feat(ai): execJsInEditor + editorPreferences + editorDocs

[abose]

Three new MCP tools that broaden the AI's reach over the editor beyond the fixed controlEditor operations. Companion to phoenix-pro's browser handlers.

mcp-editor-tools.js:

• execJsInEditor: caller-supplied JS evaluated in Phoenix's own JS space (parent window — distinct from execJsInLivePreview which targets the iframe). Same shape as execJsInLivePreview (code + optional timeoutMs, floored at 5000, default 10000, no upper limit). Description lists the __PR helpers and points the model at editorDocs before writing non-trivial JS, plus a hint to use takeScreenshot (with arbitrary CSS selector) to verify mutations visually.
• editorPreferences: list / get / set with explicit scope. Description teaches user-friendly scope names ("system-wide" / "for this project" / "just for this session") and tells the model to PICK a sensible scope per pref nature instead of reflexively offering all three. No platform-specific paths in the description — purely API-level.
• editorDocs: returns absolute apiDocsPath (the build-mirrored API reference) + apiDocsURL + featureDocsURL + sourceRepoURL. Tool DOES NOT fetch content — it's just a "where to look" index so the model Read / Greps the bundled docs and WebFetches the live ones.
• takeScreenshot description rewritten with a clean binary selector rule (live preview vs everything else) plus an arbitrary-CSS- selector option for verifying execJsInEditor mutations.
• getEditorState response now appends a fallback hint pointing the model at takeScreenshot for any UI question the JSON doesn't answer (Problems panel, sidebar, etc.).

claude-code-agent.js (system prompt + tool registration):

• allowedTools + agents.researcher / agents.coder gain the three new mcp__phoenix-editor__ tool names.
• New "live preview is" definition line so the model has a mental model of what the preview shows.
• Broadened getEditorState trigger: ALWAYS call it first on any question that references the user's current work — not just "what file am I on" but also implicit-context phrasings like "the page", "this layout", "the nav bar", "scroll down on the page". Stops the model from blind whole-codebase grep when context is one tool call away.
• Phoenix-vs-Claude-Code name-collision rule: when the user says "set / configure / change X" without naming a product, default to editorPreferences. Don't touch ~/.claude/, ~/.claude.json, or the built-in update-config skill unless the user explicitly mentions "Claude" / "Claude Code" / "SDK" / "agent". Skip and ask if genuinely ambiguous.
• Per-tool entries for execJsInEditor, editorPreferences, editorDocs added to the tool list.

build/api-docs-generator.js + .gitignore:

• copyApiDocsToSrcNode() mirrors docs/API-Reference/ → src-node/apiDocs/ in a finally block so a partial generation (e.g. one source file's jsdoc errored mid-batch) still ships whatever made it. Cross-platform via fs.cp + path.join.
• /src-node/apiDocs added to root .gitignore.

strings.js: AI_CHAT_TOOL_EDITOR_JS ("Inspecting editor"), AI_CHAT_TOOL_EDITOR_PREFERENCES ("Editor preferences"), AI_CHAT_TOOL_EDITOR_DOCS ("Editor docs").

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud