diff --git a/docs/memory/architecture/frontend.md b/docs/memory/architecture/frontend.md
index 121d53bd43..a12ff226f0 100644
--- a/docs/memory/architecture/frontend.md
+++ b/docs/memory/architecture/frontend.md
@@ -45,7 +45,7 @@
 **Library system-install section (ent#126):** the install surface is a third stacked `<section id="systems">` on `views/Library.vue`, between Agent Templates and Skills. It originally shipped as a `?tab=`-driven strip on `Templates.vue`; ent#263 renamed that page to **Library** ("installable assets for your fleet") and chose stacked sections with jump anchors over tabs/filter pills, so ent#126 conforms to the page's model rather than reintroducing a competing one — the ordering puts Systems next to Agent Templates because both *install agents* (a template creates one, a manifest creates a wired fleet), leaving Skills last as the one that configures agents that already exist. The section is hidden outright below `creator` (gated on `hasMinRole('creator')`, mirroring `POST /api/systems/deploy`) rather than shown-and-disabled, since a browse surface gains nothing from a dead panel. It hosts `components/systems/` — `SystemInstallPanel` (bundled cards / upload / paste + Preview + Deploy) → `ManifestPreview` (agents, permission topology, schedules, blockers, acknowledgement gate) → `DeployResult` (all five `status` values) — over the new domain-scoped `stores/systems.js` (Invariant #6: a *System* is a manifest-deployed name-prefix group, a *System View* is a saved tag filter — different domains). The store's `normalizeError` collapses six response shapes and switches on `status`, never the HTTP code (`partial`/`invalid` are 200; `failed` is **500 with the full report as the body**), and binds `preview` to `previewedText` so Deploy is disabled after an edit. Plain `<textarea>`: the orphaned monaco `components/YamlEditor.vue` stayed unrevived because prod CSP is `script-src 'self'` with no `unsafe-eval`/`worker-src` while the dev CSP allows `unsafe-eval` — and was deleted, with its companion `EditorHelpPanel.vue`, in the #2492 dead-component sweep (the CSP rationale is permanent, so a revival would be a rewrite either way). `views/Dashboard.vue` reads `?view=`/`?tags=` so a fresh deploy lands filtered to its own fleet. **ent#454** adds the section's gated second half: `SystemTeardownPanel` → `TeardownPreview` → `TeardownResult`, rendered only when `enterprise_features` carries `system_teardown` (the verb is an entitlement-gated module, so an OSS build's Systems section is unchanged) on top of the same `creator` gate. Its preview is a **checklist**, not a list: a member the server matched by name alone (`evidence: prefix`) may belong to a sibling system whose name shares this one's prefix, the server cannot tell which, so the badge plus the per-agent opt-out is the over-capture escape hatch — and the checked list is sent back on execute to be intersected with freshly re-resolved membership. Remove is gated on four conditions (preview current for the typed name, membership verified, consequence acknowledged via ent#126's reused `:acknowledged`/`update:acknowledged` contract under its own `teardown-ack-checkbox` id — the pattern is shared, the address deliberately is not, since both panels render on one page, ≥1 agent checked) and states which is missing. `ManifestPreview`'s "there is no un-deploy" warning now branches on the same entitlement rather than being retired on every edition — it is still exactly true where the module is absent.
 **Top-nav IA — Library (trinity-enterprise#263):** the former Templates nav entry is one **Library** entry (`views/Library.vue`, `/library`) — a single surface for installable assets: an **Agent Templates** section (the existing Starter/GitHub/Custom card grids) and a **Skills** section (`components/LibrarySkillsSection.vue` — fleet-level browse over the shared skills library, backed by the fleet-scoped `stores/skillsLibrary.js`, deliberately separate from the agent-scoped `stores/skills.js` whose KeepAlive'd Skills-tab consumer never unmounts on nav-away). Stacked sections with in-page jump anchors (no `?kind=` machinery); per-section failure isolation; per-kind empty states. The skills half reads only the existing `GET /api/skills/library` (+`/status`) surfaces and renders the #183 contract via the shared `components/skills/` chips seam consumed by both the Library and the per-agent Skills tab; assignment stays on Agent Detail (ent#182: one skill model). **No new backend endpoints.** Legacy `/templates` redirects (query+hash-preserving). See [library-page.md](../feature-flows/library-page.md).
 
-**Agent Skills tab (trinity-enterprise#754):** one tab replaced Playbooks (the agent's own skills, with Run) and the library-only Skills tab. `components/skills/SkillsTab.vue` renders **Own** and **Shared** sections of fixed-size `SkillCard.vue`s (every slot a fixed height, so content never resizes a card; verb errors replace the description slot); the library picker is `SkillAssignModal.vue`, sets open in a `BaseModal` around the unchanged `AgentSkillSets.vue`, and the #2914 conflict / warnings / superseded text open in `SkillDetailsModal.vue`. Every rule — exactly-one-section, the Run predicate (running + LIVE list + `user_invocable`), the gate line (a kind, never a person), the role/ghost/system controls, the author-mode chip relabel — is the pure `utils/skillCards.js`; data is `stores/skills.js` (+ the agent's own listing, `loadAgentList` over `GET /playbooks?last_known=true`, and `runSkill` with `async_mode` because `api.js` times out at 30 s) and the agent-scoped `stores/skillGates.js` (map first, hook probe after; write-then-reload keyed to the agent written). The tab is visible to everyone with access; `TAB_ALIASES` (now exported from `utils/agentTabs.js`) sends `?tab=playbooks` to it. `BaseBadge` gained `size="sm"` (11px) for the mode chip. `skills/ExecutionGateMarker.vue` marks a self-approved run on the Tasks row, the execution page and the Workspace reply — see [skills-tab.md](../feature-flows/skills-tab.md).
+**Agent Skills tab (trinity-enterprise#754):** one tab replaced Playbooks (the agent's own skills, with Run) and the library-only Skills tab. `components/skills/SkillsTab.vue` renders **Own** and **Shared** sections of fixed-size `SkillCard.vue`s (every slot a fixed height, so content never resizes a card; verb errors replace the description slot); the library picker is `SkillAssignModal.vue` (its draft rules — Select all / Deselect all over the listed rows only, and the confirm before a save drops an approval requirement — are the pure `utils/skillAssignDraft.js`, trinity-enterprise#539), sets open in a `BaseModal` around the unchanged `AgentSkillSets.vue`, and the #2914 conflict / warnings / superseded text open in `SkillDetailsModal.vue`. Every rule — exactly-one-section, the Run predicate (running + LIVE list + `user_invocable`), the gate line (a kind, never a person), the role/ghost/system controls, the author-mode chip relabel — is the pure `utils/skillCards.js`; data is `stores/skills.js` (+ the agent's own listing, `loadAgentList` over `GET /playbooks?last_known=true`, and `runSkill` with `async_mode` because `api.js` times out at 30 s) and the agent-scoped `stores/skillGates.js` (map first, hook probe after; write-then-reload keyed to the agent written). The tab is visible to everyone with access; `TAB_ALIASES` (now exported from `utils/agentTabs.js`) sends `?tab=playbooks` to it. `BaseBadge` gained `size="sm"` (11px) for the mode chip. `skills/ExecutionGateMarker.vue` marks a self-approved run on the Tasks row, the execution page and the Workspace reply — see [skills-tab.md](../feature-flows/skills-tab.md).
 
 **Library tabs + fleet assignment read (trinity-enterprise#384):** the three stacked sections become an `OverflowTabs` strip — **Agent Templates · Systems · Skills** — deliberately reversing ent#263's stacked-sections choice for *all* sections at once rather than special-casing Skills (a tabs-plus-stacked hybrid would leave the page with two competing models). `?tab=` URL state, with a `route.query.tab` watch so the RENDER follows the URL for any navigation that changes it — an external link, a deep link, a history entry. Reading the query once at setup (`Operations.vue`'s shape) leaves those cases changing the address bar while the panel stays put. Tab *clicks* use `router.replace`, so switching tabs deliberately does not push history entries and Back leaves the page rather than walking back through tabs; the legacy `#agent-templates`/`#systems`/`#skills` anchors migrate to the matching tab and are then cleared, with `?tab=` winning over a hash. The Systems tab stays creator-gated with **both** arms of a late-role watch, since `stores/auth.js` reports `user` until `/api/users/me` lands and a creator hard-loading `?tab=systems` would otherwise land on Templates and never recover. Panels **lazy-mount-once** (`v-if` visited + `v-show` active): a plain `v-if` refetches status+library+assignments on every switch — the #1109 teardown rationale does not apply, `stores/skillsLibrary.js` owns no poll — while a plain `v-show` mounts Skills and Systems for a Templates-only visitor; it also preserves `SystemInstallPanel`'s editor state. Each skill block now lists **which agents already hold it** (chips → `/agents/{name}?tab=skills`, bounded with a counted overflow), plus a bounded **orphaned-assignments** list for rows whose skill left the library — ent#237's revocation model is "cut a new tag without the offending skill", after which the operator's first question is *who still has it*, and a page keyed by the library listing answers "nothing" forever. Assignment itself stays a **write on the per-agent Skills tab** (ent#182: one skill model); the Library gained a read, not a second write path — the assign/unassign writes were split to ent#386 and delivery semantics remain ent#385. **OSS-core by decision (ent#384): deliberately ungated — no `requires_entitlement`, logic stays in the OSS tree.** Recorded explicitly because CLAUDE.md's default for an enterprise-tracker feature is *gated unless ruled otherwise*, so the ruling must never be inferred later from the mere fact that it merged (the ent#326 discipline). Rationale: generic fleet telemetry over OSS tables, on a page and a skills stack that already shipped OSS-core (ent#183/#236/#237/#263/#126). See [library-page.md](../feature-flows/library-page.md).
 
diff --git a/docs/memory/feature-flows/skill-assignment.md b/docs/memory/feature-flows/skill-assignment.md
index 36c5b279ec..4f98fee41f 100644
--- a/docs/memory/feature-flows/skill-assignment.md
+++ b/docs/memory/feature-flows/skill-assignment.md
@@ -6,6 +6,7 @@
 
 | Date | Changes |
 |------|---------|
+| 2026-10-22 | **trinity-enterprise#539** — the Assign skills dialog gains **Select all** / **Deselect all**, which build the draft over the listed rows only (rules in `utils/skillAssignDraft.js`; set rows, deprecated skills and unlisted assignments are left as they are), and **Save** confirms first when it would drop an approval requirement. Writes unchanged: still the one bulk PUT. Line cites below re-pointed. See [skills-tab.md](skills-tab.md). |
 | 2026-10-08 | **trinity-enterprise#754** — the assignment UI is now the **Shared skills** section of the merged Skills tab ([skills-tab.md](skills-tab.md)), visible to everyone with access; the controls stay owner/admin-only. `SkillsPanel.vue` is deleted: assigned rows became cards, the library picker moved into `SkillAssignModal.vue` ("Assign skills"), the #2914 conflict / delivery warnings / superseded text open in `SkillDetailsModal.vue`, and sets show as chips with `AgentSkillSets.vue` in a "Manage sets" dialog. Routes and store unchanged. |
 | 2026-10-07 | **trinity-enterprise#753** — every route here that writes `agent_skills` (PUT, POST, DELETE, the inject Sync, POST/DELETE set) now also keeps the agent's skill gate map in line (`_sync_gates` → `skill_gate_map_service`), and answers with a `gates` block (`gate_defaults` applied/removed, `gates_removed`, `gates_kept`, or null). A library skill whose metadata says `approval: recommended` is gated BEFORE its package is delivered; a gate on an unassigned skill goes only once its package removal completed, and an explicit gate only when a person unassigned it. See [skill-gate.md](skill-gate.md) §6. |
 | 2026-09-30 | **ent#672** — a deprecated library skill is marked and its assignment says so. `SkillsPanel.vue` shows a `deprecated` badge + "Superseded by …" line on the assigned rows (an agent already holding one shows it on load) and on the picker rows (before the tick); `POST` / `PUT` still succeed and their `delivery.skills[name].warnings` carries `deprecated[:<successor>]`, rendered as a warning line under the save note; after a manual Sync the per-skill warning list says it in words. See "Deprecated skills" below and [skill-injection.md](skill-injection.md). |
@@ -29,7 +30,7 @@ As an agent owner, I want to assign platform skills to my agents so that they fo
 ## Entry Points
 
