When cost-anomaly flags a session as a >3.5σ outlier, the next question
is "which MESSAGES were expensive?". cost-session answers that.
| Question | Skill |
|---|
| "Which sessions cost the most?" | cost-conversation |
| "Which sessions are outliers?" | cost-anomaly |
| "Which messages in THIS session were expensive?" | cost-session ← this |
Algorithm
Implementation: scripts/session.mjs.
- Resolve session jsonl:
--session-id <id> (scans ~/.claude/projects/*/)
or --latest (default; picks most-recently-modified jsonl).
- Parse all assistant messages with
usage blocks.
- Cost each message via shared PRICING (
_prices.mjs).
- Sort descending by
cost_usd, surface top-N (default 20).
- Compute p50/p90/p99 of message costs for in-session percentile context.
- Flag the top message if it's >2× the p99 — that's an in-session outlier.
Why this matters: cache writes are the silent cost
Example real session, top message:
| # | Model | In | Out | Cache W | Cache R | Cost |
| 1 | opus-4-7 | 6 | 569 | 881898 | 0 | $16.58 |
Without the Cache W column it looks like "569 output tokens cost
$16" — that's wrong by 380×. The actual cost is ephemeral 1h cache
write at opus pricing: 881,898 tokens × $18.75/1M = $16.54.
Operators reading the table see immediately: "the model wrote 881K
tokens to ephemeral cache". From there the question becomes "why did
we cache 881K tokens of context for a 6-input request?" — that's a
real engineering signal.
Drill-down workflow
# Step 1: find outliers across all sessions
cost anomaly --alert-on-outliers 1 || cost anomaly # see which session-ids
# Step 2: drill into the flagged session
cost session --session-id <flagged-id> --top 10
# Step 3: open that jsonl at the timestamp the top message reports,
# inspect the prompt + tool calls
Percentile context
Top of output:
| p50 (median) message | $0.85 |
| p90 message | $1.45 |
| p99 message | $1.74 |
Lets operators ask "is this top message a 2× outlier or a 380× one?"
without having to compute it themselves. The "top is >2× p99" footer
fires when the answer is "yes, this is an in-session outlier worth
investigating".
--since filter
Useful for drilling into a specific time range within a long session:
cost session --since 2026-06-16T13:00:00Z --top 5
Only messages with timestamp >= --since are considered.
Edge cases
- No costed assistant messages → "No costed assistant messages" + exit 0.
--session-id not found in any project's jsonls → exit 2 with error.--top must be a positive integer → exit 2.