メニュー
Router for LangChain, LangGraph, and Deep Agents skills. Load this ONCE — it tells you which sub-skill file to read for the specific task.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | langchain-stack |
| description | Router for LangChain, LangGraph, and Deep Agents skills. Load this ONCE — it tells you which sub-skill file to read for the specific task. |
| triggers | ["langchain","langgraph","deep agents","agent framework","create agent","state graph"] |
Do NOT load all sub-skills. Read ONLY the file(s) relevant to the current task.
If unsure which framework to use, read framework-selection first:
→ ~/.skills-library/langchain/framework-selection/SKILL.md (163 lines)
| When | Read |
|---|---|
| Creating agents, defining tools, basic agent loop | ~/.skills-library/langchain/langchain-fundamentals/SKILL.md (394 lines) |
| Human-in-the-loop approval, custom middleware, structured output | ~/.skills-library/langchain/langchain-middleware/SKILL.md (388 lines) |
| RAG: document loaders, splitters, embeddings, vector stores | ~/.skills-library/langchain/langchain-rag/SKILL.md (517 lines) |
| Package versions, installation, dependency management | ~/.skills-library/langchain/langchain-dependencies/SKILL.md (419 lines) |
| When | Read |
|---|---|
| StateGraph, nodes, edges, Command, Send, streaming | ~/.skills-library/langchain/langgraph-fundamentals/SKILL.md (811 lines) |
| interrupt(), approval workflows, error handling tiers | ~/.skills-library/langchain/langgraph-human-in-the-loop/SKILL.md (532 lines) |
| Checkpointers, thread_id, time travel, Store, subgraph persistence | ~/.skills-library/langchain/langgraph-persistence/SKILL.md (560 lines) |
| When | Read |
|---|---|
| create_deep_agent(), harness architecture, SKILL.md format | ~/.skills-library/langchain/deep-agents-core/SKILL.md (423 lines) |
| SubAgentMiddleware, TodoList planning, HITL interrupts | ~/.skills-library/langchain/deep-agents-orchestration/SKILL.md (471 lines) |
| StateBackend, StoreBackend, FilesystemMiddleware, CompositeBackend | ~/.skills-library/langchain/deep-agents-memory/SKILL.md (301 lines) |
~/.skills-library/langchain is a symlink to the langchain-skills submodule — don't hardcode the submodule path, use this alias so the skill stays portable if the submodule moves.
Address Greptile PR review comments — read inline comments and summary, fix issues locally, push, reply to each thread, then tag Greptile to re-score and resolve threads.
End-to-end feature development loop. You describe a feature, iterate on the plan, then the team implements, reviews, ships a PR, and handles the configured PR-review bot autonomously. Notifies you at milestones. Use ONLY when the user wants the full autonomous implement→review→PR→verify loop for a multi-file feature. Do NOT use for: quick bug fixes, single-file edits, exploratory/discussion tasks, or anything the user wants to drive step-by-step.
Fetches and debugs a Langfuse trace OR session by ID or URL. A `.../sessions/<id>` URL renders a turn-by-turn conversation transcript (USER → TOOLS → ANNIE) across all the session's traces. A trace renders a span-tree overview with auto-suggested drill targets, then drills into a chosen section to surface system prompts, tool calls, and tool results. Renders both Anthropic block-style and langchain `tool_calls`-key messages, and strips noise (tool-call ids — incl. the giant gemini `__thought__` ids — plus `additional_kwargs`/`response_metadata`/message ids). Supports comparing two spans side-by-side. Use when the user provides a Langfuse trace/session ID or URL, mentions a langfuse.* link, or asks to debug, inspect, diagnose, or compare an LLM trace or conversation. Defaults to anyformat credentials (LANGFUSE_TRACING_*) and host (langfuse.anyformat.ai); other hosts work if LANGFUSE_HOST and LANGFUSE_PUBLIC_KEY/LANGFUSE_SECRET_KEY are set.
Query and debug anyformat's Datadog observability (logs, metrics, traces, monitors, incidents) via the plugin:datadog:mcp server, accounting for our non-standard loguru log structure. Use when searching or debugging anyformat logs/metrics/traces/monitors/incidents in Datadog, or when a Datadog log search for anyformat-core/backend returns 0 results.
Commit changes grouped by logical chunks, push to a feature branch, link a Linear ticket, and create a labelled PR. Targets 'dev' if it exists on origin (anyformat convention), otherwise the repo's default branch. Triggers on: commit, push, create PR, ship it, send PR.
Address failing CI checks on the current PR — group failures by root cause, fix locally, push, then watch the re-run. Pairs with react-to-greptile (comments) for full PR clean-up.
