Benchmark harness and methodology for ContextAtlas, an MCP server that gives Claude Code a curated atlas of a user's codebase.
Status: Phase 5 reference run committed (hono, v0.1 baseline — 50–71%
tool-call reduction on architectural win-bucket prompts). v0.2 expansion
in progress — see ../contextatlas/v0.2-SCOPE.md.
src/harness/— Benchmark harness code. Implements four measurement conditions:- alpha — minimal baseline agent (Anthropic SDK tool-loop, Read/Grep/Glob/LS)
- ca — alpha + ContextAtlas MCP tools
- beta — real Claude Code CLI driven headlessly
- beta-ca — beta + ContextAtlas MCP server wired in
prompts/— Locked prompt set (24 prompts total across hono and httpx; Phase 5 reference run covered 6 per repo for hono; cobra prompts land in Step 11)configs/— Per-repo.contextatlas.ymlfor hono, httpx, and cobra benchmark targetsadrs/— Fixture ADRs for the benchmark target repositories (hono, httpx, cobra)runs/reference/— Committed reference run artifacts (Phase 5 hono run lives here)research/— Phase-by-phase research notes. Seeresearch/phase-5-reference-run.mdfor the Phase 5 synthesis.RUBRIC.md— Full methodology document (Phase 5 validated)STEP-7-PLAN.md— Initial planning document for the MVP harness (historical)
- Source code for
contextatlasitself (that lives in the main repo) - Cloned copies of benchmark target repositories (gitignored; they must be cloned manually per the setup instructions below)
Prerequisites:
- Node 20+
- npm
- ContextAtlas cloned locally at
../contextatlas(this repo'spackage.jsonreferences it via file path) ANTHROPIC_API_KEYset in your environment (for extraction and baseline agent)- Claude Code CLI installed (for Beta harness runs)
Install dependencies:
npm installClone benchmark target repositories:
mkdir -p repos
git clone --depth 1 https://github.com/honojs/hono.git repos/hono
git clone --depth 1 https://github.com/encode/httpx.git repos/httpx
git clone --depth 1 https://github.com/spf13/cobra.git repos/cobraThese are pinned to specific commits for reproducibility. See RUBRIC.md for pinned SHAs.
- Methodology:
RUBRIC.md(Phase 5 validated). - Phase 5 reference run synthesis (hono, v0.1 baseline):
research/phase-5-reference-run.md. - Reference artifacts:
runs/reference/hono/— summary, per-cell traces, run manifest. - MVP harness plan (historical):
STEP-7-PLAN.md. - v0.2 benchmark work: httpx reference run + Go benchmark target
registration (cobra). See
../contextatlas/v0.2-SCOPE.mdStream B.
See LICENSE. Same license as the main ContextAtlas repo.