Repository navigation
[Bug]: Codex paginated thread revert fails after restoring workspace files #8958
Description
Activity
Confirming the same error on Linux with a headless T3 server:
- T3:
0.0.39-nightly.20260905.1285 - Codex CLI:
0.153.4 - Action: go back to an earlier message in an existing Codex conversation.
Provider adapter request failed (codex) for thread/rollback: paginated threads do not support thread/rollbackInspection of the installed T3 bundle and its source map confirms that
CodexSessionRuntime.rollbackThreadstill callsthread/rollbackunconditionally. The staged0.0.39-nightly.20260905.1286bundle contains the same call; that candidate was inspected, not tested through the UI.A possible implementation lead: the experimental protocol schema generated directly from Codex 0.153.4 exposes:
// Replace paginated history with the prefix before this turn. thread/revert({ threadId, beforeTurnId })
beforeTurnIdexcludes that turn and every later turn. The response'sthread.turnsis always empty; retained history must be loaded throughthread/turns/list. The schema explicitly says this operation changes persisted conversation history only, not local files. This needs turn-boundary mapping and history hydration, rather than simply renaming the rollback request. We have not exercised this replacement operation end to end.The installed
CheckpointReactor.handleRevertRequestedalso restores the filesystem before requesting conversation rollback, as described in this issue. We have not established whether files were actually changed in our failed attempt.Related work: #9069 addresses workspace recovery when provider rollback fails. #9169 addresses the same unsupported operation in the retry/edit fork path using
thread/fork.lastTurnId; its scope differs from this in-place checkpoint revert.Useful regression coverage would include a paginated thread reopened before reverting, agreement between retained messages and restored files after success, continuation after restart, and a provider rejection that preserves pre-attempt files, including uncommitted edits.
Investigation and comment prepared with Codex.
- T3:
Fixed by #11338 (merged to main) — rewind uses Codex's paginated
thread/revertAPI so paginated threads no longer fail after workspace restore withpaginated threads do not support thread/rollback.
Before submitting
Area
apps/server
Steps to reproduce
Expected behavior
T3 Code should restore both the workspace and the Codex conversation to the selected point.
If Codex cannot rewind this type of thread, T3 Code should stop before changing any files and explain that the revert is unavailable.
Actual behavior
T3 Code reports this error:
The revert can fail after T3 Code has restored the workspace checkpoint. The files may move back to the selected point while the Codex conversation remains unchanged.
In
CheckpointReactor.handleRevertRequested,checkpointStore.restoreCheckpointruns beforeproviderService.rollbackConversation. When Codex rejectsthread/rollback, the operation stops after the file change.Impact
Major degradation or frequent failure
A failed revert can leave the workspace and conversation at different points. The user may continue the thread without realizing that the files have already changed.
Version or commit
T3 Code Nightly 0.0.38-nightly.20260831.1240
Environment
macOS 26.5.2, T3 Code desktop nightly, Codex CLI 0.151.0
Logs or stack traces
Workaround
No in-app workaround found. Check the Git diff after the failure before continuing the thread.