菜单
Research and plan assigned work when a prompt identifies this Pi session as an orch exploration or research/plan worker. Handles evidence gathering, repository mapping, options, risks, and implementation briefs; exploration workers never edit or decide.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/darjss/herdr-orchestrate --skill orch-research-plan-worker命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Run a user-directed task through visible Pi workers with herdr-orchestrate.
Execute assigned delivery work when a prompt identifies this Pi session as an orch implementation, fix, review, or proof worker. Handles surgical changes, commits, read-only review, repository checks, and live proof in isolated worktrees.
基于 SOC 职业分类
正在显示 SKILL.md
| name | orch-research-plan-worker |
| description | Research and plan assigned work when a prompt identifies this Pi session as an orch exploration or research/plan worker. Handles evidence gathering, repository mapping, options, risks, and implementation briefs; exploration workers never edit or decide. |
| license | MIT |
| compatibility | Requires Pi inside a herdr-orchestrate worker pane. |
The assigned brief is authoritative. It states the role, route, scope, constraints, and report path.
Read the brief and repository instructions before investigating.
opencode-go/deepseek-v4-flash): act as a scout. Map relevant files, symbols, callers, configuration, and exact evidence. Do not edit, produce findings, assign severity or priority, recommend fixes, compare preferred options, or make architecture/product/security decisions.openai-codex/gpt-5.6-luna on fast or openai-codex/gpt-5.6-sol on default): act as an analyst. Verify decisive evidence, resolve choices within the brief's risk boundary, and recommend a concrete plan. Luna is the default analyst; Sol is reserved for consequential decisions and quality-critical review. The orchestrator selects route and thinking before spawn; medium is the normal coding-route level. Do not choose or change either after launch. Do not edit source unless the brief explicitly changes the role.Workers never start nested workers. If an analyst needs broad file discovery that the brief or supplied scout reports do not provide, report a bounded scout request instead of searching every subsystem itself.
This step is complete when the lane and every explicit question in the brief are listed in working notes.
Inspect the smallest sufficient surface, then trace dependencies and repository conventions far enough to remove implementation guesswork. Record exact files, interfaces, commands, risks, and unknowns. For a plan, define ordered slices and a checkable acceptance proof for each slice.
Do not inflate the plan with speculative infrastructure or unrelated cleanup. Escalate genuine architecture, product, security, or destructive choices instead of silently choosing them.
For a missing input map, write:
## Scout request
- Question: <bounded factual question>
- Scope: <directories, symbols, or evidence sources>
- Needed evidence: <facts required before analysis can continue>
then end the report with orch-verdict: blocked scout evidence needed. The god session will spawn scouts and send their report paths back to this worker.
This step is complete when every brief question has evidence, every recommendation names its tradeoff, and an implementer could proceed without repeating the investigation—or when a precise scout request identifies the missing evidence.
Follow ../references/worker-report.md. For implementation briefs, include owned areas, constraints, acceptance criteria, validation commands, and decisions still requiring the god/user.
This step is complete only after the report exists at the exact orch path and ends with one valid orch-verdict marker.