Audit coverage

Run this after either of the other two passes, or on a repository someone else instrumented. The failure mode it hunts is silent: an uncaptured call does not error, it does not warn, it simply never appears in the dashboard. A missing route shows up as a gap you can see. A missing client shows up as nothing at all.

Both prompts are read-only. They change no code and produce a ranked list.

Caution

Absence of rows is not evidence of absence of spend. If your Metergraph total looks lower than your provider invoice, this audit is the first thing to run, not the last.

Python

Audit this Python repository's Metergraph instrumentation. Report only;
change nothing.

Find every path that reaches a model provider and say whether it is
captured:

  1. OpenAI(), AsyncOpenAI(), Anthropic(), AsyncAnthropic() or
     genai.Client() constructed but never passed through
     metergraph.wrap(), or wrapped with the result discarded.
  2. Raw HTTP calls with requests, httpx or aiohttp to api.openai.com,
     api.anthropic.com, generativelanguage.googleapis.com or a gateway
     URL.
  3. Calls made in a subprocess, Celery worker, notebook or cron
     entrypoint that never runs the setup path.
  4. Short-lived entrypoints that exit without metergraph.shutdown(), so
     queued telemetry may be lost and background work is not stopped.
  5. Work handed to a ThreadPoolExecutor that was not passed through
     metergraph.wrap_executor(), which loses the route, trace, session and
     tags.
  6. Any environment or deployment target where the SDK's app token is
     not set, since capture there is off and silent. The exact variable
     name and every other setting are at
     https://www.metergraph.dev/docs/reference/configuration/
  7. Calls inside a route() or trace() block that were left with
     content capture disabled when they did not need to be. What the
     controls do is at
     https://www.metergraph.dev/docs/guides/keep-content-private/

For each finding give the file, the line, and which of the seven it is.
Rank by how much traffic it likely represents. Do not fix anything yet.

TypeScript

Audit this TypeScript repository's Metergraph instrumentation. Report
only; change nothing.

Find every path that reaches a model provider and say whether it is
captured:

  1. new OpenAI(), new Anthropic() or new GoogleGenAI() constructed but
     never passed through mg.wrap(), or wrapped with the result
     discarded.
  2. Vercel AI SDK models used without
     mg.vercelAISDKMiddleware() in wrapLanguageModel.
  3. Functions calling a wrapped client without mg.track("name", fn).
     These will report unstable, bundler-generated names in production
     even though they look fine in development. Flag every one.
  4. Raw fetch calls to api.openai.com, api.anthropic.com,
     generativelanguage.googleapis.com or a gateway URL.
  5. Serverless handlers (Lambda, Workers, Vercel) that return without
     mg.wrapHandler, mg.bindWaitUntil or an awaited mg.flush(), so the
     queue is frozen or discarded before delivery. The rule per runtime
     is at https://www.metergraph.dev/docs/instrument/serverless/
  6. Any environment or deployment target where the SDK's app token is
     not set, since capture there is off and silent. The exact variable
     name and every other setting are at
     https://www.metergraph.dev/docs/reference/configuration/
  7. Edge or worker bundles where the SDK is imported but the setup path
     never runs.

For each finding give the file, the line, and which of the seven it is.
Rank by how much traffic it likely represents. Do not fix anything yet.

Reading the result

The seven categories are not equally urgent. In rough order of how much traffic they usually hide:

Finding	Why it matters
A deployment target with no token	Every call from that environment is invisible, and there is no error to notice
A client constructed but never wrapped	A whole call site, permanently missing
A raw HTTP call to a provider	The SDK cannot see it at all. Move it onto a client, or send rows yourself
A worker or subprocess with its own entrypoint	Often the largest single share of spend in a batch-heavy codebase
A short-lived process with no shutdown()	Rows are produced and then dropped on exit
Executor work with no context propagation	Rows arrive, but with no route, trace, session or tags
An unwrapped function in TypeScript	Rows arrive under a name a bundler invented, which changes on the next build

For a raw HTTP call you cannot route through a client, Write your own client covers sending rows to the ingest API directly. For the propagation cases, see Attribution across threads and Serverless and short-lived jobs.

Instrument a repository The full pass, if the audit found more missing than present.