I reverse-engineered Claude Code's binary to add a feature I always wanted: When context fills up - instead of nuking everything with /compact - I can now surgically strip tool calls/results and thinking blocks while keeping all actual messages intact.

The problem was: when you're deep in a session, context is 96% full, and your options are: - /compact: summarizes everything, loses detail, costs an API call - /clear: nuclear option, start over I wanted to strip the heavy garbage and leave only user/agent messages instead.

So I cracked open the binary. Claude Code is a Bun-compiled executable: ~183MB. All JS is embedded in a __BUN Mach-O section - 4,945 modules, fully minified. Even with `--bytecode`, Bun always stores the full JS source alongside. The bytecode is just a startup optimization.

What I found inside: Claude already HAS a microcompact() system. It runs before every API call, strips old tool results, saves them to disk, replaces images with placeholders. But it doesn't touch everything and leaves many of the thinking intact...

It has limitations: - 20K token minimum savings - only kicks in near context limits - keeps last 3 tool results - never removes tool_use blocks. Even after it runs, tons of tool artifacts remain in context. That's by design, but it's not enough when you're running out of space.

There's also a thinking cleanup before every API call: - filterOrphanedThinking(): removes messages that are ONLY thinking - removeTrailingThinking(): strips thinking from end of last message But thinking blocks inside messages with tool_use? Those survive. They pile up.

So I built two new commands: /microcompact: strips ALL tool calls + results + media. No thresholds, no "keep last 3". Picker lets you choose how far back to strip. /clear-thinking: strips ALL thinking blocks backward from selected message. Both instant with no LLM calls.

The meta part: I reverse-engineered Claude Code using Claude Code :) 48 sessions over several days. Built 20 tools: binary extraction, module splitting, vendor fingerprinting, AI renaming (23,105 identifiers via Gemini for ~$2), and a full round-trip pipeline.

I also built bun-demincer - a generic toolkit for cracking open ANY Bun-compiled binary. Extract embedded JS, split into modules, identify npm packages (fingerprint DB is 1,668 packages rn), deobfuscate, organize into directories and reassemble it back!

The irony: Claude couldn't finish the job - started forgetting things after another Plan mode run :) After 42 sessions it hit a wall kept looping the full pipeline instead of isolating the failing stage. Classic. Switched to Codex for the final stretch. One shot of course :)

For the Claude-specific patches (microcompact, clear-thinking) - I'm holding off on publishing them for now. Don't want to step on anyone's toes, unless @AnthropicAI are cool with it. cc: @alexalbert__

But would love these to be built-in honestly - the code paths are already there, just need manual triggers and looser thresholds. Also /compact should let you pick a message: summarize everything before it, keep everything after. Right now it just nukes the whole conversation.

All this reverse engineering really helped me to understand what's under the hood of Claude Code. I will be posting more about my explorations... Hold on for new Memory threads!

@vicnaum @banteg @claudeai everything with /compact - I can now surgically strip tool calls/results and thinking blocks while keeping all actual messages intact. You’d be destroying the cache which will cost you tons of tokens and make the model less reliable

@banteg @claudeai Yes. But it still wins over going in circles with the model that forgot what project it was working on... And actually, If I strip 195k tokens context into 5k (just user/agent conversation) - then who cares about cache?

@vicnaum @banteg @claudeai Mainly that it will usually just reread all the files you stripped out again anyway. I’m not arguing with you it’s something I was interested in exploring too but I never quite got it right

@banteg @claudeai I usually ask it to use sub-agents for file reading & modifications, and report back. but sometimes it still does things itself..

@vicnaum @claudeai I built this thing maybe this will also be helpful for you @vicnaum Basically per git repo backup of AI convos.

@claudeai Oh yeah! I had a simple skill with a script for that: But yours is on a whole new level!

@vicnaum @claudeai Cool hack, but we're essentially writing manual garbage collection for LLM context now. /compact is a grenade. This is a scalpel. But the real question is why we're still doing memory management by hand in 2026

@claudeai Exactly! Memory and context management is in such an infancy phase right now that I’m surprised this is still an untackled problem

@vicnaum @SucciMas @claudeai We’re trying @noodl_box

@vicnaum @claudeai @noodl_box How is it going?

You might've heard about this already, but this flow does the same thing and it's built in (keeps the messages, strips the tool calls): 1- Do /export and select md file 2- /new to start a new conversation 3- paste that md's path as context in the chat 4- Send any request and the previous messages will be in context The issue with this trick (and /microcompact) is you get deeper and deeper into the dumb zone of the context after a few cycles of this since you're keeping all the chat messages even the unnecessary ones. Still great job on bun-demincer.

@vicnaum @claudeai The model should be deciding this on its own.

@claudeai Ideally yes!

