diff --git a/devlog/_plan/260905_admin_token_local_ux/000_research.md b/devlog/_plan/260905_admin_token_local_ux/000_research.md
new file mode 100644
index 0000000000..ae938bf82e
--- /dev/null
+++ b/devlog/_plan/260905_admin_token_local_ux/000_research.md
@@ -0,0 +1,120 @@
+# 000 — Research: why a plain local user is shown an admin token box
+
+Two open issues describe the same wound from opposite ends.
+
+- **#3483** (juzijia, Windows 10, 2.42.0) — the admin token dialog paints an
+  empty red error notice the moment it opens, before anything is submitted.
+- **#3353** (Tao-Yida) — after upgrading to 2.40.0 a user was locked out of the
+  dashboard by a bare password box, assumed a config-loss bug, and had to have
+  an LLM read the source to learn the token was a new security feature.
+
+The user's framing is stronger than either issue, and it is the one this unit
+adopts: **a plain local user should never see that dialog at all.** When it does
+appear it is a symptom, and the box asks the user to solve a problem they did
+not cause and cannot diagnose.
+
+## How the dashboard is supposed to authenticate
+
+A loopback install never needs a typed credential. The server mints a session
+and injects it into the served document:
+
+```
+GET /opencodex-session
+  -> src/server/index.ts:2074  issueGuiSession(...)
+  -> src/server/gui-session.ts:166
+  -> meta opencodex-session-token / -csrf / -origin / -server-origin
+     (src/server/gui-static.ts:71-74)
+```
+
+`gui/src/api.ts:loadInjectedSession()` reads those tags on boot. Verified live
+against the running 2.43.0 proxy on port 10100:
+
+```text
+curl -i -H 'Host: 127.0.0.1:10100' http://127.0.0.1:10100/opencodex-session
+HTTP/1.1 200 OK
+<meta name="opencodex-session-token" content="ocx_session_CO-4g0m5B_...">
+```
+
+So on the happy path the prompt is unreachable. The interesting question is what
+happens when that mint fails.
+
+## The fallback that should not be a fallback
+
+`gui/src/api.ts:resolveTokenAfter401()` (around line 247) handles a 401 like this:
+
+```ts
+const renewed = await Promise.race([reBootstrapSessionToken(plane), watchdog]);
+if (renewed.kind === "minted") return renewed.token;
+if (renewed.kind === "failed") return null;
+const prompted = await requestAdminToken(token => verifyAdminToken(plane, token));
+```
+
+`reBootstrapSessionToken` maps **any 4xx** to `"unavailable"`:
+
+```ts
+if (!response.ok) return response.status >= 400 && response.status < 500
+  ? { kind: "unavailable" } : { kind: "failed" };
+```
+
+And `"unavailable"` is precisely the branch that raises the password box.
+
+Now read the mint conditions (`src/server/gui-session.ts:172-183`):
+
+```ts
+if (!isApiAuthRequired(config)) {
+  if (!isLoopbackHostname(host.hostname) || !isAllowedManagementOrigin(req, config)) return null;
+  ...
+}
+```
+
+with `isApiAuthRequired(config) === !isLoopbackHostname(config.hostname)`
+(`src/server/auth-cors.ts:285`).
+
+That yields the defect in one sentence: **on a loopback install the only ways to
+get a 401 from the bootstrap are a Host or Origin mismatch — a misconfiguration
+the admin token cannot fix.** Typing a token there is not a recovery path; it is
+a dead end wearing a login form.
+
+And when the bind genuinely is non-loopback, the token is real and required —
+but the dialog explains none of that, which is exactly #3353.
+
+## Why the notice is already red and empty (#3483)
+
+`gui/src/admin-token-dialog.ts:76-79` builds the error element up front:
+
+```ts
+validationError.className = "notice notice-err";
+validationError.hidden = true;
+```
+
+`hidden` works only because the UA stylesheet says `[hidden] { display: none }`,
+and that rule is the weakest one in the cascade. `gui/src/styles.css:1307` then
+says:
+
+```css
+.notice { ... display: flex; ... }
+```
+
+An author rule with an explicit `display` beats the UA `[hidden]` rule, so the
+element stays laid out. It has `notice-err` borders and padding
+(`styles.css:1355-1359`) and no text, which renders as the empty red box in the
+screenshot. The bug is a CSS cascade defect, not a logic error — which is why no
+existing test caught it: happy-dom asserts `hidden === true` happily while a real
+browser paints the box.
