| name | introspect |
| description | Confidence/assumption hard-gate before /goal handoff. Outputs confidence 0-10, assumptions, unknowns. **Blocks /goal if confidence < 6.** Cheap (Opus, ~5 turns); catches false-confidence wrong turns. |
/introspect — confidence and assumption check
A hard-gate before autonomous /goal handoff. The fastest way to catch a wrong turn is to ask the model to introspect on its own plan before executing.
Output shape
confidence: 0-10
assumptions:
- <thing treated as given>
- ...
unknowns:
- <thing the model doesn't know>
- ...
recommendation: proceed | clarify | research-more
Gate behavior
- confidence < 6 → blocks /goal handoff. Engine surfaces unknowns and asks for clarification OR runs /deepresearch to resolve.
- 6 ≤ confidence < 8 → proceeds with assumptions logged into memory.decisions.
- confidence ≥ 8 → proceeds silently.
Why it works
Opus is generally calibrated on confidence. Forcing it to enumerate assumptions and unknowns surfaces the things that would make it fail mid-/goal. Cheaper to catch them here (5 turns Opus) than mid-/goal (50+ turns).
When invoked
Auto-invoked by every umbrella between /spec(qa) and /goal. Manually invoke when you want a confidence check before a big autonomous run.
