README

Documentation: https://hypervel.org/docs/ai-sdk

Differences From Laravel

Agent and stream-event broadcast() methods deliver events immediately by default. Pass now: false to queue events; queued generation remains available. See broadcasting.

Agent broadcasts batch adjacent text/reasoning deltas by default within a 50 ms window, retaining original events for replay. Set ai.broadcasting.batching.enabled to false for individual-event delivery. See broadcast batching.

Queued AI jobs' failed() hooks accept ?Throwable so manual failures without an exception reach their callbacks. Overrides must accept null as well.

Promptable::getProvidersAndModels() returns an ordered list of [Provider|string, ?string] pairs instead of a name-keyed map, preserving provider objects and their credentials. Update overrides to return pairs. Failover retains repeated entries as separate attempts. The public Provider::formatProviderAndModelList() utility retains its name-keyed return format. See failover.

Pending requests' protected resolveProviderOptionsAndHeaders() helper accepts the provider contract so custom implementations work. PendingEmbeddingsGeneration's protected caching helpers receive precomputed identities or keys, keeping cache reads and writes consistent without repeated identity calculation. Subclasses overriding these helpers must match the Hypervel signatures; public generation and caching methods are unchanged.

RunContext::startingStep() accepts optional trailing ?string $instructions = null and array $tools = [] parameters so listeners receive the resolved step inputs. Overrides must accept these parameters too.

Custom conversation stores must implement ClaimsPendingApprovals to resume stored tool approvals safely. Claim-owned per-result recording replaces the protected ResumesToolApprovals::storeApprovalResultRecorderFor() helper, which is removed. Stores read AgentPrompt::approvalClaim() when settling the turn. Ordinary conversations and stateless resumption remain available without that capability.

RememberConversation replaces protected openTurn() with rememberTurn(): StoredTurn so the database store can persist a complete turn atomically. Update overrides to use this helper; public agent calls and standalone store methods are unchanged.

Skill discovery is cached for the fixed paths in ai.skills.cached_paths. Reload services after changing those files; skills loaded during application boot require a full restart. Other skill sources remain dynamic. Set cached_paths to [] to disable caching. See skills.

The conversation migrations use Hypervel's column types, indexes, and approval claim fields. Existing Laravel conversation data requires a schema migration before use.

Remote downloads share the HTTP client's destination checks. The protected UntrustedUrl::validate(), resolve(), and isBlocked() helpers and hostname-blocklist constants are removed; resolveUsing() remains available. URLs containing embedded credentials are rejected. Downloads are limited to 32 MiB of decoded content by default, instead of Laravel's 100 MB; set ai.remote_files.max_size in kilobytes to change the limit, or to null to disable it. See remote attachments.

Known Issues

Bedrock streaming is affected by two AWS SDK decoder issues. A valid stream can fail with a checksum error when an event's checksum arrives in parts (#3371). The final event can also be lost (#3374): token usage then reads zero, or an AWS error sent as the last event is reported as an incomplete stream. These limitations remain until the supported AWS SDK version includes the fixes.

Like Laravel AI, nested objects in tool arguments can lose their JSON type when decoded to PHP arrays: {"options":{}} can become {"options":[]}. This can fail validation in MCP or Vercel integrations that require an object. This behavior is retained for now; upstream reports and fixes will be considered during normal syncs.

Ported from: https://github.com/laravel/ai

hypervel / ai

Package info

Security

Statistics

README

Differences From Laravel

Known Issues