 ### User Configuration (UI)
-- **Component**: `src/frontend/src/components/skills/SkillAssignModal.vue:75-77` - "Save assignments" (bulk PUT), opened by **Assign skills** in the Skills tab's Shared section (`SkillsTab.vue:136`)
+- **Component**: `src/frontend/src/components/skills/SkillAssignModal.vue:107-130` - "Save assignments" (bulk PUT), with **Select all** / **Deselect all** on the list's header row (`:28-54`), opened by **Assign skills** in the Skills tab's Shared section (`SkillsTab.vue:136`)
 - **Tab Access**: Agent Detail page -> Skills tab, visible to everyone with access since trinity-enterprise#754; the assignment controls (Assign skills, Manage sets, Sync now, Unassign) are for the owner or an admin, never on the system agent. See [skills-tab.md](skills-tab.md)
 
 ### Backend API
@@ -49,7 +50,7 @@ Since trinity-enterprise#754 the assignment surface is the **Shared skills** sec
 
 | Control | File | Call |
 |---|---|---|
-| **Assign skills** → tick → **Save assignments** | `components/skills/SkillAssignModal.vue:134-146` | `store.saveAssignments(draft)` → `PUT /api/agents/{name}/skills` |
+| **Assign skills** → tick (or **Select all** / **Deselect all**) → **Save assignments** | `components/skills/SkillAssignModal.vue:236-337` (a save that drops an approval requirement confirms first, `:290-323`) | `store.saveAssignments(draft)` → `PUT /api/agents/{name}/skills` |
 | **Unassign** on a Shared card | `components/skills/SkillsTab.vue:546-564`, after its confirm (`:566-578`) | `store.saveAssignments(individualNames − name)` → the same PUT |
 | **Unassign library skill** (#2914 conflict) | `components/skills/SkillDetailsModal.vue:31-41` → `SkillsTab.onUnassign` | the same PUT |
 | **Manage sets** | `components/skills/AgentSkillSets.vue` in a `BaseModal` (`SkillsTab.vue:271-276`) | `store.assignSet` / `unassignSet` → `POST` / `DELETE /api/agents/{name}/skill-sets/{set}` |
@@ -63,7 +64,8 @@ The picker keeps its ent#235 / ent#530 / ent#672 / #2914 behaviour: the draft is
 
 - `library`, `libraryStatus`, `assigned` (the rows, with `individual` / `via_sets` / `delivery_status`), `sets`, `librarySets`; `injectionResults` + `lastInjectionAt` (the last Sync, per skill); `lastDelivery` (the last save's #2703 report).
 - Derived: `individualNames` (the draft's source — what the bulk PUT replaces), `setOnlyNames` (locked rows), `conflictNames` (#2914), `emptyReason` (`library_unconfigured` / `library_empty` / `none_assigned`).
-- The picker's dirty check compares the sorted draft with `individualNames` (`SkillAssignModal.vue:120-124`); the draft resets only when that set changes (`:132`).
+- The picker's dirty check compares the sorted draft with `individualNames` (`SkillAssignModal.vue:197-201`); the draft resets only when that set changes (`:209`).
+- Select all / Deselect all (trinity-enterprise#539) are pure rules in `utils/skillAssignDraft.js` (`bulkSelection`, read at `SkillAssignModal.vue:223-230`): they change only the draft, act only on the listed rows, never move a set row, never tick a deprecated skill or a name the PUT refuses, and keep an assignment the dialog does not list. Before the PUT, a save that removes anything re-reads the gate map and `gatedRemovals` decides whether Save confirms (`:290-318`).
 
 ### API Calls
 
@@ -79,7 +81,7 @@ lastDelivery.value = saved?.delivery ?? null
 const { data } = await api.get(`/api/agents/${agentName.value}/skills`)   // re-read the rows
 ```
 
-The dialog then emits `saved` with `deliveryText(lastDelivery, {saved: true})` and `deprecationText(lastDelivery)` (`SkillAssignModal.vue:134-146`); the tab shows it on the Shared line and arms the Sync emphasis only when delivery did not land.
+The dialog then emits `saved` with `deliveryText(lastDelivery, {saved: true})` and `deprecationText(lastDelivery)` (`SkillAssignModal.vue:325-337`); the tab shows it on the Shared line and arms the Sync emphasis only when delivery did not land.
 
 **`inject()`** (`stores/skills.js:403-432`): `POST /api/agents/{name}/skills/inject` (a repair: `force` server-side). Results are stored per skill (`injectionResults`), then the rows are re-read so a #2914 verdict that was set or cleared shows at once. A 409 reads "A skill sync is already running for this agent. Try again in a moment."
 
diff --git a/docs/memory/feature-flows/skills-tab.md b/docs/memory/feature-flows/skills-tab.md
index 0fb9189fe1..d2d10962ff 100644
--- a/docs/memory/feature-flows/skills-tab.md
+++ b/docs/memory/feature-flows/skills-tab.md
@@ -33,7 +33,8 @@ As an agent owner, I want to see, run and gate every skill my agent has — its
 |---|---|
 | `components/skills/SkillsTab.vue` | The tab: filter, in-agent enforcement line, Own section, Shared section, the three dialogs |
 | `components/skills/SkillCard.vue` | One card with fixed-height slots; renders decisions, emits verbs |
-| `components/skills/SkillAssignModal.vue` | "Assign skills": the library picker, moved out of the deleted `SkillsPanel.vue` |
+| `components/skills/SkillAssignModal.vue` | "Assign skills": the library picker, moved out of the deleted `SkillsPanel.vue`; **Select all** / **Deselect all** on the list's header row and the confirm before a save drops an approval requirement (trinity-enterprise#539) |
+| `utils/skillAssignDraft.js` | The dialog's draft rules, pure: `bulkSelection` (what each bulk button would change, and its disabled reason), `withSelectAll` / `withDeselectAll`, `gatedRemovals` (whether Save confirms), `removalConfirm` (its words) |
 | `components/skills/SkillDetailsModal.vue` | A Shared card's full story: the #2914 conflict + "Unassign library skill", every delivery warning, the injection error, the superseded text |
 | `components/skills/AgentSkillSets.vue` | Unchanged; shown inside the "Manage sets" `BaseModal` |
 | `components/skills/ExecutionGateMarker.vue` | "Ran without approval" — `badge` (list row) or `line` (detail, Workspace turn) |
@@ -48,6 +49,7 @@ As an agent owner, I want to see, run and gate every skill my agent has — its
   - meta line, last-known (warning tone): "The agent is stopped. Start the agent to run. Showing its skills as of <relative time>." or "The agent isn't answering. Showing its skills as of …"; the absolute time is on hover (`Listed <date>`).
 - **Shared skills** (`:127-267`):
   - head: count (`store.assigned.length`); for the owner or an admin (not on the system agent), on a configured library: **Assign skills** (opens `SkillAssignModal`), **Manage sets** (opens `AgentSkillSets` in a `BaseModal`), **Sync now** (disabled while stopped; primary-styled while a save did not deliver — `pendingSync`, #2703; `store.inject()`);
+  - the Assign skills dialog (`SkillAssignModal.vue`, trinity-enterprise#539): the list's header row carries "N skills in the library" and two ghost buttons. **Select all** ticks every listed skill except deprecated ones (a held deprecated skill keeps its tick), never a set's locked row and never a name the PUT refuses (`SKILL_NAME_RE`). **Deselect all** unticks every listed skill except a set-only one. Both change only the draft, and only listed rows: an assignment whose skill left the library stays, and its card's Unassign removes it. Each is disabled with its reason as the title (not loaded, no library, empty library, a save in flight, nothing to change), and enabled titles count what the click changes. Deselect all carries `data-destructive`, so the dialog never opens on it. Focus never falls out of the dialog: a `flush: 'post'` watcher over the bulk buttons' disabled states, `dirty`, `saving` and the confirm moves focus off a control that was just disabled or removed (to the bulk button that undoes it, else Close), whatever disabled it: its own click, a manual tick, Reset, Save or the confirm closing. **Save** asks first (a `ConfirmDialog`, opening on Cancel) when it would take a skill with an approval requirement off the agent, because a person's unassign drops it (trinity-enterprise#753). A save that removes anything re-reads the gate map first, so a requirement added since the tab opened counts. With the map unreadable it asks for any removal. A name an assigned set still holds does not ask (unless a held set is unresolved, when the server keeps only set-only rows), and neither does a `conflict` name;
   - set chips: up to 3 (`MAX_SET_CHIPS`) + "+N more"; purple when complete, warning with a dot when partial, unresolved or missing credentials (`setVariant` / `setTitle`), plus "Some sets need credentials" when any set lacks them; a failed sets read shows "Couldn't read this agent's sets" with a Retry (`store.loadSets()`) instead of an absent line; a click opens Manage sets;
   - the shared line (`:194-204`): the last verb's outcome (save / unassign / sync error, plus the ent#672 deprecation note), else "Last sync <time>" or "Not synced from this screen yet: statuses appear after a sync. Skills are also copied in when the agent starts.";
   - body: skeleton / `LoadFailed` / the named empty states by `store.emptyReason` (`library_unconfigured` with an admin link to `/settings?tab=agents`, `library_empty`, `none_assigned` worded by role) / "No shared skill matches the filter." / the grid.
@@ -353,6 +355,9 @@ A skipped field is warned once per `(file, field, value)` (`_SKIPPED_FIELD_WARNE
 | Run, other failure | e.g. 429 / 503 | `InlineError` on the card |
 | Sync while one runs | 409 | Shared line: "A skill sync is already running for this agent. Try again in a moment." |
 | Unassign fails | Server error | `InlineError` on the card |
+| Assign dialog: the gate map can't be re-read at Save | `GET /skill-gates` fails | Save asks for any removal: "Which skills need approval couldn't be read, so any approval requirement on them will go too." |
+| Assign dialog: the save fails after a bulk verb | PUT error | The error line in the dialog; the draft is kept, Save stays enabled |
+| Assign dialog: the library lists a name the PUT refuses | (client-side, `SKILL_NAME_RE`) | Select all skips it and its title says so; a manual tick gets the named 422 `invalid_skill_name` |
 | Hook probe, no answer | `hook: "unknown"` | Nothing (not "not ok") |
 | Self-approved read fails | Flags false (logged) | No marker |
 | `record_self_approval` write failed | No record | No marker. Named gap: the in-container hook refuses the skill inside that run anyway |
@@ -444,6 +449,8 @@ Covered by unit tests:
 | `src/frontend/tests/unit/skillGateSenders.spec.js` | Ported: the SkillsTab Run among the UI senders of a 202 / refusal (trinity#3274) |
 | `src/frontend/tests/unit/skillSets.spec.js`, `skillDeprecation.spec.js` | Ported to `SkillsTab` / the dialogs: sets, chips and Manage sets; deprecated badge, superseded line, the conflict's way out in the details dialog |
 | `src/frontend/tests/unit/skillAssignDraftSurvivesSync.spec.js` | Ported from `skillsPanelDraftSurvivesSync`: the picker draft survives a sync (`SkillAssignModal`) |
+| `src/frontend/tests/unit/skillAssignDraft.spec.js` | trinity-enterprise#539, the pure rules: what each bulk button adds or removes over every row type (plain, deprecated, a set's row, a refused name, an unlisted assignment); every disabled reason and its precedence; the counted titles, and the refused-name clause; what a save removes (set-held kept, unless the sets are untrusted); which saves confirm (gated, case-insensitive; never set-held or `conflict`, with the map read or not; any other removal with the map unread); the confirm's words |
+| `src/frontend/tests/unit/skillAssignSelectAll.spec.js` | trinity-enterprise#539, mounted through the real Teleport (a stubbed one re-creates the dialog's nodes on every patch, which made focus tests pass for the wrong reason): no write on a bulk click; the PUT body after each verb (the unlisted assignment kept, the set's row never sent); Reset after each verb; each disabled reason on the button; the dialog opens on Select all, never Deselect all; focus re-homed after a bulk click, a manual tick, Reset, Save and Confirm; the confirm names the skill, opens on Cancel and sends nothing until confirmed; the gate map re-read at save time (not for a save that only adds), a failed re-read, an unresolved set, a double click and a dialog closed mid-read; a failed save keeps the draft. 27 mutants, all killed |
 | `src/frontend/tests/unit/agentDetailTabSuperset.spec.js` | `playbooks` is an alias to `skills`; `skills` is offered to a non-owner |
 | `src/frontend/tests/unit/agentDetailGateNotice.spec.js` | `<SkillsTab>` receives the page's toast host (`:notify`) |
 | `tests/unit/test_1560_agent_redis_key_parity.py` | Existing: every `agent:` key literal is registered (`agent:skills_list:` in `CLEARED_KEYSPACES`) |
@@ -455,7 +462,8 @@ pytest tests/unit/test_ent754_*.py tests/unit/test_1560_agent_redis_key_parity.p
 # frontend
 cd src/frontend && npx vitest run tests/unit/skillCards.spec.js tests/unit/skillGatesStore.spec.js \
   tests/unit/skillsTab.mount.spec.js tests/unit/executionGateMarker.mount.spec.js tests/unit/portalSelfApprovedTurn.spec.js \
-  tests/unit/tasksPanelRunRefresh.spec.js tests/unit/baseSelectSize.spec.js
+  tests/unit/tasksPanelRunRefresh.spec.js tests/unit/baseSelectSize.spec.js \
+  tests/unit/skillAssignDraft.spec.js tests/unit/skillAssignSelectAll.spec.js
 ```
 
