@@ -89,7 +89,9 @@ Provider locking, paused-child/current-task publication, and semaphore admission
 
 Six injected legacy transition policies must produce deterministic shortest counterexamples through the same explorer: start before commit, resume before permit release, redelegation before permit release, empty current-task publication, two stale provider commits from competing snapshots, and releasing parent-transition serialization immediately after publication. The last witness must causally include first-child completion and parent publication, a second-child commit, release of the first child's scheduler permit, and then the stale first-child continuation. The checker prints the distinct reachable-state count, complete scenario/action/landmark coverage, bounds, and each named counterexample trace. It deliberately does not add a WAL, global profile projection, or scheduler state to persisted `HistoryItem` records.
 
-For #921, the execution-context matrix also exhaustively combines saved, unsaved, and locked profile selection with both possible focused views before and after the immutable handoff snapshot. Focus changes are deliberately environment inputs rather than lifecycle state: every schedule must retain the selected task-local mode, profile identity, and cloned configuration while leaving the parent context unchanged.
+For #921, the execution-context matrix verifies saved, unsaved, and locked profile selection at the handoff boundary. This proves only that delegation writes the requested task-local mode and cloned configuration into the child context. It does not prove that every downstream consumer reads that context. The checker retains a divergent-mode witness in which the child task mode differs from the shared provider mode so reader refinements can demonstrate that choosing the wrong source is observable.
+
+Issue #1623 exposed that missing refinement: `getEnvironmentDetails`, `presentAssistantMessage` tool validation, and custom-tool execution still read shared provider mode after the handoff snapshot became task-local. PR #1625 owns the production fix and adds a dedicated delegated-mode reader check plus focused adapter tests. Until that runtime PR is merged, the selector model establishes the write-side premise and the divergent-mode witness traces the unresolved read-side obligation; it must not be cited as complete #921 coverage by itself.
 
 ## Task fan-out protocol model
 
@@ -139,16 +141,16 @@ These are safety claims within the documented bounds. The checks do not claim li
 
 The following map separates issue observations from the architectural interpretation encoded here. Open issues can change after this document is written; follow each link for current status.
 
