Set viewport state with zoom clamping.

* Split DSL operations handling multi-line JSON bodies. Splits on newlines * that are OUTSIDE quoted strings and balanced brackets. This lets agents

* Try to start an arc drag. Returns true if an arc handle was hit.

* Try to start a resize from a handle hit. Returns true if a handle was hit.

* Try to start a rotation from a rotation zone hit. Returns true if hit.

Stream via Claude Agent SDK — emits text/thinking deltas as SSE

Stream via Claude Agent SDK (uses local Claude Code OAuth login, no API key needed)

* Stream via builtin provider — direct API key, no CLI tool needed. * Uses Zig NAPI addon (agent-native) with Anthropic or OpenAI-compatible provider

Stream via Codex CLI — emits text/error deltas as SSE

Stream via GitHub Copilot SDK (@github/copilot-sdk)

Stream via Gemini CLI — emits text deltas as SSE

Stream via Gemini CLI (`gemini -p -o stream-json`) — CLI handles its own auth

Stream via OpenCode SDK — emits text/thinking deltas as SSE

Stream via OpenCode SDK using event subscription for real-time streaming

Wrap an async generator with a timeout — yields values until timeout fires

* Strip "NEVER use tools" and similar instructions from system prompt * when we need Claude Code Agent SDK to use its Read tool for image analysis.

* Strip fields that don't influence code generation. Keeps request bodies small * enough that proxies don't reject them with 403/413, and reduces inp

* Strip non-standard XML-like tags that third-party models may inject * (e.g. `<minimax:tool_call>`, `<tool_call>`, `<|im_start|>`). * Preserves con

* Strip fake tool call blocks that basic-tier models (MiniMax, etc.) may emit. * These look like `{tool => "Write", args => { ... }}` and are not val

Strip raw tool-call / function-call XML that should never be shown to users

Convert any CSS color string to #rrggbb for <input type="color">.

Convert any CSS color string to #rrggbb for <input type="color">.

* Convert concatenated lowercase to kebab-case for icon name matching. * e.g. "arrowright" → "arrow-right", "chevrondown" → "chevron-down"

* Try to parse a single root PenNode with nested children from raw text. * Handles the case where weaker models output a single JSON object * instea

Same as upsertNodesToCanvas but skips sanitization (caller already did it).