TheFellow/fkyeah

This system is a pipeline orchestration engine designed to manage and execute multi-agent AI workflows, specifically supporting coding agents via the Agent Communication Protocol (ACP) and external tools through the Model Context Protocol (MCP). It provides a structured environment for defining complex, multi-turn agent interactions with features like parallel execution, context fidelity controls, and checkpoint management. The platform integrates with various LLM providers through a unified client that handles caching, circuit breaking, and cost tracking, while ensuring reliability through extensive conformance testing across different models and transport mechanisms.

The Attractor pipeline now supports the Agent Communication Protocol (ACP) via new handlers (AcpHandlers, McpHandlers) and preset configurations for Codex, Claude Code, and Gemini, allowing nodes to invoke external AI agents. This change introduces a dedicated JSON-RPC library (Codec, Correlator, Types) to manage communication with these agents, including request correlation and transport handling. Additionally, the condition expression language for edge guards has been extended to support the '==' equality operator alongside existing '=' and '!=' operators, and the DotParser lexer now accepts colons and hyphens in identifiers to support broader graph syntax.

src/Attractor · high confidence

The CLI now includes a new \checkpoint\ subcommand (inspect, mark-done, set-outcome, diff, backup) allowing users to manually manage pipeline state and node outcomes via the command line. Pipeline execution has been enhanced with configurable context fidelity modes, which control how prior stage outputs are included in LLM prompts, and automatically calculates token budgets based on reasoning effort. The default model has been updated to Claude Sonnet 5, and the CLI now exposes cost and observability data in verbose output.

src/Attractor.Cli · high confidence

The src/AcpRuntime directory now contains the core implementation for the Agent Communication Protocol (ACP) runtime. This includes a new AcpClient that manages connections via Stdio, WebSocket, HTTP+SSE, or InMemory transports, handling JSON-RPC serialization, protocol version negotiation, and session-based prompt/cancel operations. It also introduces the AcpDelegate interface and a DefaultDelegate implementation, enabling the runtime to execute file read/write operations and manage terminal sessions with configurable permission strategies (deny, auto-approve, or console prompt) and path-safety validation.

src/AcpRuntime · high confidence

Adds a new MCP (Model Context Protocol) client implementation in src/McpClient that enables connecting to external MCP servers via stdio or HTTP/SSE transports. Users can now define server configurations in JSON (specifying transport type, command/URL, arguments, environment variables, and headers), and the client handles connection lifecycle, JSON-RPC communication, and automatic discovery of available tools from those servers.

src/McpClient · high confidence

This release introduces the ACP Runtime and MCP Client libraries, adding support for Model Context Protocol integration and ACP-based coding agents. The pipeline engine now supports multi-turn coding agent nodes (via the \tab\ shape or auto-promoted \box\ nodes) with features like \thread\_id\ for session partitioning, \max\_turns\, and tool execution. Log directories have moved from \attractor-logs/\ to \.ai/attractor-logs/\, and the CLI now includes a \models\ command to list available LLM models. The solution structure has expanded to include new projects for ACP Runtime, JSON-RPC, and MCP Client, along with their respective tests.

(repo-wide) · high confidence

Added six new \.dot\ graph examples in the \examples/\ directory that define structured, multi-agent workflows for automated software development and validation. These include \cedar\_spec\_port.dot\ for porting Cedar semantics to Swift, \consensus\_task.dot\ and \consensus\_task\_parity.dot\ for multi-model plan and definition-of-done consensus, \fix\_sheets.dot\ for debugging SwiftUI sheet rendering, and \igopher\_parity.dot\, \igopher\_showcase.dot\, and \kitchensink\_parity.dot\ for auditing and implementing SwiftUI feature parity in a terminal UI renderer.

examples · high confidence

The UnifiedLlm client now includes built-in support for response caching (with configurable TTL and persistence), a circuit breaker to handle transient provider failures, and a cost ledger to track usage and spend. It also exposes optional embeddings and response-ID-based tool continuation capabilities, and validates that max\_tokens exceeds the thinking budget for reasoning models.

