Conversation
…probe
The 0.3 Agent SDK (bumped 2026-09-23) reads the append text only from
systemPrompt.append with the claude_code preset and silently ignores a
top-level appendSystemPrompt option, so none of the Phoenix instructions
reached the model. Pass systemPrompt: { type: "preset", preset:
"claude_code", append } instead.
Append a probe sentence: when a message carries "phoenix system prompt
challenge: <code>", the model replies with the code plus a fixed suffix.
The suffix is known only from the prompt and the code only from the
message, so the reply proves the prompt is in context for that request.
The EC-7 model test sends the challenge, and the builder MCP's quick
suite now runs it first so a missing prompt stops a run early.
Verified with the quick model-test suite (11 PASS, no regressions) and
two fresh-session challenge replies from Haiku.
Bootstrap's base rule sets code { white-space: nowrap } directly on the
element, so an inline path or identifier in an assistant message ran off
the panel edge instead of wrapping with the prose. Set pre-wrap on the
message and plan-body inline code rules; code blocks keep their own
pre-wrap from the pre code rule.
Registers a deferred, read-only getProblems tool on the phoenix-editor MCP server that calls the new browser peer. The description tells the model the problems come from the syntax checkers available for the file type, that respondedProviders shows which checkers answered, and that results are capped at 20 with pattern, type and provider filters. Allows the tool for the agent and its subagents, adds a line of prompt guidance, and adds the card strings.
useImage takes an optional downloadPath: a project folder, or for one photo an image file path. The description and the prompt guidance say to prefer embedding the Unsplash URLs and to download only when the user asks for local files or the use case needs them. The peer timeout allows for the download. Adds the strings for the viewer's download button, the download card and the download error cases, drops the unused request-failed string, and styles the viewer button.
…cation One helper drives every auto-closing notification: the countdown stops on mouseenter and restarts in full on mouseleave, so a user reading a toast or reaching for one of its actions does not lose it. Applies to the toasts, the anchored notifications, the inline toasts and the HUD. Adds mock-clock specs for the toast, the HUD and the inline toast that observe the timed close on the notification rather than on the CSS transition.
Extracts the "Show in File Tree" command body into an exported revealInFileTree(file): show the sidebar, switch to the Files tab, select the file and scroll it into view. When the tree does not have the node yet, as for a file just written outside the FileSystem API, it refreshes the tree and looks for the node for up to three seconds before selecting. ProjectManager gains isInFileTree(path) for that check. Adds the strings and styles for the image gallery's download toast and the search card's footer toggle.
Registers a deferred, read-only notifyUser tool that calls the new browser peer. The description tells the model to use it only when the user may not be watching the chat, that it is skipped while the AI panel is visible unless alwaysShow is set, and that autoCloseS controls how long the toast stays. Allows the tool for the main agent, adds a line of prompt guidance, the card string and the toast style.
…ch folder
Add the askInLivePreview MCP tool: the model writes only a body (uiFile,
scriptFile, params) and Phoenix draws the frame in the live preview. The
description states the contract in full: the script file runs as
function(root, phoenix, params) { <file text> }, anchor and highlight for
one element, placement out of the way by default, a theme for the frame,
hover previews and click answers.
Writes and edits into the model's UI folder under app data are allowed
without a prompt, skipped by the snapshot and edit events, and the folder
is recreated before a write. The guidance points the model at that folder
and no longer claims the editor state lists its files.
Strings for the card title, focus toggle, text field and the chat labels.
execJsInLivePreview and execJsInEditor take scriptFile and params, read
from the model's scratch folder the way askInLivePreview does: a file is
a function body that returns its value, inline code stays for a very
short throwaway, and an inline snippet with a top-level return runs as a
function body instead of failing. A file holding a whole function is
called as one.
The askInLivePreview description states the script file contract as the
template function(root, phoenix, params) { <file text> }, asks for a
visual body whose every word stays readable, and no longer names a file
list the editor state stopped reporting. The per-prompt guidance carries
the real scratch folder path so the model never has to look it up or
guess, and says never to write those files into the project.
Strings and styles for the AI panel's usage chip: the chip in the input area's mode row, its drop-up with an All tab, a tab per model and a gear tab holding the two display options, the line under each answer, and the usage shown on history rows. The subscription note reads "With a Claude subscription, this cost is included in your plan. Pricing shown is for reference only." The empty chip dims its icon and text rather than itself, so the drop-up inside it stays opaque. The askInLivePreview description now says to write both UI files before calling the tool; the smaller models were calling it first.
|
This branch has not been deployed
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.