feat: intent-aware verify loop + one-command setup (v0.6.0)

[youcefzemmar]

What

Evolves the behavioral verify loop from a generic checklist into an intent-aware verifier, and makes it install in one command — without ever trapping the user.

When an agent finishes a work turn, the loop now:

1. Grounds the check in the actual request. It parses the human prompt that opened the turn and quotes it back, so the agent verifies against what was asked, not its own restatement.
2. Tailors the verification to the work. It infers the kind of work from the diff and leads with the right check:
   • web/UI → start it (npm run dev), open the URL, take a screenshot and READ it, compare pixel-for-intent, check the console
   • API → hit the affected endpoint(s), check status + body (incl. error paths)
   • CLI → run it with realistic args, check stdout/stderr + exit code
   • library → run the tests + a smoke call of the changed function
     It surfaces the concrete run command and local URL when discoverable (from package.json scripts + ports in the diff).
3. Can never trap you. New GROUNDTRUTH_NO_LOOP=1 kill-switch instantly pauses the loop regardless of config; the per-session round cap still guarantees a turn always finishes.

Detection only tailors the guidance — groundtruth still never judges the work, so it adds no new false positives (the project's core invariant).

Also

• groundtruth setup — one idempotent command: Stop hook + verify loop + SessionEnd digest + status line, global by default. install stays for fine-grained control.
• Public API: analyze() (returns report + evidence + turn), detectWorkKind(), summarizeRequest().
• Fixes the long-standing version drift: .claude-plugin/plugin.json 0.4.0 → 0.6.0, synced with package.json.
• CHANGELOG [Unreleased] (verify loop + precision fixes) cut to 0.6.0.

Tests

+16 tests (intent classification, request capture, grounded/tailored protocol). 136 pass, typecheck + biome clean. Behaviorally smoke-tested end-to-end: hook returns exit 2 when blocking, renders the web-tailored screenshot protocol with the right URL, and the kill-switch cleanly disables blocking.

🤖 Generated with Claude Code