I hit this while pasting ~8K of DOM markup into Claude Code and iterating on selectors for a Chrome extension. About 40 minutes in, compaction fired and the summary said "user provided DOM markup" but the actual content was gone. Claude started guessing at selectors it had seen 20 minutes earlier.
The transcript with the original markup is still sitting at ~/.claude/projects/ as a .jsonl file — the compaction summary just has no pointer back to it.
I found 8+ open issues on the repo describing different symptoms of this same root cause. The proposal is to add line-range annotations to compaction summaries so Claude can surgically recover just the chunk it needs from the transcript on demand. Zero standing token overhead.
Curious if others have hit this in different scenarios or found workarounds that actually stick.
I'm fairly sure that Claude adds a note where it can find the original transcript after compactation (somewhat recently).
Fwiw I built a little CLI that could help with this, https://github.com/martinalderson/claude-log-cli. It allows Claude to search its own logs very efficiently. So I'm sure you could add something like "if the session is continued from a previous one, use claude-log cli to find users original prompt with claude-log" which would pull it out very efficiently. I built it to enable self improving claude.md files (link to the blog in the GitHub) but it's so useful for many tasks.
I honestly still don't see the point of compaction. I mean it would be great if it did work, but I do my best do minimize any potential for hallucination and a lossy summary is the most counterproductive thing for that.
If you have it write down every important information and finding along a plan that it keeps updated, why would you even want compaction and not just start a blank sessions by reading that md?
I'm kind of suprised that anyone even thinks that compaction is currently in any way useful at all. I'm working on something which tries to achieve lossless compaction but that is incredibly expensive and the process needs around 5 to 10 times as many tokens to compact as the conversation it is compacting.
Firstly, it's very useful to have your (or at least some) previous messages in. There's often a lot of nuance it can pick up. This is probably the main benefit - there's often tiny tidbits in your prompts that don't get written to plans.
Secondly, it can keep eg long running background bash commands "going" and know what they are. This is very useful when diagnosing problems with a lot of tedious log prepping/debugging (no real reason these couldn't be moved to a new session tho).
I think with better models they are much better at joining the dots after compactation. I'd agree with you a few months ago that compactation is nearly always useless but lately I've actually found it pretty good (I'm sure harness changes have helped as well).
Obviously if you have a total fresh task to do then start a new session. But I do find it helpful to use on a task that is just about finished but ran out of space, OR it's preferable to a new task if you've got some hellish bug to find and it requires a bunch of detective work.
I mean I agree the last couple of messages in a rolling window are good to include, but that is not really most of what happens in compaction, right?
> there's often tiny tidbits in your prompts that don't get written to plans.
Then the prompt of what should be written down is not good enough, I don't see any way how those tidbits would survive any compaction attempts if the llm won't even write them down when prompted.
>Secondly, it can keep eg long running background bash commands "going" and know what they are. This is very useful when diagnosing problems with a lot of tedious log prepping/debugging (no real reason these couldn't be moved to a new session tho).
I cannot really say anything about that, because I never had the issue of having to debug background commands that exhaust the context window when started in a fresh one.
I agree they are better now, probably because they have been trained on continuing after compaction, but still I wonder if I'm the only one who does not like compaction at all. Its just so much easier for an LLM to hallucinate stuff when it does have some lossy information instead of no information at all
AFIAK claude code includes _all_ messages you sent to the LLM in compactation (or it used to). So it should catch those bits of nuance. There is so much nuance in language that it picks up on that is lost when writing it to a plan.
I think your point doesn't hold up really. Telling an LLM to summarize something losslessly will loose so much more nuance than updating the plan directly every time when some useful information is gained.
That file is not even a plan but effectively a compaction as well, just better as its done on the fly only processing the last message(s) rather than expecting an LLM to catch all nuances at once over a 100-200k+ conversation.
Works fine for me in sessions that use a lot of context. My workflow is to keep an eye on the % that shows how soon it will auto compact. And either /clear and start over, or manually compact at a convenient place where I know it'll be effective.
i use https://github.com/sirmalloc/ccstatusline and when im around 100k tokens im already thinking about summarizing where we're at in the work so i can start fresh with it
it is pretty rare for me to compact, even if i let it run to 160k
--
just realized how i wouldn't think about using ccstatusline based a quick glance at its README's images. looks like this for me:
> I honestly still don't see the point of compaction.