 ### Status
@@ -465,7 +473,11 @@ Implemented (2026-10-08).
 Known limits:
 - the Overview tab's "N skills" button counts library assignments only;
 - a renamed agent that is stopped shows no own skills until it starts: rename clears the last-known copy under both names rather than moving it (PR review, deferred);
-- Shared cards show no file count or size, which the old rows did; both stay in the Assign dialog (PR review, kept).
+- Shared cards show no file count or size, which the old rows did; both stay in the Assign dialog (PR review, kept);
+- in the Assign dialog, the checkboxes stay editable while a save is in flight (only the bulk buttons wait), and a successful save resets the draft, so a tick made meanwhile is lost (trinity-enterprise#539, stated limit);
+- Save's confirm says the approval requirement goes with the skill, as the card's Unassign confirm does; on a stopped or busy agent the package removal defers, and a requirement set by hand then stays until someone clears it (security.md §26.13; stated limit, worded as the card's);
+- the library lists every directory with a SKILL.md, including names the PUT refuses: Select all skips them, a manual tick still gets the named 422 (stated limit; the listing is shared with the Library page and sets);
+- a disabled bulk button's reason is its hover title only, like Sync now's.
 
 ## Related Flows
 
@@ -483,3 +495,4 @@ Known limits:
 | 2026-10-08 | Review round (ent#754): agent-switch guards in the tab and the skills store; sections wait for the reads they depend on and name each failure; gate writes keyed on the gate the card shows; a live default approver; Unassign confirms; dialogs focus the safe action; hook states in words. Backend: a container removed mid-read is a plain 404 (no Docker text, public link included); transport errors are unreachable; the connector keeps its wording; the self-approved read is chunked; a machine Workspace viewer is never "you"; the synchronous Workspace reply carries both flags; `GET /skills/library` names `approval`. |
 | 2026-10-08 | Eyeball round (ent#754): a Run's Tasks row settles without a Refresh (the poll re-reads while a row is in flight; the highlighted row opens once); the Own meta line names each folder once (store, and the agent server scans its folder once); the Unassign confirm speaks in the future tense; the approver picker is `BaseSelect size="sm"`, so it fits the 40px row; the Own count includes kept gates; the not-synced line says statuses appear after a sync. |
 | 2026-10-09 | PR review (#3412): an unsaved Assign draft no longer carries to the next agent; the Tasks poll stops while the page is away; the public link answers fixed sentences and keeps to its listing keys; a Docker daemon fault serves the kept copy (tri-state lookup); the probe is capped at 3 s overall; a switch reads the next agent once; the text links take the focus ring; "people only" includes a person's user-scoped key; line citations re-pointed. |
+| 2026-10-22 | trinity-enterprise#539: Select all / Deselect all in the Assign dialog (rules in `utils/skillAssignDraft.js`; listed rows only, set rows and unlisted assignments untouched, deprecated skills never ticked in bulk), the confirm before a save drops an approval requirement, and focus kept in the dialog after a bulk verb or Reset. |
diff --git a/docs/memory/learnings/2026-10-22-a-control-that-disables-itself-drops-modal-focus.md b/docs/memory/learnings/2026-10-22-a-control-that-disables-itself-drops-modal-focus.md
new file mode 100644
index 0000000000..99e8cb7aea
--- /dev/null
+++ b/docs/memory/learnings/2026-10-22-a-control-that-disables-itself-drops-modal-focus.md
@@ -0,0 +1,23 @@
+## 2026-10-22 — pitfall — A control that is disabled or removed while it holds focus drops focus out of the modal, and per-verb fixes miss most of the triggers
+**Context**: trinity-enterprise#539, `components/skills/SkillAssignModal.vue`. Several controls in this dialog can disable or remove themselves while they hold focus:
+- Select all is left with nothing to add and Deselect all with nothing to remove, so each disables itself on every click.
+- Reset is `v-if="dirty"`, so it unmounts.
+- Save is disabled while it is `loading`.
+- The confirm's buttons leave with its overlay.
+
+A focused control that becomes disabled or leaves the DOM drops focus to `<body>`, outside the overlay that owns `BaseModal`'s `@keydown`: Esc stops closing the dialog and Tab is no longer trapped.
+
+The first fix was one handoff per verb. Review found the triggers it missed:
+- A manual tick: Safari and Firefox on macOS do not move focus on a mouse click, so focus stays on Select all.
+- Any save, which disables both bulk buttons.
+- Confirm, after which the confirm hands focus back to a Save that the save has just disabled.
+
+That first version also called `nextTick(cb)` before writing the state. With no flush pending, `nextTick` resolves on the already-resolved promise, so `cb` ran before the render and still read the old DOM. A re-read found it, and a mutation proved the test caught that ordering.
+
+**Lesson**:
+1. Re-home focus in ONE `flush: 'post'` watcher over every state that can disable or remove a control. Do not patch it per verb.
+2. Treat focus as lost in two cases: `document.activeElement` is a disabled control inside the dialog, or focus is on `<body>` and a control inside the dialog held it last.
+3. Track that last control with a `focusin` listener on the overlay, and reset it when the dialog opens.
+4. Send focus to the control that undoes the lost one, else to a safe control that is always present (Close).
+5. Skip the watcher while a nested confirm is open, because focus is legitimately in it.
+6. If you do schedule with `nextTick`, call it after the reactive write, never before.
diff --git a/docs/memory/learnings/2026-10-22-a-teleport-stub-recreates-nodes-on-every-patch.md b/docs/memory/learnings/2026-10-22-a-teleport-stub-recreates-nodes-on-every-patch.md
new file mode 100644
index 0000000000..d895d092f1
--- /dev/null
+++ b/docs/memory/learnings/2026-10-22-a-teleport-stub-recreates-nodes-on-every-patch.md
@@ -0,0 +1,11 @@
+## 2026-10-22 — pitfall — `stubs: { teleport: true }` re-creates the stubbed subtree on every patch, so a test of what happens to an element "in place" passes for the wrong reason
+**Context**: trinity-enterprise#539, `src/frontend/tests/unit/skillAssignSelectAll.spec.js`. The mount harness copied from `skillAssignDraftSurvivesSync.spec.js` stubs Teleport inline. Under that stub, every patch replaced the dialog's nodes, so a button that a click disables was re-created instead of being disabled in place.
+
+The focus-handoff code had two branches: "the control is gone" and "the control is disabled". The tests only ever took the first. A mutation that broke the in-place branch, which is the real browser path, left every test green. With the real Teleport it turned two tests red. An independent review found this, and a node-identity assertion (`expect(find(...).element).toBe(before)`) reproduced it: it fails under the stub and passes with the real Teleport.
+
+**Lesson**: A test that reasons about a DOM node's identity or its in-place state must mount with the real Teleport. That covers focus on a control that gets disabled, a node kept across a re-render, and scroll or selection kept on an element.
+- Query through `new DOMWrapper(document.body)`, because teleported content is not under the wrapper.
+- Use `enableAutoUnmount(afterEach)`, so dialogs from earlier tests are not left in `<body>`.
+- Pin the premise with one identity assertion before the behaviour assertion.
+
+The stub is fine for tests that only read state, such as whether a box is checked or what text is shown.
diff --git a/docs/memory/requirements/skills.md b/docs/memory/requirements/skills.md
index 98c5fec286..a5408232fe 100644
--- a/docs/memory/requirements/skills.md
+++ b/docs/memory/requirements/skills.md
@@ -329,6 +329,8 @@ entries (`actor_type=system` for the automated paths).
 - **Description**: Per-agent skill assignment from the platform library, on Agent Detail
 - **Key Features**: assigned-skills list with honest per-skill injection status (§21.4 results incl. warnings), assign/unassign against the library list, manual re-inject; agent-scoped store `stores/skills.js` (its `emptyReason` discriminator is agent-scoped — do not reuse it on fleet surfaces)
 - **Deprecated skills (trinity-enterprise#672)**: a skill the library marks `deprecated` (§21.6) carries a warning `deprecated` badge beside its name and a "Superseded by …" line on both lists — the assigned rows (so an agent already holding one shows it on load, with nothing else about it changed) and the library picker (so it is visible before ticking). Saving a draft that adds one succeeds, and a warning line under the save note names the skill and, where the successor is a skill name, what supersedes it — on every delivery outcome, a stopped agent included. A manual sync's per-skill warning list leaves the code out — that list is for what went wrong with the delivery, and the row's badge already says it. Never hidden, never refused, never a confirm dialog
+- **Select all / Deselect all (trinity-enterprise#539)**: the Assign skills dialog's list header carries the library count and two buttons. **Select all** ticks every listed skill except deprecated ones (a deprecated skill keeps the tick it has), never a skill a set assigns (it stays ticked and locked), and never a name the platform cannot assign. **Deselect all** unticks every listed skill except a set's. Both change only the draft: Save still writes, and Reset restores the saved set. Both act only on the rows the dialog lists, so an assignment whose skill left the library is kept (its card's Unassign removes it), and a filter, if one arrives, narrows both. Each is disabled, with the reason on hover, while the library is loading, unconfigured or empty, during a save, and when it has nothing to change. The dialog never opens on Deselect all, and focus never falls out of the dialog when the control holding it is disabled or removed: it moves to the other bulk button, else to Close
+- **Removal confirm (trinity-enterprise#539)**: Save asks first when it would take a skill that has an approval requirement off the agent, because a person's unassign drops it once the package has left the agent (security.md §26.13, trinity-enterprise#753). The confirm names the skills, speaks in the future tense and opens on Cancel. A save that removes anything reads the approval map fresh first; when it cannot be read, any removal asks. A skill an assigned set still holds (while every held set resolves), or one whose library copy is in `conflict`, keeps its requirement and does not ask
 - **History**: originally shipped hidden from `visibleTabs`; ent#235 rebuilt it and made it visible
 
