メニュー
Scaffold a new agent harness for open-managed-agents and register it. Use when adding a new way to drive the model loop, like the default / acp-proxy / flue harnesses.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/duyet/oma --skill new-harnessコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
Explore, clean, and analyze a dataset to answer a real question, with honest handling of uncertainty and bias. Trigger when the user asks to "analyze this data", "what does this dataset show", "explore this CSV", works with pandas/numpy on tabular data, or wants a data-driven answer to a business/research question. Profile before cleaning, sanity-check before concluding, and always report the caveats alongside the finding.
Systematically root-cause a bug instead of guessing. Trigger when the user asks to "debug this", "why is this failing", "it's broken", "reproduce this bug", "find the root cause", or pastes an error/stack trace. Reproduce, isolate, form one hypothesis at a time, fix at the root, then add a regression test — never patch the symptom you happened to notice first.
Restructure code without changing its observable behavior. Trigger when the user asks to "refactor this", "clean this up", "simplify this function", "this is hard to change", "extract this into its own function/module", or "reduce duplication here". Behavior-preserving, test-backed, one small reversible step at a time — never bundled with a feature or bug-fix change.
SOC 職業分類に基づく
SKILL.md を表示中
| name | new-harness |
| description | Scaffold a new agent harness for open-managed-agents and register it. Use when adding a new way to drive the model loop, like the default / acp-proxy / flue harnesses. |
| disable-model-invocation | true |
Scaffold a new harness named $ARGUMENTS (the string an agent selects via harness: "<name>").
Read the current contract first — open apps/agent/src/harness/interface.ts and use the live HarnessInterface, HarnessContext, and HarnessRuntime shapes. Do NOT assume a signature from memory; the interface evolves.
Study a real implementation for the pattern:
apps/agent/src/harness/default-loop.ts (DefaultHarness) — drives its own streamText/generateText loop.apps/agent/src/harness/acp-proxy-loop.ts / flue-loop.ts — meta-harnesses that hand the turn to an external runtime and make shouldCompact / compact / deriveModelContext deliberate no-ops.Create apps/agent/src/harness/<name>-loop.ts exporting a class that implements HarnessInterface. run(ctx) is required; only override the optional hooks (onSessionInit, shouldCompact, compact, deriveModelContext) where this harness must diverge from platform defaults. If you override deriveModelContext, keep its output byte-deterministic — the prompt cache invalidates on any prefix drift (see the interface comments).
Register it in apps/agent/src/index.ts, next to the existing registerHarness(...) calls:
import { <Name>Harness } from "./harness/<name>-loop";
registerHarness("<name>", () => new <Name>Harness());
Add a unit test alongside the existing harness tests, then run /preflight.