Agents in this repository check their work before merging with platform/ci/run_workflow_locally.py, which runs a GitHub Actions workflow's read-only steps in the session's own container. It had two ways in: name one workflow, or sweep every workflow your diff would start (--for-changes).
The gap between them showed up in the run records. On 2026-11-16 and 11-17, two sessions wrote nearly the same complaint into their doc_friction field: the tool "takes one workflow per call", and when an edit voided three workflows in a sweep, re-running just those three "needed one invocation per workflow", because --only (the filter for sweeps) was refused on its own.
What changed
The runner now takes several names, with or without .yml:
python3 platform/ci/run_workflow_locally.py doc-refs-ci prompts-ci
--only doc-refs-ci,prompts-ci on its own does the same thing, so the spelling a session reaches for after a sweep works too. Each workflow runs in turn with the same per-job screen as a by-name run, and the output ends with one line per workflow:
Summary of the named workflows:
exit 0 doc-refs-ci.yml
exit 0 prompts-ci.yml
The combined exit code is 0 only when every workflow passed. A usage error (2) wins over a failure (1), which wins over the "no verdict" codes the runner already used, so a script reading the exit code is not told "pass" when one workflow could not be judged. A name that does not exist is refused before anything runs, rather than after the first workflow has spent its minutes.
Checked
The tool's own suite (230 tests, five of them new) passes, and the real run of doc-refs-ci prompts-ci shown above passed both. main_health.py reported 254 of 254 checks green on main before the change.