# Agent9 Managed Tool Query Repair Parity - Owner: Managed Clinic Tool, Evidence Agent prompt, and deterministic Agent9 Provider materializer - Source: TCID-13071 replay `7be96988-68d9-43d0-b3bf-e75b1d93f944`; v12 verification replays TCID-13070 `93e36c66-f453-4360-8378-6a9626b2a20b` and TCID-13071 `03c9f361-8ca9-44cd-af99-d848c849cf2d`; v13 verification replays TCID-13070 `ef8432e2-922d-415c-8439-f0b7612befe0` and TCID-13071 `9f4efb52-7cd8-4077-a752-62f91959781b` - Implementation status: deployed in AOS `7f30f3b13c87b2b3e0691ae51be80ba389216ae3` with Agent9 Managed Clinic v13 - Effective status: `optimized_verified` ## Bounded problem Managed Clinic returned only `topsql_instance_invalid` for a local TopSQL argument rejection. The Evidence Agent did not receive the invalid field and accepted shape, and the final execution summary treated the failed call as a permanent independent objective even when a later call with the same queryId succeeded. The Managed Tool also rejected six-character Slow Query Digest references that the Kubernetes Provider Executor and the published Agent9 Tool schema both support. The v12 verification also showed that an overlong semantic query identifier was rejected even though it is only a correlation key, and that a real Agent9 Tool `runtime_error: fetch failed` was persisted but reduced to `provider_tool_call_failed` by the AOS adapter. ## Intended change Managed Clinic behaves like the ordinary Kubernetes Provider program: it returns a stable semantic `queryId`, operation, safe error classification, field and accepted shape for local validation, and bounded upstream status and detail for every HTTP 4xx. The same Evidence Agent may make at most four changed corrections in its current Turn. Calls sharing the exact `queryId` are one objective, and only their final outcome determines evidence completion and the final execution status. An unrepaired request retains its actual terminal error code instead of becoming `provider_tool_call_failed`. Managed Slow Query now accepts an exact 64-character Digest or a 6-to-63 character prefix. Dedicated uses a bounded `LIKE '%'` predicate; Starter/Essential retrieves the bounded NGM candidate list and filters it inside the ordinary Tool program, preserving the full Digest inside the Tool Sandbox while returning only the compact reference. Managed Clinic canonicalizes harmless query-id length or separator differences to one deterministic bounded artifact key instead of rejecting Provider work. When Agent9 itself reports a Tool execution failure, the adapter now projects the actual safe runtime code and message into `provider-result-v1`, and final execution retains that code rather than inventing a generic Provider failure. ## Verification The v13 read-only replays ran concurrently after both AOS and Agent9 internal production deployment. TCID-13071 completed evidence acquisition with `providerCapabilityVerified=true`, no failure code, and no `query_id_invalid` or `slow_query_digest_invalid`; its final evaluation was `ai_better`. This verifies that harmless semantic query IDs are canonicalized, compact Slow Query Digest-prefix refinement executes, and corrected caller requests do not leave a false terminal failure. TCID-13070 completed response generation and its final evaluation was `ai_better`. One real Agent9 Tool Sandbox transport failure remained, and the final execution summary retained the actual `runtime_error` emitted by Agent9 instead of replacing it with `provider_tool_call_failed`. This is verification of the bounded error-projection fix, not a claim that the independent transient Agent9-to-Tool transport failure has been eliminated. Focused contract tests cover local validation and every supported upstream HTTP 4xx class: 400, 401, 403, 404, 409, 422, and 429. Each returns the same ordinary structured Provider result with its canonical query ID, safe status/detail, and correction context where applicable; 422 has no special status. ## Next validation action Track a recurrence of Agent9 Tool Sandbox `runtime_error: fetch failed` as a separate operational reliability signal. Do not reopen this Provider-result contract optimization unless a future replay again obscures the actual code or marks a corrected caller request as terminal.