feat: first-class Foundation Models with native tools and dynamic profiles

[christopherkarani]

Summary

Makes Apple Foundation Models a first-class Swarm path for macOS/iOS and unblocks engineers who hit Conduit coupling, prompt-emulated tools, and Tool name clashes when mixing Swarm with Apple Intelligence.

Foundation Models (main change)

• Add FoundationModelsInferenceProvider that talks to LanguageModelSession without Conduit
• Bridge Swarm ToolSchema → Apple's FoundationModels.Tool using guided generation (DynamicGenerationSchema / GenerationSchema)
• Capture structured tool calls and return them to Swarm's agent loop (guardrails/observers still own execution)
• Default provider factory prefers the native FM path when the system model is available
• Deprecate Conduit-backed .foundationModels() / conduitFoundationModels() for migration only
• Document Swarm.Tool vs FoundationModels.Tool clash guidance

Dynamic Profiles (WWDC 2026–aligned)

• Add Swarm Profile, DynamicProfile, DynamicInstructions, ProfileMode, and ModeSwitchingDynamicProfile
• Re-resolve instructions, tool filters, generation overrides, and history policy every turn
• Wire via .foundationModels(profile:)
• Note: Apple's native LanguageModelSession.DynamicProfile is not in the macOS 26.2 SDK yet; this is the portable Swarm layer ready to bridge when the SDK ships

Also on this branch

• Harden workflow timeout race
• Remove stale MembraneHive import
• Version bump to 0.6.0

Test plan

• swift build
• swift test --filter DynamicProfile
• swift test --filter FoundationModelsNativeToolBridgeTests|FoundationModelsInferenceProviderTests|FoundationModelsToolCallingTests|LanguageModelSessionTool|ConduitProviderSelection|AgentDefault
• Live FM tests with SWARM_RUN_LIVE_FOUNDATION_MODELS_TESTS=1 (native tool capture + profile mode switch)
• CI green on macOS / Linux
• Manual smoke: Agent("…", inferenceProvider: .foundationModels()) { MyTool() }

Example

// First-class on-device (no Conduit)
let agent = try Agent("Be helpful.", inferenceProvider: .foundationModels()) {
    LookupTool()
}

// Dynamic profiles / baton-pass phases
enum Phase { case brainstorm, review }
let mode = ProfileMode(Phase.brainstorm)
let profile = ModeSwitchingDynamicProfile(mode: mode) { phase in
    switch phase {
    case .brainstorm:
        Profile(id: "brainstorm", instructions: "Ideate freely.",
                generation: .init(temperature: 1.0))
    case .review:
        Profile(id: "review", instructions: "Be precise.",
                history: .dropToolTranscriptAndKeepLast(count: 12),
                generation: .init(temperature: 0.2))
    }
}
let profiled = try Agent("…", inferenceProvider: .foundationModels(profile: profile))
mode.current = .review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: fff133e593

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve the existing system prompt

When a dynamic profile is used through Agent, the first message is already the agent's system prompt, but this replacement discards it and keeps only the profile/configuration instructions. In calls like Agent("You are a billing bot", inferenceProvider: .foundationModels(profile: ...)), the Foundation Models turn no longer receives the agent instructions unless the caller duplicates them in the provider configuration; merge the new instructions with the existing system content instead of replacing it.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Honor specific tool choice before bridging

This only removes tools for .none; for ToolChoice.specific(toolName:) it still bridges every resolved tool to FoundationModels. Because the later prompt text is only a hint, requests that are supposed to force search can still let the native model call calc or any other provided tool; filter effectiveTools down to the named tool for this case.

Useful? React with 👍 / 👎.