v1.42.0

Highlights

• Added Judgment Day profile rows to the OpenCode profile picker.
• Suggested OPENCODE_EXPERIMENTAL=true when OpenCode is selected and the experimental flag is missing.
• Fixed Windows Engram install flows when stale engram.exe entries shadow the managed binary on PATH.
• Fixed Windows Codex config generation by skipping the fatal /nix/store filesystem permission rule.
• Fixed Windows Engram clean-install stop-script behavior.

Validation

• Main CI passed before tagging.
• GoReleaser completed successfully.
• Release assets were uploaded for macOS, Linux, and Windows.
• Homebrew tap formula updated to 1.42.0.

Upgrade

brew upgrade gentle-ai

v1.41.0

✨ Highlights

• Reworked update experience (#875, #878, #879, #880) — the CLI now prompts for self-update by default (instead of silently auto-applying), shows a Codex-style update prompt before the Welcome screen, surfaces an informational remote advisory at launch, and caches the update check with a 6h cooldown to stop hammering the GitHub API.
• Beta channel support — opt into pre-release builds per OS via --channel beta, with the channel correctly inferred and carried through self-update on every platform, including the Windows installer.
• Profile-scoped judgment-day agents in OpenCode (#868) — judgment-day agents are now scoped per profile, and Hermes gained an ephemeral delegation skill so it orchestrates via native subagents instead of exhausting its context inline.
• Trigger-rules review system — a new declarative rule set decides when heavier review (4R) kicks in, injected into every generated agent asset, with validation and hot-path glob bindings.

🚀 Features

• update: prompt for self-update by default on the CLI; non-TTY auto-declines, GENTLE_AI_YES=1 auto-accepts (#878)
• update: Codex-style update prompt before the Welcome screen, defaulting to "Keep current" so a stray Enter is safe (#880)
• update: informational remote advisory fetched at launch and shown on Welcome (no version gate, no forced update) (#879)
• update: cache the update check with a 6h cooldown via LastUpdateCheck in state.json (#875)
• upgrade: complete the sync automatically after a gentle-ai self-upgrade instead of forcing a manual re-run (#877)
• upgrade: honor GENTLE_AI_CHANNEL when upgrading Engram (#876)
• engram: always fetch the latest engram-core and gentle-engram so new Engram releases reach users on their next upgrade (#874)
• opencode: support profile-scoped judgment-day agents
• hermes: add an ephemeral delegation skill and orchestrator guidance so Hermes delegates via native subagents (#868)
• sdd: add an automatic-mode gatekeeper that validates each phase before launching the next (#871)
• sdd: inject a trigger-rules section into all generated agent assets
• catalog/model: add the trigger-rules type model, default rule set, validator, and payments hot-path glob binding
• skills: bundle the required workflow skills

🐛 Fixes

• opencode: migrate the legacy command prompt field to template, fixing an OpenCode startup crash (ConfigInvalidError) on strict schemas
• opencode: enable judgment day in the single overlay
• tui: repair Codex and SDD-mode Back navigation in the model picker flow
• tui: align DependencyTree Back Esc/Enter paths and unify forward/back picker navigation
• tui: show update progress and wrap the advisory text correctly
• update: detect beta updates on main and infer the beta channel from pseudo-versions
• update: pass the beta channel through to the Windows installer self-update
• install: add an npm prerequisite gate and a graceful Engram stop on Windows (#866)
• install: use the -f operator in install.ps1 to support PowerShell 5.1 (#865)
• install: use portable lowercase for the go owner so --channel beta works on bash 3.2 / macOS (#892)
• install: keep the PowerShell installer ASCII-only
• sdd: scope the native dispatcher guard to file-based artifact stores so it no longer freezes Engram-backed flows
• catalog: enforce a positive MinDiffLines and prohibit 4R on everyday-always bindings

🙌 Contributors

Thanks to the community contributors in this release:

• @serberato and @skiveldev — Windows stop-script regression coverage (#884, issue #815)

Everything else is a core maintainer release.

📦 Upgrade

If Homebrew reports the tap as untrusted, trust the Gentle AI tap once before upgrading:

brew trust gentleman-programming/tap
brew upgrade gentle-ai

Or trust only the formula:

brew trust --formula gentleman-programming/tap/gentle-ai
brew upgrade gentle-ai

(or self-update on the next run)

To try pre-release builds, install on the beta channel:

curl -fsSL https://raw.githubusercontent.com/Gentleman-Programming/gentle-ai/main/install.sh | bash -s -- --channel beta

Full changelog: v1.40.2...v1.41.0

Advisory feed

Informational advisory feed consumed by gentle-ai at launch (fail-open). Edit the advisory.json asset in-place to broadcast a one-line message to users; clear the message to stop showing it. Not a version release.

v1.40.2

Changelog

• 6609179 fix(deps): pin core engram to 1.16.3
• 3dafd0f fix(upgrade): trust engram Homebrew cask

v1.40.1

Changelog

• c8f7f86 feat(agents): add 4R review agents

v1.39.4

Changelog

• f51090f fix(sync): align permission verification targets

v1.39.3

Changelog

• 8021aa1 feat(claude): configure subagent effort per phase (#818)
• a7bfcca fix(codex): improve permission profile defaults
• 940c8fb fix(opencode): remove legacy background plugin on sync

v1.39.2

✨ Highlights

• Homebrew upgrades now handle tap trust automatically (#826) — gentle-ai upgrade trusts only the Gentle AI-owned formula before running brew upgrade, matching Homebrew's new tap-trust enforcement without trusting whole taps.
• Linux Bubblewrap failures now explain the real security tradeoff (#826) — when Homebrew cannot create a rootless sandbox, Gentle AI shows namespace/AppArmor guidance without changing host policy automatically.
• TUI upgrade flows restart cleanly after updating Gentle AI itself (#826) — upgrade and upgrade+sync now hand off to the existing platform restart helper after the user acknowledges the result.

🐛 Fixes

• update: run scoped brew trust --formula gentleman-programming/tap/<tool> before Homebrew upgrades for Gentle AI-owned tools (#826)
• update: add actionable Homebrew tap-trust and Bubblewrap/rootless sandbox error guidance (#826)
• tui: quit upgrade result screens after gentle-ai itself was upgraded so the app layer can restart/re-exec safely (#826)
• install: stop swallowing real brew upgrade failures as “already latest” and surface the actual Homebrew output (#826)

🧪 Reliability

• update: add regression coverage for exact formula-level trust targets and Bubblewrap guidance wording (#826)
• tui/app: add regression coverage that Enter/Esc produce tea.QuitMsg and trigger the restart helper after a Gentle AI upgrade (#826)
• scripts: shellcheck coverage for the updated installer path (#826)

🔐 Security Notes

• Gentle AI does not auto-run sudo sysctl or relax AppArmor/kernel namespace policy.
• Bubblewrap namespace changes are shown as an explicit admin/security decision.
• HOMEBREW_NO_SANDBOX_LINUX=1 remains documented only as a final workaround because it disables Homebrew's Linux sandbox for that command.

🙌 Contributors

Core maintainer release.

📦 Upgrade

If Homebrew reports the tap as untrusted, trust only the Gentle AI formula once before upgrading:

brew trust --formula gentleman-programming/tap/gentle-ai
brew upgrade gentle-ai

If your machine already trusts the tap/formula, just run:

brew upgrade gentle-ai

(or self-update on the next run)

Full changelog: v1.39.1...v1.39.2

v1.39.1

✨ Highlights

• OpenCode now refreshes the skill registry at startup (#824) — installs a managed startup plugin that runs gentle-ai skill-registry refresh so SDD delegation gets fresh project skill paths without manual refresh.
• Repository cleanup (#822) — stops tracking the compiled gentle-ai binary in git while keeping release binaries handled by GoReleaser.

🐛 Fixes

• opencode: add a managed skill-registry.ts startup plugin using execFile with fixed argv, avoiding shell interpolation (#824)
• opencode: install the skill-registry plugin during SDD sync and remove it during SDD uninstall (#824)

🧹 Maintenance

• repo: stop tracking the compiled cmd/gentle-ai/gentle-ai binary (#822)

🧪 Reliability

• opencode: add asset, injection, and uninstall coverage for the managed startup plugin (#824)
• opencode: verified locally by removing .atl/skill-registry.md, starting OpenCode, and confirming the registry was regenerated (#824)

🙌 Contributors

Core maintainer release.

📦 Upgrade

If Homebrew reports the tap as untrusted, trust the Gentle AI tap once before upgrading:

brew trust gentleman-programming/tap
brew upgrade gentle-ai

Or trust only the formula:

brew trust --formula gentleman-programming/tap/gentle-ai
brew upgrade gentle-ai

(or self-update on the next run)

Full changelog: v1.39.0...v1.39.1

v1.39.0

✨ Highlights

• Codex now supports model + effort selection per carril (#816) — choose the model and reasoning effort used by the strong/mid/cheap SDD profiles, with the selected values actually written into the runtime Codex profile configs.
• Codex Custom mode now supports per-phase model selection (#820) — assign models per SDD phase, persist them across TUI/app/CLI sync paths, and pass both model and reasoning_effort to spawned phase agents.

🚀 Features

• codex: add per-carril model + effort selection for SDD profiles (#816)
• codex: add Custom per-phase model selection with searchable model/effort picker (#820)
• codex: render Model columns in generated SDD guidance and goldens (#816, #820)
• codex: pass model, reasoning_effort, and fork_turns: "none" during SDD phase delegation (#820)

🧪 Reliability

• codex: preserve selected carril and per-phase model assignments across run, app/TUI sync, and plain gentle-ai sync --agents codex (#816, #820)
• codex: clear stale per-phase assignments when returning from Custom mode to presets (#820)
• codex: add regression coverage for exact model rows, Custom navigation, state persistence, sync restoration, and injection paths (#820)
• codex: Judgment Day dual review completed through Round 5 with final APPROVE from both judges (#820)

🙌 Contributors

Core maintainer release.

📦 Upgrade

If Homebrew reports the tap as untrusted, trust the Gentle AI tap once before upgrading:

brew trust gentleman-programming/tap
brew upgrade gentle-ai

Or trust only the formula:

brew trust --formula gentleman-programming/tap/gentle-ai
brew upgrade gentle-ai

(or self-update on the next run)

Full changelog: v1.38.0...v1.39.0