LLM API for code review

Code review in a production application

A pull request changes authentication middleware. The review service sends the diff, nearby functions, repository security rules, and the changed-line map to codex-auto-review. The model must return only findings that identify a concrete failure path.

The service rejects comments without a file, changed line, severity, evidence, and repair. Clean pull requests are part of the evaluation set, so producing more comments is not treated as better review.

How the code review workflow operates

• Build the diff and changed-line map.
• Retrieve only directly relevant code and repository rules.
• Request structured findings with evidence and a minimal line range.
• Drop malformed and out-of-diff comments unless their dependency path is proven.
• Show accepted findings to a human reviewer.

Call codex-auto-review through LLMFly AI

Send the request from your server. Replace the example content and placeholder tool schema with data and tools from your application.

curl https://app.llmfly.ai/v1/chat/completions \
  -H "Authorization: Bearer $LLMFLY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "codex-auto-review",
    "messages": [
      {"role": "system", "content": "Review the supplied diff. Report only actionable defects introduced by this change. Each finding needs file, changed line, severity, failure path, evidence, and a minimal fix."},
      {"role": "user", "content": "Review this authentication middleware pull request against the attached repository rules."}
    ],
    "response_format": {"type": "json_object"}
  }'

Why codex-auto-review is the primary model

codex-auto-review is the purpose-specific first route currently exposed by Model Plaza for automated review.

Use Terra when the review needs broader repository reasoning; escalate a disputed high-risk finding to Sol, not every ordinary PR.

Acceptance checks for code review

Failures to handle before deployment

• Rewarding comment volume
• Using only defective examples
• Accepting vague best-practice advice
• Commenting outside changed lines without evidence

Returned output and run records

JSON findings: file, line, severity, title, evidence, failure path, and suggested fix.

Record the model ID, request ID, token usage, retries, validation result, and final disposition for every production run.