"""Example: Using ApplyPatch tool with GPT-5.1 models via direct OpenAI API.
This demonstrates adding the ApplyPatch tool to the agent and guiding the
model through a richer sequence of file operations using 'apply_patch' text:
- Create multiple files (FACTS.txt, NOTES.md)
- Apply multi-hunk edits to a single file
- Apply a single patch that touches multiple files
- Mix add / update / delete operations in one patch
Notes:
- Works with any GPT-5.1 family model (names start with "gpt-5.1").
- Uses direct OpenAI API through LiteLLM's LLM wrapper with no base_url.
- Requires OPENAI_API_KEY in the environment (or LLM_API_KEY fallback).
"""
from __future__ import annotations
import os
from pydantic import SecretStr
from openhands.sdk import LLM, Agent, Conversation, get_logger
from openhands.sdk.tool import Tool
from openhands.tools.apply_patch import ApplyPatchTool
from openhands.tools.task_tracker import TaskTrackerTool
# from openhands.tools.preset.default import register_default_tools
from openhands.tools.terminal import TerminalTool
logger = get_logger(__name__)
api_key = os.getenv("OPENAI_API_KEY") or os.getenv("LLM_API_KEY")
assert api_key, "Set OPENAI_API_KEY (or LLM_API_KEY) in your environment."
# Choose a GPT-5.1 model; mini is cost-effective for examples
default_model = "openai/gpt-5.1-codex-mini"
model = os.getenv("LLM_MODEL", default_model)
assert model.startswith("openai/gpt-5.1"), "Model must be an openai gpt-5.1 variant"
llm = LLM(
model=model,
api_key=SecretStr(api_key),
reasoning_summary=None, # avoid OpenAI org verification requirement
log_completions=True, # enable telemetry to log input/output payloads
)
# Explicitly register tool classes so Tool(name=...) can resolve
# They self-register into the global registry on import
_ = (TerminalTool, TaskTrackerTool, ApplyPatchTool)
agent = Agent(
llm=llm,
tools=[
Tool(name="terminal"),
Tool(name="task_tracker"),
Tool(name="apply_patch"),
],
system_prompt_kwargs={"cli_mode": True},
)
conversation = Conversation(agent=agent, workspace=os.getcwd())
# Compose instructions guiding the model to exercise richer ApplyPatch behavior.
prompt = (
"You have access to an apply_patch tool that edits files using unified patches. "
"Use it to perform the following sequence of operations in as few patches as "
"reasonable, while keeping each patch valid and focused:\n\n"
"1) Create two files:\n"
" - FACTS.txt containing exactly two lines:\n"
" OpenHands SDK integrates tools.\n"
" ApplyPatch can edit multiple files.\n"
" - NOTES.md containing exactly three lines:\n"
" # Notes\n"
" - Initial point A\n"
" - Initial point B\n\n"
"2) Apply a multi-hunk update to NOTES.md in a single *** Update File block:\n"
" - Change the text 'Initial point A' to 'Updated point A'.\n"
" - Append a new bullet '- Added via multi-hunk patch.' after "
"'Initial point B'.\n\n"
"3) Apply a single patch that updates BOTH FACTS.txt and NOTES.md at once:\n"
" - In FACTS.txt, append a third line: 'Multi-file patches are supported.'.\n"
" - In NOTES.md, append a final line: 'Summary: multi-file patch applied.'.\n\n"
"4) Finally, use one more patch that mixes operations across files:\n"
" - Add a TEMP.txt file containing a single line: 'Temporary file'.\n"
" - Append a line 'Cleanup step ran.' to FACTS.txt.\n"
" - Delete TEMP.txt.\n\n"
"Important rules:\n"
"- Only call the tool using the apply_patch text format between "
"'*** Begin Patch' and '*** End Patch'.\n"
"- Use '*** Add File', '*** Update File', and '*** Delete File' sections as "
"described in the GPT-5.1 apply_patch guide.\n"
"- When updating a file, include enough context lines so the patch can be "
"applied even if whitespace varies slightly.\n"
)
conversation.send_message(prompt)
conversation.run()
print("Conversation finished.")
print(f"EXAMPLE_COST: {llm.metrics.accumulated_cost}")
