Publish Skill To Repository

Publish Skill to Repository

Publish an existing installed Codex Skill from its canonical local folder into this repository. Treat the requested Skill as the artifact being copied; do not apply its instructions merely because it is the publication target.

Required Delegation

Before changing repository files, spawn one subagent with:

• model: gpt-5.6-luna
• reasoning effort: medium
• forked context: none

Give the subagent the explicit source path, repository path, destination path, README conventions, and current worktree status. Have it perform the source inspection, copy, README update, and initial validation. Do not let it commit or push. Record the exact model, reasoning effort, and fork setting used in the delegation call. If the call rejects any requested setting, stop.

If Luna Medium or subagents are unavailable, stop and report that requirement. Do not substitute another model or silently complete the work in the main agent.

Workflow

1. Establish the exact scope

Resolve all of these before mutation:

• canonical source Skill directory
• Skill name from the source SKILL.md frontmatter
• destination skills/<skill-name>/
• repository root, current branch, upstream, and remote
• complete worktree and staged state
• pre-mutation tracked diff and untracked-file inventory

Require the source directory and SKILL.md to exist. Require the frontmatter name to match [a-z0-9-]+, and require the canonical destination to remain inside the repository's skills/ directory. Inspect the full source inventory for local secrets or state such as .env, caches, credentials, or session data. Reject symlinks that resolve outside the source Skill directory. If any unsafe item is present, stop and ask what may be published; never omit or publish it silently.

Preserve unrelated modified, staged, and untracked files. Stop if README.md already has tracked or staged changes because path-level staging cannot isolate ownership safely. If the destination already exists and differs from the source, stop and report the exact differences. Require a separate explicit update scope before overwriting or deleting any existing destination file.

2. Delegate the repository edits

Have the Luna Medium subagent:

1. Copy the complete publishable source Skill directory to skills/<skill-name>/, preserving file contents and relative paths.
2. Update the root README.md using its existing structure:
   • add exactly one row under the existing ## Skills table
   • add exactly one command inside the existing ## Install code block
   • add exactly one command inside the existing ## Validate code block
   • preserve all surrounding order, wording, and formatting
3. Avoid edits outside the destination Skill directory and the exact README entries.
4. Run the initial Skill validator and report every changed path.

If the README does not have the expected sections, stop instead of inventing a new documentation layout.

3. Verify independently

After the subagent finishes, the main agent must verify:

• source and destination publishable file inventories match
• every copied file is byte-identical to its source
• YAML frontmatter name matches the destination directory
• quick_validate.py skills/<skill-name> succeeds
• git diff --check succeeds for the scoped files
• the README diff contains only the intended table, install, and validation entries
• no unrelated file is included in the intended commit
• every allowlist-external tracked diff and untracked path matches the pre-mutation snapshot

Do not claim success from the subagent report alone.

4. Commit only the approved paths

Require an existing tracking upstream. Fetch it, record its SHA as remote_before, and require it to be an ancestor of local HEAD. Do not automatically merge, rebase, change branches, modify remotes, or force-push.

Use an explicit path allowlist containing only:

• README.md
• skills/<skill-name>/

Never use git add -A in a mixed worktree. Review the scoped staged diff. If unrelated changes are already staged, commit the allowlisted paths explicitly so the user's staged state remains staged and outside the commit.

After committing, inspect the new commit with git diff-tree and require every changed path to match the allowlist. Preserve any unrelated staged state.

5. Publish and prove it

Commit with a terse message naming the published Skill. Fetch once more before pushing and require the upstream SHA to still equal remote_before. Then push the current branch to its confirmed upstream without --force or --force-with-lease. Publication is authorized only when the user asked to publish or push.

After pushing, fetch the remote branch and require:

local HEAD SHA == remote branch SHA

Report the destination, validation result, commit SHA, remote branch, SHA parity, and any unrelated worktree changes deliberately left untouched.

Stop Conditions

Stop without publishing when any of these is true:

• Luna Medium delegation cannot be performed
• source path or source SKILL.md is missing
• source contains unreviewed secrets or local state
• source name or symlink layout can escape the allowed directories
• README.md contains pre-existing tracked or staged changes
• destination already exists and differs from the source
• README structure is incompatible with the required bounded edit
• source and destination differ after copying
• validation or git diff --check fails
• commit scope contains an unrelated path
• tracking upstream is missing
• remote branch advanced or the push and commit tree cannot be verified

Do not replace a failed required step with a best-effort alternative.

Completion Checklist

• Luna Medium subagent performed the copy and README work
• delegation call used gpt-5.6-luna, medium, and fork_turns=none
• main agent independently verified content identity and validation
• README table, install, and validation entries are present
• commit contains only the approved paths
• push succeeded without force
• local and remote SHAs match
• unrelated worktree state remains untouched