• Updated package documentation for the published 0.3.x release line.

Added #

• TraceContext.fromTraceParent, fromHeaders and toHeaders: read a W3C trace-context header and continue that trace, or write the headers that continue it in the next hop. Malformed input returns null rather than throwing — a header an outside caller controls being wrong means "start a new trace", not "fail the request" — and header names are matched case-insensitively, because a lookup that misses silently starts a new trace for every request.
• TraceContext.traceState, carried through untouched. The framework never reads it and never invents it; it only refuses to drop what a backend put there.
• AgenticContext.root(traceContext:), so a server can join the trace that arrived with a request instead of starting its own.

• JsonSchema.coerce treats null for an optional, non-nullable property as not given, applying the default if there is one. Models send null for parameters they skip, and ToolExecutor used to reject those calls. null for a required property is still rejected.

• UntrustedContentLedger on AgenticContext.untrustedContent: which tools brought text into a run that the application did not write. Shared by every scope in a run, and impossible to clear.

• Shortened the package description to the 60-180 character window pana scores against. Search engines truncate anything longer, so the ten points it withheld were pointing at a real defect: the useful half of the sentence was never being shown.

Initial release of the foundation layer.

Added #

• Messaging — Message, MessageRole and a sealed ContentPart hierarchy covering text, reasoning, images, files, audio, tool calls and tool results. ConversationHistory extensions for trimming and for finding unanswered tool calls. TokenUsage with cache and reasoning accounting, and addition.
• Schema — JsonSchema, an immutable validating subset of JSON Schema matched to what LLM providers actually accept, with full-violation reporting, coerce for the near-misses models really produce, and toStrict() for guaranteed structured output.
• Errors — AgenticException and eighteen concrete failures, each with a stable code, an isRetryable answer and JSON serialisation. Errors accumulate annotations as they propagate.
• Cancellation — CancellationToken/CancellationTokenSource with callbacks, future racing, stream binding that closes the underlying socket, token merging and clock-driven deadlines.
• Resilience — RetryPolicy with budgets and Retry-After support; ExponentialBackoff, LinearBackoff, ConstantBackoff and FixedScheduleBackoff with four jitter modes; CircuitBreaker with a half-open probe phase.
• Events — AgenticEvent and BroadcastEventBus with bounded replay for late subscribers, typed on<T>() filtering and run-scoped selection.
• Telemetry — StructuredLogger with bound fields, child scopes and credential redaction by default; Tracer with spans, events, sampling and pluggable exporters.
• Primitives — web-safe monotonic Ulid, injectable Clock, Disposable and DisposableBag, checked JSON accessors, Result, typed Registry.
• Context — AgenticContext bundling run identity, logger, bus, tracer, clock, IDs, cancellation and deadlines, with scoped derivation.
• Testing — package:agentic_core/testing.dart exports FakeClock, SequentialIdGenerator, InMemoryLogSink and InMemorySpanExporter.