Context Reset Rate (CRR)

Weight in Skill: 10%. CRR is a behavioral proxy — it captures whether you manage context intentionally, rather than letting long sessions drift.

What it measures

CRR is the ratio of context-reset actions (/compact, /clear) to turns:

CRR_raw = reset_actions / turns

Unlike most metrics, CRR does not reward higher-is-better. The scoring function is a triangle centered on a healthy middle: the best score lands at the team’s typical reset cadence. Reset too rarely and the session bloats; reset too often and you destroy useful context every few turns.

Scoring (pilot defaults in apps/prism-engine/src/scoring/layer3.rs):

CRR_PEAK       = 0.15       (roughly one reset per 6–7 turns)
CRR_TOLERANCE  = 0.10       (full zero outside ±0.10 of the peak)
CRR_score      = 10 · (1 − |raw − peak| / tolerance)

These are pilot-tunable constants. Once a week of production data is in, the peak and tolerance will be re-fit to the actual healthy-team distribution.

What pushes you off-peak

Too low (< 0.05):

• 80-turn sessions that started with “explore the auth module” and have since touched billing, email, and migrations — with no /compact along the way
• Long-running “ambient” sessions left open across unrelated tasks

Too high (> 0.25):

• /compact after every turn “just in case” — throws away context the next turn needs
• /clear used as a recovery tool when the real issue is a wrong approach, not context bloat

Improving CRR

1. Reset at natural breakpoints — when a task finishes and the next is unrelated
2. /compact in long sessions around turn 30–40 if the earlier context isn’t needed
3. /clear when switching domains — don’t carry auth context into a migration task
4. Don’t reset to recover from bad prompts — add constraints or change approach instead

The /prism:cost command nudges you with a /compact reminder when it notices context growing fast without a reset.