Assistant debug/fix affordances for errors and failing checks

[josephschorr]

Summary

Adds one-click "Ask assistant to debug/fix" entry points that open the AI assistant and auto-submit a targeted debug prompt for a specific failure. The assistant already has the tools and full document state each turn, so this is purely the entry points plus one small new primitive.

Surfaces (all gated on AppConfig().aiEnabled — nothing renders/registers when AI is off)

• Schema syntax errors & failed assertions — an inline "Ask assistant to fix" CodeLens in the schema/assertions editors, mirroring the existing registerAssertionFixes pattern.
• Failing check watches — an "Ask assistant to debug" button in the Watches panel for every non-passing state (denied / errored / couldn't-run / missing-context).
• Problems panel — an "Ask assistant to fix" action on schema, assertion, and validation error rows, alongside the existing "Add check watch".

How it works

• A transient pendingPrompt on the assistant store (deliberately not persisted).
• requestAssistantDebug(prompt, source) opens the assistant dock and stashes the prompt; a usePendingPromptConsumer hook in AssistantPanel submits it once the panel isn't mid-turn. It defers only while a turn is actively streaming/executing tools — not on the terminal error state — so a request made after a failed turn still fires.
• A pure debugPrompts.ts builds the prompts (schema / assertion / validation / relationship errors, plus check watches), keeping wording consistent and testable.
• Analytics: a single playground_ai_debug_requested event tagged with { source: "editor" | "watches" | "problems" }.

Tests

• Unit: prompt builders + the isDebuggableWatchStatus predicate, the store's pending-prompt actions, and the requestAssistantDebug dispatcher.
• Browser: the affordance components (click → dock opens + prompt submitted) and the pending-prompt hook (idle, deferred-while-streaming, and fires-from-error).
• Full gate green: unit 228, browser 47, vite build clean, oxfmt --check clean, lint clean (no findings on changed files).

The inline editor CodeLens has no automated test by design, matching the existing untested registerAssertionFixes; its prompt output is covered by the builder unit tests.

Notes

• Design spec and implementation plan live under docs/superpowers/.
• The four triggers were also verified live in-app with AI enabled.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
playground	Ready	Preview, Comment	Jul 22, 2026 11:40pm