+
+## What this unit changes
+
+1. Never prompt a standalone/loopback dashboard. Tell the user what is actually
+   wrong instead. (`010`)
+2. When the prompt is legitimate, make it self-explanatory and link to a real
+   setup guide. Fix the empty notice while in there. (`020`)
+3. Triage the Windows baseline and #3320 and land what is provable. (`030`)
+4. Deliver as a stacked PR chain, admin-merged to `dev`. (`040`)
+
+## Constraints carried from the request
+
+- No repository-wide local suite on this workstation. Focused `bun test <file>`
+  plus `bun run typecheck`; heavy probes go to SSH hosts.
+- A Windows baseline is already in flight on `desktop-c795oh4` under
+  `/c/ocxwin` (lock `/c/ocxwin/.suite.lock`, shards `base-1..4`). It is read-only
+  evidence for this unit and must not be disturbed.
diff --git a/devlog/_plan/260905_admin_token_local_ux/010_suppress_local_prompt.md b/devlog/_plan/260905_admin_token_local_ux/010_suppress_local_prompt.md
new file mode 100644
index 0000000000..e887fa3219
--- /dev/null
+++ b/devlog/_plan/260905_admin_token_local_ux/010_suppress_local_prompt.md
@@ -0,0 +1,101 @@
+# 010 — Never prompt a standalone dashboard for an admin token
+
+Work-phase `wp1`. Depends on 000. Criterion `c-1`.
+
+## Problem
+
+`gui/src/api.ts:resolveTokenAfter401()` treats "the server would not mint me a
+session" as "ask the human for a token". On a loopback install those are not the
+same thing. `issueGuiSession` mints unconditionally for a loopback host with an
+allowed origin (`src/server/gui-session.ts:172-183`), so a 401 there means the
+request did not look loopback to the server — a Host/Origin/bind problem. No
+token the user can type changes that verdict, because the token is not what was
+refused.
+
+## The signal
+
+The server already states its topology on every served document:
+
+```ts
+// src/server/gui-static.ts:95
+function runtimeRoleMeta(runtimeRole: string): string {
+  return \`<meta name="opencodex-runtime-role" content="\${escapeHtmlAttribute(runtimeRole)}">\`;
+}
+```
+
+and the GUI already reads it (`gui/src/api-targets.ts:10-14`). That comment block
+is explicit that a missing tag means "standalone / older server / Vite dev", i.e.
+the safe default. This unit reuses that exact reader rather than inventing a
+second topology signal.
+
+The rule: **the admin-token prompt is for a deployment that actually requires a
+typed credential.** That is the non-loopback bind, which is the `hub` role. Any
+other role — `standalone`, `client`, or an absent tag — must not prompt.
+
+## Change
+
+In `gui/src/api-targets.ts`, add a sibling to `isConnectedRuntime()`:
+
+```ts
+/**
+ * May this dashboard ask the user to type an admin token?
+ *
+ * Only a hub does. A standalone loopback install mints its own session
+ * (src/server/gui-session.ts), so a refusal there is a Host/Origin
+ * misconfiguration that no typed token can repair — prompting for one asks the
+ * user to answer a question they did not cause and cannot diagnose (#3353).
+ * A missing tag reads as standalone, matching runtimeRoleFromDocument's
+ * existing safe default.
+ */
+export function adminTokenPromptAllowed(): boolean {
+  return runtimeRoleFromDocument() === "hub";
+}
+```
+
+In `gui/src/api.ts:resolveTokenAfter401()`, gate the prompt and record why it was
+skipped:
+
+```ts
+if (renewed.kind === "failed") return null;
+if (!adminTokenPromptAllowed()) {
+  state.promptCancelled = true;      // do not re-ask on every subsequent 401
+  reportSessionUnavailable(plane);   // surface an actionable notice instead
+  return null;
+}
+const prompted = await requestAdminToken(...);
+```
+
+`promptCancelled = true` matters: without it every failing request re-enters the
+resolution path. The existing `storeSession` already resets that flag when a
+session is later minted (`gui/src/api.ts:81`), so recovery is automatic once the
+misconfiguration is fixed.
