Tool details

Tool descriptions stay short (every model call carries them); the details are here.

• kimo_get_state: when the context gives no docId. Top-level nodes (first 50), components (first 100), other agents, your permissions (data.you).
• kimo_open: a file that fails to load gives LOAD_FAILED (untouched); a newer format opens read-only; another process's write lease gives DOCUMENT_LOCKED; files outside your allow-list ask the user.
• kimo_create {path, template?}: never overwrites (FILE_EXISTS); template seeds variables and a starter frame. Needs the files.create grant (asks otherwise).
• kimo_close: a dirty document needs save or discard (asks the user). A document open in a window stays open: only your reference is dropped.
• kimo_save {docId, baseRevision?}: temp file → fsync → rename → read back → SHA-256 compare. Reported only after the read-back matched.
• kimo_history: list shows recent op groups with their author. undo / redo act on your own origin's stack only, never on the user's or another agent's edits; with token (an undoToken) undo succeeds only when that group is on top. Replay is guarded: when later edits by others touched the same nodes the result says stale: true instead of clobbering them.
• kimo_focus: with no window it succeeds with a NO_WINDOW warning.

• kimo_query: select needs at least one of ids, under, topLevel, name (substring, case-insensitive), type, text, component (instances of), page; under: "root" is not allowed (use topLevel). depth 0–10 (elided children show as "…"). resolveVariables replaces $refs with values for theme (default: each node's effective theme); resolveInstances expands instance content as virtual nodes (instanceId/childId). include: bounds (computed boxes), problems (clipping, collapse, overflow, missing fonts/assets), geometry, origin (override origin), overrides. limit ≤ 1000, then cursor. The whole document is never dumped.
• kimo_get_variables: includeUsages counts bound nodes (slower). Edit variables with ops (create_variable, update_variable, rename_variable, delete_variable, theme_axis).

• kimo_apply_ops: see ops. Limits: 500 ops per call, of which at most 40 inserts (150 streamed: ops land while the model writes the call). Failures carry retryId (see ops). dryRun returns the would-be ops; render: true attaches a PNG of the changed top-level frames. Results leave out ops / inverse (a dry run keeps ops). data.problems lists problems in the changed layers: fix them next.
• kimo_import: svg (editable vector layers), image (hashed into assets/, placed as an image-filled rect), tokens-dtcg / tokens-css (variables, merged). Exactly one of path (absolute, allow-list applies) or text. Supports dryRun and baseRevision.
• kimo_assets: import hashes an image or font into assets/ and returns its ./assets/ reference for fills or fonts (no node is created; kimo_import kind=image places one). fonts: document, cached, local and (when allowed) Google families. Use check_font before referencing a font.
• kimo_styles {apply: true} needs scope edit plus the globals grant.

• kimo_render: the same CanvasKit engine as the canvas; always reflects every committed edit. scale ≤ 4, maxSize (longest side, default 1600) caps the image. format: "svg" returns standalone SVG per node in data.images[].svg (text laid out with the canvas fonts).
• kimo_export: needs the export grant. png (one <nodeId>.png per node) and kimo (the subtrees as a standalone file with the variables they use) need outDir. Code and token targets (react-tailwind, html-css, tokens-dtcg, tokens-css, tailwind-theme, svg, pdf) are written to outDir or returned inline (pdf as base64). Identical files are reported unchanged; nothing is written when any file conflicts. See guide/export.