dart_mcp_core 1.5.1

Production-ready pure Dart core for Model Context Protocol (MCP). Agent execution engine, LLM adapters, MCP client transports, and tool orchestration — no Flutter dependency.

1.5.1

• Tool Parameter Schema Sanitization (LLMService.sanitizeSchemaForLlm):
  • Recursively resolves and inlines $ref pointers from $defs / definitions.
  • Strips unsupported JSON schema properties ($schema, default on Gemini) to eliminate HTTP 400 schema validation errors on strict LLM providers such as Gemini 3.7 Flash and strict OpenAI-compatible endpoints.

1.5.0

• Multi-LLM Profiles & Configuration Schema (LlmConfig):
  • Added profile identification attributes: id, name, displayName, and profileId.
  • Added clone({String? newId, String? newName}) for profile duplication and migration.
  • Added execution and turn limits: maxToolCalls, maxTurns, and turnLimit.
  • Changed default maxToolIterations / maxToolCalls from 400 to 30 for safe guardrails against runaway tool loops.
  • Added resilient JSON parsing supporting snake_case (max_tool_calls, max_turns, turn_limit, api_key, base_url).
• Infinite Tool Loop Protection (ToolLoopGuard, LoopEvaluation, LoopAction):
  • Canonicalized argument hashing to detect repetitive tool calls.
  • Injects loop warnings upon 3 consecutive identical calls or alternating patterns (A-B-A-B-A-B).
  • Automatically blocks tool execution at 4+ identical calls, preventing runaway token consumption.
• Execution Limits & Session Auto-Rotation (McpExecutionLimits, SessionRotator):
  • Added McpExecutionLimits for turn thresholds (maxTurns, turnLimit, maxToolCalls).
  • Added SessionRotator for automatic rotation of session log files when turn thresholds or file size limits are exceeded.
• Bulk Serialization for EpisodicMemoryStore:
  • Added loadEntries(List<EpisodicMemoryEntry> entries, {bool replace = false}) and loadFromJsonList(List<dynamic> jsonList).
  • Added bulk toJson() and toJsonList() for persisting conversation archives.
  • Added EpisodicMemoryStore.fromJson(Map<String, dynamic> json) factory constructor.
• Dynamic Model Override in Sub-Agent Delegation (DelegateTaskTool):
  • Added optional model / modelOverride support to automatically apply model overrides to child agent's llmConfig.
• Shared Workflow Blackboard Memory (WorkflowMemoryStore, StoreMemoryTool, RecallMemoryTool):
  • In-memory key-value scratchpad for passing structured state across multi-agent workflows and sub-agents.
  • Automatically injected tools (store_memory, recall_memory) when workflowMemoryStore is provided to Agent or McpAgentEngine.

1.4.0

• Suite-Wide Architecture Synchronization:
  • Aligned package versioning with companion engine extraction (flutter_ui_mcp_embedded).
  • Strengthened pluggable LLMService.embeddedHandler and LLMService.embeddedStreamHandler interface stability.
• Shared Workflow Blackboard / Scratchpad Memory (WorkflowMemoryStore, StoreMemoryTool, RecallMemoryTool):
  • In-memory key-value scratchpad store for passing structured state across multi-agent workflows and sub-agents.
  • Automatically injected tools (store_memory, recall_memory) when a workflowMemoryStore is provided to Agent or McpAgentEngine.
  • Sub-agent clones and targets inherit the parent's workflow memory store automatically.
  • Emits AgentWorkflowMemoryEvent on storage and retrieval operations.
• Dynamic Model Override in Sub-Agent Delegation (DelegateTaskTool):
  • Added optional model / modelOverride parameter to delegate_task tool schema.
  • Automatically applies model override to child agent's llmConfig (copyWith(model: model)) for specialized sub-tasks without requiring pre-registered agent configurations.
• Bulk Serialization for EpisodicMemoryStore:
  • Added loadEntries(List<EpisodicMemoryEntry> entries, {bool replace = false}) and loadFromJsonList(List<dynamic> jsonList).
  • Added bulk toJson() and toJsonList() for persisting conversation archives to DuckDB or disk.
  • Added EpisodicMemoryStore.fromJson(Map<String, dynamic> json) factory constructor.