+
+## What the user sees instead
+
+A dismissible notice, not a form. Copy names the real cause and the real fix:
+
+> **The dashboard could not start a session.** OpenCodex is running, but this
+> page's address is not one it recognises as local. Open the dashboard at the
+> address the proxy prints on startup (usually `http://127.0.0.1:<port>`), or see
+> the dashboard access guide.
+
+`reportSessionUnavailable` is a thin, testable seam: it dispatches a
+`CustomEvent` the shell renders. It must not be a `alert()` and must not block.
+
+## Verification
+
+`bun test gui/tests/api-auth-deadline.test.ts` plus a new case:
+
+- role `standalone` (and absent tag): after 401 + `unavailable` rebootstrap the
+  injected `adminTokenPrompt` spy is **not** called, the request resolves, and a
+  second failing request does not call it either.
+- role `hub`: the spy **is** called (the legitimate path stays intact).
+
+The role must be settable per test — the tests build their own `happy-dom`
+document, so the case writes the meta tag before `installApiAuthFetch()`.
+
+## Out of scope
+
+No server change. `issueGuiSession`, `requireManagementAuth`, and the CORS
+resolvers keep their current semantics; this phase only stops the GUI from
+asking a question the server never wanted asked.
diff --git a/devlog/_plan/260905_admin_token_local_ux/020_dialog_repair.md b/devlog/_plan/260905_admin_token_local_ux/020_dialog_repair.md
new file mode 100644
index 0000000000..3e4743988a
--- /dev/null
+++ b/devlog/_plan/260905_admin_token_local_ux/020_dialog_repair.md
@@ -0,0 +1,63 @@
+# 020 — Repair the dialog itself
+
+Work-phase `wp2`. Criteria `c-2`, `c-3`, `c-4`. Landed in #3491 and #3493.
+
+Two defects in one surface, deliberately split across two PRs because they have
+nothing to do with each other beyond sharing a file.
+
+## #3483 — the empty red notice
+
+`gui/src/admin-token-dialog.ts` builds the error element up front and hides it
+with the `hidden` attribute. That works only because the UA stylesheet says
+`[hidden] { display: none }`, and `gui/src/styles.css` overrode it:
+
+```css
+.notice { ... display: flex; ... }
+```
+
+Author origin beats user-agent origin. **Specificity never enters the
+comparison** — which is why this looked like a validation false-positive and why
+a `.notice[hidden]` fix would have been the wrong shape (it would still lose to
+the three-class `.startup-runtime-notice` rule at `0,3,0`).
+
+The repository had already solved this exact problem once, in
+`gui/src/styles-combos-workspace.css`, where a bare `display: flex` left both
+tab panels painted at once. That comment block is the precedent this fix
+follows: move the display onto `:not([hidden])`.
+
+Applied to `.notice`, `.notice-warn` (used without `.notice` in several places),
+and `.notice.notice-warn.startup-runtime-notice`.
+
+### Testing this required two tests, not one
+
+happy-dom applies no author stylesheet and performs no layout, so
+`expect(alert.hidden).toBe(true)` **passes today, unfixed**, while a real browser
+paints the box. A DOM assertion cannot see this class of bug.
+
+So the DOM test asserts the notice is hidden AND empty on open (keeping the two
+halves of "no error" from drifting), and a second test reads `styles.css` and
+fails any `.notice` rule that sets `display` without the guard. The second was
+driven red by reverting the CSS before being accepted.
+
+## #3353 — the box that explained nothing
+
+The dialog's only text was a title naming an environment variable. A user who
+had never set one had no way in.
+
+Ground truth, verified in source rather than assumed:
+
+- the proxy writes the token to `getConfigDir()/admin-api-token` on first start
+  (`src/lib/admin-secrets.ts`), `0600`, matching `/^ocx_admin_[A-Za-z0-9_-]{43}$/`
+- `OPENCODEX_ADMIN_AUTH_TOKEN` replaces it entirely and is not regex-checked
+- **no CLI prints it.** `ocx doctor` deliberately reports presence without ever
+  returning a value
+
+That last point is worth stating in the docs explicitly rather than omitting:
+hunting for `ocx token` is the obvious next move, and silence about it wastes
+the user's time.
