v0.4.0

What's Changed

• Add GLM-5.2 harness support by @byapparov in #82
• chore: bump @aictrl/cli to 0.4.0 by @byapparov in #83
• feat(events): emit 5-way token breakdown + context-window utilization in message_complete by @byapparov in #87
• feat(run): emit tool_catalog NDJSON event at session start by @byapparov in #88
• fix(run): emit session_error for session failures by @byapparov in #81
• docs: add model version bump guidance by @byapparov in #84

Full Changelog: v0.3.7...v0.4.0

v0.3.7

Fixed

• Added --variant none support for OpenAI-compatible providers, including Ollama /v1 models. This sends reasoning_effort: "none" so local thinking models can disable hidden reasoning without custom model options.
• OpenAI-compatible reasoning models now expose the full reasoning-effort variant set: none, minimal, low, medium, high, and xhigh.

v0.3.6

What's Changed

• fix: preserve openai-compatible reasoning_effort option by @slegarraga in #76
• chore: bump @aictrl/cli to 0.3.6 by @byapparov in #77

New Contributors

• @slegarraga made their first contribution in #76

Full Changelog: v0.3.5...v0.3.6

v0.3.5

Supersedes partial v0.3.4. Ships #71 (run exit-code) and #72
(publish protocol resolver).

v0.3.4

Fixes the v0.3.3 publish regression (#72) and adds exit-code
propagation on auth/provider failures (#71). See PRs for details.

v0.3.3

What's Changed

• chore: remove dead packages/cli/script/publish.ts (#54) by @byapparov in #56
• ci: bump deprecated Node 20 actions before 2026-06-02 deadline (#55) by @byapparov in #59
• chore: remove dead postinstall.mjs that no-op'd on every install (#53) by @byapparov in #57
• chore(deps): bump minimatch + @modelcontextprotocol/sdk for advisories (#50) by @byapparov in #60
• ci: fail loud on npm publish errors that aren't EPUBLISHCONFLICT (#51) by @byapparov in #58
• feat(events): close NDJSON schema gaps for downstream consumers (#63) by @byapparov in #64
• chore: bump @aictrl/cli to 0.3.3 by @byapparov in #65
• chore: mark @aictrl/plugin private; stop publishing from CLI release by @byapparov in #67

Full Changelog: v0.3.2...v0.3.3

v0.3.2

Fixes

• Platform binary optionalDependencies regenerated at publish time — @aictrl/cli@0.3.0 shipped with @aictrl/cli-linux-x64 hard-pinned to 0.2.0, so vanilla upgrades silently kept running the stale native binary. The publish workflow now publishes every platform binary the build produces (linux/darwin/windows × x64/arm64 × musl/baseline variants) and regenerates @aictrl/cli's optionalDependencies from those manifests at release time, while preserving genuine runtime optionals (@parcel/watcher, bun-pty). (#46)
• Publish workflow no longer self-corrupts its npm installation — the previous npm install -g npm@latest step broke on every release since v0.3.0 because npm 11.11.0 (npm/cli#9008) removed the promise-retry dep, and installing over the runner's bundled npm tree left orphan arborist files still requiring it. Pinned to npm@11.10.1 (last release before the removal, still ≥ 11.5.0 for OIDC trusted publishing support). (#48, #49)

Note: the v0.3.1 tag on GitHub never actually published to npm due to the workflow bug above — this release supersedes it.

v0.3.1

⚠️ This release was never actually published to npm.
The publish workflow failed with a MODULE_NOT_FOUND: promise-retry error
caused by npm/cli#9008, which
silently broke the release pipeline. The GLM-5.1 support intended for
this release ships in v0.3.2,
which also contains the platform binary optionalDependencies fix from #46
and the workflow fix from #49.

Fixes

• GLM-5.1 model support — Added zai-coding-plan provider with custom loader and thinking config, enabling glm-5.1 which is currently only available via the ZhipuAI CodingPlan API.

v0.3.0

What's New

• GLM-5.1 model support — added temperature, thinking config, and test fixtures for the newly released GLM-5.1 (#39)
• Subagent tool_use events — emit tool_use events for subagent sessions, improving observability (#24)
• Headless automation improvements — improved headless automation and developer workflow guidance (#19)

v0.2.0

Changelog

0.2.0 (2026-03-10)

Features

• Progressive skill loading — Skills are no longer injected into every LLM turn. Descriptions appear in the tool schema; full content loads only when the model invokes the skill tool. (#16)
• Per-skill NDJSON events — New granular events for --format json consumers: skill_discovered, skill_loaded, skill_resource_loaded, replacing the batch skills_loaded event. (#16)
• EVENTS.md — Added documentation for all NDJSON events emitted by --format json. (#16)
• CI review via npm-installed CLI — Code review workflow now uses the published @aictrl/cli package instead of building from source. (#11)

Fixes

• CLI reliability hardening — 31 fixes across security, error handling, resource management, session reliability, data integrity, and operational quality. Includes path traversal prevention, symlink-aware containment, heredoc bypass detection, output buffer caps, MCP connection timeouts, OAuth file locking, and more. (#14)
• CI review agent isolation — Review agent now runs from an isolated git workspace to avoid interference with repo state. (#15)
• LICENSE detection — Made LICENSE file detectable by GitHub's license scanner.

Tests

• Skill event emission tests — Coverage for SkillDiscovered, SkillLoaded, and SkillResourceLoaded event lifecycle. (#17)
• Reliability regression tests — 30 test files with 2,766 assertions covering all hardening fixes. (#14)

Chore

• Removed OpenCode branding, rewritten README. (#13)