Weave Instrument

Your job is to add Weave tracing to someone's LLM or agent code so their runs show up in the W&B Weave UI. Fit the approach to their codebase instead of applying a one-size-fits-all template.

This skill works for any agent, whatever libraries it uses or none at all. Choose the approach based on two things:

• the telemetry shape the user wants;
• the structure of their code.

Never choose based on whether you recognize the framework. There are two mechanisms, and one of them always applies.

1. Session SDK (the agent-logging APIs) is the universal path. You wrap the agent's own structure with Turn, LLM, Tool, and SubAgent spans. It works for any agent, whether a custom loop, an unknown framework, or one that Weave does not map, and it always produces agent-shaped traces. Use it whenever you cannot prove that auto-instrumentation emits the shape the user needs.
2. OTEL auto-instrumentation is an optimization for libraries that Weave already recognizes. weave.init() alone captures them, with no per-call code. Use it only when the installed version of Weave auto-instruments the user's library and the shape it emits is the one they want.

Both paths produce OTEL spans. Agent-shaped spans land in Weave's Agents tab; flat call traces land in the Calls tab. The choice between the Session SDK and OTEL auto-instrumentation is driven by the shape you need and the structure of the code, not by framework identity. See references/decision.md for the full procedure.

Reference material (read the one you need)

Keep this file in context, and open a reference only when you reach the step that needs it. Each one is a precise, source-grounded cheat-sheet. Do not reconstruct the API from memory: the surface is new and easy to get subtly wrong.

• references/decision.md is the full decision procedure (shape, then structure, then probe, then strategy) and the coverage examples. Read it during "Choose a strategy" below.
• references/session_sdk_python.md covers the Python Session SDK: the classes, the context-manager pattern, the batch path, and the data types. Read it when instrumenting Python with the Session SDK.
• references/session_sdk_typescript.md covers the TypeScript Session SDK: the start* functions, the try/finally pattern, async init, and concurrency. Read it when instrumenting TypeScript or Node.
• references/otel_auto.md covers the OTEL auto path: what weave.init() captures on its own, the default-mode gotchas, and the framework caveats. Read it when OTEL auto is the chosen strategy.
• references/otel_endpoint.md covers exporting OTEL directly to Weave's endpoint (the raw-OTEL path): the environment-variable config, and how to add Weave's exporter to an app's own TracerProvider. Read it when the app already emits OTel spans or owns its OTel setup.

Workflow

Work through these steps in order. Steps 1 and 4 always run. Use your judgement for steps 2 and 3.

1. Establish the connection (always; confirm before editing widely)

A user's traces need three things:

1. the package installed;
2. credentials present;
3. a weave.init() call.

Settle these first. There is no point instrumenting code that cannot authenticate.

• Install or confirm the package. In Python, the package is weave on PyPI (pip install weave, or add it to their manifest). The Session SDK needs weave>=0.52.42. set_attributes and add_event need a newer build, so check the Python reference before relying on those. In Node, run npm install weave. The OTEL dependencies ship bundled in both, so the user adds nothing extra.
• Authentication is the user's job, not yours, so never handle their key. A W&B API key is a secret, and it must not enter your context or the repo. Tell the user to authenticate in their own terminal: set WANDB_API_KEY (from https://wandb.ai/authorize), or run wandb login (Python) or await weave.login(key) once (Node). You may check that a key is reachable (the env var is set, or ~/.netrc has an api.wandb.ai entry), but never read or print its value, and never write it into source code or into a committed .env.
• Get the project name. Ask for it as entity/project, or as just project to use their default entity. Show them the URL their traces will land at up front, https://wandb.ai/<entity>/<project>/weave, so there are no surprises about where the data goes.
• Add weave.init("entity/project") once at each real entry point: the main() function, the server startup, or the top of a script. Do not put it inside a hot loop. In Python: import weave; weave.init("entity/project"). In Node: await weave.init('entity/project'). It is async, so you must await it before any traced work. If the app installs its own global TracerProvider, weave.init() backs off and will not export to Weave; add Weave's exporter to their provider instead (see references/otel_endpoint.md). The calls are safe no-ops when init or auth is missing, so leaving instrumentation in place never breaks their program.

Before touching many files, state the plan in a sentence or two ("I'll use the Session SDK to wrap your agent loop in agent.py, and add weave.init in main.py") and let the user redirect you. Instrumenting edits their source, so a quick checkpoint beats a large surprise diff.

2. Survey the codebase

You cannot choose a strategy without knowing what they run. Find the following.

