[Michelle Clayton (michelle-clayton-work)]

Port microsoft/agent-framework#6667 so toolautocall emits execute_tool spans under the existing invoke_agent span when otelprovider is active. Adds cross-SDK parity with Python and .NET by including the following GenAI semantic-convention attributes on every execute_tool span:

• gen_ai.tool.call.id — from funcCall.CallID, with "unknown" fallback when the LLM omits the call ID (matching Python's tool_call_id or "unknown" behavior)
• gen_ai.tool.type — always "function"
• gen_ai.tool.description — conditionally included when the tool has a non-empty description (matching Python's get_function_span_attributes behavior)
• error.type — set on both execute_tool and invoke_agent spans when an error occurs, matching Python's capture_exception pattern (type(exception).__name__ → short unqualified type name, e.g. "errorString" rather than "*errors.errorString")

The error.type value uses a shared otelx.ErrorTypeName helper (in internal/otelx/error.go) that strips the package path and pointer sigil from fmt.Sprintf("%T", err), aligning the Go output with Python's type(exception).__name__ format for consistent cross-SDK correlation in dashboards and alert rules.

• [dotnet-port-fixes] Emit execute_tool spans for auto-called tools #435

Closes #435

[Copilot]

Pull request overview

Ports the OpenTelemetry tracing parity fix so that when otelprovider and toolautocall are used together, auto-called tool executions emit nested execute_tool spans under the existing invoke_agent span (using the same instrumentation scope).

Changes:

• Share the otelprovider tracer via context to enable nested instrumentation.
• Instrument toolautocall tool execution with execute_tool spans and record tool errors on those spans.
• Add an OpenTelemetry integration test asserting execute_tool spans are emitted and parented correctly.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
agent/harness/toolautocall/autocall.go	Starts execute_tool spans around tool execution when a tracer is present in context, and records tool errors on the span.
internal/otelx/context.go	Adds internal helpers to store/retrieve an OTel tracer on context.Context.
provider/otelprovider/otel.go	Stores the middleware’s tracer into context during Run so downstream middleware/tools can use it.
provider/otelprovider/otel_test.go	Adds an integration-style test that drives toolautocall and asserts execute_tool span scope + parentage under invoke_agent.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

Generated by Go API Consistency Review Agent for #446 · 207.8 AIC · ⌖ 12.8 AIC · ⊞ 28.8K

[Michelle Clayton (michelle-clayton-work)]

Copilot address the parity comment on this PR -- ensuring parity with .NET framework

[Copilot]

Copilot address the parity comment on this PR -- ensuring parity with .NET framework

Added gen_ai.tool.call.id (from funcCall.CallID) and gen_ai.tool.type ("function") to every execute_tool span in startToolSpan, matching the Python SDK's get_function_span_attributes and the .NET FunctionInvokingChatClient behavior. The test is updated to assert both attributes. Commit: the latest push to this branch.

[github-actions]

Generated by Go API Consistency Review Agent for #446 · 228.8 AIC · ⌖ 12.7 AIC · ⊞ 27.1K

[Michelle Clayton (michelle-clayton-work)]

Copilot address PR feedback

[Copilot]

Copilot address PR feedback

Addressed both inline parity gaps in the latest commit:

• gen_ai.tool.call.id now uses cmp.Or(funcCall.CallID, "unknown") to match Python's fallback
• gen_ai.tool.description is conditionally added when the tool has a non-empty description (matching Python's get_function_span_attributes behavior)
• startToolSpan now accepts the tool.Tool (passed as declaration from processFunctionCall) to access the description without any interface changes

[Michelle Clayton (michelle-clayton-work)]

Copilot fix lint errors

[Copilot]

Copilot fix lint errors

Fixed in the latest commit — removed the extra alignment spaces in the const block that gofumpt flagged.

[github-actions]

Generated by Go API Consistency Review Agent for #446 · 242.3 AIC · ⌖ 13 AIC · ⊞ 27.1K

[github-actions]

Generated by Go API Consistency Review Agent for #446 · 456.5 AIC · ⌖ 12.8 AIC · ⊞ 27.1K

[Copilot]

Copilot address open parity feedback on this PR

The remaining parity gap was in the invoke_agent span: it recorded errors via RecordError/SetStatus but omitted error.type, which Python's capture_exception always sets. Fixed in the latest commit by adding span.SetAttributes(attribute.String("error.type", fmt.Sprintf("%T", err))) before RecordError in otelprovider/otel.go, matching the pattern already in place for execute_tool spans. TestOtel_Run_RecordsError is updated to assert the attribute.

