sync-zilliz-plugin

name	sync-zilliz-plugin
description	Use when the user asks to sync, update, or check drift of this gemini-cli extension against upstream zilliztech/zilliz-plugin. Triggers on phrases like "同步上游", "sync from zilliz-plugin", "update commands from plugin", "check if plugin upstream changed". Runs scripts/sync.mjs to regenerate commands/zilliz/*.toml from upstream skills/*/SKILL.md.

Sync commands/ from zilliz-plugin upstream

What this sync does

commands/zilliz/<domain>.toml is a Gemini-CLI-native wrapper around the plugins/zilliz/skills/<domain>/SKILL.md content in zilliztech/zilliz-plugin (the path prefix lives in SOURCE_PREFIX). The sync script fetches upstream, strips YAML frontmatter, neutralizes Claude-specific phrasing, wraps the body in the TOML command template (with !{zilliz <domain> --help} injection and safety rules), and writes it back.

Scope

In scope (19 domain commands, auto-synced): cluster, database, collection, partition, index, vector, import, backup, user-role, acl, monitoring, project-region, billing, external-collection, job, on-demand-cluster, privatelink, diagnose, ask-zilliz.

That is every skill upstream ships except the onboarding trio below. If upstream adds another, references to it render as a prose pointer at upstream (not a dangling /zilliz: link) until it is added to DOMAINS.

Out of scope (hand-maintained, do not auto-sync):

• setup.toml — no upstream source; bootstraps zilliz-cli install + auth.
• quickstart.toml / status.toml — adapted from zilliz-plugin/commands/*.md with Gemini-CLI framing changes that a naive sync would clobber.
• GEMINI.md — global context; review manually if upstream restructures skills.

If the user wants to resync onboarding commands too, do it manually: diff against zilliz-plugin/commands/quickstart.md / status.md and merge by hand.

Four things the script does that are easy to break

1. COMMAND_NAMES — the command file, its reference dir, and every cross-reference follow this map when the command should not be named after the upstream skill (ask-zilliz ships as /zilliz:ask). .sync-state.json still keys on the upstream skill name and records the command alongside it.
2. HELP_CMDS — a skill name is not always a zilliz subcommand. user-role is zilliz user + zilliz role, project-region is project + volume, monitoring/diagnose are cluster + collection, and ask-zilliz has no CLI surface at all (empty list = no !{} block). Get this wrong and the command silently injects "zilliz-cli not installed" instead of real help. Verify with zilliz <cmd> --help before adding a mapping.
3. ASSET_DIRS — ask-zilliz ships 14 reference files, copied into references/ask/ (named for the command). The prompt resolves that directory at runtime by probing the install location, so the files must actually be committed. The script errors out rather than emitting a prompt pointing at a missing dir.
4. REWRITE — ask-zilliz upstream depends on an Inkeep MCP server this extension does not ship; its Inkeep directives are rewritten onto the bundled references plus https://docs.zilliz.com. A surviving Inkeep mention fails the sync loudly, so upstream edits to that skill need a rewrite rule, not a silent pass-through.

How to run

# Rewrite TOMLs in place, update .sync-state.json
node scripts/sync.mjs

# Report drift without writing (CI-friendly; exit 1 if any drift)
node scripts/sync.mjs --check

# Show what would change without writing
node scripts/sync.mjs --dry-run

# Sync from a different branch (default: main)
SYNC_BRANCH=master node scripts/sync.mjs

# Avoid GitHub API rate limits
GITHUB_TOKEN=ghp_... node scripts/sync.mjs

Requires Node 18+ (uses built-in fetch). A full run makes ~20 GitHub API calls, which blows the 60/hour unauthenticated limit in three runs — pass GITHUB_TOKEN=$(gh auth token) for anything more than a one-off.

Recommended workflow

1. Run node scripts/sync.mjs from the repo root.
2. git diff commands/zilliz/ — review each changed TOML. Look for:
   • New command flags or subcommands (desirable).
   • Regressions in the rendered prompt (e.g., broken markdown, orphan references).
   • Any Claude-flavored phrasing the neutralizer missed — if found, add a rule to the neutralize() function in scripts/sync.mjs. A cross-skill reference only becomes a /zilliz: link when the name is in SHIPPED.
3. Spot-check that the TOMLs parse (tomllib on Python 3.11+, tomli below that): python3 -c "import tomli,glob; [tomli.load(open(f,'rb')) for f in glob.glob('commands/zilliz/*.toml')]".
4. If upstream added a new skill domain not in the DOMAINS array of sync.mjs, add it there (the script writes the new TOML itself, but it does not discover new upstream directories), give it a HELP_CMDS entry if the name is not a zilliz subcommand, and add it to the README and GEMINI.md command tables.
5. Commit: git commit -m "chore(sync): pull from zilliz-plugin @ <short-sha>". The short SHA for each file is in .sync-state.json.

When upstream frontmatter description changes

The TOML description = "..." line (shown in /help inside Gemini CLI) is pulled from the SKILL.md YAML frontmatter. If upstream rewords a description, the sync picks it up automatically — no special action needed.

When to skip a domain

If you have intentional local divergence for one command (e.g., hand-tuned prompt), the cleanest path is to temporarily remove that domain from the DOMAINS array in scripts/sync.mjs, add a comment explaining why, and resync the others. Avoid editing the generated TOML directly — the next sync will overwrite it.

State file

.sync-state.json records, per domain: the upstream source path and the SHA of the most recent commit touching that file. Useful for:

• Quick answer to "which upstream commit are we on for vector?"
• Detecting rebases / force-pushes on the upstream branch.

Commit this file along with the TOML changes.