Six core capabilities, in depth
RUSH Deep Task Engine
Plan, code and verify after plan approval. Execution depends on task configuration; a separate worktree is not guaranteed for every step.
- Review and approve the task plan first
- Orchestrate execution according to dependencies and configuration
- V1 static checks, V2 unit tests, and conditional V3 integration tests
AI Chat & Inline Editing
Multi-turn sidebar chat uses selected and retrieved context. Select code and press Ctrl+K for inline rewriting, then review the diff before accepting or rejecting.
- Long-task context is limited by model windows, retrieval scope and compression strategies
- Line-by-line diff review with block-level accept granularity
- Unified credit metering — one clear bill for every AI action
FIM Smart Completion
Suggestions use context before and after the cursor. Latency depends on the model and network.
- Review suggestions before accepting
- Send cursor context to the configured completion model
- The current entry point does not promise a default SQLite persistent cache or zero-token requests
Semantic Code Graph
Project symbol indexing and dependency analysis assist retrieval without guaranteeing complete file or relationship coverage.
- Check index coverage and retrieved results
- RUSH planning can use matching call relations; review impact analysis
- Index update speed depends on project size and runtime environment
MCP Tool Connections
Connect MCP services through stdio configuration and reload them after saving; the bundled Playwright configuration still requires its local runtime.
- The current reload path supports stdio services
- Install tool commands and dependencies yourself
- Review permissions, commands and each tool's data policy
Execution Environment & Security
Local commands run in a PTY, which is not a strong isolation boundary; cloud resources require deployed services and account permissions.
- Check command permissions and back up important projects
- Recovery features do not replace Git or independent backups
- Encrypted keys do not prevent model-request context from leaving the device