# Diagnostic visual context and scoped execution efficiency - Owner: AOS Scenario Router, Scenario Diagnostic, and runner metadata/context transport. - Source baseline: `0affba36f26792f732f7bc7a48b4a3a3e94587cd`. - Implementation: PR #1416 deployed as `02bdc964` on 2026-09-09 (+08). Read-only verification is recorded below; no Jira publication or automated optimization-status promotion is claimed. The Router uses existing window/symptom/scenario fields for visible facts. A small image-request result lets the runner supply only selected original image refs from the same frozen run and continue the same Scenario Diagnostic thread. No Provider authority, cross-run context, extra query round, or downstream formatting diagnosis is added. Slow Query selection has its separate record. The explicit `noDiagnosticTimeout` read API option is a verification aid, not a claim that latency is fixed. It records ordinary diagnostic deadlines without aborting. Default automation, one-hour isolated Job/collection limits, Provider operation/auth/network limits, classifier and Judge remain unchanged. ## Verification in progress On 2026-09-09 00:04 +08, final `npm test` passed: 1,230 TypeScript tests (1,228 passed, 2 skipped, 0 failed) and 9 isolated-runner tests (all passed). The exact filtered command output is retained in the local verification artifact `final-tests.txt`; it is not a full TAP transcript. Focused tests cover actual same-thread image request/continuation, allowed original refs, no repeated image delivery, distinct problemSummary/question, known/unknown tier planning, request-scoped deadline isolation, caller cancellation, and retained network timeouts. The initial Scenario Diagnostic turn and Evidence receive the same runner-owned exact-org/cluster tier metadata, including Shared/Starter/Essential. Skipping a redundant metadata-only query does not remove those applicability facts from diagnosis; unknown tier still uses the original Provider preflight. Local real-model checks use the unmodified TCID-13103 run `7b0f4454-a70c-41d7-adcb-13fe8f2070df`: its original route, cumulative R1+R2 handoff and original CPU image. They must not manufacture an R1-only handoff or edit the Router anchor. Router image extraction is checked separately with the production routing prompt and that same frozen input. Local Codex execution is not live Agent9/Provider or a full replay; its results will be recorded separately. ### Local model results (2026-09-08, production source frozen) Artifacts: local replay directory `diagnostic-context-local-20260908.FpnNGA`, model `gpt-5.6-sol`, reasoning `high`. - Production Router prompt, original frozen input/image: 3/3 retained the db-tidb-2 excursion around 21:01–21:02 (+08:00) and peer separation. Windows differed (narrow event versus wider displayed chart), but none substituted the 20:32 common peak. Runtime was 35.9 / 54.0 / 30.1 seconds. - Original finalization-only simulator plus on-demand image: all 3 trials requested the original graph and distinguished the 21:02 node event from the earlier global TopSQL slice. All retained unresolved initiating workload attribution. Trial 2 nevertheless mislabeled local 21:02 as UTC in its observation and refinement request, so this is **not a 3/3 time-semantics pass**. The complete first results remain in `diag-image-trials.json` without replacement. - That inherited finalization-only entrypoint does not provide the production initial Scenario Diagnostic turn's frozen ticket context (including +08:00 cutoff). A separate check now runs the production initial evidence-request prompt from the same frozen input, then same-thread original-handoff finalization with on-demand images. This completed on 2026-09-09 00:09 +08 in `diag-full-context-trials.json`: all 3 trials requested the original image during the initial turn, continued the same thread, retained db-tidb-2's approximately 13:02 UTC / 21:02 local event, and explicitly distinguished the earlier 12:32 UTC / 20:32 local global TopSQL sample. None used that earlier sample to establish the cause of the later node increment or exclude the local hotspot. The bounded original-image/time criterion passed 3/3; no production prompt was changed to correct the first test's UTC error. - This is not a full root-cause pass: trials 1 and 3 still label the earlier cluster-wide SQL baseline confirmed and rank the node-local session/TSO explanation below it, while acknowledging the missing onset-local TopSQL comparison. Trial 2 labels workload attribution correlated and the local diagnosis suspected. Whether ticket-level synthesis foregrounds the actual node incident instead of the older background mechanism remains an explicit full-replay observation, not a hidden success claim or an additional repair implemented here. These checks test image extraction, local image-request/continuation and use of frozen evidence only. They do not exercise live Agent9 image materialization, Provider execution, new queries, or end-to-end latency. Those remain release replay checks. Old status/error wording is not silently promoted to case success. ## Post-deployment verification (2026-09-09) See the [three-case verification report](../replay-verification/2026-09-09-diagnostic-context-efficiency.md) for exact runs, fingerprints, scopes, timing, and retained limitations. TCID-13103 now preserves the reported node/time throughout actual acquisition and the final answer, but still has instance TopSQL gaps and causal-strength risks. TCID-13139 completes both selected diagnostic scenarios with a partial workload acquisition result. TCID-13132 successfully exercises live on-demand original-image continuation, but still lacks the target SQL, equal-row runtime contrast, and TiFlash resource curve; both local diagnoses remain inconclusive. Same-session reuse did not eliminate repeated Skill reads in the observed TCID-13103 refinement; do not claim that efficiency problem verified. No `optimized_verified` or ticket-wide repaired status is set, and no Jira/DB/Base publication is authorized. Keep the original failed local trial evidence above.