skills
langchain-ai/langchain-skills/config/skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- 32d6228affab
- Files read
- 53
This skill pack teaches AI coding agents how to build, test, and deploy LangChain, LangGraph, and Deep Agents applications, and includes helper scripts for parallel task batching and evaluation. It contains no malicious code or severe flaws, and its main risks are a starter guide that turns on LangSmith cloud tracing without asking and a swarm batching mode where untrusted text in one item can influence neighboring items' results.
What it can reach
- Files
- Skills guide the coding agent in scaffolding project directories and .env files; the swarm runtime reads input files and writes JSONL state tables via agent file tools, and eval-engineering scripts read JSON schemas and SQLite databases (read-only) and write JSON reports.
- Network
- Several skills instruct the agent to fetch live documentation from docs.langchain.com and configure applications that call LLM provider APIs, TypeSafe, and LangSmith cloud.
- Shell commands
- Quickstart, CLI, and evaluation skills instruct the agent to install packages and run commands with pip, uv, uvx, npm, npx, langgraph, mda, harbor, and Docker.
- Secrets
- Skills instruct configuring environment variables and .env placeholders for model provider API keys, LANGSMITH_API_KEY, TYPESAFE_API_KEY, and DATABASE_URL while advising not to paste live secrets into chat.
Findings
-
LowStarter primer tells agents to enable LangSmith cloud tracing by default
config/skills/ecosystem-primer/SKILL.md:116
The ecosystem-primer skill is marked INVOKE FIRST for every LangChain, LangGraph, or Deep Agents project and instructs the agent to always set LANGSMITH_TRACING=true for observability. Unlike the quickstart skills, which skip LangSmith unless requested, this causes scaffolded applications to send prompts, retrieved documents, and tool outputs to LangChain's hosted LangSmith servers whenever a LANGSMITH_API_KEY is present in the user's environment or .env file.
Fix: Make LANGSMITH_TRACING=true opt-in after asking the user, matching the quickstart skills, and note that tracing uploads prompts and tool inputs/outputs to LangSmith cloud.
-
LowSwarm auto-batching combines multiple items into one prompt without delimiter escaping
config/skills/swarm/scripts/batching.ts:204
When a swarm table has more than 10 rows and batchSize is omitted, up to 50 rows are grouped into a single model or subagent call. renderItemsBlock concatenates raw item values using unescaped [id] line prefixes, and unpackBatchResults trusts the returned id fields. If one untrusted record or file contains forged [other_id] lines or prompt injection text, it can spoof or corrupt the structured results merged into up to 49 other items in the same batch, or influence a shared tool-using subagent across batched files.
Fix: Default to batchSize: 1 when processing untrusted records or running tool-enabled subagents, and encode batched items with structured delimiters rather than raw [id] lines.
-
LowTypeScript dependency templates use unpinned latest tags in package.json
config/skills/langchain-dependencies/SKILL.md:214
The TypeScript project templates in langchain-dependencies specify "latest" instead of a bounded semver range for deepagents, @langchain/tavily, and @langchain/pinecone. Fresh installs without a lockfile will pull whatever release is currently tagged latest on npm, bypassing major-version bounds and increasing exposure to breaking or compromised releases.
Fix: Replace "latest" specifiers in package.json examples with bounded semver ranges and recommend committing a lockfile.
