ccompactor/ccompactor

CCompactor

Extract any coding agent's session into a compact, verified, provenance-linked handoff that any other agent can continue from.

📖 ccompactor.github.io · 📦 npm

ccompactor list                       # what sessions exist
ccompactor find "auth migration"      # find one by topic
ccompactor extract claude:last        # the handoff artifact
ccompactor handoff claude:last --to codex --run

Same idea as sctxx, in TypeScript.

Install and run

npm i -g ccompactor     # the tool
ccompactor doctor

Or from source, to work on it:

npm install          # workspaces; installs commander, ink and react
npm run link         # builds, then points `ccompactor` at this checkout
npm run dev -- doctor   # build and run in one step

See DEVELOPING.md for the loop, the checks and how to try it on real sessions.

Standalone binaries (no Node needed) for macOS arm64/x64, Linux x64/arm64 with a .deb, and Windows x64/arm64 are attached to each release.

Either way, ccompactor update moves you to the newest release: it re-runs the npm install for a global npm install, and downloads and replaces the binary in place for a standalone one. It reports what it would do without --check changing anything, and it refuses to touch a source checkout.

Or without linking:

npm run ccompactor -- doctor
node packages/ccompactor/dist/cli.js list --any-project

Releasing is automated — see RELEASING.md. The one manual step is an npm Automation token stored as the NPMTOKEN secret, because a normal Publish token is gated behind an OTP prompt that CI cannot answer.

Status

All phases of SPEC.md are implemented: CLI, discovery, three adapters, ledgers, artifact, verify, expand, handoff, skill and TUI. No Anthropic-derived code is present — an earlier vendored copy was never imported and has been deleted; see NOTICE.

ccompactor bench also exists, as a research instrument rather than a feature: it measures whether a handoff hands anything off, and it is how the recency-tail arm was measured and then deleted. A previously published comparison against sctxx has been withdrawn — the two runs were scored on question sets sharing 3 of 21 questions, so the per-arm percentages were not comparable. On the 22 questions they do share, sctxx answers 14 and ccompactor