v1.5.0 — Check before you trust

A stamp costs ~15 tokens. Re-verifying costs 15,000.

This release closes the loop that makes stamping worth it: agents (and humans) can now read trust state in one line instead of re-verifying work that was already verified.

New

akf check — the one-line agent trust check

$ akf check auth.py
OK trust=0.65 agent=claude-code evidence=test_pass age=1d claims=1

• OK (exit 0) — fresh stamp with verified evidence; build on it
• LOW / STALE (exit 1) — unverified, decayed, or modified after stamping; re-verify
• UNSTAMPED (exit 2) — no metadata
• --json for structured output; evidence upgrades authority (tests/CI/human review); also available as the 10th MCP tool check_file and as akf.check_file() in Python

akf init — zero-effort, default-on setup

One command wires: git post-commit stamping (via the new akf hook group), a Claude Code PostToolUse auto-stamp hook, and the AKF protocol stanza in your agent rules files (CLAUDE.md / AGENTS.md / .cursorrules / copilot-instructions). Idempotent; --only to scope.

akf stamp --preset memory|skill

• memory: 30-day trust half-life — stale agent memories decay below threshold and report LOW automatically instead of poisoning future sessions
• skill: supply-chain provenance for skill/plugin files; STALE = the file on disk is not the file the publisher stamped
• Recipes: agent memory · skill provenance

akf scan --badge

Shields.io endpoint JSON for any repo ("14% stamped · trust 0.76") — this repo now wears its own badge, auto-refreshed on every push.

Improved

• akf certify --format markdown (GitHub Action PR comments): headline verdict, failures first, evidence column, row cap for large repos
• Sidecar .akf.json files are self-documenting ("spec": "https://akf.dev")

Fixed

• Markdown frontmatter embedding destroyed nested YAML (tags lists, tool metadata) — non-AKF frontmatter is now preserved verbatim
• akf init --git-hooks installed a hook calling a command that didn't exist
• stamp_file silently dropped extra claim fields (kind, decay_half_life)
• "reviewed by @user" evidence now classifies as human_review as documented

Install

pip install akf          # Python — 1.5.0
npm install akf-format   # TypeScript — 1.5.0

Or paste the install block from the README to your agent and let it set itself up.