@vicnaum @claudeai Yep, context adjustments are the single biggest unlock and why you should generally create and use your own harnesses. OpenAI if you are reading this, you are missing out by not enabling more functionality here in the vanilla instance

@vicnaum @claudeai rip limits

@vicnaum @claudeai there's dynamic context pruner for opencode that does the same automatically during the session

@vicnaum @banteg @claudeai That’s great. The tool calls take the most space

@banteg @claudeai exactly! It was around 150k of tool calls - crazy..

@vicnaum @claudeai @bcherny add this to Claude Code

@vicnaum @claudeai You reverse engineered an open source binary?

@claudeai Its not really opensource - its still minified (later I found it on NPM, but would had saved me a few minutes on extracting the JS from bun-binary)

@vicnaum @claudeai yesss i’ve also often felt the need for a saving grace when at 90% context. you could, alternatively, build a script that runs on the transcript’s JSONL, without touching the Claude binary?

@claudeai Yeah, thats what I was doing before, and also asked another instance of Claude to break it down to me, see what takes most context and help us decide what to trim.

@vicnaum can you post the script version?

i was doing it on 2.1.63, but many new versions came out. I still have the transfer functionality to move my patches to newer versions - but havent tried it on latest yet. will probably post the patch soon

@vicnaum @claudeai How did you give Clawd a hat???????? This is very important information I need.

@claudeai I painted it in :) but pistol didn't turn out that good Wanted to make it a pirate )

@vicnaum @claudeai Fascinating work. ~5000 modules is revealing CC as sloppy patchwork. Pi agent main binary has ~20 direct deps and <400 unique transitive packages. Comparing apples to oranges here in terms of functionality out of the box. Just to illustrate how bloated CC is

@claudeai not really, its more of minification artifacts, I’d say - we can never tell what the original source was

@vicnaum @claudeai Thanks for clarifying. We think that the number is plausible from React ink deps alone.

@vicnaum @claudeai Mastra has an interesting concept called Observational Memory where 2 external models curate the transcript at runtime which basically results in unlimited context, preserving important details

@vicnaum @claudeai Context limit hits, most people restart the whole conversation. This guy cracked open the binary and built what he needed. That’s the real AI skill gap

@vicnaum @claudeai @claudeai hire this person

@vicnaum @claudeai anthropic ships this as a feature in 2 weeks and this becomes the most well-documented wasted weekend ever

@claudeai Let them do :) I have much more features in my mind - if that’s the only way I can contribute to Claude - so be it 😁

@vicnaum @claudeai The number of ~5000 dependencies for a CLI is staggering. Worth a good rant:

@vicnaum @claudeai Hack the tool until it fits the workflow

@vicnaum @claudeai I’ve asked Claude to build a .md context file from current session and load it into a new session. It’s simple and I’ve had great results doing that.

@claudeai Yeah, I was also doing that, and even have a skill for this: But I wanted something native, and still keep the tool-calls/thinking from some later messages.

@vicnaum @claudeai That's classic.. bro you're lucky you didn't have to go through what I did back then haha

@vicnaum @claudeai Yo, this is genius. Well done bro 👏

@vicnaum @claudeai It sounds like you turned a feature limitation into an opportunity for more precision. Reminds me of when I had to deal with similar limitations in my app, and I ended up designing a custom data compression algorithm to prioritize certain data types.

150k tokens of tool calls sitting in context is wild. The fact that Claude already has microcompact() internally but caps it with a 20K minimum and "keep last 3" feels like they were being cautious about cache invalidation costs. Stripping everything down to 5k of pure conversation and letting the model rebuild its understanding from that is a bold tradeoff, but honestly probably better than watching it loop on forgotten state. The bun-demincer tool sounds more generally useful than the Claude patches themselves.

@vicnaum @claudeai

@vicnaum @claudeai the dawn of self-editing code

@vicnaum @claudeai Very cool feature.

@vicnaum @claudeai Cool!

@vicnaum @claudeai Can I install your version of claude code with this feature? Absolutely need it

@claudeai i think I will post a patch soon that anyone can install

@vicnaum @claudeai 😂😂😂

@vicnaum @claudeai This sounds like it would have side affects? Like dementia

@claudeai not really, quite the opposite - if you compact at this stage - it will forget much more than if you keep the conversation intact, but remove the tools

@vicnaum @claudeai

@vicnaum @claudeai Smart approach. The built-in compaction heuristics always felt too conservative for long sessions. Surgically stripping tool artifacts while keeping the conversation intact is the fix power users need.

this solves my biggest claude code frustration tbh lost count of how many times ive hit 96% context in a deep debugging session and had to choose between /compact (lose everything) or start over the message picker approach is perfect. sometimes you just need to clear tool noise from the middle of a session, not nuke the whole conversation

@claudeai Here I've made it a skill also: