Add model observation inference attributes

[Pujarini Mohapatra (biswapm)]

Summary

Adds the privacy-safe OpenTelemetry attributes required by the Model Observation pipeline.

• Capture gen_ai.response.model when the provider returns the canonical served model.
• Capture gen_ai.usage.cache_read.input_tokens.
• Capture gen_ai.usage.cache_write.input_tokens.
• Normalize the inference gen_ai.operation.name value to lowercase OTel values, including chat.
• Capture model endpoint information as server.address and an optional numeric server.port.
• Support attributes supplied when the inference scope starts or recorded later when the response becomes available.
• Add focused tests covering the emitted attribute names and value types.

Validation

The repository’s targeted inference tests, build, formatting, and linting checks pass.

[github-actions]

Performance comparison

Threshold: regressions >15.0% on gating scenarios fail the build. Higher ops/s is better; positive Δ means the PR is slower.

Scenario	Gating	Baseline (ops/s)	Candidate (ops/s)	Δ %	Status
azure_monitor_log	yes	36,584.5	36,865.0	-0.76%	✅
azure_monitor_span	yes	275,406.2	270,892.6	+1.67%	✅
otel_log	no	43,293.8	42,235.1	+2.51%	✅
otel_span	no	81,096.4	81,486.3	-0.48%	✅

[Copilot]

🟡 Changes recommended

Non-chat operation names are normalized to invalid values without underscores.

1 open finding

• Use canonical operation names for non-CHAT span values · New

What changed in this PR

Adds model-observation attributes to A365 inference spans.

Changes:

• Records response-model and cache-token attributes at startup or later.
• Emits numeric endpoint ports.
• Normalizes inference operation names and adds focused tests.

File	Description
tests/​a365/​test_inference_scope.py	Tests new inference attributes.
src/​microsoft/​opentelemetry/​a365/​core/​inference_scope.py	Emits and updates model-observation attributes.
src/​microsoft/​opentelemetry/​a365/​core/​inference_call_details.py	Adds response and cache-token fields.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🔵 Needs a closer look

Multiword inference operations are normalized without required underscores, producing invalid OTel values.

1 open finding

• Use canonical operation names for non-CHAT span values

🧠 Review effort: Balanced

[Copilot]

🟡 Changes recommended

Span names still use legacy PascalCase operations instead of the newly canonicalized values.

1 open finding

• Use normalized operation value in GenAI span names · New

1 resolved since last review

• Use canonical operation names for non-CHAT span values

🧠 Review effort: Balanced

[Pujarini Mohapatra (biswapm)]

Fixed in bc8a5eb: \�ctivity_name\ now builds from the same normalized operation name used for \gen_ai.operation.name, so span names use canonical values (e.g. \ ext_completion gpt-4o). Added a span-name assertion to the canonical-operation-names test.

[Copilot]

🟡 Changes recommended

Normalized non-chat span names can cause incorrect agent-invocation baggage enrichment.

1 open finding

• Recognize canonical GenAI operation names to prevent span misclassification · New

1 resolved since last review

• Use normalized operation value in GenAI span names

🧠 Review effort: Balanced

[Copilot]

🔵 Needs a closer look

Normalized non-chat operation names can cause processor misclassification and incorrect endpoint baggage propagation.

1 open finding

• Recognize canonical GenAI operation names to prevent span misclassification

🧠 Review effort: Balanced

[Pujarini Mohapatra (biswapm)]

Fixed in bba0434: GEN_AI_PROCESSOR_OPERATION_NAMES now also includes the canonical lowercase operation-name spellings (chat, ext_completion, generate_content) alongside the legacy PascalCase values, so span-name-based inference (_operation_name_from_span_name) recognizes canonical inference span names like ext_completion gpt-4o instead of falling through to inherited gen_ai.operation.name baggage (e.g. invoke_agent). This prevents invoke_agent-only attributes (caller agent identity, server.address/server.port) from leaking onto inference spans.

Added a regression test ( est_canonical_inference_span_name_does_not_inherit_invoke_agent_baggage) covering chat/ ext_completion/generate_content span names with inherited invoke_agent baggage, verifying common identity attributes are still enriched while invoke_agent-only attributes are withheld. Full suite: 604 passed, 25 skipped.

[Copilot]

🟢 Approval recommended

The focused changes include regression coverage and have no identified blocking issues; tests were not executed during this review.

0 open findings

1 resolved since last review

• Recognize canonical GenAI operation names to prevent span misclassification

🧠 Review effort: Balanced

[Nikhil Navakiran (nikhilNava)]

Reviewed canonical operation normalization, span-name consistency, response/cache attributes, numeric endpoint ports, and regression coverage. No blocking findings for supported inputs.