Skip to content

omp-flow 0.6.0

Latest
Latest

Choose a tag to compare

View all tags
@Andyduck-ops Andyduck-ops released this 28 Oct 05:56
v0.6.0
a239666
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

What's new

  • Less interruption: agents stop only for a decision that belongs to the human and is not yet
    made, such as value or scope trade-offs, QbD calibration, blocker routing, Spec promotion, and
    high-risk effects nobody asked for. What the request already asks for is done. Repair inside a
    Work runs on its own for at most three rounds, each re-reviewed by a different actor.
  • Two new Skills: omp-flow-adr keeps decision records in .omp-flow/wiki/decisions/ (based on
    MADR 4.0.0), and omp-flow-restate restates the human's purpose, end state, and limits before work
    starts (based on teach-back, commander's intent, and related methods). The Wiki can archive
    Concepts by date and reason under .omp-flow/wiki/archive/.
  • Session maps no longer vanish: Claude Code moves hook context over 10,000 characters into a
    file and shows the model a 2,000-character preview. The three maps now share a 9,000-character
    budget; a long map is cut with a pointer to its file instead of being dropped.
  • Simpler dispatch: actorId is a readable name Main chooses and need not equal a Harness id;
    the shared Skill maps each role to each Harness's agent. Claude agent definitions deny only the
    Agent tool and inherit every other tool, MCP tools included.
  • Upgrade note: if you customized .claude/settings.json, update keeps your copy; remove its
    SubagentStart entry that points to inject-agent-identity.py by hand.

更新内容

  • 少打扰人：只在决定属于人、而人还没做出时才停下来问，例如价值或范围取舍、QbD 校准、blocker
    路由、Spec 晋升，以及没人要求过的高风险操作。请求里已经要求的事直接做；Work 内的修复会自动进行，
    最多三轮，每轮由不同 actor 复核。
  • 新增两个 Skill：omp-flow-adr 在 .omp-flow/wiki/decisions/ 中记录决策（以 MADR 4.0.0 为基础），
    omp-flow-restate 在动手前用自己的话复述人的目的、完成状态和限制（依据 teach-back、指挥官意图等
    沟通方法）。Wiki 支持按日期和原因归档到 .omp-flow/wiki/archive/。
  • 开局地图不再丢失：Claude Code 会把超过 10,000 字符的 Hook 上下文存成文件、只给模型 2,000 字符
    预览。现在三张地图共用 9,000 字符预算，超长时截断并给出文件路径，不再整张丢弃。
  • 派发更简单：actorId 是主会话起的可读名字，不再要求等于 Harness 原生 id；共享 Skill 给出每个
    role 对应的各 Harness agent。Claude 的 agent 定义只禁用 Agent 工具，其余工具（包括 MCP）照常继承。
  • 升级提示：如果你改过 .claude/settings.json，update 会保留你的版本；请手动删除其中指向
    inject-agent-identity.py 的 SubagentStart 配置。

Also: Windows CI is green again; tests now resolve temporary roots the way a Harness does, and sleep start resolves the repository root once.

npm install -g omp-flow@0.6.0
cd <your-project>
omp-flow update --dry-run
omp-flow update
Assets 2
Loading