# fake_api

> **Tool `tools/llm/fake_api.py`** (program, Models). A stand-in for the Claude API and the claude command, so the tests spend nothing.

Run it with `./gameframe run fake_api ARGS` or `python3 tools/llm/fake_api.py`. [Source on GitHub](https://github.com/tanepiper/game-frame/blob/main/tools/llm/fake_api.py).

## Its own help

This is the help text in the tool itself, as `./gameframe tool fake_api` prints it.

```text
A stand-in for the Claude API and the claude command, so the tests spend nothing.

Usage:
  fake_api.py serve PORTFILE     serve on a free port on 127.0.0.1 and write the port to
                                 PORTFILE; point GAMEFRAME_LLM_ENDPOINT at it
  fake_api.py claude ARGS...     act as `claude -p --output-format json`: print a JSON result

The server answers POST /v1/messages, and creates, reports and returns Message Batches
(every batch has ended at once). Under a JSON schema (output_config.format) the answer fills
the schema: each string is "fake_<property>", each enum its first value, each number 0 and each
boolean true. Otherwise it echoes
the start of the last user message. The claude mode answers "fake run: " and the start of the
prompt (read from stdin) and the number of CLAUDE* variables it was given, with a cost of 0.01. It writes FAKE_ANSWER, when set, to answer.txt in the current
folder (the tests of skillcheck.py use it).

What it does not do: model anything. It checks the shape of requests only as far as the
tools need, and it keeps batches in memory until it stops.
```