<div align="center">

# lcc — Local Context Compiler

**Compact coding-agent context without summarizing it away.**<br>
lcc cuts what the model does not need, keeps everything else byte for byte, and writes a receipt.

[![PyPI](https://img.shields.io/pypi/v/local-context-compiler.svg)](https://pypi.org/project/local-context-compiler/)
[![Python](https://img.shields.io/pypi/pyversions/local-context-compiler.svg)](https://pypi.org/project/local-context-compiler/)
[![CI](https://github.com/lucasmartins-ai/lcc/actions/workflows/ci.yml/badge.svg)](https://github.com/lucasmartins-ai/lcc/actions/workflows/ci.yml)
[![MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
[![GitHub stars](https://img.shields.io/github/stars/lucasmartins-ai/lcc?style=social)](https://github.com/lucasmartins-ai/lcc/stargazers)

<img src="https://raw.githubusercontent.com/lucasmartins-ai/lcc/main/demos/stash.gif" alt="lcc measure-session on a 91-message Claude Code session, then lcc compact with the stashed mode, then lcc recall returning the exact lines of a removed file read" width="820">

</div>

<!-- mcp-name: io.github.lucasmartins-ai/lcc -->

## `/compact` vs lcc

Claude Code's `/compact` replaces your session with a summary written by a model. Summaries
drop the exact error, path or line the agent needs two turns later. lcc does the opposite: old
tool outputs leave the context but stay one `recall` away, and every user and assistant message
is kept as it was.

| One real Claude Code session (193 messages, 75 tool calls) | tokens | user and assistant messages | removed tool outputs |
| --- | ---: | --- | --- |
| before | 57,083 | 43, as written | — |
| `/compact` | one summary | rewritten into the summary | gone |
| **lcc** (`rules`, `stashed`) | **28,807 (−49.5%)** | **43 / 43 byte for byte** | 33, each one `recall` away |

Measured with `lcc measure-session` (PyPI 1.2.0, offline, exact `o200k_base` token count;
re-run 2026-10-23). On `main`, whose stash note now states lines and size, the same session
gives 28,983 (−49.2%). That session is one of the best cases. **Across all 299 of my long sessions** (100+ messages each,
36.3 M tokens) the median cut was **21.8%** (middle half 15% to 31%, best 65%), with 0 messages
lost (`python benchmarks/research/long_sessions.py`, 2026-10-23, `main`). These are all my sessions: [run it on yours](#send-us-your-numbers) and post the result.

In a pre-registered test on 172 real requests ([evidence](docs/EVIDENCE.md)), an agent answering
from the `stashed` context plus `recall` did as well as with the full context on file recall
(+0.006, 95% CI −0.023 to +0.034) with 36.6% fewer tokens. It called `recall` for 90% of the
questions about removed content. A `/compact`-style summary of the same sessions got 73% of exact values
right (our reimplementation of `/compact`, written by Opus); full and verbatim context got 98%.

## Install

```sh
pip install "local-context-compiler[tiktoken]"   # the CLI is `lcc`
lcc measure-session "$(ls -t ~/.claude/projects/*/*.jsonl | head -1)"   # your latest session
```

The second line reads your most recent Claude Code session, runs every mode offline and prints
counts only. Nothing is changed and nothing leaves the machine.

**Claude Code plugin** (replaces the `/compact` summary). Function hooks are early access, so
this needs **Claude Code 2.1.274+ and the flag below**:

```sh
export CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1
claude plugin marketplace add lucasmartins-ai/lcc
claude plugin install lcc@lcc --config provider=rules --config rulesMode=stashed
```

Allow `mcp__plugin_lcc_lcc__recall` so the agent can get removed outputs back. When a pass
fails or removes too little, the plugin hands over to the built-in summary and says so in a
toast. Details: [Claude Code guide](docs/CLAUDE_CODE.md).

**MCP** (Claude Code, Cursor, Claude Desktop, any stdio client):

```json
{"mcpServers": {"lcc": {"command": "lcc", "args": ["mcp"]}}}
```

Tools: `compact`, `compact_transcript`, `recall`, `inspect`, `prepare`, `explain`, `intake`
([MCP guide](docs/MCP.md)).

