Iterating locally

The fast local loop before you submit

Don't burn queued runs to experiment. Install the ha CLI, run ha login and ha clone, then iterate against a task subset with ha run --local — your own OpenRouter key + harbor, pass count and cost in seconds.

The loop


   edit harness ──▶ ha run --local (your key + harbor) ──▶ read pass + cost
        ▲                                                            │
        └────────────────── did it get cheaper / pass more? ─────────┘

Step	What you do	What you get
1	Tweak a skill, agent profile, or routing rule	—
2	Run `ha run --local` with your key	Pass count + billed cost on the local subset
3	Compare to your last run	Signal: cheaper? more passes?
4	Repeat until happy, then `ha submit`	—

Read the local run as a signal, not a score

The local run is ranking signal — it tells you whether a change helped. It is not your final number. The official dollar score only comes from the full ranked run an EC2 worker executes on org-funded infra after ha submit, because absolute cost/pass shifts between a local subset and the full ranked set. Use the local run to steer; trust the scored run for the result.