Currently my mental model is every token Claude generates gets added to the context window. When it fills up there is no way forward. If you are going to get a meaningful amount of work done before the next compaction they have to delete most of the tokens in the context window. I agree after compaction it's like dealing with something that's developed a bad case of dementia, but you've run out what is the alternative?
> why would you even want compaction and not just start a blank sessions by reading that md?
If you look at "how to use Claude" instructions (even those from Anthropic), that's pretty much what they do. Subagents for example are Claude instances that start set of instructions and a clean context window to play with. The "art of using Claude" seems to be the "art of dividing a project into tasks, so every task gets done without it overflowing the context window".
This gives me an almost overwhelming sense of déjà vu. I've spent my entire life writing my code with some restriction in mind - like registers, RAM, lines of code in a function, size of PR's, functions in a API. Now the restriction is the size of the bloody context window.
> I'm working on something which tries to achieve lossless compaction but that is incredibly expensive and the process needs around 5 to 10 times as many tokens to compact as the conversation it is compacting.
I took a slightly different approach. I wanted a feel for what the limit was.
I was using Claude to do a clean room implementation of existing code. This entails asking Claude to read an existing code base, and produce a detailed specification of all of its externally observable behaviours. Then using that specification only (ie, without reference to the existing program, or a global CLAUDE.md, or any other prompts), it had to reliably produce a working version of the original in another language. Thus the specification had to include all the steps that are needed to do that - like unit tests, integration tests, coding standards instructions on running the compiler, and so on, that might normally come from elsewhere.
Before proceeding, I wanted to ensure Claude could actually do the task without overflowing its context window - so I asked Claude for some conservative limits. The answer was: a 10,000 word specification that generated 10,000 lines of code would be a comfortable fit. My task happened to fit, but it's tiny really.
When working with even a moderate code base, where you have CLAUDE.md, and a global CLAUDE.md for coding standards and what not and are using multiple modules in that code base so it has to read many lines of code, you run into that 10,000 words of prompt, 10,000 lines of code it has to read or write very quickly - within a couple of hours for me. And then the battle starts to split up the tasks, create sub-agents, yada-yada. In the end, they are all hacks for working around the limited size of the context window - because, as you say, compaction is about as successful for managing the context window as the OOM killer is for managing RAM.
It's interesting that you find compaction trivial. I think it's one of the most important tasks, to the point where I use Amp these days because its "handoff" feature is so much nicer than CC's compaction.
It's a lot of little things + polish more than any big change.
Amp has a first-class concept of threads, a tree of sessions. This is really nice for long work on related features, it tracks which threads are spawned from others. When you type /handoff, it asks you for a goal for the new thread, then summarizes your existing context with respect to that goal, and opens a new thread with just that context.
This makes it really easy and pleasant to spin up new sessions to do relatively focused tasks, which keeps your context usage low and the model smarter. It also enables some really nice use cases like opening up an old thread where you built a feature two weeks ago, then spinning up a new one to do a modification.
You can do all of this with Claude Code but it's just clunkier and in my experience hasn't worked nearly as well, e.g. I find the compactions tend to be full of a lot of useless stuff, or miss something important compared to the handoffs.
is it similar to the concept of "branch off from here" feature in some chat UIs where you can continue one convo in different directions? but does amp keep each thread in a separate worktree/isolated env and let you choose which one to merge?
set Anthropic base URL in CC to your proxy server and map each model to your preferred models (I keep opus↔opus but technically you can do opus↔gpt-5.3, etc.). then check the incoming messages for the string that triggers compaction (it's a system prompt btw) and modify that message before it hits the LLM server.
The transcript with the original markup is still sitting at ~/.claude/projects/ as a .jsonl file — the compaction summary just has no pointer back to it.
I found 8+ open issues on the repo describing different symptoms of this same root cause. The proposal is to add line-range annotations to compaction summaries so Claude can surgically recover just the chunk it needs from the transcript on demand. Zero standing token overhead.
Curious if others have hit this in different scenarios or found workarounds that actually stick.
reply