Repository navigation
feat: Add decisions prompt support - #2769
Luca Forstner (lforst) wants to merge 2 commits into
Conversation
Luca Forstner (lforst)
commented
Nov 10, 2026
Luca Forstner (lforst)
commented
- Build decisions prompts with text or message inputs and predicate, choice, or score questions. Template variables are rendered in both inputs and questions.
- Support saved and code-defined decisions prompts in scorers, eval parameters, and CLI exports.
- Define the decisions schema in the SDK so support does not depend on generated platform type updates.
Codex Review Summary: ✅ CompletedThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![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: e2e23057c5
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 (@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 (@codex) address that feedback".
| content: z.union([ | ||
| z.string(), | ||
| z.array(chatCompletionContentPartTextSchema), | ||
| ]), |
There was a problem hiding this comment.
For decisions prompts with array-valued message content, this schema accepts only Chat Completions { type: "text", text } parts, whereas the Decisions request contract uses input_text and input_image (see js/src/vendor-sdk-types/openai-decisions.ts and the existing image requests in e2e/scenarios/openai-instrumentation/scenario.impl.mjs). Consequently, valid saved Decisions inputs fail parsing in loadPrompt() or build(), while accepted text parts are returned unchanged by the renderer and produce invalid requests when passed to decisions.create(). Use the Decisions content-part schema and render text parts while preserving image parts.
Useful? React with 👍 / 👎.
