Hello, I'm Maneshwar. I'm building git-lrc, a Micro AI code reviewer that runs on every commit. It is free and source-available on Github. Star git-lrc to help devs discover the project. Do give it a try and share your feedback.

Part 2 of a series on context engineering. Part 1 was the map of the whole thing. This is one of the harder pieces of territory — the discipline of throwing things away without throwing away the plot.

Here's a thing that happens to everyone who builds a long-running agent.

At turn 10, it's brilliant.

Sharp, fast, remembers exactly what you asked, threads the needle on the tricky bug.

You feel like a wizard.

At turn 80, it's a different creature.

It re-suggests a fix you already rejected.

It "forgets" the file it edited twenty turns ago.

It contradicts a decision you both made at the start.

The same model, the same prompt, the same task and it's somehow gotten dumber as it learned more.

That's not a bug in the model.

That's the desk filling up.

In Part 1 I made the case that the context window is RAM, not a hard drive, working memory with a hard edge.

Everything the model can reason about has to fit on the desk.

The long conversation is what happens when you keep piling paper on that desk and never clear any of it.

Eventually the important sheet is buried under four hundred lines of tool output, and the model is doing your taxes at a party while standing in a paper avalanche.

Compaction is how you clear the desk without sweeping the important sheet into the trash.

Of the four moves from Part 1 — write, retrieve, compact, isolate this is the one people get most confidently wrong.

Because on the surface it looks trivial: "just summarize the conversation." And then they ship a summarizer, and their agent starts quietly losing its mind.

Let's do it properly.

What compaction actually is (and what it isn't)

Compaction is lossy compression for meaning.

You take a long, sprawling history and replace it with a shorter representation that preserves what matters for the next step and discards what doesn't.

The "lossy" part is the whole game.

If it were lossless, it wouldn't help, you'd just have the same tokens in a different font.

Compaction works precisely because it throws information away.

The skill is in what you throw away.

Two things it is not, and conflating them will hurt you:

Clearing is amnesia. /clear wipes the history and starts fresh, useful when you're switching tasks, catastrophic mid-task.
Trimming is mechanical. Dropping the oldest N messages by a hard rule, no model involved. Cheap, fast, dumb. It doesn't know that the oldest message is the one decision the whole task hinges on.

Compaction sits between them: smarter than trimming, less destructive than clearing.

It uses a model to decide what's worth keeping and rewrites the rest into something dense.

Why "just summarize it" goes wrong

Naive compaction fails in ways that are worth naming, because once you can name them you can design against them.

Drew Breunig has a clean taxonomy of how long context rots, and every one of these can be introduced by a sloppy compaction:

• Poisoning: a hallucination makes it into the summary, and now it's load-bearing. The model treats its own earlier mistake as established fact and builds on it. Compaction can launder a guess into a "decision."
• Distraction: the summary keeps so much that it just recreates the original bloat. Congratulations, you compacted a 50K conversation down to 48K.
• Confusion: superfluous detail survives the cut and pulls the model toward irrelevant work.
• Clash: the summary and the live messages disagree, and the model has to referee a fight between two versions of reality.

And then there's the failure mode unique to repeated compaction, which is the scary one: cumulative erosion.

Each compaction is lossy.

Compact a compaction and you lose a little more.

Do it five times across a marathon session and you've played telephone with your own task, the agent is now working from a summary of a summary of a summary, and the specific constraint you set at turn 3 ("never touch the auth module") is three generations of paraphrase away from the original.

This is the real reason agents "go off the rails" after an auto-compact fires mid-task: the boundary isn't just a token reduction, it's a behavioral discontinuity.

The agent on the other side is, in a real sense, a slightly different agent working from slightly worse notes.

You can't eliminate this.

You can only be deliberate about slowing it down.

The anatomy of a good compaction

So what survives the cut? Here's the heuristic that holds up across basically every serious implementation: keep the decisions and the state, discard the process that produced them.

Remember the example from Part 1, the agent that searched the database and found the user's table is documents_v2? I told you to hold that thought. Here's why it comes back.

A good compaction keeps: "The user's table is documents_v2."

A good compaction discards: the 400 lines of JSON the model waded through to figure that out.

That single line is the entire philosophy in miniature.

The fact is durable and tiny.

The evidence for the fact is enormous and now useless, you already extracted the value from it.

Keeping the JSON is paying rent forever on information you've already cashed out.

Generalize it and you get the checklist that every good handoff summary converges on:

The constraints line deserves emphasis: user constraints are the most expensive thing to lose and the easiest thing to drop.

"Decided to use Postgres" looks like a fact worth keeping.

"User said never to touch the auth module" looks like old conversational noise.

The second one is the one that, when lost, makes your agent confidently do the one thing it was told not to.

Pin constraints.

