diff --git a/docs-ai/065-bundled-agent-skills/001-action.md b/docs-ai/065-bundled-agent-skills/001-action.md
index 5ac4a024..50019e65 100644
--- a/docs-ai/065-bundled-agent-skills/001-action.md
+++ b/docs-ai/065-bundled-agent-skills/001-action.md
@@ -61,9 +61,10 @@
 - Settings reports a successful link with a toast only (the chip changes state); the plan's
   "mirroring the CLI install row" did not decide this, and a modal per link would be noise.
 - `agentSkills` is created and cleared by `SettingsFeature.setSelection`, not by
-  `AppFeature.setSelection` as the plan's `agentProfiles` analogy implied: a grandparent
-  mutation happens after the child `ifLet` has run, so it cannot cancel in-flight link effects
-  (K3 review round 1).
+  `AppFeature.setSelection` as the plan's `agentProfiles` analogy implied: the grandparent's
+  `Reduce` runs before the `Scope` into `SettingsFeature`, so its mutation lies outside the
+  child `ifLet`'s transition boundary and cannot cancel in-flight link effects (K3 review
+  round 1, wording corrected in round 2).
 - The section shows detected targets only, as planned, but the "no target detected" state adds an
   explicit pointer to `prowl skills install --target …` since Settings never creates folders.
 
diff --git a/docs-ai/065-bundled-agent-skills/005-k3-settings-agent-skills.md b/docs-ai/065-bundled-agent-skills/005-k3-settings-agent-skills.md
index 975a3185..463e7a26 100644
--- a/docs-ai/065-bundled-agent-skills/005-k3-settings-agent-skills.md
+++ b/docs-ai/065-bundled-agent-skills/005-k3-settings-agent-skills.md
@@ -61,9 +61,10 @@ shared installer so both surfaces always report the same status.
 
 Adversarial review round 1 (sibling reviewer, brief and findings kept outside the repository)
 found no P0/P1 and one P2: the child state was created and cleared by `AppFeature.setSelection`,
-the grandparent, which mutates `settings.agentSkills` *after* `SettingsFeature`'s `ifLet` has
-already run for that action — so the `ifLet` never observed a non-nil → nil transition and an
-in-flight install/uninstall effect survived the section switch; its delayed completion then hit
+the grandparent, whose core `Reduce` runs before the `Scope` into `SettingsFeature` — so the
+nested optional was already nil when `SettingsFeature`'s `ifLet` ran, the `ifLet` never observed
+a non-nil → nil transition of its own, and an in-flight install/uninstall effect survived the
+section switch; its delayed completion then hit
 nil child state (a TCA runtime warning in Debug) and was dropped without the promised refresh,
 toast, or failure alert. The reviewer reproduced it with a suspended-install probe. The lifecycle
 now lives in `SettingsFeature.setSelection` (create on `.commandLineTool` when absent, clear
@@ -71,6 +72,11 @@ otherwise), so `ifLet` cancels the child's effects with the state; `AppFeature`
 `agentSkills`. Pinned by an `AppFeature` test that suspends install/uninstall on a `TestClock`,
 switches to General, and requires every effect to be gone.
 
+Round 2 re-verified the fix (including re-selecting the same row, reopening Settings mid-action,
+and creating a replacement child) and found no P0/P1/P2; its one P3 was that this record and the
+action log had described the reducer order backwards ("after the `ifLet` had run"), corrected
+above. The review loop closed there.
+
 ## Verification
 
 - TDD RED: 25 missing-symbol errors for `SkillInstallClientTests` before the client existed;