# OP Factual Child Handoff - Optimization ID: `op-factual-child-handoff-2026-08-10` - Owner: auto-oncall-services runner - Source reference: Focus #606 post-acquisition replays for `APID-29265` and `APID-29232` - Implementation status: merged and deployed to staging - Effective status: `optimized_unverified` ## Bounded problem The Wetech helper already returned compact provider results and the runner materialized their artifacts, but the child handoff observation copied only a generic statement such as “returned usable provider evidence.” The following tool-free parent is prohibited from reading those artifacts, so facts such as metric extrema, affected instance, record count, and TopSQL numeric rows were unavailable to its judgment. The parent then spent later rounds asking for information that had already been acquired. This record owns only the factual OP handoff and the boundary for what counts as a later refinement round. It does not change query-ID selection, provider authentication, target resolution, the four-round maximum, Jira publication, Judge semantics, attachment handling, required capabilities, or Focus lifecycle. ## Candidate change - Keep the provider compact report as the full bounded audit artifact. - Add an allowlisted, at-most-1,000-character factual suffix to each handoff observation containing useful metric counts/extrema, selected series entity labels, and dashboard row numeric decision fields. - Never copy raw SQL text, plan text, full digests, credentials, tokens, or arbitrary provider strings into the handoff. - Use one batch's capacity of up to six queries for already-known supporting checks. Provider transport retry remains helper-owned inside that batch. - Allow another round only for a genuinely new load-bearing question that can change the diagnosis; restating or summarizing facts already present in the handoff is not refinement. ## Verification evidence Local contract tests verify that realistic metric and TopSQL compact results reach the parent as decision facts while SQL, plan, digest, and unallowlisted strings remain absent. The complete TypeScript test suite passes with 697 tests. Staging deployed commit `d3db6e5a1d711006821b6410f06c41b8ce9d0955` through [aisupport run 31353565480](https://github.com/tidbcloud/aisupport/actions/runs/31353565480) and [app-delivery PR 76608](https://github.com/tidbcloud/app-delivery/pull/76608). The live Deployment and optimizer CronJob both use that image. The Deployment reported 1/1 updated, ready, and available replicas; the Pod was Ready with zero restarts, and `/readyz` returned HTTP 200. - Read-only `APID-29265` replay run `1fbfce6e-e715-44b8-8519-67268d17607f` completed in one Wetech round, down from the prior three rounds. Its five-query batch produced three factual observations containing metric counts, extrema, and instances. The parent directly used those facts, returned a valid terminal `required=false/question=null/stopCondition=null` object, and response generation completed with verified provider capability. - Read-only `APID-29232` replay run `728b7831-50ab-4bc0-8d49-afd41fcea823` completed in two Wetech rounds, down from the prior three-round schema failure. Round one used one five-query batch and exposed the new load-bearing need for store-to-instance CPU and placement alignment. Round two used another five-query batch, after which the parent returned the valid terminal null refinement and response generation completed with verified provider capability. - Both requests used `mode=read`; publication, database, evaluation, Feishu, and Jira writes were `not_requested`. No customer-visible or internal Jira comment was changed. These runs verify the runtime factual handoff and terminal refinement behavior, but the optimization remains `optimized_unverified` until the corrected drafts are refreshed into the existing Jira internal comments and read back with `public=false`. ## Unresolved gaps - Jira internal-comment refresh and readback remain unresolved for both replay cases. - A provider result that lacks every allowlisted decision field remains usable only through its compact artifact and may still require a genuinely changed query. - This change does not guarantee a diagnosis or promote partial evidence to a causal conclusion. - It does not alter Clinic handoff materialization; the observed failures and current implementation scope are OP/Wetech. ## Next validation action With explicit Jira-write authorization, update each existing auto-oncall Jira internal comment in place from its verified read-only replay draft, then read back the exact comment ID and verify `public=false`. Do not rerun provider analysis merely to regenerate the already verified bodies. Promote this problem to `optimized_verified` only after both comment refreshes succeed.