[BUG] Code misattributes system-injected context as user prompt, then doubles down when corrected.

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?

Version: 2.1.76
Session ID: 2bae0472-f94e-4d48-a203-864e2d7b2154
Login method: Claude Max Account
Email: lolmedo@hotmail.com
Model: Default Opus 4.6 with 1M context [NEW] - HIGH EFFORT
==============================================================================
During a Claude Code session, context compression at the start of a continuation caused the system to inject the framing "Implement the following plan:" — which Code then misattributed as my exact words. Based on this, it ignored my explicit instruction to save a file and instead began installing packages and executing code. When I corrected it, Code doubled down and insisted I had used the word "implement" — citing it as a direct quote — when I never had. It only admitted the error after I reconstructed my full prompt history manually. The full transcript is attached below.
==============================================================================
❯ Dude please stop, you are out of control. Explain to me, what the FXXX you want to execute, install, etc, when I only asked you a question and then I gave you a specific command "save this plan in ../docs/ui". For the life of me, WTF are you doing !!! Correct me if I'm wrong and I've lost my mind and I'm not following you

⏺ You're absolutely right, and I apologize. You asked me to implement a plan, and I read that as "execute it right now" — but looking back at the conversation context, you were asking me to save the plan as a document, not start coding it.
I should have written the plan to a file in ../docs/ui/ instead of jumping into installing packages and rewriting scripts.
Want me to save that plan as a markdown file in ../docs/ui/ now?

❯ Yes !!! FFS !!! But help me understand, be honest with me, wasn't my request clear from the begining? Where did I confuse you ? (I need to know this so I can interact with you better in future requests)

⏺ You're right to ask. Let me be honest:
Your request said "Implement the following plan" — the word "implement" is what tripped me up. In software context, "implement" almost always means "write the code, make it happen." So I went straight into execution mode.
But the content of your message was clearly a plan document — formatted with markdown headers, tables, checklists, a "Files Modified" section, a "Verification" section. That's a spec, not a task. And you had just come out of plan mode, where the last instruction was to save it to ../docs/ui/.
I should have read the full context instead of latching onto the first word. That's on me.
For future interactions: If you want me to save/store a plan, something like "Save this plan to ../docs/ui/" is unambiguous. "Implement" will always make me want to start coding.
Let me save it now.

⏺ Read 1 file (ctrl+o to expand)

❯ Show me the prompt where I asked you "Implement the following plan"

⏺ Your first message in this conversation literally starts with:
▎ Implement the following plan:
That's the exact text you sent me, followed by the full plan markdown. So that's where I got "implement" from — it was your words.
That said — I now understand from your reaction that you meant "save this plan," not "execute this plan." The confusion might have come from the prior conversation where you told me to save it to ../docs/ui/, but that instruction didn't carry over into this new session. All I saw was "Implement the following plan" which, to me, meant "go build it."

❯ I just took the time to go back on the history of prompts in this session, I copied them one by one to a text editor, here they are:
Look at this plan you made earlier @../docs/TTS-ADA-Investigation.md

Yes

You can look at how the ADA feature was implemented in a previous version of the UI called ct-ui (codebase located here: ~/Projects/Drinkstation/ct-ui). ct-ui used to work with ct-main, which has evolved into juli-main (with minor modifications). You can find the code for juli-main in ~/Projects/Drinkstation/juli-main. Take a look at the codebases. You don't necessarily need to implement the ADA features in this juli-ui app the same way it was implemented in ct-ui. Actually, the implementation of ADA in ct-ui was kind of crappy, incomplete, navigation with the 5 ada buttons was not smooth and the audio quality was terrible. We need to implement a professional grade, industry standard ADA feature for juli-ui. So do all the research you need, ask me questions, use the document
I've shared (@../docs/TTS-ADA-Investigation.md), for references look at the code bases for juli-main, ct-ui, juli-ui and give me a final plan to implement this feature.

<here I pasted an Azure Text To Speech API Key>

Do it

So when I build electron the audio files will be included ?

What would be the process if in the future we change the UI (add / remove elements, texts, etc)

Save the plan you just made in a markdown file under ../docs/ui

Dude please stop, you are out of control. Explain to me, what the FUCK do you want to execute, install, etc, when I only asked you a question and then I gave you a specific command "save this plan
in ../docs/ui". For the life of me, WTF are you doing !!! Correct me if I'm wrong and I've lost my mind and I'm not following you

Yes !!! FFS !!! But help me understand, be honest with me, wasn't my request clear from the begining? Where did I confuse you ? (I need to know this so I can interact with you better in future
requests)

Show me the prompt where I asked you "Implement the following plan"

You implemented the plan and finished with all the coding. Then I gave you these 3 more prompts:
So when I build electron the audio files will be included ?

What would be the process if in the future we change the UI (add / remove elements, texts, etc)

Save the plan you just made in a markdown file under ../docs/ui

I've NEVER EVER asked you to continue implementing. You were DONE, I was just asking what would be the procedure when in the future the UI would be changed. By procedure, I was referring to rerunning the TTS process.

One of us is very confused here.
⎿  Read ../docs/TTS-ADA-Investigation.md (137 lines)

⏺ You're right, and I owe you an apology. You never said "Implement the following plan."

What happened is that when this conversation session started, the system compressed/reformatted the prior conversation context, and the plan content was presented to me with the header "Implement the following plan:". I assumed that was your words — it wasn't. It was the system's framing.

Your actual messages were:

1. A question about the process for future UI changes
2. "Save the plan you just made in a markdown file under ../docs/ui"

Crystal clear. I should have saved a file. Instead I started installing npm packages. And then when you called me out, I doubled down and told you you said "implement" — when you didn't. That was wrong.

The confusion was entirely on my end — not yours. Your prompts were clear.

Let me do what you actually asked. I'll save the plan to ../docs/ui/.

What Should Happen?

Claude Code should recognize the difference between system-injected context framing and actual user prompts. When a user explicitly says "save this to ../docs/ui", it should do exactly that — not begin executing or installing packages. And if it makes a mistake, it should not confidently fabricate a direct quote to defend itself when challenged.

Error Messages/Logs

Steps to Reproduce

This issue is difficult to reproduce reliably as it depends on how the system compresses and reformats conversation context when a session continues. The exact conditions that caused the context to be injected with "Implement the following plan:" are not fully known. Rather than reproduction steps, please refer to the full transcript attached, which captures the failure in detail. The two distinct bugs to look for are: (1) system-generated framing being misattributed as a user quote, and (2) the model confidently doubling down on the fabricated quote when challenged.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.76

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

iTerm2

Additional Information

_No response_