fix(instrumentation-py): repair Claude Agent SDK, Cohere, and CrewAI OTel 2.x spans

[Nightingalelyy]

Summary

• split current Claude Agent SDK operation spans into common-only agent parents and connected canonical chat children, preserving sessions, usage/cache data, tool calls, and tool execution
• preserve structured Cohere rerank input/output with reversible reference-counted upstream helper patching
• normalize CrewAI tool definitions, calls, arguments, and tool-result identity into valid canonical JSON
• add the supporting respan-tracing exporter regression coverage and a patch release intent for all four changed packages

Validation

• Claude Agent SDK: 29 passed, 1 skipped

• Cohere: 12 passed

• CrewAI: 17 passed, 1 skipped

• respan-tracing exporter: 22 passed

• release inventory and release intent validation passed

• final example marker: otel2-fix-py-group-11-20260816T113500Z

• MCP audit: 13 complete traces / 39 full records; every tree and record inspected, with matching counts, resolved parents, successful statuses, exact token ownership, canonical tool schemas/calls/results, and no errors, duplicate IDs, or orphan spans

• Contribution re-audit against every current document under contribution/ passed after cef5a27: Cohere text-completion input is normalized to canonical llm.request.type=chat while retaining the text log type.

• Fresh correction validation: Claude Agent SDK 29 passed / 1 skipped, Cohere 12 passed, CrewAI 17 passed / 1 skipped; Cohere wheel, release inventory/intents, formatting/import, and diff checks passed; marker otel2-pr-contract-corrections-20260818T133300Z produced 1 connected Cohere trace / 2 full records, both inspected.

External follow-ups

• None documented.

Companion examples

• fix(examples-py): validate Claude Agent SDK, Cohere, and CrewAI OTel 2.x spans respan-example-projects#57

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.