Add decode-only study mode support

src/aituner/llm.py

@@ -15,6 +15,15 @@ def build_prompt(
     state: StudyState,
     capability_profile: dict[str, Any] | None,
 ) -> str:
+    objective_notes: list[str] = []
+    if study.trace.request_mode == "decode_only":
+        objective_notes.append(
+            "This study is decode-only. The engine uses a KV decode benchmark connector, so TTFT is informational only unless an explicit TTFT rule is configured."
+        )
+    if study.slo.ttft_rule is None:
+        objective_notes.append("There is no TTFT SLO for this study.")
+    if study.slo.tpot_rule is None:
+        objective_notes.append("There is no TPOT SLO for this study.")
     history = []
     for trial in state.trials[-study.llm.max_history_trials :]:
         history.append(
@@ -57,6 +66,7 @@ def build_prompt(
                 },
                 "trace": {
                     "window_id": study.trace.window_id,
+                    "request_mode": study.trace.request_mode,
                     "input_length_filter": (
                         {
                             "min_input_tokens": study.trace.input_length_filter.min_input_tokens,
@@ -88,6 +98,7 @@ def build_prompt(
                 "target_pass_rate": study.slo.target_pass_rate,
                 "ttft_rule": study.slo.ttft_rule,
                 "tpot_rule": study.slo.tpot_rule,
+                "objective_notes": objective_notes,
             },
             default=lambda value: value.__dict__,
             ensure_ascii=False,

src/aituner/spec.py

@@ -237,6 +237,7 @@ class TraceSpec:
     windows_path: str
     window_id: str
     trace_file_override: str | None
+    request_mode: str
     u_field: str
     timestamp_field: str
     max_concurrency: int
@@ -251,12 +252,16 @@ class TraceSpec:
     def from_dict(cls, data: Mapping[str, Any]) -> "TraceSpec":
         max_requests = data.get("max_requests_per_probe")
         synthetic_prompt_cap = data.get("synthetic_prompt_cap_tokens")
+        request_mode = str(data.get("request_mode") or "chat").strip().lower()
+        if request_mode not in {"chat", "decode_only"}:
+            raise SpecError("trace.request_mode must be one of: chat, decode_only.")
         return cls(
             windows_path=_require_str(data.get("windows_path"), context="trace.windows_path"),
             window_id=_require_str(data.get("window_id"), context="trace.window_id"),
             trace_file_override=str(data["trace_file_override"]).strip()
             if data.get("trace_file_override")
             else None,
+            request_mode=request_mode,
             u_field=str(data.get("u_field") or "sampling_u").strip(),
             timestamp_field=str(data.get("timestamp_field") or "timestamp").strip(),
             max_concurrency=_require_int(

src/aituner/worker.py

@@ -71,6 +71,7 @@ def _latency_summary(
     tpot_values = [float(item.tpot_ms) for item in outcomes if item.tpot_ms is not None]
     return {
         "observed_request_count": len(outcomes),
+        "request_mode": study.trace.request_mode,
         "ttft_ms": _metric_summary(ttft_values),
         "tpot_ms": _metric_summary(tpot_values),
         "failed_reason_counts": _reason_counts(evaluations),