feat(agent): runStreaming(_:images:) for vision

CHANGELOG.md

@@ -9,6 +9,7 @@ All notable changes to SwiftAgentKit will be documented in this file.
 - **ContextSift: stop the `artifact_read` retrieval loop.** A large *active* tool result was bounded to `maxActiveResultChars` and spilled to an artifact with a "use artifact_read" hint; when the model then called `artifact_read`, that retrieval result was *itself* re-truncated and re-spilled — so the model kept calling `artifact_read` to "get the full output" that never fully surfaced (observed as ~6 repeated `artifact_read` calls until `maxTurns` ran out). Fix: retrieval tools (`artifact_read`, `artifact_search`) are now exempt from active-display truncation and receipt spilling — their output is shown in full (the tools already page via `offset`/`limit`). Also raised the default `maxActiveResultChars` 2000 → 8000 so ordinary tool outputs fit inline without any artifact round-trip. Regression coverage added (an active `artifact_read` result larger than the bound is shown in full, never re-truncated).
 
 ### Added
+- **`runStreaming(_:images:)`** — a streaming overload that accepts `[LLMImage]` for vision-capable models (attaches the images to the user turn; the ReAct loop is otherwise identical). The no-image `runStreaming(_:)` now delegates to it.
 - **Self-improving skills — persist skills the agent authors at runtime.** New `AgentSkillStore` protocol + `FileAgentSkillStore` (one human-editable markdown file per skill: `# name` / `Triggers:` / instructions) and a built-in `LearnSkillTool` (`learn_skill`). Set `agent.skillStore = …` and the agent (a) loads previously authored skills into its `SkillRegistry` and (b) auto-registers `learn_skill` — so it can turn a recurring task, or a corrected mistake, into a reusable keyword-triggered skill that fires immediately (added to the live registry) and persists across sessions. Mirrors the `memoryStore` → `RememberTool` pattern. Regression coverage added (store round-trip; `learn_skill` persists + activates on a matching query).
 - **`SwiftAgentKitTools` — a new opt-in product of ready-made native tools.** Import it and register the ones you want; the app supplies the confirmation UI/policy via `AgentCallbacks.onToolConfirmation`.
   - **Filesystem** (Foundation-only, cross-platform): `read_file` (paged), `write_file` (append/overwrite, `requiresConfirmation`), `list_dir`, `search_files` (name and/or content substring, bounded traversal).

Sources/SwiftAgentKit/Core/Agent.swift

@@ -1079,6 +1079,12 @@ public final class Agent: @unchecked Sendable {
     ///   text, but on the final turn the original deltas have already been
     ///   streamed — so a rewrite won't retroactively change what the caller saw.
     public func runStreaming(_ query: String) -> AsyncThrowingStream<String, Error> {
+        runStreaming(query, images: [])
+    }
+
+    /// Streaming variant that accepts images for vision-capable models. The images
+    /// are attached to the user turn; the rest of the ReAct loop is identical.
+    public func runStreaming(_ query: String, images: [LLMImage]) -> AsyncThrowingStream<String, Error> {
         return AsyncThrowingStream { [weak self] continuation in
             let task = Task { [weak self] in
                 guard let self else {
@@ -1088,7 +1094,7 @@ public final class Agent: @unchecked Sendable {
                 do {
                     _ = try await self.runLoop(
                         query: query,
-                        images: [],
+                        images: images,
                         onText: { continuation.yield($0) }
                     )
                     continuation.finish()