1.3.0

• Typed Multimodal Parts (ChatContentPart):
  • Introduced ChatContentPart hierarchy (TextPart, ImagePart, DocumentPart, AudioPart) for rich multimodal message representations.
  • Native base64/URL/bytes image payload forwarding across OpenAI (ContentPart.imageUrl), Anthropic Claude (InputContentBlock.image), Google Gemini (Part.inlineData), and Ollama (images).
  • Added ergonomic factories: ChatMessage.user(), ChatMessage.assistant(), and ChatMessage.system().
• Episodic Memory Store & Recall (EpisodicMemoryStore, RetrieveMemoryTool):
  • Prevents permanent information loss during context window auto-compression by stashing raw turn history in an indexed episodic archive before summarization.
  • Injects retrieve_memory tool so the model can inspect exact past code snippets, hashes, and tool outputs on demand.
  • Emits AgentMemoryRetrievedEvent when the model recalls archived turns.
• Sub-Agent Delegation (DelegateTaskTool):
  • Native task delegation enabling primary agents to delegate focused sub-tasks to specialized sub-agents or clean agent clones (assignee: 'clone').
  • Multi-level recursion guards (isSubAgent, delegationDepth) preventing infinite delegation loops.
  • Real-time delegation event lifecycle: AgentSubAgentStartedEvent, AgentSubAgentProgressEvent, and AgentSubAgentCompletedEvent.
• Dynamic Runtime Skill Activation (SkillDefinition, ActivateSkillsTool, DeactivateSkillsTool):
  • Allows grouping system prompts and tools into modular SkillDefinition units.
  • Non-active skills only publish concise one-line descriptions in the system prompt, keeping token footprints minimal.
  • Injects activate_skills and deactivate_skills tools for the model to toggle toolsets on-the-fly. Emits AgentSkillsChangedEvent.
• Agent Evaluation Harness (EvalTask, EvalGrader, EvalRunner, TrajectoryRecorder):
  • Added end-to-end evaluation framework for agent accuracy, regression testing, and model benchmarking.
  • Included standard evaluators: ExactMatchGrader, RegexGrader, and ToolCallGrader.
  • Added TrajectoryRecorder to serialize full multi-turn conversation trajectories to and from JSON.

1.2.0

• Agent Lifecycle Hooks (AgentHook):
  • Added lifecycle interceptors: beforeToolCall, afterToolCall, beforeModelCall, and afterModelCall.
  • Added ToolCallDecision (proceed, deny, modify) for security guardrails, parameter sanitization, and enterprise compliance.
• Structured Agent Planning (PlanState, PlanStep, WriteTodosTool):
  • Native multi-step agent planning and tracking.
  • AgentPlanEvent stream emitted through McpAgentEngine.agentEvents for reactive UI progress visualization.
  • Optional enablePlanning flag on Agent automatically injecting the write_todos planning tool.
• Enhanced MCPToolResult:
  • Added requestStop flag to gracefully terminate execution loops upon task completion.
  • Added metadata map for rich telemetry, timestamps, and custom tool payloads.
  • Added ergonomic factories: MCPToolResult.text, MCPToolResult.error, MCPToolResult.stop, and MCPToolResult.image.
• Context Window Auto-Compression (ContextCompressor):
  • Pure Dart token estimation and auto-compression utility designed for SLMs and long-running sessions.
  • Automatically summarizes earlier conversation turns via the LLM when approaching context limits while preserving recent turns and system instructions.
  • Emits AgentContextCompressedEvent with token reduction metrics (originalTokens, compressedTokens, compressedTurnCount).
  • Pluggable via McpAgentEngine(contextCompressor: ...) or Agent(contextCompressor: ...).
• Progressive MCP Tool Discovery:
  • Added McpListToolsTool and McpCallToolTool to support workspaces with hundreds of MCP tools without blowing LLM context windows.

1.1.6

• OpenAI-Compatible & Mistral HTTP Response Patching:
  • Introduced _OpenAiCompatiblePatchClient to patch HTTP responses and SSE streams (text/event-stream) for OpenAI-compatible endpoints (such as OpenRouter, vLLM, local LLMs) and Mistral.
  • Sanitizes usage.prompt_tokens_details and usage.completion_tokens_details by stripping null and non-integer values, preventing FormatException / ParseException in openai_dart's requireJsonInt (such as on cache_write_tokens).
  • Automatically injects missing type: "function" into tool calls.
  • Coerces empty list content: [] to null in streaming delta and completion messages.

1.1.5

• Robust Multi-Encoding File Support (fs_read_file, fs_replace_text):
  • Implemented automatic encoding fallback (UTF-8 -> Latin-1 / ISO-8859-1 / Windows-1252 -> UTF-8 with allowMalformed: true) in BaseCodingTool, FsReadFileTool, and FsReplaceTextTool.
  • Fixes crash (FileSystemException: Failed to decode data using encoding 'utf-8') when reading or editing legacy embedded C/C++ source code, header files, or non-UTF-8 workspace documents.

