loop-testing v0.1.1

Post-v0.1.0 hardening: CI, Codex unattended execution, and validation of the decision-triage rule.

• CI — GitHub Actions runs the full test suite (plugin-manifest JSON validation + tests/run-all.sh) on push, tags, PRs, and manual dispatch.
• Codex unattended driver (skills/loop-testing/scripts/unattended-codex.sh) — the Codex counterpart to unattended-loop.sh. codex exec is single-shot, so a long loop can end a session before converging; the driver relaunches codex exec to resume from STATE.md until a terminal status, with a per-session wall-clock watchdog, the same no-progress / max-sessions / max-minutes circuit breakers, and read-only protection of the installed skill during full-access sessions. 16 stub-driven tests.
• Codex full-loop validated — a real Codex session drove a complete QA loop to legitimate convergence (round 5, two clean rounds) with no mechanism-layer gate, on prompt discipline alone.
• Decision-triage rule validated — with the hardened skill installed, a planted "silently accept a negative value" ambiguity was correctly filed as NEEDS_CONFIRMATION + routed to a decision record, instead of being silently accepted. The earlier cross-platform miss was traced to a stale Codex install (F7), now documented: re-run install-codex.sh after updates.
• Fixes — driver arg parsing no longer hangs on a trailing value-flag (fixed in both drivers, regression-guarded).

Known limitations updated in README: Codex has no mechanism-layer gate (prompt discipline only, n=1 convergence evidence); the triage rule errs toward over-filing; Codex degrades MoA to single-model unless external access is explicitly authorized; the driver's cross-session resume path is stub-tested but not yet exercised in a real multi-session run.

loop-testing v0.1.0

First release: dual-platform autonomous QA loop skill.

• Claude Code plugin — /plugin marketplace add sdsrss/loop-testing → /plugin install loop-testing; Stop-hook mechanism layer auto-loads (verified in both /plugin install and --plugin-dir modes).
• Codex — bash install/install-codex.sh (same SKILL.md, verified in a real codex-cli session).
• MoA decision engine — zero-dep Node ≥ 20, OpenAI/OpenRouter, proxy-first, calibrated top-tier defaults, token-consumption guards, graceful degradation.
• Unattended headless driver — skills/loop-testing/scripts/unattended-loop.sh resumes claude -p sessions until natural convergence (env sanitization + watchdog + circuit breakers).

Acceptance: full unbounded runs on a CLI and a REST API sample both converged naturally in 4 rounds — 4/4 seeded bugs found on each, 14 additional real bugs discovered, every fix replay-verified with an atomic commit. See CHANGELOG.md for details and known limitations.