hermes-dsh-collab
将 DeepSeek Harness 接入您的 Hermes 流程：提供 dispatch-spec 模板、模型层级路由、质量门禁和 Git 单一写入者规则，作为可安装的技能包。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:Cavan-Ou/hermes-dsh-collab#459f8753a5e2dc3e2455182b9c89dda79202e655Hermes Dsh Collab 以 hermes-dsh-collab 发布，当前版本为 0.3.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Hook DeepSeek Harness into your Hermes pipeline: automated dispatch, execution, and verification — with quality gates that don't trust self-reports.
简体中文版见 README.zh.md
Your AI assistant can work all day. The question is whether you can walk away.
This skill makes it safe to: Hermes writes the dispatch spec, DSH executes it, Hermes verifies it — and you only step in when a quality gate actually fails. Distilled from a real 14-day pipeline (30 commits, 7/7 stages shipped without rework).
Dependencies: DSH 0.1.x (headless profile) + any orchestrator agent (built and tested with Hermes).
Hermes is your personal assistant. DSH is a capable executor. The gap is the operating contract between them: what a good dispatch spec looks like, which model tier to use for which stage, who is allowed to commit, and how verification actually happens.
Most pipelines skip that contract and pay for it in rework. Recent work (COPE) shows planner/executor separation works — but only when the executor's behavior is controlled. This skill encodes those controls, from a pipeline where they were tested:
reasoning: max) for routine stages and for looking at images (the main model is natively multimodal — no vision tier, no vision patch), Pro for multi-file refactors and long synthesis (its channel is currently down upstream; the table says so instead of pretending otherwise). When in doubt: try Flash once — rework means escalate.cd <project> && dsh --profile headless "task" writes straight back. No /tmp mirrors, no patch handoffs.read chokes on binaries while read_image needs the model to declare image input, stale backend processes invalidate green tests…# via dsh plugin (bundle — recommended)
dsh plugin --profile headless add github:Cavan-Ou/hermes-dsh-collab
# or copy the skill pack directly (lightweight, any profile)
git clone https://github.com/Cavan-Ou/hermes-dsh-collab
cp -r skills/hermes-dsh-collab "$DSH_HOME/skills/" # default: ~/.dsh/skills/
Either way, the next DSH session loads it automatically (bundle registers a skill provider; the copy is picked up by the skills scanner).
Verify with three checks:
hermes-dsh-collab in its skillsskills/hermes-dsh-collab/
├── SKILL.md # judgment guide: blocking rules → decision tables → What NOT to do
└── references/ # loaded on demand; keeps the guide lean
├── spec-template.md # copyable dispatch-spec template (3 iron rules + no-commit clause)
├── model-routing.md # tier table, patch mechanics, and the 2 patch traps
├── quality-gates.md # 4-step gate commands + rework/escalation chain
└── pitfalls.md # canonical list of 12 real incidents (symptom → cause → fix)
Shaped like DeepSeek's own repo skills (see .agents/skills in deepseek-ai/deepseek-harness): guidance rather than a checklist · points at sources of truth instead of restating them · a dedicated "What NOT to do" section.
Install forms: both supported — bundle (
dsh plugin add, official distribution path, verified on DSH 0.1.0-rc.6) and direct copy to$DSH_HOME/skills/(lightweight, no build).
| Result | Evidence |
|---|---|
| Quality gates catch systemic errors automatically | In one long-synthesis run (74 design docs, ~600K tokens), verification surfaced 3 inverted negative-features and a "universe feature" (hairline borders: 78–100% across all five style families) — mistakes that would take hours of manual review and would likely slip through |
| Routine stages on Flash ship without rework | 3/3 stages — TDD red→green, all gates green, real-browser walkthrough passed |
| The skill changes behavior | Live session: after loading, DSH refuses to commit and restates the single-writer rule with reasons |
Where you (the human) step in: only on the escalation chain — a quality gate fails, or the same stage retries twice. Everything else runs unattended.
SKILL.md — the judgment guide itself (read this first)REPORT.md — build report: design decisions, self-test methodology (isolated DSH_HOME verification), open itemsRoadmap: bundle packaging (dsh plugin add support) · per-workspace failure isolation · live routing-table refresh from the observations card · English mirror of references
Portability note: the skill's "sources of truth" point at workspace-specific paths (e.g.
~/.dsh/profiles/headless/*.patch.yml). The rules are portable — update the paths to match your workspace.
This skill is meant to grow from real use. Hit a pitfall not listed here? Open an issue with the incident (symptom → root cause → fix). PRs welcome in Chinese or English.
MIT
Built by running a real pipeline, not by reading docs.