 ### 22.3 Library Page — Fleet Skills Browse (trinity-enterprise#263)
diff --git a/docs/user-docs/automation/skills-and-playbooks.md b/docs/user-docs/automation/skills-and-playbooks.md
index 4eafa2d56b..9a8b47c1ec 100644
--- a/docs/user-docs/automation/skills-and-playbooks.md
+++ b/docs/user-docs/automation/skills-and-playbooks.md
@@ -98,7 +98,7 @@ Both surfaces write the same per-agent assignment, so there is one skill model w
 
 #### From the agent's Skills tab
 
-Library skills assigned to the agent are the cards under **Shared skills**, each with its version short-SHA, description, and the outcome of its last delivery. The owner or an admin manages them from the section's header. **Assign skills** opens the library picker: everything available, with its declared dependencies. Tick or untick skills and click **Save assignments**. The save delivers straight away, and the note under the button says what happened:
+Library skills assigned to the agent are the cards under **Shared skills**, each with its version short-SHA, description, and the outcome of its last delivery. The owner or an admin manages them from the section's header. **Assign skills** opens the library picker: everything available, with its declared dependencies. Tick or untick skills and click **Save assignments**. **Select all** and **Deselect all**, above the list, tick or untick every listed skill in one click. Nothing is saved until you click **Save assignments**, and **Reset** puts the ticks back. Select all leaves deprecated skills as they are, so tick one by hand to assign it. Neither button changes a skill that a set assigns, or an assignment whose skill is no longer in the library: remove that one from its card. A save that would remove a skill that needs approval asks first, because its approval requirement goes with it. The save delivers straight away, and the note under the button says what happened:
 
 | Note | Meaning |
 |------|---------|
diff --git a/src/frontend/src/components/skills/SkillAssignModal.vue b/src/frontend/src/components/skills/SkillAssignModal.vue
index 91099f873c..75a4c78f24 100644
--- a/src/frontend/src/components/skills/SkillAssignModal.vue
+++ b/src/frontend/src/components/skills/SkillAssignModal.vue
@@ -7,31 +7,64 @@
     list the bulk PUT replaces; a set-only member is ticked and locked; a
     deprecated skill is marked before it is ticked; a draft survives a sync
     that re-reads the same assignment set.
+    Select all / Deselect all build the draft in one click, over the rows
+    listed here only; Save asks first when it would drop an approval
+    requirement. The rules are utils/skillAssignDraft.js.
   -->
   <BaseModal
     :model-value="modelValue"
     aria-label="Assign shared skills"
     panel-class="relative w-full max-w-xl rounded-lg bg-white p-6 shadow-xl dark:bg-gray-800"
     @update:model-value="$emit('update:modelValue', $event)"
+    @focusin="onFocusIn"
   >
     <h3 class="text-[18px] font-[650] text-gray-900 dark:text-gray-100">Assign shared skills</h3>
     <p class="mt-1 text-[12.5px] text-gray-600 dark:text-gray-300">
-      {{ store.library.length }} skill{{ store.library.length === 1 ? '' : 's' }} in the library.
       Tick to assign, then save. Assigned skills are copied into the agent now if it is running, or when it next starts.
     </p>
 
+    <!-- The list's header row. Both buttons are always drawn, and only their
+         disabled state changes, so nothing here moves the list. -->
+    <div class="mt-3 flex flex-wrap items-center justify-between gap-x-3 gap-y-1">
+      <span class="text-[12.5px] tabular-nums text-gray-600 dark:text-gray-300" data-testid="skill-assign-count">
+        {{ listed.length }} skill{{ listed.length === 1 ? '' : 's' }} in the library
+      </span>
+      <div class="flex items-center gap-1">
+        <BaseButton
+          ref="selectAllButton"
+          variant="ghost"
+          size="sm"
+          :disabled="bulk.select.disabled"
+          :title="bulk.select.title"
+          data-testid="skill-assign-select-all"
+          @click="selectAll"
+        >Select all</BaseButton>
+        <!-- The removal verb: never where the dialog opens (data-destructive). -->
+        <BaseButton
+          ref="deselectAllButton"
+          variant="ghost"
+          size="sm"
+          :disabled="bulk.deselect.disabled"
+          :title="bulk.deselect.title"
+          data-testid="skill-assign-deselect-all"
+          data-destructive
+          @click="deselectAll"
+        >Deselect all</BaseButton>
+      </div>
+    </div>
+
     <ul
-      class="mt-3 max-h-[50vh] overflow-y-auto divide-y divide-gray-200 dark:divide-gray-750 rounded-lg border border-gray-200 dark:border-gray-750"
+      class="mt-2 max-h-[50vh] overflow-y-auto divide-y divide-gray-200 dark:divide-gray-750 rounded-lg border border-gray-200 dark:border-gray-750"
       data-testid="skill-assign-list"
     >
-      <li v-for="s in store.library" :key="s.name" class="px-4 py-3">
+      <li v-for="s in listed" :key="s.name" class="px-4 py-3">
         <label
           class="flex items-start gap-3"
-          :class="store.setOnlyNames.has(s.name) ? 'cursor-not-allowed' : 'cursor-pointer'"
-          :title="store.setOnlyNames.has(s.name) ? `Assigned via ${store.viaSets(s.name).join(', ')} — unassign the set to remove it` : undefined"
+          :class="isLocked(s.name) ? 'cursor-not-allowed' : 'cursor-pointer'"
+          :title="isLocked(s.name) ? `Assigned via ${store.viaSets(s.name).join(', ')} — unassign the set to remove it` : undefined"
         >
           <input
-            v-if="store.setOnlyNames.has(s.name)"
+            v-if="isLocked(s.name)"
             type="checkbox"
             checked
             disabled
@@ -72,11 +105,28 @@
     </ul>
 
     <div class="mt-4 flex items-center gap-3">
-      <BaseButton :disabled="!dirty" :loading="store.saving" loading-label="Saving…" @click="onSave">
+      <BaseButton
+        :disabled="!dirty"
+        :loading="store.saving"
+        loading-label="Saving…"
+        data-testid="skill-assign-save"
+        @click="onSaveClick"
+      >
         Save assignments
       </BaseButton>
-      <BaseButton v-if="dirty" variant="ghost" @click="resetDraft">Reset</BaseButton>
-      <BaseButton variant="secondary" class="ml-auto" @click="$emit('update:modelValue', false)">Close</BaseButton>
+      <BaseButton
+        v-if="dirty"
+        variant="ghost"
+        data-testid="skill-assign-reset"
+        @click="resetDraft"
+      >Reset</BaseButton>
+      <BaseButton
+        ref="closeButton"
+        variant="secondary"
+        class="ml-auto"
+        data-testid="skill-assign-close"
+        @click="$emit('update:modelValue', false)"
+      >Close</BaseButton>
     </div>
     <p
       v-if="notice && notice.text"
@@ -91,6 +141,18 @@
       class="mt-1 text-[12.5px] text-status-warning-700 dark:text-status-warning-400"
     >{{ notice.deprecation }}</p>
     <p v-if="store.error" class="mt-2 text-[12.5px] text-status-danger-700 dark:text-status-danger-400">{{ store.error }}</p>
+
+    <!-- Principle 19: a save that drops an approval requirement restates it
+         first, and opens on Cancel (the card's Unassign does the same). -->
+    <ConfirmDialog
+      :visible="!!pendingRemoval"
+      :title="removalCopy.title"
+      :message="removalCopy.message"
+      :confirm-text="REMOVAL_CONFIRM_TEXT"
+      variant="warning"
+      @confirm="confirmRemoval"
+      @cancel="pendingRemoval = null"
+    />
   </BaseModal>
 </template>
 
@@ -99,10 +161,15 @@ import { computed, ref, watch } from 'vue'
 import BaseBadge from '../base/BaseBadge.vue'
 import BaseButton from '../base/BaseButton.vue'
 import BaseModal from '../base/BaseModal.vue'
+import ConfirmDialog from '../ConfirmDialog.vue'
 import SkillContractChips from './SkillContractChips.vue'
 import { deps, supersededLine, DEPRECATED_TITLE } from './contract'
 import { useSkillsStore } from '../../stores/skills'
+import { useSkillGatesStore } from '../../stores/skillGates'
 import { deliveryText, deprecationText } from '../../utils/skillDelivery'
+import {
+  bulkSelection, withSelectAll, withDeselectAll, removedNames, gatedRemovals, removalConfirm, REMOVAL_CONFIRM_TEXT,
+} from '../../utils/skillAssignDraft'
 
 const props = defineProps({
   modelValue: { type: Boolean, default: false },
@@ -114,7 +181,17 @@ const props = defineProps({
 const emit = defineEmits(['update:modelValue', 'saved'])
 
 const store = useSkillsStore()
+const gatesStore = useSkillGatesStore()
 const draft = ref([...store.individualNames])
+// A save that would take a skill with an approval requirement off the agent
+// asks first (a person's unassign drops it, trinity-enterprise#753):
+// {names, gatesKnown} while the confirm is open.
+const pendingRemoval = ref(null)
+
+// The rows this dialog lists. The list and both bulk buttons read this one
+// value, so a filter added later narrows all three at once.
+const listed = computed(() => store.library)
+const isLocked = (name) => store.setOnlyNames.has(name)
 
 // ent#530: the draft is the INDIVIDUAL list — what the bulk PUT replaces.
 const dirty = computed(() => {
@@ -134,7 +211,116 @@ watch(() => [...store.individualNames].sort().join('|'), resetDraft)
 // outlives an agent switch, and two agents with the same assignments give the
 // watcher above nothing to see, so an unsaved tick would be offered, and
 // saved, on the next agent. Closing without saving discards it.
-watch(() => props.modelValue, (open) => { if (open) resetDraft() })
+watch(() => props.modelValue, (open) => {
+  if (!open) return
+  resetDraft()
+  pendingRemoval.value = null
+  lastFocus = null
+})
+
+// trinity-enterprise#539: Select all / Deselect all change the draft only;
+// Save still writes, and Reset still restores the saved set.
+const bulk = computed(() => bulkSelection({
+  rows: listed.value,
+  draft: draft.value,
+  locked: store.setOnlyNames,
+  loaded: store.sharedLoaded,
+  emptyReason: store.emptyReason,
+  saving: store.saving,
+}))
+
+const selectAllButton = ref(null)
+const deselectAllButton = ref(null)
+const closeButton = ref(null)
+
+function selectAll() {
+  draft.value = withSelectAll(draft.value, bulk.value.toAdd)
+}
+
+function deselectAll() {
+  draft.value = withDeselectAll(draft.value, bulk.value.toRemove)
+}
+
+// Focus can sit on a control this dialog then disables (a bulk button with
+// nothing left to change, every button while a save runs) or removes (Reset
+// once the draft is clean, the confirm's buttons). It then falls to <body>,
+// outside the overlay that owns Esc and Tab. After every change that can do
+// that, put it back: on the bulk button that undoes the one that lost it,
+// else on Close. `lastFocus` is the last control inside the dialog that held
+// focus, for when the one that had it is already gone.
+let lastFocus = null
+function onFocusIn(event) {
+  lastFocus = event.target
+}
+
+watch(
+  () => [bulk.value.select.disabled, bulk.value.deselect.disabled, dirty.value, store.saving, pendingRemoval.value],
+  () => {
+    const close = closeButton.value?.$el
+    if (!props.modelValue || pendingRemoval.value || !close) return
+    const active = document.activeElement
+    let from = null
+    if (active && active.disabled && close.closest('[role="dialog"]')?.contains(active)) from = active
+    else if ((!active || active === document.body) && lastFocus) from = lastFocus
+    if (!from) return
+    const select = selectAllButton.value?.$el
+    const deselect = deselectAllButton.value?.$el
+    const undo = from === select ? deselect : from === deselect ? select : null
+    ;(undo && !undo.disabled ? undo : close).focus()
+  },
+  { flush: 'post' },
+)
+
+const removalCopy = computed(() => (pendingRemoval.value
+  ? removalConfirm(pendingRemoval.value)
+  : { title: '', message: '' }))
+
+// Names an assigned set still holds, so a save only demotes them. Trusted
+// only when every held set resolved: with one unresolved, the server keeps
+// just the set-only rows (`db.set_agent_skills` fails closed).
+function setHeldNames() {
+  if (!store.setsLoaded || store.setsError || store.sets.some((s) => s.status === 'unresolved')) return new Set()
+  return new Set(store.assigned.filter((s) => s.via_sets?.length).map((s) => s.skill_name))
+}
+
+// One decision at a time: a second click while the gate map is re-read is the
+// same click.
+let deciding = false
+
+async function onSaveClick() {
+  if (deciding) return
+  const agent = store.agentName
+  if (removedNames({ saved: store.individualNames, draft: draft.value, setHeld: setHeldNames() }).length) {
+    // Decide on a fresh map. The tab read it when it opened, and a save (the
+    // library's default requirements), another session or an MCP call may
+    // have changed it since.
+    deciding = true
+    try {
+      await gatesStore.load(agent)
+    } finally {
+      deciding = false
+    }
+    if (agent !== store.agentName || !props.modelValue) return
+  }
+  const removal = gatedRemovals({
+    saved: store.individualNames,
+    draft: draft.value,
+    setHeld: setHeldNames(),
+    gated: new Set(gatesStore.gates.map((g) => String(g.skill_name || '').toLowerCase())),
+    conflict: store.conflictNames,
+    gatesKnown: gatesStore.hasLoaded && !gatesStore.error,
+  })
+  if (removal) {
+    pendingRemoval.value = removal
+    return
+  }
+  onSave()
+}
+
+function confirmRemoval() {
+  pendingRemoval.value = null
+  onSave()
+}
 
