一键导入
orchestrator
Manage parallel HoP GitHub issue agents: show work status, suggest next issues, generate prompts, and clean idle sessions.
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。
菜单
Manage parallel HoP GitHub issue agents: show work status, suggest next issues, generate prompts, and clean idle sessions.
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
Pipeline simplificado: documento fonte → extração de conhecimento → padrões reutilizáveis → classificação contra repositório → canonical docs. Focado em gerar docs canônicos a partir de um único documento (currículo, análise, transcript). NÃO gera skills, exercícios ou integração profunda de currículo — apenas canonical docs + atualização de tags/aliases + system-of-record. Dispara com: 'canonicalize this document', 'extract canonical from', 'doc to canonical', 'create canonical docs from', 'documento para canonical', 'extrair canônicos de'.
Implementa uma camada de selecao de contexto model-agnostic e vendor-independent. Define um formato padronizado de contexto que qualquer modelo pode consumir, um Context Router que resolve queries de qualquer agente contra o grafo relacional e storage em tiers, e Vendor Adapters que traduzem o formato agnostico para o formato nativo de cada modelo. Transforma contexto no ativo mais duravel da organizacao — portavel entre modelos, sessoes, e provedores. Dispara com: 'neutral selection', 'model-agnostic context', 'vendor-independent context', 'context format standard', 'context router', 'multi-model context', 'portable context', 'cross-model context', 'camada neutra', 'contexto agnostico', 'formato de contexto', 'vendor adapter', 'neutral context layer', 'context portability', 'model migration context'.
Torna o retrieval de contexto budget-aware: cada candidato a retrieval e ranqueado por razao valor/custo antes de ser injetado no contexto. Implementa um Information Value Predictor que estima reducao de incerteza por candidato, um Token Cost Estimator que computa custo em tokens, e um Utility Feedback Loop que aprende com uso real — quais itens recuperados foram efetivamente referenciados pelo modelo. Previne o memory feedback loop onde o sistema construido para resolver o problema de memoria se torna o motor da degradacao. Contrapoe diretamente o Link 4 (inert memory feedback) do Agent Degradation Loop. Dispara com: 'selection-budgeted retrieval', 'budgeted retrieval', 'retrieval budget', 'cost-benefit retrieval', 'information value predictor', 'utility feedback retrieval', 'retrieval ranking', 'orcamento de retrieval', 'busca budgetada', 'recuperacao com orcamento', 'value-cost retrieval', 'retrieval utility feedback'.
Implementa armazenamento de contexto em tres tiers (hot/warm/cold) com promocao e democao baseadas em relevancia. Mantem o conjunto de trabalho ativo deliberadamente pequeno na hot tier (cache in-memory), move contexto recentemente relevante para warm tier (NVMe, baixa latencia), e arquiva historico completo na cold tier (object storage, alta latencia). O Tier Orchestrator executa transicoes baseadas em scores de relevancia e prefetch preditivo antes de cada passo de raciocinio. Previne context rot (cada token acumulado degrada qualidade do passo seguinte) e separa preocupacao de armazenamento de preocupacao de selecao. Dispara com: 'tiered context', 'tier storage', 'context tiers', 'hot warm cold', 'tiered storage', 'context promotion', 'context demotion', 'tier orchestrator', 'promocao de contexto', 'democao de contexto', 'armazenamento em tiers', 'tiered context storage', 'context tier management'.
Operacionaliza um curriculo de autonomia para agentes: controla a proporcao entre rollouts supervisionados (professor/humano) e rollouts auto-gerados (agente) com um parametro lambda, gates de prontidao por classe de tarefa, e uma progressao explicita observe→assist→own. Aplica importance sampling para corrigir a distribuicao mista. Previne cold-start collapse (rollouts puros do agente antes de aprender recuperacao) e autonomy stagnation (agente nunca pratica recuperacao autonoma). Usar ao implantar um novo agente em producao, ao fazer transition de fluxo manual para agentico, ao calibrar o grau de supervisao de um agente existente, ou quando o agente apresenta comportamento fragil em cenarios nao-supervisionados. Dispara com: 'autonomy curriculum', 'curriculo de autonomia', 'lambda schedule', 'teacher mixing', 'rollout sampling', 'autonomy dial', 'autonomy progression', 'observe assist own', 'readiness gate', 'student rollout', 'teacher-student mix', 'gradual autonomy', 'agente semi-supervisionado', 'autonom
Separa o sinal de melhoria do agente em dois componentes independentes: magnitude (o quanto o modelo acredita que uma mudanca importa, extraida de self-distillation delta, log-ratio, atencao, ou confianca interna) e direcao (para onde a mudanca deve ir, determinada por verificador externo, testes, ou revisao humana). Combina magnitude × direcao em um plano de correcao ponderado: gaste esforco de correcao onde o agente tem conviccao FORTE e o verificador confirma a direcao; reduza ou escale quando a direcao e incerta. Previne information leakage (agente aprende a imitar formato sem substancia) e overconfidence collapse (self-distillation puro sem verificacao externa). Usar ao projetar loops de melhoria de agente, ao implementar self-distillation com verificacao, ao calibrar feedback de producao, ou quando o agente produz outputs com formato correto mas conteudo errado. Dispara com: 'magnitude direction', 'verifier split', 'trust but verify', 'confidence direction', 'correction weight', 'self-distillation verif
| name | orchestrator |
| description | Manage parallel HoP GitHub issue agents: show work status, suggest next issues, generate prompts, and clean idle sessions. |
| license | MIT |
| compatibility | opencode |
| metadata | {"audience":"all-agents","workflow":"github","priority":"high"} |
I coordinate parallel HoP agent sessions across GitHub issues. I do not implement product/code changes. I fetch state, summarize active work, suggest the next issue, generate prompts for worker sessions, and clean up idle or completed sessions.
