菜单
Full feature development cycle with herdr-managed agent panes. Ensures an ai-workers tab exists with coder, researcher, and critic panes, then orchestrates implement-review loop via pi-link. Use when you want a persistent multi-agent workspace managed by herdr.
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
Delegate deep domain analysis, architecture planning, or complex decisions to your own scope's expert agent (claude-opus). Ensures the expert is running for YOUR scope. Use for architectural decisions, multi-step planning, tradeoff evaluation, domain modeling. Never borrows agents from other scopes.
Delegate a research task to your own scope's researcher agent. Ensures ai-workers panes are running for YOUR scope, then sends the request to <scope>@researcher. Use for codebase exploration, external references, pattern lookup. Never borrows agents from other scopes.
Delegate a code review to your own scope's critic agent. Ensures ai-workers panes are running for YOUR scope, then sends the request to <scope>@critic. Use for code review, diff review, quality checks. Never borrows agents from other scopes.
Analyze colleague feedback for relevance, objectivity, and practical value. Use when evaluating code review comments, architectural feedback, or any critique from a peer before deciding what to act on.
Full feature development cycle with multi-agent collaboration — researcher, coder, and reviewer. Use when implementing a non-trivial feature that benefits from research, iterative coding, and independent review.
Delegate a task to the most suitable connected agent via link_prompt and return the result to the user. Use when you want to route a request to a specialist colleague on the pi-link network.
| name | team-implement |
| description | Full feature development cycle with herdr-managed agent panes. Ensures an ai-workers tab exists with coder, researcher, and critic panes, then orchestrates implement-review loop via pi-link. Use when you want a persistent multi-agent workspace managed by herdr. |
| argument-hint | <what to implement> |
$ARGUMENTS
Call link_list.
Case A — "Not connected": Tell the user:
"Please type
/link-connectthen/link-name <project>@leadin this terminal, then re-run."
Stop here and wait.
Case B — Connected, but name has no @ (e.g. t-a1b2):
Tell the user:
"Please type
/link-name <project>@leadto identify yourself, then re-run."
Stop here and wait.
Case C — Connected with a proper name (contains @, marked (you)):
Extract scope: kaizen@lead → scope = kaizen. Continue.
Rule: scope = the part before
@in YOUR OWN name. Never use workers from a different scope, even if they appear idle inlink_list.study@coderis NOT your worker if your scope istmp4.
echo $HERDR_ENV
If HERDR_ENV=1 → herdr available. Otherwise note it and continue — Phase 2 will handle the fallback.
Your workers are: <scope>@coder, <scope>@researcher, <scope>@critic.
This step is mandatory. Do not skip it even if other scopes' workers appear idle in link_list.
ai-workers-setup "<scope>" "$PWD"
If not on PATH:
python3 ~/.agents/skills/team-implement/scripts/ai-workers-setup "<scope>" "$PWD"
Reading output:
| Line | Meaning |
|---|---|
all agents already idle — nothing to do | workers ready, continue |
missing agents: <scope>@coder, ... | starting them now |
reusing stored panes | correct panes reused, no new ones created |
<name> ready | agent reached idle |
On error (exit non-zero):
| Error | Action |
|---|---|
No workspace found matching cwd | open this project directory in herdr first |
Timeout waiting for <name> | check the pane manually |
nested herdr is disabled | add nested_herdr = true to ~/.config/herdr/config.toml, reload |
No supported multiplexer detected | run pi from inside herdr/zellij/tmux |
Stop and show the error to the user if script exits non-zero.
Evaluate $ARGUMENTS. Identify ambiguities. Ask clarifying questions before writing any code.
Delegate to <scope>@researcher via link_prompt. Ask for:
Then wait for the response — link_prompt is synchronous, you receive the result directly.
Draft a minimal implementation plan with clear acceptance criteria.
Check jj revision matches the task scope. If not, create/describe a new one.
Maximum 3 iterations. You must complete this loop fully — do not stop early to tell the user "check the tab".
Step 1. Send plan (or fix list) to <scope>@coder via link_prompt.
link_prompt is synchronous — wait for [<scope>@coder] response before continuing.
Step 2. After receiving coder's response, additionally confirm idle via herdr:
herdr agent wait "<scope>@coder" --status idle --timeout 300000
Step 3. Read full coder output:
herdr agent read "<scope>@coder" --source recent-unwrapped --lines 80
Step 4. Send to <scope>@critic via link_prompt for review.
Include: task context, acceptance criteria, list of changes, any skipped items with reasons.
Wait for [<scope>@critic] response.
Step 5. Confirm idle:
herdr agent wait "<scope>@critic" --status idle --timeout 120000
Step 6. Read critic output:
herdr agent read "<scope>@critic" --source recent-unwrapped --lines 80
Step 7. Triage findings (colleague-comment rubric):
All findings resolved → exit loop. Accepted findings remain → send fix list to coder, repeat from Step 1.
After 3 iterations with unresolved items: make the final call yourself, document reasoning, move on.
Deliver to the user: