DSH HUB
HomePlugin StorePlugin PacksCommunityRankingsResourcesPublish Guide
Plugin source
Back to catalog

xu-jin-cs /

xu-jin-cs/dsh-skills

Topic repository only

Mechanical gates + skills that make LLM coding agents keep their promises — agent capability evaluation (9-dim radar/rank reports), parallel dispatch, architecture mapping. Plain SKILL.md, zero lock-in. Custom Source-Available license.

★ 1 Stars1 Forks1 IssuesN/A Community rating0 Confirmed installs
View on GitHubProject homepage
READMESource: main@943053cb

dsh-skills

License: Custom Source-Available Last commit Python 3.10+ Format: SKILL.md

English | 中文说明

Important: This repository is NOT open source under the MIT License. This is a Source-Available license. Source code is available for personal non-commercial study and local use only. Publicly forking and releasing modified versions, commercial distribution, and SaaS services are forbidden without written permission.

本仓库并非 MIT 开源协议,为源码可用许可。源码仅供个人非商用学习、本地运行。未经书面授权,禁止公开 Fork 并发布修改版本、商用分发、搭建 SaaS 服务。

Mechanical gates and reusable skills that make LLM coding agents keep their promises — capability evaluation, parallel dispatch, architecture cartography, and a full PM workflow. Plain SKILL.md files, zero lock-in: compatible with DeepSeek Harness, Claude Code, Kimi Code, Codex, and any agent that reads the standard skill format.

Why this repo

Most agent skills hope the model follows instructions. This repo moves judgment from the model to scripts: spec-driven gates mechanically verify what the agent claims (A = pass / B = block with violations), parallel fan-out goes through an audited switch, and architecture mapping is deterministic computation instead of full-repo re-reading.

The lead hook is agent-eval — agent/skills capability evaluation: read-only collection of every skill on your machine → 9-dimension scoring against a frozen baseline → strength/weakness grading (blocking / high-risk / optimization) → overall ranking → radar chart + leaderboard + HTML report. Evaluation tooling for agents is the gap; the RAG/ETL engine (Xj-engine) plays support.

60-second quickstart

# 1. Install all five skills (symlinks into your skill discovery root)
curl -fsSL https://raw.githubusercontent.com/xu-jin-cs/dsh-skills/main/scripts/dsh-skill.sh | bash -s -- install --all --with-deps

# 2. Score all your local agent skills → radar + leaderboard + HTML report
python3 ~/.dsh/dsh-skills/agent-eval/scripts/eval_agents.py --mode generic

# 3. Watch a mechanical gate verify this repo's own README (run from the repo root)
python3 gate-switch/scripts/gate_switch.py --spec examples/gate-switch/demo_spec.json --set target=README.md

Sample outputs are checked in: radar · leaderboard · more in examples/.

Skills

Skill What it does
agent-eval Agent/skills capability evaluation & visual reports. Read-only collection → 9-dimension frozen-baseline scoring → three-tier strength/weakness verdicts with action guidance → overall ranking → dark-tech radar chart + rank board + HTML report.
archmap Architecture cartography agent (self-contained Python engine). Zero-arg full/lite auto-routing; requirement text → precise impact surface (file/function/route level); diff mode: zero-LLM line-level impact + import closure + test selection + change ledger; sync incremental baseline refresh. Deterministic computation instead of full-repo reading — massive token savings.
gate-switch Universal probabilistic-execution gate skeleton (evidence-family engine, zero deps). Cures three LLM chronic failures: skipped steps / half-done checklists / fabricated "done" claims. Write what must be true as a spec JSON; the engine mechanically verifies each check — A passes, B blocks with the violations as the reason. 7 frozen check primitives (file_exists / json_field / glob_count / grep_count / mtime_after / script_exit …) + ready-made gate instances + an L3 framework-gate template. New scenario = new spec, zero engine changes.
parallel-dispatch Master rules for parallel dispatch & sub-agent clones. ≥2 independent subtasks trigger parallel fan-out by default; two-axis decision (scale: light clone / task-breakdown / engine-level × count: subagent / grouped / workflow); scene auto-matching, minimal probe, merge checkpoints — all through the audited dispatch_switch SPDT gate (A/B verdict, no handwritten decisions).
expert-clinic Self-contained diagnostic workflow agent for hard problems. Single file, zero dependencies — works as a system prompt on any platform. Fixed pipeline: intake → expert triage → 7-step deduction (locate the gap → root cause → candidate fixes → pseudo-solution detection → converge) → plan selection (fewest steps wins) → four-force logic check → ROI assessment → consultation report.

