zero-gpu: skip the forced think block for triage labels

BUILD.json

@@ -1,5 +1,5 @@
 {
-  "built_at": "2026-10-29T15:02:32",
+  "built_at": "2026-10-29T15:10:15",
   "commit": "e9f5ce0",
   "files": 53,
   "with_store": false,

traceforge/config.py

@@ -119,6 +119,7 @@ class Config:
     triage_model: str = ""      # empty = server default
     triage_max_chars: int = 24_000  # per-window char budget (≈6k tokens)
     triage_max_tokens: int = 4000   # per-window label emit ceiling
+    triage_think: bool = True       # False = skip the model's think block
     triage_overlap_messages: int = 4
     triage_workers: int = 8     # parallel window requests
     # Parallel traces in the pre-pass (2026-10-23). The pre-pass used to be

traceforge/space/local_llm.py

@@ -288,7 +288,7 @@ def _prefetch_error() -> str:
 
 @_gpu
 def generate(messages: list[dict], *, max_tokens: int = 512,
-             temperature: float = 0.2) -> tuple[str, dict]:
+             temperature: float = 0.2, think: bool = True) -> tuple[str, dict]:
     """One chat completion on the local model (inside a GPU slice)."""
     if _stub():
         return '{"ok": true}', {"stub": True}
@@ -303,6 +303,11 @@ def generate(messages: list[dict], *, max_tokens: int = 512,
     model, tok = _load()
     prompt = tok.apply_chat_template(messages, tokenize=False,
                                      add_generation_prompt=True)
+    if not think and prompt.rstrip().endswith(" thinking"):
+        # The template opens a think block for us; closing it immediately
+        # skips the reasoning pass.  At ~16 tok/s on the MIG slice the
+        # reasoning costs more than the answer it improves.
+        prompt = prompt.rstrip() + "\n\n\n\n"
     enc = tok(prompt, return_tensors="pt", truncation=True,
               max_length=LOCAL_MAX_INPUT_TOKENS).to(LOCAL_DEVICE)
     n = max(1, min(int(max_tokens or 512), LOCAL_MAX_NEW_TOKENS))
@@ -406,6 +411,7 @@ class _Handler(BaseHTTPRequestHandler):
                     messages,
                     max_tokens=int(payload.get("max_tokens") or 512),
                     temperature=float(payload.get("temperature") or 0.2),
+                    think=bool(payload.get("think", True)),
                 )
         except Exception as e:
             global LAST_ERROR

traceforge/space/pipeline.py

@@ -89,6 +89,9 @@ def space_config(data_dir: Optional[str] = None, **overrides: Any) -> Config:
         # Labels are compact JSON: a 4000-token ceiling let the model ramble
         # for 3264 tokens (199 s) and blow the slice.
         cfg.triage_max_tokens = _env_int("TRACEFORGE_LOCAL_TRIAGE_TOKENS", 600)
+        # The template forces a </think> block; closing it up front keeps the
+        # whole budget for the labels.
+        cfg.triage_think = False
         cfg.rewrite_max_tokens = _env_int("TRACEFORGE_LOCAL_REWRITE_TOKENS",
                                           1000)
         cfg.judge_max_tokens = _env_int("TRACEFORGE_LOCAL_JUDGE_TOKENS", 1200)

traceforge/triage.py

@@ -361,6 +361,10 @@ def _classify_window(
             "temperature": cfg.triage_temperature,
             "max_tokens": cfg.triage_max_tokens,
         }
+        if not cfg.triage_think:
+            # Only the local server understands this; remote endpoints get
+            # the field stripped so their schema validation never sees it.
+            payload["think"] = False
         if not model:
             payload.pop("model")
         payload.update(variant)