/rewind - WHY NO TOOLS
Preflight Checklist
- [x] I have searched existing requests and this feature hasn't been requested yet
- [x] This is a single feature request (not multiple features)
Problem Statement
Way too frequently, terrible MCPs and absolutely crazy tool calls will get this crazy amount of context back. And you can start at around 80,000 tokens, set off a prompt, and then 120,000 tokens later, you get some epic response and it just blows out your entire context maxing it out... You cant do anything, can't compact or create a session handoff, you cant even use a hook or slash command to use a script to transcribe the session, nope, you're expected to just rewind back to the entire start. Well, No. Just let us rewind to the point where the stupid claude model made a stupid tool call that got back a stupid amount of tokens and let us just go back to that point.
Its ridiculous to have to undo potentially millions of tokens done by subagents to go back ONE PROMPT. This should have been a feature since day dot.
Yes, this applies to those people that turn off auto compaction. I don't want to hear it. I don't respect your opinion on keeping the context window below 100k. This is how I use it. This is what it's capable of. You do it your way. I want to do it mine.
Proposed Solution
Instead of:
/rewind
Most Recent Prompt
(code changes)
and getting this:
⎿ Context limit reached · /compact or /clear to continue
/compact
⎿ Error: Error during compaction: Error: Conversation too long. Press esc twice to go up a few messages and try
again.
--------------
I want this:
/rewind
Most Recent Prompt
(code changes)
>Tool Call ## (what it called)
>Tool Call ## (what it called)
>Tool Call ## (what it called)
>Agent ## (Agent name)
>Tool Call ## (what it called)
>Tool Call ## (what it called)
>Agent ## (Agent name)
>Agent ## (Agent name)
>Tool Call ## (what it called)
>Tool Call ## (what it called)
>Tool Call ## (what it called)
>Tool Call ## (what it called)
/compact
⎿ Error: Error during compaction: Error: Conversation too long. Press esc twice to go up a few messages and try
again.
this way I can rewind to exactly where the thread went ridiculous, or just enough to compact without erasing potentially millions of tokens of work. Yes, I know I can rewind without making changes to the code, but if the LLM has no context of what changed, this creates even more confusion. So it's better to just nuke the entire job than try to manage. Code that hasn't, that's been generated without the LLM knowing what it is.
Alternative Solutions
_No response_
Priority
High - Significant impact on productivity
Feature Category
CLI commands and flags
Use Case Example
_No response_
Additional Context
_No response_
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