Justify or remove: auditing what the agent always carries

## Spark

Digest, week of 2026-07-25: a pass over what Claude Code always carries — trimming the skills index out of CLAUDE.md, scoping the permission allowlist to behaviour-changing rules, committing read-only sensor permissions per repo, and putting `rtk` on notice to justify itself with measured savings or be removed.

## Why it could be interesting

Every always-loaded token is a standing tax paid on every turn, and the usual response is to add rather than audit. The sharper claim is in alunduil-chezmoi#444: a tool already installed in the bootstrap now has to prove its savings or leave. Applying a justify-or-remove ratchet to your own tooling is rarer than adopting the tool was.

## Open questions

- [ ] How do you actually measure rtk's savings without a controlled comparison?
- [ ] What's the rule for context that belongs in a skill body vs the always-on file — is "does a skill already carry it" sufficient?
- [ ] Does the committed `.claude/settings.json` contract hold for cloud/remote sessions where host config doesn't travel?
- [ ] Is there a measurable before/after on turn cost from the trim?

## Source material

- https://github.com/alunduil/alunduil-chezmoi/pull/432 (closes #431)
- https://github.com/alunduil/alunduil-chezmoi/pull/430
- https://github.com/alunduil/alunduil-chezmoi/issues/428, https://github.com/alunduil/alunduil-infrastructure/issues/310
- https://github.com/alunduil/alunduil-chezmoi/issues/429
- https://github.com/alunduil/alunduil-chezmoi/issues/444
- Archived: https://read.readwise.io/read/01ky3rh0m5yr8anhwvnv5pj9nm, https://read.readwise.io/read/01ky5k9gtsqzm28sv4nqv7996n
- Related: #66, #68