Static (LLM-free) detection of duplicate / similar code in Python and TypeScript (TSX), aimed at catching an LLM re-implementing something that already exists. Input can be a git diff checked against existing source.
crates/duplicatecode-engine– Rust engine (tree-sitter based)crates/duplicatecode-cli–duplicatecodeCLIdataset/– benchmark: coding tasks + diverging implementations (seedataset/README.md)
uv tool install duplicatecode # from PyPI (prebuilt binary, no Rust needed)
uvx duplicatecode scan . # or run without installing
curl -fsSL https://raw.githubusercontent.com/bmsuisse/duplicatecode/main/install.sh | sh # standalone binaryBinaries for Linux/macOS/Windows (x86_64, plus aarch64 on Linux/macOS) are also attached to each
GitHub Release. Releases are automatic: bump
version in [workspace.package] in Cargo.toml, merge to main, and CI tags and publishes it.
cargo build --release
duplicatecode units src/ # list extracted units
git diff origin/main | duplicatecode diff --repo . # check added code against the repo
duplicatecode scan packages/ services/api # find duplicate groups inside one or more folders (monorepo)
duplicatecode scan . --exclude '**/generated/**' --skip-tests --fail-on-found # CI-style run
duplicatecode scan . --pairs --json # machine-readable pairs instead of groups
duplicatecode bench --dataset dataset [--file-level] [--mutations] [--negatives <other repo>]Units (functions, methods, classes, arrow-function components) are extracted with tree-sitter and
normalized (identifiers/literals abstracted; comments, docstrings and type annotations dropped).
Each pair gets several signals: token 4-/2-gram overlap, node-kind histogram, literals, API/attribute
names, name subwords, callees, and statement-level multiset/alignment scores. combined is a weighted
sum (0.6 structure, 0.2 name, 0.2 callees). Constructors and dunder methods are ignored by default.
An inverted index over k-grams keeps matching against large corpora fast.
- Ranking is good (right match is top-1 in ~70–95% of cases) but absolute detection at a 1% false-positive rate is modest for independently written code: ~35–50% (function level) of same-task pairs.
- Comparing whole files (which removes helper-splitting noise) gives 76–100% recall, mostly thanks to names; structure-only signals reach ~35–55% on loose/hard tasks. Statement-level features are roughly on par with token k-grams; a fitted logistic model does not beat the hand weights in cross-validation.
- Mechanical rewrites (rename, reorder, temp variable, logging, dead code) are handled well individually (86–100% of pairs ≥ 0.7); all combined drops the mean score to ~0.58, mostly because renames zero the name signal.
Before comparing, debug output (print, logging.*, console.*) is dropped; Python comprehensions are
read as the equivalent loop; x += 1 as x = x + 1; a temp variable returned right after its definition
is inlined; operators are canonicalized (===/==/is, and/&&, None/null, const/let).
Mutation benchmark (bench --mutations): each of rename, statement swap, temp variable, logging, dead code and
loop-to-comprehension keeps >= 93% of pairs at score >= 0.7; all combined 40% (renames zero the name signal).
Self-scans of three internal repositories (Python + TypeScript), 189 pairs judged by reading both units (TRUE_DUP / PARTIAL / BOILERPLATE / FALSE):
- Raw scores are a weak signal: only ~23% of pairs above 0.55 were true duplicates, 40% incl. partial; below 0.65 almost none. Real duplicates were almost all exact copies with matching names.
- Dominant false alarms: react-query/fetch wrappers differing by endpoint, per-entity CRUD endpoints, one-line repository delegators, DTO/model classes, per-file test fixtures, generated clients.
- Filters added from this: field-only classes, constructors/dunders, generated files, tiny units (unless near-identical with the same name), test code (unless identical), minimum name similarity (0.3).
--profile copies(default) re-weights signals for copy-paste (name 0.36, statements 0.27, literals 0.18); leave-one-repo-out AUC improved on all three repos (0.71/0.69/0.95 -> 0.75/0.79/0.96). At threshold 0.6 the judged sample gives ~50% true duplicates and ~76% incl. partial while keeping ~70% of the true ones. These numbers are partly in-sample (weights and filters were chosen on the same pairs) — re-judge a fresh sample before trusting them. A structure-heavy--profile reimplexists for renamed re-implementations (use with--min-name 0) but has no real-repo precision data yet.
A third, untouched sample of 57 pairs (nothing was tuned on it): 30% true duplicates, 47% incl. partial
(OneSales 0/24, MDMApp 11/24, CCMT2 6/9 true). The 50%/76% above was optimistic because it was measured on the
pairs used to choose the weights. Test code is about half of the noise but also holds real copies
(25% true either way), so it is kept by default; --skip-tests drops it.
42 real helper functions from the three repos were described neutrally (no names/code) and re-implemented from scratch by Haiku and Sonnet without seeing the repo. Fraction where the original is found (copies profile) and where the best other match also crosses the threshold:
| threshold | Sonnet found | Haiku found | both | other-code alarm |
|---|---|---|---|---|
| 0.3 | 67% | 60% | 63% | 21% |
| 0.4 | 62% | 33% | 48% | 8% |
| 0.5 | 45% | 14% | 30% | 1% |
| 0.6 | 24% | 5% | 14% | 0% |
So diff (checking new code) defaults to threshold 0.4 while scan (existing copies) defaults to 0.6. About half of
independent re-implementations are caught; the rest are genuinely different code. The structure-heavy reimpl
profile is not better on this test.
