skill — registry / market / lifecycle
| Command | Notes |
|---|---|
list / market / get <id> / market <id> |
Browse registry + market. |
names <id...> |
Resolve ids → names in one call; unknown ids are skipped, not an error. |
content <id> [--raw] [--draft] / read-file <id> <rel-path> [--raw] [--draft] / history <id> |
Progressive read + versions. |
files <id> [path] [--tree] [--draft] |
File listing. No path = root; a directory path = its direct children; --tree = whole hierarchy under path (or the whole skill when path is omitted). |
execute <id> --entry '<shell>' [--timeout <s>] [--raw] [--exit-code] |
Run the skill in the platform sandbox. Omit --timeout to leave the limit to the sandbox (300s by default, 3600s max); the client waits out the sandbox's maximum either way. |
set-status <id> <status> |
unpublish | published | offline. |
register <dir> [--source custom|internal] [--extend-info <json>] |
Zip a local skill dir → multipart register. SKILL.md must have frontmatter (name/description). --source defaults to custom, matching the backend's own default (default:"custom" validate:"oneof=custom internal"), so the registered result is unchanged from omitting it. |
download <id> [out.zip] [--draft] / install <id> [dir] |
Save archive / download + unzip. |
update-metadata <id> --body <json> / update-package <id> <dir> |
Edit metadata / replace package (zip). |
republish <id> --version <v> / publish-history <id> --version <v> |
Republish / publish a historical version. |
--draft — which version you read
Without it, reads hit the published release; with it, the draft (current,
possibly unpublished) version. They differ whenever a skill has been edited but
not republished, so pick deliberately. An unpublished skill answers 404 skill not found without --draft. Permissions differ too: draft reads want
view/modify, published reads want execute/public_access/view.
--raw — file text instead of a link
content/read-file return JSON holding a pre-signed object-store URL whose
host (minio.…svc.cluster.local) only resolves inside the cluster — useless
from a laptop. --raw writes the file's own bytes to stdout instead, so
read-file <id> <path> --raw > file reproduces the file byte for byte. Binary
files are refused with a pointer to install, never dumped as mojibake.
--raw and --draft compose. Draft reads are served inline by the backend in
one request; published reads fall back to fetching the archive once per run and
serving every later read from it.
Running a skill
openbkn skill execute <id> --entry 'python scripts/generate_page.py --list-styles'
openbkn skill execute <id> --entry 'python scripts/generate_page.py --style brutalist' --raw > page.html
openbkn skill execute <id> --entry 'python check.py' --exit-code # exit status = sandbox's
The platform uploads the package into a sandbox session and runs --entry from
its work dir, so scripts reach bundled files by relative path
(styles/brutalist/tokens.json). --raw writes the run's stdout straight
through (stderr stays on stderr) so output can be redirected into a file;
--exit-code is opt-in, since it changes what the process returns. A sandbox
that reports mocked=true exits 125: nothing failed, but nothing ran either,
and --exit-code && next-step must not read that as a pass. An exit code the
shell cannot represent (>255) becomes 1 rather than truncating to 0.
Check mocked in the response: true means the sandbox stubbed the run and the
skill never executed. --raw prints that warning to stderr so it can't
contaminate a redirected capture.