| """Shared worker-agent harness for the GitSwarm Appendix C replication. |
| |
| Worker LLM: moonshotai/Kimi-K3 via HF Inference Providers (OpenAI-compatible). |
| Tool protocol: every assistant reply must be a single JSON object |
| {"tool": "<name>", "args": {...}}. The harness executes the tool and returns |
| the result as the next user message. |
| """ |
|
|
| import json |
| import os |
| import re |
| import time |
|
|
| from huggingface_hub import InferenceClient |
|
|
| MODEL_ID = os.environ.get("WORKER_MODEL", "moonshotai/Kimi-K3") |
| PROVIDER = os.environ.get("HF_PROVIDER") or None |
|
|
|
|
| def make_client(): |
| kwargs = {"model": MODEL_ID} |
| if PROVIDER: |
| kwargs["provider"] = PROVIDER |
| return InferenceClient(**kwargs) |
|
|
|
|
| TOOL_RE = re.compile(r"\{.*\}", re.DOTALL) |
|
|
|
|
| def parse_tool_call(text): |
| """Extract the first JSON object from a reply; return (tool, args) or None.""" |
| if text is None: |
| return None |
| |
| m = TOOL_RE.search(text) |
| if not m: |
| return None |
| try: |
| obj = json.loads(m.group(0)) |
| except json.JSONDecodeError: |
| return None |
| if not isinstance(obj, dict) or "tool" not in obj: |
| return None |
| args = obj.get("args", {}) |
| if not isinstance(args, dict): |
| args = {} |
| return obj["tool"], args |
|
|
|
|
| class Usage: |
| """Accumulates token usage across a run so we can report API cost.""" |
|
|
| def __init__(self): |
| self.calls = 0 |
| self.prompt_tokens = 0 |
| self.completion_tokens = 0 |
|
|
| def add(self, usage): |
| self.calls += 1 |
| if usage is not None: |
| self.prompt_tokens += getattr(usage, "prompt_tokens", 0) or 0 |
| self.completion_tokens += getattr(usage, "completion_tokens", 0) or 0 |
|
|
|
|
| def run_worker_episode(system_prompt, tools, max_turns=250, temperature=0.0, |
| usage=None, client=None, max_retries=5): |
| """Run one worker episode. tools: {name: fn(args) -> (result_str, done)}. |
| |
| Returns the name of the tool that ended the episode (or "max_turns"). |
| """ |
| if client is None: |
| client = make_client() |
| messages = [{"role": "system", "content": system_prompt}] |
| done_tool = None |
|
|
| for turn in range(max_turns): |
| reply = None |
| for attempt in range(max_retries): |
| try: |
| resp = client.chat.completions.create( |
| model=MODEL_ID, |
| messages=messages, |
| max_tokens=2048, |
| temperature=temperature, |
| ) |
| reply = resp.choices[0].message.content |
| if usage is not None: |
| usage.add(getattr(resp, "usage", None)) |
| break |
| except Exception as e: |
| if attempt == max_retries - 1: |
| raise |
| time.sleep(min(60, 5 * (2 ** attempt))) |
| call = parse_tool_call(reply) |
| if call is None: |
| messages.append({"role": "assistant", "content": reply or ""}) |
| messages.append({"role": "user", "content": ( |
| 'Your reply was not a valid tool call. Reply with exactly one ' |
| 'JSON object: {"tool": "<tool name>", "args": {...}}')}) |
| continue |
| name, args = call |
| messages.append({"role": "assistant", "content": json.dumps({"tool": name, "args": args})}) |
| if name not in tools: |
| messages.append({"role": "user", "content": ( |
| f"ERROR: unknown tool '{name}'. Valid tools: {sorted(tools)}.")}) |
| continue |
| try: |
| result, done = tools[name](args) |
| except Exception as e: |
| result, done = f"ERROR: tool '{name}' failed: {e}", False |
| if done: |
| done_tool = name |
| break |
| messages.append({"role": "user", "content": result}) |
|
|
| return done_tool or "max_turns" |
