`# Maestro Orchestrator — Multi-Agent Orchestration Framework
English / 日本語
Security Policy / セキュリティポリシー(日本語)
A research and educational repository for studying how simulated agents coordinate: when to continue, pause or stop, how to hand decisions to human review, and whether outputs agree with logs.
- Maestro: simulations of agent coordination, check ordering, audit records and checkpoints.
- Tasukeru: an advisory workflow that organizes static-analysis and repository-specific findings for human review.
Phase 1F passed its basic simulation tests, but subsequent in-depth validation identified several issues requiring correction.
These issues are being addressed in V04. The final freeze remains on hold until the fixes have been re-tested and reviewed.
Successful demos and individual tests do not establish completion or safety of Phase 1F as a whole. The mock experiments do not verify real-AI judgment, real-user identity or approval, or durable STOP and duplicate-use protection across separate GitHub Actions runs. Each experiment's documentation defines its scope.
Start with the introductory HITL demo v0.1 to see hold, mock approval and resume behavior. Install Git and Python; the repository's Python CI targets 3.10 and 3.11. This demo needs no additional packages or AI API key.
git clone https://github.com/japan1988/multi-agent-mediation.git
cd multi-agent-mediation
python -B demo_hitl_flow_v0_1.pyUse python3 instead of python if required by your environment.
The output shows WAITING_HITL, LOCAL_HOLD and completion after simulated clarification and approval. These responses are fixed in the script; it does not wait for user input. The task labeled as an external send only records state and text. It sends nothing.
Implementation: demo / HITL simulator
- This is a research prototype. A test
PASSor successful CI run is not execution permission, final adoption or a production safety guarantee. - The design under study separates mediation proposals, safety checks and execution authority, with humans retaining the final decision. Do not treat mock approval as real approval.
- Tasukeru Analysis saves review material and comment drafts in its summary and artifacts. It does not automatically post PR comments, repair source files, create branches, PRs or commits, push, or merge.
- GitHub Actions workflows involve network activity, including GitHub API reads, dependency retrieval and auditing, and artifact storage. Their communication scope differs from the local demo above.
- Review only targets you own or are explicitly authorized to review, for defensive and research purposes. Do not use real secrets, personal information or confidential data in examples. See the Security Policy and its Japanese translation; the English policy takes precedence if they differ.
- Hashes and audit records support integrity checks. They do not establish publisher identity, prevent changes or guarantee whole-system safety.
Software code is under Apache License 2.0. Documentation, diagrams and research materials are under CC BY-NC-SA 4.0.
| Topic | Reference |
|---|---|
| Documentation, diagrams and reference files | Documentation index |
| Document simulator architecture, audit and resume | Architecture documentation |
| Actual Tasukeru Analysis configuration | Analysis workflow |
| Mock mediation and independent verification boundaries | Mediator boundary demo |
| Mediator, Gate, Framework and Observer comparison | MGF comparison demo |
| Earlier simulator catalog and extended explanations | Pinned snapshot of the previous README |
| Conceptual workflow for internal verification in the private Maestro build | Workflow diagram (English) |
Historical and version-specific documents are reference records. Distinguish them from the current implementation and validation status of Phase 1F.
Development of the Maestro Orchestrator research framework is continuing incrementally.
We are refining the simulation environment, reviewing safety controls, and preparing a local sandbox environment to support future validation and testing.
Some experimental components remain under internal development and are not yet included in the public repository.
Progress is being made step by step, with safety and verification taking priority over development speed.
The existing safety boundaries and human-review requirements remain unchanged.
Further technical details and validation results will be shared when appropriate.
This update confirms ongoing development. It does not announce a completed implementation, successful final validation, or a release.
Maestro Review Core is a Windows library for .NET Framework 4.8 that verifies approval signatures, checks approval expiry, and compares evidence SHA-256 hashes. A result of REVIEW_SATISFIED means the review conditions were met. Execution authority remains NOT_GRANTED.
The bundled tests passed against the distribution DLL in an offline Windows 11 VM. A rebuild from the included C# sources matched the distribution DLL byte for byte after zeroing the build timestamp and module identifier (MVID) in comparison copies. These results apply to Review Core v0.1.
We are exploring the integration of Tasukeru and Mediator as part of the ongoing development of a transparent and verifiable multi-agent orchestration framework.
The research focuses on traceable decision evidence, consistent safety checks across agent interactions, and a clear separation between review outcomes and execution authority.
Human oversight and final approval remain essential design principles. The framework is intended to support reviewable decisions without automatically granting execution authority.
This work is currently in the specification and validation-case preparation stage. Integration is not yet complete, and the combined framework has not undergone final validation.