> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tensor9.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Debugging with Coding Agents

> Give your existing coding agent a bounded, reproducible adapter investigation.

You can ask a coding agent, such as Codex or Claude, to organize request evidence, compare it with the service reference and reduce a reproduction. Use the same public interfaces you would use yourself. Access to Tensor9's source tree is not a prerequisite for this workflow.

## Prepare an evidence bundle

Provide the product and SDK versions, origin service and operation, selected backend, redacted configuration, relevant tuning tags, error and response headers. Include the resource state or setup needed to reproduce the issue.

Remove authorization headers, session tokens, signed URLs, secret payloads and unrelated application data. Replacing a secret with a placeholder is preferable to asking an agent not to print a secret it already received.

## Set the execution limits

Use a task like this:

```text theme={null}
Investigate this adapter request using the attached sanitized evidence.

First identify the installed Tensor9 CLI version and inspect its public help.
Compare the operation and selected backend with the service reference.
Separate observations from hypotheses and list any missing state.

Do not access production credentials, send requests to production endpoints,
create or delete cloud resources, change configuration, or retry the captured
request. Use only the isolated test resources I explicitly name.

Use tensor9 explain with the documented isolation and side-effect controls.
Do not replace an explanation with a live request replay.

Report the commands used, observations, remaining uncertainty and the smallest
safe next test. Ask before performing any action outside these limits.
```

For an isolated test, name the exact resources and credentials the agent may use and what cleanup it may perform. “Debug the adapter” is not sufficient authorization to modify a live backend.

## Use explain

Have the agent invoke the Tensor9 CLI rather than recreate adapter logic from its own assumptions. It must preserve relevant resource state and distinguish a profile explanation from an observed request result. See [understanding behavior with explain](/cloud-adapter/debugging/explain).

## Review the result

Require evidence for the conclusion: the exact operation, configuration and observed outcome. A generated explanation is not proof of compatibility. Check that any proposed fix respects the service's supported fields and the [configuration migration rules](/cloud-adapter/operations/overview).