Skip to content

LLLL Guard 0.2.0: one engine, no silent passes

Choose a tag to compare

View all tags
@JettyMarquis JettyMarquis released this 03 Oct 07:56
· 16 commits to main since this release
Immutable release. Only release title and notes can be modified.
v0.2.0
218bace
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

LLLL Guard 0.2.0: one engine, no silent passes

Guard is now a single TypeScript engine (@layrix.ai/llll-guard, Node.js 20 or newer). The bash guard script is a small launcher that finds Node and the engine; it holds no rules. This release is a transition release: it is warn-first (see below), and 1.0 will fail closed.

What changed

The gate no longer passes when it could not look. Exit code 0 is a pass, 1 is a block, 2 means the guard could not run, and 2 is never a pass. Before: the bash script crashed or passed silently on macOS's bash 3.2, the release gate printed "Nothing to scan" and exited 0, and an unreadable policy or a git error could end in a pass.

Push gate

  • The pushed range comes from git's pre-push input (new branches, tags, force-pushes, deletions, several refs), not from the current branch's upstream.
  • Secrets: provider-format keys and private keys (including PKCS#8 and PGP) are never downgraded; values published on purpose (NEXT_PUBLIC_*, env templates) are reported, not blocked, by comparing the value itself. Docs and env templates are scanned.
  • Overrides work: token-scoped to one finding on one version of one file, expiring (14 days by default), confirmed in a terminal, logged. HARD rules cannot be overridden.
  • Changes to llll.policy.json, .guardignore or llll.whitelist.json raise PG-S011. The policy is read from the remote's default branch, so a push cannot switch the gate off for itself.

Release gate: reads the files npm pack would publish (with --ignore-scripts), every file, any encoding; inline source maps, internal files, secrets; a file it could not read is a finding (RG-S008).

Hook: works in git worktrees and with core.hooksPath; keeps your own pre-push hook and runs it after the guard; no network npx; uninstall-hook restores it. husky, overcommit and lefthook are not chained: the installer prints the line to add.

Installers: Claude Code links the skill files only; opencode and Codex installers can upgrade and --uninstall, never truncate your config, and back files up.

Skill: one folding rule (round(N/2)), the credential downgrade rule matches the engine, corrected scan patterns, allowed-tools narrowed to specific commands plus llll-guard push|release, and /llll guard runs the engine first.

Upgrading

npm install -g @layrix.ai/llll-guard      # Node.js 20 or newer
cd your-project && llll-guard install-hook

If you installed the old bash hook, git pull in ~/.llll and run ~/.llll/guard doctor --fix; the old hook keeps working and tells you to reinstall it.

Warn-first. If Node or the engine cannot be found when you push, the hook warns loudly, writes .llll/logs/guard-not-run.log and lets the push through. Set LLLL_GUARD_STRICT=1 to stop the push instead. Failing closed becomes the default in 1.0.

Known gaps

Legal and factual corrections in the examples and sample audits are pending an expert review; SKILL.md is still one large file.

Assets 3
Loading