**CLI**, on a session file or any document:

```sh
lcc compact ~/.claude/projects/<project>/<session>.jsonl -q "finish the parser fix" \
  --mode tool-calls --provider rules --rules-mode stashed -o compacted.json -r report.json
lcc recall <id>                                   # a removed output, hash-checked

lcc compact notes.md -q "what is wrong with the booking flow?" \
  --provider mechanical -o compacted.md -r report.json
lcc explain report.json --source notes.md         # why each block was kept or dropped
```

## Send us your numbers

```sh
lcc measure-session ~/.claude/projects/<project>/<session>.jsonl --markdown
# Codex CLI rollouts are auto-detected from their session_meta record:
lcc measure-session /path/to/rollout.jsonl --markdown
# or a whole project directory: per-mode median and p90 reduction over its sessions
lcc measure-session ~/.claude/projects/<project>/ --markdown
```

It runs every mode on one of your sessions, offline, and prints **counts only**: messages, tool
calls, tokens before and after, messages kept, time. No text, path or tool output. For exact
counts, install the `tiktoken` extra and cache its tokenizer once (lcc itself never downloads):
`python -c "import tiktoken; tiktoken.get_encoding('o200k_base')"`. Otherwise the report says
`approximate` and why. Paste it in
[Discussions → Results](https://github.com/lucasmartins-ai/lcc/discussions/categories/results).
Add `--input-map` to see where the session's billed input goes instead: cache reads and
writes, and the share taken by tool results, tool-call inputs, injected instructions and the
fixed system prompt and tool schemas (counts only, from the usage records in the file).
Already compacting with lcc? `lcc share-stats` turns your local reports into the same kind of
counts-only summary ([SHARE_STATS.md](docs/SHARE_STATS.md)), or use the
[results issue form](https://github.com/lucasmartins-ai/lcc/issues/new?template=results.yml).
Every number from someone other than me makes the claims above less mine and more true.

## Why lcc

- **Verbatim.** Kept spans are byte-identical to the source. Nothing is summarized or rewritten.
- **Recoverable.** `stashed` moves removed outputs to a local content-addressed store
  (`~/.cache/lcc/stash`); `recall` re-hashes on read.
  The placeholder, `[removed: 120 lines, 4.3 KB · lcc recall <id>]`, states lines and size, not
  tokens, so it is identical on every machine.
- **Auditable.** Every pass writes a JSON report; `lcc explain` gives the reason for every drop.
- **Local.** The default judges run on your CPU in plain Python. No API key, no network.

## Providers

| Provider | Runs | When it is used | Report says |
| --- | --- | --- | --- |
| `specialist-v3` | local, CPU | default (`auto`) whenever transcript blocks are present: the Claude Code hook, `compact_transcript`, `--mode tool-calls` | `semantic_guarantee: none` (a rank, not a judgment) |
| `mechanical` | local, CPU | `--provider mechanical`; plain text under `auto` (a configured TypeSafe key does not change that unless `LCC_AUTO_REMOTE=1`) | `semantic_guarantee: none` |
| `rules` (`lossless`, `recoverable`, `stashed`) | local, no model | `provider: rules` for transcripts | `lossless`: only provable duplicates removed; `recoverable`: also outputs that can be re-read; `stashed` (recommended): removed outputs stay byte-recoverable through `recall` |
| `jev` | remote (TypeSafe API) | opt-in: `--provider jev` / `provider: jev`, needs `TYPESAFE_API_KEY` | semantic judgment or explicit degradation |
| `laya` | local model | deprecated; runs only when named | named model and fallback state |

`specialist-v3` details and limits: [SPECIALIST.md](docs/SPECIALIST.md). Rules modes:
[TOOL_CALLS.md](docs/TOOL_CALLS.md). Jev is an optional path you name explicitly
(`--provider jev`); it sends the objective and context blocks to the TypeSafe API, and
`LCC_DISABLE_NETWORK=1` disables it. `auto` never picks it on its own: having a key is not
consent to send context off-machine, so plain text under `auto` stays on `mechanical` unless
you set `LCC_AUTO_REMOTE=1`. `compact` writes a `relevance-compaction-1.2` report;
check `provider_used` and `degraded` before trusting the output.

## Benchmarks

**Real sessions (held-out, INFERIDO).** In plain words: when a ranker has to drop part of a
session, how often does it keep what the agent actually used next? 1.0 would be always. A
one-shot final test on 193 items from 36 of the maintainer's sessions that started after the
training cutoff. Retention AUC over drop budgets of 0–60% of candidate tokens:

| ranker | recall AUC | all-kept AUC |
| --- | ---: | ---: |
| specialist-v3 | **0.899** | **0.434** |
| mechanical | 0.829 | 0.238 |
| recency | 0.818 | 0.306 |
| bm25 | 0.803 | 0.270 |
| random | 0.689 | 0.138 |

The labels are next-action evidence from one maintainer's Claude Code sessions, which favours
v3's training target. The sessions are private; only aggregates are published
([SPECIALIST.md](docs/SPECIALIST.md)).

