langsmith-fetch-skill

為什麼要用

核心特性

Pulls traces directly from LangSmith Studio via CLI
Auto-triggers on debugging-flavored prompts
Tracks failures + root causes
Latency + token usage analysis
Tool-call and memory operation review
Session export

即時演示

實際使用效果

langsmith-fetch-skill.replay ▶ 就緒

0/0

安裝

作者儲存庫 · Claude Skills 官方文件

使用場景

實戰用法： langsmith-fetch-skill

Diagnose a slow agent without opening LangSmith UI

👤 LangChain/LangGraph devs ⏱ ~20 min intermediate

何時使用： An agent that used to be snappy is now slow; you want root cause.

步驟

Ask Claude to diagnose

Why is my agent <name> slow? Pull the last 20 traces from LangSmith and find the hotspot.✓ 已複製

→ Skill fetches traces; reports latency by step
Drill in

Focus on trace <id> — break down tool-call times.✓ 已複製

→ Per-tool latency with the outlier called out

結果： A specific step to optimize, not a hunch.

注意事項

Optimizing the average when tail latency is the problem — Ask for p95/p99 specifically

搭配使用： filesystem

Root-cause an agent failure from a trace id

👤 Anyone running LangGraph in production ⏱ ~15 min intermediate

何時使用： A user reported an error; you have the trace id.

步驟

Fetch + analyze

Debug trace <id>. What was the immediate error, what's the root cause in the graph?✓ 已複製

→ Error step + upstream decision that led to it

結果： A fix targeting the actual decision point, not the symptom.

Find which node burns the most tokens

👤 Teams optimizing LLM spend ⏱ ~30 min intermediate

何時使用： Monthly review of agent spend.

步驟

Aggregate token usage

Across last 7 days' traces, rank graph nodes by total token usage.✓ 已複製

→ Ranked table with % share
Propose tightening

For the top 3, propose prompt-compression or model downgrade options.✓ 已複製

→ Targeted opt list

結果： Measurable spend reduction without chasing tiny savings.

組合

與其他 MCP 搭配，撬動十倍槓桿

langsmith-fetch-skill + filesystem

Save exports alongside the agent code for lineage

Export the debugging session to debugging/traces/<date>.json.✓ 已複製

langsmith-fetch-skill + github

Attach trace summary to a bug report

Summarize trace <id> and post as comment on GitHub issue #42.✓ 已複製

工具

此 MCP 暴露的能力

工具	輸入參數	何時呼叫	成本
fetch_recent	project, limit	Broad diagnosis	LangSmith quota
fetch_trace	trace_id	Deep dive on one run	LangSmith quota
analyze_failures	traces	Post-fetch analysis	0
export_session	traces	Archive for later	0

成本與限制

運行它的成本

API 配額: LangSmith: depends on plan; Studio has generous quotas
每次呼叫 Token 數: Low — traces are structured
費用: Free skill; LangSmith free tier covers most devs
提示: Use fetch_recent with a narrow window, not full history

安全

權限、密鑰、影響範圍

最小權限： LangSmith: read on your project(s)

憑證儲存： LANGCHAIN_API_KEY in env

資料出站： Only to LangSmith (where the traces already live)

切勿授予： Write to LangSmith — skill only reads

Traces can include user inputs and outputs — treat them as user data under your privacy policy.

故障排查

常見錯誤與修復

langsmith-fetch: command not found

Install LangSmith CLI or ensure it's on PATH

驗證： which langsmith-fetch

Empty results

Check project name and time window; traces are project-scoped

替代方案