[Vu Linh Truong (vulinh1704)]

Fixes #151.

Claude Code's Edit result names no outcome. It is the file's text (filePath, oldString, newString, originalFile, structuredPatch, ...), and a Write result carries the whole file as content. A success has no exit status, so the guard read that text with ERR_RE. Edits to a document that mentions assert, failed or an ...Error counted as error outputs and were tagged [err#...] in the judge's digest (#151).

An edit that does not apply (text not found, file not read first) comes as PostToolUseFailure. So an Edit, MultiEdit or Write that came back as PostToolUse succeeded, and claude_success_code now gives it 0, the same way it gives 0 to a covered shell call since #128. The history import's envelope replay uses the same function, so it reads edits the same way. ERR_RE, SpiralState and response_text are unchanged.

I did not take the fix suggested in the issue (return "" from response_text for a dict with no known key). It covers Edit, but not Write, whose content is a known key and is the whole file. It would also change what every host folds for an unknown dict.

The tests are in a new tests/test_loopguard_edit_results.py, next to test_loopguard_productive_turns.py. They send Edit, MultiEdit and Write results whose file text contains assert, failed and AssertionError, shaped as Claude Code 2.1 logs Edit and Write results. A failed Edit and a failing test run still count as error outputs.

Verified on Ubuntu 24.04 (WSL), Python 3.12, on main at a71684c:

• Before the fix, the new file reports 2 failed and 1 passed. Four successful edits gave error_outputs == 4.
• After the fix, it reports 3 passed.
• python -m pytest tests -rfEs --tb=line reports 739 passed and 16 skipped. python -m pytest tests experiments/tests reports 1132 passed and 20 skipped. Each skip needs a tool the machine does not have (a real Codex, Hermes or opencode CLI, aider) or Windows.

Risk: only Claude Code edits reported as successes change, and they could not have failed. A failed edit still arrives as PostToolUseFailure and is read by its error text, as before.

[Saksham Uboweja (SakshamUboweja)]

Thanks for this! #152 fixes the same issue (#151) and is now merged. Edits to a file that mentions failures no longer count as errors. Closing this as a duplicate.