0.2.0

Added

• The plugin asks for its settings when it is enabled: the TypeSafe key (kept in the
  Keychain), the claims_done threshold, the decision log, shadow mode, and a regex for
  your own check command. Each arrives as CLAUDE_PLUGIN_OPTION_<NAME> and wins over the
  plain environment variable.
• Subagent work counts. The hook reads the subagents/ files next to the session
  transcript, keeps the ones carrying this turn's prompt id, and merges them into the turn
  by timestamp, so a turn that delegated its edits no longer looks like one that changed
  nothing.
• Shadow mode: the whole pipeline runs, the turn still ends, and a stop it would have
  blocked prints jev-belay would have blocked this turn: and the reason through
  systemMessage, which Claude Code shows to you and not to the model.
• The closing message is read from the Stop payload's last_assistant_message when the
  host sends it, so the 300 ms re-read is paid only on older hosts.
• belay.mjs stats (verdict counts, calls, cost, latency p50 and p90, the last 14 days),
  belay.mjs last (the last decision, rendered), and belay.mjs doctor (key, host
  version, hook registration, transcripts readable, a round trip through a fake; --live
  makes one real call).
• Two skills: /jev-belay:why explains the last decision, /jev-belay:doctor reads the
  doctor's output back and names the fix for anything that failed.
• Belt 2 reads nextest, deno, ruff, mypy, mocha, rspec, minitest, phpunit, playwright,
  swift, ctest and biome summaries, plus two compiler shapes, rustc's error[E0119]: and
  go's file.go:12:3:, which decide fail when no runner summary follows them. A compile
  error used to read as a passing check in a cargo check or a go build.
• test/runners/ holds a captured or documented output per runner belt 2 claims to know,
  passing and failing, so a belt change has something to fail against.
• A check command of your own: set the check option (or JEV_BELAY_CHECK) to a regex
  and belt 1 tests it next to the built-in runners. An invalid regex is ignored.

Changed

• tui.mjs holds the live view, the replay, stats, last and doctor. belay.mjs is
  the hook and imports it lazily, so the hook path never parses the TUI.
• .env.example is gone. Nothing ever read it, and the README told people to copy it.
• The claims_done threshold is 0.70, up from 0.65: on the re-extracted corpus, with
  subagent turns folded in, 0.70 is the lowest cutoff holding wrong blocks under 2%, and
  it catches the same 7 of 12 false dones.
• A check regex that matches the empty string is ignored like a broken one, since .*
  would make every Bash call a passing check. An RSpec error outside the examples is a
  failure whatever the failure count says. The rustc rule needs the numbered diagnostic or
  cargo's closing line, so git push is no longer quoted as a failed check.

Fixed

• A threshold that is not a number no longer disables blocking silently; it falls back to
  the default.
• The plugin install never delivered a key. The README pointed at a .env file that
  neither the hook nor Claude Code reads, so a plugin user got "no key", exit 0, forever,
  with nothing saying so.
• A compile error in a cargo check or a go build read as a passing check, which sent
  Jev a state saying a check passed and put the same claim in the nudge.

Measured

• The ablation on the re-extracted corpus (2,694 stops, 477 reaching the gate, 17.7%),
  same 100-stop audit slice with 12 false dones, direct API, jev-1.13.0, 2026-09-20:
  AUROC 0.777 for the wording alone, 0.886 with the evidence gate, 0.976 for the full
  hook, interval 0.915 to 1.000. On the 76 gated stops: 0.967 against 0.844. At 0.70:
  8 blocks in 100, 7 right. Median call 1,222 input tokens, $0.00005, 346 ms.