| name | skill-publisher |
| description | Publishes a local skill directory to the API hub catalog and the GCS bundle store. Runs the four-step author pipeline end to end, packing the source tree into a deterministic .skill zip, signing the manifest with an ed25519 private key, uploading the zip to GCS, and registering the signed manifest with API hub as an API plus Version plus Spec triple along with its four attribute values. Idempotent by design; re-running with identical inputs produces byte-identical output. Includes a bootstrap mode for first-time self-publication. |
| license | Apache-2.0 |
| compatibility | opencode, antigravity, gemini-cli |
| metadata | {"pipeline_scripts":["scripts.pack_skill","scripts.sign_skill","scripts.upload_skill","scripts.register_skill"]} |
skill-publisher
⚠️ Runtime requirements (read this first)
This skill requires its bundled venv wrapper. publish.sh
internally invokes the four scripts/* Python modules
(pack_skill, sign_skill, upload_skill, register_skill),
which depend on cryptography, google-auth, requests, and
pyyaml. Those are installed into a per-user venv by the
installer (bin/install-skill-publisher.sh). They are not
available to the system python3.
Always invoke publish.sh via the wrapper, never via bare
bash:
${SKILL_DIR}/bin/run-with-venv.sh <args>
bash ${SKILL_DIR}/scripts/publish.sh <args>
The Command blocks in this file use the wrapper form. Do not
"simplify" them — publish.sh reads $PYTHON from its
environment and the wrapper is the only thing that sets it
correctly. If you invoke publish.sh directly, the four
sub-scripts will fail with ModuleNotFoundError.
If the wrapper at ${SKILL_DIR}/bin/run-with-venv.sh is missing,
the skill was installed by a pre-v0.1.4 installer. Re-run
bin/install-skill-publisher.sh to regenerate the wrapper.
Use this skill when the user asks to "publish a skill", "ship this
skill to API hub", "upload and register my skill", "release the
skill", or any close paraphrase. This is the single entry point for
author-side distribution; it wraps the four scripts/* CLIs from
the upstream source repo into one orchestrated sequence with
fail-fast semantics.
What this skill does NOT do
- It does not create the GCS bucket. The bucket must already
exist (see repo README §2).
- It does not create API hub attribute definitions. Run
python3 -m scripts.update_taxonomy once per project first
(see repo README §4).
- It does not generate or rotate signing keys. The
ed25519 private key path is an input.
- It does not install the published skill into any local
runtime. That is
skill-finder's job.
If any prerequisite is missing the underlying script returns a
non-zero exit code and this skill surfaces it verbatim and stops.
Inputs
| Variable | Source | Required | Notes |
|---|
SKILL_SRC | User / agent | yes | Path to a skills/<name>/ source directory containing manifest.yaml and (at minimum) SKILL.md. |
SKILL_OUT | Operator | no | Output path for the packed zip. Defaults to /tmp/${name}-${version}.skill derived from the manifest. |
GCS_BUCKET | Operator (${GCS_BUCKET}) | yes | GCS bucket name (no gs:// prefix). Must match the bucket portion of the manifest's gs_uri. |
SIGNING_PRIV_KEY | Operator (${SIGNING_PRIV_KEY}) | yes | Path to the raw 32-byte ed25519 private key. See the upstream README's author-flow section for how to generate one. |
APIHUB_PROJECT | Operator (${APIHUB_PROJECT}) | yes | GCP project hosting the API hub instance. |
APIHUB_LOCATION | Operator (${APIHUB_LOCATION}) | yes | API hub region (e.g. us-central1). |
REPO_ROOT | Auto | no | Root of the upstream source checkout. Defaults to the current working directory. The four scripts/* modules must be importable from there. |
BOOTSTRAP | User / agent | no | If set to 1 or true, runs the bootstrap path that publishes itself. See "Bootstrap mode" below. |
Runtime dispatch
This skill ships one orchestration contract that runs on multiple
runtimes. Pick the invocation path that matches your runtime; the
behaviour and stdout contract are identical.
| Runtime | How you invoke the publish script |
|---|
| OpenCode | Use the !bash`` injection blocks below. OpenCode auto-executes them on SKILL.md load. |
| Gemini CLI | Use your bash tool. Run the matching Command block, substituting ${SKILL_DIR} with the install path (~/.gemini/skills/skill-publisher for global; <project>/.agents/skills/skill-publisher for workspace) and the operator-controlled env vars from the table above. |
| Antigravity | Same as Gemini CLI but ${SKILL_DIR} is ~/.gemini/antigravity/skills/skill-publisher for global installs. |
| Any other runtime | Same as Antigravity / Gemini CLI: invoke via whatever bash mechanism the runtime provides. |
Mode 1: Publish a third-party skill
Use this when the user wants to publish any skill other than
skill-publisher itself.
