Claude Code: plan annotations are silently dropped when the plan is approved (approve path never sends feedback to the agent)

[mangxy]

Summary

When reviewing a plan in the annotation UI for Claude Code, any annotations added before clicking Approve are silently dropped — the agent never sees them. Annotations only reach the agent when the plan is denied. The Gemini CLI approve path already forwards feedback via systemMessage; the Claude Code approve path does not.

Environment

• plannotator 0.27.4 (also present in the 0.27.3 source I checked)
• Claude Code on WSL2, PermissionRequest hook on ExitPlanMode

Steps to reproduce

1. Enter plan mode in Claude Code, have the agent submit a plan (ExitPlanMode).
2. The plannotator review page opens. Add one or more line annotations.
3. Click Approve (with annotations still present).
4. The plan is approved and execution continues.

Actual behavior

The agent receives only the allow decision. None of the annotations are delivered into the conversation — the agent never learns the feedback existed. The annotations are written to <plans-dir>/<plan>.annotations.md, but nothing surfaces them to the agent.

Expected behavior

On approve, forward the collected feedback to the agent (e.g. as systemMessage), mirroring the Gemini CLI path.

Root cause

server/index.ts (0.27.4):

• Gemini approve path (line 2249) — feedback IS forwarded:

  console.log(result.feedback ? JSON.stringify({ systemMessage: result.feedback }) : "{}");

• Claude Code approve path (lines 2263–2292) — outputs hookSpecificOutput with decision.behavior: "allow" (+ updatedInput / updatedPermissions) but never includes result.feedback. Only the deny branch of the Claude Code path includes feedback (as decision.message).

Suggested fix

In the Claude Code approve branch, attach the feedback, e.g.:

console.log(
  JSON.stringify({
    hookSpecificOutput: {
      hookEventName: "PermissionRequest",
      systemMessage: result.feedback || undefined,
      decision: { /* existing allow output */ },
    },
  })
);

The current UX makes "annotate + approve" (notes for the agent to keep in mind while executing) indistinguishable from "plain approve", which reads as data loss rather than intent.

Workaround

Deny the plan to have feedback delivered, or manually paste the contents of <plan>.annotations.md into the conversation.