+
+The dialog now carries a help paragraph plus a link to a new
+"Finding the admin token" anchor, styled like the existing in-app docs links
+(`target="_blank"`, `rel="noreferrer"`, accent colour) per
+`gui/src/pages/dashboard-dialogs.tsx`. Copy landed in all nine locales;
+`gui/tests/i18n-locales.test.ts` enforces key parity.
diff --git a/devlog/_plan/260905_admin_token_local_ux/030_windows_baseline.md b/devlog/_plan/260905_admin_token_local_ux/030_windows_baseline.md
new file mode 100644
index 0000000000..a0371ea812
--- /dev/null
+++ b/devlog/_plan/260905_admin_token_local_ux/030_windows_baseline.md
@@ -0,0 +1,82 @@
+# 030 — Windows baseline: what the failures actually were
+
+Work-phase `wp3`. Criterion `c-5`. Evidence, not a landed fix.
+
+A full Windows suite was already running on `desktop-c795oh4` when this unit
+started (`/c/ocxwin/repo` at `00834d710`, shards `base-1..4`). It was read-only
+evidence for this unit and was not disturbed. A second run on Bun `1.4.0`
+followed (`v140-1..4`).
+
+## The headline
+
+| run | Bun | fail |
+|---|---|---:|
+| `base-1..4` | 1.3.14 | 179 |
+| `v140-1..4` | 1.4.0 | 25 |
+
+Shard 1 alone went from 100 `(fail)` lines to 2. **The Windows suite was not
+telling us about 179 product defects; it was mostly reporting one harness
+failure over and over.**
+
+## Why 161 of them were one bug
+
+`tests/preload.ts` calls `acquireTestRunLock` (line 42) BEFORE it arms the guard
+with `OCX_TEST_HOME_GUARD=1` (line 64). On Windows the lock path needs the
+effective account SID, and under 4-shard load that lookup timed out:
+
+```text
+CodexUserIdentityRefusal: Windows effective-account lookup timed out.
+  at powershellValue (src/codex/user-identity.ts:252)
+  at resolveWindowsSid (src/codex/user-identity.ts:261)
+  at resolveDefaultTestRunLockPath (scripts/test-run-lock.ts:227)
+  at tests/preload.ts:42
+```
+
+The worker then ran with the guard permanently off, and every test that asserts
+"this helper is only available under the repository test preload" failed as a
+cascade: 44 in `codex-reset-credit-operation-ledger`, 68 in
+`codex-reset-credit-recovery`, 49 in `lab-fabric-task`.
+
+**That cascade had teeth.** With the guard down, two suites reached live
+machine state instead of being refused:
+
+- `windows-elevation-spawn.test.ts:89` expected `launcherPid: null` and got
+  `18144` — a real PowerShell process was launched.
+- `service.test.ts:1283`/`:1307` expected "refusing to mutate the
+  machine-global Windows Task Scheduler from an armed test process" and instead
+  got real scheduler-registration outcomes.
+
+So the ordering in `preload.ts` is not only noisy, it is the difference between
+a refused test and one that touches the developer's own Task Scheduler. Worth
+fixing on its own merits, independently of the Bun version that exposed it.
+
+## What survives on Bun 1.4.0
+
+25 failures in three groups:
+
+- `multi-account auth store` — 22 of the 25, one file. Not yet diagnosed.
+- `ocx v2 keep-native-v1` — 2. Dirac classified the `base` occurrence as a test
+  artifact: the product CLI exited 0 and the V2 disable took effect; only the
+  spy compares raw argv, and Windows `.cmd` invocation goes through the ComSpec
+  wrapper (`src/lib/win-exec.ts:79`), which is correct behaviour.
+- `cli wiring > interactiveGuardOk ... when cwd is unlinked` — 1.
+
+## Not fixed here, and why
+
+This unit's authority is the admin-token UX. None of the surviving failures are
+in that surface, and each needs its own reproduction on Windows before a fix is
+more than a guess — the `base` run's evidence is contaminated by the guard
+cascade, so a fix written against it would be written against an artifact.
+
+The honest carry-forward is three separate units:
+
+1. Arm `OCX_TEST_HOME_GUARD` before `acquireTestRunLock`, or make the SID lookup
+   fail closed instead of proceeding unguarded. Highest value: it is a safety
+   defect, not just a flake.
