docs: dashboard disable-login design + save session-resilience tasklist snapshot

This commit is contained in:
Joseph Doherty
2026-06-16 07:47:31 -04:00
parent c446bef64f
commit 0856cd4f93
2 changed files with 209 additions and 0 deletions
@@ -0,0 +1,141 @@
# Dashboard "Disable Login" Dev Flag — Design
**Date:** 2026-06-16
**Status:** Approved (brainstorming) — ready for implementation plan.
## Goal
A config flag that **disables login in the gateway dashboard**. When enabled, every
request is auto-authenticated as a fixed dev user (default **`multi-role`**) holding
**both** dashboard roles (`Administrator` + `Viewer`), so no login form, cookie, or LDAP
bind is involved and the whole UI behaves as a signed-in multi-role admin. Default
**off**. Mirrors the sister project OtOpcUa's `Security:Auth:DisableLogin` feature.
## Why / scope
Speeds up dashboard testing against the remote dev boxes (10.100.0.48, wonder) with no
sign-in round-trip and no GLAuth dependency. Scope is the **dashboard cookie web surface
only** — the gRPC API-key auth path (`authorization: Bearer mxgw_…`) and its scopes are a
separate auth model and are **untouched**.
## Background (current dashboard auth, verified)
- Dashboard auth is a **single cookie scheme** `MxGateway.Dashboard` registered in
`Dashboard/DashboardServiceCollectionExtensions.cs::AddGatewayDashboard`
(`AddAuthentication("MxGateway.Dashboard").AddCookie(...)`), plus a bearer scheme
`MxGateway.Dashboard.HubToken` (`HubTokenAuthenticationHandler`) for SignalR hubs.
- Real login: `/login``DashboardAuthenticator.AuthenticateAsync` → shared
`ILdapAuthService` bind/search → `IGroupRoleMapper<string>``CreatePrincipal` builds a
`ClaimsPrincipal` (`ZbClaimTypes.Name`/`Username`/`DisplayName` + one `ZbClaimTypes.Role`
per role + `LdapGroupClaimType` group claims; identity authType = the cookie scheme,
nameType = `ZbClaimTypes.Name`, roleType = `ZbClaimTypes.Role`) → cookie sign-in.
- Authorization: a custom `DashboardAuthorizationHandler` evaluates
`DashboardAuthorizationRequirement`. Policies: `ViewerPolicy` (AnyDashboardRole),
`AdminPolicy` (AdminOnly), `HubClientsPolicy` (cookie **or** hub-token scheme,
AnyDashboardRole).
- Roles: exactly two — `DashboardRoles.Admin` (`"Administrator"`) and
`DashboardRoles.Viewer` (`"Viewer"`).
- **Existing escapes (important):** `DashboardAuthorizationHandler` already short-circuits
when `Authentication.Mode == Disabled` or when `Dashboard.AllowAnonymousLocalhost`
(default **true**) and the request is loopback. **But both only `context.Succeed(...)`
the authorization requirement — they do not mint an authenticated principal.** So
`HttpContext.User.Identity.IsAuthenticated` stays false, `Identity.Name` is null, and
role-gated `AuthorizeView` write affordances stay **hidden**. That is precisely why they
do not deliver the "logged-in multi-role admin" experience this feature needs.
## Approach (chosen: always-authenticating handler under the cookie scheme name)
When the flag is **on**, **replace the `.AddCookie(...)` registration with a custom
`AuthenticationHandler` registered under the *same* scheme name**
(`DashboardAuthenticationDefaults.AuthenticationScheme` = `"MxGateway.Dashboard"`). Its
`HandleAuthenticateAsync` **always returns `AuthenticateResult.Success`** with the fixed
dev principal (configured username, both roles), shaped identically to what
`DashboardAuthenticator.CreatePrincipal` produces. `UseAuthentication()` stamps that
principal on `HttpContext.User` for **every** request.
Registering under the cookie scheme name (not a new name) is the load-bearing detail: the
`ViewerPolicy`, `AdminPolicy`, and `HubClientsPolicy` all resolve through that scheme via
`DashboardAuthorizationHandler`'s role check, so they pass with **no policy or page
changes**. The HTTP pipeline (Razor pages, admin endpoints), the Blazor circuit
(`AuthorizeView`, `[CascadingParameter] AuthenticationState`), and the SignalR hubs are
all covered by the single `HttpContext.User` seam. Because the handler authenticates every
request, the feature is inherently **global** (all clients, including remote browsers) —
the agreed scope.
