flutter_gemma_agent 0.2.3

On-device agentic skills (SKILL.md + tool-calling loop) for flutter_gemma. Opt-in package adding text/JS/native-intent/MCP skill executors.

0.2.3

• feat: AgentSession.ask / AgentLoop.run accept an optional imageBytes — vision-capable agent turns (requires supportImage: true).
• feat: McpClient advertises MCP protocol revision 2025-11-25 + negotiated-version header on every request.

0.2.2

• AgentLoop delegates its tool loop to core generateChatResponseWithTools (single balancing impl); fed-back tool-response now carries the model's original call name.

0.2.1

• AgentSession.ask now streams the final answer token-by-token (VoiceSession parity); balance pending tool-calls on a mid-stream error.

0.2.0

• Add optional isCancelled to AgentLoop.run / AgentSession.ask for barge-in / cancellation.

0.1.0

• Initial pure-Dart foundation: Skill model + SKILL.md parser (Gallery-compatible).
• Sealed SkillResult (text/image/widget/webview/error) + SkillExecutor probe-chain abstraction.
• SkillRegistry (select/discovery string) + in-memory SecretStore (never in prompt).
• Agent loop over existing function-calling: 4 built-in tools, parallel-call dispatch, maxIterations guard.
• AgentSession facade (ask → AgentEvent stream) with two-stage skill discovery + injected secrets.
• TextSkillExecutor (0 deps): text-only persona skills resolve to a no-op acknowledgement.
• McpSkillExecutor + McpClient: call MCP tools over Streamable HTTP with a per-call permission hook (default-deny).
• NativeIntentExecutor: whitelist of 6 OS intents (email/SMS/calendar/notification/date) with param validation, behind user confirm.
• AgentChatView: chat widget over the AgentEvent stream with a collapsible progress panel + inline image/webview/widget results.
• Adaptive SkillManagerView + McpManagerView (bottom sheet on narrow, side panel on wide).
• SecretEditorDialog, McpToolCallPermissionDialog, SkillTesterView, and the add-skill security disclaimer.
• Bundled 8 starter skills (Gallery, Apache-2.0): JS/intent/text covering all four mechanisms.
• AssetSkillSource: load bundled SKILL.md skills + wire JsSkillExecutor to their bundled HTML.
• Verified on Android, iOS, macOS, and Windows; the agent is not supported on web yet (browser LLM runtimes don't reliably emit tool calls).