**Synthetic, reproducible** (`python benchmarks/run_judge_bench.py`, about 20 s, no network).
Nine hand-written edge cases: supersession, contradiction, config change, exception rule,
pt/en mixed, prompt injection in a tool output, large git diff, old root cause, re-read.

| raw ranking | specialist-v3 | mechanical | bm25 | recency |
| --- | ---: | ---: | ---: | ---: |
| cases with every must-keep block kept at a 30 / 50 / 70% cut (of 9) | 8 / 7 / 7 | 9 / 9 / 9 | 9 / 9 / 7 | 8 / 6 / 3 |

| full pass, default settings | must-keep kept | 302 blocks | 1,138 blocks | peak RSS |
| --- | ---: | ---: | ---: | ---: |
| `specialist-v3`, tool-call mode | 9/9 | 181 ms | 3.4 s | 71 MB |
| `rules` lossless | 9/9 | 2 ms | 16 ms | 55 MB |

These cases are lexically easy, so the lexical baselines do well; v3 misses the Portuguese
case. Timings are from macOS arm64 under heavy machine load. Method, every arm and the limits:
[BENCHMARKS.md](docs/BENCHMARKS.md). Rules modes on real sessions:
[REAL_SESSION_BENCHMARKS.md](docs/REAL_SESSION_BENCHMARKS.md).

Reduction depends on how much of a context is noise. Below a 25% gain, the report marks the
pass as not worth it, and the Claude Code plugin keeps the built-in summary.

## What it does not do

- It does not check that a model will answer correctly. A structural `PASS` is not an answer
  quality result.
- No RAG, embeddings or vector store.
- Protection rules and bounded restoration reduce, but do not rule out, lost evidence; see the
  [safety model](docs/adr/0014-minimum-sufficient-context.md).
- `specialist-v3` was trained on Claude Code transcripts. Other agents and plain documents are
  out of distribution; plain text uses `mechanical` unless you opt in to Jev.

## Docs

| Topic | Link |
| --- | --- |
| Five-minute tour of the CLI | [QUICKSTART.md](docs/QUICKSTART.md) |
| Claude Code plugin | [CLAUDE_CODE.md](docs/CLAUDE_CODE.md) |
| MCP server | [MCP.md](docs/MCP.md) |
| Tool-call compaction and rules modes | [TOOL_CALLS.md](docs/TOOL_CALLS.md) |
| Aider chat history as input | [AIDER.md](docs/AIDER.md) |
| The local ranker | [SPECIALIST.md](docs/SPECIALIST.md) |
| Benchmarks | [BENCHMARKS.md](docs/BENCHMARKS.md), [REAL_SESSION_BENCHMARKS.md](docs/REAL_SESSION_BENCHMARKS.md), [benchmarks/](benchmarks/README.md) |
| Python API (`lcc.msi.compile`) | [msi-api.md](docs/lcc/msi-api.md) |
| Architecture and decisions | [architecture.md](docs/architecture.md), [ADR 0010](docs/adr/0010-deterministic-first-preparation-model-assistance.md), [ADR index](docs/adr/README.md) |
| Opt-in and deprecated providers | [JEV.md](docs/JEV.md), [LAYA.md](docs/LAYA.md) |
| Research archive | [paper.md](research/paper.md), [results-master.md](research/results-master.md), [TRANSCRIPT_AB.md](benchmarks/research/TRANSCRIPT_AB.md) |
| Privacy | [PRIVACY_POLICY.md](docs/PRIVACY_POLICY.md) |

## Development

Python 3.11+.

```sh
python -m pip install -e '.[dev,tiktoken]'
pytest -q && ruff check . && mypy
npm test                                   # Node API and hook tests
python benchmarks/run_judge_bench.py --check --out /tmp/judge-bench
```

The PyPI package named `lcc` is unrelated; this one is `local-context-compiler`.
Contributing: [CONTRIBUTING.md](CONTRIBUTING.md). Security: [SECURITY.md](SECURITY.md).
Changes: [CHANGELOG.md](CHANGELOG.md).

## License

[MIT](LICENSE). The optional Laya integration carries Apache-2.0 attribution in
[NOTICE](NOTICE). Maintained by [LookADev](https://lookadev.com).