# Diagnostic Agent Role Unification - Optimization ID: `diagnostic-agent-role-unification-2026-08-10` - Owner: auto-oncall-services diagnostic runner - Source reference: operator architecture review on 2026-08-10 - Implementation status: merged and deployed to staging - Effective status: `optimized_unverified` ## Bounded problem Runtime terminology and ownership had diverged. Wetech used a plan-only agent followed by runner execution and a separate diagnostic turn, while Clinic and Manus still had a scenario agent orchestrate provider evidence inside its own flow. That made provider choice alter agent responsibilities and encouraged incorrect reasoning about which stage owned diagnosis. This record owns only diagnostic role and provider-orchestration consistency. It does not change provider authorization, target selection, supported query operations, causal admission rules, Jira publication, or Focus lifecycle. ## Candidate change - Standardize the runtime roles as Scenario Router, Evidence Agent, runner-owned Provider Executor, Scenario Diagnostic Agent, and Final Synthesis Agent. - Apply the same sequence to Clinic, Wetech, and Manus. - Keep provider differences limited to exact target/auth contract, query plan and result schema, supported operations, helper, and serial or concurrent execution. - Scope OP serialization to runner-owned readiness and Provider Executor calls; do not serialize Evidence Agent planning or Scenario Diagnostic Agent reasoning merely because the provider transport is stateful. - Reuse one Evidence Agent thread for refinements within a scenario and start a fresh offline Scenario Diagnostic Agent after each materialized round. - Retain old `child-evidence-handoff` wire identifiers only for compatibility, not as agent-role names. ## Verification evidence Focused TypeScript tests cover all three providers, plan-only Evidence Agent behavior, runner-owned execution, separate Scenario Diagnostic Agent turns, same-thread evidence refinement, fresh diagnostic turns, and final tool-free synthesis. The complete suite passed with 709 tests on the release branch. - Source PR: `tidbcloud/aisupport#732`, merged as `9763e77e`. - Staging image PR: `tidbcloud/app-delivery#76716`. The exact release image was deployed and was later superseded by descendant image `2ff0796c`, which contains `9763e77e`; the live Pod became Ready at `2026-08-10T13:31:37Z`. - Wetech replay `APID-29232`, run `2e6f9473-1b48-460e-b513-2250fa075597`: the trace used one Router, one Evidence Agent thread across three rounds, a runner-owned Wetech auth and batch executor, a fresh Scenario Diagnostic Agent per round, and tool-free final synthesis. Evidence acquisition completed with provider capability verified. The scenario converged by ruling out material Region/Leader distribution imbalance as the cause of the CPU spread. - Clinic replay `ST-2865`, run `571f090c-5030-4229-b443-7a3004696208`: Router and plan-only Evidence Agent ownership matched the new flow, but runner validation rejected the generated 12-query plan with `clinic_evidence_plan_query_invalid` before Provider Executor output was materialized. The final result correctly remained `human_required`, but Clinic runtime equivalence was not verified. - Both replays used `mode=read`. Neither run ID exists in `auto_oncall_runs`; Jira, database, evaluation, and Feishu publication were not requested. ## Prior-change audit - PR #721 (`a9bab52c`) correctly separated Wetech planning from scenario diagnosis and reused its evidence thread, but used inconsistent acquisition/parent terminology and did not apply the same runner-owned orchestration to Clinic and Manus. This is the concrete earlier gap corrected here. - PR #723 (`5ac06093`) scoped browser temporary files to the current run. That filesystem isolation remains valid independently of agent-role naming. - PR #724 (`4bdff00b`) moved Wetech auth/readiness to deterministic runner code, and PR #725 (`3995421b`) moved Wetech batch execution outside the agent sandbox. Both changes match the now-explicit Provider Executor boundary. - PR #729 (`d0427d09`) fixed canonical Wetech target alias validation. It is a target handoff fix and does not depend on which agent owns diagnosis. The audit found no evidence that the incorrect verbal role description made the auth, temporary-file, provider-execution, or alias fixes invalid. The material design defect was provider-specific orchestration and ambiguous role names, not those operational boundaries. ## Unresolved gaps - Wetech runtime ownership is verified by one staging replay, but Clinic is blocked by a generated field-level plan-contract mismatch (`kind`/`query` and operation parameters); its query IDs were valid. Manus has not been replayed. Equivalent behavior across all three providers is therefore not verified. - The Wetech case needed three evidence rounds. Rounds reused the same Evidence Agent thread as designed, but cumulative Evidence Agent input grew from 80,499 to 136,129 to 216,754 tokens. This is observed cost evidence, not part of the role-unification acceptance result. - No Jira internal comment has been refreshed by this change. ## Next validation action Resolve the Clinic field-level plan-contract mismatch as the separate bounded provider-plan-contract recovery problem, then replay one Clinic and one Manus case. Keep this record `optimized_unverified` until those provider paths pass and any applicable Jira internal-comment refresh is separately authorized and verified.