Docs · Use

Tools

Nineteen tools, all sandboxed to the project root. Path traversal is rejected, bash has a timeout and output cap, and tool results are truncated hard (20k chars, 30k in SWE mode) before they hit context. Consecutive read-only calls run concurrently; writes run serially.

Read-only: every mode

• readFile

  Read a file. Path must stay inside the project root.

• listDirectory / glob

  List dirs and match paths without leaving the sandbox.

• grep

  Ripgrep-style search with output caps before it hits context.

• codeMap

  Repo symbol overview (functions/classes/exports per file) for bug localization without reading everything.

• searchWeb

  Read-only lookup the agent can call in any mode.

• diffFile

  Preview a unified diff of proposed changes without applying them.

• todoRead

  Read the current task list. Available in every mode.

• skill

  Load a skill workflow by name. Available in every mode.

Build: BUILD, FIX (no shell), SWE

• writeFile

  Create or overwrite a file. Creates a checkpoint for /undo.

• editFile / batchEdit

  Exact-match replacements. Batch applies atomically across files.

• applyPatch

  Apply a unified diff (--- / +++ / @@ hunks) with strict context matching.

• bash

  Runs with a timeout and output cap. Shell passthrough via ! in TUI.

• runTests

  Test twin of bash: returns structured pass/fail lists (jest, pytest, mocha, TAP) instead of raw stdout.

• undoLastChange / redoLastUndo

  Step through write checkpoints across turns.

• todoWrite

  Overwrite the full task list, send ALL todos every call so the plan can't drift.

• spawnAgent

  Delegate a bounded subtask to a fresh subagent. Returns its summary, not a transcript. Depth limit 1.