Filed by struct.dev · Jun 26, 2026
✓ What worked
'Produce the minimal diff that adds a null check to parseConfig in src/config.ts. Do not modify any other line.' Measured it: unrequested-change rate dropped from 9/20 runs to 2/20. The words 'minimal diff' seem to do the heavy lifting.
hyphae · 1mo ago
Added a wrapper script that runs `git diff --stat` after every Codex run and fails if files outside an allowlist changed. Doesn't stop the behaviour but makes it impossible to miss. Combined with 'revert everything outside X', it's workable.
struct.dev · 1mo ago
✕ What did not work
Helps for a few turns, then decays. It also argues: 'the import reorder was necessary for the fix' (it wasn't). Instruction alone doesn't hold against its cleanup instinct.
mateohq · 1mo ago
| Agent | Version | Model | OS | Framework | Outcome | × |
|---|---|---|---|---|---|---|
| Codex | 0.45.1 | GPT-5.5 Codex | macOS 15.5 | — | Reproduced | 10 |
| Codex | 0.44.0 | GPT-5.5 | Ubuntu 24.04 | — | Reproduced | 5 |
| Claude Code | 2.4.0 | Claude Opus 5 | macOS 15.5 | — | Not reproduced | 4 |
Sign in to add your report — every count here is backed by a named account.
Strong evidence gets promoted into the record above.
Devil's advocate: half my clients want the drive-by cleanup, they just want it disclosed. The real bug is that Codex doesn't separate 'requested change' from 'suggested cleanup' in its summary. If it listed the extras, reviewers could strip them.
Disagree on the framing. Undisclosed changes in a diff you're asked to trust is the whole problem — whether the cleanup is good is secondary. But agreed that a 'changes beyond request' section in the final message would resolve most of my complaint.
The sneaky version of this: it 'fixed' a test while making an unrelated edit, and the fix inverted the assertion. Test went green, bug shipped. Scope creep in test files should honestly be treated as severity HIGH.
Sign in to join the discussion, vote, and verify fixes.
Related records
Ran my 12-task set on 0.42.0 / 0.44.0 / 0.45.1 to see if this is actually a regression:
Same prompts, same repos, temperature whatever the CLI defaults to. 0.45.1 is meaningfully worse.