Dev Release

RAI Release: Bump + Tag (local) → Push + Publish (human-gated)

Claude Code treats plugin.json version as the plugin cache key — if you ship code changes without bumping, end users silently keep the cached copy. All three manifests that carry the plugin version must stay in sync.

The helper script handles the mechanical, local parts (bump manifests, commit, create the tag). It deliberately does not push or publish — those steps happen under human review via the publishing procedure, so the releaser can review repo state and shape release notes before anything user-visible lands on origin.

Assumptions

This workflow targets the canonical repo and assumes:

• You're on main in a clone of RelationalAI/rai-agent-skills, with origin

pointing at it.

• You have push access to main and permission to create tags + releases on

origin. Only RAI maintainers do — if git push origin prompts for credentials or gh release create 403s, stop.

• gh is installed and authenticated.
• Working tree has no uncommitted tracked changes (untracked files are fine).

Running against a fork or a feature branch will put the release on the wrong branch/repo. Don't do it.

Files carrying the plugin version

The plugin version lives in exactly three files — the official plugin metadata for each supported host (Claude Code, Codex, Cursor). Keep them identical. Any other version field in the repo is out of scope for this skill.

• plugins/rai/.claude-plugin/plugin.json
• plugins/rai/.codex-plugin/plugin.json
• plugins/rai/.cursor-plugin/plugin.json

> Adding or removing a plugin host (e.g. Gemini, a new marketplace) means > this list changes. When that happens, update both: > - This skill — the list above, the count in the intro, the wc -l expected > value in the Manual path verification step, and the OLD/NEW bump loop. > - The script — the FILES array in contrib/dev-release/scripts/release.sh. > > A release run with an out-of-date FILES array will silently skip the new > manifest.

Recommended path — run the helper script

From the repo root:

# default: patch bump
contrib/dev-release/scripts/release.sh

# or pick a bump level
contrib/dev-release/scripts/release.sh minor
contrib/dev-release/scripts/release.sh major

# or specify an exact version
contrib/dev-release/scripts/release.sh 1.2.3

# preview without executing
contrib/dev-release/scripts/release.sh --dry-run minor

The script runs the preflight checks from Assumptions above, asserts that origin points at RelationalAI/rai-agent-skills (no fork misrouting), verifies HEAD is not already at the latest release tag (no double-bump after a partial prior run), reads the current version from the manifests (and fails if they've drifted), computes the next version, prints a plan, and prompts before executing the bump → commit → tag sequence. All three steps are local — nothing is pushed.

Pushing the bump commit and tag, and creating the GitHub release, happen in the publishing procedure below, under human review. This keeps everything user-visible — the commit on main, the public tag, and the release — behind the same go-ahead gate.

Publishing the GitHub release (agent-driven, human-gated)

After release.sh finishes, the bump commit and tag exist only locally — nothing is on origin. The agent pushes both and creates the release, but only after surfacing context, collecting any release-note commentary from the user, and getting an explicit go-ahead.

Agent steps:

1. Surface the context so the user can decide what they want to say.

Summarize:

• The commit range since the previous tag

(git log ..v$NEW --oneline).

• User-visible changes the agent noticed (skill renames, new skills,

breaking path changes, etc.).

• Anything not obvious from commit subjects alone.

1. Invite commentary with a single open question — don't enumerate

mechanisms. Something like: "Want to add any commentary to the release notes, or should I publish with just the auto-generated commit log?"

Based on the user's response, pick the matching gh release create invocation from the reference table below. Do not ask the user to choose between mechanisms; the agent owns that mapping.

| User intent | Invocation (append -R RelationalAI/rai-agent-skills) | |--------------------------------------------------------------------|--------------------------------------------------------| | Short paragraph, provided inline in chat | gh release create v$NEW --generate-notes --notes "" — text becomes header, auto-notes appended | | Longer notes drafted (or to be drafted) in a file | gh release create v$NEW --generate-notes --notes-file | | Wants to review/edit the full text before it goes live | gh release create v$NEW --generate-notes --draft, then user edits via web UI or gh release edit and publishes | | No commentary — just ship the auto commit log | gh release create v$NEW --generate-notes |

1. Wait for explicit go-ahead before touching origin. Do not push the tag

or publish on inferred consent — a "sounds good" about the commentary is not the same as approval to ship.

1. On go-ahead, push the bump commit + tag together, then create the release:

``bash git push origin HEAD v$NEW gh release create v$NEW ... # the invocation picked in step 2 `` Pair these so the commit, tag, and release all land together.

1. After publishing, spot-check:

• gh release view v$NEW -R RelationalAI/rai-agent-skills to confirm body

and assets. Add --web to open in the browser.

• Flag anomalies (truncated notes, wrong tag target, unexpected assets)

before closing the loop with the user.

Never push the tag or run gh release create without explicit go-ahead — both are user-visible and hard to un-publish cleanly.

Recovery: push succeeded but release not created

If git push origin HEAD v succeeded but gh release create failed, the tag is on origin with no release attached. Re-run the gh release create variant chosen in step 2 whenever you're ready. Do NOT re-run release.sh; its tag-at-HEAD preflight will correctly refuse.

Recovery: script aborted mid-run, or publishing never happened

If the script created the bump commit + tag locally but the publishing procedure never ran (or was deferred), HEAD will be at the new tag. Re-running release.sh will refuse (the tag-at-HEAD preflight catches this). Either:

• Pick up the publishing procedure from step 1 whenever you're ready —

everything is still local, no rework needed.

• Or unwind: git tag -d v and git reset --hard HEAD^ to drop the

local tag and bump commit, then retry the script.

Manual path (reference)

Use these steps only if the script can't run — e.g. you need a one-off change between bump and tag, or you're on a machine where the script's preflight is too strict.

1. Decide the next version

Tags follow vMAJOR.MINOR.PATCH.

| Change | Bump | |------------------------------------------------|-------| | Bug fix, wording, small cleanup | patch | | New skill, new capability inside a skill | minor | | Breaking change, plugin rename, skill removal | major |

Find the latest tag:

git tag --sort=-v:refname | head -1

2. Update the three manifests

OLD=1.0.10; NEW=2.0.0
for f in plugins/rai/.claude-plugin/plugin.json \
         plugins/rai/.codex-plugin/plugin.json \
         plugins/rai/.cursor-plugin/plugin.json; do
  sed -i.bak "s/\"version\": \"$OLD\"/\"version\": \"$NEW\"/" "$f" && rm "$f.bak"
done

Verify every non-example manifest is on the new version — expect exactly 3 matches. Any other count means drift or a missed manifest:

grep -rn "\"version\": \"$NEW\"" .claude-plugin .agents .cursor-plugin plugins | grep -v /examples/ | wc -l
# Expect: 3

3. Commit and tag (all local)

git add -u
git commit -m "Release v$NEW"
git tag -a "v$NEW" -m "v$NEW"

Nothing is pushed yet. Pushing the bump commit, pushing the tag, and creating the GitHub release all happen together in the publishing procedure (step 4), under human review.

4. Publish the GitHub release

Publishing follows the same agent-driven, human-gated flow as the scripted path — see Publishing the GitHub release above. The agent surfaces context, invites optional commentary, waits for go-ahead, and picks the matching gh release create invocation.

When NOT to use

• Docs-only changes that don't ship to end users — a README typo doesn't need a

plugin-cache bust.

• Changes to version fields outside the three official plugin manifests

listed above — this skill is scoped to the supported plugin hosts.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: RelationalAI
• Source: RelationalAI/rai-agent-skills
• License: Apache-2.0
• Homepage: https://relational.ai

Install and usage instructions live in the source repository linked above.