agentleFS
Sign inStart free
MrinankaMrinanka/aegiscontext/CLAUDE.md

aegiscontext

CLAUDE.md1 starsChanged 11 days agoApache-2.0View on GitHub

An open-source (Apache-2.0) Claude Code plugin (mod) that hides secrets before the model sees them. Spec: @docs/engineering/spec.md. Plan: @docs/engineering/plan.md. Progress: docs/engineering/progress.md.

The file

We split this CLAUDE.md into the rules it gives an agent

10 rules. About 0 are common, 6 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈6

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  • could be automatic: a setting could enforce it
  1. Very specific:

    If a step spans sessions, commit a step <n> (wip): ...

    line 11
  2. Distinctive:

    Ask with AskUserQuestion, recommended option first and labelled "(Recommended)".

    line 19
  3. Very specific:

    Don't reopen an accepted decision in docs/engineering/decision-log.md without saying so.

    line 21
  4. Distinctive:

    Write the tests first, then the code.

    line 25
  5. Distinctive:

    Never change or delete a test to make it pass.

    line 26
  6. Distinctive:

    If the spec is unclear or seems wrong, stop and ask.

    line 27
  7. Very specific:

    Commit only in /ship (or a wip checkpoint I ask for).

    line 37
  8. Distinctive:

    Never force-push, amend published commits or rewrite history.

    line 38
  9. Distinctive:

    Never use real secrets. Ever.

    line 42
  10. Very specific:

    Check the plugin: claude plugin validate ./plugin

    line 51

≈ “Common” is an estimate: at least 1 in 5 of the 26,980 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
CLAUDE.md

AegisContext

An open-source (Apache-2.0) Claude Code plugin (mod) that hides secrets before the model sees them. Spec: @docs/engineering/spec.md. Plan: @docs/engineering/plan.md. Progress: docs/engineering/progress.md.

Workflow

  • 7 steps (docs/engineering/plan.md, "The 7 steps"); one step per session.
  • Loop: /plan-step (in plan mode) → I approve → tests first, then code → /verify → fix → /ship → /clear.
  • Steps 3, 6 and 8 end with an audit: /security-review and /code-review high before /ship.
  • If a step spans sessions, commit a step <n> (wip): ... checkpoint when I ask; no PR until /ship.
  • Spec and plan: propose a change, edit only after I say yes, and log it as a decision.
  • Parked features go to spec §23; bugs and unfinished bits go to progress "Known issues". Nothing deferred lives only in session notes.
  • [Step 0] items in the spec are unconfirmed. Don't build on them until docs/engineering/findings-step0.md confirms them.

Questions

  • Research first (docs/, code, the mod API types named by the plugin-authoring skill, official docs, a quick scratch experiment). Don't ask what you can find out.
  • Ask with AskUserQuestion, recommended option first and labelled "(Recommended)". Give each option's main pro and con.
  • Back the recommendation with at least one source actually read or run this session (file:line, spec §/R-id, D-id, findings row, docs URL, test output). Never invent sources or numbers. Label anything from memory "(unverified)". If evidence is weak, say so and propose the cheapest way to get it.
  • Don't reopen an accepted decision in docs/engineering/decision-log.md without saying so.

Rules

  • Write the tests first, then the code. Run the tests before saying you're done.
  • Never change or delete a test to make it pass. If a test seems wrong, stop and ask.
  • If the spec is unclear or seems wrong, stop and ask. Don't guess.
  • TypeScript only. No runtime libraries in the plugin; everything that runs is in this repo.
  • Dev dependencies (build and test tools) are pinned with a lockfile. Don't add one without asking.
  • The tool must never contact the network and must never write a secret to disk unencrypted (logs, reports and temp files included). One exception, only while the optional put-back setting is on (off by default, D-096): the tool may hand Claude Code's own Write or Edit a value to save into a file that already holds it. This rule is about the product. Claude may read official docs online while researching.
  • Code copied from anywhere else must be credited in THIRD_PARTY_LICENSES before it is merged. Today that is only the Gitleaks rules.
  • Cross-platform: the maintainer develops on Windows. Scripts are Node, not bash. Build paths with path.join, never hard-coded / or \.

Git

  • One branch per step: step-0 … step-8, from develop.
  • Commit only in /ship (or a wip checkpoint I ask for). /ship pushes the step branch (no force) and opens a PR into develop. I review and merge; Claude never merges.
  • Never force-push, amend published commits or rewrite history.

Fake secrets in tests

  • Never use real secrets. Ever.
  • Build fake keys at test time by joining pieces ("ghp_" + "abc..."), so GitHub secret scanning doesn't flag the repo.
  • Logs, notes and the docs/ files never contain a secret, not even a fake one written whole.

Commands

  • Tests: npm test (Node's built-in node:test, runs tests/**/*.test.ts directly; Node 24, see .nvmrc)
  • Typecheck: npm run typecheck (tsc -p ., no emit; covers src/, tools/, tests/, not plugin/ or experiments/)
  • Install dev tools: npm ci --ignore-scripts (exact versions from package-lock.json)
  • Check the plugin: claude plugin validate ./plugin
  • Plugin tests: claude plugin test ./plugin
  • Step 0 probe: claude plugin test ./experiments/step0-probe; run it with claude --plugin-dir "<repo>/experiments/step0-probe" from a scratch folder (log: <cwd>/.step0-probe/)
  • Install the pinned Claude Code for the kit: npm ci --ignore-scripts --prefix tools/leak-test/claude-pin (about 320 MB, its own lockfile)
  • Leak test kit: npm run leak-test:baseline (empty tool: every fake must be seen; add -- --require-pinned to refuse another version) and npm run leak-test (the tool in plugin/: no fake may be seen). -- --only <scenario> runs one. Exit codes: 0 pass, 1 leak or blind, 2 invalid run

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free