Skip to main content

Executor.sh comparison (executor-cmp-001)

Side-by-side token comparison against executor.sh on a GitHub PR list — the benchmark behind Both Sides of Context Compression.

Tokenizer: cl100k_base. focus=input (model output omitted from headline ratios).

Headline numbers (live, vercel/next.js)

LayerExecutorClawQLNotes
Layer 11,044 (homepage) / 115 (live execute)394 (codemode)394 does not beat live 115
Layer 2143,466 (pulls.list, full REST)907 (execute + fields)158× on tool result alone
Combined144,510 (pub L1) / 143,581 (live L1)1,301111× / 110×

Layer 2 is 100% of Executor's input and 70% of ClawQL's on this task.

Live flamegraph (public)

Static snapshot — same HTML as mcp-api-adapter's built-in demo:

Regenerate after measurement changes: npm run generate:executor-cmp-static (from repo root).

Reproduce

# Fixture (CI shape)
npm run benchmark:executor-comparison

# Live GitHub + ClawQL
BENCHMARK_LIVE=1 CMP_GITHUB_REPO=vercel/next.js CMP_PER_PAGE=30 \
  npm run benchmark:executor-comparison

# Live both arms (Executor CLI + ClawQL)
BENCHMARK_LIVE=1 CMP_GITHUB_REPO=vercel/next.js CMP_PER_PAGE=30 \
  EXECUTOR_BIN=/path/to/executor EXECUTOR_CWD=/path/to/executor-cwd \
  EXECUTOR_GITHUB_PULLS_PATH=github.user.githubMain.pulls.list \
  npm run benchmark:executor-comparison

Multi-turn series: npm run benchmark:executor-comparison:multiturn and npm run benchmark:executor-comparison:uniform.

Raw artifacts (repo)

What not to claim

  • 862× multi-provider workflow — different metric; do not blend into this comparison
  • Fixture 953 vs 23 — harness sanity only, not for headlines
  • ClawQL 883 / 3,548 — context tiers, not Executor chart parity