The session starts. Press Play or Next to follow it step by step.
claude-sonnet-5.5Context not measured yet
0:00.00 / 7:08.20Paused
The clock shows recorded time. Long idle gaps are compressed in the chart and playhead.
Click or drag the chart ruler to seek; Space on the chart plays or pauses.
Whole session
Chart controls
Scroll or pinch to zoom at the pointer; drag, Shift-scroll or scroll sideways to pan;
click or drag the ruler to seek. The lanes under the ruler sort every recorded event by what the agent was doing;
click a tick to jump to it. While replaying, work running at the playhead is outlined
and work yet to start fades. Click a bar to expand it to the full width: only its
own work stays below, and the work that caused it stacks above as a grey trail.
Click a trail bar, press Escape or zoom out past it to step back out; Fit shows the
whole session again.
With the chart focused, arrow keys move between bars, W and S zoom, A and D pan, 0
fits, Enter expands the selection, and Space plays or pauses. Shaded breaks fold idle
stretches longer than a minute.
Session events
Where the time went
Active for 2m 30s of 7m 08s. The largest share went to bash: 45% of active time.
Activity
Time
Of active
Spans
Failed
Each moment counts once, split evenly across the innermost work running then. Colors run
from red to violet in order of time taken. Select an activity to pick it out in the chart.
Compare recorded runs
Compare the same TodoMVC task across models or input sizes. Width is elapsed time,
not context capacity. The illustrative sample is not evidence.
How to make a fair comparison
Use the same starter commit, prompt, Copilot CLI version, tools, permissions and acceptance
tests. Compare one model with small and large input payloads, then repeat with another model.
Change one variable at a time; keep reasoning settings consistent and record cache conditions.
Repeat each condition. A larger context window alone does not mean more tokens were sent.
Tools, output length and idle time also affect the duration. Context snapshots are occasional
measurements, not a continuous peak; only per-call usage records establish input size.
Choose 2–6 Copilot CLI events.jsonl files (100 MiB combined). Claude and OpenAI models
are supported when run through Copilot CLI; native Claude Code and Codex logs use different
schemas. Files stay in this browser. Importing another set replaces the comparison.
No comparison logs loaded. Choose at least two runs; no model calls will be made.
Aligned at the first prompt; idle gaps included. Equal widths mean equal time. Colors match work kinds across runs.