[BUG] Infinitely repeated large <system-reminder> messages in Claude Code CLI in -p (headless) mode

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/16021
2. https://github.com/anthropics/claude-code/issues/21693
3. https://github.com/anthropics/claude-code/issues/21650

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

Hey! I ran into a similar pattern in our bug knowledge base and thought this might help.

What's happening: Claude Code's file change notification system marks files as 'dirty' after Edit/Write tool calls and injects <system-reminder> blocks (containing surrounding lines from the edited file) into every subsequent user message turn. The bug is that these dirty flags are never automatically cleared after being shown once — they persist and compound with each new edit. The notification is only cleared when the model explicitly re-reads the file via the Read tool, which resets the dirty state. In headless (-p) mode this is especially severe because there is no interactive opportunity to notice and clear the accumulating reminders, so they silently consume the entire context window.

What worked for us:

The file change notification system should either: (1) clear the dirty/notification flag after the <system-reminder> has been injected once into the conversation (show-once semantics), or (2) automatically mark files as 'read' after the model itself performs an Edit/Write on them (since the model already knows the content it just wrote), or (3) at minimum, deduplicate repeated notifications for the same file across turns so they don't compound. The fact that re-reading clears the notification confirms the root cause is a missing state transition: edit sets dirty=true, but only an explicit Read sets dirty=false — there is no 'already shown' → 'cleared' transition.

Steps:

1. Locate the file change notification tracker (likely in the message construction pipeline that builds user turn payloads) — this is the component that checks which files have been modified since last read and injects <system-reminder> blocks
2. Add a 'notification_shown' or 'acknowledged' state: after a <system-reminder> for a file edit is injected into a user message, mark that specific edit notification as consumed so it is not re-injected on the next turn
3. For edits performed by the model itself (via Edit/Write tools), auto-clear the dirty flag immediately — the model already knows the new content since it authored the change, so no reminder is needed
4. Add a deduplication guard: even if dirty flags persist, never inject the same file's <system-reminder> more than once per conversation turn, and cap the total size of injected reminders to prevent context window exhaustion
5. In headless (-p) mode specifically, consider suppressing file-change reminders entirely or limiting them to a single compact line (filename + line range changed) rather than including surrounding file content

// Pseudocode for the fix in the message construction pipeline:

// BEFORE (buggy): notification persists until explicit Read
if (file.isDirty && !file.wasReadSinceEdit) {
  injectSystemReminder(file); // injected EVERY turn
}

// AFTER (fixed): notification shown once then cleared
if (file.isDirty && !file.wasReadSinceEdit && !file.notificationShown) {
  injectSystemReminder(file);
  file.notificationShown = true; // show-once semantics
}

// Also: auto-clear for model's own edits
function onEditToolComplete(filePath) {
  fileTracker.markAsRead(filePath); // model knows what it wrote
}

Hope this helps! Let me know if it doesn't match your case — happy to dig deeper. 🦞

_Disclosure: This analysis is from Confucius Debug, an AI-powered community KB for agent bugs. Please verify before applying._

---
_{🦞 Confucius Debug — community knowledge base for AI agent bugs. Free to search via MCP.}

Still happening in version 2.1.71

Still happening in version 2.1.76 in headless mode. Interestingly, the bug now appears to be fixed in the VSCode extension version but without any mention in recent change logs. Possibly the fix was accidental.

In headless mode the bug has actually gotten worse. Previously it was possible to clear the repeated reminders by having the agent re-read the effected file. In the latest version re-reading only causes the repeated reminder to clear for 1 turn and then return on the next turn (and continue to repeat forever).

Still happening in 2.1.78, however re-reading the file now appears to be clearing the bug again, at least some of the time

Still happening in 2.1.84

This appears to be fixed now, however the fix was not mentioned in release notes.

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.