openai
tools: truncate lines in session history
This PR applies the truncation policy to session history, and is a minimal fix for #7395. I chose not to simply omit gitignored files in capture_existing_untracked since that seems to be a deliberate design decision. I have changed the warning for format_large_untracked_warning somewhat, which may address similar reports like #7405, #7369, #7313 and #7067.
What
- Persisted response items now use a single processing pass that truncates tool outputs according to the existing
TruncationPolicy. The truncated copy is what goes into history/rollout; live output during the turn remains unchanged. - Reworded the large-untracked warning to accurately state that undo snapshots will stage large untracked/ignored dirs and may be slow.
Why
- Large tool outputs were being inlined in rollout JSONL leading to session bloat. Bounding the persisted copy keeps session files small while retaining head/tail context.
- Warning now reflects current behavior: snapshots still stage untracked/ignored content, so the cost is time/disk, not JSONL size.
How
-
context_manager/history.rs: centralized process_response_items, ensured processed items are used for persistence. -
codex.rs,compact.rs: persistence now uses processed items. Live events are unchanged.
@jif-oai thank you for the comments! I've addressed them and moved the processing back into the context manager.
After some internal discussion, we've decided to pass on this change.
As noted above, the root cause of the ghost_commit issue was already fixed separately. And more importantly, the proposed change introduces truncation at a point where it doesn't belong.
I get what you guys mean, especially with changes like https://github.com/openai/codex/commit/349734e38d91c6d70b7dd04e3e9c5d5319132323 and https://github.com/openai/codex/commit/2222cab9eafb5e8abd100be465515a4c6be028f2. This PR was raised a few days before these fixes came in. Thanks for the feedback @jif-oai, @etraut-openai!