- New Platform.gd autoload: is_ios_safari() detects iPhone/iPad Safari
via JavaScriptBridge UA check (no Chrome/CriOS), gating all iOS memory paths
- ModelCache.release_model(url, etag): evicts a model from in-memory cache
on iOS Safari when a unit leaves the board; no-op on other platforms so
disk cache is always preserved for fast reload
- unit_mesh_builder: stores resolved model URL+ETag per unit (_unit_model_urls,
_unit_model_etags); calls ModelCache.release_model in the to_remove loop so
PackedScene refs are freed as soon as the unit is pooled or freed
- ContentCache.get_model_url_for_definition: stub for looking up a card's
Model3D CDN URL without a loaded unit dict (foundation for future eviction paths)
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
Three bugs fixed for abilities like Demonate Leviathan (SummonToken + TileFilter):
1. Server: AbilityUsesThisTurn was incremented (and costs paid) before effect-level
TileFilter validation ran. Added PreValidateClientTileTargets() called before the
use-count increment so an invalid tile pick wastes no uses and no sacrificed tokens.
2. Server: SummonToken silently break'd when TileFilter or token-affinity check failed,
giving the client no error. Changed both to throw GameException("INVALID_TILE").
3. Client: tile_passes_filters() only checked the target selector's filters, not the
effect-level TileFilter (e.g. SummonToken.tileFilter). Added _collect_effect_tile_filters()
and TileAffinityMatches handling so invalid tiles are never highlighted or accepted.
Also includes: Art2dUrl/ModelUrl deltas in GameStateDiffer/GameMessages/GameEngine,
unit_mesh_builder token-model URL fallback + shader/tint helpers, ModelCache v3 bump,
Game.gd unit delta merging for art2dUrl/modelUrl.
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
Adds evict_all() to ModelCache which sets _shutting_down before flushing
_pending with null callbacks, then clears _mem and _job_queue. This
prevents in-flight HTTP completions from firing into a freed game scene
and ensures each new game starts with a clean WASM heap. Called from
game_screen._exit_tree() after all signal disconnections.
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
Caps simultaneous GLB parses and HTTP downloads at _MAX_JOBS=4 via a
shared job budget. Disk-cache hits when slots are full queue as
parse_from_disk:true and drain correctly without re-downloading. Replaces
the no-op get_signal_connection_list().clear() with CONNECT_ONE_SHOT.
Adds _shutting_down flag so evict_all() can null-out pending callbacks.
Adds debug-gated logging on all hot paths.
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
When admin replaces a model/art file via the CDN browser, the S3 ETag is
now captured and stored on CardAsset.ContentETag. The ETag is included in
the content sync DTO (CardAssetDto) so clients receive it on every sync.
Client-side ModelCache now keys its disk cache on sha256(url + "#" + etag)
instead of sha256(url), so updating a file at the same CDN path busts the
disk cache automatically. A _CACHE_VER="v2" sentinel purges old-scheme
.bin files on first launch after the client update.
