AIReasoningCore

AIReasoningCore is a native Swift AI runtime for iOS and macOS. Its public model is PiAILanguageModel, which conforms directly to AnyLanguageModel.LanguageModel and is used through LanguageModelSession.

The package deliberately contains no coding sandbox, shell, CLI-agent bridge, iSH runtime, provider compatibility server, Host capability framework, or second language-model protocol.

Modules

The package ships one library target:

AIReasoningCore
├── PiAILanguageModel
├── AnyLanguageModel ↔ pi-ai-swift mapping
├── Provider streaming and replay state
├── Reasoning and tool-call continuation
└── Provider asset delivery callback

External capabilities are ordinary AnyLanguageModel.Tool values owned and injected by the Host. Conversation persistence, browser/filesystem/network access, asset storage, and asset presentation are not shipped by this package.

Reasoning effort

Select a typed effort through the standard custom generation options seam:

var options = GenerationOptions()
options[custom: PiAILanguageModel.self] = .init(reasoningEffort: .high)
let response = try await session.respond(to: "Explain the tradeoff", options: options)

Use the selected ProviderModel.supportedReasoningEfforts from the runtime catalog to populate a picker. nil leaves the provider default unchanged; .off requests disabled reasoning and is offered only where representable. Unsupported selections fail explicitly in pi-ai-swift. Provider/model mappings remain owned by pi-ai-swift, not by the app or Core.

The typed seam is integrated with pi-ai-swift/main; the verified revision and acceptance boundaries are recorded in Docs/UPSTREAM_GATES.md.

iOS smoke verification

The deterministic iOS smoke app exercises the Core-owned provider adapter path: text and structured responses, reasoning, external Tool continuation, transcript checkpoints, replay metadata, and provider asset delivery. The Tool fixture is a small deterministic echo implementation; it is not a shipped Host capability. The app writes a machine-readable report that the runner reads back from the app container:

./Scripts/test-ios-smoke-simulator.sh

See Smoke/AIReasoningSmoke/README.md for the exact acceptance and claim boundaries. The baseline requires no API key or public network service.

Development setup

Dependencies are resolved through Swift Package Manager. Run:

swift test

pi-ai-swift tracks the latest commit on its public main branch. SwiftPM records the revision used by a particular checkout in Package.resolved; run package resolution/update before compatibility verification so that snapshot advances to the current remote main commit.

Provider runtime status

PiAILanguageModel accepts an injected PiAIProviderRuntime.ProviderRuntime. The adapter, mapping, structured output, reasoning, cancellation and Tool continuation are verified with a deterministic runtime. The iOS Smoke app also has a manual Live Provider Console for pi-ai-swift catalog, authorization, streaming, function-call, image-input, and image-output acceptance. Live results remain environmental evidence and do not replace deterministic fixtures.

AnyLanguageModel integration limits

• The maintained fork uses FM27-shaped reasoning transcript entries and explicit streaming error policies. With .preserveTranscript, callers can retain the latest reasoning/tool checkpoint and await waitForResponseCompletion() before saving. No partial answer is committed as a successful response. Non-streaming failures have no partial checkpoint channel; tool side effects are not rolled back.
• Reasoning text, opaque signature, redaction and provider metadata survive transcript persistence. Answer text stays separate. Signed answer text, tool thought signatures, response identity and other opaque assistant state still lack full transcript representation.
• Session-generated response transcript entries always use empty asset IDs. Provider assets are delivered to the optional PiAILanguageModel onAsset callback. The Host owns storage, presentation, and retention; a provider asset fails explicitly when no callback is installed.
• Formal dependency pins consume the published maintained fork. See upstream gates for exact revisions and verification.

License

GPL-3.0-or-later. See LICENSE and NOTICE.