v0.5.0

contextcost v0.5.0

Measure what a repository costs an AI coding agent to read — and prove how
much of it is waste by measuring again.

uvx contextcost .          # no install, no config, results in seconds

Highlights

1. Measures the token cost of every file an agent could read (estimate
   with a published ±14% error bound; --accurate gives exact cl100k_base
   counts).
2. Proposes cuts — only where the filesystem itself proves a file isn't
   hand-written work: lockfiles, minified bundles, vendored deps, generated
   code, recorded fixtures.
3. Re-measures with the proposal applied — the saving it reports is an
   observed difference between two walks of your repository, never a sum of
   guesses.

New in v0.5.0

• GitHub Action (CAOShurong/contextcost/action): posts (and keeps
  updated) a per-PR comment measuring what a pull request does to the
  repository's context budget — "+41,797 tokens, 92% of it a lockfile" as a
  measurement, not a guess from filenames.
• MCP server mode: contextcost mcp serves estimate(repo) and
  propose(repo) over stdio JSON-RPC so Claude Code, Cursor and Codex can
  call it directly (copy-paste configs).
• --delta BASE: measure a change's context cost against any git revision —
  contextcost . --delta main, or against a tagged release, no second clone.
• --fail-over BUDGET: CI gate that exits 4 when the measured total exceeds
  BUDGET tokens.

Real numbers from real repositories

Repository	Tokens to read	Measured saving	Waste
moby/buildkit	14.6M	13.1M	89.5%
jesseduffield/lazygit	5.8M	4.5M	77.8%
plotly.js	63.8M	26.8M	42.0%
astropy	7.9M	212K	2.7%
mikefarah/yq	420K	3.0K	0.7%

The spread is the finding: there is no universal waste percentage — only
your repository's number. Full case studies with per-commit reproduction:
docs/case-studies.

MIT, Python 3.9+, zero dependencies.
Questions: Discussions.