Tool waterfall

Which step moved after the tool or model swap, and by how much.

The tool waterfall is the per-step latency attribution in the evidence report. Added and removed steps are labelled. A step with no usable duration stays unknown. It is never drawn as a 0 ms bar.

node src/cli.mjs fixtures/positive-outcome-cost-changed.json --text

Demo from the positive fixture

This block is the engine output for fixtures/positive-outcome-cost-changed.json. Bars share one scale. A missing duration would be labelled, not drawn as zero.

conediff report (contract v1)
baseline  gpt-4o (run run-gpt4o)
candidate claude-haiku (run run-haiku)

Outcome: changed "pass" -> "fail"
Cost: decreased by 0.009 (-0.009, -75%)
Latency: increased by 800 (+800, 44.4444%)

Latency attribution (per step, ms):
  generate     base 1400 ███████████████·········  cand 2180 ████████████████████████  +780 slower
  retrieve     base  300 ███·····················  cand  320 ████····················  +20 slower
  tool.search  base  100 █·······················  cand  100 █·······················  unchanged

Unknown fields: 0

toolwaterfall.dev is a name for this feature. It is not a second company and it does not have its own repository. DNS for that alias is not cut over; the comparison still runs locally through conediff.

Run the CLI Checkout