All five skills are first-class top-level directories — clone and use, or install --all. A 57-skill bundle also ships in Xj-rules/store-package (store-package-full.zip / -lite.zip).

Xj-agent — full-lifecycle PM workflow

A self-contained PM orchestration skeleton (13 nodes: pm_bootstrap → spm → pm_prd_confirm → dpm → [ui_designer ∥ test_lead_design] → fe → be → pm_quality_gate → test_lead_full → ops → qa → process_audit → retro), wired by default to this repo's Xj-engine and pluggable via env vars. Ships with 11 role skills under Xj-agent/agents/ so the workflow runs out of the box.

pip install -r Xj-agent/pm/requirements.txt
python3 Xj-agent/pm/scripts/flow_kernel.py routes --rules Xj-agent/pm/flow.yml --node be

Xj-engine

The standalone engine (engine.kernel.et / xj-engine CLI): ETL + task domain + bridge executor, local database, pip install -e . installable. See Xj-engine/.

Compatibility

Agent Skill directory
DeepSeek Harness ~/.dsh/skills or ~/.agents/skills (watcher hot-reload)
Claude Code ~/.claude/skills
Kimi Code ~/.agents/skills
Codex ~/.codex/skills

Skills are plain SKILL.md + YAML frontmatter; only the discovery root differs between agents.

Install

One-liner (recommended, no clone needed)

# List all skills
curl -fsSL https://raw.githubusercontent.com/xu-jin-cs/dsh-skills/main/scripts/dsh-skill.sh | bash -s -- list

# Install one skill (symlinked into ~/.dsh/skills, hot-reloaded by DSH's watcher)
curl -fsSL https://raw.githubusercontent.com/xu-jin-cs/dsh-skills/main/scripts/dsh-skill.sh | bash -s -- install archmap

# Everything + dependencies
curl -fsSL https://raw.githubusercontent.com/xu-jin-cs/dsh-skills/main/scripts/dsh-skill.sh | bash -s -- install --all --with-deps

One-click downloads without a clone

# Just agent-eval (standalone installer: downloads only that skill, wires deps, smoke-tests)
curl -fsSL https://raw.githubusercontent.com/xu-jin-cs/dsh-skills/main/agent-eval/install.sh | bash
  • 57-skill store pack: store-package-full.zip / store-package-lite.zip
  • Whole repo zip: GitHub Code → Download ZIP, or curl -fsSLO https://github.com/xu-jin-cs/dsh-skills/archive/refs/heads/main.zip

First run shallow-clones this repo to ~/.dsh/dsh-skills (override with DSH_SKILLS_HOME); all later commands run locally.

Already cloned

git clone https://github.com/xu-jin-cs/dsh-skills.git
cd dsh-skills
./install.sh                # interactive picker
./install.sh archmap        # a specific skill
./install.sh --all          # everything

CLI subcommands (scripts/dsh-skill.sh): list / install (--copy, --target DIR, --with-deps) / uninstall / update / doctor.

Principles

  1. Lightweight — rule-type skills are a single SKILL.md; engine-type skills are self-contained with explicit requirements.txt.
  2. Universal — no private engine logic, no backend lock-in; standard SKILL.md works across agent hosts.
  3. Auto-trigger — triggers live in each skill's description; the host injects them into the session catalog for scene matching.

Methodology

The governance philosophy behind these skills, plus a 27-case battle retrospective: Mechanical Gates for LLM's Verbal Promises (中文版) — the Mandatory-Completion Gate meta-method, L1/L2/L3 gate levels, skeleton-freeze discipline, and the "1 proven case + N named siblings" generalization gate.

Contributing

Every release passes a pre-publish compatibility checklist (portable paths / zero credentials / no private deps / hygiene scan). See 「上传必看」 in README_zh.md.

License

Custom Source-Available license — see https://github.com/xu-jin-cs/dsh-skills/blob/main/LICENSE (Chinese version: https://github.com/xu-jin-cs/dsh-skills/blob/main/LICENSE-zh.md). Not MIT, not CC BY-NC-SA, not an OSI-approved open source license.

Allowed: view, study, and run the Software locally for non-commercial, personal internal use; internal modifications for your own use.

Forbidden without prior written permission (contact: https://github.com/xu-jin-cs): publishing modified versions as a separate public repository; selling, commercially distributing, or sublicensing the Software or its derivatives; providing SaaS, online, or other commercial services based on it.

—/ 5

No ratings yet

Manifest verification required

Commit 943053cb543e

Community comments

No comments yet. Be the first to write one.

DSH HUB

A community index for DSH plugins. Not an official GitHub or DeepSeek AI product.

CommunityResourcesAPIAbout