Skip to main content

​
Diff

cmd/crewship/cmd_diff.go takes two existing run IDs, fetches their state plus the last assistant message of each linked chat, and renders a terse before/after view. Distinct from eval compare, which re-runs an eval scenario on two tiers — diff works strictly on what already happened. Use cases:
  • “Did v2 of my routine actually fix the bug?” → diff before/after
  • “What changed between Friday’s prod run and Monday’s regression?”
  • “Did the agent give two different answers to the same prompt?”
Both runs are fetched in parallel; partial failures surface as run-a <id>: <err> or run-b <id>: <err>.

​
crewship diff <run-a> <run-b>

FlagTypeDefaultEffect
--fullboolfalseShow full assistant output. Without it, each side is truncated to 1 KiB with an ellipsis.

​
Examples

crewship diff r_abc r_def
crewship diff r_abc r_def --full
crewship diff r_abc r_def --format json     # both RunDetails + a_text/b_text in one object

​
Output

Side-by-side headers (agent, status, started, finished) with status colours (green/red/yellow), then a minimal line-by-line diff of the assistant output: 
Run A  r_abc
  agent  : viktor
  status : COMPLETED
  ...

Run B  r_def
  agent  : viktor
  status : FAILED
  ...

── Output diff ──
  Reviewed 3 files.
- All checks pass.
+ Linter warnings on src/auth.go:42.
The diff is intentionally not Myers — for run comparison users want the gist, not patch-applicable hunks.

​
See also