`SignInAsync`/`SignOutAsync` are no-ops (no cookie to write or clear; the next request
re-authenticates through the handler).
**Alternatives rejected:** (2) mint the principal inside the existing
`DashboardAuthorizationHandler` bypass branches — authorization runs after authentication,
so `HttpContext.User` is set too late for the Blazor auth state, and two seams must agree
(this is essentially today's half-feature); (3) a pipeline middleware plus a stubbed
`AuthenticationStateProvider` — two components to keep in sync, and a page request still
302s to `/login` unless `HttpContext.User` is also set.
## Components
### 1. Config surface — two new fields on `DashboardOptions` (`MxGateway:Dashboard:*`)
- `DisableLogin` (bool, default **false**).
- `AutoLoginUser` (string, default **`"multi-role"`** — this project's GLAuth
Administrator test user). Used as `Name`/`Username`/`DisplayName` of the minted
principal; blank falls back to `"multi-role"`.
"All permissions" = principal minted with **both** `DashboardRoles.Admin` and
`DashboardRoles.Viewer`.
### 2. `DashboardAutoLoginAuthenticationHandler`
`AuthenticationHandler<AuthenticationSchemeOptions>` implementing
`IAuthenticationSignInHandler`. Mirrors OtOpcUa's `AutoLoginAuthenticationHandler`, adapted
to this project's claim shape (`ZbClaimTypes.*`, `DashboardRoles.*`). Always `Success`;
SignIn/SignOut no-ops.
### 3. Wiring in `AddGatewayDashboard`
Read `MxGateway:Dashboard:DisableLogin` directly from `IConfiguration` at registration
time (the same idiom OtOpcUa uses, since scheme registration precedes options binding).
- On → `AddScheme<AuthenticationSchemeOptions, DashboardAutoLoginAuthenticationHandler>(
"MxGateway.Dashboard", _ => {})` in place of `AddCookie`; the `HubToken` scheme stays
registered unchanged.
- Off → existing `AddCookie(...)` path unchanged.
### 4. Safety
- Default **off**.
- A **loud one-time startup `LogWarning`** ("DASHBOARD LOGIN DISABLED
(MxGateway:Dashboard:DisableLogin=true) — every request authenticated as '{user}' with
full permissions (Administrator, Viewer). Dev/test only; never enable in production.")
via the same options `PostConfigure<ILoggerFactory>` idiom OtOpcUa uses.
- The existing `AllowAnonymousLocalhost` / `Authentication.Mode == Disabled` escapes are
left untouched — `DisableLogin` is orthogonal (it changes *authentication*, minting a
principal, not authorization bypass); when it is on the authorization handler's normal
role-check branch succeeds, so the bypass branches simply do not matter.
## Error handling / edge cases
- Blank `AutoLoginUser` → falls back to `"multi-role"` (handler never mints a nameless
principal).
- `/login` still renders when the flag is on but is pointless (the user is already
authenticated); `POST /login`'s `SignInAsync` is a no-op. `/logout` is likewise a no-op.
No redirect added (YAGNI).
- No interaction with the gRPC API-key path — that auth is entirely separate.
## Testing
- **Handler unit test:** `HandleAuthenticateAsync` returns `Success`;
`principal.Identity.IsAuthenticated`, `Identity.Name == AutoLoginUser`,
`IsInRole("Administrator")` && `IsInRole("Viewer")`; blank-user fallback.
- **Wiring / integration (`WebApplicationFactory`):** with `DisableLogin=true`, an
`AdminPolicy`-gated endpoint returns 200 with **no** cookie, and a `/hubs/*` negotiate
authorizes; the startup warning is emitted.
- **Regression:** with the flag off (default), the real cookie handler is still registered
and existing dashboard auth tests pass.
## Docs to update in the same change
- `docs/GatewayConfiguration.md` — new `MxGateway:Dashboard:DisableLogin` /
`AutoLoginUser` options.
- The dashboard design doc (`docs/GatewayDashboardDesign.md`).
- The CLAUDE.md dashboard-auth note (alongside the `AllowAnonymousLocalhost` mention).
## Scope / verification
Gateway-server-side only (.NET 10, x64) — builds and tests entirely on macOS. No worker,
no `.proto`, no client, no gRPC changes.
