CodeEraser

· English | 中文

An eraser against LLM-induced code & document entropy.

What it is

Long-lived LLM-assisted codebases drift the same way: the same function implemented twice, the same paragraph pasted into three files, updates that arrive as appends, files that only ever grow. CodeEraser stops that drift at the moment of writing and gates it in CI, with no model in the loop anywhere. Two refusals happen at write time, before the content reaches disk. A write that would introduce an exact T1/T2 clone (duplication the replaced content did not already carry) is denied at PreToolUse, with the region it duplicates named and the ordering that passes taught; a write leaving a file over 750 lines, or over the line its [[rules.class]] declares, is denied the same way. Everything else is a report or a gate: the Stop audit refuses the turn, ce precommit and ce commitmsg refuse the commit, and the CI exit codes refuse the merge.

Scope. Judged languages: Python, TypeScript/TSX, Rust, Go, Haskell and Markdown (seven language codes over six tree-sitter grammars). Size-only arm: js/mjs/cjs/jsx, css/scss/less, html/htm, vue, svelte, sh/bash, yml/yaml. They enter the size gates, the hard budget and the ratchet, never a semantic verdict. Faces: CLI · GUI (eleven screens) · Claude Code plugin (three hooks, one skill, one command, sixteen read-only MCP tools) · pre-commit · CI.

How it works, and what is different about it

• Interception at the instant of writing. Every file's normalized tokens (identifiers → ID, literals → LIT, comments dropped) are winnowed with k = 25, w = 26, so any shared run of 50+ tokens is guaranteed a shared fingerprint. The fingerprints live in a SQLite WAL index kept by a lazy per-project daemon; the PreToolUse probe answers in 48 ms p50 / 58 ms p95 on a two-file fixture, and the whole plugin chain in 0.50 s p95 as last measured on 2026-08-29, before the tombstone leg joined it. The guard charges only novel duplication: matches the replaced content already carried are subtracted, so on the live-stream reading it misfires on none of 719 production probes (0.00 per 500); the 2,761-event replay's full-file-write reading charges the 32 split-a-file intermediate states at 7.03 per 500; both readings are ledgered in FPR-REPLAY.
• Two clone layers, one verdict owner. T1/T2 is the hot path above. T3 is a cold path: structural fingerprints and MinHash/LSH (128 permutations, 32 bands × 4 rows) generate candidates without dropping a pair that could pass, and the Haskell core computes Zhang–Shasha tree edit distance and accepts at TSED ≥ 0.85, in exact integer arithmetic.
• Documentation duplication that survives rewording. NFC-normalized words, 5-word shingles, MinHash/LSH candidates, then an exact Jaccard ≥ 0.80 or a 50-word verbatim run, judged in the core with exact rationals.
• Liveness that is named, not guessed. Per-language resolution ladders (imports, re-exports, doc links, assets, package roots) feed a rung-filtered graph; SCCs, reachability from entry roots and a four-way verdict (unreferenced/unreachable × private/public) come back with a confidence code derived from the unresolved-site ledger. Beside it, the mention universe (every identifier in every text file, stored only as fnv1a64 hashes) yields the unmentioned declaration advisory, which never turns a gate red.
• Same-role advice from the repository itself, no model. The index's own facts become term bags (names, shapes, callees, docs, structure, literals), scored with integer BM25 (k1 = 6/5, b = 3/4) and judged in the core: candidates in one exact order, plus a role bit that holds only when the name, callee and shape channels agree. --widen adds the in-repo PPMI associative view. Advisory by construction: no exit code, no gate, no hook block.
• Structure as a measured thing. Eight axes (geometry, naming diversity, mixing, misplacement, documentation coverage, stale docs, redundancy, modularity), Tsallis-2 entropy per directory, chi-squared divergence from a declared layout, and split-ROI pricing with four cost legs (crossing references, clone cuts, churn crossings, a new-file φ) or a cohesion alibi.
• A check score that cannot be gamed by moving lines. The gate's own axes (size, complexity, clones, documentation duplication, dead code, churn, cycles) each charge violation mass over opportunity as floor(1000·v/(v+n)), and the weighted fold lands on 0–1000. The ADR-006 ratchet tightens every ceiling automatically; growth needs the tolerance max(+2 %, +10 lines) or a named re-establish (CE_ACCEPT_BASELINE=1), and a knob edit stops ce check by name instead of moving every line.
• Time as a first-class signal. Theil–Sen slope over the last 512 score points (a single wild point cannot drag a median); churn = added − surviving lines by blame; the join lattice combines similarity, graph position and churn into merge / delete / churn-hotspot with reason bits and confidence.
• Erase with a safety predicate, not a heuristic. Three classes (verbatim doc twin, whole-unit T1 twin whose copy is dead, confident non-public dead file), seven frozen reason codes, a 4,096-row cap, and a convergence re-plan that fails if any applied verdict survives.
• Tombstone residue, judged as a conjunction. Narrating a name the same change erased leaves a residue: a struck-through or (removed) label, or a sentence carrying a backward-looking mark (no longer, used to, 曾经 …) together with the erased name. Rust measures the changeset's surfaces (the names it erased, its new labels, its new sentences); the core judges them sentence by sentence (a mark and an erased name together; a label needs the name alone) against a floor and the class's [tombstone] budget. PreToolUse, the Stop audit and ce precommit / ce commitmsg speak it, each only at the class's [tombstone] tier and only when the core answers over. Changelog-role documents are exempt by path, by ledger shape, by a segment with three version witnesses or by a [tombstone] ledger declaration, and the false-positive rate is replayed on git history before any default moves (FPR-TOMBSTONE).
• Deterministic by construction. No RNG or clock in any judgment; golden fixtures compared byte for byte; configuration crosses as facts, never as names.