1.1.4

• Web & WASM Platform Abstraction for Coding Tools:
  • Added conditional export for coding tools with coding_tools_stub.dart, allowing dart_mcp_core to compile seamlessly in Web and WASM environments where dart:io is unavailable.
• Multimodal Image & Attachment Support:
  • Added multimodal image attachment handling across LLM providers:
    • OpenAI: ContentPart.imageUrl with base64 data URI format.
    • Anthropic: InputContentBlock.image base64 blocks with media type detection (image/png, image/webp, image/gif, image/jpeg).
    • Google Gemini: Part.bytes for image attachments.
• Model Reasoning & Commentary Preservation:
  • In McpAgentEngine, preserved model reasoning and commentary text in conversation history alongside tool calls to prevent repeating thought loops on reasoning models (e.g., DeepSeek).
• Tool Iteration Limit Increase:
  • Increased default maxToolIterations from 100 to 400 in McpAgentEngine and LlmConfig.
• Dependency Upgrades:
  • Upgraded openai_dart to ^11.0.1.

1.1.3

• Upgraded AI SDK Dependencies:
  • Upgraded googleai_dart to ^13.0.0 (Dart SDK >=3.12.0).
  • Upgraded openai_dart to ^10.0.0 (GPT Image 2.5 support, extensible ImageSize value class).
  • Upgraded anthropic_sdk_dart to ^9.0.0 (Dart SDK >=3.12.0).
  • Upgraded ollama_dart to ^4.0.0 (extended thinking union support).

1.1.2

• Conversation Session Resumption & Trajectory Preservation:
  • Added initialMessages property to Agent definition and JSON serialization (toJson / fromJson), allowing multi-turn REPL and CLI agents to save, restore, and continue session transcripts seamlessly.
  • Enhanced AgentFinalResultEvent to carry the full List<ChatMessage> messages trajectory generated across all tool calls and responses during execution.
• Configurable Tool Iteration Limit (maxToolIterations):
  • Added configurable maxToolIterations (default: 100, up from 10) to LlmConfig, Agent, and McpAgentEngine.
  • Added graceful final response synthesis when the tool iteration limit is reached instead of halting execution abruptly.
  • Added automatic recovery prompting if an LLM finishes tool calling with an empty text response.
• Token Usage Reporting & Metric Events:
  • Added AgentUsageEvent stream event to McpAgentEngine providing promptTokens, completionTokens, and totalTokens emitted after each LLM call.
  • Added LLMUsage model attached to LLMResponse capturing usage returned by LLM providers or calculated via fallback token estimation.
• Enhanced Coding Tools:
  • Included full 10-tool autonomous engineering tool suite in CodingTools.createAll(): fs_find, fs_list_dir, fs_read_file, fs_write_file, fs_replace_text, fs_create_dir, fs_move, fs_delete, terminal_exec, and fetch_web.
  • Improved FsReadFileTool with intelligent pagination and safe 800-line chunk truncation, providing clear range hints for large files.
• Bug Fixes:
  • Cleaned up nullable type assertions in OpenAI provider usage mapping.

1.1.0

• New Native Coding Tools & Engine Helpers:
  • Added FsFindTool (fs_find) for discovering files and directories across workspaces with .gitignore awareness.
  • Added FsReadFileTool (fs_read_file) for line-numbered and paginated file inspection.
  • Added FsWriteFileTool (fs_write_file) for creating and overwriting files with automatic directory resolution.
  • Added FsReplaceTextTool (fs_replace_text) for surgical, unique search-and-replace edits.
  • Added TerminalExecTool (terminal_exec) for cross-platform workspace command execution (e.g. dotnet build, dart test, git status).
  • Added CodingTools.createAll() factory helper to bundle coding tools for CLI, Desktop, and Server environments.

1.0.2

• Updated package documentation and branding references to flutter_mcp_suite.

1.0.1

• Expanded LLM SDK dependency version constraints to support anthropic_sdk_dart 9.x, googleai_dart 13.x, openai_dart 9.x, and ollama_dart 3.x.

1.0.0

• Initial stable production release under the flutter_mcp_suite ecosystem.
• Rebranded and modernized from mcp_playground_dart.
• Production-ready Model Context Protocol (MCP) client transports (HTTP/SSE & stdio subprocesses).
• LLM service adapters for OpenAI, Anthropic, Gemini, and Ollama.
• Agent execution loop and tool orchestration engine.

