turns

Per-turn timing analysis: average, median (p50), 95th percentile, and max duration per project.

Usage

claudex turns [-p/--project <substr>] [-l/--limit <n>] [--json]

Flags

Flag	Default	Description
-p, --project <substr>	—	Filter by substring match on the project path.
-l, --limit <n>	20	Maximum projects to show.
--json	off	Emit JSON.

Example

# Global p95 leaderboard
claudex turns --limit 10

# One project's timing profile
claudex turns --project claudex

# p95 as a single number
claudex turns --project claudex --json | jq '.[0].p95_duration_ms'

Columns

Column	Source
Project	Decoded project name.
Turns	Number of turns measured.
Avg	Arithmetic mean duration (ms).
p50	Median duration (ms).
p95	95th-percentile duration (ms).
Max	Longest single turn (ms).

JSON shape

[
  {
    "project": "/Users/you/projects/claudex",
    "turn_count": 1420,
    "avg_duration_ms": 4132.0,
    "p50_duration_ms": 2811.0,
    "p95_duration_ms": 17420.0,
    "max_duration_ms": 98233
  }
]

What counts as a "turn"?

A turn is the wall-clock interval between one user message and the next assistant message in the same session. Long tool-calling round-trips within one turn add to that turn's duration.

Notes

• No --no-index. Turn durations are derived during ingest; the file-scan fallback would re-parse every session just to compute them. Run claudex index if you want a fresh measurement.
• Outliers. A single 10-minute thinking turn will spike the max but barely move p50. Report both.