# Current Configuration Agent9 Retry Latency - Optimization ID: `current-config-agent9-retry-latency-2026-08-27` - Owner: Evidence Agent configuration-plan translation and same-turn repair - Source: TCID-13091 read replay `6e4c6f2e-aa70-479e-8b36-26947ba6eeaf` - Implementation status: deployed in Auto-oncall SHA `2725d5270ae239bf2db016ae67aaa84678962f53` and Managed Clinic v19 - Effective status: `optimized_verified` ## Bounded problem The deployed replay correctly reduced evidence acquisition to one logical configuration query with four selectors and returned 64, 64, 32, and 32. The Evidence Agent nevertheless invoked `managed_auto_oncall_clinic` three times. The first two single-query calls failed with `configuration_key_invalid`; only the third reached Provider transport and completed through Clinic 403 followed by CloudConfig fallback. The scenario stage took 165.1 seconds and the complete read replay took 249.5 seconds. Clinic batch execution was not serial across four parameters. All four selectors were returned by one successful configuration operation. The remaining latency is the cost of two avoidable Agent9 Tool turns plus the fixed Router, Scenario Diagnostic Agent, reconciliation, and Final Response turns. ## Implemented repair - Managed Clinic v19 exposes only `issueKey`, `orgId`, and `queries` to the Evidence Agent. The single-operation path remains implementation-only for existing deterministic callers. - One configuration objective carries all requested selector objects in one `configKeys` array. The Tool schema and Evidence Agent prompt include the exact `{component, key}` shape and prohibit string selectors or per-key Tool calls. - Clinic-to-CloudConfig fallback and both credentials remain inside the Managed Tool. The Evidence Agent cannot select CloudConfig as another operation. - A configuration selector validation error permits at most one changed correction rather than the generic four-attempt query repair budget. Local verification passed the complete 1,001-test suite. The configuration contract test executes one batch entry with all four TCID-13091 selectors, observes Clinic 403, and receives the complete CloudConfig fallback result. The post-deploy read replay `86c7bfc5-8959-422b-affa-c120ddcc8a16` verified the same path against the live services. The Evidence Agent made one Skill read and one `managed_auto_oncall_clinic` call. That call contained one configuration query with four selectors; no `configuration_key_invalid` result or correction call occurred. Managed Clinic v19 recorded Clinic HTTP 403 followed by a complete helper-owned CloudConfig fallback and returned 64, 64, 32, and 32. The Provider-facing Evidence Agent turn fell from 110.8 seconds to 70.2 seconds (-40.6 seconds, -36.6%). The complete scenario stage fell from 165.0 seconds to 147.8 seconds (-17.2 seconds). This verifies the bounded validation-retry optimization. ## Unresolved gaps - The full diagnostic pipeline remains too slow for a simple factual lookup. Its Agent-runtime span improved only from 241.9 seconds to 238.1 seconds, and observed HTTP wall time varied from 249.5 seconds to 256.7 seconds. The Scenario Diagnostic Agent and Router reconciliation were 23.4 and 11.2 seconds slower in the verification replay, respectively, offsetting most of the query-turn gain. - The raw invalid selector values from the pre-change replay remain redacted in durable Agent9 operation history. Operation count and the stable validation error code are observable and sufficient for this bounded criterion. ## Next validation action Treat total diagnostic-pipeline latency as a separate problem with its own earliest-owner evidence. Do not repeat the configuration batching repair: the live replay already proves one valid helper call and substantially lower Provider-facing Evidence Agent latency.