[Alexander Kaplan (tokaplan)]

Motivation & Context

PR #9361 detached tool-call payload attributes from function-invocation metrics, but the metric snapshot still copies every base tool-span attribute. As a result, gen_ai.tool.description (which may contain sensitive information) and the per-invocation gen_ai.tool.call.id remain histogram dimensions.

This follow-up keeps span telemetry unchanged while preventing sensitive and high-cardinality span-only metadata from flowing into metrics.

Description & Review Guide

• What are the major changes? Build function-duration metric attributes from an explicit allowlist containing operation name, tool name, tool type, and the framework function name. Continue adding error.type only on failures.
• What is the impact of these changes? Tool descriptions and call IDs remain available on execute_tool spans but no longer become metric dimensions. Existing low-cardinality dimensions and failure classification remain unchanged.
• What do you want reviewers to focus on? Confirm that the allowlist preserves useful aggregation dimensions while preventing future span-only attributes from entering the histogram implicitly.

Related Issue

N/A - telemetry correctness and cardinality follow-up to #9361.

AI Assistance

• No material AI assistance was used.
• This is an AI-assisted contribution. I reviewed, understood, and verified all submitted content and accept responsibility for it.

AI assistance details: Copilot assisted with implementation, regression tests, and the PR description.

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• This PR links to an agreed issue with no competing open PR, or the Related Issue section documents a trivial-change or repository-automation exception.
• This is not a breaking change. If it is a breaking change, add the breaking change label (or add "[BREAKING]" to the title prefix, before or after the language prefix) — a workflow keeps the label and title prefix in sync automatically.

[Copilot]

🟢 Approval recommended

The focused change preserves aggregation dimensions and error classification, with regression coverage for excluded metadata.

No open issues 🎉

What changed in this PR

This follow-up to #9361 keeps span-only tool metadata out of Python function-duration metrics without changing span telemetry.

Changes:

• Explicitly allowlists four metric dimensions while preserving failure classification.
• Extends regression assertions to exclude tool descriptions and call IDs from metrics.

File	Description
python/​packages/​core/​tests/​core/​test_tools.py	Verifies metric exclusions and preserved span attributes.
python/​packages/​core/​agent_framework/​_tools.py	Replaces copied span attributes with an explicit metric allowlist.

_{Review effort: Balanced}

[Bicheng (Kenneth) (BichengWang)]

Reviewed a29302a7b1377ec8e5aa98be473722616980a106. The allowlist retains the four aggregation dimensions and keeps error.type on failures, while leaving call IDs and descriptions on spans. I found no correctness blocker in the diff.

Offline validation on Python 3.13.16 with OpenTelemetry SDK 1.45.1: all 151 tests in test_tools.py pass, and Ruff 0.17.0 lint/format checks pass for both changed files. The four affected regression cases fail against the base implementation because the call ID is still a metric attribute. I also exercised the actual SDK with two invocations having distinct call IDs and descriptions: the head produces one histogram series with count 2, and both spans retain their respective metadata. This was a focused review; I did not run the full core suite or type-checking matrix.

One nonblocking coverage suggestion: preserve that SDK aggregation case as a test using InMemoryMetricReader. The existing histogram mocks verify individual attributes, while a real reader also verifies that changing call IDs/descriptions does not split the metric into multiple time series.

[Eduard van Valkenburg (eavanvalkenburg)]

Thank you for identifying and implementing this telemetry follow-up. We opened #9592 as the maintained replacement; it uses the same explicit allowlist approach, adds the explicit dict[str, str] annotation required by the current Pyright check, strengthens the regression assertions around the complete metric attribute shape, and is rebased on the latest main. Closing this as a duplicate so review can continue there. Thanks again for the contribution.