feat: add openai responses provider

[mapleluvr]

Summary

• Add an OpenAI Responses provider type with template exposure, runtime request mapping, response parsing, streaming support, and provider path tests.
• Preserve the existing OpenAI chat-completions provider path with regression coverage.
• Include the current root master checkpoint/subagent runtime updates already present locally before this merge.

Test Plan

• uv run pytest tests/test_openai_source.py -q -k "openai_responses or responses"
• uv run pytest tests/test_dashboard.py::test_provider_templates_include_openai_responses_type -q
• node --test tests/providerUtils.test.mjs from dashboard/
• uv run ruff check astrbot/core/provider/manager.py astrbot/core/provider/sources/openai_source.py tests/test_openai_source.py

[sourcery-ai]

Sorry @mapleluvr, your pull request is larger than the review limit of 150000 diff characters

[gemini-code-assist]

Code Review

This pull request introduces a persistent subagent runtime and an agent group collaboration system, enabling long-lived conversation contexts and multi-agent workflows. Key additions include a dedicated skill tool for runtime-aware instruction loading, management tools for agent groups, and support for the OpenAI Responses API. Feedback focuses on performance and robustness: specifically, implementing chunked reading for file hashing to prevent OOM errors, adding null checks in API response parsing, replacing expensive deep copies with shallow copies in payload preparation, and improving the efficiency of the token-based history pruning logic to avoid quadratic complexity.

[gemini-code-assist]

Reading the entire file into memory with read_bytes() to calculate a hash can lead to high memory consumption or Out-Of-Memory (OOM) errors if the workspace contains large files. It is safer to read the file in chunks. Additionally, ensure that this functionality is accompanied by corresponding unit tests.

        sha256_hash = hashlib.sha256()
        with path.open("rb") as f:
            for byte_block in iter(lambda: f.read(8192), b""):
                sha256_hash.update(byte_block)
        digest = sha256_hash.hexdigest()

References

1. New functionality, such as handling attachments, should be accompanied by corresponding unit tests.

[gemini-code-assist]

The _get_response_field method does not check if item is None before calling getattr(). This will cause an AttributeError if the input item is None, which is a likely scenario when parsing nested optional fields from API responses (e.g., usage.input_tokens_details might be missing in some responses).

Suggested change

	def _get_response_field(item: Any, field: str, default: Any = None) -> Any:
	if isinstance(item, dict):
	return item.get(field, default)
	return getattr(item, field, default)
	@staticmethod
	def _get_response_field(item: Any, field: str, default: Any = None) -> Any:
	if item is None:
	return default
	if isinstance(item, dict):
	return item.get(field, default)
	return getattr(item, field, default)

[gemini-code-assist]

Using copy.deepcopy(payloads) on every request can be very expensive in terms of CPU and memory, especially when the payload contains large base64-encoded images for vision tasks. Since _prepare_chat_payload typically returns a fresh dictionary, a shallow copy (payloads.copy()) should be sufficient and much more efficient for the subsequent pop operations. Additionally, ensure that this new attachment-related functionality is accompanied by corresponding unit tests.

Suggested change

	payloads = copy.deepcopy(payloads)
	payloads = payloads.copy()

References

1. New functionality, such as handling attachments, should be accompanied by corresponding unit tests.

[gemini-code-assist]

This loop performs token counting on the entire message history in every iteration while pruning. This results in $O(N^2)$ complexity relative to the number of messages. While typically acceptable for short conversations, it can become a performance bottleneck if max_persisted_turns is configured to a large value or if messages are very long. Consider estimating the number of tokens to remove or using a more efficient pruning strategy.