feat(pi-tui): add opt-in render profiler#1462
Conversation
Record per-frame render cost and stdout write timing behind the PI_TUI_RENDER_PROFILE env flag to help diagnose terminal render stutter, especially on Windows. Disabled by default.
🦋 Changeset detectedLatest commit: 26811f5 The changes in this PR will be included in the next version bump. This PR includes changesets to release 1 package
Not sure what this means? Click here to learn what changesets are. Click here if you're a maintainer who wants to add another changeset to this PR |
commit: |
![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: 26811f5068
ℹ️ About Codex in GitHub
Codex has been enabled to automatically 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 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| if (this.pending.length === 0) return; | ||
| const chunk = this.pending.join(""); | ||
| this.pending = []; | ||
| fs.promises.appendFile(this.filePath, chunk).catch(() => {}); |
There was a problem hiding this comment.
When a profiling run reaches 100 records, this drops the chunk from pending and starts a fire-and-forget appendFile; close() only sync-flushes the current pending array, while the normal Kimi shutdown path calls TUI.stop() and then process.exit() from apps/kimi-code/src/cli/run-shell.ts. In sessions long enough to cross the threshold, records already handed to this async write can therefore be lost or written out of order before the process exits, which makes the opt-in profile incomplete exactly for the long render-stutter captures it is meant to diagnose.
Useful? React with 👍 / 👎.
1 participant
Related Issue
No linked issue. This PR adds the diagnostic instrumentation described below; it is a precursor to a fix rather than the fix itself.
Problem
On Windows, the kimi-code TUI can make the desktop feel laggy while streaming output — for instance, opening File Explorer or switching windows stutters. Today we have no visibility into per-frame render cost or stdout write behavior, so the root cause is being guessed. Windows ConPTY / Windows Terminal stdout writes are substantially more expensive than Unix pseudoterminals, and the renderer does not yet react to that cost.
What changed
Adds an opt-in render profiler to pi-tui, gated behind the
PI_TUI_RENDER_PROFILEenvironment variable. It is disabled by default and near-zero overhead when off. When enabled it writes newline-delimited JSON to~/.pi/agent/pi-render-profile.jsonl(or a custom path via the env value), recording for every frame:writereturningfalse);The profiler buffers records and flushes asynchronously during the render loop, with a synchronous flush on shutdown, so the instrumentation does not add blocking I/O to the hot path. Unit tests and a changeset are included.
This is instrumentation only and changes no default behavior. The captured data is meant to drive the follow-up fix (likely backpressure-based frame dropping and/or render throttling for unfocused windows) once measured on Windows.
Checklist
gen-changesetsskill, or this PR needs no changeset.gen-docsskill, or this PR needs no doc update.