Release v0.22.3 · tuanle96/agent-harness-kit

Summary

Bugfix patch for three open issues. All fixes ship as source templates + regression tests.

#40 — skill-permission-guard no longer bricks orchestration skills. The guard stops treating skill_invoked telemetry as a sticky active skill, and a skill with no policy entry now defers to task/default tool gating instead of clamping to the read-only default allow list (so e.g. deep-research can issue its own Workflow/Agent calls). Note: per-skill enforcement is now dormant unless AHK_ACTIVE_SKILL is set, since nothing in the kit emits skill_active/skill_started yet.
#38 — Codex SubagentStop advisor detection. The hook separates advisor role candidates from identity candidates and probes nested payload metadata (metadata.agent_name, agent.role, tool_input.agent.name, …). UUID-only payloads with no advisor role still write no proof (anti-forgery preserved). For a deterministic Codex advisor proof, use .harness/scripts/codex-advisor-run.mjs.
#39 — Flaky e2e-codex-cli reviewer artifact probe. The driver splits the deterministic runtime-surface smoke from the live-LLM reviewer-artifact proof; the surface-read probe is non-blocking unless AHK_E2E_CODEX_REQUIRE_SURFACE_READ=1.

Also corrects CHANGELOG bookkeeping for v0.22.2 (#35 + active-task evidence isolation).