apigee/devrel/skill-publisher

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`:

# CORRECT - the wrapper exports PYTHON to point at the venv:
${SKILL_DIR}/bin/run-with-venv.sh <args>
 
# WRONG - publish.sh will fall back to system python3 which
# will hit ModuleNotFoundError mid-pipeline:
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

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.

Mode 1: Publish a third-party skill

Use this when the user wants to publish any skill other than skill-publisher itself.

Steps

1. 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.

1. 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.

1. 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.

1. **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}"

1. 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.

1. 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

1. 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.

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.

1. 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

1. Surface stdout verbatim (same as Mode 1 step 5).

1. 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):

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