• The languages: Python, TypeScript/Node, or both.
• The LLM and agent libraries. Grep the imports and read the dependency manifest, and note the exact package. The names differ by language, so check the manifest that applies. Step 3 is where you resolve whether the installed Weave auto-instruments each one, and whether it emits agent shape or flat calls, so do not assume from the name.
  • Python (pyproject.toml / requirements.txt): openai, anthropic, langchain, openai-agents (import agents), claude_agent_sdk, google.adk, crewai, llama_index, and so on.
  • TypeScript (package.json): openai, @anthropic-ai/sdk, @google/genai, @openai/agents, @anthropic-ai/claude-agent-sdk, langchain, llamaindex, and so on.
• The agent's shape. Is there a hand-rolled loop (a while that calls the model, dispatches tools, and repeats)? Where is each model call? Where are tools dispatched? Is there delegation to sub-agents? These map directly onto Turn, LLM, Tool, and SubAgent.
• Existing OTel or telemetry. Look for any TracerProvider, opentelemetry setup, or a competing tracing vendor. This affects where you place init, and whether auto-capture already flows somewhere.

3. Choose a strategy, then instrument

Open references/decision.md for the full procedure. The short version is to decide by shape, probe for coverage, and fall back to the universal path.

1. What shape does the user want? An agent-shaped tree (the Agents tab), or flat model-call traces (the Calls tab)?
2. What is the agent's structure? The turn boundary, the model calls, tool dispatch, sub-agents, and any streaming or concurrency. This is what you will instrument, and it exists whatever the library is.
3. Probe auto-coverage. Does the installed Weave auto-instrument their library (read the registry: INTEGRATION_MODULE_MAPPING in Python, integrations/hooks.ts in TypeScript), and does it emit the shape from step 1? Do not trust a memorized framework list.
4. Choose. If auto-coverage emits the wanted shape, use OTEL auto (weave.init() alone; do not also hand-wrap, or you will double-log). Otherwise, use the Session SDK and map the structure from step 2. If the app already emits OTel spans or owns a TracerProvider, use raw OTEL export (references/otel_endpoint.md). A mixed approach is normal. Apply any caveats that references/otel_auto.md triggers (the gotcha lookup).

When you instrument with the Session SDK, the mapping is the heart of the work.

• Turn is one cycle of the top-level agent loop handling one user input, and it opens its own trace root. Wrap the loop body, not the whole program.
• LLM is one model API call. Wrap the call site, and record the input messages, the output, and the token usage when it is available.
• Tool is one tool or function execution. Wrap the dispatch, and record the name, the arguments, and the result.
• SubAgent is a delegated or nested agent invocation. Wrap the sub-call.

Map to the agent's real semantic boundaries. Not every helper function is a Tool, and a retry of one model call is still one LLM. Over-instrumenting produces noisy, misleading traces.

Never change behavior. This is instrumentation, so the program must do exactly what it did before. Preserve return values and exceptions. Keep spans closed on every path. Python context managers and TypeScript try/finally do this for you, so use them rather than bare start/end calls, which can leak an open span when an exception is raised. Do not reorder the user's logic or gate it on a span. If you cannot wrap cleanly without restructuring their code, prefer the batch path (Python log_turn or log_session; see the reference) or flag it to the user, rather than forcing an invasive rewrite.

4. Verify

Instrumentation you cannot see is worthless, so confirm that traces actually arrive.

• If the app is runnable and credentials are present, run a minimal path (their quickstart, a smoke test, or a tiny script that exercises one turn) and confirm a trace shows up at the project URL. Both the Python and Node weave.init() print View Weave data at https://wandb.ai/<entity>/<project>/weave at startup, so watch for that line. (The env-var OTEL export prints nothing, so open the project URL there.)
• The strongest check is the span shape, not the presence of code. Where you can run a smoke path, capture the emitted spans with an in-memory OTel exporter (no credentials needed) and confirm the operations are what you intended: a gen_ai.operation.name of invoke_agent for turns and sub-agents, chat for model calls, and execute_tool for tools, with the turn nesting its LLM and tool spans. This is also how you confirm that an auto strategy emits agent shape rather than flat 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 View Weave data at https://wandb.ai/.../weave line that weave.init() prints, and a trace in the Agents or Calls tab.
• If nothing shows up, the usual causes are auth not actually set in the run's environment, the Python plain-openai default-mode gotcha, init placed after the work it should wrap, or (in Node ESM) the missing --import=weave/instrument preload. The OTEL reference lists these.

Hygiene checklist

• The API key is never touched, read, printed, or committed.
• Behavior is unchanged: the same outputs, the same exceptions, and spans always closed.
• No double-instrumentation. Do not add Session SDK spans around calls a framework already auto-captures, or you will get duplicate traces.
• weave.init() appears once per entry point, ordered correctly relative to any user OTel setup.
• The version is pinned high enough for the APIs you used.
• The user knows the destination URL and how to verify.