Task · SKT-0153.08

Source and model Bedrock application inference profiles per agent

Description

doc-0004 W7: source profile metric and log shapes from a real deployment (read with owner approval) or keep the field gated with a cantfind PENDING. Generation model stays provider plus model name.

Acceptance Criteria

Definition of Done

Implementation Plan

Root captures two temporary per-agent profiles with bounded calls, reads available telemetry, deletes and verifies absence; gate any uncaptured fields.

Implementation Notes

Owner 2026-10-20: the shape that matters is per agent (each agent’s own application inference profile), not per team; team stays a tag. Owner approved one AWS capture: create synthkit-capture-* application inference profiles, bounded Converse calls, read CloudWatch, metric stream and invocation logs, delete the profiles. Root only.

Two temporary profiles created serially, two successful Converse calls each (maxTokens 32), then both deleted; post-delete reads ResourceNotFound. No logging or stream configuration changed. Waiting for eventual CloudWatch metric evidence; uncaptured log/stream fields will remain gated.

Final Summary

Delivered per-agent application-profile mechanism in fbf50d825f09fc265f05fce3538d2e9b0a742a7b. Five capture-backed API bases use profile ModelId and independent per-period gauge statistics; public duplicate-identity red-green and idle-volume checks pass, legacy inventory unchanged. Raw profile OTLP, logs and collector enrichment remain gated under SK-130; offline projection is not live exporter proof. Regeneration, aggregate review with zero findings and full composed check including race passed at pushed tip 802deabd549fa00c0c473ccfbaa882a23891adbc.

References

View the source file on GitHub