Fix gpt-oss empty turns: reasoning_effort=low via chat-template kwarg

At the harmony default (medium) effort, gpt-oss-20b spent the whole 512-token
budget in its analysis channel and never emitted assistantfinal, so the
normalizer dropped the output and Oona sat every turn out (trace health gate
measured 10/10 empty; she appeared in 1/10 attract frames). A plain
"Reasoning: low" line in the system message does nothing (the template routes
system content to the developer block), so the knob is passed server-side as
a chat-template kwarg, baked via TTW_REASONING_EFFORT into the engine image
(deployed; TypeError fallback for older vLLM). Live-verified: Oona returns
clean JSON with 4 valid offers; republished traces show 0/10 empty for
OpenAI (all four labs healthy). README deploy line updated.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

README.md

@@ -112,7 +112,7 @@ pip install -r requirements.txt
 # No GPU, dummy agents (for trying the UI):
 TTW_DUMMY=1 python app.py
 # The multi-model council on Modal (deploy each engine, then enable the council):
-TTW_APP_NAME=ttw-serve-gptoss   TTW_MODEL=openai/gpt-oss-20b           TTW_CUDA_DEVEL=1                  python -m modal deploy serve_council.py
+TTW_APP_NAME=ttw-serve-gptoss   TTW_MODEL=openai/gpt-oss-20b           TTW_CUDA_DEVEL=1 TTW_REASONING_EFFORT=low python -m modal deploy serve_council.py
 TTW_APP_NAME=ttw-serve-minicpm  TTW_MODEL=openbmb/MiniCPM3-4B          TTW_CUDA_DEVEL=1 TTW_TRUST_REMOTE=1 python -m modal deploy serve_council.py
 TTW_APP_NAME=ttw-serve-nemotron TTW_MODEL=nvidia/Nemotron-Mini-4B-Instruct TTW_CUDA_DEVEL=1              python -m modal deploy serve_council.py
 TTW_APP_NAME=ttw-serve-qwen     TTW_MODEL=AdmiralTaco/ttw-trader-0.5b  TTW_CUDA_DEVEL=1                  python -m modal deploy serve_council.py

serve_council.py

@@ -39,6 +39,15 @@ CUDA_DEVEL = os.environ.get("TTW_CUDA_DEVEL", "0") == "1"
 CUDA_TAG = os.environ.get("TTW_CUDA_TAG", "12.4.1-devel-ubuntu22.04")
 # Some models ship custom modeling code (MiniCPM3); vLLM needs explicit opt-in.
 TRUST_REMOTE = os.environ.get("TTW_TRUST_REMOTE", "0") == "1"
+# Chat-template reasoning effort (gpt-oss harmony: low/medium/high; empty = leave
+# the template default). At the default (medium) gpt-oss-20b spends the whole
+# per-turn budget in its analysis channel and never emits `assistantfinal`, so
+# the council's harmony normalizer drops the output and Oona sits every turn out
+# (measured: 10/10 empty turns). Deploy the gptoss engine with
+# TTW_REASONING_EFFORT=low to reach the final answer inside the budget. A plain
+# "Reasoning: low" line in the system message does NOT work: the template routes
+# system content to the developer block; the knob must be a template kwarg.
+REASONING_EFFORT = os.environ.get("TTW_REASONING_EFFORT", "")
 
 app = modal.App(APP_NAME)
 
@@ -69,6 +78,7 @@ def build_image() -> modal.Image:
                 "TTW_MAX_LEN": str(MAX_LEN),
                 "TTW_DTYPE": DTYPE,
                 "TTW_TRUST_REMOTE": "1" if TRUST_REMOTE else "0",
+                "TTW_REASONING_EFFORT": REASONING_EFFORT,
             }
         )
     )
@@ -113,7 +123,19 @@ class Engine:
         params = self.SamplingParams(
             temperature=temperature, top_p=0.9, max_tokens=max_tokens
         )
-        outputs = self.llm.chat(conversations, params)
+        effort = os.environ.get("TTW_REASONING_EFFORT", "")
+        if effort:
+            try:
+                outputs = self.llm.chat(
+                    conversations, params,
+                    chat_template_kwargs={"reasoning_effort": effort},
+                )
+            except TypeError:
+                # Older vLLM without chat_template_kwargs: serve without the
+                # knob rather than fail the whole engine.
+                outputs = self.llm.chat(conversations, params)
+        else:
+            outputs = self.llm.chat(conversations, params)
         return [o.outputs[0].text for o in outputs]