Repository navigation
Releases
ORNELYON v0.4.2 Public Preview
ORNELYON v0.4.2 Public Preview
This is a Public Preview, not a stable release. v0.4.2 is a maintenance release: it fixes what v0.4.1 showed wrongly and clarifies what is and is not verified. It adds no feature.
Main fixes
- The GUI header no longer shows the old name or a wrong version. The page now takes its version from
package.json. - The session token is embedded in the page more safely, and every token that worked before still works.
- Old compiled files can no longer end up in the distribution ZIP, and building a package no longer deletes other versions' packages.
- The README (English and Japanese) now keeps apart what is implemented and what was checked with a real agent for the nine handoff routes, and states the conditions a handoff must meet.
- The README, SECURITY.md, UNINSTALL.md and the implementation report were brought in line.
Verification status
- 286 offline tests pass (fake agents and a local fake model; no paid call).
- With a real Claude Code (subscription sign-in, read-only, isolated folder): one single task and one Claude Code to Claude Code handoff succeeded. The handoff text hash matched what was sent, and the receiving agent's answer agreed with the file. There is no record that the receiving agent opened the file, so that is not proven.
- Not tested: any handoff involving Codex, handoffs between different agents, handoffs involving a local LLM, file-editing runs, macOS, Linux and Node.js 20.
- Gemini CLI is not implemented.
- Whether ORNELYON may drive Codex CLI or Claude Code CLI automatically under a subscription sign-in is NOT CONFIRMED (see the README).
- Implemented does not mean verified: see the "Handoff routes" section of the README.
Download
ORNELYON-0.4.2.zip — 196,549 bytes
SHA-256: c4b6f3bd1869707414aba7399c173160ca458287b785ecf2719cebc152ef728c
ORNELYON v0.4.1 - Public Preview
ORNELYON v0.4.1 - Public Preview
Overview
ORNELYON is a small, local harness. It runs one task through the agent you choose (OpenAI Codex CLI, Claude Code CLI, or an OpenAI-compatible local model) from a local GUI or from the command line. Every run is recorded in a local history with three separate judgements: Execution, Verification and Acceptance. A successful run is not reported as verified unless the checks you chose pass.
This release is a Public Preview. It is not a stable release.
Main features
- One task, one agent of your choice; manual switching, mutual review, rule-based routing with fallback (L0 to L3).
- Three separate judgements per task (Execution, Verification, Acceptance). Acceptance is recorded only by you.
- Checks you choose: file existence and hash, JSON parse, trusted commands registered in
config.json, a file-scope rule over the whole workspace, and an optional independent review. - Task Passport and manual handoff to another agent, only when you ask. Handoff text is sanitised and labelled as untrusted data.
- Optional read-only MCP server (
ornelyon mcp) with a scope you set at start. It cannot start agents or change anything. - Efficiency Benchmark computed from your own history. It reports only what agents reported, and labels fake or synthetic data.
- Experimental Research command, limited to enabled agents and
maxCalls.
What it does not do
- It does not sign in for you, and does not store API keys or sign-in tokens.
- No self-improvement loop, no session resume, no L4 cooperative execution, no WSL or container isolation.
- No Gemini CLI support (not implemented).
First install
Requirements: Node.js 20 or newer. The version verified so far is Node.js 24 on Windows 11.
- Download
ORNELYON-0.4.1.zipfrom the release page and check its SHA-256 (see below). - Unzip it into a new folder.
- Windows: double-click
start-ornelyon.bat. macOS/Linux: runsh start-ornelyon.sh(not tested on a real machine). - Open the link printed in the terminal, go to Setup, choose a preset and complete setup.
- Read the billing section of the README before you use a Codex or Claude Code sign-in.
Command line: node dist/interface/cli.js doctor.
Verification and independent audit (facts only)
- Offline tests: 281 (
npm test). They use fake agents, a localhost fake model, hand-built history and the official MCP SDK client. No paid or metered agent call was made. - Real agent runs: v0.3.0 and v0.3.1 had one read-only smoke task and read-only handoffs with Claude Code. v0.4.0 and v0.4.1 had no real-agent run.
