Measured and estimated

Every token or cost value in AgentLens is either measured or estimated, and the output says which.

You see	It means
12,345	Measured: the client reported this number (usage of a request, a cost it computed).
~12,345	Estimated: AgentLens computed it, from a file size or from a price table.
n/a	The client does not record it. It is not zero and it is never guessed.
$1.50 + ~$2.00	A cost with a measured and an estimated part.

What is estimated, and how

• The size of text (a skill, an instruction file, a tool definition) is not reported by any client in isolation. AgentLens estimates it as characters divided by four. ADR 0001 explains the choice. With --exact the text of skills is counted by the Anthropic API instead.
• Cost the client did not report is computed from a bundled price table and marked ~. The table only contains prices checked against the vendor's list, with source and date (ADR 0002); a model without a price is listed as unpriced, never guessed.
• Overhead is a size times a number of requests, so it is an estimate whenever the size is.

What is measured

Token usage of a request (input, output, cache read and write), the context size at each request, cost that the client reported, tool outcomes and durations.

In files and in the MCP server

In the application and on the command line the tilde carries the marking. In --format json and in the results of the MCP server, every count that can be an estimate is an object with count and measured: the load, baseline and carry of a skill, the tokens and overhead_tokens of a context row, the idle cost of an unused item and the skill overhead of compare. Counts taken from the usage a client reports (requests, input and output tokens, cache tokens, the context size in flow and timeline) are measured by definition and stay plain numbers; so does the money split into measured_usd and estimated_usd.

A sum is measured only if every part of it is: a context row that holds one item whose size was estimated is an estimate.

In CSV the counts are plain integers, and the flags are columns at the end of the table (baseline_measured, carry_measured, overhead_measured, tokens_measured, overhead_measured, skill_overhead_measured), so a file that is read by position keeps its columns.