 async function onSave() {
   if (await store.saveAssignments([...draft.value])) {
diff --git a/src/frontend/src/utils/skillAssignDraft.js b/src/frontend/src/utils/skillAssignDraft.js
new file mode 100644
index 0000000000..ae846dfac6
--- /dev/null
+++ b/src/frontend/src/utils/skillAssignDraft.js
@@ -0,0 +1,160 @@
+/**
+ * The Assign skills dialog's draft rules (trinity-enterprise#539):
+ * Select all / Deselect all, and whether Save must ask first.
+ *
+ * Pure, so `tests/unit/skillAssignDraft.spec.js` runs every row type and every
+ * reason; `components/skills/SkillAssignModal.vue` wires them.
+ *
+ * The draft is the agent's INDIVIDUAL assignments (ent#530), and both buttons
+ * act on the rows the dialog lists, nothing else:
+ *   - a set's row is ticked and locked: never added to the draft, never removed;
+ *   - an assignment whose skill left the library has no row, so it stays
+ *     (operator ruling, 2026-10-22: its card's Unassign removes it);
+ *   - Select all never ticks a deprecated skill, and leaves the tick a
+ *     deprecated skill already has (operator ruling, 2026-10-22);
+ *   - Select all never ticks a name the PUT would refuse: the library lists
+ *     every directory with a SKILL.md, unvalidated, and one refused name fails
+ *     the whole save.
+ * A filter, if one arrives, narrows `rows`, and both buttons follow it.
+ */
+import { SKILL_NAME_RE } from './skillCards'
+
+/** Why a button is disabled: each says what is missing, never just "unavailable". */
+export const BULK_TITLES = Object.freeze({
+  loading: 'The library is still loading',
+  unconfigured: 'No skills library is configured',
+  empty: 'The library has no skills yet',
+  saving: 'Wait for the save to finish',
+  allTicked: 'Every skill listed here is already ticked',
+  deprecatedLeft: 'The skills left unticked are deprecated: tick one by hand to assign it',
+  unassignableLeft: "The skills left unticked have names the platform can't assign",
+  noneTicked: 'No skill listed here is ticked',
+  onlySetTicked: 'Only skills a set assigns are ticked here: unassign the set to remove them',
+})
+
+export function selectAllTitle(count, { skipsUnassignable = false } = {}) {
+  const except = skipsUnassignable ? "deprecated ones and names the platform can't assign" : 'deprecated ones'
+  return `Tick every skill listed here except ${except} (${count} to tick)`
+}
+
+export function deselectAllTitle(count) {
+  return `Untick every skill listed here except ones only a set assigns (${count} to untick). `
+    + 'Saving then unassigns them, and asks first if that would drop an approval requirement.'
+}
+
+function selectReason({ loaded, emptyReason, saving, toAdd, unticked }) {
+  if (!loaded) return BULK_TITLES.loading
+  if (emptyReason === 'library_unconfigured') return BULK_TITLES.unconfigured
+  if (emptyReason === 'library_empty') return BULK_TITLES.empty
+  if (saving) return BULK_TITLES.saving
+  if (toAdd.length) return null
+  if (!unticked.length) return BULK_TITLES.allTicked
+  return unticked.some((s) => s.deprecated) ? BULK_TITLES.deprecatedLeft : BULK_TITLES.unassignableLeft
+}
+
+function deselectReason({ loaded, saving, toRemove, hasLocked }) {
+  if (!loaded) return BULK_TITLES.loading
+  if (saving) return BULK_TITLES.saving
+  if (toRemove.length) return null
+  return hasLocked ? BULK_TITLES.onlySetTicked : BULK_TITLES.noneTicked
+}
+
+/**
+ * What each button would change, and whether it may run now.
+ * @param {object} p
+ * @param {Array<{name: string, deprecated?: boolean}>} p.rows  the rows the dialog lists
+ * @param {string[]} p.draft  the individual names ticked now
+ * @param {Set<string>} p.locked  names present only through a set
+ * @param {boolean} p.loaded  the agent's assignments and the library have answered
+ * @param {string|null} p.emptyReason  `store.emptyReason`
+ * @param {boolean} p.saving  a save is in flight
+ */
+export function bulkSelection({
+  rows = [], draft = [], locked = new Set(), loaded = false, emptyReason = null, saving = false,
+} = {}) {
+  const ticked = new Set(draft)
+  const editable = rows.filter((s) => !locked.has(s.name))
+  const unticked = editable.filter((s) => !ticked.has(s.name))
+  const toAdd = unticked.filter((s) => !s.deprecated && SKILL_NAME_RE.test(s.name)).map((s) => s.name)
+  const toRemove = editable.filter((s) => ticked.has(s.name)).map((s) => s.name)
+  const hasLocked = rows.some((s) => locked.has(s.name))
+  const skipsUnassignable = unticked.some((s) => !s.deprecated && !SKILL_NAME_RE.test(s.name))
+
+  const selectWhy = selectReason({ loaded, emptyReason, saving, toAdd, unticked })
+  const deselectWhy = deselectReason({ loaded, saving, toRemove, hasLocked })
+  return {
+    toAdd,
+    toRemove,
+    select: { disabled: Boolean(selectWhy), title: selectWhy || selectAllTitle(toAdd.length, { skipsUnassignable }) },
+    deselect: { disabled: Boolean(deselectWhy), title: deselectWhy || deselectAllTitle(toRemove.length) },
+  }
+}
+
+export function withSelectAll(draft, toAdd) {
+  return [...draft, ...toAdd]
+}
+
+export function withDeselectAll(draft, toRemove) {
+  const drop = new Set(toRemove)
+  return draft.filter((name) => !drop.has(name))
+}
+
+/**
+ * The saved names this save would take off the agent. A name an assigned set
+ * still holds is only demoted, so it stays; pass an empty `setHeld` when the
+ * sets cannot be trusted (one is unresolved), because the server then keeps
+ * only rows that are set-only.
+ */
+export function removedNames({ saved = [], draft = [], setHeld = new Set() } = {}) {
+  const keep = new Set(draft)
+  return [...saved].filter((name) => !keep.has(name) && !setHeld.has(name))
+}
+
+/**
+ * The removed names whose approval requirement a save would put at risk: a
+ * person's unassign drops it once the package has left (trinity-enterprise#753).
+ * A name in `conflict` is the agent's own skill, which stays, and the server
+ * keeps its requirement, so it never asks. With the approval map unread, every
+ * other removal asks.
+ * @returns {{names: string[], gatesKnown: boolean} | null}
+ */
+export function gatedRemovals({
+  saved = [], draft = [], setHeld = new Set(), gated = new Set(), conflict = new Set(), gatesKnown = false,
+} = {}) {
+  const removed = removedNames({ saved, draft, setHeld }).filter((name) => !conflict.has(name))
+  if (!removed.length) return null
+  if (!gatesKnown) return { names: removed, gatesKnown: false }
+  const names = removed.filter((name) => gated.has(name.toLowerCase()))
+  return names.length ? { names, gatesKnown: true } : null
+}
+
+const LISTED_NAMES = 5
+
+function nameList(names) {
+  const shown = names.slice(0, LISTED_NAMES).map((name) => `/${name}`)
+  const more = names.length - shown.length
+  if (more > 0) return `${shown.join(', ')} and ${more} more`
+  if (shown.length === 1) return shown[0]
+  return `${shown.slice(0, -1).join(', ')} and ${shown[shown.length - 1]}`
+}
+
+export const REMOVAL_CONFIRM_TEXT = 'Save and unassign'
+
+/** The confirm's title and message, in the future tense (the Unassign confirm's rule). */
+export function removalConfirm({ names, gatesKnown }) {
+  const one = names.length === 1
+  const list = nameList(names)
+  if (gatesKnown) {
+    return {
+      title: one ? `Unassign /${names[0]}?` : `Unassign ${names.length} skills that need approval?`,
+      message: one
+        ? `${list} needs approval. Saving will remove it from this agent, along with its approval requirement.`
+        : `${list} need approval. Saving will remove them from this agent, along with their approval requirements.`,
+    }
+  }
+  return {
+    title: one ? `Unassign /${names[0]}?` : `Unassign ${names.length} skills?`,
+    message: `Saving will remove ${list} from this agent. Which skills need approval couldn't be read, `
+      + `so any approval requirement on ${one ? 'it' : 'them'} will go too.`,
+  }
+}
diff --git a/src/frontend/tests/unit/skillAssignDraft.spec.js b/src/frontend/tests/unit/skillAssignDraft.spec.js
new file mode 100644
index 0000000000..9098806a73
--- /dev/null
+++ b/src/frontend/tests/unit/skillAssignDraft.spec.js
@@ -0,0 +1,218 @@
+/**
+ * trinity-enterprise#539 — the rules behind Select all / Deselect all in
+ * "Assign shared skills" and the confirm before a save drops approval
+ * requirements (`utils/skillAssignDraft.js`), executed over every row type and
+ * every reason. The wiring — what the buttons do to the real draft, what Save
+ * then sends, focus, the confirm — is `skillAssignSelectAll.spec.js` (mounted).
+ */
+import { describe, it, expect } from 'vitest'
+import {
+  bulkSelection, withSelectAll, withDeselectAll, removedNames, gatedRemovals, removalConfirm,
+  BULK_TITLES, selectAllTitle, deselectAllTitle,
+} from '../../src/utils/skillAssignDraft'
+
+// One row of every kind the dialog lists. `locked` is held only through a set;
+// `_scratch` is a library directory whose name the PUT refuses (the listing
+// does not validate names). The draft also holds `gone`, an assignment whose
+// skill left the library, so the dialog has no row for it.
+const ROWS = [
+  { name: 'plain-on' }, { name: 'plain-off' },
+  { name: 'dep-on', deprecated: true }, { name: 'dep-off', deprecated: true },
+  { name: 'locked' }, { name: 'both' },
+  { name: '_scratch' },
+]
+const LOCKED = new Set(['locked'])
+const SAVED = ['plain-on', 'dep-on', 'both', 'gone']
+const sel = (over = {}) => bulkSelection({
+  rows: ROWS, draft: SAVED, locked: LOCKED, loaded: true, emptyReason: null, saving: false, ...over,
+})
+
+describe('what each button would change', () => {
+  it('Select all adds only rows that are unticked, not deprecated, assignable and not a set\'s', () => {
+    expect(sel().toAdd).toEqual(['plain-off'])
+  })
+
+  it('Deselect all removes every ticked listed row except a set\'s, deprecated ones included', () => {
+    expect(sel().toRemove).toEqual(['plain-on', 'dep-on', 'both'])
+  })
+
+  it('both keep every name the dialog does not list', () => {
+    const s = sel()
+    expect(withSelectAll(SAVED, s.toAdd)).toEqual(['plain-on', 'dep-on', 'both', 'gone', 'plain-off'])
+    expect(withDeselectAll(SAVED, s.toRemove)).toEqual(['gone'])
+  })
+
+  it('Deselect all then Select all does not re-tick a deprecated skill the agent held', () => {
+    const cleared = withDeselectAll(SAVED, sel().toRemove)
+    const again = withSelectAll(cleared, sel({ draft: cleared }).toAdd)
+    expect(again).not.toContain('dep-on')
+    expect([...again].sort()).toEqual(['both', 'gone', 'plain-off', 'plain-on'])
+  })
+
+  it('a set\'s row never enters the draft, so a library of them has nothing to add', () => {
+    const s = bulkSelection({ rows: [{ name: 'locked' }], draft: [], locked: LOCKED, loaded: true })
+    expect(s.toAdd).toEqual([])
+    expect(s.toRemove).toEqual([])
+  })
+})
+
+describe('Select all — enabled, or the reason it is not', () => {
+  it.each([
+    ['the library has not loaded', { loaded: false }, BULK_TITLES.loading],
+    ['no library is configured', { rows: [], emptyReason: 'library_unconfigured' }, BULK_TITLES.unconfigured],
+    ['the library is empty', { rows: [], emptyReason: 'library_empty' }, BULK_TITLES.empty],
