菜单
Use when a project hardcodes LLM model IDs (claude-*, gpt-*, gemini-*) that go stale, when the user wants current model IDs resolved at runtime, or on /model-registry [install|migrate|status|refresh-fallback]. Works in TypeScript/Next.js, Python, and any project that can fetch JSON.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Audit a repo's agent-facing documentation layer and dispatch the right fix — instruction-file topology (AGENTS.md canonical? CLAUDE.md/GEMINI.md @import it?), always-loaded token cost, CodeMap freshness and line-number accuracy, broken pointers, changelog-vs-code version drift. Read-only; reports first, then offers remediation in dependency order. Use when asked to "check this repo", "audit the agent docs", "is my CLAUDE.md/AGENTS.md set up right", "are the docs stale", "check the code map", "health check this project", or when onboarding an unfamiliar repo and you want to know whether its docs can be trusted. Also on /repo-doctor. NOT for code review, tests, dependency or security audits.
Audit and rightsize CLAUDE.md / AGENTS.md and skills per Anthropic's Claude 5 context-engineering guidance (Thariq, 2026-07-24) — cut rules the model can judge, extract detail to on-demand files/skills, dedupe layers, and verify. Use when users mention slow /resume, CLAUDE.md being too long, reducing token usage, context optimization, rightsizing instructions, auditing CLAUDE.md, or following up on /doctor output. Triggers on "optimize CLAUDE.md", "rightsize my instructions", "my resume is slow", "CLAUDE.md is too big", "audit my CLAUDE.md", "apply the context-engineering rules", "slim down my instructions".
Sync Claude Code config (CLAUDE.md, agents, skills, commands, instructions, settings, statusline, guard hooks, plugins, MCP servers, CC version) AND the shared cross-CLI house-rules — canonical ~/.config/agents/AGENTS.md plus the per-CLI instruction files that @import it (CLAUDE.md, GEMINI.md) and Codex/Copilot/agy/Pi wiring — from THIS machine to any SSH-reachable host so every AI coding CLI works the same there. Discovers targets from ~/.ssh/config. Use on /sync-claude-config [host], or when asked to "sync claude config to <host>", "mirror my agents/skills to <machine>", "make Claude Code (or my agents) work the same on <host>". Push-only (local → remote); never copies credentials or machine state.
Restructure a project's (or the whole machine's) AI-agent instruction files so AGENTS.md is the single canonical guide that every coding CLI follows — Claude Code, Codex, Copilot, Antigravity/agy, Gemini, Pi — with CLAUDE.md and GEMINI.md reduced to thin @import pointers plus tool-specific extras. Use this whenever the user wants non-Claude agents to behave the same as Claude, mentions consolidating or unifying AGENTS.md / CLAUDE.md / GEMINI.md, says things like "make all my agents follow the same rules", "share my house rules across CLIs", "make Codex/Copilot/Gemini work like Claude", or wants to clean up duplicated/drifting instruction files. Works on the LOCAL project by default, then asks whether to also do the GLOBAL ~/.claude machine-wide setup. Trigger it even if the user doesn't name the skill.
Use when bumping a version, cutting a release, or when a CHANGELOG is stale, missing versions, or disagrees with the code — and before opening any PR, per the versioning rule in AGENTS.md. Also on /version-manager [check|backfill|release], or when the user says "bump the version", "cut a release", "the changelog is old". Auto-detects version locations in any stack (pyproject, package.json, VERSION file, module constants, README display) — no per-repo config.
Create a Discord post (announcement, release recap, changelog summary, community update) for the current project and save it as a markdown file under a Discord/ directory in the project root. Use this whenever the user asks to "create a Discord post", "write a Discord announcement", "make a Discord update", "post this to Discord", "draft a Discord message", or wants a community-facing summary of changes, a release, or news for a project — even if they don't say the word "skill". This skill only drafts and saves the post to a file; it never sends anything to Discord.
| name | model-registry |
| description | Use when a project hardcodes LLM model IDs (claude-*, gpt-*, gemini-*) that go stale, when the user wants current model IDs resolved at runtime, or on /model-registry [install|migrate|status|refresh-fallback]. Works in TypeScript/Next.js, Python, and any project that can fetch JSON. |
Wire a project to the published model registry so it resolves current LLM model IDs at runtime instead of hardcoding IDs that go stale.
https://chameleonlabs-model-registry.s3.us-east-1.amazonaws.com/models/latest.json (public, refreshed daily ~3 AM by agents/model-discovery/ in the ChameleonLabs repo)python/src/model_registry/client.py) + TypeScript (typescript/registry.ts), both zero-dependency and vendorableMODEL_REGISTRY_URL points clients at a different registry| Key | Meaning |
|---|---|
families[provider][family] | primary lookup — current model ID, e.g. families.anthropic.opus → "claude-opus-4-8" |
families_detail[provider][family] | {active, stable?, preview?} channels |
available[provider] | full curated ID list |
capabilities[model_id] | {context_window} |
unmatched[provider] | IDs with no family (never migrate these to family lookups) |
schema_version, fetched_at, providers_ok | validate schema_version >= 2; fetched_at age = staleness |
Families: anthropic opus|sonnet|haiku; openai gpt|gpt-mini|gpt-nano|gpt-pro|chat; gemini pro|flash|flash-lite.
/model-registry <verb>)No args → show this verb table and ask which to run.
install — detect project type, then:
pip install "chameleonlabs-model-registry @ git+https://github.com/Chameleon-Labs-LLC/model-registry-client#subdirectory=python" OR vendor client.py (single file, stdlib-only — prefer vendoring for projects with no dep management). TypeScript → vendor registry.ts (fetch raw from the repo). Other languages → port from the canonical client; behavior contract below.curl -sf <registry-url> -o <project>/model-registry.fallback.json — place it beside the project's central models module (e.g. next to llm_models.py / lib/ai/). If installing via pip, also add the git URL line to requirements.txt/pyproject.toml.fallback_path= / createRegistryClient({fallback})).amplify.yml; other CI → equivalent; plain Python apps → document manual refresh-fallback.migrate — grep for hardcoded IDs (claude-[a-z0-9.-]+|gpt-[a-z0-9.-]+|gemini-[a-z0-9.-]+|o[0-9]-), usually centralized in llm_models.py / models.ts / constants. For each: map to resolve(provider, family) if it belongs to a family lineage — older/dated variants map to their line (gpt-4o → gpt, claude-sonnet-4-20250514 → sonnet, gpt-4.1-mini → gpt-mini); suffixes -mini/-nano/-pro pick the matching family. Leave alone anything in unmatched, pinned-for-a-reason IDs (ask if unsure), and non-chat models (tts/transcribe/image/embedding). Keep sync call sites reading the bundled fallback JSON; only async paths get live data. Propose diffs before applying; run the project's tests/tsc after.
status — fetch live registry; report fetched_at age, diff vs the project's bundled fallback (family IDs that changed), and any remaining hardcoded IDs from the migrate grep.
refresh-fallback — re-run the snapshot curl over the project's fallback file; show the family-level diff.
1-hour in-memory TTL cache; 5s fetch timeout; honor MODEL_REGISTRY_URL; validate schema_version >= 2 + families present. Failure ladder: serve cache (even expired) → serve bundled fallback → only then error. With a fallback configured, never throw — model pickers must always work. Log failures to stderr/console. Ports use the language's idiomatic embed for the fallback (e.g. Rust include_str!); everything else above is the contract.
with { type: "json" } (ERR_IMPORT_ATTRIBUTE_MISSING otherwise).preBuild phase so each deploy ships a fresh last-known-good.lib/ai/registry.ts) — don't reinstall there.