Skip to content

v0.52.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 15 Jun 12:17
v0.52.0
b2909e7
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Highlights

See the exact prompt your backend rendered. A new opt-in GenerationConfig.captureRenderedPrompt makes the orchestration layer emit a .promptRendered(text:) GenerationEvent immediately before the first token — the formatted template string for local backends (GGUF/MLX), or the most-recent user message for cloud backends (#1879). It is off by default to avoid retaining sensitive prompt content, and is advisory metadata only — no chat-message state is mutated.

var config = GenerationConfig()
config.captureRenderedPrompt = true   // opt in

for await event in backend.generate(prompt: prompt, systemPrompt: system, config: config) {
    if case let .promptRendered(text) = event {
        print("Rendered prompt sent to the model:\n\(text)")
    }
}

Batteries-included context compression. DefaultCompressionPolicy ships three ready-made policies so hosts no longer hand-roll trimming: .truncating (drop oldest), .extractive (zero-inference scored selection with optional head-pinning to fight "lost in the middle"), and .anchored (inference-backed summary of old turns prepended to a verbatim recent tail) (#1885). Pass a tokenizer for a guaranteed budget, or omit it for an advisory chars/4 heuristic.

// Zero-inference scored compression, pinning the oldest 15% of the budget.
let policy = DefaultCompressionPolicy.extractive(
    headBudgetFraction: 0.15,
    contextSize: 8192,
    tokenizer: backend.tokenizer   // omit → advisory budget
)

Model-lifecycle & inference observability. A KeepAlivePolicy now idle-auto-unloads resident models (UnloadReason.idleTimeout) (5ca06aa); a ResidentModelStatus snapshot exposes what's loaded plus queuedRequestCount for backpressure-aware UIs (#1880); and InferenceMetric moves into ManifoldInference with FoundationBackend now reporting metrics through it (072c5dc).

Headless model selection. ModelSelection is now a headless, UI-free selection surface, and a public ModelPicker sample is built directly on it (#1873, #1877). This collapses the long-standing #1312 dual-write — selecting a model now clears the mutually-exclusive endpoint synchronously, so consumers can drop the binding-mirroring workaround (see breaking changes).

⚠ Breaking changes (pre-1.0)

Synchronous selection dual-write collapse (#1873) — ModelLoadCoordinator.dispatchLoad(_:) gains a defaulted drivesChatSeams: parameter (source-compatible, allowlisted in the API-break gate). Consumers that hand-mirrored ChatViewModel.selectedModel into ModelRegistry.selectedModel to work around #1312 can delete that mirroring — selection clears the mutually-exclusive endpoint synchronously now.

Features

Auto-injected tool system prompt (#1874) — templates that don't render tools natively now get a ToolSystemPromptBuilder injected automatically, so tool calling works on local non-Gemma models without the host hand-rendering tool definitions into the system prompt (closes #1856).

EmbeddingBackend capabilities + config contract (#1887) — EmbeddingBackend gains capability advertisement, and the GenerationConfig hint-vs-guarantee rule is now documented on the type (closes #1834).

Public ModelPicker sample + UI gap closers (#1877) — a ready-to-adopt picker over ModelSelection, plus attachImage / togglePin parity fixes (closes #1298, #1300).

Documentation

Recorded the June 2026 CI-cost reductions in the CLAUDE.md hygiene section (#1872).

Continuous Integration

Batched README-snippet compiles into one package and collapsed redundant push-to-main runs within a merge burst (#1870, #1871), and removed the glassbox-live-e2e job from the nightly workflow (#1878).

Assets 2
Loading