Harden Murph chat continuity#109
Conversation
|
@codex Review |
|
To use Codex here, create a Codex account and connect to github. |
|
@codex review |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 0fe4e59141
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| const path = stringField(output, "path") ?? stringField(input, "path"); | ||
| const url = normalizePageUrl( | ||
| stringField(output, "url") ?? |
There was a problem hiding this comment.
The continuity extractor currently builds a page artifact from input.path even when the tool call did not succeed, because artifactFromTool never checks tool.status and falls back to the input path. In runs where phantom_create_page or phantom_preview_page fails, this can still add a "user-visible page artifact" for a page that was never created, which then misleads later answers about page links. Please gate artifact creation on successful tool results (or otherwise require success-validated output fields) before including it in continuity context.
Useful? React with 👍 / 👎.
| if (typeof content === "string") return content.includes(PHANTOM_CONTEXT_OPEN_TAG); | ||
| if (!Array.isArray(content)) return false; |
There was a problem hiding this comment.
The Murph context cleaner removes any user message whose text merely contains <phantom_chat_context>. If a real user message includes that literal token (for example while debugging or quoting prior output), it will be silently stripped from conversation history on subsequent turns whenever this transform is enabled, causing context loss and altered behavior. Consider matching only the exact synthetic wrapper format (or a dedicated metadata marker) instead of a substring match.
Useful? React with 👍 / 👎.
2 participants
Summary
Validation