Cold Cache Detection

Cold-cache detection

TL;DR: A "cold restart" is a request that rewrote almost the whole context to the cache when it should have read it from the cache. The widget counts these and estimates the money they wasted. (H)

The rule

A main-thread request is cold when all 6 checks pass. Each check compares it with the previous main-thread request in the same session.

#	Check	Why
1	Not a subagent	Subagents have their own cache
2	Same session	Needs a previous request to compare with
3	ctx > 20,000	Small contexts aren't worth flagging
4	cr < 10% of ctx	Almost nothing came from the cache
5	Same model	A model switch rewrites legitimately
6	ctx >= 80% of previous ctx	About the same size, so not a /clear or /compact

Checks 5 and 6 separate "the cache expired" from "you started over on purpose".

Overspend

over = write cost − cost of reading the same tokens from the cache
     = (c5 × 1.25 + c1 × 2.0) × price_in − (c5 + c1) × price_read

Worked example

Opus 5.5 (in $4, read $0.20). The previous request had 150k context. This request:

Value
ctx	160,000 ✅ > 20k, ✅ ≥ 120k (80% of 150k)
cr	5,000 ✅ < 16,000 (10% of 160k)
c1	153,000

write = 153,000 × 2.0 × 4 / 1M   = $1.224
read  = 153,000 × 0.20 / 1M      = $0.031
over  = $1.224 − $0.031          = $1.19

Replying within the TTL would have saved $1.19 on that single request.

When it's flagged

Where	Trigger
Menu "Today" warning	≥ 3 cold restarts today (warn_cold_restarts_at_least)
Session report	Any cold restart. Severity is "med" if the total overspend is above $5

What to do: reply before the TTL runs out, or run /compact before you step away.

---

Next: Menu advice →