| name | verify |
| description | Verify claude-commit changes by driving the cco CLI end-to-end on real staged changes - launch/drive recipe and gotchas. |
Verifying claude-commit
The surface is the cco CLI (bin/cco.ts). Tests and typecheck are CI's job;
verification means running the CLI against real staged changes.
Recipe
- Stage something (the change under verification works):
git add -A
- Happy path:
bun run bin/cco.ts --dry-run --no-spinner --verbose
- Prints chunk count, per-call cost, each summary, and the final message(s).
Dry-run never commits.
- A successful run exercises both pipeline stages (summary model and final
model) through the Claude Agent SDK and the bundled
claude binary.
- Error-surface probe: add
--model-summary bogus-model-xyz and expect a
clean error naming the model, exit code 1. Put the bogus model on
--model-summary (not --model-final) so the run fails before any paid
summary call.
- Interactive mode:
bun run bin/cco.ts --dry-run -i (OpenTUI; needs a TTY).
Gotchas
- Needs staged changes; the pipeline errors out otherwise.
- Requires Claude Code subscription auth; an API key is refused unless
allowApiKey is enabled.
- When checking exit codes, do not pipe through
head - $? reports the last
command in the pipeline, not cco.
- Model aliases resolve inside the bundled
claude binary (e.g. sonnet ->
claude-sonnet-5). To see the resolved model, context window, and cost, run
./node_modules/.bin/claude -p 'Reply OK' --model <m> --output-format json
and read the modelUsage key.
