v1.21.0

What's New

This release adds first-class Kimi support, improves Antigravity setup behavior, documents Qwen Code, and introduces an external profile sync strategy for OpenCode.

Changes

Features

• Kimi support: Added Kimi as a first-class agent with SDD assets, modular prompt injection, MCP integration, and Kimi-specific uv preflight/remediation aligned with the real install flow.
• OpenCode external profile sync: Added support for syncing OpenCode with external profile managers while preserving orchestrator prompt behavior where needed.

Fixes

• Antigravity setup: Corrected the setup slug and initialized settings through the proper Gemini-compatible flow.
• Kimi install flow: Fixed Kimi preflight behavior so unsupported platforms fail with the right error, fresh installs fail early when uv is missing, and already-installed Kimi setups do not require uv for component-only flows.

Documentation

• Qwen Code docs: Added Qwen Code to the supported agents documentation.
• OpenCode docs: Documented external profile manager support.

Merged Pull Requests

• #251 — feat: kimi support
• #317 — fix(antigravity): correct setup slug and initialize settings
• #312 — docs: add Qwen Code to supported agents documentation
• #320 — feat(opencode): support external profile sync strategy

Upgrade

brew upgrade gentle-ai

Links

• Full changelog

v1.20.1

Bug Fix (community contribution by @MrMonkey09)

• Windows: Install/Sync no longer fail with "Access Denied" — WriteFileAtomic called fsync on the parent directory after every file write, which NTFS rejects with ACCESS_DENIED. Now tolerates only that specific ErrPermission on Windows while keeping all other errors fatal. (#293)
• Windows: symlink tests skip gracefully — tests requiring SeCreateSymbolicLinkPrivilege now skip with a clear message instead of failing on unprivileged Windows builds
• CONTRIBUTING.md — documents how to enable symlink privileges on Windows (Developer Mode, admin, or Group Policy)

Install / Upgrade

brew upgrade gentle-ai
gentle-ai sync

v1.20.0

What's New

Kiro IDE Support (#283)

Full first-class support for Kiro IDE as the 10th supported agent:

• Native SDD subagents (10 phase agents via steering files)
• Dedicated Kiro model configuration flow in the TUI
• Per-phase model assignment with KiroModelAssignments persisted in state.json
• Split-root layout support (~/.kiro/skills/ + project .kiro/)
• Platform-agnostic paths (macOS, Linux, Windows)
• New documentation: docs/kiro.md

Granular Uninstall Flow (#241)

New managed uninstall via TUI and CLI:

• Partial uninstall — remove specific components from specific agents
• Full uninstall — remove all managed configuration
• Clean Install — uninstall + sync (reset to fresh state)
• Binary removal — optionally delete the gentle-ai binary itself
• Homebrew-managed binaries detected and handled gracefully
• Backup snapshot created before any destructive operation
• Symlink rejection on all file operations for security

TUI: gentle-ai → Uninstall
CLI: gentle-ai uninstall --agent claude-code --component sdd

See Usage — CLI Flags (uninstall) for all flags.

Install / Upgrade

brew upgrade gentle-ai
gentle-ai sync

Links

• Full changelog
• Documentation
• OpenCode SDD Profiles guide
• Engram command reference

v1.19.1

Bug Fix

• Engram upgrade on Linux — fixed "text file busy" (ETXTBSY) error when upgrading engram while it is running as an MCP server. The binary download now uses atomic rename instead of direct overwrite, so the running process keeps its file descriptor to the old inode while new executions pick up the updated binary.

Install / Upgrade

brew upgrade gentle-ai
gentle-ai sync

v1.19.0

What's New

New Agents

• Qwen Code — full integration with Qwen Code agent (#263)
• Kilo Code — full integration with Kilo Code agent (#161)

New Command

• gentle-ai uninstall — safely remove Gentle AI managed files with --dry-run, --force, --purge-memory, and --keep-backups flags (#285)

Bug Fixes

• Upgrade reporting — engram dev builds and version-unknown tools are now surfaced in the upgrade report instead of being silently omitted. Version detection has a 10s timeout to prevent hangs. Dry-run summary no longer counts manual-hint items as pending upgrades (#286)
• OpenCode AGENTS.md preservation — persona injection now uses marker-based sections instead of replacing the entire file, preserving user-authored content. Legacy persona stripping hardened across all agent strategies (#280)
• Model assignment persistence — model assignments now persist correctly across sync runs (#265)
• Empty upgrade feedback — TUI now shows feedback when upgrade results are empty (#282)

Documentation

• Added OpenCode Profiles guide, Engram reference, and improved README discoverability

Install / Upgrade

brew install Gentleman-Programming/tap/gentle-ai   # first install
brew upgrade gentle-ai                              # upgrade
gentle-ai sync                                      # sync configs

v1.18.4

Bug Fix

• Neutral persona no longer leaks Argentine Spanish — Shared infrastructure assets (engram protocol, SDD orchestrator, Windsurf workflows) contained Rioplatense expressions (acordate, ¿Seguimos?, dale) that caused LLMs to infer voseo context and respond with "vos" even in neutral persona mode. All shared assets now use standard pan-Hispanic Spanish. Gentleman persona retains all Rioplatense expressions as designed.

