perf(save): skip backward-byte scan on hot path

[mikasenghaas]

Summary

save_new_outputs calls truncate_malformed_trailing_line on every append. The helper walks the file backwards one byte at a time (seek(); read(1) per byte) to find the previous \n and validate the trailing row.

That walk only matters once per run — to clean up a partial row a crashed prior write may have left behind. After the first successful append the file ends in \n and the walk does nothing.

On a networked filesystem (BeeGFS / NFS) each read(1) is a network round trip. For an ~80 KB multi-turn bfcl-v3 rollout the scan does ~80,000 syscalls per save → tens of seconds of wall-time. Because save_new_outputs is awaited inline via asyncio.to_thread inside the as_completed loop, the client's main coroutine parks while the scan runs and no new dispatches happen.

Change

Hoist the truncate call out of save_new_outputs and run it exactly once, at the resume site in Environment.generate, right after load_outputs. Steady-state appends no longer pay the scan cost; crash recovery still works because the resume path runs it before any append.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes when malformed trailing JSONL cleanup occurs (now only on resume), which could affect correctness if callers append to an existing partially-written results.jsonl without going through the resume path. Otherwise the change is localized and primarily performance-motivated.

Overview
Speeds up incremental result saving by removing the backward byte-scan/truncation step from the hot append path in save_new_outputs, so steady-state results.jsonl appends no longer pay an O(file-size) tail check.

The malformed-trailing-line cleanup helper is promoted to public (truncate_malformed_trailing_line) and is now invoked once during Environment.generate resume (after load_outputs) to repair any partial final row before future appends; tests are updated to reflect that the caller performs the truncation explicitly.

^{Reviewed by Cursor Bugbot for commit b7f4df0. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Move malformed trailing-line truncation out of save_new_outputs onto the call site

• truncate_malformed_trailing_line is renamed from a private helper to a public function in save_utils.py, and save_new_outputs no longer calls it internally.
• environment.py now calls truncate_malformed_trailing_line explicitly on results.jsonl during the resume-evaluation path, before appending new outputs.
• Behavioral Change: callers of save_new_outputs are now responsible for ensuring a valid JSONL boundary before appending; the function itself no longer truncates.

^{Macroscope summarized b7f4df0.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit b7f4df0. Configure here.}

[cursor]

Truncation skipped when crash precedes metadata creation

Medium Severity

truncate_malformed_trailing_line now only runs inside the resume branch, which requires is_valid_eval_results_path to return True (both results.jsonl and metadata.json must exist). If a crash occurs after save_new_outputs partially writes to results.jsonl but before save_metadata creates metadata.json, a retry with the same results_path takes the fresh-start else branch, skips truncation, and appends new rows directly after the partial data — creating an unrecoverable mid-file corruption that load_outputs will raise on rather than skip.

Additional Locations (1)

• verifiers/utils/save_utils.py#L793-L796

 

^{Reviewed by Cursor Bugbot for commit b7f4df0. Configure here.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR changes crash-recovery behavior by moving truncation logic to only run in the resume branch. An open review comment identifies a potential data corruption scenario when crashes occur before metadata creation. The behavioral change and unresolved bug concern warrant human review.

^{You can customize Macroscope's approvability policy. Learn more.}