+68
View File
@@ -0,0 +1,68 @@
# Saved Task List — Session Resilience Epic
> Snapshot taken 2026-06-16, before switching to the dashboard disable-login feature.
> This is the in-flight epic from `docs/plans/2026-06-15-session-resilience.md`.
## How to resume
```
/superpowers-extended-cc:executing-plans docs/plans/2026-06-15-session-resilience.md
```
The authoritative resume state lives in
`docs/plans/2026-06-15-session-resilience.md.tasks.json` (tasks 112 completed,
1328 pending). This file is just a human-readable mirror.
## Status
**12 of 28 tasks complete** (Phases 12 + reconnect core of Phase 3). All completed
work is merged to `main` (commit `c446bef`, pushed to origin).
### Completed — Phase 1 (Foundation)
- ✅ Task 1 (#108): Add OwnerKeyId to the session
- ✅ Task 2 (#109): SessionEventDistributor skeleton
- ✅ Task 3 (#110): Bounded replay ring buffer
- ✅ Task 4 (#111): Rewire AttachEventSubscriber + EventStreamService onto distributor
- ✅ Task 5 (#112): Per-subscriber backpressure isolation
- ✅ Task 6 (#113): Dashboard broadcaster becomes a distributor subscriber
### Completed — Phase 2 (Multi-subscriber fan-out)
- ✅ Task 7 (#114): Remove validator block + add subscriber cap option
- ✅ Task 8 (#115): Subscriber-lease collection + cap enforcement
- ✅ Task 9 (#116): Multi-subscriber end-to-end test (FakeWorkerHarness)
### Completed — Phase 3 (Reconnect core)
- ✅ Task 10 (#117): Proto — ReplayGap signal
- ✅ Task 11 (#118): Detach-grace session retention
- ✅ Task 12 (#119): Replay-on-reconnect + emit ReplayGap
### Pending — Phase 3 finish
- ⏳ Task 13 (#120): Owner re-validation on reconnect — blockedBy 12, 1
- ⏳ Task 14 (#121): Client ReplayGap handling — all 5 clients — blockedBy 10
- Carry the per-language presence-check idiom note for `optional` message fields.
- ⏳ Task 15 (#122): Reconnect integration test (fake worker) — blockedBy 12
### Pending — Phase 4 (Per-session dashboard ACL)
- ⏳ Task 16 (#123): gRPC session-owner gate + all-sessions admin scope — blockedBy 9, 1
- ⏳ Task 17 (#124): Session Tag + dashboard group-to-tag config — blockedBy 9
- ⏳ Task 18 (#125): EventsHub per-session ACL + hub-token tag claim — blockedBy 17
- Open decision: Viewer default (admin-sees-all vs strict per-session).
- ⏳ Task 19 (#126): ACL tests incl. live LDAP users — blockedBy 18
### Pending — Phase 5 (Orphan-worker reattach)
- ⏳ Task 20 (#127): Stable gateway-instance id + stable pipe naming — blockedBy 19
- ⏳ Task 21 (#128): Adoption manifest store (SQLite) — blockedBy 20
- ⏳ Task 22 (#129): Proto — worker adopt/reconnect frame — blockedBy 21
- ⏳ Task 23 (#130): Worker phone-home reconnect loop + self-terminate — blockedBy 22 (net48/x86, windev)
- ⏳ Task 24 (#131): Gateway adoption — re-open pipes, nonce-validate, reject impostors — blockedBy 23
- ⏳ Task 25 (#132): Resync adopted worker + ReplayGap to subscribers — blockedBy 24, 12
- ⏳ Task 26 (#133): EnableOrphanReattach flag (default off) + terminator fallback — blockedBy 24
- ⏳ Task 27 (#134): Gateway-restart reattach round-trip (WINDEV + live worker) — blockedBy 25, 26
- ⏳ Task 28 (#135): Documented-rule reversals + stillpending refresh — blockedBy 27
## Notes
- Phase 5 reverses the documented "Gateway restart does not reattach orphan workers"
rule (CLAUDE.md) — this was explicitly approved during design.
- Two deferred follow-ups noted earlier: dashboard visibility of `DetachedAtUtc` on
`DashboardSessionSummary`.
- Worker (net48/x86) tasks build/test on windev; everything else builds on macOS.