# Agent9 Multi-turn Evidence Status

<!-- auto-oncall-optimization-record
{
  "schemaVersion": "auto-oncall-optimization-record-v1",
  "optimizationId": "agent9-multiturn-evidence-status-2026-08-21",
  "problems": [
    {
      "problemId": "agent9-complete-provider-evidence-projected-failed",
      "summary": "A successful Agent9 multi-turn Provider acquisition was projected as failed when the later Scenario Diagnostic Agent returned bounded_failure.",
      "status": "optimized_verified",
      "applicability": "Agent9-backed multi-turn diagnostics with completed Managed Provider Tool results and an unresolved local diagnosis.",
      "issueKeys": ["ST-2865"],
      "sourceRefs": ["https://tidb.atlassian.net/browse/ST-2865#problem:agent9-complete-provider-evidence-projected-failed"],
      "signalTextIncludesAny": [
        "provider_evidence_acquisition_incomplete",
        "providerCapabilityVerified false after completed managed tool calls",
        "bounded_failure overwrote complete provider acquisition"
      ],
      "observedThrough": "2026-08-21",
      "effectiveAt": "2026-08-21T07:39:57Z",
      "verifiedAt": "2026-08-21T07:49:44Z",
      "attemptedAction": "Derive Agent9 multi-turn evidenceAcquisition from each routed Evidence Agent Managed Provider Tool result while ignoring Core Tools and keeping Scenario diagnosis status independent.",
      "evidenceUrl": "https://github.com/tidbcloud/aisupport/actions/runs/32459210493"
    }
  ]
}
-->

- Owner: Auto-oncall execution-state projection
- Source: Agent9-backed `ST-2865` read-only replay
- Implementation status: deployed from `769cf9231d805f72db6cf9a1b44d12f510502ddb`
- Effective status: `optimized_verified`

## Bounded problem

The Agent9 Evidence Agent completed one Drive file read and three Clinic Managed
Tool calls. All three compact Provider reports and query-coverage entries were
complete, but the execution projection inspected only the legacy single-turn
artifact. It then reused the later Scenario Diagnostic Agent's
`bounded_failure` as evidence-acquisition state and emitted
`provider_evidence_acquisition_incomplete`.

## Candidate change

Derive Agent9 multi-turn acquisition from the exact Managed Provider Tool calls
in every routed Scenario Evidence turn. Ignore ordinary Agent9 Core Tools,
preserve partial or failed Provider results, and keep unresolved diagnosis
independent from Provider execution.

## Verification

The passive Agent9 runner became Ready on the exact deployed image at
`2026-08-21T07:39:57Z`. Read-only `ST-2865` replay
`a0dd273e-f3d1-4302-8a21-5d101f40b24a` then completed with HTTP 200. Across
two Evidence rounds, eleven Clinic Managed Tool calls completed and the
cumulative handoff had complete query acquisition with no Provider errors.
The final execution projection was `status=complete`,
`providerCapabilityVerified=true`, and `failureCodes=[]`; Router, Final
Response, and the independent Judge also completed. The run requested no Jira,
database, evaluation, or Feishu write.

## Remaining gaps

None for this bounded execution-status defect. Later regressions should rerun
the same read-only projection oracle without requiring identical diagnostic
wording.