0.3.2

• Upgraded AI SDK dependencies:
  • anthropic_sdk_dart to ^8.0.0
  • googleai_dart to ^12.0.1
  • openai_dart to ^8.1.0
  • ollama_dart to ^2.6.1
  • yaml to ^3.1.4
  • test to ^1.32.0
• New Pure Dart Examples & Skills:
  • Git Staged Diff Reviewer & PR Generator (examples/dart/git_review_skill): Automated git inspection workflow that reads staged diffs, runs static analysis, and produces structured pull request markdown reviews (PR_REVIEW.md).
  • API Endpoint Health & Schema Prober (examples/dart/api_tester_skill): REST health-checking agent benchmarking endpoint latency, validating JSON schema payloads, and generating diagnostic audit reports (API_DIAGNOSTICS.md).
  • SQLite Database Analyst & Executive Reporter (examples/dart/sqlite_analyst_skill): Autonomous database exploration agent that introspects relational schemas, executes aggregation queries, and exports business intelligence summaries (DATABASE_REPORT.md).

0.3.1

• Fixed repository URL in pubspec.yaml to point to GitHub main branch.

0.3.0

• Windows Stdio MCP Server Execution: Switched Windows NPX server launching to direct node.exe execution with npx-cli.js, eliminating cmd.exe stdin pipe stalls during MCP initialize handshakes.
• Mistral Streaming: Fixed _MistralPatchClient buffering live SSE event streams, enabling real-time chunk streaming.
• Added LocalMcpRuntime.findNpxCli and _resolveWindowsExe helpers for robust tool detection on Windows.
• Unified 0.3.0 release across the MCP Playground ecosystem.

0.2.2

• Added SkillImporter.collectNeededTools() — collects all tool names referenced by a skill manifest (declared tools + per-step enabledToolNames).
• Added SkillImporter.buildUnresolvableToolsWarning() — builds a human-readable warning listing missing tools with capability/install/registry details.
• McpAgentEngine.registerAgentFromManifest() now tracks resolved capability tools and emits an AgentLogEvent warning when unresolvable tools are detected.

0.2.1

• Added McpAgentEngine.registerAgentFromManifest() — builds and registers an Agent directly from a parsed SkillManifest, handling tool-declaration → McpServerConfig mapping and prompt-step conversion. Accepts optional dartTools, workingDir, and serverOverrides.
• Updated SkillExporter.toSkillMd() compatibility tag from mcp_playground to "Universal" to match agentskills.io TealKit standard.
• Added dartTools parameter to registerAgentFromManifest() for Dart-native local tool injection (weather, charts, SSH).

0.2.0

• BREAKING: SKILL.md export/import format changed to agentskills.io TealKit-compatible standard (compatibility:, metadata:, workflow:, agents:). Legacy custom format (system_prompt:, prompts:, tools:) is deprecated.
• Added SkillManifest, SkillPromptStep, SkillToolDeclaration models with 3-tier portability (capability, local, external).
• Added SkillExporter — converts SavedPlaygroundSetup/conversation → SKILL.md YAML frontmatter.
• Added SkillImporter — parses SKILL.md (both TealKit and legacy formats) → SavedPlaygroundSetup with multi-prompt step conversion.
• Added SkillStorageAdapter abstract interface with StoredSkillInfo for pluggable skill ZIP persistence.
• Added yaml dependency for robust YAML parsing.

0.1.4

• Upgraded dependencies: anthropic_sdk_dart to ^6.0.0, googleai_dart to ^9.0.0, ollama_dart to ^2.4.0, http to ^1.6.0.

0.1.3

• Refactored streaming support for Small Language Models (SLMs) and embedded models.
• Added dynamic static delegates embeddedHandler and embeddedStreamHandler to plug on-device inference engines without hard code dependencies.
• Added a new headless CLI example embedded_example demonstrating on-device Hugging Face GGUF model downloading, offline inference, local weather tool execution, and token-by-token terminal stream outputs.

0.1.2

• Expose agentEvents stream in McpAgentEngine and return a reactive event stream from runAsync for asynchronous streaming execution.
• Add support for loading LlmConfig hyperparameters (temperature, maxTokens, topP, topK, repeatPenalty) inside provider clients.
• Isolate local stdio client dependencies from web/wasm builds via default stub conditional imports.
• Protocol and logging compliance updates.

0.1.1

• Fixed WASM and Web compilation by integrating universal_io for subprocess/file wrappers.
• Added pure Dart McpChangeNotifier to core clients to remove Flutter dependencies.
• Added comprehensive self-contained package example for pub.dev.

0.1.0

• Initial release of the pure-Dart core package.
• Built-in multi-LLM SDK adapters (OpenAI, Claude, Gemini, Ollama, Mistral).
• HTTP/SSE stateful Model Context Protocol (MCP) clients with authentication and health-monitoring reconnection.
• Desktop stdio process execution client for Node.js and Python MCP servers.
• Core McpAgentEngine supporting sub-prompt orchestration pipelines, iterative tool loops, and loop-protection.