[BUG] Claude Desktop 1.6608.0 (macOS): All rewind attempts rejected with "Can't rewind to this message" — regression after auto-update

Status Closed — not planned

Maintainer reply None cached

Activity 9 comments · opened May 8, 2026 · closed May 8, 2026

Preflight Checklist

[x] I have searched existing issues and this hasn't been reported yet
[x] This is a single bug report
[x] I am using the latest version of Claude Code

What's Wrong?

In Claude Desktop 1.6608.0 on macOS, every rewind attempt is rejected with the popup "Can't rewind to this message" — including in a freshly created empty conversation with only 2 user messages exchanged. The rewind picker renders normally and shows anchor candidates, but selecting any anchor triggers the rejection popup.

This affects every conversation — fresh, resumed, native, or imported — and every message position, not just the first message. So this is a much broader failure than:

#50780 (Desktop first-message-only)
#48955 (CLI-origin messages via /desktop only)

The same machine's terminal Claude Code (CLI) rewind continues to work correctly, isolating the regression to the Desktop client.

What Should Happen?

Selecting an anchor in the rewind picker should restore the conversation (and optionally code) to the point before that message, as it did before the 1.6608.0 update.

Error Messages/Logs

~/Library/Logs/Claude/main.log shows two repeating patterns on every rewind attempt today:

2026-05-08 15:34:14 [info] LocalSessions.rewind: sessionId=local_0614a145-... target=e88e4bcb-...
      at dE.rewindSession (/Applications/Claude.app/Contents/Resources/app.asar/.vite/build/index.js:2420:4427)
      at Object.rewind (/Applications/Claude.app/Contents/Resources/app.asar/.vite/build/index.js:2506:4220)
2026-05-08 15:34:14 [info] [Rewind] No rewind point for local_0614a145-... target=e88e4bcb-... (assistantUuid=none, hasText=true)

2026-05-08 15:40:35 [info] [Rewind] target=9c70ad22-... → assistantUuid=ac2cc93a-... not on active chain for local_c0799054-... (chain size 2/14)

Earlier in the day on a long-running conversation: chain size 4/178 — only 4 of 178 messages on active chain. New empty conversations show the assistantUuid=none, hasText=true variant.

All rewind calls today route through LocalSessions.rewind (client-side path).

Steps to Reproduce

Open Claude Desktop 1.6608.0 on macOS
Start a fresh conversation (Cmd+N)
Send a short message ("hi") and wait for response
Send another short message and wait for response (now there are 2 user + 2 assistant messages)
Press Esc Esc (or click rewind button) — picker renders with anchors
Select any anchor
Popup: "Can't rewind to this message"

Reproduces 100% on every conversation tested.

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Pre-1.6608.0 — regression appeared after Claude Desktop auto-update on 2026-05-07.

Claude Code Version

Claude Desktop 1.6608.0
(CFBundleShortVersionString from /Applications/Claude.app/Contents/Info.plist, mtime 2026-05-07 14:33)

Platform

Anthropic API

Operating System

macOS

Additional Context

All [Rewind] log entries today route through LocalSessions.rewind (client-side path), suggesting a client-side validation regression rather than a server-side rejection.
assistantUuid=none, hasText=true indicates the user message lacks a paired assistant UUID in client memory — Desktop now treats this as a hard reject. Previous behavior was apparently more lenient.
The not on active chain (chain size N/M) pattern suggests parentUuid chain validation has been tightened, rejecting messages whose ancestor chain the client cannot fully resolve.
Confirmed plugin SessionStart hooks (superpowers v5.1.0) are not the cause: bug persists after disabling plugin hooks and restarting Desktop.
Same machine, same day, same account: terminal Claude Code (CLI) rewind works correctly.

Related issues (similar but distinct)

#50780 — Desktop cannot rewind first message (this bug rejects every message)
#48955 — Desktop rewind breaks for CLI-origin messages (this bug rejects native Desktop messages too)
#44828 — isMeta:true filter (this bug surfaces a different rejection path: assistantUuid=none and not on active chain)

View original on GitHub ↗

9 Comments

Epikoding · 3 months ago

Additional data point: same Desktop client also fails to follow CLI-side rewind results

On the same machine and same Desktop 1.6608.0 client, I tested the inverse path — rewind initiated from terminal Claude Code on a session that Desktop also has open — and Desktop fails to reflect the CLI rewind result. JSONL analysis shows the in-file branch structure is correct, so this looks like a Desktop-side rendering/sync gap rather than a CLI bug.

