These features are available in 3.0 without migration — none require changes to existing configs. Adopt them incrementally.
Press Shift+Tab to toggle Plan mode. The Plan agent analyzes your request, creates a step-by-step plan, and executes it with checkpoints between steps.
On first open, CLI prompts you to trust the workspace. In untrusted workspaces, the agent restricts capabilities (no shell, limited writes) until you explicitly trust it.
Use /compact to manually trigger context compaction when your session gets long. The engine generates a structured summary (preserving task status, file paths, key decisions, and next steps) and persists it as a checkpoint marker. On session reload, only the summary loads rather than replaying full history.
Powers installed via the IDE are automatically detected by CLI sessions. Install a power once in the IDE, and it's available everywhere.
Create reusable agent capabilities in .kiro/skills/<name>/SKILL.md. Skills are automatically discovered and available to all agents in the workspace.
Steering documents now support front matter with inclusion modes:
--- inclusion: fileMatch fileMatchPattern: "**/*.tsx" --- # React Guidelines ...
Modes: always, fileMatch, manual, auto. (The auto mode is determined by the engine based on context — not directly configurable in file front matter.)
Save and restore sessions for portability across machines or for sharing context. Session data is stored in ~/.kiro/sessions/.
Add a .kiroignore file to exclude paths from agent context. Works like .gitignore — the agent won't read or index matched files.
Sub-agents run complex tasks in parallel with isolated contexts. You interact with the primary agent; sub-agents execute behind the scenes. In the IDE, a visual tree shows progress across all branches.
Out of the box, without any permissions.yaml, the agent has these defaults:
fs_read ./**), common git read-only commands (git status, git log, git diff, etc.), system info commands (pwd, whoami, uname)~/.kiro/settings/ or .kiro/settings/ (prevents the agent from modifying its own trust boundaries)ask)Add a permissions.yaml to pre-approve or block specific operations.
Markdown format with embedded MCP servers, permissions, resources, and a full system prompt as the document body.
CLI 3.0 includes these built-in tools (use short aliases in agent profiles and permissions):
| Tool | Description |
|---|---|
read / read_file | Read file contents |
write / fs_write | Create or edit files |
glob / list_directory | Find files by pattern |
grep / grep_search | Search file contents |
shell | Execute terminal commands |
web_fetch | Fetch URL content |
web_search | Web search |
code | Code Intelligence — LSP-powered go-to-definition, find references, symbol search |
knowledge | Semantic indexing and search across sessions |
todo | In-session task tracking (gated — IDE visualization coming) |
Note: Checkpoints are not available in CLI 3.0. They remain an IDE-only feature. Use session export for state preservation.
| Gap | Detail |
|---|---|
| AL2 not supported | Kiro CLI 3.0 does not run on Amazon Linux 2. Use CLI 2.x if your environment requires AL2. |
| Classic mode not supported | The legacy non-TUI mode (kiro-cli chat without the TUI) does not support the v3 engine. Use the TUI. |
| Session resume | V3 sessions cannot be resumed in V2. If you switch back to the V2 engine, previously created V3 sessions will not be available. |
New features in 3.0