AI Code Completion Service · View 04 of 21 · People and journeys
The trough is the finding
- The sentiment line bottoms out at Wait, not at Read. A wrong suggestion costs a glance; a slow one costs the thought the developer was holding.
- A cold prefix cache adds ~180 ms at p95 and is perceived as a broken editor rather than a slower model. That perception is why session affinity is the critical design decision rather than an optimisation.
What answers each trough
- Wait → session affinity to a warm prefix, plus a deadline that returns empty rather than late (views 15, 13).
- Read → degeneracy detection and suffix deduplication before display (view 08).
- Live with it → accepted-character retention at 120 s as the canonical quality metric, not acceptance (views 16, 18).
Assumptions
- 60 ms client debounce; ≥ 55% of requests are cancelled before display.
- Displayed single-line acceptance ≥ 28%; accepted-character retention ≥ 70% at 120 s.