+    ['a save is in flight', { saving: true }, BULK_TITLES.saving],
+    ['a save is in flight and nothing is left', { saving: true, draft: ROWS.map((r) => r.name) }, BULK_TITLES.saving],
+    ['every row is ticked', { draft: ROWS.map((r) => r.name) }, BULK_TITLES.allTicked],
+    ['only a set\'s rows are listed', { rows: [{ name: 'locked' }], draft: [] }, BULK_TITLES.allTicked],
+    ['only a deprecated row is left', { draft: [...SAVED, 'plain-off', '_scratch'] }, BULK_TITLES.deprecatedLeft],
+    ['only an unassignable row is left', { draft: [...SAVED, 'plain-off', 'dep-off'] }, BULK_TITLES.unassignableLeft],
+    ['deprecated and unassignable rows are left', { draft: [...SAVED, 'plain-off'] }, BULK_TITLES.deprecatedLeft],
+  ])('disabled when %s', (_, over, title) => {
+    expect(sel(over).select).toEqual({ disabled: true, title })
+  })
+
+  it('an agent with nothing assigned yet can still Select all', () => {
+    expect(sel({ draft: [], emptyReason: 'none_assigned' }).select.disabled).toBe(false)
+  })
+
+  it('enabled, its title counts what it will tick, and names what it skips', () => {
+    // `_scratch` is unticked and refused by the PUT, so the title says it is skipped.
+    expect(sel().select).toEqual({ disabled: false, title: selectAllTitle(1, { skipsUnassignable: true }) })
+    expect(sel({ draft: [] }).select.title).toBe(selectAllTitle(3, { skipsUnassignable: true }))
+    // With no such row left unticked, only the deprecated exception is named.
+    expect(sel({ draft: [...SAVED, '_scratch'] }).select.title).toBe(selectAllTitle(1))
+    expect(selectAllTitle(1)).not.toMatch(/can't assign/)
+    expect(selectAllTitle(1, { skipsUnassignable: true })).toMatch(/can't assign/)
+  })
+})
+
+describe('Deselect all — enabled, or the reason it is not', () => {
+  it.each([
+    ['the library has not loaded', { loaded: false }, BULK_TITLES.loading],
+    ['a save is in flight', { saving: true }, BULK_TITLES.saving],
+    ['nothing is ticked', { rows: [{ name: 'plain-off' }], draft: [] }, BULK_TITLES.noneTicked],
+    ['the draft holds only a name the dialog does not list', { rows: [{ name: 'plain-off' }], draft: ['gone'] }, BULK_TITLES.noneTicked],
+    ['only a set\'s row is ticked', { rows: [{ name: 'locked' }, { name: 'plain-off' }], draft: [] }, BULK_TITLES.onlySetTicked],
+    ['the library is empty', { rows: [], draft: [], emptyReason: 'library_empty' }, BULK_TITLES.noneTicked],
+  ])('disabled when %s', (_, over, title) => {
+    expect(sel(over).deselect).toEqual({ disabled: true, title })
+  })
+
+  it('enabled, its title counts what it will untick', () => {
+    expect(sel().deselect).toEqual({ disabled: false, title: deselectAllTitle(3) })
+  })
+})
+
+describe('the titles say what the rule is', () => {
+  it('Select all names the listed rows and the deprecated exception', () => {
+    // "listed here": a filter, if one arrives, narrows what both buttons act on.
+    expect(selectAllTitle(2)).toMatch(/listed here/)
+    expect(selectAllTitle(2)).toMatch(/deprecated/)
+  })
+
+  it('Deselect all names the rows it leaves and what a save then does', () => {
+    expect(deselectAllTitle(2)).toMatch(/listed here/)
+    // A row both assigned on its own and through a set IS unticked; only a
+    // set-only row is left, so the title must not promise more.
+    expect(deselectAllTitle(2)).toMatch(/only a set assigns/)
+    expect(deselectAllTitle(2)).toMatch(/asks first if that would drop an approval requirement/)
+  })
+
+  it('every reason reads differently', () => {
+    const all = Object.values(BULK_TITLES)
+    expect(new Set(all).size).toBe(all.length)
+  })
+})
+
+describe('what a save takes off the agent', () => {
+  it('saved names left out of the draft, except ones a set still holds', () => {
+    expect(removedNames({ saved: ['a', 'b', 'c'], draft: ['a'], setHeld: new Set(['c']) })).toEqual(['b'])
+  })
+
+  it('with the sets untrusted (an empty setHeld), a set-held name is removed too', () => {
+    expect(removedNames({ saved: ['a', 'c'], draft: ['a'], setHeld: new Set() })).toEqual(['c'])
+  })
+
+  it('nothing when the draft keeps every saved name', () => {
+    expect(removedNames({ saved: new Set(['a', 'b']), draft: ['b', 'a', 'x'] })).toEqual([])
+  })
+})
+
+describe('which saves ask first', () => {
+  const base = {
+    saved: ['plain-on', 'dep-on', 'both', 'gone'],
+    draft: ['gone'],
+    setHeld: new Set(['both']),
+    gated: new Set(),
+    conflict: new Set(),
+    gatesKnown: true,
+  }
+
+  it('none when nothing is removed', () => {
+    expect(gatedRemovals({ ...base, draft: base.saved })).toBe(null)
+  })
+
+  it('none when nothing removed has an approval requirement', () => {
+    expect(gatedRemovals(base)).toBe(null)
+  })
+
+  it('a removed skill with one, matched on the lower-cased gate key', () => {
+    expect(gatedRemovals({ ...base, saved: ['Plain-On', 'gone'], gated: new Set(['plain-on']) }))
+      .toEqual({ names: ['Plain-On'], gatesKnown: true })
+  })
+
+  it('not a skill a set still holds: the save only demotes it, and its requirement stays', () => {
+    expect(gatedRemovals({ ...base, gated: new Set(['both']) })).toBe(null)
+  })
+
+  it('not a skill in conflict: the server keeps that requirement', () => {
+    expect(gatedRemovals({ ...base, gated: new Set(['dep-on']), conflict: new Set(['dep-on']) })).toBe(null)
+  })
+
+  it('every removal when the approval map could not be read', () => {
+    expect(gatedRemovals({ ...base, gatesKnown: false }))
+      .toEqual({ names: ['plain-on', 'dep-on'], gatesKnown: false })
+  })
+
+  it('none when the map could not be read but nothing is removed', () => {
+    expect(gatedRemovals({ ...base, draft: base.saved, gatesKnown: false })).toBe(null)
+  })
+
+  it('never a conflict name, even with the map unread: the agent\'s own skill and its requirement stay', () => {
+    expect(gatedRemovals({ ...base, conflict: new Set(['dep-on']), gatesKnown: false }))
+      .toEqual({ names: ['plain-on'], gatesKnown: false })
+    expect(gatedRemovals({ ...base, saved: ['dep-on', 'gone'], conflict: new Set(['dep-on']), gatesKnown: false }))
+      .toBe(null)
+  })
+})
+
+describe('what the confirm says', () => {
+  it('one skill', () => {
+    expect(removalConfirm({ names: ['a'], gatesKnown: true })).toEqual({
+      title: 'Unassign /a?',
+      message: '/a needs approval. Saving will remove it from this agent, along with its approval requirement.',
+    })
+  })
+
+  it('several, in a bounded list', () => {
+    expect(removalConfirm({ names: ['a', 'b'], gatesKnown: true })).toEqual({
+      title: 'Unassign 2 skills that need approval?',
+      message: '/a and /b need approval. Saving will remove them from this agent, along with their approval requirements.',
+    })
+    const seven = removalConfirm({ names: ['a', 'b', 'c', 'd', 'e', 'f', 'g'], gatesKnown: true })
+    expect(seven.title).toBe('Unassign 7 skills that need approval?')
+    expect(seven.message.startsWith('/a, /b, /c, /d, /e and 2 more need approval.')).toBe(true)
+  })
+
+  it('an unread map says why it asks', () => {
+    expect(removalConfirm({ names: ['a'], gatesKnown: false })).toEqual({
+      title: 'Unassign /a?',
+      message: "Saving will remove /a from this agent. Which skills need approval couldn't be read, so any approval requirement on it will go too.",
+    })
+    expect(removalConfirm({ names: ['a', 'b', 'c'], gatesKnown: false })).toEqual({
+      title: 'Unassign 3 skills?',
+      message: "Saving will remove /a, /b and /c from this agent. Which skills need approval couldn't be read, so any approval requirement on them will go too.",
+    })
+  })
+})
diff --git a/src/frontend/tests/unit/skillAssignSelectAll.spec.js b/src/frontend/tests/unit/skillAssignSelectAll.spec.js
new file mode 100644
index 0000000000..10243518e0
--- /dev/null
+++ b/src/frontend/tests/unit/skillAssignSelectAll.spec.js
@@ -0,0 +1,442 @@
+// @vitest-environment jsdom
+/**
+ * trinity-enterprise#539 — Select all / Deselect all in "Assign shared skills",
+ * and the confirm before a save drops approval requirements, MOUNTED.
+ *
+ * The rules are executed in skillAssignDraft.spec.js. This proves the wiring a
+ * rule test cannot see: what each button does to the real draft, what Save then
+ * SENDS (a checkbox assertion cannot see the assignment the dialog does not
+ * list, or the set's locked row), when Save asks first, and where keyboard
+ * focus goes. Only `@/api` is mocked; the stores and the component are real.
+ *
+ * The dialog renders through the REAL Teleport, into `document.body`, and is
+ * queried there. Stubbed inline, every patch re-created the dialog's nodes, so
+ * a button that a click disables was REPLACED rather than disabled in place,
+ * and the focus tests passed for the wrong reason (review of this PR).
+ *
+ * Two harness facts shape the assertions: VTU's `trigger` does nothing on a
+ * disabled element, so a disabled state is read off `element.disabled`, never
+ * "proven" by a click that goes nowhere; and jsdom leaves focus on a control
+ * that becomes disabled, so focus is asserted on its exact target.
+ *
+ * The controls render only inside this dialog, which only the owner or an
+ * admin can open: `skillsTab.mount.spec.js` ("a viewer is offered no
+ * assignment controls") holds that.
+ */
+import { describe, it, expect, beforeEach, afterEach, vi } from 'vitest'
+import { mount, flushPromises, enableAutoUnmount, DOMWrapper } from '@vue/test-utils'
+import { setActivePinia, createPinia } from 'pinia'
+import { nextTick } from 'vue'
+
+const { api } = vi.hoisted(() => ({
+  api: { get: vi.fn(), post: vi.fn(), put: vi.fn(), delete: vi.fn() },
+}))
+vi.mock('@/api', () => ({ default: api }))
+vi.mock('@/stores/auth', () => ({
+  useAuthStore: () => ({ role: 'admin', isAuthenticated: true }),
+}))
+
+import SkillAssignModal from '../../src/components/skills/SkillAssignModal.vue'
+import { useSkillsStore } from '../../src/stores/skills'
+import { useSkillGatesStore } from '../../src/stores/skillGates'
+import { BULK_TITLES, selectAllTitle, deselectAllTitle } from '../../src/utils/skillAssignDraft'
+
+enableAutoUnmount(afterEach)
+
+const AGENT = 'agent-x'
+const LIBRARY = [
+  { name: 'plain-on', description: 'held' },
+  { name: 'plain-off', description: 'not held' },
+  { name: 'dep-on', description: 'retired, held', deprecated: true, superseded_by: 'plain-on' },
+  { name: 'dep-off', description: 'retired, not held', deprecated: true, superseded_by: null },
+  { name: 'locked', description: 'held only through the set kit' },
+  { name: 'both', description: 'held on its own and through kit' },
+]
+const row = (skill_name, over = {}) => ({ skill_name, individual: true, via_sets: [], delivery_status: null, ...over })
+// `gone` is assigned on its own, but its skill has left the library: the
+// dialog has no row for it.
+const HELD = [
+  row('plain-on'), row('dep-on'), row('both', { via_sets: ['kit'] }),
+  row('locked', { individual: false, via_sets: ['kit'] }), row('gone'),
+]
+const SAVED_TICKS = ['both', 'dep-on', 'plain-on']
+const gate = (skill_name) => ({ skill_name, approver: 'primary', approver_reachable: true, origin: 'set' })
