菜单
Delegate scoped coding tasks to AI workers through Foreman. Use when a coding task can be split into independent implementation slices that should run in isolated git worktrees while Codex remains architect and QA.
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。
基于 SOC 职业分类
Use when the user asks about Eidos, Eidos AGI, platform status, `eidos do`, closeout, cleanup before done, docket-task orchestration, evidence verification, praxis learning, gateway routing, vault/secrets, plugins, or which Eidos AGI specialist system should handle a task. This skill tells Codex to call the installed `eidos` CLI first; MCP and plugins should only point to CLIs, while CLIs provide progressive reveal of the deeper tool graph.
Investigate current context before predicting likely futures, failures, user objections, and concrete changes. Use when Zoltar needs to research an app, plugin, architecture, workflow, plan, decision, or implementation and answer what is likely to go wrong, what evidence supports it, what complaint or failure is being prevented, and what should change today.
Answer a specific foresight decision question with researched evidence, a clear judgment, likely futures, user complaint prediction, and concrete changes. Use when Daniel asks whether an idea, implementation, plugin, architecture, or workflow is likely to be regretted, whether it is the right thing in the right place with the right authority, or what decision should be made given future effects.
Compatibility entrypoint for Zoltar foresight. Use when Daniel mentions Zoltar, asks to predict the future, asks for second-order effects, asks what he will complain about, or wants future-backed product/spec changes; route the work through researched evidence, judgment, structured output, doer/checker handoff, and self-improvement notes rather than generic prediction.
Maintain and check the Eidos marketplace when plugins must work in both Claude Code and Codex. Use when adding, refactoring, auditing, or debugging marketplace plugins, dual-host packaging, .claude-plugin and .codex-plugin manifests, root marketplace entries, source/store/cache drift, install proof, or host-neutral plugin payloads.
Challenge whether a Zoltar prediction, product idea, plugin, workflow, category framing, or architecture is overfit to existing market patterns. Use when Daniel asks for a Challenger Matrix, anti-consensus review, category redefinition, frontier user needs, non-consensus futures, Doubter voice, or whether an answer is too derivative, market-shaped, competitor-shaped, or copied from current agent/product patterns.
| name | delegate-with-foreman |
| description | Delegate scoped coding tasks to AI workers through Foreman. Use when a coding task can be split into independent implementation slices that should run in isolated git worktrees while Codex remains architect and QA. |
Use Foreman when the user explicitly wants Codex to parallelize coding work through AI coding workers, or when a repo task has independent file-scoped implementation slices.
claude-emux when the user may need to watch or interrupt Claude Code through Emux, or when Foreman should use the normal interactive Claude Code surface; claude only when the headless metered claude -p path is intentional; codex for hard reasoning/QA; gemini for broad alternate passes; aider for narrow patch tasks; opencode for alternate agent/model runs; gemma4 for throttled local Gemma-family runs; and smoke only for plumbing tests.foreman_delegate once per independent task.foreman_delegate as fire-and-return: record the returned worker_id, run_id, worktree path, and log path.
For claude-emux, also record the returned emux.head_command, emux.capture_command, and emux.interrupt_command.foreman_status for a cheap lifecycle check and foreman_tail for bounded recent logs while workers run.foreman_collect to review changed files and diff.foreman_finalize.Foreman's MCP server exposes more tools than an initial broad tool search may load. If only foreman_delegate, foreman_list, or foreman_control_jobs are visible, do not conclude the plugin lacks lifecycle tools. Search again for the exact names:
foreman_status foreman_tail foreman_collect foreman_finalize
Then use the newly loaded tools for status, logs, diffs, and finalization. This protects the plugin from partial tool-index exposure while keeping the same fire-and-return workflow.
For Slack or other remote-adjacent callers, prefer the control-plane path when available: submit a structured job with foreman control-submit, let a local or rented-machine agent lease it with foreman agent-run --once, then report back from control-status. This keeps Foreman reusable while preserving the same isolated worktree execution model.
Goal:
Acceptance criteria:
Files in scope:
Files out of scope:
Verification command:
Notes:
Keep worker specs narrow. Do not delegate broad architecture, secrets, deploys, money movement, destructive git actions, or human-only approvals.