chore(weave): add `weave-instrument` agent skill by andrewtruong · Pull Request #7370 · wandb/weave

andrewtruong · 2026-06-24T05:33:48Z

Adds the weave-instrument skill which helps an agent add Weave to a Python or TypeScript LLM codebase based on how their code is structured. Supported instrumentation includes the OTEL auto-instrumentation, Session-based manual implementation, and a generic OTEL export.

Doc:

Instrument your agent with the `weave-instrument` skill

weave-instrument is an agent skill that adds Weave tracing to an existing LLM or
agent codebase. Point your agent at your project and ask it to add Weave using this skill.
It will:

wire up weave.init();
set up authentication; and
instrument your code so your runs show up in the Weave UI.

It works for Python and TypeScript!

What it does

The skill picks the right tracing approach for your code instead of applying a fixed
template. It decides based on the kind of traces you want and the structure of your code,
not on which framework you use, so it works for any agent: a known framework, an in-house
one, or a plain loop with no framework at all. It can:

Wrap your agent with the Session SDK (Turn, LLM, Tool, and SubAgent spans) to
produce agent-shaped traces in the Weave Agents tab. This is the universal path and
works for any agent.
Enable OTEL auto-instrumentation for libraries Weave already recognizes, where a
single weave.init() call captures everything with no per-call code.
Send an existing OpenTelemetry pipeline to Weave when your app already emits OTel
spans or owns its own tracer.

It then verifies that traces actually arrive, and tells you the exact URL where your data
will land.

How to use it

Install the skill. Put the weave-instrument folder wherever your coding agent loads
skills from. For Claude Code, that's ~/.claude/skills/ (available in every project) or a
project's .claude/skills/ (just that project). Your agent discovers it automatically.
Authenticate to W&B in your own terminal. Set WANDB_API_KEY (from
https://wandb.ai/authorize) or run wandb login. The skill never handles your key.
Ask your agent to instrument your project, in plain language. For example:
- "Add Weave to my agent."
- "Get my agent into the Weave Agents tab."
- "Trace my tool calls with the Weave Session SDK."
- "Set up Weave logging for this app."

Your agent reads your code, proposes a short plan, makes the edits, and asks you to confirm
the project name (entity/project).

What you get

Tracing that matches how your agent is built, added without changing its behavior: the same
outputs, the same exceptions, and spans that always close. Your runs appear at
https://wandb.ai/<entity>/<project>/weave.

codecov · 2026-06-24T05:46:24Z

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

drtangible · 2026-06-26T19:58:21Z

+  (`pyproject.toml` or `requirements.txt`, and `package.json`). Note the *exact* libraries: `openai`,
+  `anthropic`, `langchain`, `openai-agents` (`import agents`), `claude_agent_sdk`, `google.adk`,
+  `crewai`, `llama_index`, and so on. Step 3 is where you resolve whether the installed Weave


These package names are bit different in package.json.

drtangible · 2026-06-26T20:07:56Z

+2. **What is the structure?** The turn boundary (one input maps to one cycle), the model calls, tool
+   dispatch, sub-agents, and any streaming or concurrency. This is what you instrument, and it exists
+   whatever the library is.
+3. **Probe auto-coverage.** Do not trust a memorized list, because it rots. Check whether the


because it rots

v dramatic

drtangible · 2026-06-26T20:08:41Z

+## Procedure
+
+1. **What shape is wanted?** An agent-shaped tree goes to the Agents tab. Flat model-call traces
+   ("just my LLM calls") go to the Traces tab.


Looks like this might be getting renamed? https://github.com/wandb/core/pull/46847

drtangible · 2026-06-26T20:09:44Z

+- **Agent-shaped, lands in the Agents tab, no extra code:** the OpenAI Agents SDK (`import agents`),
+  the Claude Agent SDK (`claude_agent_sdk`), and Google ADK (`google.adk`, with an import-order
+  caveat).


Should we detail the specific js packages name like we do below as well?

drtangible · 2026-06-26T20:11:09Z

+The Session SDK models Session, then Turn (`invoke_agent`), then a set of LLM (`chat`), Tool
+(`execute_tool`), and SubAgent (`invoke_agent`) spans. It covers loops, unknown frameworks, streaming
+(hold the LLM span open, accumulate, then close it), sub-agents, async work (Python contextvars, or
+TypeScript `runIsolated`), and post-hoc logging (`log_turn` or `log_session`).


