The same model can now be offered by several providers (e.g. gemma-4-31b-it on GWDG and Google). Models are identified by <provider_id>/<model_id>; previously the second provider's entry failed to sync.
php hawki init and php hawki run -build no longer run an unattended composer update when outdated packages are detected. Setup and build now only perform a deterministic composer install; upgrading packages is a deliberate decision. Previously the silent update could jump packages across major versions (e.g. laravel/ai 0.10 → 1.3), breaking the app with errors that first surfaced in unrelated commands like models sync.
Self-service access token creation was unavailable in the profile view: it still checked the pre-2.5.0 sanctum.* config keys instead of the new external_access.* keys, so the AccessTokens panel never appeared even with ALLOW_EXTERNAL_COMMUNICATION and ALLOW_USER_TOKEN_CREATION enabled.
Security: Fixed a privilege escalation in group chat invitations. Any room member (including viewers) could create an invitation for themselves with the admin role and accept it to become a room admin, or overwrite other users' pending invitations. Sending invitations is now restricted to room admins, invitation roles are validated, and accepting an invitation no longer changes the role of an existing member. Thanks to Artur Chlebicki for responsibly reporting this issue!
The room controls are now fully restricted for room admins.
Removing or leaving a room now deletes the member's pending invitations for that room, and leaving a room no longer fails halfway.
An admin leaving a room no longer deletes the whole room when other admins remain. The last admin of a room can't leave it anymore, only delete it.
Members can no longer change their own role, and the role of HAWKI can't be changed. Members can only be added to a room through invitations.
Model enrichment (pricing, context windows, feature flags) stopped updating because LiteLLM discontinued their catalog API. Lookups always fell back to the static files. The catalog is now read from LiteLLM's GitHub file.
Fixed attachments never reaching the AI conversation when STORAGE_DISK uses an s3-compatible server (e.g. garage, MinIO): uploaded files were never persisted from the temp area because whole "directory" moves are silently ignored by s3 drivers, so the attachment was never linked to the message. Each file is now copied individually instead.
The s3 disk in config/filesystems.php now fails loudly on storage errors and works on s3 servers without ACL support (see Internals section for details).
The check:storage artisan command now respects the configured S3_ENDPOINT and path-style addressing instead of always probing amazonaws.com.
Tool-call progress never reached the UI during streaming because the stream handler matched the tool-call data class instead of the stream event. Tool calls are now reported for every provider as tool_call/tool_result packets.
MCP tools failed because processes were decoupled using autoSse=true (the default). New default is autoSse=false.
Per-provider AI request timeouts via the adapter settings in model_providers.php: request_timeout (default 60 seconds, 0 = no limit) limits the complete answer of non-streaming requests. stall_timeout (default off) aborts a stream that receives no data for this many seconds, before or mid-stream — like HAWKI 2.4, streams are never cut off while they receive data, and silent streams are no longer aborted after 60 seconds. Supported drivers for stall detection: openai-compatible, ollama. GWDG ships with the HAWKI 2.4 behaviour (request_timeout=0, stall_timeout=120, via GWDG_REQUEST_TIMEOUT/GWDG_STALL_TIMEOUT).
Large chat messages fail to encrypt with a RangeError: too many arguments provided for a function call: the client-side encryption encoded the ciphertext with one String.fromCharCode.apply() over every byte, which exceeds the browser's argument limit at roughly 100k bytes. Base64 conversion now runs in 32k chunks, in both the encryption utils and the legacy bundle.
Multiple Mermaid diagram previews failed to render at the same time: generated diagrams could receive duplicate render IDs, causing the renderer to silently draw only one and skip the rest. Fixed upstream in markstream 2.0.15.
ai_models.model_id is now unique per provider instead of globally; system_models references models via the new ai_model_id column, system_models.model_id was removed.
Inside HAWKI a model is only referenced by its provider-qualified identifier <provider_id>/<model_id> (or its numeric id). This includes the JSON:API model_id of ai-models and system-models, the system-modelsmodel_id filter, chat requests on /req/streamAI, and the model stored with messages and usage records.
A migration converts the stored model references (ai_conv_msgs.model, messages.model, usage_records.model) to the qualified identifier. References to models that no longer exist are left unchanged.
The external app API (/api/ai-req) still accepts a plain model_id as long as only one provider offers it; otherwise the request is rejected with 400 and asks the client to use the <provider_id>/<model_id> format. A model that does not exist is rejected with 400 as well.
New ModelFinder (App\Services\Ai\Models\Identifiers\ModelFinder) resolves model references, AiModelIdentifier builds and parses the qualified form.
Provider keys in model_providers.php must not contain /, the config sync reports an error for them.
The model lists (config/model_lists/*.php) keep their format; the config sync combines the provider key with the model ID.
System models in model_providers.php (system_models, default_models, and the *_MODEL env variables) keep their format and also accept <provider_id>/<model_id>. If a plain model ID is offered by multiple providers, the config sync uses the model that was created first and logs a warning recommending the qualified ID (e.g. openAi/gpt-5-nano).
ai:tools:assign --model and usage:top-users --model accept the qualified identifier; a plain model ID matches the model of every provider offering it.
Breaking for custom code: AiModelRepository::findOne() / findOneOrFail() and ModelFinder::find() no longer resolve a plain model_id, pass <provider_id>/<model_id> or use AiModelRepository::findOneByProvider(). SystemModel::$model_id was removed, use ai_model_id / SystemModel::model().
The laravel/ai version constraint is now ^0.10.1 (was >=0.10.1), so dependency updates stay within semver-compatible releases instead of silently allowing a major jump. Lock file refreshed to laravel/ai v0.10.3.
LiteLLM catalog: downloaded once into a shared cache and filtered per provider, kept fresh via ETag revalidation every 5 minutes. During upstream outages cached data is served for up to 24 h, then the static files take over.
The static_llm_data updater no longer deletes the old files when the catalog is unreachable.
Added RoomRepository and InvitationRepository (App\Services\Chat\Room\Repositories). Room and invitation queries in RoomService, InvitationController, StreamController, the roomAdmin/roomEditor middlewares, the room broadcast channel and the migrate:avatars command now go through these repositories instead of Eloquent statics.
Replaced the unused POST /req/room/addMember/{slug} endpoint, which could add members without a room key, with POST /req/room/changeMemberRole/{slug} (route name web.roomEditMember unchanged).
Storage: driver-agnostic file persistence (app/Services/Storage/AbstractFileStorage.php): persistTemporaryFile previously moved the whole temp directory, which only works on local disks (via rename()); s3-style drivers treat directories as key prefixes and silently ignore such moves. It now copies each file individually (allFiles() + copy() + one deleteDirectory()), which Laravel's contract guarantees for every driver — a server-side CopyObject on s3, a native copy on local disks.
Storage: hardened s3 disk config (config/filesystems.php): throw => true turns failed Flysystem operations into logged exceptions instead of unchecked false returns (the reason the original bug stayed invisible); retain_visibility => false skips the GetObjectAcl pre-check before each copy(), which ACL-less s3 servers (garage, Cloudflare R2) reject.
Storage: check:storage now honors custom s3 endpoints (app/Console/Commands/CheckStorageConnection.php) by forwarding the disk's endpoint/use_path_style_endpoint to its S3Client instead of always probing amazonaws.com.
AbstractLaravelAgent passes the timeout explicitly to the Laravel AI package: send() the request timeout, sendStreaming() the stall timeout (or no limit when stall detection is off). The new AgentTimeouts value object reads and validates both values. For streamed requests, Guzzle uses its PHP stream handler, where timeout is a per-read timeout and curl options are ignored.