Steps
-
Resolve the source path. Take the value from SKILL_SRC
verbatim if provided. Otherwise ask the user once: "Which skill
directory do you want to publish?" Expect a path like
skills/apigee-policy-top10. Do not paraphrase the answer.
-
Validate the inputs before invoking bash. Verify that the
following are all set and non-empty: SKILL_SRC, GCS_BUCKET,
SIGNING_PRIV_KEY, APIHUB_PROJECT, APIHUB_LOCATION. If any
is missing, REFUSE to invoke and tell the user which one is
missing. Do not guess.
-
Reject shell metacharacters in SKILL_SRC. The path is
interpolated into a bash command via POSIX double quotes, which
do not suppress $(...) command substitution or backtick
expansion. If SKILL_SRC contains any of
` $ \ " ; & | < > ( ) { }
REFUSE to invoke and ask the user to rephrase.
-
Invoke the publish script using the runtime path from the
table above. The exact command is:
Command (OpenCode auto-runs this; on Antigravity / Gemini
CLI run it via your bash tool):
!bash ${SKILL_DIR}/scripts/publish.sh \ --src "${SKILL_SRC}" \ --bucket "${GCS_BUCKET}" \ --priv-key "${SIGNING_PRIV_KEY}" \ --project "${APIHUB_PROJECT}" \ --location "${APIHUB_LOCATION}"
-
Surface the script's stdout verbatim. The script prints
one [skill-publisher] step N/4 ... line per pipeline step,
followed by the GCS URI and the API hub resource path on
success. Reproduce every line as-is; do NOT summarize. The
lines are a stable contract for log parsing.
-
On non-zero exit, stop. Surface the failing step's stderr
verbatim. Do not retry, do not "fix" the manifest, do not
re-sign with a different key. The four underlying scripts have
distinct exit codes (see "Exit code mapping" below); use them
to give the user an accurate diagnosis.
Mode 2: Bootstrap — publish skill-publisher itself
Use this when the user wants to publish skill-publisher for the
first time, or republish it after editing its own source.
The bootstrap problem
skill-publisher cannot use its own installed copy to publish
itself: at first run the installed copy does not yet exist in the
API hub catalog. Once published, subsequent versions of
skill-publisher can be republished by the currently-installed
copy — but the very first publish is a chicken-and-egg situation.
Steps
-
Confirm the user actually means bootstrap. Ask once: "Do
you want to publish skill-publisher itself (bootstrap mode)?
This is normally only done by repo maintainers." If the user
says no, switch to Mode 1.
-
Validate the inputs exactly as Mode 1 step 2. SKILL_SRC
for bootstrap is always the path to the in-repo
skills/skill-publisher/ directory; default to that if the
user does not specify.
-
Invoke with --bootstrap. The bootstrap flag tells
publish.sh to use the repo-local scripts/publish.sh (this
file) directly via the four sibling Python modules in
${REPO_ROOT}/scripts/, rather than relying on any installed
copy. The exact command is:
Command:
!bash ${SKILL_DIR}/scripts/publish.sh \ --src "${SKILL_SRC:-skills/skill-publisher}" \ --bucket "${GCS_BUCKET}" \ --priv-key "${SIGNING_PRIV_KEY}" \ --project "${APIHUB_PROJECT}" \ --location "${APIHUB_LOCATION}" \ --bootstrap
-
Surface stdout verbatim (same as Mode 1 step 5).
-
On success, advise the user to re-run with Mode 1. Tell
them: "skill-publisher is now in the catalog. Any future
publish — including republishing skill-publisher itself — can
use the installed copy via skill-finder + this skill, without
--bootstrap."
Exit code mapping
publish.sh exits with the exit code of the first failing
step. The mapping below is the union of the four underlying
scripts' exit codes (see each script's docstring for detail):
| Exit code | Meaning | Where it comes from |
|---|
0 | All four steps succeeded. | n/a |
1 | User error: bad CLI args, missing file, invalid YAML, ADC unavailable, or empty required env var. | any step |
2 | System error: filesystem write failure, GCS network error, GCS 404 (bucket not found), or API hub network error. | pack / upload / register |
3 | Either (a) cryptographic error during sign (priv key unreadable / wrong length) OR (b) GCS IAM denial (403) during upload OR (c) API hub IAM denial (403) during register. The failing-step log line in stdout disambiguates. | sign / upload / register |
4 | API hub rejected the registration because the four attribute definitions are not initialised. Run python3 -m scripts.update_taxonomy once and retry. | register only |
5 | Packaging policy violation: scripts/common/ has the wrong file set (missing or extra files), or the source dir is missing SKILL.md. | pack only (originally exit 3 in pack_skill; remapped here to disambiguate from sign/IAM). |
The [skill-publisher] step N/4 FAILED line printed by
publish.sh immediately before exit tells the user which step
failed and what the underlying script's raw exit code was, so the
remap above is auditable, not opaque.
