[MODEL] Claude repeats its previous response after an idle pause and stops reading the newest message

### Preflight Checklist

- [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Amodel) for similar behavior reports
- [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

### Type of Behavior Issue

Claude ignored my instructions or configuration

### What You Asked Claude to Do

No specific prompt triggers it. It happens in normal conversation.
The trigger is not the prompt, it is the pause: I leave the conversation idle for a while, from minutes to hours, then come back and send a new message. Any message.
Environment: Claude Code desktop app on Windows 11, Opus 5. Also reproduced on Opus 4.6, so it is not model-specific. I use voice output.

### What Claude Actually Did

1. I send a new message after the idle pause.
2. Claude answers with the content of its PREVIOUS response, repeated, instead of answering the new message.
3. Telling it "you are repeating yourself" inside that same conversation does nothing. It keeps repeating.
4. It never recovers. The only workaround is opening a brand new chat and losing the whole working thread.
5. With voice output the effect is worse: the voice re-reads the previous answer before the new one, and it stacks up.
Frequency: daily since 2026-08-13. Ten separate conversations in nine days. The worst session had around 850 messages, which fits the context-summarisation theory: the idle pause seems to reload the session from the context summary, and something breaks there.
Note: the existing issues about repeated text (#10533, #52020, #60829) are terminal rendering bugs, where the same text is re-painted on screen. This is not that. This is the model generating its previous answer again as its actual response.
Accessibility impact: I use voice output because reading is difficult for me. It is not a convenience, it is how I use the product. When this bug hits, the product becomes unusable for me, not merely degraded.

### Expected Behavior

Claude should have answered the new message I just sent, using the conversation context, instead of re-emitting its own previous response.
If the session context is reloaded from a summary after an idle period, that reload should not cause the model to treat its own last answer as the thing to produce again.

### Files Affected

```shell
None. This is not a file-modification issue.
```

### Permission Mode

I don't know / Not sure

### Can You Reproduce This?

Sometimes (intermittent)

### Steps to Reproduce

1. Start a normal conversation in the Claude Code desktop app and work in it for a while, until the thread is long.
2. Leave the conversation idle. Minutes or hours, without sending anything.
3. Come back and send a new message. Any message.
4. Claude answers with its previous response again instead of answering the new message.
5. From that point it never recovers within that conversation.
Longer threads make it far more likely. It is much rarer in a short, freshly opened chat.

### Claude Model

Opus

### Relevant Conversation

```markdown

```

### Impact

High - Significant unwanted changes

### Claude Code Version

2.1.232 (Claude Code)

### Platform

Anthropic API

### Additional Context

The strongest clue I have: the bug correlates with idle time, not with any particular prompt. Continuous conversations are fine. Conversations resumed after a pause are the ones that break. This points at whatever happens when a session is rehydrated from its context summary.
It also correlates with thread length. Short, freshly opened chats almost never show it.
Once the loop starts, there is no in-conversation recovery. The user cannot break it from inside.

No source files, tests, or entry points are identified; start by reproducing the issue in the Claude Code desktop app using a long conversation, an idle pause, and a new message. Compare resumed sessions with continuous and short chats, including voice output. Done means the new message is answered with the current conversation context and the previous response is not repeated or re-read.

Written by the indexing model from the issue text.