Hunt startup and worker operation

Deploy the matching UndauntedInternalServer.dll and deploy-server build together. The deploy server creates a private per-launch readiness marker automatically and waits up to 90 seconds before returning a travel address. Newer native servers write the marker only after InitListen succeeds. Existing installs with an older DLL are also supported: the deploy server verifies that the assigned UDP port is actually bound before matchmaking can return IN_PROGRESS. GAMESERVER_READY_DIR may override the marker directory.

Launches retain their configured spacing and memory reservations, but one cold world no longer blocks every following launch. Client queue/status requests wait at most 1.5 seconds for an allocation (the party leader’s initial join allows 2.5 seconds), then keep reporting matching until it ends. Temporary worlds allow 180 seconds for the first player to finish connecting. After a player connects, the native empty-world timer resets and allows 30 seconds after the last player leaves. Unavailable host capacity still has a bounded one-minute queue; more time cannot create CPU, RAM or free ports.

Recovering from failed matchmaking

The metagame bounds an allocation request, including its JSON response body, with MATCHMAKING_DEPLOY_TIMEOUT_MS (240000 milliseconds by default). A stalled request ends that candidate with FAILED. It does not automatically allocate another server: a request that timed out may already have started a remote process. A new explicit join can retry. Configure the deadline above the worker startup grace and launch spacing; it is an upper bound, not an extra delay on successful hunts.

Ready candidates carry the deployment’s immutable session ID. A cached, shared /gameservers observation invalidates a dead session even if another process has reused its port. The deploy server’s complete flag distinguishes an empty fleet from an unavailable worker. Missing sessions are treated as dead only in a complete snapshot; positive observations also work when another worker is unavailable. Older or unreachable deployments fall back to a fixed reuse lease anchored to the first travel response, rather than extending it on every retry. Party rejoin retention remains bounded separately. Status polling never starts a replacement for a failed allocation by itself.

Ramsgate, Training and tutorial joins install their pending candidate before waiting on deployment. Completion updates that exact candidate only. A newer hunt or an explicit cancellation cannot be overwritten by the older response. Repeated joins also retain the original 20-second park deadline; they cannot keep restarting that pause.

Expected players and native compatibility

A queued player’s place requires a join or status poll within 45 seconds. Expired players are removed immediately before a public queue is sent to deployment, including retries after a capacity refusal. The remaining active players retain their candidates and region. Allocation and loading use separate grace periods; stopping matchmaking polls during a normal client load does not immediately remove the player.

The inspected 1.4.4 executable sends the complete expected-player roster to /candidate/player/alive. Its response handler changes the roster and expected count only for a different, positive count; it ignores an empty array. Its separate KeepAlivePlayerStatusEndpoint key is absent. Therefore the expected list itself cannot be used as evidence that those players connected.

The matched native DLL supplies its deployment session and a bounded snapshot of connected player IDs through the existing authenticated HTTP request hook. Engine objects are read on the game thread, and the HTTP hook reads only copied strings. Pending or unidentified connections, a stale snapshot, or missing native support leave the observation incomplete. The metagame preserves unresolved players in that case. Complete observations retain connected players, allow an initial 180-second loading grace, and allow a 45-second grace for a previously connected player’s disappearance. This permits a nonempty roster of actual arrivals to replace one that still contains abandoned players. Unknown sessions after a metagame restart retain their roster conservatively.

The native channel index also observes the existing actor/channel setter so a new actor’s first lookup does not scan every channel opened before it. Positive lookups still validate current array slots; removal, replacement and reentrant assignment are covered by the regression suite. If the setter hook cannot be installed, lookup keeps its conservative fallback. See CHANNEL_LOOKUP_TESTING.md for the exact binary checks and performance regression.

Update the central metagame, the deploy servers on each worker, and the server DLL together. A client-side DLL replacement alone does not update a remote hunt host. Test private and public hunts, a party retry, a player disconnect during loading, and a worker exit before promoting the build. Source tests and a native startup smoke test do not establish playable end-to-end client loading times.

The installation has two servers: the main host and one overflow worker connected by an OVERFLOW_DEPLOYSERVER_URL loopback SSH tunnel. Existing hunts are not live-migrated between hosts. Both use the main metagame’s account and save database.

For native builds use MSVC 14.44 or newer. On machines with multiple v143 versions, pass /p:VCToolsVersion=14.44.35207 to MSBuild. The older 14.36 compiler rejects the generated SDK’s uninstantiated static_assert(false) templates.

Store

The central metagame owns the store and inventory; workers use that same backend. Set STORE=free to enable the shipped catalog. Offers currently cost zero. See UndauntedMetagame/STORE_PRICING.md before changing earned-currency prices. No real-money payment processing is provided.

Set STORE_CATALOG_PROFILE=curated30 for the 30-cosmetic storefront listed in UndauntedMetagame/src/vendor/store_curated_30.json. Set STORE_REPEATABLE_TOKENS=0; the curated profile excludes bounty-token bundles even if that flag is enabled. Other offers cannot be purchased by bypassing the listing and requesting their SKU. The separate Hunt Pass endpoint remains available.

To revoke the earlier free-store purchases, stop both hosts’ game processes and the metagame, back up the database, then run scripts/revoke-free-store.cjs <UTC-cutoff> from the metagame directory with its protected environment loaded. This previews the recorded grants. Add --apply to revoke them. It preserves pre-owned and separately earned items, removes only recorded store instance IDs/stack quantities, and revokes entitlements only where their source is the matching store SKU. Append-only audit markers prevent a second removal. Unredeemed pre-cutoff tokens expire. Paid purchases abort the operation. The command checks that character and escalation progress is unchanged; it is not an account rollback.