# Agent9 Diagnostic Role Agents

<!-- auto-oncall-optimization-record
{
  "schemaVersion": "auto-oncall-optimization-record-v1",
  "optimizationId": "agent9-diagnostic-role-agents-2026-08-24",
  "problems": [
    {
      "problemId": "agent9-diagnostic-roles-share-capabilities",
      "summary": "All diagnostic roles used one Agent9 Agent, so Router, Scenario Diagnostic, and Final Response Turns inherited Provider Skills and Managed Tools that only the Evidence Agent needs.",
      "status": "optimized_verified",
      "applicability": "Agent9-backed Auto-oncall diagnostic multi-turn runs.",
      "issueKeys": ["TCID-13071"],
      "sourceRefs": ["auto-oncall-run:a64e970a-c16f-410e-ad27-824c8012ba13#problem:shared-role-agent-capabilities"],
      "signalTextIncludesAny": [
        "agent9 diagnostic roles share one agent",
        "non evidence role inherits managed tools"
      ],
      "observedThrough": "2026-08-24",
      "effectiveAt": "2026-08-24T09:45:00Z",
      "verifiedAt": "2026-08-24T09:58:40Z",
      "attemptedAction": "Bind Scenario Router, Evidence Agent, Scenario Diagnostic Agent, and Final Response Agent to four distinct Agent9 Agents; install Provider Skills and Managed Tools only on the Evidence Agent.",
      "evidenceUrl": "https://internal.us-west-2.agent.mem9.ai/api/sessions/sess_f3b0cccbcc8245b892e1f79991c594f3"
    }
  ]
}
-->

- Optimization ID: `agent9-diagnostic-role-agents-2026-08-24`
- Owner: Agent9 adapter role-to-Agent binding
- Source: TCID-13071 read-only run `a64e970a-c16f-410e-ad27-824c8012ba13`
- Implementation status: deployed to the normal Agent9-backed AOS Deployment
- Effective status: `optimized_verified`

## Bounded problem

One Agent9 Agent owned every diagnostic Turn. That made Provider capability part
of Router, Scenario Diagnostic, and Final Response runtime configuration even
though those roles are offline and tool-free by contract.

## Candidate change

Use one explicit Agent9 Agent ID per diagnostic role. Preserve Primary Analysis
on the existing compatibility Agent. Install the six shared Auto-oncall Skills
and the Clinic and Manus Managed Tools only on the Evidence Agent.

## Verification evidence

Read-only run `f724218c-1620-4f95-81ef-4fba2e137058` returned `status=ok` and
proved the live role mapping:

- Router Session `sess_f3b0cccbcc8245b892e1f79991c594f3` used
  `agent_b0f561f8fa3bb508083325d3ace4dc98` for both assignment and reconciliation.
- Evidence Sessions `sess_f969d1917045422ca828328271373a34` and
  `sess_0964ae27ce9b4e70a27e614913374aa4` used
  `agent_8a187d1f3645a482ed8861274d25f5fd`; Clinic operations appeared only there.
- Scenario Diagnostic Sessions used
  `agent_0f75560bc77477b5fc9c25c2876b4adb`.
- Final Response Session `sess_6dc28a8fdf214626925587bcb34de2f4` used
  `agent_9aa6d1d0bc0ec6a7695bf29dc4dc26b4` and completed with only `llm.complete`.

Router, Scenario Diagnostic, and Final Response Agents expose zero Managed
Tools. The Evidence Agent exposes Clinic v10 and Manus v4 plus the six shared
Auto-oncall Skills.

At `2026-08-24T10:02:45Z`, the normal `auto-oncall-services-agent9`
Deployment was `1/1 Ready` on image `8ffeb7a8`; its rendered Pod environment
contained all four verified role Agent IDs. Polling remained active and the
worker and bot watchdog continued to report healthy idle cycles.

## Unresolved gaps

- The run was not faster overall: it took 12 minutes 2 seconds versus 11 minutes
  8 seconds before the split. One Evidence branch still reached its 180-second
  cap, one Scenario finalization hit a transient LLM provider failure, and
  Router reconciliation varied upward. These are not claimed as fixed here.
- The Evidence Agent still reports the legacy sandbox-template probe warning;
  capability separation correctly confines it to Evidence but does not repair it.

## Next validation action

Monitor the next naturally selected live job for the same role correlation;
track Evidence timeout and sandbox-template readiness as separate bounded
problems.