- Not tested: real Codex tasks, real file edits, real failures, live MCP connections from Claude Code or Codex, macOS, Linux, Node.js 20 on a real machine.
- Internal audit: findings F1 to F8 (workspace boundary, incomplete scope checks, Research path, MCP history completeness, Local LLM error handling, benchmark known/unknown tokens, MCP write side effects, stale verification results) were fixed and covered by regression tests in
src/test/audit-fixes.test.ts. - An independent re-audit was run by Codex in a separate session on HEAD
9e85c6e. Final technical verdict: READY_WITH_LIMITATIONS. (An earlier HOLD verdict was recorded before the F2 and F8 fixes; it is superseded.) - Windows with Node.js 24: 281 tests passed.
- Not verified: other operating systems, and real agent connections (Codex or Claude Code).
- Whether automated use under a subscription sign-in complies with the providers' terms: NOT CONFIRMED.
- This is not a certification by an external audit body.
- The audited commit is
9e85c6e. The release commit differs from it only in documentation files and the packaging list (which now includes CONTRIBUTING.md and SECURITY.md in the ZIP). No runtime code differs.
Known limits
- Platform: Windows with Node.js 24 is the only verified combination.
- Verification sees only what the snapshots can observe. Writes outside the workspace, changes undone before the next snapshot, and processes that keep writing after the final check are not covered.
- Verification commands are only as independent as their scripts. Keep check scripts outside the agent's allowed files.
- Allowed-file checks detect changes after a run; they do not prevent them.
- Cost is shown only when an agent reports it. ORNELYON does not estimate prices.
- Whether automated use of a subscription sign-in is allowed is not confirmed. See the README billing section.
License
MIT. See LICENSE (Copyright (c) 2026 amane3914).
Modification, forks and feedback
Forks, modifications, redistribution, pull requests and feedback are welcome under MIT. You do not have to report your changes. Bug reports and questions: see the README section "Reporting a problem". Security problems: see SECURITY.md.
Downloads and checksums
ORNELYON-0.4.1.zip— the runtime package (built bynpm run package). Use this one to run ORNELYON.- SHA-256:
622296e830729e558d0b3e19cf300ae546e5357fc03e796e639bd25a6276b8bc - GitHub also generates source archives (zip and tar.gz) of the tag automatically. Those contain the source tree, not the runtime package, and they are not the same file as the ZIP above.
ORNELYON v0.4.1 - Public Preview(日本語)
概要
ORNELYON は小さなローカルハーネスです。利用者が選んだ Agent(OpenAI Codex CLI、Claude Code CLI、OpenAI 互換のローカルモデル)で、ローカルの GUI またはコマンドラインからタスクを 1 件実行します。実行ごとに履歴を残し、Execution、Verification、Acceptance の 3 つを別々に判定します。選んだ検査が通らない限り、成功した実行も検証済みとは表示しません。
