v0.22.0

[0.22.0] - 2026-06-04

Added

• Work on multiple tasks at once in a single workspace. A new task registry
  (.harness/state/active-tasks.json = { focused, open[] }, driven by
  .harness/scripts/active-tasks.mjs — open / focus (alias switch) /
  close / list / focused / clear) keeps several tasks open with exactly
  one focused task. Opening a second task no longer clobbers the first, and
  switching is cheap. The focused id is mirrored into the legacy
  .harness/state/active-task.txt, so the skill/task permission guard, the Stop
  completion gate, and the advisor runtime-proof keep operating single-task — on
  the focused task — with no change to their logic; completion gates apply only
  to the focused task. SessionStart now surfaces the open-task landscape, and
  the model is backward-compatible (a project with only active-task.txt is read
  as one open + focused task). See .harness/docs/task-management.md.

Changed

• The skill/task permission guard now ships in warn (advisory) mode by
  default. Generated .claude/settings.json env and the Codex hook wrapper
  set AHK_SKILL_PERMISSIONS_MODE=warn, so out-of-policy tool calls are logged
  to .harness/bypass.log instead of blocked — the permission matrix stays out
  of the way during normal coding while keeping an audit trail. The guard's
  enforcement mechanism is unchanged; set AHK_SKILL_PERMISSIONS_MODE=enforce
  in .claude/settings.json env (or remove the key) to make it block again.
  Existing projects pick this up on agent-harness-kit upgrade. The always-on
  destructive-path rails (bash-guard, edit-guard) are unaffected.

---

Full history: CHANGELOG.md
Install: npx agent-harness-kit@0.22.0 init