agents_core 0.3.3

A Dart library for orchestrating multi-agent AI workflows with LM Studio integration. Create agents, manage conversations, execute Python in Docker, and coordinate multi-step pipelines.

Changelog

0.3.3 — 2026-03-27

Features

Smart Loop Detection

• New LoopDetectionConfig — immutable configuration class controlling loop detection thresholds: maxConsecutiveIdenticalToolCalls (default 3), maxConsecutiveIdenticalOutputs (default 3), and similarityThreshold (default 0.85). Supports const construction, value equality, and toString().
• New LoopDetector — stateful detector that tracks consecutive identical tool-call sequences (via sorted fingerprints) and near-identical text outputs (via bigram Sørensen–Dice similarity). Call recordToolCalls() / recordOutput() per iteration, then check() to get a LoopCheckResult. Includes reset() for reuse across tasks.
• New LoopCheckResult — result type with isLooping flag and human-readable reason. Provides a static LoopCheckResult.ok constant for the non-looping case.
• LoopDetector.bigramSimilarity() — static utility that computes the Sørensen–Dice coefficient over character bigrams for fuzzy string comparison.

ReActAgent Integration

• ReActAgent now accepts an optional loopDetectionConfig parameter. When provided, a LoopDetector is created per run() invocation and checks for repetitive tool-call sequences or near-identical outputs after each iteration. Detected loops stop the agent with stoppedReason: "loop_detected".

AgentLoop Integration

• AgentLoop now accepts an optional loopDetectionConfig parameter. When provided, producer outputs are tracked each iteration and the loop stops early with stoppedReason: "loop_detected" if repetitive patterns are found.
• AgentLoopResult gains a stoppedReason field ("accepted", "max_iterations", or "loop_detected") and two new convenience getters: loopDetected and an updated reachedMaxIterations that excludes loop-detection stops.

0.3.2 — 2026-03-27

Examples

• New bugfix_pipeline.dart — realistic 5-step end-to-end bugfix workflow demonstrating ReActAgent with file-context tools, AgentLoopStep.dynamic produce-review loops, AgentStep.dynamic with conditional execution, custom buildProducerPrompt for reviewer-feedback injection, and StepResult pattern matching for post-run inspection. Seven specialised agents collaborate across triage → root-cause analysis → fix → regression tests → PR summary, all orchestrated by a single Orchestrator.run() call.

0.3.1 — 2026-03-26

Bug Fixes

LM Studio Client

• Fixed HTTP request body encoding in LmStudioHttpClient — replaced request.write() with request.add(utf8.encode(...)) in both postStream and the internal _sendRequest helper. The previous implementation could corrupt non-ASCII characters (e.g. Unicode prompts) because write() uses the platform default encoding, which is not guaranteed to be UTF-8.

0.3.0 — 2026-03-26

Previously, orchestrator pipelines only supported single-agent steps, forcing produce-review loops to be managed outside the pipeline. This release introduces a step hierarchy that lets you mix single-agent tasks and iterative review cycles in the same orchestrator, enabling end-to-end workflows like "research → develop → review" without glue code. Existing AgentStep usage is fully backward-compatible.

Migration Notes

OrchestratorResult.stepResults type change

stepResults changed from List<AgentResult> to List<StepResult>. Code that accessed AgentResult properties directly needs to use the common StepResult accessors or unwrap via pattern matching:

// Before (0.2.x)
for (final agentResult in result.stepResults) {
  print(agentResult.output);
  print(agentResult.stoppedReason);
}

// After (0.3.0) — common accessors work on all subtypes
for (final stepResult in result.stepResults) {
  print(stepResult.output);     // available on all StepResult subtypes
  print(stepResult.tokensUsed); // available on all StepResult subtypes

  // Type-specific access via pattern matching
  if (stepResult is AgentStepResult) {
    print(stepResult.agentResult.stoppedReason);
  } else if (stepResult is AgentLoopStepResult) {
    print(stepResult.accepted);
    print(stepResult.iterationCount);
  }
}

Orchestrator.steps type widened

steps changed from List<AgentStep> to List<OrchestratorStep>. Existing code that passes a List<AgentStep> continues to work without changes since AgentStep now extends OrchestratorStep.

Features

Orchestrator Step Hierarchy

• OrchestratorStep — abstract base class for all pipeline steps with taskPrompt and optional condition guard. Enables polymorphic step pipelines.
• AgentLoopStep — new step type that embeds a produce-review loop directly inside an Orchestrator pipeline. Supports static and dynamic (AgentLoopStep.dynamic) task prompts, custom prompt builders, and configurable maxIterations.
• StepResult — abstract base for step results with uniform output and tokensUsed accessors.
• AgentStepResult — wraps AgentResult from a single-agent step.
• AgentLoopStepResult — wraps AgentLoopResult from a produce-review loop step, with accepted and iterationCount convenience accessors.

Orchestrator Refactoring

• Orchestrator.steps now accepts List<OrchestratorStep> (was List<AgentStep>), enabling mixed pipelines of AgentStep and AgentLoopStep.
• OrchestratorResult.stepResults changed from List<AgentResult> to List<StepResult> — use is AgentStepResult or is AgentLoopStepResult for type-specific access.
• AgentStep now extends OrchestratorStep (backward-compatible — existing AgentStep usage continues to work unchanged).

Examples

• New orchestrator_with_agent_loop.dart — 3-step pipeline mixing AgentStep with AgentLoopStep.
• New feature_development_pipeline.dart — realistic 5-stage software development pipeline with PersistingAgent decorator, dynamic prompts, conditional steps, and AgentLoopStep with custom prompt builders.

