Highlights
/init-deep - Hierarchical Knowledge Base Generator
Automatically generates nested AGENTS.md files across your project. Analyzes directory complexity to create appropriately-sized documentation at the right depth, enabling AI agents to automatically discover relevant context for their current work.
/init-deep # Analyze and generate hierarchical AGENTS.md
/init-deep --create-new # Force create from scratch
/init-deep --max-depth=2 # Limit directory depth- Use
disabled_commandsconfig to disable builtin commands
Comment Checker - Token-Efficient Code Review
Updated default prompt to reduce unnecessary token usage. Custom prompts are also supported:
{
"comment_checker": {
"custom_prompt": "Found {{comments}} comments. Review briefly, fix only critical issues."
}
}DCP Recovery - Aggressive Context Pruning
When token limit is exceeded, DCP (Dynamic Context Pruning) now runs FIRST to attempt session recovery before compaction. Personally, I've never had a session break while using this feature, but it's still experimental so it's provided as a flag:
{
"experimental": {
"dcp_for_compaction": true
}
}New Features
Todo Continuation Enforcer Redesign
Complete state machine architecture for reliable task continuation:
- Version-token system prevents race conditions
errorBypassmode persists until user message- Plan mode agents (
plan,planner-sisyphus) now properly skipped - Minimum injection interval throttling (10s)
Enhanced Antigravity Token Refresh
Robust OAuth token handling with retry logic:
- Up to 3 retries with exponential backoff
- Proper error classification (
invalid_grant, network errors) - Retryable error detection (429, 500+)
Bug Fixes
- fix(dcp): correct storage path to match OpenCode's actual location
- fix(anthropic-auto-compact): use correct MESSAGE_STORAGE path
- fix(background-agent): prevent memory leak - completed tasks now removed from Map
- fix(session-manager): convert blocking sync I/O to async
- fix(session-recovery): improve empty message index search
- fix(lsp): improve error messages when LSP server is not installed
- fix: defer module-level side effects to prevent Bun 1.3.5 + macOS 15 segfault
Improvements
- Add ctx.metadata() calls for session navigation UI
- CLI: librarian/explore model fallback based on installer settings
- AGENTS.md knowledge base documentation files added
Commits
- bfb5d43 Add AGENTS.md knowledge base documentation files
- 385e8a9 Add builtin-commands feature with init-deep command and disabled_commands config option
- 7daabf9 Add ctx.metadata() calls for session navigation UI in background/subagent tasks
- 5fbcb88 fix(todo-continuation-enforcer): persist errorBypass mode until user sends message
- daa5f6e fix(todo-continuation-enforcer): redesign with version-token state machine
- 4d66ea9 fix(lsp): improve error messages when LSP server is not installed (#305)
- 7b7c143 fix(dcp): correct storage path to match OpenCode's actual location
- e3be656 fix: disable todo-continuation for plan mode agents (#303)
- c11cb2e fix: defer module-level side effects to prevent Bun 1.3.5 + macOS 15 segfault (#301)
- 195e8dc refactor(todo-continuation-enforcer): improve state machine and injection logic
- 284e7f5 fix(anthropic-auto-compact): use correct MESSAGE_STORAGE path for session messages
- 465c9e5 feat(comment-checker): pass custom_prompt to CLI
- 18d134f fix(background-agent): prevent memory leak - completed tasks now removed from Map (#302)
- 092718f fix(thinking-block-validator): handle text content parts in message validation
- 19f504f fix(session-recovery): improve empty message index search with expanded range
- 49f3be5 fix(session-manager): convert blocking sync I/O to async for improved concurrency
- 6d6102f fix(anthropic-auto-compact): sanitize empty messages before summarization
- 1d7e534 Upgrade @code-yeongyu/comment-checker from ^0.6.0 to ^0.6.1
- 17b7dd3 feat(cli): librarian/explore model fallback based on installer settings (#299)
- 889d80d feat(anthropic-auto-compact): run DCP first on token limit errors before compaction
- 87e229f feat(auth): enhance Antigravity token refresh with robust error handling and retry logic
Thank you to 1 community contributor: