Prompt quality (Ask ACE / Agent)
ACE keeps Run/Debug Ask ACE replies on-topic and avoids ecommerce/auth digressions.
Ask ACE prompts
| Case | Marker | Behavior |
|---|---|---|
| Exited 0 | [ACE_ASK_RUN_SUCCESS] |
Ask mode, summarize output only |
| Failure | [ACE_ASK_RUN_FAILURE] |
Agent (if plan allows), fix this error only |
Both prompts forbid inventing auth/seller/ecommerce features.
Follow-up offers
- Success → arguments / logging / breakpoint
- Failure → apply a minimal patch + F5 again
- General Agent/Ask → tied to the open file / current topic (no hard-coded seller/auth tips)
Extra guards
- Ask ACE uses a fresh chat, empty history, and minimal context (target file only)
- Off-topic auth/seller paragraphs are stripped from Ask ACE replies when detected
- Chat shows a short Ask ACE label (not the raw instruction tags)
- Success Ask ACE does not show “Switch to Agent to apply changes”
Quick test
- Open
desktop-ide/samples/phase3-run-debug/hello.js→ F5 → Ask ACE
Pass: short success summary, no auth/sellers talk
Pass: user bubble likeAsk ACE to summarize this successful run · hello.js
Pass: no “Switch to Agent” CTA footer - Open
fail-on-purpose.js→ F5 → Ask ACE
Pass: quotes the error, proposes a fix for that file only - Agent rename in
greet.js
Pass: reply stays about the rename (no random auth checklist)
