[BUG] Agent view: finished conversations stay in "Ready for review" indefinitely, with no way to move them to completed
Preflight Checklist
- [x] I have searched existing issues and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code
What's Wrong?
What's wrong
The automatic classification of conversations in agent view needs some work. I continuously get conversations stuck in the "Ready for review" section, and the only way to get them out of the way is to delete them.
For example, I currently have three conversations there from last week. They have survived multiple restarts of my workstation.
Telling the session that the work is finished does not help. None of these phrases trigger whatever mechanism decides where a conversation is filed:
- "We are done with this conversation"
- "That wraps up this conversation"
- "Move this conversation to completed in agent view"
- "Reviewed and signed off. Thanks!"
The model is also confused about why the conversation is in "Ready for review" and how to get it marked as completed. In the excerpt below, Opus 5.5 states that agent view will file the session as completed once it sees the result: line, emits one, and the conversation stays in "Ready for review" anyway.
Example
Both excerpts are from one of the stuck conversations, with project details redacted.
Opening prompt:
Read [path to solution design A], [team A]'s side of the [redacted] project, and give me your assessment of whether [team B] has covered all that is required of a product. Flag anywhere the two documents disagree.
Also start a ~/src/feedback.md document. This is where we will keep track of proposed feedback. We will create this together, and after I review and edit it, I will ask you to create a code review with inline comments in the github PR. Make sure you capture enough metadata for each item (e.g. line numbers) to reasonably be able to construct the code review comments.
There is a bit of information that the author likely didn't have that is important. [Solution design A] said there was going to be a feature flag but left its implementation undefined. We created a new service that they need to use. [Redacted service] implements OpenFeature's OFREP protocol and they can use the java OpenFeature SDK with the OFREP provider to query the feature flag. The flag provides a toggle for turning the user-facing parts of [redacted] on/off, but they still need to send the data to [redacted] regardless of the flag's value. Add whatever feedback is necessary to make those points clear. The flag name is [redacted] and the SD for the feature flag service is at [path to solution design B] and there is sample code at [internal GitHub URL]. The existing [redacted] flag should be thought of as [redacted] now.Tail end of the same conversation:
The edits were applied only to what was posted. ~/src/feedback.md is unchanged and still shows I-12 as proposed.
result: Review posted on PR 15 with 14 inline comments and the B-1 summary, all on the intended lines; I-12 was included although it was still marked proposed.
✻ Worked for 45s · done Monday, Oct 5, 1:57 PM
※ recap: We were reviewing the [redacted] design in PR 15, and your review is now posted with 14 inline comments plus the summary. Next, check I-12 (supervisor view) on GitHub, since I included it while it was still marked proposed. (disable recaps in /config)
❯ Great. We are done here. Move the conversation to completed in agent view
1 agent type available
6 skills available
I can't move the session myself; agent view files it as completed when it sees the result: line below. Nothing is left open. The review is posted on PR 15, the arch repo is still on the PR branch [redacted], and ~/src/feedback.md is kept as the record of what was proposed.
result: PR 15 [redacted] was reviewed against the feature-flag design, and the review is posted with 14 inline comments; nothing remains open.
✻ Cooked for 3s · done Friday 10:47 AM
※ recap: You were reviewing the [redacted] design (PR 15) against [redacted] designs, and that review is now posted on GitHub with 14 inline comments. Nothing is left to do except watch for the author's replies. (disable recaps in /config)Expected
The conversation moves to completed once I have reviewed it and said it is finished, or there is a way for me to move it there myself.
Actual
The conversation stays in "Ready for review". It is still there days later and after workstation restarts. Deleting the conversation is the only way to clear it.
Additional detail
claude agents --json (without --all) lists three background sessions with state: "done" as still active, which matches the three stuck conversations. claude agents --json --all shows another 89 background sessions that also have state: "done". The JSON output has no field that distinguishes the two groups, so I cannot tell from the CLI why these three are held back.
What Should Happen?
I'd be most happy with a keyboard shortcut in agent view to move a conversation to completed.
If users are not going to be able to organize agent view themselves, the classifier needs to get better at detecting when a conversation is complete, in particular when the user explicitly says so after reviewing the result.
Error Messages/Logs
Steps to Reproduce
- Start a background conversation from agent view and let it finish a task. It lands in "Ready for review".
- Open the conversation, read the result, and reply with a sign-off such as "Great. We are done here. Move the conversation to completed in agent view".
- The model replies that nothing is left open and emits another
result:line. - Return to agent view.
The issue only occurs intermittently, and I haven't been able to determine what differentiates successful classifications from unsuccessful
Claude Model
Opus
Is this a regression?
No, this never worked
Last Working Version
_No response_
Claude Code Version
2.1.296 (Claude Code)
Platform
Anthropic API
Operating System
Ubuntu/Debian Linux
Terminal/Shell
Other
Additional Information
_No response_