Evidence: the same task, run twice

One seven-step task, two identical copies of the seed; the only variable is whether CodeEraser is in the loop — the write-time guard, the Stop audit, and, once the audit refuses, the eraser acting on its own plan. Both runs still end red — not on the same things.

The task (add discounts, a compact report, CSV and JSON output, money formatting in the API) is replayed by a scripted agent on demo/seed, a small invoicing service in Python and TypeScript. The seed is measured first, so every finding below was written by the task. Each loop then runs to its end; with nothing in the loop nothing refuses anything, so that one ends at the last write. Every verdict is the verbatim output of ce, and both trees are measured by the same six commands.

The two denied writes are copies of an existing helper. The compact renderer that slipped past is the honest boundary, since a full-file rewrite copies its own blocks and nothing is novel at write time. It is what the Stop audit refuses the turn over, naming both blocks; the repair that answers it is the only write made because a gate asked rather than because the task did. What stays red is what a person has to settle: invoicer/invoice.py is 93 lines against a tolerated ceiling of 61, which the ratchet holds open for a named re-establish instead of absorbing silently, and two files are unreferenced: the new page nothing links to, and the renderer the CLI stopped importing on its way to JSON. Both transcripts, both SVGs and the JSON behind this table are generated by demo/run.js and re-checked byte for byte in CI (demo_replay). The first real interception, recorded the day it happened: T1-INTERCEPT.

Two moments from close up on the same seed; the second adds one ce.toml declaration.

A copied helper, refused before the file exists. Move 1 of the run above, on its own. The reason names the region the content duplicates and the ordering that would pass, so the refusal is actionable rather than a veto.

$ Write invoicer/discount.py
✗ ce: content for <work>/invoicer/discount.py duplicates 1 indexed region(s): invoicer/money.py:1-18 (89 tokens). Reuse the existing implementation instead of re-writing it. Moving it? Trim the source region first: the probe verifies against the current tree, and the same write then passes.

One line, two mouths. ce.toml puts invoicer/** on file_lines_fail = 40. The write-time guard refuses the write that would cross it, and ce scan grades the same tree against the same number — one declaration, read by the hook and by CI.

$ Write invoicer/invoice.py
✗ ce: this write leaves <work>/invoicer/invoice.py at 93 lines, past the hard budget of 40 (plan §4.1). Split the file instead of growing it.
$ ce scan .
FAIL invoicer/invoice.py:1 file-lines = 51 (limit 40) [invoicer/invoice.py]
warn invoicer/report.py:1 file-lines = 35 (limit 30) [invoicer/report.py]
scanned 9 files / 19 functions — 1 warn, 1 fail -> FAIL (failed: hard_line)

Latency · v1.7.0

Every value is generated from contracts/bench/bench.json; the test rejects hand edits to this block. The current release, v1.7.1, ships the same measured code as the newest row and gets none of its own. Full replay notes and per-version series · Complete website dashboard

Latency rows are release-build replays on one fixed host, comparable version to version only. The precision and recall points are frozen with their evaluation ledgers (EVAL-SET) and rendered on BENCH; comparators (jscpd, similarity-*) are named with the exact version measured.

