Files
Dune/docs/SEAT_AI_PERMISSION.md
T
Quintine b21b2bc378 Add one-use consent for a named player to start seat AI
Let a human owner authorize another human in the started room to start a fixed difficulty once within 24 hours. The owner retains the seat, cookie, private hand, recovery setup and takeback control. Consent is explicit even while online; neither host status nor inactivity authorizes a takeover. Activation enters the existing paced private-view AI path without strategic changes.

Add additive migration 0006, both exact session fences, owner/delegate-only metadata in the authenticated room snapshot, atomic CAS activation and consumption, expiry and terminal retry receipts. Replacement hides older consent; own AI changes invalidate unused consent. Recovery or handover of either bound seat invalidates the grant. Future-version guards prevent stale computed state from overwriting a racing action. Exact browser retry records retain the original operation/version across refresh and cannot rearm old consent. Shared HTTP continuation preserves coalescing across ordinary and control routes.

Independent review found no material remaining issue. Revoke messaging accurately handles activation winning the race. Browser QA exercised grant/revoke, preserved consent after refresh, reciprocal fixed-profile activation, consumed receipts, keyboard takeback and private ownership. The panel now remains expanded after success and action buttons are at least 44px. Responsive screenshot acceptance remains open because the browser viewport override targeted a failed secondary-origin tab; desktop controls and no-overflow width were verified. The QA room is stopped at a genuine human setup decision with both seats taken back.

Types, lint, all 4906 offline tests, production build and all 45 HTTP tests pass on the stable source fingerprint. Seven new offline cases cover client consent/retry validation and real SQLite privacy, restart, one-use races, future versions, takeback, expiry and both-seat recovery/handover; two new HTTP cases cover authentication, injection, fixed-profile continuation and takeback. The first broad run exposed 68 legacy VM fixtures missing the new production dependency; their mappings were updated and the real shared continuation remains exercised. The historical 0005 migration test verifies backfill before applying 0006 for current code.

All 1174 original rooms, 2456 seats and every prior recovery/entry/handover/migration row remain unchanged. SQLite integrity and foreign keys pass; migration and QA added 88 rooms and 188 seats. No reset, server restart, usage check or automation was performed.

The current Sites integration has no supported durable Cron/DO provisioning path. Document this platform blocker rather than adding an inert scheduled export; unattended resumption and broader network-failure acceptance remain unfinished. Full rules, component, AI strength and publication gates remain closed.

Verified source tree: 2d86fc611e9ef4f5a10db10f96687a31db1ec238908ecf1e9333c50fac9ccaa8
2026-09-20 16:30:37 +10:00

66 lines
3.8 KiB
Markdown

# Permission to start another player's AI
An owning human can authorize one named human player in the same started room to
enable a chosen AI difficulty once, within 24 hours. This is optional operational
consent. It does not change board-game rules, transfer ownership or disclose the
owner's private view. It can be used while the owner is online; no inactivity
detector or host privilege substitutes for consent.
## Connected behavior
Below the table, **Let a player start AI for your seat** offers a named player and
Easy, Medium, Hard or Brutal. The owner can replace or revoke an unused grant.
The delegate sees its fixed difficulty and expiry and can activate it once.
Both retain their existing credentials, recovery kits and private views. Public
history names the activating delegate. The owner uses the existing **Take back
control** button; completed AI decisions remain in the game.
The normal `setAutopilot` control branch preserves gameplay and sealed choices.
Activation queues the existing persisted, 1.5-second AI continuation and does not
implement new AI strategy. Setup/playing human seats are supported. Lobby,
finished, permanent-bot and already-controlled owner seats reject new consent or
activation. A delegate must be a different authenticated human in the room.
## Consent and saved continuation
Migration `0006_thick_imperial_guard.sql` adds a separate consent table. Each grant
binds its identifier, owner, delegate, chosen difficulty and expiry to both exact
current seat-session hashes. Only the owner and delegate receive its public
metadata, through the same SQL read that authenticates the viewer and loads the
room. Neither hash is projected. Recovering or transferring either seat makes
old consent unusable; a new owner or delegate must establish fresh consent.
Creation, replacement, revocation and activation use room-version and credential
predicates. A D1 transaction consumes the grant only after its exact versioned
room update succeeds. Taking back or changing one's own AI control invalidates
unused consent in the same transaction. A losing concurrent action cannot leave
the grant consumed without activation, or reactivate a seat after takeback.
Grant identifiers remain durable receipts. Retrying creation can report its
original result but never renews or rearms a used, expired, revoked or replaced
grant. Retrying activation reports prior use without repeating control changes.
Replacement hides older grants from the current permission list while retaining
their receipts. Expiry is enforced by server time.
The browser saves the exact operation, grant identifier and original version in
tab storage before dispatch. An uncertain result offers an explicit exact retry;
refresh restores it. Retry never substitutes a newer version or a new grant.
Clearly rejected operations release the retry record. Explicit abandonment
explains that discarding local details cannot cancel a committed request. No
cookie, recovery secret or private hand is stored in this record.
## Verification and remaining work
`seat-ai-delegation-recovery.test.ts` exercises real SQLite migrations and the
production store, including one-use custody, both session fences, rejected-action
immutability, races and restored continuations. The client tests verify exact
retry storage and target/profile injection rejection. HTTP tests cover session,
origin, private projection, fixed-profile activation, background work and
takeback. Independent review, browser evidence and final check/build/HTTP results
belong to the private source-bound checkpoint and Git message.
This is a bounded prototype of preauthorized absence handling. An unprepared lost
seat still cannot be commandeered by a host. Full network-failure acceptance and
durable unattended scheduling remain unfinished. No mode or publication gate is
opened, and strategic AI work remains deferred until non-AI feature completion.