adk evals runs: --format json is ignored, prints ANSI-coloured human output instead

[johnf]

Summary

adk evals runs --format json ignores --format and prints the human-readable, ANSI-coloured listing instead of JSON. The flag is documented in the subcommand's own help output.

Versions

adk CLI 2.0.3, @botpress/evals 2.0.3, Linux, node 26.

Reproduction

$ adk evals runs --help
...
Options
  --latest  Show the latest run
  --limit <n>  Max runs to list
  -v, --verbose  Show full details for all evals in a run
  --format <format>  output format (json)

$ adk evals runs --latest --format json | head -c 200 | cat -v
  ^[[1mRun^[[22m 3141202d4a4a41bfa470ea84fe
  ^[[90m7/25/2026, 4:34:07 PM^[[0m

  ^[[31mM-bM-^\M-^W^[[0m ^[[1mstate-probe-multiturn^[[0m^[[22m  ^[[90m10.1s^[[0m  ^[[90m3/6 assertions^[[0m ...

Same with -v, and the same when redirected to a file rather than a pipe (so it is not the stdout-flush issue in #38):

$ adk evals runs --latest --format json -v > out.json; node -e "JSON.parse(require('fs').readFileSync('out.json','utf8'))"
SyntaxError: Unexpected token '', "
  [1mRun[2"... is not valid JSON

ANSI escapes are emitted even when stdout is not a TTY.

Expected

Either the documented JSON document on stdout, or — if JSON output was never implemented for this subcommand — remove --format from its help so it does not advertise a mode it lacks.

Why it matters

adk evals --format json works, so run history is the only way to recover a report you did not redirect at the time, and it cannot be consumed programmatically. Minor on its own; noting it because it also blocked a re-test of #38 — the 104-eval report here is ~970 KB, and runs was the only way to re-emit a payload that large without paying for another suite run.

Found while investigating #40.