Agent patterns

You're Claude Code, working in my repository.

Task
Add a minimal agent loop to this project on Meta Model API's Responses API: read_file, write_file and run_tests function tools, each call run here and sent back as function_call_output with previous_response_id, repeated until the model answers, with the agent-patterns guards: a turn cap, a token budget, and a stop when one call repeats three times.

Read this first
Meta's docs page "Agent patterns" as Markdown: https://dev.meta.ai/docs/cookbook/agent-patterns.md
It's the source of truth for request and response shapes. The same page for people: https://dev.meta.ai/docs/cookbook/agent-patterns

Setup
- Base URL: https://api.meta.ai/v1
- Auth: send Authorization: Bearer $MODEL_API_KEY, reading MODEL_API_KEY from the environment. Never hard-code, print or commit the key.
- Model: muse-spark-1.3-contributor (Contributor mode: Meta may use what you send to train future models; drop -contributor for Standard mode)
- Python: the OpenAI SDK (pip install openai) with base_url="https://api.meta.ai/v1" and api_key=os.environ["MODEL_API_KEY"] works, or plain requests.

Facts to get right
- POST /v1/responses with tools [{type: function, name, description, parameters}]; each function_call has a call_id and arguments as a JSON string.
- Send results as input [{type: function_call_output, call_id, output}] with previous_response_id. Instructions and tools don't carry over: resend them every call.
- parallel_tool_calls: false keeps it to one tool per turn, the basic loop's contract.
- Muse Spark often writes a line of narration beside its tool call.
- For long runs, have the model keep a STATE.md, and edit files with exact-match replacements rather than whole-file writes.

Starter code (Python)
import json


def agent(client, task, tools, run_tool, max_turns=15):
    """client: OpenAI on https://api.meta.ai/v1. run_tool(name, args) returns text, errors included."""
    opts = {"model": "muse-spark-1.3-contributor", "tools": tools, "parallel_tool_calls": False}
    r, seen = client.responses.create(input=task, **opts), []
    for _ in range(max_turns):
        calls = [i for i in r.output if i.type == "function_call"]
        if not calls:
            return r.output_text
        seen += [(c.name, c.arguments) for c in calls]
        if any(seen.count(key) >= 3 for key in seen):
            raise RuntimeError("stuck: the same call three times")
        outputs = [{"type": "function_call_output", "call_id": c.call_id,
                    "output": run_tool(c.name, json.loads(c.arguments))} for c in calls]
        r = client.responses.create(previous_response_id=r.id, input=outputs, **opts)
    raise RuntimeError("turn cap reached")

Done when
- The loop fixes a failing test in a sample repo and stops when the tests pass.
- A run that repeats the same call three times stops with a stuck message.
- The key comes from MODEL_API_KEY and never appears in code, logs or commits.
- API errors (4xx, 429) surface their message instead of failing silently.