Skip to content

emiliolugo/clawmark

Repository files navigation

clawmark

Local Rust CLI to compare agent instruction files (CLAUDE.md / AGENTS.md variants) on real code.

clawmark runs two or more variant files against a bundled 5-task SWE-bench Lite smoke set, invokes Claude or Cursor per variant, evaluates patches with the official SWE-bench harness in Docker, and prints a leaderboard.

Commands: doctor · run · report

Quickstart

1. Check prerequisites

cargo run -- doctor

You need Rust, Docker, Python 3.11+, swebench, and git. For agents: claude and/or cursor-agent on PATH (only the backend you use must be installed and authenticated).

2. Create variant files (inside your working directory)

mkdir -p variants
echo "Make the smallest fix. Do not run the full test suite." > variants/a.md
echo "Make the smallest fix. Prefer targeted tests only." > variants/b.md

3. Run the benchmark

cargo run -- run \
  --a variants/a.md \
  --b variants/b.md \
  --model sonnet \
  --timeout-secs 600 \
  --out out

This runs 2 variants × 5 tasks = 10 agent invocations. --out must not already exist.

4. Read the report (also printed at the end of run)

cargo run -- report --out out
cargo run -- report --out out --show-patches   # include patches (20 lines each)

Common variations

Different model per variant:

cargo run -- run --a variants/a.md --b variants/b.md \
  --model sonnet --model-b haiku --timeout-secs 600 --out out

Claude vs Cursor (A=claude, B=cursor):

cargo run -- run --a variants/a.md --b variants/b.md \
  --model sonnet --agent-b cursor --timeout-secs 600 --out out

Three or more variants:

cargo run -- run \
  --variant alpha=variants/a.md \
  --variant beta=variants/b.md \
  --variant gamma=variants/c.md \
  --variant-model alpha=sonnet \
  --variant-model beta=haiku \
  --variant-model gamma=sonnet \
  --timeout-secs 600 \
  --out out

Concurrent tasks within a variant pass:

cargo run -- run --a variants/a.md --b variants/b.md \
  --model sonnet --timeout-secs 600 --parallel 3 --out out

Release binary:

cargo build --release
./target/release/clawmark run --a variants/a.md --b variants/b.md --model sonnet --timeout-secs 600 --out out

Documentation

Contributing

See CONTRIBUTING.md. Open an issue before large changes.

License

MIT

About

clawmark runs standardized benchmarks on agent configs

Resources

License

Contributing

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages