# Symptom-specific Scenario Diagnostic causal admission

<!-- auto-oncall-optimization-record
{
  "schemaVersion": "auto-oncall-optimization-record-v1",
  "optimizationId": "symptom-specific-causal-admission-2026-09-05",
  "problems": [{
    "problemId": "cpu-attribution-blocked-by-transaction-mechanism-checklist",
    "summary": "CPU workload causality was gated by transaction shape and request-to-latency linkage regardless of the diagnostic question.",
    "status": "optimized_unverified",
    "applicability": "CPU hotspot workload attribution supported by aligned node-local TopSQL contribution and corroborating process CPU changes or recovery.",
    "issueKeys": ["APID-29251"],
    "sourceRefs": ["https://tidb.atlassian.net/browse/APID-29251#problem:cpu-causal-admission"],
    "signalTextIncludesAny": ["CPU attribution blocked by transaction mechanism checklist"],
    "observedThrough": "2026-09-05",
    "effectiveAt": "2026-09-05T06:27:40Z",
    "verifiedAt": null,
    "attemptedAction": "Use symptom-specific causal evidence in Scenario Diagnostic; retain linked transaction evidence for DML-to-COMMIT attribution, partial CPU contribution limits, and neutral Evidence and wording-only Final Response ownership.",
    "evidenceUrl": "replay://APID-29251/run/6232c2de-bc46-47b3-a037-30c76251f50e"
  }]
}
-->

- Owner: Scenario Diagnostic production prompts and diagnosis SOPs.
- Source: user review of APID-29251 offline frozen-fact model comparison.
- Implementation: source PR #1342 merged as
  `b80e1efb438300f6d4d859869b74037392057bbb`; deployed, not case-verified.
- Local verification (2026-09-05): TypeScript compilation passed; focused
  diagnosis/SOP/optimization-record tests passed 100/100; full suite passed
  1089/1089. These checks do not establish a changed real-case model outcome.
- Evidence boundary: the APID-29251 comparison used summarized historical facts,
  not an original paired route/handoff or a fresh end-to-end replay. No Judge
  result is being changed. TCID-12986 partial contribution remains a control.
- Next validation: establish an authoritative historical cluster mapping before
  authorizing another replay, or use an original paired route/handoff for a
  bounded Scenario Diagnostic regression. Do not guess a target or count an
  upstream-blocked replay as proof of the causal-admission change.

## Deployment and replay (2026-09-05)

- CI run [33949209173](https://github.com/tidbcloud/aisupport/actions/runs/33949209173)
  first failed the 25-ms streamed-iterator cleanup test (expected 1, got 0).
  The first release was cancelled before Skill/application updates. An unchanged
  CI rerun passed; no test or runtime deadline was modified.
- Unified release [33949431694](https://github.com/tidbcloud/aisupport/actions/runs/33949431694)
  succeeded, publishing `op-diagnosis` v9. app-delivery PRs
  [79320](https://github.com/tidbcloud/app-delivery/pull/79320) and
  [79321](https://github.com/tidbcloud/app-delivery/pull/79321) merged.
- Deployment and root endpoint confirmed the exact source SHA above. Pod
  `auto-oncall-services-agent9-7f55476d5-vb2gp` was Ready with zero restarts;
  `/`, `/healthz`, and `/readyz` returned HTTP 200.
- One `mode=read` APID-29251 replay returned HTTP 200 in **160.195 seconds**:
  `6232c2de-bc46-47b3-a037-30c76251f50e`, Agent9 backend with observed
  Router/Final model `gpt-5.6-sol`, high reasoning.
- The historical input's `亚太02` did not resolve to an exact Wetech target.
  Router selected `workload-hotspot`, but its entity set was empty; the runner
  recorded `diagnostic_wetech_target_missing`, zero scenario duration, and no
  Scenario Diagnostic thread. Provider execution and the changed causal
  admission were **not exercised**.
- Final decision: `human_required`, with internal target confirmation required.
  Judge: `different_but_valid`, eligibility `ai_human_required`, compared with
  human comment `23304357` proposing consolidation with APID-29231. This is not
  proof of improved workload diagnosis and is not the earlier frozen-fact test.
- Jira, database, evaluation persistence, and Feishu publication were all
  `not_requested`. No case answer was overwritten and no second replay started.
- Local API response and trace bundle:
  `/home/nizheming/.cache/aos-model-comparisons/20260905-scenario-causal-deployed-replay/`.