External Repo Intake

External-repository intake

When the user asks to "import", "pull", "vendor", "cherry-pick", or "borrow from" an external GitHub repository (anything not already registered in the sync table of the upstream-sync skill), follow this workflow before touching any local file. This skill covers the first intake. Re-syncing repositories already registered uses the separate upstream-sync skill.

1. Classify the operation

We do not fork, submodule, or add runtime dependencies. The only intake mode for this marketplace is vendoring: a one-shot or tracked copy of upstream content into our tree, with attribution preserved and the content adapted to our conventions.

Combinations are normal (the wshobson/agents intake used cherry-pick plus hybrid merge plus new files across codebase-cleanup and e2e-testing-patterns; its multi-agent generic core, the reverse-engineering vendor, and the e2e-testing-patterns vendor were all later delegated back upstream, and the codebase-cleanup vendor was retired outright in marketplace 19.0.0).

2. Decide the four dimensions

Before writing any file, lock down each dimension. State them back to the user via AskUserQuestion whenever a meaningful choice exists.

3. License compliance gate

Block before fetching:

1. Read the upstream LICENSE file. The four expected outcomes:
   • MIT / BSD / ISC / Apache-2.0: proceed; preserve attribution header in every derived file.
   • MPL-2.0: proceed for documentation-only content; flag to the user before importing source code.
   • GPL-2.0 / GPL-3.0 / AGPL: STOP. Ask the user explicitly; the marketplace is MIT and incompatible licensing must be a conscious decision.
   • No license / proprietary: STOP. Do not import.
2. For Apache-2.0 specifically, check whether upstream has a NOTICE file. If it does, preserve its contents alongside the derived files.
3. Attribution header on every derived file (new or merged section):

   <!--
   Portions of this file are derived from <owner>/<repo>
   (https://github.com/<owner>/<repo>), <SPDX-license> License.
   Snapshot YYYY-MM-DD.
   -->
   

4. Fetch and inspect (read-only)

Use gh api repos/<owner>/<repo>/contents/<path> with --jq '.content' | base64 -d. Save everything to .upstream-scratch/<repo>/ (excluded from commits). Read every fetched file before writing local files. Count and assess size before proposing the merge plan.

5. Adapt to local conventions

Before saving any derived file, scan for and rewrite:

• Dash-aside construct ("X — Y — Z" / "X -- Y -- Z" / "X - Y - Z" bracketing a clause): replace with sentences, parentheses, or colons. Never substitute one dash form for another.
• Emoji: remove if the destination plugin's existing files have none.
• Upstream-specific cross-references: rewrite [reference/foo.md](foo.md) style links to point at the local destination path (or remove if the target was not imported). Rewrite {{template_vars}} and references to upstream-only commands.
• Namespace prefixes: rewrite upstream <their-plugin>:X skill references to the local <our-plugin>:X equivalent, or drop them when we vendor no equivalent.
• Stale tool names in team-related imports: Teammate / Task tool to spawn -> Agent tool; explicit TeamCreate / TeamDelete steps -> rewrite to implicit team formation (the team forms when the first teammate is spawned) and automatic cleanup at session end (both tools were removed in Claude Code 2.1.178).

6. Wire the new content into existing agents and commands

Importing content that no agent reads is wasted work. After saving derived files:

1. Add a ## References Library (or equivalent) entry in the host plugin's SKILL.md that indexes every new/extended reference with a one-line topic description.
2. Update any command or agent that should now consult the new references. Add explicit Read plugins/<plugin>/skills/.../<file>.md instructions in the relevant prompt sections.
3. Avoid preloading discipline: the consumers must read references on-demand, not all upfront. State this in the wiring text.

7. Decide on tracking and re-sync

If the sub-mode is "ongoing sync" (or "snapshot but worth tracking for re-import"), append a row to the sync table in .claude/skills/upstream-sync/SKILL.md, with:

• Plugin (and sub-skill, if applicable) plus license tag for non-MIT sources
• Upstream repo plus the specific subpath
• Full list of derived local files and any merged sections

Then append the matching gh api fetch loop to that skill's "How to sync a plugin" code block. Do this even for snapshots; it makes a future re-import a one-command operation rather than archaeology.

If the sub-mode is "inspiration only", do NOT add a sync-table row. Add an inline note in the affected file describing what was adopted from where, but no sync entry.

Never add a sync-table row for anything listed under "Deliberately not vendored" in CLAUDE.md. Those areas were handed back to their upstreams on purpose.

8. Version bump and commit

• Bump every plugin whose version in marketplace.json had content added.
• Bump metadata.version (minor bump for first-time intake of a new upstream; patch bump for follow-up reworks of an existing intake).
• Single commit with the imported files, the local edits, the SKILL.md wiring, the sync-table update in .claude/skills/upstream-sync/SKILL.md, and the version bumps together.
• Commit message: Cherry-pick / Vendor / Import <subject> from <owner>/<repo> (v<new>) with a short description block listing new files, merged sections, license, and attribution date.

9. Verification before push

• grep derived files for any leftover upstream-only references, stale tool names, or dash-aside constructs.
• Validate marketplace.json JSON syntax.
• git status shows nothing in .upstream-scratch/ staged.
• git diff --stat to sanity-check scope.