Latency Breakdown

Preview in your theme

status='ready' — 8 runs on one shared scale, a 3.5 s budget (dashed line, flagged rows, footer count) and server-computed p50/p95 over 12,480 runs, which the footer names instead of pretending they came from these 8. Hover or focus a segment for its detail; click to pin it (Esc releases); pick a phase in the legend to rank the runs by it. Tab reaches the bars once, then the arrow keys walk them.

8 runs · slowest 7.37 s

median per phase — pick one to rank the runs by it

chat.completions · 50772egpt-4o-mini
1.67 s
chat.completions · 1ad37agpt-4o-mini
1.37 s
agent.step · e8ce0aclaude-sonnet-4
7.37 sover budget
chat.completions · bfd33bclaude-sonnet-4
2.98 s
summarize.batch · f8bddellama-3.1-70b
3.45 s
agent.step · 8ca1cegpt-4o
3.13 s
chat.completions · b9bcb8gpt-4o
3.8 sover budget
rerank.query · ccc6cellama-3.1-70b
251 ms

p501.84 s

p954.21 s

2 of 8 over the 3.5 s budget

Total latency · nearest rank over 12,480 runs

phase=null · onSelectRequest: not fired yet

scale='share' with defaultPhase='toolMs' — every bar stretched to full width, so magnitude is discarded and only composition is left. Same 8 runs: the tool-bound one is obvious here and invisible in a totals ranking. Note the footer switches to the isolated phase and drops the server stats, which only describe total latency.

8 runs · slowest 7.37 s

Ranked by Tools — press it again for the whole run.

median per phase — pick one to rank the runs by it

agent.step · e8ce0aclaude-sonnet-4
5.12 sof 7.37 s
agent.step · 8ca1cegpt-4o
840 msof 3.13 s
chat.completions · 50772egpt-4o-mini
0 msof 1.67 s
chat.completions · 1ad37agpt-4o-mini
0 msof 1.37 s
chat.completions · bfd33bclaude-sonnet-4
0 msof 2.98 s
summarize.batch · f8bddellama-3.1-70b
0 msof 3.45 s
chat.completions · b9bcb8gpt-4o
0 msof 3.8 s
rerank.query · ccc6cellama-3.1-70b
0 msof 251 ms

p500 ms

p955.12 s

Tools · nearest rank over these 8 runs

A 41 s outlier next to runs of 60–200 ms. Every fast row is under half a percent of the scale and its queue is a fraction of a pixel — the floor gives each non-zero phase ~3px, borrowed from its own row so the bar's total width still tells the truth about the scale. The percentages in the detail panel are computed from the raw milliseconds, never from the drawn width.

5 runs · slowest 41.14 s

median per phase — pick one to rank the runs by it

research.agent · 1061d3claude-sonnet-4
41.14 sover budget
chat.completions · 25b1b2gpt-4o-mini
162 ms
chat.completions · 0eb00egpt-4o-mini
136 ms
classify.intent · dc5b54llama-3.1-70b
63 ms
embed.query · df48ddllama-3.1-70b
21 ms

p50136 ms

p9541.14 s

1 of 5 over the 12 s budget

Total latency · nearest rank over these 5 runs

A single run: 'p95' is that one run, so the footer says 'over this one run' rather than quoting a percentile it cannot have. The median comparison in the detail panel stays hidden below 3 runs for the same reason.

1 run · slowest 2.35 s

median per phase — pick one to rank the runs by it

chat.completions · 1f7aeagpt-4o
2.35 s

p502.35 s

p952.35 s

Total latency · nearest rank over this one run

14 runs with maxRows={6} and order='slowest'. The disclosure states the exact count it is holding back and expanding it never rescales the bars — the shared scale is measured over every run, not over the visible slice.

14 runs · slowest 7.37 s

median per phase — pick one to rank the runs by it

agent.step · e8ce0aclaude-sonnet-4
7.37 s
chat.completions · b9bcb8gpt-4o
3.8 s
summarize.batch · f8bddellama-3.1-70b
3.45 s
agent.step · 8ca1cegpt-4o
3.13 s
chat.completions · bfd33bclaude-sonnet-4
2.98 s
chat.completions · 6bf7adgpt-4o-mini
2.61 s

p502.08 s

p957.37 s

Total latency · nearest rank over these 14 runs

status='loading' — skeleton rows on the same grid template as the real ones, so nothing jumps when the data lands.

status='empty' — the window held no runs. A different sentence from the one a 'ready' envelope with an empty list gets.

No runs in this window

Widen the time range, or send a request to start collecting timings.

status='error' — the timings failed to load, with a retry the host owns.

Latency data unavailable

The trace store returned 503 for this window. request-id: 7f2c-9ad1

status='ready' with an empty list — a contract slip rather than a state, so it says so in its own words instead of silently rendering a chart with no rows. Every state on this page comes straight from the contract; not one of them waits on a timer.

No runs in this window

The panel was handed an empty list.

Loading preview…

This component is part of Pro

The install command, prompt and concept map are included with a Pro license — one purchase unlocks every Pro component, forever.

Unlock with Pro