[BUG] Plan-file inline comments ("Leave a comment for Claude") don't reach the model on resume
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?
Summary
The plan-review UI shows "Leave a comment for Claude…" on hover for each line of the plan file. Comments left this way are visible in the client's right pane but are never delivered to the model — not in the approved-plan payload, not in the plan file on disk, and not through any other channel the model can access.
What does work
Direct file edits made by the user in the plan review UI do come through (they modify the file on disk). So the model can see changes the user types directly into the plan text.
Comments do not. They live only in the client UI.
Tested twice in one session
First test (organic): User left 4 substantive comments on a dashboard plan (field corrections, scope changes). Model couldn't see any of them. User had to screenshot each comment one-by-one and paste into chat.
Second test (controlled): Model wrote 3 marked lines ("Comment test line A/B/C") explicitly asking the user to leave comments. User confirmed comments were visible in the right pane. After approval, all 3 lines came back clean — no comments attached.
Impact
Plan comments are the only structured way for a user to give line-level feedback on a plan without rewriting the plan text. When they don't work:
The model builds the wrong thing (missing field corrections, scope changes, business logic overrides)
The user has to re-communicate feedback through screenshots or chat — defeating the purpose of the feature
Trust in the planning workflow erodes ("why does it say 'for Claude' if Claude can't read it?")
Suggested fix
Either:
(a) Merge comments into the plan file on approval as inline annotations (e.g. <!-- user: comment text --> next to the target line)
(b) Append a ## User Comments section to the approved-plan payload in the ExitPlanMode tool result, with line references
(c) Deliver them as a follow-up user message after plan approval
Option (a) is the most durable — it survives the model re-reading the file later.
Environment
Claude Code (desktop app, macOS)
Model: Claude Opus 4.6 (1M context)
Date: 2026-04-15
What Should Happen?
The model can read the comments after approval — either embedded in the plan file, appended to the ExitPlanMode tool result, or delivered as a user message.
Actual: The model receives the plan with zero trace of the comments. The plan file on disk is byte-identical to what the model wrote pre-approval. The approved-plan payload in the ExitPlanMode result also omits them. The model has no tool or API to access the client's annotation layer.
Error Messages/Logs
Steps to Reproduce
Enter plan mode — model writes a plan to the plan file
Model calls ExitPlanMode
In the plan review UI, hover on any line and click "Leave a comment for Claude…"
Type a comment and click "Comment"
Confirm the comment appears in the right-side annotation pane
Click Approve
Expected: The model can read the comments after approval — either embedded in the plan file, appended to the ExitPlanMode tool result, or delivered as a user message.
Claude Model
None
Is this a regression?
I don't know
Last Working Version
_No response_
Claude Code Version
Claude 1.2773.0 (884b37) 2026-04-15T14:32:09.000Z
Platform
Anthropic API
Operating System
macOS
Terminal/Shell
Terminal.app (macOS)
Additional Information
_No response_
14 Comments
Found 3 possible duplicate issues:
This issue will be automatically closed as a duplicate in 3 days.
🤖 Generated with Claude Code
Plan-file inline comments labeled "Leave a comment for Claude" that are never delivered to the model is a broken UI contract — the label explicitly promises the model will see the comment, and the feature does not deliver on that promise.
This is a trust failure at the UX layer: users make decisions based on the assumption that labeled UI elements do what they say. "Leave a comment for Claude" creates a reasonable expectation that Claude will receive the comment. When the feature silently discards that comment, the user's plan decisions, scope corrections, and business logic overrides are invisibly lost.
The two proposed fixes are both correct — choose one:
Option A: Merge comments into plan file on approval. Inject comments as HTML-style annotations (
<!-- user comment: text -->) adjacent to the target line in the plan file before committing it. The model reads the plan file and sees the annotations as part of the plan content. Advantage: simple, no new payload field needed. Disadvantage: the model must be instructed to handle these annotations, or it may treat them as code or ignore them.Option B: Append a User Comments section to the ExitPlanMode tool result. The approval payload sent to the model when a plan is accepted includes a structured
user_commentsarray:[{"line": 3, "text": "Use snake_case for field names"}, ...]. The model receives these alongside the approved plan and applies them as additional constraints. Advantage: explicit, structured, harder to miss. Disadvantage: requires API change.The controlled test you ran (model wrote "Comment test line A/B/C" then comments came back clean) is definitive evidence — the feature is completely broken, not just intermittently broken. All comments are silently discarded.
Fix priority: high. This is the only structured way for users to give line-level feedback without rewriting the plan. Silent discard of user feedback is worse than not having the feature at all.
I got the same problem
Ask me to approve the plan in the plan review mode.Type this as prompt to claude code and it will work
Can confirm this works as a workaround for now. If this is intentional its extremely counterintuitive.
This worked thanks. I think the issue is you perhaps accept a plan or steam past it and yet if it is open on the right hand side, is allows you to still add comments for a now accepted plan. I guess adding a comment should force a review of the plan.
+1 — reproduces on Windows (Claude Code desktop app, Opus 4.7 1M context).
Reproed today (7 May 2026). User left inline comments via the hover "Leave a comment for Claude…" UI on a plan file in the right Plan pane. Comments rendered locally with selection highlights — confirmed visible to user via screenshot. They never reached the model.
After re-entering plan mode and re-reading the plan file, the model returned byte-identical content to what it had originally written — zero markdown changes, zero comment annotations, no metadata sidecar accessible to any read tool.
Same behaviour as the OP describes for macOS, so the
platform:macoslabel probably under-scopes this — the gap is in the comment-delivery channel, not in OS-specific rendering.Strongly endorse suggested fix option (a) — merging comments as inline
<!-- user: ... -->annotations into the plan file on approval. It's the only option that survives:Currently the only working workaround the model can offer is "Ctrl+G → edit the plan markdown in your default editor" — which wastes the user's time relative to the hover UI they were told would work. Today's session burned ~4 turns of the user explaining + the model re-reading the plan file before the model figured out (via web search) that this was a known bug, not user error.
God this UI is something I really want to use... but these sort of bugs are like Anthropic didn't use the UI themselves. Another +1 for me. The workaround does work... but it's jus fundamentally broken UI.
think @bcherny ever looks at issues? I'm starting to just give up on providing feedback cause i feel like they go into this black hole... never to be addressed
Closing for now — inactive for too long. Please open a new issue if this is still relevant.
This appears to have been fixed. I just checked and claude was able to see inline comments now.
I'm on macOS 26.5 with VSCode 1.125.1 and it is still happening.
Still happening on MacOS 26.6 in Claude Desktop plan mode.
Still reproduces in July 2026 on the VS Code extension (0.1.10) — flagging in case it warrants reopening, since the impact is ongoing.
Surface: Plan View, "Select any text to leave a comment for Claude."
What happened: I left two inline comments anchored to text in the plan; both saved and persisted as ①/② markers in the editor. Neither reached the model — the assistant confirmed it received only the prose I typed in the chat box, never the anchored comment text. A third comment reproduced the sync loss from #43503: I could type it, but Plan View stopped updating and the comment was lost (I retried several times, same effect). Net effect: all three review comments were silently dropped; the only way through was to retype them in the chat.
Environment:
Screenshot: it shows that the first 2 comment-attempts seem to have worked:
<img width="1150" height="960" alt="Image" src="https://github.com/user-attachments/assets/48a1764f-6aa6-4a02-883d-79596c273a7a" />
Screenshot: I asked Claude whether it got my comments, and it said "No":
<img width="1658" height="262" alt="Image" src="https://github.com/user-attachments/assets/41084fea-062b-4e8f-92dd-844b6e21b7d8" />