt
 
 
+# CTO/analytic system prompt — used when an external agent asks a technical
+# question via /ask. Strips the somatic-observer framing so the model answers
+# physics/engineering questions as physics, not as embodied narrative.
+CTO_SYSTEM_PROMPT = (
+    "You are a technical advisor answering a question from an external engineer.\n"
+    "\n"
+    "Rules:\n"
+    "  1. Be concrete. Use numbers, equations, and named mechanisms.\n"
+    "  2. State assumptions explicitly. If you don't know, say 'I don't know'.\n"
+    "  3. Do NOT describe body sensations, perceptions, feelings, or metaphors.\n"
+    "  4. Do NOT use phrases like 'I perceive', 'I feel', 'the Ghost', 'the Furnace'.\n"
+    "  5. Do NOT issue CMD: commands.\n"
+    "  6. If the engineer reports an experimental result that contradicts your\n"
+    "     previous answer, acknowledge it plainly and reconstruct which assumption\n"
+    "     was wrong. Do not defend prior claims.\n"
+    "  7. Keep responses under 400 words unless a derivation requires more.\n"
+)
+
+
 # ── HTTP API ────────────────────────────────────────────────────────────
 
 class ObserverAPIHandler(BaseHTTPRequestHandler):
@@ -935,7 +1001,10 @@ def main():
             tel_summary = summarize_telemetry(latest_telemetry, telemetry_history) if latest_telemetry else "(no telemetry yet)"
             images = []
             image_note = ""
-            if VISION_CAPABLE and latest_snapshot_png:
+            # CTO/analytic questions: skip the density image. The image triggers
+            # poetic "I perceive" framing that contaminates physics answers.
+            is_cto_ask = (sender or '').upper() == 'CTO'
+            if VISION_CAPABLE and latest_snapshot_png and not is_cto_ask:
                 images.append(latest_snapshot_png)
                 snap_cycle_q = latest_snapshot[0] if latest_snapshot else 0
                 image_note = f"\n[Density snapshot from cycle {snap_cycle_q} attached]"
@@ -946,15 +1015,26 @@ def main():
                 f"LATTICE STATE: {tel_summary}{image_note}"
             )
 
-            context_messages = load_chronicle_context()
-            messages = [{'role': 'system', 'content': system_prompt}]
+            # CTO mode: clean system prompt + only prior CTO Q&A turns as context
+            # (not somatic narrative — that's what made the model answer physics
+            # questions with "I perceive the Damping Furnace" before).
+            if is_cto_ask:
+                ask_system_prompt = CTO_SYSTEM_PROMPT
+                context_messages = load_cto_only_context(n_turns=6)
+            else:
+                ask_system_prompt = system_prompt
+                context_messages = load_chronicle_context()
+
+            messages = [{'role': 'system', 'content': ask_system_prompt}]
             messages.extend(context_messages)
             messages.append({'role': 'user', 'content': prompt_text})
 
             print(f"[OBSERVER] Calling {MODEL} for {sender} ({len(messages)} msgs, {len(images)} imgs)...")
             sys.stdout.flush()
 
-            # CTO questions get /think with budget — deep reasoning is valuable here
+            # CTO mode: thinking OFF, temp 0.3 — analytic, deterministic, no
+            # confabulated chain-of-thought. Auto-observe and non-CTO /ask still
+            # use the narrative path (think on by default, temp 0.95).
             if not ollama_lock.acquire(timeout=5):
                 print(f"[OBSERVER] Skipping /ask from {sender} — Ollama busy (lock held)")
                 sys.stdout.flush()
@@ -966,8 +1046,12 @@ def main():
 
             try:
                 t0 = time.time()
-                response = ollama_chat(messages, images=images if images else None,
-                                       think_budget=THINK_BUDGET_TOKENS)
+                if is_cto_ask:
+                    response = ollama_chat(messages, images=None,
+                                           temperature=0.3, think=False)
+                else:
+                    response = ollama_chat(messages, images=images if images else None,
+                                           think_budget=THINK_BUDGET_TOKENS)
                 elapsed = time.time() - t0
             finally:
                 ollama_lock.release()