fix(preprocess): deterministic Path-A for pre-validated harness + nudge prose-only agent turns

[iraj465]

Summary

Two robust, workload-agnostic fixes for the v3 preprocess + agent loop. Both were found while running the gpt-oss-120b fused_moe kernel-optimization flow in mixed mode, but neither is specific to that workload — they are general correctness fixes.

1. Pre-validated harness → deterministic Path-A bypass

run_preprocess_v3 always drove preprocess through the LLM orchestrator, even when the caller supplied a pre-validated harness. The orchestrator's Step-0 "shapes pre-check" then either:

• diverted a shape-bearing task to the harness-generator (regenerate a harness from scratch), or
• simply failed to converge (100+ LLM steps),

…burning the entire preprocess budget (900s soft cap) without ever producing a benchmark_baseline, so the kernel run aborted.

A pre-validated harness already encodes its authoritative shapes internally — the whole A1 sequence (collect_baseline → collect_profile → render_commandment) is deterministic and needs no LLM. _run_prevalidated_path_a() runs it directly and returns PreprocessResult(path_taken="A"), preserving the existing worktree-bypass validate_harness gate. Profiling stays advisory/non-fatal (matches the orchestrator escape-hatch contract). Opt-out: GEAK_NO_PREVALIDATED_BYPASS=1. A prompt-level exemption in the Step-0 classifier is kept as a secondary guard.

Result: preprocess completes in ~260s instead of timing out at 900s.

2. Prose-only agent turns are nudged, not silently accepted

DefaultAgent.parse_action returned {"output":"","returncode":0} for a turn with no fenced bash and no tool call, and the final if all_action["output"] or all_action["returncode"] == 0: accepted it as a successful no-op. A model that believes it already finished (e.g. narrates "Done." / "tasks submitted") then repeats that prose every step with no corrective signal, looping until the step limit.

Observed in the heterogeneous task-planner: 143 prose turns → 0 tool calls → LimitsExceeded. Fix: track whether any action (bash / tool / skill) actually dispatched; if not, raise FormatError (a NonTerminatingException) so the model is nudged to emit a real action. Multi-action turns still raise the existing format error. test_empty_actions_handling still passes (the nudge is non-terminating → the next turn submits).

Test plan

• tests/agents/test_default.py — new test_prose_only_turn_is_nudged_not_silently_accepted; full suite (13) passes
• tests/run/test_preprocess_v3_bugfixes.py — new test_prevalidated_harness_bypasses_llm_orchestrator + test_prevalidated_bypass_opt_out_env; full suite (16) passes
• End-to-end: gpt-oss-120b fused_moe mixed-mode run now completes preprocess (263s) and the planner submits tasks that dispatch to workers

🤖 Generated with Claude Code