+
+let state
+function respond(over = {}) {
+  state = {
+    library: LIBRARY, assigned: HELD, status: { configured: true }, sets: [],
+    gates: [], gateMapFails: false, ...over,
+  }
+  api.get.mockImplementation((url) => {
+    if (url === '/api/skills/library/status') return Promise.resolve({ data: state.status })
+    if (url === '/api/skills/library') return Promise.resolve({ data: state.library })
+    if (url === `/api/agents/${AGENT}/skills`) return Promise.resolve({ data: state.assigned })
+    if (url === `/api/agents/${AGENT}/skill-sets`) return Promise.resolve({ data: state.sets })
+    if (url === `/api/agents/${AGENT}/skill-gates`) {
+      if (state.gateMapFails) return Promise.reject({ response: { status: 500, data: { detail: 'down' } } })
+      const answer = () => ({ data: { agent_name: AGENT, gates: state.gates, approvers: [] } })
+      return state.gateMapHold ? state.gateMapHold.then(answer) : Promise.resolve(answer())
+    }
+    return Promise.resolve({ data: [] })
+  })
+}
+
+// The tab loads both stores; the dialog reads them.
+async function mountModal({ load = true } = {}) {
+  if (load) {
+    await useSkillsStore().load(AGENT)
+    await useSkillGatesStore().load(AGENT)
+  }
+  const w = mount(SkillAssignModal, {
+    props: { modelValue: true },
+    attachTo: document.body,
+    global: { stubs: { SkillContractChips: true } },
+  })
+  await flushPromises(); await nextTick()
+  return w
+}
+
+// Teleported: the dialog lives in <body>, not under the wrapper.
+const page = () => new DOMWrapper(document.body)
+const tid = (id) => page().find(`[data-testid="${id}"]`)
+const selectAll = () => tid('skill-assign-select-all')
+const deselectAll = () => tid('skill-assign-deselect-all')
+const save = () => tid('skill-assign-save')
+const reset = () => tid('skill-assign-reset')
+const close = () => tid('skill-assign-close')
+const box = (name) => page().find(`[data-testid="skill-assign-list"] input[type="checkbox"][value="${name}"]`)
+const ticked = () => page().findAll('[data-testid="skill-assign-list"] input[type="checkbox"][value]')
+  .filter((b) => b.element.checked).map((b) => b.element.value).sort()
+const sent = () => [...api.put.mock.calls.at(-1)[1].skills].sort()
+const writes = () => api.put.mock.calls.length + api.post.mock.calls.length + api.delete.mock.calls.length
+const gateReads = () => api.get.mock.calls.filter(([url]) => url === `/api/agents/${AGENT}/skill-gates`).length
+
+async function settle() {
+  await flushPromises(); await nextTick(); await flushPromises()
+}
+async function click(wrapper) {
+  await wrapper.trigger('click')
+  await settle()
+}
+
+beforeEach(() => {
+  setActivePinia(createPinia())
+  for (const f of Object.values(api)) f.mockReset()
+  respond()
+  api.put.mockResolvedValue({ data: { delivery: null } })
+})
+
+describe('Select all', () => {
+  it('ticks what it may, keeps every tick it found, and writes nothing', async () => {
+    await mountModal()
+    expect(ticked()).toEqual(SAVED_TICKS)
+    expect(save().element.disabled).toBe(true)
+
+    await click(selectAll())
+
+    expect(ticked()).toEqual(['both', 'dep-on', 'plain-off', 'plain-on'])
+    expect(box('dep-off').element.checked).toBe(false)              // deprecated and not held: left to the operator
+    expect(tid('skill-locked-locked').element.checked).toBe(true)   // the set's row is still ticked…
+    expect(tid('skill-locked-locked').element.disabled).toBe(true)  // …and still locked
+    expect(save().element.disabled).toBe(false)                     // dirty, as after a manual tick
+    expect(reset().exists()).toBe(true)
+    expect(writes()).toBe(0)
+  })
+
+  it('then Save sends the additions, the held deprecated skill and the unlisted assignment, never the set\'s row', async () => {
+    const w = await mountModal()
+    await click(selectAll())
+    await click(save())
+
+    expect(api.put).toHaveBeenCalledTimes(1)
+    expect(api.put.mock.calls[0][0]).toBe(`/api/agents/${AGENT}/skills`)
+    expect(sent()).toEqual(['both', 'dep-on', 'gone', 'plain-off', 'plain-on'])
+    expect(w.emitted('saved')).toHaveLength(1)                     // the same notice path as any save
+  })
+})
+
+describe('Deselect all', () => {
+  it('unticks every listed skill it may, deprecated included, and writes nothing', async () => {
+    await mountModal()
+    await click(deselectAll())
+
+    expect(ticked()).toEqual([])
+    expect(tid('skill-locked-locked').element.checked).toBe(true)   // the set's row stays, locked
+    expect(save().element.disabled).toBe(false)
+    expect(writes()).toBe(0)
+  })
+
+  it('then Save keeps the assignment the dialog does not list', async () => {
+    await mountModal()
+    await click(deselectAll())
+    await click(save())
+
+    expect(tid('confirm-dialog').exists()).toBe(false)              // nothing removed carries a requirement
+    expect(sent()).toEqual(['gone'])                               // `both` stays through its set, server-side
+  })
+
+  it('with nothing unlisted, Save persists an empty list', async () => {
+    respond({ assigned: [row('plain-on'), row('plain-off')] })
+    await mountModal()
+    await click(deselectAll())
+    await click(save())
+    expect(sent()).toEqual([])
+  })
+})
+
+describe('Reset still restores the saved set', () => {
+  it.each(['select', 'deselect'])('after %s all', async (verb) => {
+    await mountModal()
+    await click(verb === 'select' ? selectAll() : deselectAll())
+    expect(ticked()).not.toEqual(SAVED_TICKS)
+
+    await click(reset())
+    expect(ticked()).toEqual(SAVED_TICKS)
+    expect(save().element.disabled).toBe(true)
+  })
+})
+
+describe('disabled, with the reason on the button', () => {
+  it('both wait for a library that has not loaded', async () => {
+    await mountModal({ load: false })
+    for (const b of [selectAll(), deselectAll()]) {
+      expect(b.element.disabled).toBe(true)
+      expect(b.attributes('title')).toBe(BULK_TITLES.loading)
+    }
+  })
+
+  it.each([
+    ['unconfigured', { status: { configured: false } }, BULK_TITLES.unconfigured],
+    ['empty', { library: [] }, BULK_TITLES.empty],
+  ])('Select all names a library that is %s', async (_, over, title) => {
+    respond(over)
+    await mountModal()
+    expect(selectAll().element.disabled).toBe(true)
+    expect(selectAll().attributes('title')).toBe(title)
+  })
+
+  it('enabled, each title counts what the click will change', async () => {
+    await mountModal()
+    expect(selectAll().attributes('title')).toBe(selectAllTitle(1))    // plain-off
+    expect(deselectAll().attributes('title')).toBe(deselectAllTitle(3)) // plain-on, dep-on, both
+  })
+
+  it('both wait while a save is in flight, then come back', async () => {
+    let finish
+    api.put.mockImplementation(() => new Promise((resolve) => { finish = resolve }))
+    await mountModal()
+    await click(selectAll())
+    await click(save())
+
+    for (const b of [selectAll(), deselectAll()]) {
+      expect(b.element.disabled).toBe(true)
+      expect(b.attributes('title')).toBe(BULK_TITLES.saving)
+    }
+    finish({ data: { delivery: null } })
+    await settle()
+    expect(selectAll().element.disabled).toBe(false)
+    expect(deselectAll().element.disabled).toBe(false)
+  })
+})
+
+describe('keyboard: focus never falls out of the dialog', () => {
+  it('the dialog opens on Select all; Deselect all is marked so it never does', async () => {
+    await mountModal()
+    expect(document.activeElement).toBe(selectAll().element)
+    expect(deselectAll().attributes()).toHaveProperty('data-destructive')
+  })
+
+  it('with nothing to tick, it opens on the list, not on Deselect all', async () => {
+    respond({ library: [LIBRARY[0]], assigned: [row('plain-on')] })
+    await mountModal()
+    expect(selectAll().element.disabled).toBe(true)
+    expect(document.activeElement).toBe(box('plain-on').element)
+  })
+
+  it('a button that disables itself hands focus to the other one', async () => {
+    await mountModal()
+    const selectEl = selectAll().element
+    selectEl.focus()
+    await click(selectAll())
+    expect(selectAll().element).toBe(selectEl)                     // disabled in place, not re-created
+    expect(selectEl.disabled).toBe(true)
+    expect(document.activeElement).toBe(deselectAll().element)
+
+    await click(deselectAll())
+    expect(deselectAll().element.disabled).toBe(true)
+    expect(document.activeElement).toBe(selectAll().element)
+  })
+
+  it('…and to Close when the other one has nothing to do', async () => {
+    respond({ library: [LIBRARY[2]], assigned: [row('dep-on')] })    // a held deprecated skill, nothing else
+    await mountModal()
+    deselectAll().element.focus()
+    await click(deselectAll())
+    expect(selectAll().element.disabled).toBe(true)
+    expect(document.activeElement).toBe(close().element)
+  })
+
+  it('a manual tick that leaves Select all nothing to do moves focus off it', async () => {
+    await mountModal()
+    selectAll().element.focus()                                    // a mouse click on a box leaves it there (Safari)
+    await box('plain-off').setValue(true)
+    await settle()
+    expect(selectAll().element.disabled).toBe(true)
+    expect(document.activeElement).toBe(deselectAll().element)
+  })
+
+  it('Reset hands focus to Close rather than dropping it with the button', async () => {
+    await mountModal()
+    await click(selectAll())
+    reset().element.focus()
+    await click(reset())
+    expect(reset().exists()).toBe(false)
+    expect(document.activeElement).toBe(close().element)
+  })
+
+  it('Save hands focus to Close while it saves', async () => {
+    api.put.mockImplementation(() => new Promise(() => {}))         // never answers: the save stays in flight
+    await mountModal()
+    await click(selectAll())
+    save().element.focus()
+    await click(save())
+    expect(save().element.disabled).toBe(true)
+    expect(document.activeElement).toBe(close().element)
+  })
+})
+
+describe('Save asks first when it would drop an approval requirement', () => {
+  it('names the skill, opens on Cancel, and sends nothing until confirmed', async () => {
+    respond({ gates: [gate('plain-on')] })
+    await mountModal()
+    await click(deselectAll())
+    await click(save())
+
+    expect(tid('confirm-dialog').exists()).toBe(true)
+    expect(tid('confirm-dialog-title').text()).toBe('Unassign /plain-on?')
+    expect(tid('confirm-dialog-message').text())
+      .toBe('/plain-on needs approval. Saving will remove it from this agent, along with its approval requirement.')
+    expect(document.activeElement).toBe(tid('confirm-dialog-cancel').element)
+    expect(api.put).not.toHaveBeenCalled()
+
+    await click(tid('confirm-dialog-cancel'))
+    expect(tid('confirm-dialog').exists()).toBe(false)
+    expect(ticked()).toEqual([])                                    // the draft is kept
+    expect(api.put).not.toHaveBeenCalled()
+
+    await click(save())
+    await click(tid('confirm-dialog-confirm'))
+    expect(api.put).toHaveBeenCalledTimes(1)
+    expect(sent()).toEqual(['gone'])
+  })
+
+  it('decides on a fresh gate map, not the one read when the tab opened', async () => {
