utils/detect-dupe produces clusters of suspected duplicate functions (exact + fuzzy). On a real codebase those clusters mix real duplicates with false positives - pattern-matching boilerplate, generic accessors, shared-shape glue code that happens to canonicalize alike. Manually triaging clusters is the bottleneck for actually deleting duplicates.

find-dupe feeds each cluster's actual source to Claude (via the das-claude daspkg, module anthropic/anthropic) and asks it to partition the cluster into real-duplicate groups vs false positives, with a one-line reason. Output is JSON (machine, for CI gates / future tooling) + Markdown (human, for review).

[!] Sends source code to Anthropic's API. Each cluster's full function bodies (verbatim, with file paths and line numbers) are uploaded as part of every judge call. Do not run this tool on proprietary, confidential, or otherwise restricted code unless your project's data-handling policy permits sending that source to Anthropic. Anthropic's API terms apply (see anthropic.com/legal). Use --dry-run to preview cluster size and token estimates without making any API calls.

The default model is Claude Haiku 4.5 - cheap, fast, and reliable for "are these the same function?" judgments. Use --model sonnet for harder clusters.

# from the project root
bin/daslang utils/daspkg/main.das -- install --root utils/find-dupe

This fetches das-claude per .das_package and unpacks it under utils/find-dupe/modules/. If you skip this step, bin/daslang utils/find-dupe/main.das fails at compile time with module 'anthropic/anthropic' not found - re-run the install command above.

Set ANTHROPIC_API_KEY so daslang's non-interactive subshells can see it. On macOS the easiest path is ~/.zshenv (sourced by every zsh invocation, interactive or not):

echo 'export ANTHROPIC_API_KEY="sk-ant-..."' >> ~/.zshenv

~/.zshrc works for terminals you opened yourself but not for GUI-launched processes (VS Code, Claude Code's MCP server) - those inherit env from launchctl. ~/.zshenv covers both.

Verify the wiring with the built-in smoke test:

bin/daslang utils/find-dupe/main.das -- --smoke-test

Expect:

Running das-claude smoke test (model=claude-haiku-4-5-20251001)...
Reply: OK
Tokens: in=15 out=4
Smoke test PASSED.

1. Run detect-dupe to produce a JSON report:

   bin/daslang utils/detect-dupe/main.das -- -p <paths> --json ./dupes.json

2. Dry-run find-dupe for a cost preview (no API calls):

   bin/daslang utils/find-dupe/main.das -- --input ./dupes.json --dry-run -v

   Output: cluster counts, estimated input/output tokens, dollar estimate.

3. Live run to produce verdicts:

   bin/daslang utils/find-dupe/main.das -- --input ./dupes.json -v

   Writes find_dupe_verdicts.json (machine) and find_dupe_verdicts.md (human) to ./find-dupe-out/ (override the directory with --out, the basename with --out-basename).

Run from the project root. detect-dupe records source paths relative to its cwd. find-dupe extracts each function's body from disk using those paths, so its cwd must match.

Cross-platform paths. Examples use cwd-relative paths (./dupes.json) so they work on Linux, macOS, and Windows alike. If you'd rather drop intermediates in the OS tempdir, substitute your shell's convention - $TMPDIR/dupes.json (zsh/bash on macOS), /tmp/dupes.json (Linux), or %TEMP%\dupes.json (PowerShell/cmd) - none of those are required by the tool.

The default sort puts fuzzy pairs first (highest AI value), then larger clusters, then larger total span - useful with --max-clusters and when interrupting a long run.

find_dupe_verdicts.json - top-level fields:

Each VerdictRow carries cluster_id, kind (exact/fuzzy), similarity (for fuzzy), members (with file/line/end_line), verdict (real/partial/false_positive/skipped/error), groups (the partition), false_positives (indices), reason, and per-call token usage.

find_dupe_verdicts.md - human-readable, per-cluster sections with verdict tag, reason, member list, partition, and the original source (collapsed under <details>).

Two MCP tools shell out to this CLI so AI assistants can run the full pipeline without leaving the chat:

• judge_duplicates - takes a detect-dupe JSON report path; returns the verdict envelope.
• find_dupe - convenience: runs detect-dupe against paths, then judges the resulting clusters.

Both surface a structured "anthropic daspkg not installed" error with the install command when the subprocess fails for that reason. See doc/source/reference/utils/find_dupe.rst and utils/mcp/README.md for parameter details.

bin/daslang dastest/dastest.das -- --test utils/find-dupe/tests

Hermetic - no API calls. Validates parse_verdict, classify_verdict, extract_source, and read_input_report. Test fixtures use daslib/fio's create_temp_file_result(prefix, ext) so they're cross-platform clean. Wired into run_utils_tests via utils/CMakeLists.txt.

Approximate, as of 2026-04:

A typical cluster (3-4 functions, 50 lines total) consumes ~1 KTok in + 0.3 KTok out -> ~$0.0025 on Haiku. Use --dry-run first if a run looks large.

• utils/detect-dupe/README.md - the cluster-producing pipeline whose JSON we consume.
• the daslang helper bot in the das-claude package repo - the source pattern this tool was built from.