# Preserve Parent Semantic Autonomy - Optimization ID: `parent-semantic-autonomy-2026-08-05` - Owner: auto-oncall-services - Source references: staging replays for `TCID-12973` and `APID-29231` - Implementation status: deployed to staging - Effective status: `optimized_verified` ## Bounded problem The scenario parent owns diagnosis and has broader context than its evidence child. Its terminal packet was structurally valid and cited canonical evidence, but backend cross-field rules treated one parent-authored classification as a packet-level failure. The runner then replaced the whole result with an empty bounded-failure packet, discarding unrelated observations and artifact refs. ## Implementation - Keep schema shape, scenario identity, canonical evidence refs, bounded size, safe text, execution state, and checkpoint integrity as hard validation. - Replace the live parent response with `auto-oncall-diagnostic-scenario-decision-v1`. The parent returns only its status, observations, counterevidence, mechanism, alternative, single monotonic workload-attribution conclusion, monitoring-semantics signal, severity, and handoff. The attribution state is exactly one of `causal_confirmed`, `identity_proven`, `correlated`, `unresolved`, or `not_applicable`, so conflicting identity, gate, and causal states cannot be emitted independently. - Remove `source`, `provenance`, `kind`, `linkage`, `completeness`, objective gates, execution, bounded failures, and artifact inventory from the live parent response schema. The runner merges scenario identity, role, entity/window, current child handoff, execution state, and artifact inventory without comparing the two owners' semantic outputs. - Treat parent conclusions as inputs to the higher-level synthesis. The final synthesis may correct an unsupported conclusion using broader context; the runner does not reject the scenario or rewrite the parent's diagnosis. - Retain the old typed packet parser only for legacy packet and checkpoint compatibility. Its local reconciliation is not used by the new live response schema. ## Local verification - TypeScript typecheck and build pass. - The targeted diagnostic scenario, runner, and multiturn suites pass all 99 tests. - All 623 repository tests pass. - The live decision schema has no source/provenance pair or independently authored objective gate, so the two replayed contradiction shapes cannot be expressed by a new parent response. - A parent decision is merged with route and child-owned context while keeping all parent observations and canonical evidence refs. Extra legacy fields such as `provenance` are rejected structurally because they have no owner in the decision contract. - Branch-local `mode=read` replay `APID-29231` run `959ed41f-be6d-4fa6-8174-7e792ebee42b` returned a terminal `auto-oncall-diagnostic-scenario-decision-v1`, wrote a complete `query-execution` checkpoint with five child-owned artifact refs, and did not emit `diagnostic_workload_identity_proof_invalid`. - Branch-local control replay `ST-2865` run `3cf089c5-4761-4a7f-8c16-7987bf4cefa2` produced a terminal parent decision while its child handoff remained partial. The follow-up local runner change now preserves that partial execution state independently while retaining the parent decision and child-owned artifact refs. - Branch-local `TCID-12973` run `c17c4128-2e8b-43c3-85a6-89b79a40b20d` reached the scenario reserve boundary before a terminal parent decision. Its partial checkpoint preserved four child artifacts and did not emit the former provenance validator error, but it does not verify the new decision schema for that case. - Deployed `mode=write` replay `TCID-12973` run `60f1bb4f-3f2b-4382-ba4e-3f1e2c41ea0f` produced a complete child handoff, five artifacts, a terminal parent decision, and complete final synthesis. It did not emit provenance, proof, objective-gate, or packet mismatch errors. DB run `120017`, Feishu Base, and existing private Jira comment `23327373` were updated successfully. - Deployed `mode=write` replay `APID-29231` run `de874fd8-c11c-416f-8b85-f57d59996f02` preserved seven Wetech artifacts and provider gaps, returned a terminal parent decision and complete synthesis, and did not emit `diagnostic_workload_identity_proof_invalid`. DB run `120018`, Feishu Base, and existing private Jira comment `23279460` were updated successfully. - Both Jira comments were read back through Jira Platform and Service Desk APIs with the same comment IDs, expected bodies, and `public=false`. ## Unresolved gaps - Provider acquisition gaps remain valid parent inputs rather than backend packet failures. APID-29231 still had bounded Wetech operation errors, but they no longer erased the complete observations and artifact inventory. - The separate finalization-availability optimization owns synthesis timeout behavior and is independently verified by the same deployed source-case replays. ## Next validation action Monitor for new post-verification parent packets rejected by cross-field semantic validation. A structurally invalid packet remains an error; a valid parent conclusion that disagrees with narrower child metadata must continue to reach higher-level synthesis instead of failing the scenario.