Idempotency
All four underlying scripts are idempotent by design:
pack_skill.py writes the zip with sorted entry order, so
sha256(zip) is stable across builds.sign_skill.py produces byte-identical output for identical
inputs (ed25519 is deterministic per RFC 8032; YAML is dumped
with sort_keys=True).upload_skill.py overwrites the object at the same name —
re-running uploads the same bytes to the same key.register_skill.py reads first, only POSTs/PATCHes on diff;
re-running with no changes makes only GET calls.
Therefore re-running publish.sh with the same inputs is safe
and observable as zero mutating API calls after the first
successful publish. The skill does NOT need a "dry-run" mode of
its own; pass --dry-run through to register_skill if you want
to skip the registration step entirely (currently not wired —
file an issue if needed).
Security notes
- The ed25519 private key path is the only secret this skill
touches. It is read by
sign_skill.py, never logged, never
uploaded. The key file should have mode 0600.
- The signature is computed over the canonical manifest bytes
(see
scripts/common/canonical.py). Any post-signing edit to
the manifest invalidates the signature; skill-finder will
refuse to install the skill.
- The
signing_key_id field is the sha256 of the public key,
not the private key. It is safe to publish. skill-finder
cross-checks it against installed trust roots
(keys/*.pem); a manifest signed by a key not in
the trust root is rejected client-side. Multiple keys can be
installed side-by-side to support multi-publisher orgs and
zero-downtime key rotation.
- GCS uploads use ADC. The operator must have
storage.objects.create on the target bucket. The skill does
not run a permission pre-flight; it relies on the upload itself
to surface 403 as exit code 3.
Common rationalizations
| Rationalization | Why it fails here |
|---|
| "I'll skip the pack step — the zip already exists on disk." | sign_skill.py writes zip_sha256 from a fresh sha256 of the zip bytes. If the zip on disk is stale (e.g. a SKILL.md edit hasn't been re-packed), the manifest commits to a hash that does not match the bytes skill-finder will fetch from GCS, and every install fails signature verification. Always run pack-then-sign in one sequence. |
| "I'll re-sign without re-packing — only the manifest changed." | Same trap. If the manifest text changed but the zip bytes also changed (because pack_skill would have noticed a source-tree edit), zip_sha256 is wrong. The pipeline is a unit; do not split it. |
| "I'll upload before signing — saves a round trip." | If signing fails (bad key, invalid manifest), the GCS object now points at a zip whose manifest is unsigned. Any client that fetches it gets a manifest that fails schema validation. Sign first, upload second, register third — this order is enforced by publish.sh. |
| "I'll register before uploading — the catalog can wait for the zip." | skill-finder fetches the zip immediately after the manifest passes signature check. A registered manifest pointing at a non-existent gs_uri returns 404 to every install attempt. Upload before register. |
"I'll use gcloud storage cp instead of upload_skill.py." | The script uses ADC + the GCS JSON API directly (the runtime tree is capped at four packages, so no google-cloud-storage dependency). It also has structured exit codes that this skill remaps. Using gcloud breaks the exit-code contract and the offline test suite. |
"I'll skip update_taxonomy.py — the attributes will auto-create." | They will not. register_skill.py PATCHes attribute values by reference to attribute definitions that must already exist. Without update_taxonomy.py having run once, the PATCH returns 400 and this skill exits 4. The first publish in a new project always needs update_taxonomy.py first. |
Layout
skills/skill-publisher/
SKILL.md # this file
manifest.yaml # unsigned template; sign-step rewrites
# zip_sha256, signing_key_id, signature
# in place
scripts/
publish.sh # the bash orchestrator that runs the
# four scripts/* modules in order
publish.sh is the entire runtime surface of this skill. It
shells out to:
python3 -m scripts.pack_skill --src <SKILL_SRC> --out <SKILL_OUT>python3 -m scripts.sign_skill --manifest <SKILL_SRC>/manifest.yaml --zip <SKILL_OUT> --priv-key <SIGNING_PRIV_KEY> --in-placepython3 -m scripts.upload_skill --zip <SKILL_OUT> --bucket <GCS_BUCKET> --object-name <name>-<version>.skillpython3 -m scripts.register_skill --manifest <SKILL_SRC>/manifest.yaml --project <APIHUB_PROJECT> --location <APIHUB_LOCATION>
The four module paths are stable; if the repo ever renames them,
update publish.sh and bump the skill version.