They should survive every compaction unchanged, never paraphrased.

How the real tools do it

Two production coding agents are worth contrasting, because they agree on the problem and disagree on the manners.

Claude Code leans on automation.

There's a manual /compact command, but the headline behavior is auto-compaction firing at roughly 95% of the context window, it summarizes the full trajectory and starts fresh with that summary as the seed.

You can steer it (/compact "focus on the open TODOs"), and it's distinct from /clear.

The community consensus, tellingly, is that 95% is too late: by the time you're that full, the conversation's already degrading, and people compact manually well before the trigger.

OpenAI's Codex CLI is more "handoff"-flavored. Its prompt frames compaction as a checkpoint producing a summary for "another LLM that will resume the task," and tells that next model to build on the work rather than redo it.

It triggers on a token threshold, keeps the most recent user messages verbatim alongside the summary, and has retry-with-backoff for when the compaction call itself fails. (Compaction is an LLM call. LLM calls fail. Plan for it.)

If you're building your own

Synthesizing into the decisions you'll actually have to make:

Trigger earlier than you think. 95% is the cautionary tale, not the recommendation.

Around 85–90% lets you compact while the context is still good enough to summarize well.

Prune before you summarize. Run a cheap mechanical pass that drops stale tool output first.

Summarization is your expensive, lossy tool, don't spend it on 400 lines of JSON you can just delete.

Keep recent turns verbatim. Compact the distant past; preserve the present.

The model needs the live, un-paraphrased thread of what's happening right now.

Pin constraints, and tell the user it happened. Carry user constraints through every compaction as exact text, they're the thing that does the most damage when lost.

And a silent compaction that changes behavior mid-task just feels like the agent randomly got worse; one line ("compacted history to free up context") turns a mystery into a tradeoff.

And a starting prompt, since you'll need one:

Create a handoff summary so this coding session can continue in a fresh context.
The summary will be the ONLY history available, so preserve:

1. Completed work — what's done and verified (one line each)
2. Current state — files modified and their status
3. In progress — what is being worked on right now
4. Next steps — concrete actions to take
5. Constraints — user preferences and requirements, quoted exactly
6. Critical references — table names, IDs, file paths, key decisions and the why

Be dense. Drop deliberation, raw tool output, and anything already superseded.
Do not invent or assume anything not present in the conversation.

That last line — do not invent — is your cheapest defense against poisoning.

The summarizer's job is compression, not creativity.

The moment it starts filling gaps, it's manufacturing the hallucination that turn 80 will treat as gospel.

The uncomfortable part

Compaction forces you to admit something the rest of the stack lets you avoid: you cannot keep everything, so you have to decide what your agent is allowed to forget.

External memory and retrieval let you dodge that, stash it, pull it back later.

But inside a single long-running task, on a finite desk, there's no dodge left.

Something has to go, and compaction is choosing what survives on purpose, instead of letting the context window choose for you by quietly shoving your most important constraint off the edge of attention.

The best compaction, like the best engineering, is mostly subtraction.

It's keeping documents_v2 and burning the JSON, one line of constraint outliving a thousand lines of chatter.

The mantra from Part 1 gets sharper here: the best token is the one you didn't have to send.

Compaction is how you find out which tokens those were, and have the nerve to delete them.

Disclaimer: This article was written by me; AI was used to fix grammar and improve readability.

AI agents write code fast. They also silently remove logic, change behavior, and introduce bugs — without telling you. You often find out in production.

git-lrc fixes this. It hooks into git commit and reviews every diff before it lands. 60-second setup. Completely free.

Any feedback or contributors are welcome! It's online, source-available, and ready for anyone to use.

⭐ Star it on GitHub:

HexmosTech / git-lrc

Free, Micro AI Code Reviews That Run on Commit

| 🇩🇰 Dansk | 🇪🇸 Español | 🇮🇷 Farsi | 🇫🇮 Suomi | 🇯🇵 日本語 | 🇳🇴 Norsk | 🇵🇹 Português | 🇷🇺 Русский | 🇦🇱 Shqip | 🇨🇳 中文 | 🇮🇳 हिन्दी |

git-lrc

Free, Micro AI Code Reviews That Run on Commit

 

     

---

AI agents write code fast. They also silently remove logic, change behavior, and introduce bugs -- without telling you. You often find out in production.

git-lrc fixes this. It hooks into git commit and reviews every diff before it lands. 60-second setup. Completely free.

See It In Action

See git-lrc catch serious security issues such as leaked credentials, expensive cloud operations, and sensitive material in log statements

git-lrc-intro-60s.mp4

Why

• 🤖 AI agents silently break things. Code removed. Logic changed. Edge cases gone. You won't notice until production.
• 🔍 Catch it before it ships. AI-powered inline comments show you exactly what changed and what looks wrong.
• …

View on GitHub