Agent Skills
› r0b0tlab/hermes-concurrent-agents
› hermes-concurrent-agents
hermes-concurrent-agents
GitHub管理Hermes并发智能体团队，实现单目标分解为持久化、受监督的协作流程。支持任务分发、状态监控、人工交互及结果验证，确保工作边界清晰且产出可追溯。
触发场景
需要协调多个智能体共同完成一个复杂目标
需要将单一任务拆分为独立并行的子任务
需要对智能体执行过程进行监督和干预
安装
npx skills add r0b0tlab/hermes-concurrent-agents --skill hermes-concurrent-agents -g -y
SKILL.md
Frontmatter
{
"name": "hermes-concurrent-agents",
"version": "2.0.2",
"description": "Supervise bounded Hermes teams on one host."
}
hermes-concurrent-agents
Use HCA when one goal should become a small, persisted, supervised Hermes team with exact ownership and one evidence-backed result.
Preconditions
- Hermes Agent is installed, authenticated, and configured.
- Stable contract lane: Hermes
0.18.2 / 2026.7.7.2. tmuxis available.- Workers and the authoritative Kanban board remain on one host.
- Any model endpoint, local or remote, is already configured through Hermes.
Primary workflow
hca init --preset generic-linux --model <id> --source-profile default
hca doctor
hca run \
--source-profile default \
--acceptance "The requested result is verified" \
"Complete the goal"
hca run-status <run-id>
hca collect <run-id>
One-step work uses one worker. Fan-out requires multiple criteria plus
--independent-criteria; set a bounded --concurrency only when work is truly
independent.
Human interaction
hca respond <run-id> <question-id> "answer"
hca stop <run-id>
Never answer a different run's question or treat blocked/cancelled work as success. Stop is confirmation/approval-gated and signals only exact HCA-owned process groups.
Principles
- Hermes Kanban is task/run truth; HCA owns only its bounded graph and mappings.
- Reserve capacity and a concrete slot before claim.
- Workers cannot create or dispatch unrelated graph work.
- Durable fan-out uses controller-created Kanban children; worker delegation is disabled by default.
- Unknown telemetry is conservative, never unlimited.
- Completion requires evidence, barriers, review state, and worker cleanup.
- HCA does not provision models, copy credentials, normalize providers, or own fallback.
- Remote model endpoints are supported; remote agent placement is unsupported.
Verification
Before release or a substantial upgrade:
scripts/release-check.sh --full
Read docs/running-a-team.md, docs/support-matrix.md, docs/security.md, and
docs/migration.md for the complete contracts.
版本历史
- fa17feb 当前 2026-07-24 12:12
