v5.3.0

5.3.0 (2026-01-12)

Bug Fixes

• agent: filter JSON from error extraction to prevent false matches (b5833aa)
• agents: use modelLevel instead of model in git-pusher-agent (646da10)
• clear require-await lint errors (c96c188)
• conductor: add robustness to prevent silent hang on hook failure (5453958)
• inject completion agent after template loading (0d92a36)
• openai: Codex CLI structured output support (999c55b)
• orchestrator: honor --pr in foreground runs (710466a)
• persist task status and stale recovery (56f51e0)
• preflight: use cross-platform command detection for Windows (adf599c), closes #47
• race: poll log file before checking stale output + auto-clean stale locks (78486e2)
• replace cpu-blocking spin-wait with async file locking (f5bfb0d)
• task-executor: handle stale task status when watcher dies (026357b)
• testers: require behavioral testing, not code review (27f9e08), closes #828 #829 #830
• validator: recognize zero-length fallback as valid role handling (474ca9b)

Features

• add minModel setting and markdown file input (e689472), closes #42
• add model override and fix --pr completion agent (36fb714)
• add multi-provider CLI support (claude/codex/gemini) (301fc5f)
• validators: add debugging methodology checks to prevent shortcuts (85bc462)