🤖 fix: bound Shiki worker highlighting#3352
Conversation
|
@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: 2299b3276a
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".
Keep syntax highlighting optimistic by rendering plain text first while bounding worker highlight calls with a 5s budget. Timeout recovery now terminates and recycles the highlight worker, remembers pathological inputs to avoid repeated worker kills, and avoids main-thread Shiki fallback for worker runtime failures. Remove the 32KB diff highlight cap so human-scale files can attempt highlighting under the same worker budget. --- _Generated with `mux` • Model: `openai:gpt-5.5` • Thinking: `xhigh` • Cost: `$5.98`_ <!-- mux-attribution: model=openai:gpt-5.5 thinking=xhigh costs=5.98 -->
ammar-agent
force-pushed
the
fix/highlight-worker-budget
branch
from
2299b32 to
4d2d33d
Compare
|
@codex review Please take another look. I addressed the queued-request timeout poisoning by serializing worker highlight calls on the client side so the timeout clock starts only when a payload reaches the front of the queue. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 4d2d33d3a0
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".
|
@codex review Please take another look. I now resolve/revalidate the worker proxy when each queued highlight job starts, so jobs queued behind a recycled worker talk to the fresh worker instead of a terminated Comlink proxy. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 75cfcff509
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".
Resolve the worker proxy when each queued highlight job starts instead of when highlightCode is first called. This keeps later queued chunks from calling into a terminated Comlink proxy after an earlier chunk times out and recycles the worker. --- _Generated with `mux` • Model: `openai:gpt-5.5` • Thinking: `xhigh` • Cost: `$5.98`_ <!-- mux-attribution: model=openai:gpt-5.5 thinking=xhigh costs=5.98 -->
ammar-agent
force-pushed
the
fix/highlight-worker-budget
branch
from
75cfcff to
b225f7c
Compare
|
@codex review Please take another look on the updated branch. The latest head preserves the queued-worker revalidation fix and keeps the local/remote branch in sync for readiness checks. |
Guard worker.onerror recycling so asynchronous error events from an already-replaced highlighter worker cannot terminate the healthy current worker. --- _Generated with `mux` • Model: `openai:gpt-5.5` • Thinking: `xhigh` • Cost: `$5.98`_ <!-- mux-attribution: model=openai:gpt-5.5 thinking=xhigh costs=5.98 -->
|
@codex review Please take another look. I added an active-worker guard so stale error events from replaced highlighter workers cannot recycle the current healthy worker. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 43084163b7
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".
Guard worker.onerror recycling so asynchronous error events from an already-replaced highlighter worker cannot terminate the healthy current worker. --- _Generated with `mux` • Model: `openai:gpt-5.5` • Thinking: `xhigh` • Cost: `$5.98`_ <!-- mux-attribution: model=openai:gpt-5.5 thinking=xhigh costs=5.98 -->
|
@codex review Please take another look. The branch now includes the stale-worker guard: worker creation is captured in |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: cb11a6eede
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".
Add a pre-worker synchronous budget that still allows human-scale 10k-line chunks, while skipping generated/minified hunks before renderer-thread join and structured-clone work can stall the UI. --- _Generated with `mux` • Model: `openai:gpt-5.5` • Thinking: `xhigh` • Cost: `$5.98`_ <!-- mux-attribution: model=openai:gpt-5.5 thinking=xhigh costs=5.98 -->
|
@codex review Please take another look. I added a pre-worker sync budget for oversized/generated diff payloads while explicitly allowing 10k-line human-scale chunks through that budget. |
|
Codex Review: Didn't find any major issues. More of your lovely PRs please. ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
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". |
Release the Comlink proxy before terminating a recycled highlighter worker so Bun does not later surface an unhandled InvalidStateError when Comlink tries to release an already-terminated endpoint. --- _Generated with `mux` • Model: `openai:gpt-5.5` • Thinking: `xhigh` • Cost: `$5.98`_ <!-- mux-attribution: model=openai:gpt-5.5 thinking=xhigh costs=5.98 -->
|
@codex review Please take another look. I added a Comlink proxy release before worker termination to avoid Bun surfacing unhandled InvalidStateError after worker recycling. |
|
Codex Review: Didn't find any major issues. More of your lovely PRs please. ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
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". |
1 participant
Summary
Keep syntax highlighting optimistic while preventing pathological Shiki/TextMate inputs from wedging the renderer. Highlighting still attempts human-scale diffs/code blocks, but worker calls now have bounded runtime, oversized generated/minified hunks fall back before expensive renderer-thread preprocessing, and recovery no longer leaves stale or dead workers in the queue.
Background
A live Mux instance had a
shiki-highlighterworker pegging one renderer thread after it became unresponsive inside native Oniguruma matching. BecausehighlightCode()awaited the Comlink call without a timeout, the worker singleton stayed wedged forever and every later highlight request queued behind it.Implementation
highlight()calls.worker.onerrorrecycling against stale workers and release the Comlink proxy before termination to avoid unhandledInvalidStateErrorin Bun.join/structured-clone work can stall the renderer.Validation
make static-checkbun test src/browser/utils/highlighting/make typecheckRisks
Low-to-medium, limited to syntax highlighting. Pathological or generated/minified inputs now render as plain text after the sync/runtime budget is exceeded, and the fingerprint cache can conservatively skip highlighting for a rare collision, but the underlying code/diff text remains visible and copyable.
Pains
git push/git fetchto GitHub hung repeatedly in this workspace, so branch updates after the initial push were applied through the GitHub Git Database API while keeping the local branch ref synchronized to the API-created commits.Generated with
mux• Model:openai:gpt-5.5• Thinking:xhigh• Cost:$5.98