0.2.1 — 2026-03-26

• Expand README.md with detailed AgentLoop usage and examples

0.2.0 — 2026-03-26

Features

Agent Loop

• AgentLoop — producer/reviewer orchestration loop that iterates a producer agent and a reviewer agent until an approval pattern is matched or maxIterations is reached.
• AgentLoopIteration — immutable record of a single loop iteration capturing index, producerResult, and reviewerResult.
• AgentLoopResult — aggregated result with iterations, approved flag, total duration, and combined errors.
• New example/agent_loop.dart demonstrating a developer + QA review loop.

Bug Fixes

File Context

• Fixed FileContext._resolve to use Uri.file instead of Uri.parse so that workspace paths containing spaces are handled correctly without percent-encoding artifacts.

0.1.2 — 2026-03-25

Improvements

LM Studio Client

• LmStudioHttpClient now correctly throws LmStudioHttpException for 4xx client errors instead of misclassifying them — improves error handling for authentication failures, not-found, and rate-limit responses.
• Retry logic refined to only retry on transient (5xx/network) errors, not client errors.

Bug Fixes

• Fixed flaky retry-related test expectations caused by timing sensitivity in exponential backoff verification.

0.1.1 — 2026-03-25

Add API_KEY to AgentsCoreConfig

Configuration

• Optional apiKey parameter on AgentsCoreConfig — sent as a Bearer token in the Authorization header when the LM Studio server requires authentication. Readable from AGENTS_API_KEY via fromEnvironment(). Masked in toString() output to prevent accidental credential leakage.
• AgentsCoreConfig.copyWith() supports clearApiKey to explicitly remove an API key.

0.1.0 — 2026-03-25

First feature-complete release of agents_core.

Features

Agent Framework

• Agent abstract base class with run(String task, {FileContext? context}) method.
• SimpleAgent — single-round chat completion agent.
• ReActAgent — multi-turn Reason + Act loop with tool calling, configurable maxIterations and maxTotalTokens budget.
• AgentResult — structured output with output, tokensUsed, toolCallsMade, filesModified, and stoppedReason.

LM Studio Client

• LmStudioClient — high-level typed API for LM Studio's OpenAI-compatible endpoints (chatCompletion, chatCompletionStream, chatCompletionStreamText, completion, completionStream, listModels).
• LmStudioHttpClient — HTTP transport with automatic retry and exponential backoff (maxRetries, configurable delay).
• SseParser — Server-Sent Events stream transformer that handles multi-line data and [DONE] sentinels.

Data Models (OpenAI-compatible)

• ChatMessage and ChatMessageRole enum (system, user, assistant, tool).
• ChatCompletionRequest / ChatCompletionResponse / ChatCompletionChoice.
• ChatCompletionChunk / ChatCompletionChunkChoice / ChatCompletionDelta for streaming responses.
• CompletionRequest / CompletionResponse / CompletionChoice.
• CompletionUsage — token usage tracking.
• ToolDefinition and ToolCall / ToolCallFunction for function calling.
• LmModel — model listing response.

Configuration

• AgentsCoreConfig — central configuration with lmStudioBaseUrl, defaultModel, requestTimeout, dockerImage, workspacePath, and logger.
• AgentsCoreConfig.fromEnvironment() factory — reads LM_STUDIO_BASE_URL, AGENTS_DEFAULT_MODEL, AGENTS_DOCKER_IMAGE, AGENTS_WORKSPACE_PATH, and AGENTS_REQUEST_TIMEOUT_SECONDS from environment variables.
• AgentsCoreConfig.copyWith() for immutable modifications.
• Logger abstraction with StderrLogger and SilentLogger implementations.

File Context

• FileContext — sandboxed file-system abstraction with read, write, append, delete, exists, and listFiles (with glob filtering).
• Path traversal protection on all file operations.
• Pre-built tool definitions: readFileTool, writeFileTool, listFilesTool, appendFileTool, and createHandlers() factory.

Orchestrator

• Orchestrator — sequential agent pipeline with shared FileContext.
• AgentStep — static or dynamic (AgentStep.dynamic) task prompts with optional condition guards.
• OrchestratorResult — collects stepResults, duration, and errors.
• OrchestratorErrorPolicy — stop (default) or continueOnError.

Docker Integration

• DockerClient — run containers, check availability, pull images.
• DockerRunResult — captures stdout, stderr, and exitCode.

Python Execution

• PythonToolAgent — pre-configured ReActAgent with Docker-based Python execution and optional file tools.
• PythonExecutionTool — tool definition and handler factory for running Python code in sandboxed Docker containers.

Quick Functions

• ask() — one-shot chat completion that manages client lifecycle.
• askStream() — streaming one-shot chat completion.
• Conversation — stateful multi-turn wrapper with send(), sendStream(), setSystemPrompt(), and clearHistory().

Exception Hierarchy

• AgentsCoreException — library base exception.
• LmStudioHttpException — non-2xx HTTP responses.
• LmStudioApiException — structured API errors with isModelNotFound, isContextLengthExceeded, and isRateLimited helpers.
• LmStudioConnectionException — transport failures with socketError, httpError, timeout, and fromException factories.
• DockerNotAvailableException / DockerExecutionException.
• FileNotFoundException / PathTraversalException.
• SseParseException — malformed SSE data.

0.0.1

• Initial project scaffold.