Changes:
- CdnUploadResult: add string? ETag field
- S3CdnClient: capture ETag from PutObjectAsync response
- CardAsset: add string? ContentETag property
- Migration 20260318000001_AddAssetContentETag: add content_etag text column
- ContentMessages: add string? ContentETag to CardAssetDto
- ContentDtoMapper: map ContentETag in CardAsset → CardAssetDto
- CdnBrowser.razor: UploadFileToKeyAsync returns Task<string?> (ETag);
ReplaceAsync calls PropagateAssetETagAsync which updates CardAsset rows
and touches CardDefinition.UpdatedAt to trigger a version bump
- ModelCache.gd: versioned disk key + cache migration + etag parameter on
load_model() and evict()
- game_screen.gd: pass model_asset.contentETag to load_model / retry chain
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
**Feature 1 — Model retry (client)**
- `ModelCache.gd`: add `evict(url)` to remove failed entries so re-download is possible
- `game_screen.gd`: when GLB load fails, `_schedule_model_retry` queues a 60-second
retry (up to 10 attempts); `_do_model_retry` evicts + re-downloads and swaps the
placeholder mesh for the real model if/when it succeeds
**Feature 2 — Rejoin on login (server + client)**
- Server: `GameGetActiveSessionHandler` (`game.get_active_session`) — returns the
caller's active session ID + phase, or null when no session exists
- Client: `init_screen.gd` checks for an active session after content sync; shows
a "Rejoin / Concede" dialog when one is found; rejoin navigates to GameScreen with
the session ID already set so `send_ready` → `AlreadyStarted` → `_fetch_state`
**Feature 3 — 5-minute inactivity detection (server + client)**
- `GameSession`: add `LastActionAt` column + EF migration
(`20260317000001_AddLastActionAt`)
- `GameHandlerBase.SaveStateAsync`: sets `session.LastActionAt` on every action
- `GameInactivityMonitorService`: background poll every 60 s; when a session hasn't
had an action in ≥ 5 min, pushes `game.inactive_warning` to the stalled player and
`game.opponent_inactive` to their opponent (once per session per server lifetime)
- `GameForceConcedeHandler` (`game.force_concede`): lets the waiting player claim
the win; validates caller is not the active player and `LastActionAt` is stale
- `Game.gd`: new `opponent_inactive` / `inactive_warning` signals + push handlers +
`send_force_concede()` helper
- `game_screen.gd`: handles `opponent_inactive` with Accept/Wait dialog;
handles `inactive_warning` with informational popup
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
- AbilitySystem: implement WithinRangeFilter (Manhattan distance min/max) and
IsAdjacentToSourceFilter in FilterMatches — both were falling through to
_ => true, so range constraints on targeting were silently ignored
- Program.cs: add GET /cdn/{**path} proxy endpoint with
Access-Control-Allow-Origin: * so browser web clients can fetch GLB models
(Ceph RadosGW / cdn.aeternalisgames.com has no CORS headers)
- ModelCache.gd: on web export, rewrite cdn.aeternalisgames.com URLs to go
through hub.aeternalisgames.com/cdn proxy; disk/memory cache still keyed
on the original CDN URL for consistency across platforms
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
- ModelCache._download: log result enum, HTTP code, body size,
Content-Type header, and first 16 bytes for every download — enough
to diagnose why Godot receives non-GLB data despite CDN serving valid GLB
- Moved disk-cache write to AFTER a successful parse so corrupt bytes
are never written; disk cache read now deletes + re-downloads on bad parse
- game_screen: replace raw `body` lambda captures with weakref to
eliminate "Lambda capture freed" noise (behavior unchanged)
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
- ModelCache: switch GLTFDocument.append_from_buffer (was append_from_file
which returned ERR_BUSY/43); validate GLB magic bytes; explicit signal
disconnect before queue_free to avoid freed-lambda errors
- ContentCache: add tile model fetch (content.tile_models), disk cache at
user://content/tile_models.json, get_tile_model_url() API
- game_screen: _add_tile now loads tile CDN model via ModelCache when bound
- Server: ContentGetTileModelsHandler (content.tile_models msg), registered
in ServiceRegistration; TileModelDto / GetTileModelsResponse added to
ContentMessages
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>
Client:
- Added ModelCache autoload (autoloads/ModelCache.gd)
- Downloads .glb files from CDN via HTTPRequest
- Disk cache: user://remote_model_cache/<sha256>.glb
- Memory cache: PackedScene per URL for instant re-instantiation
- Concurrent requests for same URL queue and share one download
- Loads via GLTFDocument.append_from_file + generate_scene
- game_screen.gd: _add_unit_mesh now looks up cardDefinitionId
→ ContentCache.get_definition() → finds Model3D asset URL
→ calls ModelCache.load_model() to async replace placeholder geometry
- Placeholder (sphere/cylinder/box) shown immediately
- Real model swaps in on download completion
- Guards against body being freed before model arrives
- Added _unit_bodies dict for O(1) body lookup
- project.godot: registered ModelCache autoload
Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)
Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>