v3.14.5: Trustworthy exit codes, working platform setup, stable audit chains

What's New

Results you can trust in scripts and CI. Commands that fail now exit non-zero: platform verify, plugin install/remove, commands whose initialization fails, and task submit with an unknown domain or no --wait. aqe platform setup works again for all eight supported platforms.

Codex setup that works from a global install. aqe init --with-codex installs the Codex hooks and QE skills whether aqe comes from a global install, a local install or npx. --codex-guidance modes apply to an existing AGENTS.md and leave your own text exactly as it was.

An audit chain that stays valid. Concurrent writers (an MCP server plus hooks, or two agent hosts) no longer fork the chain. brain import records the merge instead of splicing in foreign audit rows.

Learning that lands in the right place. With an embedder configured, the pattern search index binds in every process, and the provenance guard still refuses unverified embeddings. Running from a subfolder uses the project's root store, and RUVECTOR_USE_* overrides take effect.

A quieter CLI. Internal startup logging follows AQE_LOG_LEVEL / LOG_LEVEL (default warn) on stderr, and aqe memory commands no longer boot the whole fleet.

SemVer plugins. Plugin versions, dependency ranges and minAqeVersion follow npm SemVer, and a rejected plugin no longer stops the others from loading.

Upgrade notes cover task submit --wait, Codex guidance on an existing AGENTS.md, per-submodule stores, RUVECTOR_USE_*, LOG_LEVEL, and stricter plugin versions: see docs/releases/v3.14.5.md.

Thanks to @pacphi for ten precise, reproducible issue reports, and to @rudycelekli for most of the fixes.

Getting Started

npx agentic-qe init --auto

See CHANGELOG for full details.