Choosing and calling coding models

Define the coding task first

Completion, single-file Q&A, repository changes, debugging, and long-running agents are different workloads. Define visible files, tools, languages, test commands, execution time, and network permission before selecting a model with the required context and tools.

Recommended workflow

1. Supply only the code and constraints needed for the task.
2. Ask the model to locate evidence before proposing a minimal patch.
3. Run formatters, static checks, and tests in isolation.
4. Feed failures back with bounded loops and token budgets.
5. Require human review for security, migrations, and irreversible operations.

API shape

Simple generation can use Chat Completions or Responses. File search, shell, and editing require structured tool calls with validated arguments. A model proposing a command does not grant execution authority; enforce path, timeout, and network boundaries in the tool layer.

Evaluation

Build a fixed set from real defects and track patch application, test pass rate, first-pass success, retries, latency, and total cost. Safe refusal and recognition of incomplete requirements are correct outcomes.

Common failures

Context pollution, test leakage, dependency mismatch, and blind trust in generated code are often more important than model choice. Treat diffs, tests, and runtime evidence—not the model's completion claim—as acceptance.