-| Issue and directly observed evidence                                                                                                                                                                                                                                                                           | Derived protocol rule                                                                                                                                                                             | Production transition and current check                                                                                                                                                                                                                                                                    |
-| -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| [#1469](https://github.com/Zoo-Code-Org/Zoo-Code/issues/1469): the issue report states that a barrier-controlled two-host run reproduced an old child completion clearing a newer handoff 25/25 times.                                                                                                         | Completion is conditional on the parent still awaiting that exact child; a live-linked child must remain owned by its parent.                                                                     | `completeDelegatedChild` rejects stale authoritative input. The lifecycle explorer checks that reducer rule, while the shared-store explorer reproduces the cross-host stale-cache counterexample with an exact causal witness.                                                                            |
-| [#1021](https://github.com/Zoo-Code-Org/Zoo-Code/issues/1021): an in-flight `saveClineMessages` can restore parent/root IDs after abandonment cleared them.                                                                                                                                                    | Detachment should be monotonic: later lifecycle work must not reattach an abandoned child.                                                                                                        | `abandonDelegatedChild` clears both sides. The shared-store explorer proves the detach commit occurs, then reproduces a refreshed-cache delta that preserves interrupted status while restoring stale live-task lineage.                                                                                   |
-| [#1453](https://github.com/Zoo-Code-Org/Zoo-Code/issues/1453), under user report [#1279](https://github.com/Zoo-Code-Org/Zoo-Code/issues/1279): CI observed `TaskCompleted` before restart-visible API history once; 120 local repetitions did not reproduce it.                                               | Completion implies restart-visible assistant history. Delayed or failed writes keep completion pending, and cancellation settles readiness without starting stale retries or emitting completion. | The completion persistence explorer checks the bounded event-ordering and cancellation contract for standalone and delegated tasks. Focused `Task` and `AttemptCompletionTool` tests cover the production adapter; `restart-persistence.test.ts` verifies visibility through a fresh extension host.       |
-| [#921](https://github.com/Zoo-Code-Org/Zoo-Code/issues/921): delegation across parallel tabs lacks coverage for different view-local mode/profile state.                                                                                                                                                       | Delegation must bind an explicit immutable execution-context snapshot rather than read whichever view is focused later.                                                                           | `selectHandoffExecutionContext` creates the production snapshot. The provider handoff checker exhaustively combines saved, unsaved, and locked profiles with focus changes before and after selection; focused selector and provider tests cover lookup fallback and child creation.                       |
-| [#920](https://github.com/Zoo-Code-Org/Zoo-Code/issues/920): issue analysis identifies a missing cross-instance history-update test and potential lost writes.                                                                                                                                                 | Distinct task writes must not overwrite one another, and same-task conflicts need an explicit merge/ownership rule.                                                                               | The shared-store explorer checks distinct-task writes and same-record independent deltas. Cross-instance store tests retain production API coverage, and the synchronized real-filesystem smoke test exercises the actual lock/write path without claiming exhaustive filesystem proof.                    |
-| [#369](https://github.com/Zoo-Code-Org/Zoo-Code/issues/369) and [#372](https://github.com/Zoo-Code-Org/Zoo-Code/issues/372): planned fan-out keeps a parent live while a child runs and requires completion routing by explicit parent ID, single-writer result readiness, permit release, and orphan cleanup. | Persisted `delegated` status is ownership, not proof that the parent instance is suspended. Completion must route by IDs; scheduler resources and live-instance state need separate invariants.   | The fan-out protocol checker models a live parent, two siblings, per-child permits/readiness, explicit delivery, parent loss, and orphan cleanup. `TaskScheduler` and ID-routed completion tests cover shipped primitives; production fan-out remains disabled pending live-parent integration and E2E.    |
-| [#1468](https://github.com/Zoo-Code-Org/Zoo-Code/issues/1468): a late chunk from one request combined tool identity with arguments from another request; rerun passed.                                                                                                                                         | Every stream accumulator needs a request/task generation key, and late events cannot mutate another scope.                                                                                        | Separate protocol. The [native tool-call parser request-scope model](./native-tool-call-parser-scoping-model.md), whose source of truth is `scripts/check-native-tool-call-parser-scoping.ts`, exhaustively replays bounded production-parser interleavings without adding fields to this lifecycle model. |
-| [#612](https://github.com/Zoo-Code-Org/Zoo-Code/issues/612): the CLI copied a status union and omitted `interrupted`.                                                                                                                                                                                          | Lifecycle vocabulary should have one type owner.                                                                                                                                                  | `HistoryItemStatus` is derived from `HistoryItem`, and production/checker transitions share `taskLifecycle.ts`; consumers should import rather than copy the union.                                                                                                                                        |
