메뉴
Delegate implementation to codex (coder) via the herdr-agentchat plugin and drive a two-pane conversation to completion.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
issue が build に投入できる形かを検分し、verdict (build-ready / needs-plan / needs-fix) と指摘を返す。起票には使わない (/issue)。PR のスクリーニングには使わない (/preview)。
Inspect whether an issue is in shape to hand to build, returning a verdict (build-ready / needs-plan / needs-fix) and the findings. Do NOT use to file an issue (use /issue) or to screen a PR (use /preview).
critic-design による敵対的批判を伴う設計探索。生き残った案を構造化 plan にまとめ、自己点検して呼び出し元に返す。plan の永続先は issue の Plan 節が唯一。計画意図のないコードベース調査には使わない (代わりに /research)。
Design exploration with adversarial critique by critic-design. Assembles the surviving approach into a structured plan, self-checks it, and returns it to the caller. The issue's Plan section is the plan's only persistent home. Do NOT use for codebase investigation without planning intent (use /research instead).
構造化されたタイトルと本文で GitHub Issue を生成する。単独で成立し、前段を要求しない。challenge / research の成果物が会話にあれば本文の根拠に使い、/think の plan 下書きがあれば `## Plan` 節へ移設する。issue 番号を渡すと、起票済みで Plan 節を持たない issue へ plan を転記する。
Generate GitHub Issue with structured title and body. Standalone; requires no upstream stage. When challenge / research artifacts exist in the conversation, they feed the body's evidence; when a /think plan draft exists, it is transferred into the `## Plan` section. Given an issue number, it transfers a plan into a filed issue that has no Plan section.
| name | use-cli-herdr |
| description | Delegate implementation to codex (coder) via the herdr-agentchat plugin and drive a two-pane conversation to completion. |
| when_to_use | codex と連携, coder に委譲, coder に任せる, ペア実装, herdr で 2 agent, leader として実装依頼, agentchat, 2 ペイン会話, send で依頼 |
| allowed-tools | Bash Read |
| user-invocable | false |
Run inside a herdr pane (HERDR_ENV=1) with the thkt.agentchat plugin installed. The send script lives at ~/.config/herdr/plugins/github/thkt.agentchat-*/actions/send.sh (the glob resolves uniquely). Do not use codex exec or launch codex directly. That leaves the conversation and completion-relay machinery, so reports stop arriving.
herdr agent rename "$HERDR_PANE_ID" leaderherdr plugin action invoke thkt.agentchat.start-coder (codex starts in a right split in no-approval mode; check results with herdr plugin log list --plugin thkt.agentchat)herdr agent read coder --source visible shows Press t to trust, run herdr agent send-keys coder t; when it shows Press enter to view hooks, run herdr agent send-keys coder esc. A body sent while a dialog is showing gets swallowed, so confirm a bare input prompt before Phase 2bash ~/.config/herdr/plugins/github/thkt.agentchat-*/actions/send.sh --reply-to leader coder "<instruction>"
Include target files, completion criteria, and write-scope constraints in the instruction. --reply-to leader embeds the reply command for coder into the body automatically.
| exit | Meaning | Response |
|---|---|---|
| 0 | coder started | Wait for the report |
| 3 | Identical resend | Do not resend |
| 5 | coder is blocked | Awaiting human approval. Do not send |
| 6 | Wake failed (body may sit unsubmitted in its input box) | Do not resend; check the screen with herdr agent read coder |
| 7 | Start not observed (delivered if already working) | Check state with herdr agent get coder. Do not resend |
Coder's turn completion arrives automatically as an [auto-relay] message. No polling is needed; continue your own work until it arrives. When coder's own send report and the auto-relay both arrive, prefer the send report. Blocked transitions are toast-notified to the human, so the leader does not stand in for approvals. After receipt, verify (run tests, check the diff) before the next instruction or the report to the human.