src/UnifiedLlm · high confidence

The CodingAgent now supports configurable token limits via a new MaxTokens field in SessionConfig (defaulting to 16384) and introduces a ToolCallHook that can intercept tool calls to return custom results or errors. Additionally, the tool registry now includes an IsCacheable flag on registered tools, and parallel tool dispatching explicitly switches to the thread pool before execution.

src/CodingAgent · high confidence

The conformance suite now includes tests for additional models (GPT-5.4, GPT-5.5, GPT-5.6 variants, GPT-6 Astra, Claude Opus 4.7, Claude Opus 4.8, and Claude Sonnet 5), doubling the model matrix from 72 to 144 tests. To support these tests, the Docker image now builds and includes mock fixtures (MockLlmServer, MockMcpServer, MockAcpAgent) and example pipelines. Test execution reliability is improved by replacing static sleeps with a port-readiness check for mock servers and adding a warmup run of the Attractor CLI to absorb cold-start costs before timed tests.

conformance · high confidence

The validation engine now enforces a broader set of pipeline quality checks, including warnings for unrecognized attributes (with suggestions), loop session pollution, missing scope gates or build gates, missing timeouts on tool steps, measure-only validation steps, strict review gate formatting, and scratch path consistency. It also detects conflicting session attributes as errors and supports suppression of certain warnings via explicit attributes like scope\_gate and requires\_green\_build. The schema command now documents these new rules and attributes, ensuring users can validate their pipelines against these updated standards.

conformance/02-validation · high confidence

Add ACP conformance tests for stdio, permission, parallel, and in-memory modes; Add MCP conformance tests for stdio, discovery failures, and HTTP+SSE transports; Add conformance tests for LLM fidelity modes and streaming reassembly; Added conformance tests for GPT-5.4, Claude Opus 4.7, and GPT-5.5; Added mock test fixtures for ACP, LLM, and MCP servers; Added regression and acceptance tests for ACP Runtime; Added regression tests for MCP Client transport and SSE parsing; Added unit tests for JSON-RPC codec and correlator logic; Enhanced parallel branch handling and context tracking; Expanded test coverage for Attractor pipeline engine features; Expanded test coverage for CodingAgent session lifecycle and live API integration; Expanded test coverage for UnifiedLlm core behaviors; New conformance tests for context fidelity, attribute interpolation, and custom outcomes; New conformance tests for execution behaviors and updated retry semantics.

The project has migrated to the .NET 10.0 target framework and introduced several new components: a new \AcpRuntime\ library for agent communication, a \McpClient\ for Model Context Protocol support, and a \JsonRpc\ library. The \Attractor\ engine now references these new libraries and includes additional modules for MCP handling and ACP presets. The \UnifiedLlm\ library has been expanded with new files for observability, model cataloging, costing, circuit breaking, and caching. Additionally, the \analyzers\ directory now includes a sentinel project to download \G-Research.FSharp.Analyzers\ version 0.22.0, and the \Attractor\ and \UnifiedLlm\ projects now treat the \FS3569\ warning (non-tail-recursive functions marked with \\[\<TailCall\>\]\) as an error. Test projects have been updated to include new test files and references, and the \UnifiedLlm.Tests\ project now references \Microsoft.Extensions.TimeProvider.Testing\ version 10.0.0.

(dependencies) · high confidence

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

• CAI 59 → 59 (-0.4)
• Rubric changed (rubric-2026.09.15 → rubric-2026.10.1) — scores are not directly comparable.

• Code Health 55 → 55 (+0.0)
• Architecture 96 → 98 (+2.2)
• Maturity 83 → 82 (-0.7)
• Readiness 45 → 44 (-0.9)
• Security 99 → 99 (-0.2)

• Documentation: no installation or build instructions (README.md)
• Documentation: no usage examples (README.md)

• Documentation: no project overview (README.md)
• Flaky test: UnifiedLlm.Tests::UnifiedLlm.IntegrationTests.UnifiedLlm.IntegrationTests.circuit breaker middleware opens and later recovers after cooldown

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

TheFellow/fkyeah was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.