+    await mountModal()                                              // read: nothing gated
+    state.gates = [gate('plain-on')]                                // gated since, elsewhere
+    const before = gateReads()
+    await box('plain-on').setValue(false)
+    await click(save())
+    expect(gateReads()).toBe(before + 1)
+    expect(tid('confirm-dialog').exists()).toBe(true)
+    expect(api.put).not.toHaveBeenCalled()
+  })
+
+  it('a second click while the map is re-read is the same click', async () => {
+    await mountModal()
+    let release
+    state.gateMapHold = new Promise((resolve) => { release = resolve })
+    await box('plain-on').setValue(false)
+    await save().trigger('click')
+    await save().trigger('click')
+    release()
+    await settle()
+    expect(api.put).toHaveBeenCalledTimes(1)
+  })
+
+  it('a dialog closed while the map is re-read neither saves nor asks', async () => {
+    const w = await mountModal()
+    let release
+    state.gateMapHold = new Promise((resolve) => { release = resolve })
+    await box('plain-on').setValue(false)
+    await save().trigger('click')
+    await w.setProps({ modelValue: false })
+    release()
+    await settle()
+    expect(api.put).not.toHaveBeenCalled()
+  })
+
+  it('reads nothing extra for a save that only adds', async () => {
+    await mountModal()
+    const before = gateReads()
+    await click(selectAll())
+    await click(save())
+    expect(gateReads()).toBe(before)
+    expect(api.put).toHaveBeenCalledTimes(1)
+  })
+
+  it('asks for a manual untick too: it is the same consequence', async () => {
+    respond({ gates: [gate('plain-on')] })
+    await mountModal()
+    await box('plain-on').setValue(false)
+    await click(save())
+    expect(tid('confirm-dialog').exists()).toBe(true)
+    expect(api.put).not.toHaveBeenCalled()
+  })
+
+  it('does not ask when the set still holds the gated skill', async () => {
+    respond({ gates: [gate('both')] })
+    await mountModal()
+    await box('both').setValue(false)
+    await click(save())
+    expect(tid('confirm-dialog').exists()).toBe(false)
+    expect(api.put).toHaveBeenCalledTimes(1)
+  })
+
+  it('does ask for it when another held set is unresolved: the server then keeps only set-only rows', async () => {
+    respond({
+      gates: [gate('both')],
+      sets: [{ name: 'kit', status: 'ok' }, { name: 'gone-set', status: 'unresolved', reason: 'missing_upstream' }],
+    })
+    await mountModal()
+    await box('both').setValue(false)
+    await click(save())
+    expect(tid('confirm-dialog-title').text()).toBe('Unassign /both?')
+    expect(api.put).not.toHaveBeenCalled()
+  })
+
+  it('asks for any removal when the approval map cannot be read', async () => {
+    await mountModal()                                              // read fine when the tab opened…
+    state.gateMapFails = true                                       // …and fails now
+    await box('dep-on').setValue(false)
+    await click(save())
+    expect(tid('confirm-dialog-message').text()).toContain("couldn't be read")
+    expect(api.put).not.toHaveBeenCalled()
+  })
+
+  it('confirming keeps focus in the dialog while the save runs', async () => {
+    respond({ gates: [gate('plain-on')] })
+    api.put.mockImplementation(() => new Promise(() => {}))
+    await mountModal()
+    await box('plain-on').setValue(false)
+    save().element.focus()
+    await click(save())
+    await click(tid('confirm-dialog-confirm'))
+    expect(tid('confirm-dialog').exists()).toBe(false)
+    expect(api.put).toHaveBeenCalledTimes(1)
+    expect(document.activeElement).toBe(close().element)
+  })
+})
+
+describe('a failed save', () => {
+  it('keeps the bulk draft and says why', async () => {
+    api.put.mockRejectedValue({ response: { data: { detail: 'Could not reach the agent' } } })
+    await mountModal()
+    await click(selectAll())
+    await click(save())
+
+    expect(ticked()).toContain('plain-off')
+    expect(document.body.textContent).toContain('Could not reach the agent')
+    expect(save().element.disabled).toBe(false)
+  })
+})
diff --git a/tests/registry.json b/tests/registry.json
index 2675f9547a..3ee8ca4160 100644
--- a/tests/registry.json
+++ b/tests/registry.json
@@ -6314,7 +6314,7 @@
         "security",
         "mcp"
       ],
-      "description": "Anti-poisoning gates on agent-written user memory (MEM-001): the write-boundary screen through the real router (each of memory_prompt_marker / memory_instruction / memory_tool_directive / memory_link_in_instruction refused with its code and the line the agent sent, nothing stored, the log carries rule and line never the text; folded spellings \u2014 zero-width, full-width, Hangul filler, combining marks, split across a line break, case; a benign table of facts, preferences, open loops, plain links and the false positives an adversarial review found; the cheap misses it found), the exemption keyed on the hit (an old hit kept is not counted again; old lines reordered, re-contextualised or respelled into a new hit are refused; a split across old and new is refused), copy-back normalisation (the rendered block stored clean; a whole-block copy-back keeps only the agent-notes section), the non-creating previous-notes read over real SQLite, AC 4 (a channel write resolves ent#549's one-to-one speaker stamp; a group turn writes no one's memory), the trust envelope on the real format_user_memory_block (nonce'd open/close lines, every splitlines() line margined incl. \\r / NEL / U+2028, forged ends and platform structure only ever margined, control characters removed, summary inside, decisions after the reminder, render failure drops memory, fresh nonce, a Hypothesis property), wiring through the seat-run and channel caller prompts, the summarizer (a tripping line is dropped and the rest kept, all lines dropped keeps the previous summary, each message clamped, the conversation margined inside its prompt), seat decisions (any line terminator refused, markers / instructions / an MCP tool id / a slotted URL refused with a sentence receipt, business-rule decisions accepted, legacy rows rendered on one line, the same one-to-one rule in _seat_for, the block's framing line pinned against the guard), the Telegram history defang, AST pins (the router stores only what prepare_write returned; the summarizer screens before it stores; one agent-notes writer across src/backend), the structural-heading list pinned against the composed prompt, totality (Hypothesis) and time budgets on adversarial and summarizer-sized input."
+      "description": "Anti-poisoning gates on agent-written user memory (MEM-001): the write-boundary screen through the real router (each of memory_prompt_marker / memory_instruction / memory_tool_directive / memory_link_in_instruction refused with its code and the line the agent sent, nothing stored, the log carries rule and line never the text; folded spellings — zero-width, full-width, Hangul filler, combining marks, split across a line break, case; a benign table of facts, preferences, open loops, plain links and the false positives an adversarial review found; the cheap misses it found), the exemption keyed on the hit (an old hit kept is not counted again; old lines reordered, re-contextualised or respelled into a new hit are refused; a split across old and new is refused), copy-back normalisation (the rendered block stored clean; a whole-block copy-back keeps only the agent-notes section), the non-creating previous-notes read over real SQLite, AC 4 (a channel write resolves ent#549's one-to-one speaker stamp; a group turn writes no one's memory), the trust envelope on the real format_user_memory_block (nonce'd open/close lines, every splitlines() line margined incl. \\r / NEL / U+2028, forged ends and platform structure only ever margined, control characters removed, summary inside, decisions after the reminder, render failure drops memory, fresh nonce, a Hypothesis property), wiring through the seat-run and channel caller prompts, the summarizer (a tripping line is dropped and the rest kept, all lines dropped keeps the previous summary, each message clamped, the conversation margined inside its prompt), seat decisions (any line terminator refused, markers / instructions / an MCP tool id / a slotted URL refused with a sentence receipt, business-rule decisions accepted, legacy rows rendered on one line, the same one-to-one rule in _seat_for, the block's framing line pinned against the guard), the Telegram history defang, AST pins (the router stores only what prepare_write returned; the summarizer screens before it stores; one agent-notes writer across src/backend), the structural-heading list pinned against the composed prompt, totality (Hypothesis) and time budgets on adversarial and summarizer-sized input."
     },
     {
       "file": "../src/mcp-server/src/tools/memory.test.ts",
@@ -6325,6 +6325,27 @@
         "security"
       ],
       "description": "write_user_memory over a fake TrinityClient: a backend 422 refusal reaches the agent as a success:false result carrying its code, message and line; a stored write passes through unchanged; the published description names all four refusal codes and the margin to drop."
+    },
+    {
+      "file": "frontend/tests/unit/skillAssignDraft.spec.js",
+      "feature": "abilityai/trinity-enterprise#539",
+      "added": "2026-10-22",
+      "categories": [
+        "frontend",
+        "unit"
+      ],
+      "description": "trinity-enterprise#539: the pure rules behind Select all / Deselect all in Assign shared skills (utils/skillAssignDraft.js) over every row type: Select all adds only unticked, non-deprecated, assignable rows that are not a set's; Deselect all removes every ticked listed row except a set's; both keep a name the dialog does not list. Every disabled reason and its precedence, the counted enabled titles, and what a save removes (set-held kept unless the sets are untrusted), which saves must confirm first (a removed skill with an approval requirement, case-insensitive; never a set-held or conflict name; any other removal with the approval map unread), plus the confirm's words and the refused-name title clause. Mutation-proven."
+    },
+    {
+      "file": "frontend/tests/unit/skillAssignSelectAll.spec.js",
+      "feature": "abilityai/trinity-enterprise#539",
+      "added": "2026-10-22",
+      "categories": [
+        "frontend",
+        "unit",
+        "mount"
+      ],
+      "description": "trinity-enterprise#539, MOUNTED through the real Teleport: Select all / Deselect all change only the draft (no write), Save then sends the expected body (the unlisted assignment kept, the set's row never sent), Reset restores the saved set, every disabled state carries its reason, the dialog never opens on Deselect all, focus is re-homed whenever a control it sat on is disabled or removed (bulk click, manual tick, Reset, Save, Confirm), and Save confirms (Cancel focused) before dropping an approval requirement, deciding on a gate map re-read at save time (failed read, unresolved set, double click and a dialog closed mid-read covered). Mutation-proven (27 mutants)."
     }
   ]
 }