dsh-coder-preset

English | 中文

dsh-coder-preset

A DeepSeek Harness bundle that adds the Coder agent preset and 35 Matt Pocock engineering skills. The skills are scoped to Coder sessions, so installing it does not change the skill catalog of other agents in the profile.

• 35 skills: 13 model-invocable (loaded automatically) and 22 user-invoked (/name)
• Coder-only: a standard session cannot load them
• No install-time code and no runtime dependencies: configuration, Markdown, and a few skill-owned scripts and templates
• Vendored from mattpocock/skills at commit 5beb35b (v1.3.1 plus 47 commits)
• An engines.dsh gate, a two-build sha256 comparison and an install smoke test that mounts the skill catalog through DSH

Install

1. Download dsh-coder-preset-1.2.1.tgz from the releases page.
2. In DSH, open the plugin manager, install a local bundle, and select the tarball.
3. Restart DSH.
4. Create a session and select the Coder agent.

CLI-managed profiles (not the Electron-managed desktop profile):

dsh plugin --profile <profile> add /path/to/dsh-coder-preset-1.2.1.tgz

On Windows, an upgrade can fail with EPERM while DSH is running. Disable the bundle in the plugin manager, install the new version, then enable it again.

Usage

• Model-invocable skills load automatically when a task matches. You can also type the command, for example /tdd.
• User-invoked skills are hidden from the model. Type them yourself, for example /ask-matt, /to-spec, or /chief-of-staff.
• Run /setup-matt-pocock-skills once per project before the first engineering flow.
• Skills are visible only in Coder sessions.

Common commands:

Command · What it does /ask-matt · Routes the current situation to the right skill or flow /grill-with-docs · Sharpens a plan or design while writing ADRs and glossary entries /to-spec · Turns the conversation into a spec on the issue tracker /to-tickets · Breaks a spec into tracer-bullet tickets with blocking edges /implement · Implements a spec or a set of tickets /tdd · Drives test-first development: red, green, refactor /code-review · Reviews changes on the Standards and Spec axes /pr · Writes the PR body from the template /retro · Runs a retrospective on a coding session

All 35 skills

User-invoked (22) · Model-invocable (13) ask-matt, chief-of-staff, grill-me, grill-with-docs, handoff, implement, implement-spec, improve-codebase-architecture, loop-me, retro, setup-matt-pocock-skills, setup-ts-deep-modules, teach, to-questionnaire, to-spec, to-tickets, triage, wait-what, wayfinder, writing-beats, writing-fragments, writing-shape · code-review, codebase-design, diagnosing-bugs, domain-modeling, grilling, migrate-to-shoehorn, pr, prototype, research, setup-pre-commit, tdd, wizard, writing-for-agents

Compatibility

Verified on Windows with DSH 0.2.1-rc.3: install, the 13 model-invocable skills in a Coder session, /ask-matt, and standard-session scoping.

CI installs the built tarball into a throwaway DSH profile, checks the composed preset rows, evaluates the shipped skill-directory expression with DSH's own loader, and mounts the real skill registry and filesystem provider over the result. It asserts 35 skills (13 model-invocable, 22 user-invoked) and loads sample bodies, so a broken bundledSkillDir or skill provider fails CI. A model turn needs credentials and stays out of scope; CI proves the catalog, not an agent turn.

The bundle declares engines.dsh >=0.2.1-rc.3 <0.3.0. DSH checks that range when it installs a bundle, and an out-of-range bundle is refused with a named requirement and a dsh plugin allow-version remedy; npm run verify-gate proves that refusal end to end with a deliberately incompatible build, and the refused bundle contributes no preset. The range accepts every 0.2.x release because semver treats a 0.2.x patch as compatible, and it rejects 0.3.0. A DSH upgrade inside the range still mounts the pinned preset, so npm run verify-dsh diffs all 32 plugin rows, nested rows and module names included, and reports any drift by name. Widen the range in package.json only after verify-dsh and the smoke test pass against the new DSH.

verify-dsh and verify-gate are manual checks that need an importable DSH, so CI does not run them. CI is a single Ubuntu job with Node 24: validate, unit tests, two builds compared by sha256, then the install smoke test against the pinned DSH CLI. GitHub Actions are pinned to commit SHAs by hand.

• Issue-tracker skills need gh, glab, or the local markdown tracker configured by /setup-matt-pocock-skills.
• /implement-spec needs git worktrees and is not end-to-end verified.
• On Windows, bash-based skills such as /wizard need Git Bash or WSL.
• /ask-matt references /clear, which DSH does not have; start a new session instead.
• Three upstream skills are not shipped: claude-handoff needs the Claude Code CLI, git-guardrails-claude-code needs Claude Code hooks, and scaffold-exercises needs the ai-hero CLI.

Build from source

npm ci
npm run validate
npm test
npm run build

npm run smoke builds the tarball, installs it into a temporary DSH profile, checks the composed preset rows, and mounts the skill catalog through DSH's own registry. It needs pnpm and the DSH CLI, for example npm install -g @deepseek-ai/dsh@0.2.1-rc.3. npm run build writes the tarball and dist/SHA256SUMS; npm run validate checks every vendored skill file against skills.manifest.json, so an edited skill body fails before it can ship.

npm run verify-dsh checks an installed DSH against the engines.dsh gate and diffs all 32 pinned plugin rows, nested rows and module names included, against the shipped standard preset; run it when you upgrade DSH. npm run verify-gate builds a deliberately incompatible copy of the bundle and asserts DSH refuses it, which is the negative half of that gate.

These checks need an importable DSH. The Windows desktop app keeps DSH inside app.asar, which plain Node cannot import, so install the CLI, or run the scripts with ELECTRON_RUN_AS_NODE=1 and pass the desktop executable as --dsh; the resolver maps it to resources/app.asar/dsh/node_modules/@deepseek-ai/dsh. An asar path used without that runner is reported with the same remedy. In a confined agent session where npm test cannot spawn its test-runner children, npm run test:in-process runs the same test files in one process.

License and credits

MIT. The 35 skills are vendored from mattpocock/skills by Matt Pocock under the MIT license. See NOTICE for attribution and for the upstream skills that are not included. Not affiliated with or endorsed by DeepSeek or Matt Pocock.