+2. Diagnose `multi-account auth store` on Windows.
+3. Decide whether `keep-native-v1` should assert on parsed argv rather than the
+   raw ComSpec string.
+
+Issue #3320 (non-ASCII account names misclassifying a valid scheduler task) is
+adjacent to (1) — both are Windows identity handling — but it is `needs-info`
+and was not reproduced here, so it stays open.
diff --git a/devlog/_plan/260905_admin_token_local_ux/040_delivery_record.md b/devlog/_plan/260905_admin_token_local_ux/040_delivery_record.md
new file mode 100644
index 0000000000..7c9e486298
--- /dev/null
+++ b/devlog/_plan/260905_admin_token_local_ux/040_delivery_record.md
@@ -0,0 +1,64 @@
+# 040 — Delivery record
+
+Work-phase `wp4`. All three PRs merged to `dev` with admin authority.
+
+| PR | merge SHA | what |
+|---|---|---|
+| [#3491](https://github.com/lidge-jun/opencodex/pull/3491) | `85e42117c` | #3483 empty red notice (CSS cascade) |
+| [#3496](https://github.com/lidge-jun/opencodex/pull/3496) | `3e65218ba` | never prompt a local dashboard |
+| [#3493](https://github.com/lidge-jun/opencodex/pull/3493) | `8b961b198` | #3353 dialog guidance + docs anchor |
+
+Ancestry proven for each: `git fetch origin dev` then
+`git merge-base --is-ancestor <sha> FETCH_HEAD` exit 0.
+
+Issues #3483 and #3353 closed with landing comments.
+
+## The stack-parent close race
+
+#3492 was the original middle PR. Merging #3491 with `--delete-branch` removed
+`codex/admin-token-empty-notice`, which was #3492's base, and GitHub auto-closed
+it. A closed PR cannot be retargeted, and it could not be reopened because its
+base branch no longer existed. #3496 carries the identical commits rebased onto
+`dev`; #3492 records the supersession.
+
+**For the next stack: retarget children to `dev` BEFORE merging the parent, or
+merge without `--delete-branch` and delete the branch afterwards.** The order
+that looks natural — merge, then restack — is the one that loses the PR.
+
+## Trailing CI, classified
+
+Two failures on green-otherwise heads, both proven flakes rather than assumed:
+
+- `test 3/4` on `6da1269f8`: `tests/responses-state.test.ts:1464` through
+  `fallbackPendingResponseSpills` (`src/responses/state.ts:714`), which is the
+  `remaining <= 0` deadline-budget branch in a 2 MB shutdown-spill test — a
+  loaded-runner timeout. The diff touches no `responses/` file; the file passes
+  141/141 locally twice; it passed on rerun.
+- `macos` on `87abc9152`: `tests/shutdown-launcher.test.ts:147`. That PR changes
+  only GUI copy, i18n, and docs — no runtime code at all. Passes 3/3 locally;
+  passed on rerun.
+
+Both were rerun to green rather than merged over.
+
+## Verification actually run
+
+No repository-wide local suite (prohibited for this task). Instead:
+
+- `bun run typecheck` — clean
+- `cd gui && bun test` — 1366 pass / 0 fail across 220 files
+- `bun test tests/gui-static.test.ts tests/server-management-auth.test.ts tests/server-auth.test.ts` — 143 pass / 0 fail
+- `bun run lint:gui`, `bun run privacy:scan` — clean
+- exact-head CI green on `d9a1afc71`, `6da1269f8`, `87abc9152` before each merge
+
+The full GUI suite earned its place: it caught 7 tests in `api-auth-memory`
+that assumed the prompt always fires, which the focused files did not cover.
+
+## One correction worth recording
+
+The first implementation gated the prompt on `runtimeRole === "hub"`. A review
+lane checking "would this lock anyone out?" found that it would: `standalone` +
+`hostname: "0.0.0.0"` is an operator who deliberately exposed the dashboard and
+must type the token, and `tests/server-management-auth.test.ts` already proves
+that bind mints no session. The role is topology; the question is the bind. The
+shipped predicate is `isApiAuthRequired`, the same one the server gates the mint
+with, and a regression now covers the exposed-standalone case explicitly.