The TS SDK doesn't have these log* fns, worth mentioning?

drtangible · 2026-06-26T20:17:38Z

+    for (const tc of msg.tool_calls ?? []) {
+      const tool = weave.startTool({name: tc.function.name, args: tc.function.arguments, toolCallId: tc.id});
+      try { tool.result = await runTool(JSON.parse(tc.function.arguments)); }
+      finally { tool.end(); }
+    }


I believe this is in the wrong block -- msg wouldn't be in scope here.

drtangible · 2026-06-26T20:20:55Z

+  } finally { turn.end(); }
+} finally { session.end(); }


Not sure if it's worth running this through prettier so the agent has example of code more similar to what users' might have set up w/their repos?

drtangible · 2026-06-26T20:21:48Z

+   `WANDB_API_KEY`, or call `await weave.login(apiKey)` once. The user does this; never hard-code a
+   key.
+
+## Canonical pattern (try/finally, because JS has no `with`, so close in `finally`)


Adding to our pod sync list: we should talk about some more friendly APIs to support this.

drtangible

🔥

drtangible · 2026-06-26T21:09:53Z

+  calls, since registry membership alone does not prove it.
+- If you cannot run it (because of missing provider keys or heavy setup), give the user an exact
+  copy-paste command to run themselves, and tell them what to look for: the printed
+  `https://wandb.ai/.../weave` link, and a trace in the Agents or Calls tab.


We call out the parameterized https://wandb.ai/<entity>/<project>/weave thing elsewhere, should we be consistent?

drtangible · 2026-06-26T21:15:05Z

+  test, or a tiny script that exercises one turn) and confirm that a Weave URL is printed and a trace
+  shows up at the project URL.


I don't see any urls when running the TS examples -- we have a printCallLink setting that appears only relevant to some op-related stuff. Following up in Slack, since the skill seems to be relying on this.

andrewtruong force-pushed the andrew/weave-instrument-skill branch 2 times, most recently from ffa1c9b to 2b1f673 Compare June 26, 2026 16:50

andrewtruong marked this pull request as ready for review June 26, 2026 19:40

andrewtruong requested a review from a team as a code owner June 26, 2026 19:40

drtangible reviewed Jun 26, 2026

View reviewed changes

drtangible approved these changes Jun 26, 2026

View reviewed changes

drtangible reviewed Jun 26, 2026

View reviewed changes

andrewtruong added 2 commits June 26, 2026 17:48

fix(weave): print Weave data URL on Node init

db7e8af

chore(weave): add weave-instrument Claude Code skill

52bf68a

andrewtruong force-pushed the andrew/weave-instrument-skill branch from 0f3e658 to 52bf68a Compare June 26, 2026 21:51

andrewtruong requested a review from a team as a code owner June 26, 2026 21:51

andrewtruong changed the base branch from master to andrew/ts-init-print-weave-url June 26, 2026 21:51

Base automatically changed from andrew/ts-init-print-weave-url to master June 26, 2026 22:03

andrewtruong changed the title ~~chore(weave): add weave-instrument Claude Code skill~~ chore(weave): add weave-instrument agent skill Jun 30, 2026

andrewtruong merged commit 095c41b into master Jun 30, 2026
110 checks passed

andrewtruong deleted the andrew/weave-instrument-skill branch June 30, 2026 00:38

github-actions Bot locked and limited conversation to collaborators Jun 30, 2026

Navigation Menu

Uh oh!

chore(weave): add `weave-instrument` agent skill - #7370

chore(weave): add `weave-instrument` agent skill#7370
andrewtruong merged 2 commits into
masterfrom
andrew/weave-instrument-skill

andrewtruong commented Jun 24, 2026 •

edited

Loading

Uh oh!

codecov Bot commented Jun 24, 2026

Uh oh!

drtangible Jun 26, 2026

Uh oh!

drtangible Jun 26, 2026

Uh oh!

drtangible Jun 26, 2026

Uh oh!

drtangible Jun 26, 2026

Uh oh!

drtangible Jun 26, 2026

Uh oh!

drtangible Jun 26, 2026

Uh oh!

drtangible Jun 26, 2026

Uh oh!

drtangible Jun 26, 2026

Uh oh!

drtangible left a comment

Uh oh!

drtangible Jun 26, 2026

Uh oh!

drtangible Jun 26, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

		test, or a tiny script that exercises one turn) and confirm that a Weave URL is printed and a trace
		shows up at the project URL.

Uh oh!

Conversation

andrewtruong commented Jun 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Instrument your agent with the weave-instrument skill

What it does

How to use it

What you get

Uh oh!

codecov Bot commented Jun 24, 2026

Codecov Report

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

drtangible left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

andrewtruong commented Jun 24, 2026 •

edited

Loading

Instrument your agent with the `weave-instrument` skill