Repository navigation
TuckBit's first tagged public release. It packages everything public since 0.1.2 and adds the opt-in Headroom companion tools. No token, cost, speed or accuracy savings are claimed. Installing the skill does not prove that a model will activate it.
Highlights
- Headroom companion (opt-in). TuckBit decides what context to fetch, Headroom (Apache-2.0) compresses it, and the Observatory measures what happened. See Using TuckBit with Headroom.
scripts/headroom_compare.pycompares sessions you select, run without and with a compressor. It reports observed counters, the cache fraction and an optional configured cost scenario. It flags cases where fewer input tokens did not mean fewer uncached tokens or a lower configured cost.scripts/json_densify.pyis a stdlib-only schema+CSV densifier for uniform JSON arrays. It returns the dense form only when it decodes to the same parsed value; otherwise you get the original back unchanged.
- Context reread profiling.
trace_profiler.pyclassifies first reads, unchanged and changed rereads, and post-mutation rereads. - Observatory local preview. A loopback-only dashboard with a synthetic demo and explicit file import. It does no background capture.
- Reproducible offline task bundles, multilingual onboarding (14 locales), and a refreshed README and install prompt.
- The metadata is synchronized: all six distribution manifests, the canonical skill and the locale skills are at 0.4.0.
Full details are in CHANGELOG.md.
Install
npx skills add MakeItEzzz555/tuck-bit --list # inspect first
npx skills add MakeItEzzz555/tuck-bitReversible project-scoped preview: python3 scripts/install.py install --scope project --project /path/to/project --client codex --dry-run. See compatibility.
Validation
Hosted CI passed on the release commit across Ubuntu and macOS with Python 3.10 and 3.13 (run 40867092642). All checks are no-model. Every step in .github/workflows/tests.yml passed locally too: doctor, manifest sync, README flow check, 83 Node Observatory tests, i18n check, 479 Python unit tests, public privacy audit, smoke test, routing corpus, offline bundle (24/24), and router/registry/trace smoke tests. Review hardening in #21: the comparison excludes estimated counters and refuses a session selected in both arms even when one copy is incomplete. Both CLIs refuse files over 4 MiB before reading them whole, and expand() rejects duplicate schema keys.
Known limitations
- Real cross-provider skill activation and native efficacy are not established yet (help wanted: #4, #5).
- The Headroom comparison is observational. Sessions are user-selected rather than matched tasks, and success is not assessed. A smaller character count is not a token or cost result.
- The standalone Agent Skill install is instruction-only. The Python utilities need a full source checkout.