Platform
Which tokens did not need to be spent.
The Token Ledger says what the tokens cost. The usage scan says which of them were avoidable. Both read the same local transcripts, and neither makes a network request.
Rebuilt from disk
Model use and API-equivalent spend from ~/.claude/projects and ~/.codex/sessions. Reading them needs no proxy: the cost figures come from local files, and nothing is proxied to read them.
Nine detectors
Cache hit rate, context carried, repeated reads, oversized results, thinking share, idle resumes, unused connectors, subagent fan-out, instruction-file length.
Priced on the right meter
Each finding says which meter its tokens sat on — fresh input, output or cache read — so the money line uses the rate that traffic was actually billed at.
Measuring cost does not belong in the prompt traffic path
That is the whole reason this reads transcripts instead of proxying anything. Nothing sits between an agent and its model provider, so there is nothing new to trust, nothing to go down, and no second copy of your prompts.
Agent Access is the one part that holds a connection, and it is a different thing: a local endpoint for an agent’s tool calls to your linked accounts, off until you turn it on. Your prompts and the model traffic never pass through it. What Agent Access is, in full.
Content is never stored. A repeated read is spotted by hashing the result in memory and keeping the hash. The per-file cache is keyed by size and modification time, so a finished session is parsed once, ever.
What the detectors look at
| Detector | Fires when | Estimate |
|---|---|---|
| Cache hit rate | median below 70% in a session | modeled from what a warm cache would have served |
| Idle resumes and mid-session model switches | rebuilds over 50K tokens | measured — the fresh input and cache writes on the turns that followed |
| Context carried | any session above 150K | measured, then halved, because a split session still carries a handoff |
| Oversized single results | above the read threshold | measured size, two thirds assumed recoverable by a scoped read |
| Byte-identical results | the second copy in one session | measured — the extra copies |
| Instruction file length | CLAUDE.md or AGENTS.md over 200 lines | modeled at about 11 tokens a line, per session |
| Declared connectors never called | any declared connector with no calls in the scanned window | no figure claimed |
| Subagent fan-out | three or more in a session | no figure claimed |
| Thinking as a share of output | above 25% | measured, priced as output |
Where a figure would be a guess, none is given. Carried context priced as fresh input would overstate it tenfold, which is the mistake the Estimate column exists to prevent.
Two scores, kept apart
The usage score — Spending habits in the app — sits beside the doctrine score rather than inside it. One measures safety and the other measures waste; blending them dilutes both. Subagent transcripts are counted but not judged — a subagent has no conversation of its own to keep warm, so scoring its cache hit rate would be meaningless.
Right-sizing is arithmetic, not advice
The comparison prices the traffic each model actually ran against the next model down the ladder — Opus against Sonnet, Sonnet against Haiku. It is arithmetic on tokens already spent. The note beside each swap says what the transcripts suggest about whether the work would survive it — a high thinking share is the worst candidate, heavy tool use the most plausible one — and the view says plainly that a smaller model changes the answers as well as the bill.