Four Sonnet agents (report only, ~80 tool calls, max 30 groups) hunted duplicates in two repos: two with plain read/grep, two with this CLI. All distinct groups (82) were then judged blind by independent agents that read the code.
| groups | true dup. | precision (true / incl. partial) | pooled recall (true) | tool calls | |
|---|---|---|---|---|---|
| MDMApp, LLM only | 25 | 15 | 60% / 80% | 60% | ~27–33 |
| MDMApp, with CLI | 26 | 20 | 77% / 92% | 80% | 7 |
| OneSales, LLM only | 28 | 10 | 36% / 71% | 67% | ~33–42 |
| OneSales, with CLI | 21 | 10 | 48% / 86% | 67% | 9 |
Only 18 of 82 groups were found by both, so the approaches are complementary. The agents' reports are capped at 30
groups, so the raw tool is a better measure of recall: scan --threshold 0.6 (defaults) finds 38 of the 40 judged
true duplicates (25/25 MDMApp, 13/15 OneSales) and 23 of the 25 that the LLM-only agents found independently.
What it still misses: a differently-written picker function (same purpose, different code) and a formatFileSize
variant with different constants. Fixed after this test: tiny same-name exact copies (min_tokens 20 -> 8, near-exact
same-name rule) and multi-line module-level values such as export const queryClient = new QueryClient({...}).
Same task and judging as above (blind judges, pooled recall against all distinct confirmed true duplicates: 37 in MDMApp,
18 in OneSales). review + skills/duplicate-code-review/SKILL.md were built from what the first runs missed.
"cost index" = tokens x relative price (Haiku assumed 1/3 of Sonnet per token; estimate).
| repo | approach | groups | precision true / incl. partial | recall | tokens | cost index |
|---|---|---|---|---|---|---|
| MDMApp | Sonnet, no CLI | 25 | 60% / 80% | 41% | 86k | 259 |
| MDMApp | Sonnet + CLI | 29 | 79% / 93% | 54% | 59k | 177 |
| MDMApp | Haiku + CLI + skill (r1 / r2) | 31 / 40 | 90% / 90% · 70% / 82% | 65% / 70% | 86k / 86k | 86 |
| OneSales | Sonnet, no CLI | 28 | 36% / 71% | 56% | 131k | 393 |
| OneSales | Sonnet + CLI | 21 | 52% / 90% | 61% | 62k | 187 |
| OneSales | Haiku + CLI + skill (r1 / r2) | 11 / 30 | 82% / 91% · 33% / 67% | 50% / 50% | 88k / 74k | 74–88 |
r1 = first skill (top-60 groups only); r2 = skill with --brief paging. Read: on MDMApp Haiku + CLI beats Sonnet alone on
recall (70% vs 41%) and precision at about a third of the cost; on OneSales it is at par with Sonnet alone on recall
(50% vs 56%, one group) with better-or-equal precision at about a fifth of the cost, but below Sonnet + CLI. Broadening (r2)
buys recall at the price of precision on OneSales. Small samples; judges and finders are the same model family.
Round 3 (after LIKELY-NOISE tagging and the line-diff view; hard cap 30 groups): MDMApp 29 groups, precision 69% / 86%,
recall 51%, 69k tokens; OneSales 12 groups, precision 75% / 83%, recall 42%, 89k tokens. Recall pooled against 41 (MDMApp)
and 19 (OneSales) confirmed duplicates. Run-to-run variation of Haiku (59–63% / 51% on MDMApp, 42–47% on OneSales) is as
large as the differences between skill/tool versions, so the tag and diff view are not shown to help measurably; averaged over
three runs Haiku + CLI + skill lands at ~58% (MDMApp) and ~45% (OneSales) recall versus 37% / 53% for Sonnet alone and
49% / 63% for Sonnet + CLI, at roughly a fifth to a third of the cost of Sonnet alone. Note the 30-group cap is itself a ceiling:
the confirmed pool holds 41 real duplicates in MDMApp.
Pooled confirmed true duplicates: 48 (MDMApp), 35 (OneSales); every group anyone reported was judged blind by reading the code.
| repo | arm | groups | precision true / incl. partial | recall | tokens | cost index (Haiku = 1/3 price) |
|---|---|---|---|---|---|---|
| MDMApp | Sonnet, no CLI | 65 | 42% / 65% | 65% | 174k | 522 |
| MDMApp | Sonnet + CLI + skill | 69 | 62% / 86% | 77% | 72k | 214 |
| MDMApp | Haiku + CLI + skill | 59 | 73% / 86% | 67% | 102k | 101 |
| OneSales | Sonnet, no CLI | 64 | 30% / 56% | 54% | 153k | 459 |
| OneSales | Sonnet + CLI + skill | 80 | 32% / 71% | 66% | 90k | 270 |
| OneSales | Haiku + CLI + skill | 47 | 38% / 68% | 40% | 85k | 84 |
| both | tool list only, IDENTICAL+NEAR-COPY, minus LIKELY-NOISE |
173 / 278 | (judged subset: 66% / 85% MDMApp, 38% / 64% OneSales) | 88% / 83% | 0 | 0 |
- With the same budget Haiku + CLI matches Sonnet alone on MDMApp (67% vs 65% recall, better precision) at ~1/5 of the cost; on OneSales it is below (40% vs 54%) at ~1/5.5 of the cost. Sonnet + CLI is best on recall in both.
- The unfiltered tool list already covers 83–88% of the pool: the agents' job is pruning, and their lower recall comes from what they choose to report (they judge from brief lines and read few members), not from what the tool misses. Precision of the raw list is unknown beyond the judged subset (which is biased towards groups an agent reported).
- Caveats: judges and finders are the same model family; the pool only contains duplicates someone found; single runs per arm (Haiku varies by +-10 points between runs).