Install, run, update

Installer. Each release ships five GUI installers (NSIS setup.exe / AppImage / dmg) bundling the GUI, ce and the ce-core judgment core. The Windows installer puts the install dir on PATH and runs ce setup; AppImage and dmg users run ce setup once themselves; it finds Claude Code, wires the plugin below and says whether that ce's directory is on PATH. The fifteen artifacts (ce and ce-core per target plus the installers) and SHA256SUMS are unsigned by decision; verify with sha256sum -c --ignore-missing SHA256SUMS.

Homebrew / winget. The release generates a Homebrew formula and the three winget manifests under packaging/ from the same pinned manifest it verified, never edited by hand, and publishes them when the repository's tap and winget tokens are configured: brew install skymanbp/codeeraser/codeeraser (macOS and Linux, ce + ce-core from the pinned assets) and winget install skymanbp.CodeEraser (the Windows installer, once winget-pkgs merges).

Claude Code plugin. ce setup registers the marketplace at this repository's release branch (fast-forwarded to every published tag, so an install follows releases, not main) and installs the plugin; by hand: /plugin marketplace add skymanbp/CodeEraser@release, then /plugin install codeeraser@codeeraser. The starter resolves ce and ce-core by pin: a matching local or PATH copy, then a pinned download, then an unverified PATH binary that says so.

CLI only, or from source. Download ce-<ver>-<platform> and ce-core-<ver>-<platform> (x86_64-windows / x86_64-linux / aarch64-macos / x86_64-macos / aarch64-linux), rename them ce / ce-core and put them side by side on PATH; or cargo install codeeraser and place a ce-core beside it; or build both with the pinned Rust toolchain (rust-toolchain.toml) and GHC 9.14.1 + cabal: cd core && cabal build all && export CE_CORE_BIN=$(cabal list-bin ce-core), then cd .. && cargo install --path cli. Core resolution is one chain everywhere: CE_CORE_BIN → a ce-core sibling → PATH; --core <path> wins.

Console output, --help and the hooks' own refusal sentences are English by default and Chinese under --lang zh, CE_LANG=zh or the project's [ui] lang = "zh" in ce.toml, in that order of precedence, and --help reads only the first two; JSON schemas and the FAIL/pass vocabulary are never translated. [[rules.class]] in ce.toml gives one glob set its own size and complexity lines and ratchet tolerance (0 = may not grow), and the same line is read by the score, the ce scan ladder and the PreToolUse budget (ce.toml reference).

Updating. Release builds are two-phase: draft assets hashed, the pins committed to plugin/bin/manifest.env, and only then the tag verifies the same bytes (RELEASE), which is what ce update and the tag job's installer check both read. ce update reads the latest tag and that tag's committed manifest.env; the verdict is the exit code, and --yes acts only where nothing else keeps a ledger of the binary. A copy the plugin bound is re-pinned by /plugin update codeeraser; a cargo install by cargo install codeeraser; the GUI app itself by the installer --installer saves. The plugin's SessionStart line announces a newer release once a day (CE_UPDATE_CHECK=0 turns that off); the GUI has an update screen; /codeeraser:update runs the check from Claude Code.

Three faces, one product

Every capability is claimed once in this table, the sets are derived from the code (clap's enum, the Tauri roster, the MCP catalog, hooks.json, plugin/commands, plugin/skills), and a CI gate (face_parity) refuses a face nobody wrote down or a claim nobody shipped. Deliberate omissions are rows, not silence.

Tech stack, design and philosophy

• Rust 1.94.1 (edition 2024). The codeeraser crate carries tree-sitter 0.27 with six grammars, rusqlite 0.40 (bundled SQLite, WAL, index schema 16 / GRAPH_REV 15 / MENTION_REV 2), the ignore walker, interprocess named pipes / Unix sockets, clap, serde, sha2 for the updater's pins.
• Haskell (GHC 9.14.1, GHC2021, -Wall -Werror): ce-core, every judgment family, a frozen dependency graph.
• Tauri 2 GUI over the same crate, vanilla JavaScript in the webview, no build step; NSIS / AppImage / dmg bundles with ce and ce-core as sidecars.
• One wire. ce ↔ core is NDJSON over stdio with SemVer negotiation (proto 7.1.0, twelve families); the per-project daemon speaks its own protocol (2.1.0) over interprocess; a protocol-major skew is a named refusal, never a guess.
• Design rules. ADR-001 Rust frontend · ADR-002 Haskell judges, never parses · ADR-003 lazy daemon, 30-minute idle exit, fail-open hooks · ADR-004 cheap PreToolUse, deep Stop, CI as backstop · ADR-005 two clone layers · ADR-006 shrink-only ratchet · ADR-007 pinned distribution · ADR-008 policy as Haskell data · ADR-009 documentation facts derived, never hand-typed. The plan is the contract: DEVELOPMENT_PLAN.
• Philosophy. Measure in Rust, decide in Haskell, render everywhere else. Codes cross the wire; each face owns its sentences. Nothing on any surface asks a model anything. Hooks fail open and say so. A guard class reaches deny only with its own false-positive record in CHANGELOG. Documentation is generated or gated: the CLI and config references, the fifteen-booklet methodology with machine-checked citations, the numbers this page derives from the code, the two diagrams above, the bench block, the demo, the site's terminal block and its GUI screenshots, the parity table, the NOTICE. This repository is its own first user, and every push runs the six product gates on this tree.

