Keel Skills Changelog

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/quitohooded/keel-skills/llms.txt

Use this file to discover all available pages before exploring further.

​

[0.5.0] — 2026-06-30

• Renamed the three permission levels from the code names L1 / L2 / L3 to plain words: a goal (was L1), a method (was L2), and a green light (was L3). Other jargon was humanized too: mechanical propagation → “following through on a green light you already have”; four-step test → “four-step check”; source-of-truth artifact → “source-of-truth file”. Hot zone was kept (already intuitive).
• Applied across the skills, SPEC.md (bumped to spec 0.2 with a terms migration note), both READMEs, the policy template and packs, the enforcement hook’s user-facing reason strings, the examples (l3-brake.md → green-light-brake.md), the docs site, the diagrams, and the marketing copy.
• Historical changelog entries below are left as-is — they describe what shipped at the time, under the old names.

​

[0.4.0] — 2026-06-30

• PreToolUse hook (hooks/enforce-policy.cjs) — deterministic enforcement. Inspects every tool call before it runs and returns allow / ask / deny against the SPEC §4 hot defaults plus your policy’s concrete hot paths and commands. ask is the L3 event (an explicit human-approval prompt); in a non-interactive run (KEEL_NONINTERACTIVE=1 / CI) a hot action is denied since no human can grant L3. Standing approvals in the policy let scoped exceptions through. Fails open on malformed hook input so it never wedges a workflow. Dependency-free.
• Audit trail — every decision is appended to .keel/audit.jsonl (timestamp, tool, matched rule, verdict).
• AGENT_POLICY.md machine-readable block — an optional fenced ```keel-policy block (flat lists: hot_paths, hot_commands, standing_allow_commands, standing_allow_paths) that the hook parses, kept in the same single file as the prose. Documented in SPEC.md §7.1; added to the template.
• SPEC.md — §7.1 (machine-readable block) and §8.1 (enforcing implementations, optional & stricter) added, both runtime-neutral. The model is explicit that enforcement is a backstop, not a security boundary.
• Tests — hooks/enforce-policy.test.cjs, a dependency-free runner (16 cases) wired into CI (validate.yml).
• Docs — root and plugin READMEs gain a “two layers (judgment + enforcement)” section and the not-a-security-boundary caveat; authorization-protocol skill notes the hard backstop.
• No change to the existing skills’ decision logic or the SessionStart hook.

​

[0.3.0] — 2026-06-19

• License → MIT. Relicensed from source-available (Attribution, No-Resale) to MIT to enable free adoption, forking, and reimplementation. Attribution is preserved via the MIT notice and a new NOTICE file. This unblocks the distribution model the project’s growth strategy depends on.
• SPEC.md — the authorization model (L1/L2/L3, four-step test, hot zones, mechanical propagation) and the AGENT_POLICY.md format are now specified runtime-neutral, with a Conformance section, so they can be cited and reimplemented outside Claude Code (the “open standard” play).
• English is now the primary repo language (README.md); the Spanish version is preserved as README.es.md.
• policies/ — ready-made AGENT_POLICY.md packs for common stacks (web-app-deploy, nextjs-vercel, supabase) plus a registry index.
• examples/ — l3-brake.md (a concrete before/after of the brake) and demo-script.md (a recordable 60-second demo script).
• CONTRIBUTING.md — how to add policy packs, build Keel-compatible implementations, and improve the spec.
• STRATEGY.md — the growth strategy and roadmap (internal source of truth).
• No behavioral change to the three skills, the command, or the hook.

​

[0.2.0] — 2026-06-19

• SessionStart hook (hooks/hooks.json + hooks/inject-policy.cjs) — when a project has an AGENT_POLICY.md, its contents are injected into context at the start of every session, so the policy no longer depends on the agent choosing to read it. Cross-platform (runs via Node); stays silent when no policy is present.
• authorization-protocol — Step 0 now notes the policy may already be in context via the hook.
• context-discipline — now points to AGENT_POLICY.md’s “where state and decisions get recorded” section instead of leaving the location to guesswork.
• Docs — README embeds the authorization-model diagram (assets/authorization-flow.svg); DISTRIBUTION.md updated (repo already published, hook added to layout/checklist).

​

[0.1.0] — 2026-06-17

• authorization-protocol skill — three-level authorization model (broad mandate / mechanism / explicit scoped approval), four-step safety test, hot-zone defaults, mechanical-propagation rule, tie-breakers.
• model-delegation skill — model tiering by task type, subagent depth limit (2), no-self-escalation rule, no-subagent-approvals rule, cheapest-first tool ladder, encapsulate-repetition guidance.
• context-discipline skill — files as source of truth, when to stop expanding a session, resumable handoff requirements.
• /keel-skills:policy-init command and AGENT_POLICY.template.md — per-project configuration that keeps the framework generic and the buyer’s data separate.