autonomy-stat

Analyze how long a Claude Code agent ran on its own per turn, from a session JSONL log, and write an interactive HTML report (two time-series charts + a per-turn table).

When to invoke

• User wants to evaluate agent autonomy / self-running time.
• User asks how long the agent worked per turn, or to chart a session over time.
• User points at a project (~/foo) or a session id and asks to "analyze" it in this autonomy sense.

Usage

skills/autonomy-stat/scripts/autonomy-stat.py stat <session-id|project-dir> <outfile.html> [--lang en|ja]

The first argument accepts any of:

• a session UUID — e.g. 66af0915-9b59-4b1e-88a3-ff312e0c2ed3 (searched across all ~/.claude/projects/*/)
• a full path to a *.jsonl
• a project-directory name in the -home-user-dir form — e.g. -home-user-myproject. The latest session in that project is used.

--lang sets the report's UI language: en (default) or ja. Only the UI labels are translated; the analyzed data (input prompts etc.) is never translated.

Example:

skills/autonomy-stat/scripts/autonomy-stat.py stat -home-user-myproject /tmp/myproject.html
skills/autonomy-stat/scripts/autonomy-stat.py stat -home-user-myproject /tmp/myproject.html --lang ja

Open the resulting HTML in a browser. The user runs their own HTTP server — do not launch one.

Output

A self-contained dark-mode HTML (no external dependencies) with:

• Top chart (blue) — self-running time excluding tool-wait: pure model thinking/generation time per turn.
• Bottom chart (orange) — self-running time including tool-wait (model work + tool execution + approval waits).
• Each chart auto-scales its own y-axis (nice steps: 1m/2m/5m/.../1h/2h/...); x-axis is per-day, local time.
• Hover any point for a per-turn breakdown; a full table follows below.

The gap between the two charts is the tool-wait time. Read the blue chart for "time the agent actually spent reasoning"; the orange/blue gap reveals long tool executions or unattended approval waits.

How it works (algorithm)

1. Drop replayed history (drop_replayed): /compact and resume re-inject old log lines with their original (stale) timestamps. Any entry whose timestamp jumps backward more than REPLAY_TOLERANCE_MS below a running monotonic clock is discarded as a replay.
2. Detect genuine human input (is_genuine_user): a user entry that is not a tool result (toolUseResult), not meta, not a sidechain, not a harness injection (origin.kind, <task-notification>, <command-*>, <system-reminder>, the compaction continuation summary). These mark turn boundaries.
3. Split into turns: consecutive human inputs (queued) collapse into one boundary (start = last of them). A turn's activity = assistant messages and tool results until the next genuine input. system/queue/command events never extend a turn's end.
4. Classify the gap before each activity event into three buckets:
   • next event is a tool result → toolwait (a tool was running / awaiting approval); counted in full.
   • next event is an assistant message and gap ≤ IDLE_THRESHOLD_MS → work (normal generation/thinking).
   • next event is an assistant message and gap > IDLE_THRESHOLD_MS → work capped at the threshold, the excess is sleep (e.g. /loop sleeps, unattended idle) and is excluded. Rationale: a long gap before an assistant message has no running tool to justify it, so it is idle, not work. A long gap before a tool result is a real (possibly slow) tool, so it counts.
5. Two series per turn:
   • active_excl = work (tool-wait excluded) → blue
   • active_incl = work + toolwait (tool-wait included) → orange
6. Between-turn idle (idle_before_sec) = next turn's start − this turn's end. This is the agent sitting stopped while the human is away; never counted as self-running (shown for reference only).

Tunable constants (top of the script)

Tool-wait gaps are intentionally not capped — instead both the with-tool-wait and without-tool-wait series are shown so the user can judge.

Notes / caveats

• Timestamps are UTC in the log; the HTML renders them in the browser's local time zone.
• Sessions heavily compacted/resumed can be majority replay; step 1 is what makes the per-turn durations trustworthy.