weave-add-tracing

Add W&B Weave Tracing

Instrument an existing AI application, run a representative task, and verify that the resulting Weave trace explains the application's behavior. Preserve the application design unless the user explicitly requests a refactor.

Operating Rules

• Verify current SDK and integration behavior with official W&B Weave documentation before relying on exact APIs or auto-patching claims.
• Add observability without changing prompts, control flow, retries, tool behavior, or user-visible output.
• Initialize Weave once at the application boundary. Reuse an existing valid weave.init() call.
• Prefer a small trace tree of meaningful workflow, model, retrieval, and tool boundaries over decorating every helper.
• Do not duplicate calls already captured by a supported provider or framework integration.
• Trace only serializable inputs and outputs. Summarize or reference large objects and documents.
• Never send credentials, access tokens, private notes, unpublished documents, or unnecessary personal data to Weave.
• Do not introduce weave.Model, datasets, assets, scorers, or evaluations unless the user separately requests them.

Workflow

1. Preflight

1. Detect the project language, package manager, entrypoint, and normal test or run command.
2. Search for existing Weave setup and classify the task:
   • Fresh: no Weave integration exists.
   • Evolve: weave.init(), @weave.op, or framework tracing already exists.
3. Identify model providers, agent frameworks, tools, retrieval, callbacks, retries, and async boundaries.
4. Confirm the target Weave project from existing configuration or team/project. Do not put credentials in source code.
5. Read references/codebase-analysis.md for a repository-wide integration scan.

2. Build a Compact Trace Map

Present the proposed root and child Calls before editing. Include file, symbol, reason, and capture behavior.

run_agent()                         root workflow Op
|- provider model call             automatic integration
|- retrieve_context()              custom Op
`- execute_tool()                  custom Op