本リリースは Public Preview(公開プレビュー) です。安定版ではありません。
主な機能
- タスクごとに Agent を 1 つ選ぶ。手動切替、相互レビュー、ルールベースの振り分けとフォールバック(L0〜L3)。
- タスクごとに 3 つを別々に判定。Acceptance は利用者だけが記録します。
- 利用者が選ぶ検査:ファイルの存在とハッシュ、JSON の構文、
config.jsonに登録した信頼済みコマンド、ワークスペース全体を対象にした変更範囲の規則、任意の独立レビュー。 - Task Passport と、利用者が依頼したときだけの別 Agent への手動引き継ぎ。引き継ぎ文は整えられ、未検証のデータとして明示されます。
- 任意の読み取り専用 MCP サーバー(
ornelyon mcp)。公開範囲は起動時に指定し、Agent の起動や変更はできません。 - 自分の履歴から計算する効率ベンチマーク。Agent が報告した値だけを使い、偽データや合成データには印を付けます。
- 実験的な Research コマンド。有効な Agent と
maxCallsに従います。
できないこと
- 代わりにサインインすること、API キーやサインイン用トークンを保存すること。
- 自己改善ループ、セッションの再開、L4 の協調実行、WSL やコンテナによる隔離。
- Gemini CLI への対応(未実装)。
はじめ方
必要なもの:Node.js 20 以上。これまで検証したのは Windows 11 と Node.js 24 です。
ORNELYON-0.4.1.zipをダウンロードし、SHA-256 を確認します(下記)。- 新しいフォルダに展開します。
- Windows は
start-ornelyon.batをダブルクリック。macOS/Linux はsh start-ornelyon.sh(実機未確認)。 - ターミナルに表示されたリンクを開き、Setup でプリセットを選んで Complete setup。
- Codex または Claude Code のサインインを使う前に、README の課金に関する節を読んでください。
コマンドライン:node dist/interface/cli.js doctor
検証と独立監査(事実のみ)
- オフラインテスト:281 件(
npm test)。偽の Agent、localhost の偽モデル、手組みの履歴、公式 MCP SDK クライアントを使用。有料・従量の Agent 呼び出しは行っていません。 - 実 Agent の実行:v0.3.0 と v0.3.1 で、Claude Code による読み取り専用の確認タスクと読み取り専用の引き継ぎを 1 件ずつ。v0.4.0 と v0.4.1 では実 Agent を実行していません。
- 未実施:実 Codex タスク、実ファイル編集、実際の失敗、Claude Code / Codex からの MCP 実接続、macOS、Linux、Node.js 20 の実機確認。
- 内部監査:F1〜F8(ワークスペース境界、不完全な scope 検査、Research の経路、MCP の履歴の完全性、Local LLM のエラー処理、ベンチマークの既知・未知の扱い、MCP の書き込み副作用、古くなった検証結果)を修正し、
src/test/audit-fixes.test.tsの回帰テストで確認しています。 - Codex による、別セッションでの独立再監査を実施(対象 HEAD
9e85c6e)。最終の技術判定:READY_WITH_LIMITATIONS。(F2・F8 修正前の HOLD は途中判定であり、後続の再監査で置き換えられています。) - Windows/Node.js 24 で 281 件のテストが成功。
- 未検証:他の OS、実 Agent(Codex・Claude Code)への接続。
- サブスクリプション認証による自動実行の規約適合性:未確認(NOT CONFIRMED)。
- 外部の監査機関による認証ではありません。
- 監査対象のコミットは
9e85c6eです。リリースのコミットは、文書とパッケージの一覧(ZIP に CONTRIBUTING.md と SECURITY.md を含める変更)だけが異なります。実行コードの差分はありません。
既知の制限
- 動作確認:Windows と Node.js 24 のみ。
- 検証は、スナップショットが観測できる範囲だけを見ます。ワークスペース外への書き込み、次の観測までに元へ戻された変更、最終検査の後も書き続けるプロセスは対象外です。
- 検証コマンドは、そのスクリプトと同じだけしか独立していません。検査用スクリプトは Agent の許可ファイルの外に置いてください。
- 許可ファイルの検査は、実行後に検出するもので、防止はしません。
- 費用は Agent が報告した場合のみ表示します。価格の推定はしません。
- サブスクリプションのサインインによる自動実行が許されるかは 未確認 です。README の課金に関する節を参照してください。
ライセンス
MIT。LICENSE を参照(Copyright (c) 2026 amane3914)。
改変・フォーク・フィードバック
フォーク、改変、再配布、プルリクエスト、フィードバックは MIT のもとで歓迎します。変更を報告する義務はありません。不具合や質問は README の「問題の報告」、セキュリティ上の問題は SECURITY.md を参照してください。
ダウンロードとチェックサム
ORNELYON-0.4.1.zip:実行用パッケージ(npm run packageで作成)。ORNELYON を動かすにはこのファイルを使います。- SHA-256:
622296e830729e558d0b3e19cf300ae546e5357fc03e796e639bd25a6276b8bc - GitHub はタグのソースアーカイブ(zip / tar.gz)を自動生成します。それらはソースツリーであり、実行用パッケージとは別のファイルです。