Fix: max_input_tokens extracted from YAML config and passed as the separate run_eval param (was going into gen_kwargs where truncation never saw it -> raw 128k text sent to gateway -> 400)

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

evalharness/cli.py

@@ -652,10 +652,15 @@ def _cmd_eval_run(args) -> int:
                     else:
                         _print_phase(_console, _idx, total_runs, _name, msg)
                 _gen_kw = {**bench_cfg, **(getattr(args, '_gen_override', {}) or {})}
+                # max_input_tokens must be a SEPARATE run_eval param (it drives
+                # truncation in assemble(), not a gen_kwarg the adapter sees) --
+                # extract it from the YAML-derived dict
+                _mit = _gen_kw.pop('max_input_tokens', 0) or getattr(args, 'max_input_tokens', 0)
                 report = asyncio.run(run_eval(
                     ds, model_spec, concurrency=args.concurrency, limit=args.limit,
                     limit_per_task=args.limit_per_task,
                     gen_kwargs=_gen_kw or None,
+                    max_input_tokens=_mit,
                     checkpoint=args.resume,
                     judge_spec=_compose_judge_spec(args), env=args.env,
                     api_key=getattr(args, 'api_key', ''),