Setup

One session jsonl at ~/.claude/projects/<proj>/<sid>.jsonl shared between Desktop and CLI
Desktop entries: entrypoint=claude-desktop, version=2.1.128
CLI entries: entrypoint=cli, version=2.1.133

Reproduction

From Desktop: send 리와인드 테스트 → response received
Switch to terminal CLI on the same session
CLI: rewind (Esc Esc) to the point before 리와인드 테스트 — works on CLI side
CLI: send 리와인드 테스트2 → response received (fork branch added)
CLI screen: only the 리와인드 테스트2 branch visible (correct — active chain)
Desktop screen: both 리와인드 테스트 and 리와인드 테스트2 plus their responses are shown as if on a single linear chain (incorrect)

JSONL tree (key user/assistant nodes)

time           type       uuid        parent      entry
07:10:52       user       dbce28ca    881d84a5    claude-desktop   ← "리와인드 테스트"
07:10:55       assistant  e4ea2c23    dbce28ca    claude-desktop
07:10:55       assistant  77c4c8ae    e4ea2c23    claude-desktop
─── (CLI rewind here) ───
07:11:27       user       30bad87f    881d84a5    cli              ← "리와인드 테스트2"
07:11:38       assistant  cf447317    c5d4d8cb    cli

Both dbce28ca and 30bad87f carry parentUuid=881d84a5 — the correct in-file fork structure (cf. #19451). CLI rendering only the new branch is correct; Desktop rendering both as a flat sequence indicates Desktop is not walking parentUuid chains when reconciling this jsonl with its conversation view.

Unresolved parent UUIDs in the same file

3b37ca03, 23c159b0, 881d84a5 (the fork point itself), c9c5541e, c5d4d8cb are referenced as parents but absent from the file. This matches the chain size N/M log pattern reported above (e.g., chain size 4/178). When most parents cannot be resolved, Desktop's chain validation seems to either reject rewinds outright (the main bug in this report) or, in the inverse direction, fail to truncate forked-off branches.

Relation to this report and to #48955

This is a different failure mode than the popup rejection in the original report — here Desktop renders messages but renders forked-off branches as if on the active chain.
Both behaviors point at the same underlying gap: Desktop 1.6608.0 is not maintaining a coherent parentUuid chain view of session jsonl. Forks become invisible in one direction (no rewind possible), and forks become flattened in the other direction (CLI rewind not reflected).
Possibly related to #48955, but distinct: #48955 covers Desktop-initiated rewind not propagating to the local jsonl; this is the inverse — CLI-initiated rewind structure (which is correctly recorded in the jsonl) not being reflected in Desktop's view.
Same machine's terminal CLI (v2.1.133) handles both directions correctly.

alextes · 3 months ago

workaround.
if you use "fork from here" button it lets you rewrite the message you want to rewind to. you can then archive / delete the original convo.

side notes:

the design on fork is a bit weird. it would make more sense for the button to sit on an answer. as you'd like want to fork either mid conversation or after the last answer. fancy version would put the button between. but on the question and to then allow the user to edit the question makes little sense imo.
if it turns out this isn't some niche combination bug but broken for everyone, can you test releases before you put them out please 😅 .

Epikoding · 3 months ago

Resolution: not a Desktop regression — third-party plugin trigger

After full bisection of my own configuration, the rewind rejection is caused by a third-party plugin (double-shot-latte from the superpowers-marketplace), not Claude Desktop 1.6608.0 itself.

Bisection summary

Minimal ~/.claude/settings.json (env={}, defaultMode=default, all plugins disabled, hooks={}) → rewind works.
Plugins re-enabled one group at a time → broke when double-shot-latte was active.
Removing double-shot-latte while keeping every other variable (env, permissions, all other plugins, hooks) at original values → rewind works.

Likely mechanism

double-shot-latte registers a Stop hook (hooks/run-hook.cmd claude-judge-continuation) that runs after every assistant response. On Desktop 1.6608.0 this appears to add metadata or entries to the conversation that cause Desktop's LocalSessions.rewind validator (assistantUuid=none, hasText=true / not on active chain) to reject every anchor.

The earlier hypothesis of a Desktop 1.6608.0 client-side regression was wrong — the same plugin set on a previous Desktop build presumably also produced fragmented chains, but the older validator did not reject them. Whether the new validator's strictness is itself a regression that should be relaxed is a separate question; the immediate cause of "Can't rewind to this message" in my environment is the plugin's Stop hook output.

Closing

Closing as not-a-bug for the Desktop client. Apologies for the false alarm. Keeping the related JSONL/parentUuid analysis comment above for reference, since it documents what fragmented chains look like in the wild.

If anyone hits the same popup, check whether you have a plugin whose Stop hook injects content that ends up in the session jsonl.

stefansimik · 3 months ago

Same problem here, after any try to rewind, gettting this warning:

I don;t have any plugins installed, just pure clean Claude Desktop

jajibhee · 3 months ago

Same issue here
<img width="331" height="157" alt="Image" src="https://github.com/user-attachments/assets/cab8c474-5fcd-479f-a0d6-4811b923796c" />

limbermicah · 3 months ago

have the same issue (always getting error) and I have no plugins.

<!-- Uploading "Screenshot 2026-05-18 at 10.36.46 AM.png"... —>

rapastranac · 3 months ago

Same issue here, I cannot even rewind to the last query.

b-hermes · 3 months ago

Same issue here. I was asking on how to defend against some supply-chain attacks and now I'm stuck with the API/guard rail error and cannot rewind.

github-actions[bot] · 1 month ago

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.