Fixes · last 7 days
$33.02 a week is avoidable. 2 changes, ranked by what they save.
Each figure is Tracekit's own tips detector run on your agents' signed records, at Anthropic's list prices of 2026-11-05. Mark a fix done and the same detector measures the days after it, beside the days before.
Compact at about 300k tokens, not at 958k
Every step sends the whole context again. claude-code run 296a0914's median step carried 497k tokens of it: $0.11 before it did anything.
Since it was marked done · 2026-11-16 00:44 UTC by Dana ReyesNot a controlled test
Better so far. The after side is under a day: wait for more before you trust either way.
The one change
Run /compact with a one-line focus at each task boundary (a PR merged, a release tagged), and at about 300k tokens; start a new session for unrelated work.
/compact keep the open task, its files and the decisions madeBehind it
- 296a0914420 signed model turns · context 497k tokens at the median, 958k at most
- 078bc37b300 signed model turns · context 221k tokens at the median, 399k at most
- c718c64b260 signed model turns · context 371k tokens at the median, 759k at most
- 3ab9eb36180 signed model turns · context 183k tokens at the median, 398k at most
How it is worked out
- claude-code run 296a0914 (claude-opus-5-5): 420 steps; its context was 497k tokens at the median and 958k at most; it compacted 1 time, at a median 958k, down to 70k.
- Its input side as it ran (context read from cache, written to it, or sent uncached): $46.76, $0.22 per million tokens of context a step on average (the chart's right axis).
- Replayed on the same growth, compacting whenever it passed 300k: 3 compactions, each down to 70k, one summary call (20k output tokens at $20/M plus a re-read) and a cache write of what is left. Every other step keeps its real cache behaviour (a rebuild after idle time, or a call with no cache, re-pays its whole replayed context): $23.00.
- Would have saved: $46.76 - $23.00 = about $23.76; across the 4 main sessions of 50+ steps, about $33.02 of $97.83.
- 300k is the largest compaction point that keeps 80% of the best saving tried: compacting more often saves a little more and loses more detail.
- Token counts of the 1,160 steps these figures rest on: 1,160 from signed records.
- An estimate: not counted is work redone when a summary drops a detail. Time isn't counted either.
From signed records · detector: tips context · Anthropic list prices of 2026-11-05
Done 2026-11-16 · measuring1 held call nobody answered
Tracekit held 1 call for a person; 1 never got an answer, and the agent went on without them.
The one change
TK-A056 held 1 of them (its rule). See each one waiting, with the commands that answer it:
tracekit approvals listBehind it
- 296a09141 held, never answered · TK-A056 1, TK-F004 1
How it is worked out
- Held calls: 1 signed decisions `ask` in the window, the agent's retry of a held call counted with it, as the run review counts them; 0 answered (a signed approval or rejection), 1 never.
- Time blocked: no transcript holds these calls, so how long each one waited isn't known.
- Where each ran: the transcript that holds its call (a subagent's file), else the record's agent; a Claude Code run's records decide a subagent's calls under the session's run.
From signed records · detector: tips holds · Anthropic list prices of 2026-11-05
Cache rebuilds after long waits
It needs how long each cache write lasted (5 minutes or an hour) and when each blocking call started and ended. Only Claude Code's transcripts record those; signed records don't, so Tracekit Cloud doesn't estimate it.
Detector: tips cache · needs what only a transcript records
Can't be priced from signed records