LLM API for code generation

Code generation in a production application

A product team receives a small feature ticket: add CSV export to an existing reporting page without changing its public API. The model gets the ticket, repository conventions, the affected files, and the test command. It returns a unified diff and the tests that prove the feature works.

LLMFly AI is used only for generation. The application still checks the patch in a disposable worktree, runs lint and tests, and sends failures back as a new message. Nothing is merged from the first response.

How the code generation workflow operates

• Retrieve the ticket, conventions, affected files, and test command.
• Ask for a unified diff plus new or changed tests.
• Apply the patch only inside a disposable worktree.
• Run deterministic validation; return failures for one bounded repair attempt.
• Send the passing patch to human review.

Call gpt-5.6-terra 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": "gpt-5.6-terra",
    "messages": [
      {"role": "system", "content": "You are editing an existing repository. Return only a unified diff. Preserve public APIs, follow the supplied conventions, and include tests for every acceptance criterion."},
      {"role": "user", "content": "Add CSV export to the report page. Files and acceptance criteria follow below."}
    ]
  }'

Why gpt-5.6-terra is the primary model

Terra is the default because this workflow needs repository reasoning and usable patches, but not Sol-level effort on every ticket.

Use gpt-5.4 for smaller isolated edits; escalate to Sol when the change crosses architecture boundaries or Terra fails the same acceptance test twice.

Acceptance checks for code generation

Failures to handle before deployment

• Testing only toy prompts
• Accepting code that compiles but misses requirements
• Giving candidates unequal context
• Ignoring review and repair cost

Returned output and run records

A unified diff that applies cleanly, plus tests covering every acceptance criterion.

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