Known limits

v1.7.0 scope. The v2.29 plan re-opened, item by item, the deferred work the 2026-08-31 close-out had ruled not-done (plan v2.22, 45 items across three bundles), keeping only what earns its cost. The limits below are the positions that survived.

Limits. PreToolUse shapes behaviour; it is not a security wall (shell writes bypass it; the Stop audit and CI are the backstops). Hooks fail open on internal errors and record the degradation. Binaries are unsigned.

• Languages. AST-based judgment uses the six grammars above; Markdown is judged by the documentation and graph rules without a tree-sitter grammar; JSDoc and Rust /// are comments, not docstrings; T4 clones are not promised. churn, join and trend are minute-scale.
• Advice, never a verdict. Symbol-level liveness is an advisory and never a verdict; ce deadcode says so on its own last line. The same-role advisor is an advisor and never a verdict: ce similar, MCP similar_units and the GUI similar screen rank and rebut nothing. ce similar always exits 0, ce check never reads the family, and the Stop audit's advisory row only reaches the observe ledger. A guard class stays observe until it carries its own false-positive record.
• Lines this product does not draw. The complexity axis ships with no hard line at all. cognitive_fail defaults to 0, so a tangled function warns and nothing more until a repository declares one, and the write-time hook never judges complexity. ce structure has no score floor, so that family reports and gates nothing.
• Distribution and setup. A release builds five targets since v1.7.0 (x86_64-windows, x86_64-linux, aarch64-macos, x86_64-macos, aarch64-linux); on the two newer ones a plugin starter reading a manifest older than that finds an empty pin and falls back to a PATH ce or a source install. ce setup wires nothing when the account running it is not the logged-in user (exit 13); run it unelevated from your own account. Judging this repository needs the cli/tests submodule seated (it is a reader of the tree, never a measured part).
• Tombstone residue. It counts single-word names, so a common word erased as an identifier can bind a sentence genuinely about that word; the repository's own [tombstone] terms list is the way out. It measures CJK names only at word boundaries: a wide Chinese name is seen where it stands alone as a heading, a list lead or an identifier, and is bound in prose by substring.
• Score comparability. Scores are not comparable across a [[rules.class]] switch, across the v0.7.3 → v1.0.0 density-charge rescoring, across the v1.2.0 → v1.3.0 test-submodule move, across the v1.3.x → v1.4.0 recursion increment, or across the v1.6.0 → v1.7.0 re-denomination of the clone and docdup axes (files touched over each axis's own universe, with docdup pairs reaching ce check for the first time). ce structure's own 0–1000 is likewise not comparable across v1.6.0 → v1.7.0: the modularity axis is new, and every axis enters the equal-weight fold. Scores are also not comparable across a named re-establish that moves the baseline's softLine: the size axis charges against that repo-relative line rather than a constant, and this repository's has moved with every named re-establish since 304 at v0.7.3 (372 at v1.4.1, three points apart when both lines are applied to the v1.4.1 tree).

Documentation

• CLI reference · ce.toml reference (generated from the binary and the config schema; drift reddens CI) · Methodology (fifteen booklets, cited to implementation lines) · structure axes · size advisory · erase contract · GUI reference · plugin · demo
• DEVELOPMENT_PLAN · EVAL-SET · EVAL-SET-SIMILAR · FIELD-TEST · BENCH · PERF-BUDGET · FPR-REPLAY · FPR-TOMBSTONE · FPR-L2 · T1-INTERCEPT · contracts/VERSIONING.md · docs/RELEASE.md (wire SemVer and the two-phase release runbook)
• Website: codeeraser.dev · how it works · stack · bench

License

Apache-2.0; see LICENSE. Third-party inventory in NOTICE (regenerated and gated in CI). The test suite lives in skymanbp/CodeEraser-tests; clone with --recurse-submodules. "CodeEraser"™ is a trademark of skymanbp; per Apache-2.0 §6 the license covers the code, not the name.