Conversation
The JS SDK was split out of @anthropic-ai/claude-code in v2 and moved to @anthropic-ai/claude-agent-sdk. The CLI binary still ships under @anthropic-ai/claude-code (used by Phoenix's terminal "claude" command); the SDK now lives in its own package with the same query() signature and SDKResultMessage shape, so the rest of claude-code-agent.js is unchanged beyond the import swap. - @anthropic-ai/claude-code: ^1.0.0 → ^2.1.118 - @anthropic-ai/claude-agent-sdk: new ^0.2.126 - zod: ^3.25.76 → ^4.0.0 Lock files regenerated by npm install.
Two complaints about the takeScreenshot tool:
1. Claude was sometimes shooting the whole editor window when the
user obviously meant "how does the rendered preview look".
2. The user-visible label said "Screenshot of full page" which was
misleading — the default capture is the entire Phoenix editor app
window (toolbar + sidebar + code area + panels), not a rendered
web page.
Tightening on both fronts:
- mcp-editor-tools.js: tool description now starts "ALMOST ALWAYS pass
a selector — capturing the full editor returns a busy image full of
editor chrome that's hard to reason about" and explicitly lists the
trigger phrases ("how does it look", "is the page rendering",
"check the preview") that should fire #panel-live-preview-frame.
Selector schema unchanged — the IDs were already documented, the
recommendation just had to be louder.
- claude-code-agent.js system prompt: same trigger-phrase nudge plus
a last-resort line: when the user asks about something in the
editor you can't identify from getEditorState, take a screenshot
(no selector) to see what they're looking at.
- strings.js: rename AI_CHAT_TOOL_SCREENSHOT_FULL_PAGE to
AI_CHAT_TOOL_SCREENSHOT_FULL_EDITOR with value "the full editor"
so the chat bubble reads "Screenshot of the full editor" instead
of the misleading "Screenshot of full page".
execJsInLivePreview was in the "no timeout" bucket on the rationale that user-supplied JS can legitimately take a while. The downside: when the live preview is wedged or just slow to settle, the agent hangs indefinitely on "Inspecting preview" with no way to recover. Add a timeoutMs parameter to the tool. The model picks a value that fits the snippet it's running; the call is raced against the timeout and either resolves or surfaces a deterministic timeout error so the agent can move on. - _execPeerWithTimeout now accepts an optional overrideMs that wins over the static EXEC_PEER_TIMEOUT_MS map. - _resolveCallerTimeout floors at 5000ms (no point picking a tighter timeout — the preview frame may still be settling). No upper limit; a user can legitimately request a long-running inspection. - Default 10000ms when timeoutMs is omitted.
…rome-devtools
Two adjacent failure modes the previous prompt didn't catch:
1. When the user asked editor-context questions ("can you see the
open file", "in this page", "on the screen"), Claude sometimes
answered from generic Claude knowledge — "No, I don't have
visibility into your editor" — instead of calling getEditorState.
2. When Claude DID decide it needed to see the page, it sometimes
reached for the user's chrome-devtools MCP (which opens a fresh
separate browser session) instead of phoenix-editor.takeScreenshot
(which captures the live preview inside Phoenix, reflecting the
user's actual unsaved edits).
Add an upfront block to the system prompt that:
- Explicitly anti-patterns the "I can't see..." refusal — call
getEditorState / takeScreenshot / execJsInLivePreview instead.
- Explicitly steers ALL preview interactions (screenshots, JS eval,
DOM inspection, viewport resize, reload) to the phoenix-editor MCP
rather than chrome-devtools or other browser MCPs. Only fall back
to a non-Phoenix browser context when the user explicitly asks.
…itorState fallback hint Bundled set of editor-tool refinements driven by support traces: - getEditorState now reports inDesignMode (true when the code editor is hidden and the live preview is expanded full-bleed). The model needs this to interpret "I can't see your code edits" cases — design mode hides the code view, and that's a separate fact from the active file. - getEditorState response now appends a fallback hint pointing the model at takeScreenshot (no selector) for any UI question the JSON doesn't answer (e.g. "what's in the Problems panel" / "what does this sidebar say" — UI panels not represented in the state object). - takeScreenshot description rewritten to a clean binary rule — "rendered live preview" → selector='#panel-live-preview-frame', "anything else (Problems panel, file tree, toolbar, any Phoenix UI)" → no selector. Earlier "ALMOST ALWAYS pass a selector" framing was over-discouraging the legit no-selector case where the user is asking about Phoenix UI. - controlEditor gains a toggleDesignMode operation with an `enabled` boolean. The op description spells out what design mode is (full live preview, code editor hidden, content-focused browser-like view) so the model picks it for the right intents. - System prompt: explains what the live preview actually is (the rendered view of the active HTML/CSS/JS/SVG/Markdown file), so Claude has a mental model before reading the per-tool sections.
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.