Replacements

• acordate → removed (neutral recordar already present)
• ¿Seguimos? → ¿Continuamos? (all 7 editor-specific SDD orchestrators)
• dale / Dale → de acuerdo / De acuerdo (Windsurf workflow + orchestrator approval gates)

All changes are prompt/markdown only — zero Go code modifications.

v1.18.3

Bug Fixes

• Strict TDD Mode now enforced across sub-agent boundaries — The orchestrator injects strict_tdd status directly into sdd-apply and sdd-verify sub-agent prompts instead of relying on self-discovery via engram search. Added hard gate in sdd-apply: no silent fallback to Standard Mode when TDD is active. sdd-verify respects orchestrator-injected TDD mode as authoritative.

• Apply-progress preserved across batches — sdd-apply now uses read-merge-write semantics instead of blind topic_key upsert. The orchestrator passes previous apply-progress reference to continuation batches. Sub-agents must read and merge prior progress before saving, preserving cumulative task state across all batches.

All changes are prompt/markdown only — zero Go code modifications.

Closes #206, Closes #209

v1.18.2

Bug Fix

• Fixed background-agents plugin failing to load in OpenCode 1.3.13 — The plugin used ESM named imports from unique-names-generator, which publishes CJS as its main entry point without an exports field. OpenCode's plugin loader (unlike Node.js) cannot extract named exports from CJS modules. Switched to default import + destructure.

Closes #229

v1.18.1

v1.18.1 — Agent Builder Patch

Includes all v1.18.0 features (Agent Builder) plus post-merge fixes from Copilot review and bug reports.

Bug Fixes

• Configure models → Claude → Back navigation — pressing Back in the Claude model picker when reached from "Configure models" menu now correctly returns to the config menu instead of navigating through the install flow
• skill-registry not deployed to OpenCode — added to skill catalog and foundation presets so it gets installed to all agents
• Welcome menu "Create your own Agent" — now properly disabled with "(no agents)" suffix when no engine binaries are detected
• Generation cancel via Esc — Esc now actually cancels in-flight generation instead of being blocked
• Prompt screen hint — clarified that Tab continues, Enter adds newlines
• Installed agents passed to prompt — generation prompt now includes target agent context
• SDDNewPhase PhaseName — placeholder set before prompt composition so generation knows about the new phase
• Builtin name conflict warning — preview screen shows warning banner when name conflicts with a built-in skill
• Registry duplicate handling — updates existing entry instead of appending duplicates on re-install
• Rollback marks all results as failed — InstallResults correctly reflect rollback state
• Stderr capture in all engines — error messages now include the engine CLI's stderr output
• SDD injection wired — InjectSDDReference now called during installation for non-standalone modes
• Skill paths use filepath.Join — cross-platform safe path construction
• Builtin skills derived from catalog — conflict detection stays in sync automatically
• Nil guard before install — prevents panic if Generated is nil on preview screen

Upgrade

brew upgrade gentle-ai

v1.18.0

v1.18.0 — Agent Builder: Create Your Own Sub-Agents

Highlights

New feature: "Create your own Agent" — a guided 6-step TUI flow that lets you build custom AI sub-agents from natural language descriptions, preview the generated SKILL.md, and install it across all your configured agents in one shot.

How It Works

1. Select "Create your own Agent" from the Welcome menu
2. Choose which installed AI engine generates your skill (Claude Code, OpenCode, Gemini CLI, or Codex)
3. Describe what you want your agent to do in natural language
4. Optionally configure SDD integration (standalone, phase support, or new pipeline phase)
5. Preview the generated SKILL.md with full metadata
6. Install — the skill is deployed to every configured agent automatically

What's New

• Agent Builder TUI — 8 new screens with full j/k navigation, multi-line textarea input, scrollable preview, and progress indicators
• Generation Engine — GenerationEngine interface wrapping installed AI CLI tools (claude --print, opencode run, gemini -p, codex exec) with stderr capture for clear error messages
• Cross-Agent Installer — atomic skill writer with rollback on partial failure, deploys to Claude Code, OpenCode, Gemini CLI, Codex, and more
• Custom Agent Registry — JSON registry at ~/.config/gentle-ai/custom-agents.json with builtin name conflict detection (auto-derives from catalog)
• SDD Integration — custom agents can support existing SDD phases or add new ones to the pipeline, with marker-based system prompt injection
• Generation Cancel — Esc cancels in-flight generation via context cancellation
• skill-registry skill — now deployed to all agents (was missing from OpenCode)

Bug Fixes

• Fixed: "Configure models" → Claude models → Back would navigate through install flow instead of returning to config menu
• Fixed: Welcome menu "Create your own Agent" now properly disables when no engine binaries are detected
• Fixed: Preview screen shows install error message when installation fails
• Fixed: Registry creates parent directory on first use
• Fixed: Preview scroll uses arrow keys, j/k navigates action bar (help text updated)

Install / Upgrade

brew upgrade gentle-ai

Or run gentle-ai and select "Upgrade tools".