Process metrics for coding-agent sessions, derived from the JSONL transcripts Claude Code already writes.
Measuring whether an agent did a good job needs a grader, and graders are the component that keeps failing — six documented incidents across three projects in this lab, every one an interpretation step going wrong.
Process metrics need no interpretation. They are counts:
| metric | question it answers |
|---|---|
tool_calls_per_turn |
how much work happens between two human messages |
reads_before_first_edit |
did it explore before implementing |
repeated_commands |
is it thrashing on the same command |
error_rate |
how often tool calls come back failed |
seconds_to_first_write |
did it plan, or dive straight in |
edit_revisits |
how many times the same file was rewritten |
None require ground truth. None require judging output quality. Every one is derived from the transcript at generation time rather than reconstructed after the fact — which is the rule the graders in this lab broke.
Rules written in prose (CLAUDE.md, output styles, memory) are hypotheses about
agent behaviour. Each was added on a date. With ~/.claude under version
control, rule-introduction dates are known, and these metrics can be compared
before and after. If a rule does not move its metric, the rule does not bind.
That is the measurement; this package produces the numbers for it.
uv run python -m agent_telemetry report # summary across all sessions
uv run python -m agent_telemetry export out.parquet
Reads from $CLAUDE_TELEMETRY_ARCHIVE/raw (default ~/claude-telemetry/raw),
which is populated by ~/.claude/bin/archive-transcripts.sh.
Transcripts were verified on 2026-07-28 to contain credential-shaped strings. This package reads them locally and its outputs carry only metadata — tool names, counts, timestamps, paths — never tool output bodies. Do not commit raw transcripts or exported frames containing them.