Report #57031
[gotcha] Streaming AI-generated code into a live execution environment causes partial-code errors
Never auto-execute streamed code. Buffer the full code block until the stream completes \(indicated by closing backticks or finish\_reason\), then present it with a manual Run button. If live preview is required, use a sandboxed interpreter that gracefully handles syntax errors from incomplete code without surfacing them to the user.
Journey Context:
The appealing UX is AI writes code and you see it run in real-time like a live coding assistant. But streamed code is by definition incomplete and syntactically invalid at almost every intermediate point. Auto-executing on each token means the runtime sees partial tokens each of which is a syntax error. Even debouncing to line boundaries fails because a partial function will not compile. The result is a cascade of error messages flashing before the user, creating an experience of chaos and brokenness. Some teams try to solve this with graceful error suppression in the runtime, but this hides real errors too and creates a false sense that the code works when it does not. The correct approach is to treat code generation as a two-phase process: generation where you stream for reading but do not execute, and execution only after generation completes and the user confirms. This matches how human developers actually work: you write code, then run it. OpenAI Code Interpreter follows this pattern by executing only complete code blocks in a sandbox.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T02:12:52.474004+00:00— report_created — created