fix(agentscope): improve data attached to spans

[starfolkai]

Summary

Align the AgentScope integration with the SDK integration instrumentation requirements and replace mock-heavy coverage with cassette-backed tests across the supported AgentScope 1.x and 2.x matrix.

Changes

Span shape and metadata

• Move model tool definitions, tool_choice, and structured-output configuration out of LLM span input and into metadata.
• Normalize provider names to canonical pricing slugs (openai, anthropic, google, dashscope, ollama, and trinity).
• Keep request metadata to an explicit configuration allowlist instead of forwarding arbitrary kwargs that may contain credentials or transport details.
• Materialize inline multimodal message content through the shared _normalize_chat_messages helper.
• Continue creating integration spans through the local start_span() helper so every span carries agentscope-auto instrumentation provenance.

Metrics and streaming

• Parse usage through the shared OpenAI-style usage helpers, including nested cached/reasoning token details when reported.
• Derive total tokens from prompt and completion tokens when AgentScope does not expose a total directly.
• Record time_to_first_token and keep streaming spans open until the returned iterator is consumed.
• Add a version-gated AgentScope 1.x patcher for OpenAIChatModel._parse_openai_stream_completion_response. AgentScope 1.x consumes OpenAI's terminal usage-only chunk without yielding it, so the patcher captures those metrics without changing the user-visible stream and passes them to the outer model span through a ContextVar.
• Leave AgentScope 2.x on its native path because it surfaces the terminal usage chunk itself.

Errors and lifecycle behavior

• Log exception instances directly with span.log(error=exc) while preserving provider exception propagation.
• Verify repeated setup calls do not add another wrapper layer, including compatibility with AgentScope 1.x's own decorated methods.

Tests and cassettes

• Replace three fake/mock streaming tests with real VCR-backed AgentScope coverage.
• Add per-version streaming cassettes and assertions for accumulated output, TTFT, prompt/completion/total tokens, and one LLM span per model call.
• Add per-version provider-error cassettes and verify the exception originates from OpenAI, preventing missing-cassette/VCR exceptions from falsely satisfying the propagation test.
• Tighten existing span assertions for canonical provider metadata, token metrics, tool placement, parent/child relationships, and instrumentation provenance.
• Extend the auto-instrument subprocess test to assert metadata.provider == "openai".

Test plan

• nox -s "test_agentscope(latest)" -- --vcr-record=none
• nox -s "test_agentscope(1.0.0)" -- --vcr-record=none
• Focused recording and playback for streaming and provider-error tests on both matrix versions
• nox -s pylint
• Targeted pre-commit hooks, including stale-cassette validation

Created by abhijeet

Slack thread