Agents are tracked by issue number. For example, "Agent #123" means the session working on GitHub issue #123.
Load this skill when:
Gather these in parallel when possible:
gh issue list --state open --json number,title,labels,assignees --limit 100
gh pr list --state open --json number,title,headRefName,isDraft,url --limit 50
git worktree list --porcelain
gh label list --limit 200
Present a compact table:
## Dashboard
| # | Issue Title | Labels | Assignee | PR | Worktree | Status |
|---|-------------|--------|----------|----|----------|--------|
| 123 | Fix tenant artifact paths | priority:P1 | bot | #130 draft | .worktrees/123-... | active |
| 124 | Dashboard health panel regression | - | - | - | - | ready |
Mark issues with agent:working as active. Put blocked issues at the bottom. If the repo has no blocked label, treat only explicit issue labels/status as blockers.
Do not assume historical label systems. Discover labels first with gh label list.
Priority order:
agent:working.blocked if that label exists.priority:P0 -> priority:P4.blocked by #N, check that blocker is closed before suggesting it.gh issue view <BLOCKER_N> --json state --jq '.state'
Output:
Suggested next: #123 "Fix tenant artifact paths" (priority:P1, unassigned, no open blockers)
When the user says "assign #N", generate a prompt for a new worker session.
Work on HoP issue #N: "<title>".
Load skill `issue-start` and claim issue #N.
After setup, implement the issue in the created worktree. Follow:
- `AGENTS.md`
- `docs/system-of-record.md`
- `agents/manifest.yaml`
- `README.md`
- Relevant `docs/canonical/`, `docs/decisions/`, `docs/guides/`, and `docs/evidence/`
- `DESIGN.md` for dashboard/UI work
Use the minimum viable change. Respect tenant isolation for `.runtime/` and `artifacts/`. If crossroad files are touched, follow `.github/PULL_REQUEST_TEMPLATE.md`, `.github/CODEOWNERS`, and `docs/guides/crossroad-change-policy.md`.
When implementation is complete, load skill `issue-review`. Fix blocking findings and rerun review.
When review passes and the user explicitly confirms merge, load skill `issue-finish`.
Do not add extra implementation instructions unless the issue requires them.
When the user reports "Agent #N idle", inspect before cleanup.
N=<issue_number>
gh issue view "$N" --json state,title,labels,assignees
gh pr list --state all --search "#$N" --json number,state,mergedAt,headRefName,url --limit 20
git worktree list --porcelain
Clean up only that issue's worktree/branch.
N=<issue_number>
WORKTREE=$(git worktree list --porcelain | awk -v n="/$N-" '/^worktree / {path=$2} path ~ n {print path; exit}')
if [ -n "$WORKTREE" ] && [ -d "$WORKTREE" ]; then
if [ -n "$(git -C "$WORKTREE" status --porcelain)" ]; then
echo "Worktree has uncommitted changes; cleanup skipped: $WORKTREE"
else
BRANCH=$(git -C "$WORKTREE" rev-parse --abbrev-ref HEAD 2>/dev/null || true)
git worktree remove "$WORKTREE"
if [ -n "$BRANCH" ] && git show-ref --verify --quiet "refs/heads/$BRANCH"; then
git branch -D "$BRANCH"
fi
fi
fi
gh issue edit "$N" --remove-label "agent:working" 2>/dev/null || true
git remote prune origin
Release the claim so another session can continue.
gh issue edit "$N" --remove-label "agent:working" 2>/dev/null || true
gh issue edit "$N" --remove-assignee "@me" 2>/dev/null || true
gh issue comment "$N" --body "Agent session ended without completion - $(date -u +%Y-%m-%dT%H:%M:%SZ). Issue returned to queue."
Do not delete a worktree containing uncommitted work unless the user explicitly approves. Report its path for manual inspection.
Run Phase 2 and present the next suggestion.
## Agent #N - Cleaned Up
Issue #N: <closed | returned to queue | needs manual inspection>
Worktree: <removed | retained at path>
Branch: <deleted | retained>
Label: agent:working released
Suggested next: #M "Next issue title" (<labels/status>)
If a worker cannot resolve blocking findings after three materially different attempts:
blocked if the label exists, otherwise add a comment only and surface that the label is missing.agent:working unless the user wants the issue retained by that agent.if gh label list --limit 200 --json name --jq '.[].name' | grep -qx "blocked"; then
gh issue edit N --add-label "blocked"
fi
gh issue comment N --body "Blocked: <brief reason>. Needs manual intervention."
issue-finish handles merge/close after confirmation.agent:working without user confirmation.agent:working labels when an agent stops and the issue is returned to the queue.gh issue list --state open --json number,title,labels,assignees --limit 100
gh pr list --state open --json number,title,headRefName,isDraft,url --limit 50
git worktree list --porcelain
gh label list --limit 200
gh issue edit N --remove-label "agent:working"