+| Issue and directly observed evidence                                                                                                                                                                                                                                                                                         | Derived protocol rule                                                                                                                                                                             | Production transition and current check                                                                                                                                                                                                                                                                    |
+| ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| [#1469](https://github.com/Zoo-Code-Org/Zoo-Code/issues/1469): the issue report states that a barrier-controlled two-host run reproduced an old child completion clearing a newer handoff 25/25 times.                                                                                                                       | Completion is conditional on the parent still awaiting that exact child; a live-linked child must remain owned by its parent.                                                                     | `completeDelegatedChild` rejects stale authoritative input. The lifecycle explorer checks that reducer rule, while the shared-store explorer reproduces the cross-host stale-cache counterexample with an exact causal witness.                                                                            |
+| [#1021](https://github.com/Zoo-Code-Org/Zoo-Code/issues/1021): an in-flight `saveClineMessages` can restore parent/root IDs after abandonment cleared them.                                                                                                                                                                  | Detachment should be monotonic: later lifecycle work must not reattach an abandoned child.                                                                                                        | `abandonDelegatedChild` clears both sides. The shared-store explorer proves the detach commit occurs, then reproduces a refreshed-cache delta that preserves interrupted status while restoring stale live-task lineage.                                                                                   |
+| [#1453](https://github.com/Zoo-Code-Org/Zoo-Code/issues/1453), under user report [#1279](https://github.com/Zoo-Code-Org/Zoo-Code/issues/1279): CI observed `TaskCompleted` before restart-visible API history once; 120 local repetitions did not reproduce it.                                                             | Completion implies restart-visible assistant history. Delayed or failed writes keep completion pending, and cancellation settles readiness without starting stale retries or emitting completion. | The completion persistence explorer checks the bounded event-ordering and cancellation contract for standalone and delegated tasks. Focused `Task` and `AttemptCompletionTool` tests cover the production adapter; `restart-persistence.test.ts` verifies visibility through a fresh extension host.       |
+| [#921](https://github.com/Zoo-Code-Org/Zoo-Code/issues/921): delegation across parallel tabs lacks coverage for different view-local mode/profile state. Follow-up [#1623](https://github.com/Zoo-Code-Org/Zoo-Code/issues/1623) proves a correct snapshot is insufficient when downstream readers use shared provider mode. | Delegation must bind an explicit immutable execution-context snapshot, and every mode-sensitive consumer must read the child task's context rather than whichever provider/view is focused.       | This checker proves the write-side selector matrix and retains a divergent task/provider-mode witness. Runtime reader correctness is intentionally owned by [PR #1625](https://github.com/Zoo-Code-Org/Zoo-Code/pull/1625), its delegated-mode reader check, and focused environment/tool tests.           |
+| [#920](https://github.com/Zoo-Code-Org/Zoo-Code/issues/920): issue analysis identifies a missing cross-instance history-update test and potential lost writes.                                                                                                                                                               | Distinct task writes must not overwrite one another, and same-task conflicts need an explicit merge/ownership rule.                                                                               | The shared-store explorer checks distinct-task writes and same-record independent deltas. Cross-instance store tests retain production API coverage, and the synchronized real-filesystem smoke test exercises the actual lock/write path without claiming exhaustive filesystem proof.                    |
+| [#369](https://github.com/Zoo-Code-Org/Zoo-Code/issues/369) and [#372](https://github.com/Zoo-Code-Org/Zoo-Code/issues/372): planned fan-out keeps a parent live while a child runs and requires completion routing by explicit parent ID, single-writer result readiness, permit release, and orphan cleanup.               | Persisted `delegated` status is ownership, not proof that the parent instance is suspended. Completion must route by IDs; scheduler resources and live-instance state need separate invariants.   | The fan-out protocol checker models a live parent, two siblings, per-child permits/readiness, explicit delivery, parent loss, and orphan cleanup. `TaskScheduler` and ID-routed completion tests cover shipped primitives; production fan-out remains disabled pending live-parent integration and E2E.    |
+| [#1468](https://github.com/Zoo-Code-Org/Zoo-Code/issues/1468): a late chunk from one request combined tool identity with arguments from another request; rerun passed.                                                                                                                                                       | Every stream accumulator needs a request/task generation key, and late events cannot mutate another scope.                                                                                        | Separate protocol. The [native tool-call parser request-scope model](./native-tool-call-parser-scoping-model.md), whose source of truth is `scripts/check-native-tool-call-parser-scoping.ts`, exhaustively replays bounded production-parser interleavings without adding fields to this lifecycle model. |
+| [#612](https://github.com/Zoo-Code-Org/Zoo-Code/issues/612): the CLI copied a status union and omitted `interrupted`.                                                                                                                                                                                                        | Lifecycle vocabulary should have one type owner.                                                                                                                                                  | `HistoryItemStatus` is derived from `HistoryItem`, and production/checker transitions share `taskLifecycle.ts`; consumers should import rather than copy the union.                                                                                                                                        |
 
 The issue-derived cases intentionally map to bug classes rather than issue-specific flags. In particular, stale event ownership, monotonic terminal/detached state, explicit scope, and single-writer boundaries generalize to future concurrent task work.