Bug Description

_process_batch_worker() intentionally discards prompts whose trajectories have has_any_reasoning == False, but those prompt indices are never added to completed_prompts. On --resume, they are retried forever.

Affected files / lines

• batch_runner.py:441-447
• batch_runner.py:493-497

Why this is a bug

These prompts are not failures; they are explicit discard decisions. Because they are omitted from completed_in_batch, resume treats them as unfinished work and re-spends tokens on them indefinitely.

Minimal Reproduction

Monkeypatch _process_single_prompt() to return:

{
    "success": True,
    "trajectory": [{"role": "assistant", "content": "x"}],
    "reasoning_stats": {"has_any_reasoning": False},
    "tool_stats": {},
    "metadata": {},
    "completed": True,
    "api_calls": 1,
    "toolsets_used": [],
}

Then run:

out = _process_batch_worker((1, [(0, {"prompt": "hi"})], tmpdir, set(), {"verbose": False}))
print(out["discarded_no_reasoning"], out["completed_prompts"])

Observed output:

• discarded_no_reasoning == 1
• completed_prompts == []

Expected Behavior

Once a prompt is explicitly discarded for no reasoning, resume should treat that prompt as completed/dispositioned and not re-run it forever.

Actual Behavior

Discarded prompts never enter the completed set, so --resume retries them on every run.

Suggested Investigation Direction

Record discarded prompt indices separately (or include them in completed_prompts) so the resume checkpoint represents all terminal outcomes, not just saved trajectories.