fix(runner): assistant follow-ups — event rate limit, Markdown rendering, Ask AI tooltip (DEV-2047)

[qunabu]

Three follow-ups to #119.

1. /api/chat/event gets its own rate-limit bucket

A regression I introduced in #119 while fixing the unbounded-dimensions finding, and merged before the review comment landed. The event route reused checkChatRateLimit, so Apply/Undo incremented the same per-IP counters as paid questions:

• applying an edit silently spent the user's question quota;
• users near the 8/min cap lost chat_edit_applied counts — skewing the acceptance metric the panel exists to show;
• anyone could exhaust an IP's chat budget by spamming the free route.

Route	Per minute	Per day
/api/chat	8	120
/api/chat/event	60	600

Verified: 12 consecutive events all return 204 (the shared bucket refused after 8), and a question afterwards still succeeds.

2. Answers render as Markdown

The panel shipped with a placeholder renderer that understood fenced code and paragraphs and nothing else, so every **bold**, inline code and bullet list appeared as literal asterisks and backticks. The content was right; the presentation made it look broken.

There is now a real renderer for the subset answers actually use — headings, ordered/unordered lists, bold, italic, inline code, links, fenced code. It builds React elements and never touches innerHTML, so model output cannot inject markup, and links are restricted to http(s). A Markdown library would have meant a new dependency in the authoring bundle and, in most cases, an HTML sink to sanitise. Unsupported syntax falls through as text.

3. The Ask AI button explains itself

"Ask AI" is a label, not an invitation — nobody clicks a chat button to find out it can edit their code. A hover/focus tooltip now names the three things worth knowing: scoped to the open example, changes arrive as an edit you can apply, answers are grounded in the docs with links. It also states up front that nothing is applied without you, which is the question people ask before letting an assistant near their code, and it hides once the panel is open.

---

Note

Medium Risk
Separate rate-limit buckets fix a real quota/abuse regression on a public API; the Markdown renderer touches untrusted model output but only allows http(s) links and avoids innerHTML.

Overview
Rate limiting: POST /api/chat/event no longer shares per-IP counters with paid /api/chat. checkChatRateLimit now takes a bucket (chat vs event) with separate limits (8/120 vs 60/600 per min/day), so Apply/Undo analytics do not consume question quota and the free route cannot exhaust an IP’s LLM budget.

Authoring UX: The toolbar Ask AI control is extracted into AskAiButton with a hover/focus tooltip that explains example-scoped Q&A, doc grounding, and apply/undo. Chat replies swap the inline Markdownish helper for a shared Markdown component that renders headings, lists, bold/italic, fenced code, and http(s) links via React (no innerHTML).

Docs in example-chat.md document the second bucket.

^{Reviewed by Cursor Bugbot for commit 362d98b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 362d98b. Configure here.}

[cursor]

Underscore italics break snake_case

Medium Severity

The _italic_ alternative in INLINE matches any underscore pair, including spaces and mid-identifier characters, so phrases like first_name and last_name render as italicized spans instead of literal text. Chat answers and user turns both go through Markdown, so ordinary snake_case column names get mangled unless wrapped in backticks.

Additional Locations (1)

• runner/apps/authoring/src/markdown.tsx#L57-L59

 

^{Reviewed by Cursor Bugbot for commit 362d98b. Configure here.}