Prompt optimizer

The prompt optimizer rewrites a vague or hedged prompt into a specific, scoped, actionable one — before you send it to the agent. Trigger it with Cmd+Shift+Enter (macOS) / Ctrl+Shift+Enter (Linux), or click the ✨ sparkle button in the input toolbar.

1. Type a prompt in the dashboard textarea
2. Press Cmd+Shift+Enter (or click ✨)
3. Dark overlay appears: "✨ Optimizing prompt…" — textarea becomes read-only and dimmed
4. The rewritten text replaces the textarea content — review before sending
5. Press Enter to send (no auto-send)

On error or unchanged text, the original is restored silently. Typical overhead: 2–5 seconds.

Optimization helps most when the original is:

• Vague — "fix that thing" → gets scope, files, and constraints added
• Under-specified — "add auth" → gets required components and edge cases
• Missing structure — a wall of text → gets numbered steps
• Uncertain — hedged with "maybe" or "if you could" → replaced with direct verbs

Optimization is unnecessary when the prompt is already:

• Specific and scoped
• Short and unambiguous ("What's 2+2?")
• Referencing a specific file the agent should already understand

The optimizer knows this — it returns the original unchanged when there's nothing worth adding.

The optimizer's system prompt encodes six rules (earlier wins on conflict):

1. Never change your intent — no added requirements, no invented values
2. Return unchanged if the prompt is already specific, scoped, and actionable
3. When rewriting, add what you skipped: scope, constraints, structure, uncertainty handling
4. Replace hedging ("maybe", "if you could") with direct verbs
5. Add "preserve existing behavior" when the change is to existing work
6. Never exceed min(3 × original length, 250 words)

Before:

fix the zoom bug in the dashboard

After:

Fix the zoom bug in the dashboard chat input.

Scope:
- `website/src/hooks/useZoom.ts` — cleanup CSS custom properties on unmount
- `website/src/components/ChatInput.tsx` — cache the `effectiveVh()` result per render

Constraints:
- Preserve existing zoom-in and zoom-out gestures
- Do not change the public API of `useZoom`
- Run `npm run build` and `npm run test` before completing

If you find related issues while fixing this, list them but do not fix them in this change.

Same intent, but the agent now knows the files, the constraints, and where to stop.

The optimizer receives:

• Your prompt in <original_prompt> tags
• The last ~10 conversation messages (truncated to 2,000 chars)

Context makes the optimizer smarter when you're mid-conversation: it picks up file names, error messages, and prior scope from the recent history. In a fresh session, it works from the prompt alone.

The optimizer runs on a dedicated _optimizer session with the kirocrew-lite agent — a small, tools-disabled agent tuned for single-shot rewrites.

• Independent from the shared _bg session (which handles title generation and folder categorization)
• No tool calls — any tool permission request during optimization is rejected outright
• 30-second asyncio.wait_for wraps the entire operation (session claim + stream + release)
• LLM output has "UNCHANGED" detection built in — if the LLM decides the prompt is fine, the original is returned

On timeout or error, the original text is silently restored. The optimizer never fails loudly.

Output is passed through redact_credentials() and redact_exfiltration_urls() before it reaches your textarea. If the optimizer ever surfaces a credential pattern from context, it's redacted.