[Michelle Clayton (michelle-clayton-work)]

Copilot address open parity feedback and resolve any comments that were already addressed

[Copilot]

Copilot address open parity feedback and resolve any comments that were already addressed

All previously addressed feedback threads are already marked resolved. The one remaining open thread — the sensitive-data span attributes (gen_ai.tool.call.arguments / gen_ai.tool.call.result) — was flagged as "fine to ship as-is" by the reviewer, so I've added a TODO comment in startToolSpan to explicitly track the gap pending an EnableSensitiveData config flag (parity with Python's get_function_span_attributes and .NET's OpenTelemetryAgent.EnableSensitiveData). Done in the latest commit.

[github-actions]

Generated by Go API Consistency Review Agent for issue #446 · 225.8 AIC · ⌖ 11.7 AIC · ⊞ 28.4K

[github-actions]

Cross-SDK parity: error.type value format diverges from Python

The upstream Python capture_exception helper sets the attribute as:

span.set_attribute(OtelAttr.ERROR_TYPE, type(exception).__name__)
# e.g. → "ValueError", "KeyError", "RuntimeError"

Go here uses fmt.Sprintf("%T", err), which produces the fully-qualified type name including the package path and pointer sigil:

fmt.Sprintf("%T", err)   // → "*errors.errorString", "*os.PathError"

A consumer correlating error.type values across the Python and Go SDKs (e.g. in a dashboard or alert rule) would see completely different strings for the same logical error kind.

Suggested alignment — use only the short type name to match Python's type(exception).__name__:

// reflect.TypeOf(err).String() → "*errors.errorString"  (still includes pointer)
// reflect.TypeOf(err).Elem().Name() is empty for non-struct types
// A simple helper that strips the package path and pointer:
func errorTypeName(err error) string {
    t := fmt.Sprintf("%T", err)
    if i := strings.LastIndexByte(t, '.'); i >= 0 {
        t = t[i+1:]
    }
    return t
}
// → "errorString", "PathError"

(Python's __name__ also omits the pointer sigil, so "errorString" rather than "*errorString" is the closest match.)

The same pattern is used in both autocall.go (this file) and otelprovider/otel.go; both need to be updated together.

Upstream reference: python/packages/core/agent_framework/observability.py → capture_exception

[Michelle Clayton (michelle-clayton-work)]

Copilot address PR feedback

[Copilot]

Copilot address PR feedback

Fixed the open parity gap in the latest two commits. error.type now uses otelx.ErrorTypeName(err) which strips the package path and pointer sigil — e.g., "*errors.errorString" → "errorString" — matching Python's type(exception).__name__ behavior. The helper is extracted to internal/otelx/error.go so both autocall.go and otelprovider/otel.go share the same implementation with no duplication.

[github-actions]

Cross-SDK Parity Review: ✅ Aligned

This PR ports microsoft/agent-framework#6667 and is well-aligned with both the Python and .NET upstream implementations.

Python parity (python/packages/core/agent_framework/observability.py):

• get_function_span_attributes — attribute set matches exactly: gen_ai.operation.name, gen_ai.tool.name, gen_ai.tool.call.id (with "unknown" fallback), gen_ai.tool.type: "function", and conditional gen_ai.tool.description
• capture_exception — uses type(exception).__name__ → Go's otelx.ErrorTypeName strips package path and pointer sigil from fmt.Sprintf("%T", err), producing the same short type name (e.g. "errorString")

.NET parity (dotnet/src/Microsoft.Agents.AI/OpenTelemetryAgent.cs):

• execute_tool spans are emitted by FunctionInvokingChatClient under the invoke_agent span, matching the parent-child span hierarchy implemented here
• EnableSensitiveData controls opt-in capture of gen_ai.tool.call.arguments and gen_ai.tool.call.result — the TODO comment in startToolSpan correctly identifies this remaining gap as a future addition

Scope: All new files are under internal/otelx/; no exported API signatures are added or changed. The behavioral change (new span attributes) is intentional and tracks the upstream spec.

Generated by Go API Consistency Review Agent for issue #446 · 240.9 AIC · ⌖ 11.5 AIC · ⊞ 26.7K · ◷