Your production-grade dev team for DeepSeek Harness.
Vibe coding makes fast fashion: use it once, throw it away. dsh-auto-dev ships products you can launch and maintain.
English · 简体中文
Targets the Web UI of DeepSeek Harness 0.2.1-rc.1.
- Vibe coding: one model writes code; you babysit everything else.
- dsh-auto-dev: a whole dev team. Production-grade results, no coding skills required.
- Vibe coding: works in the demo, breaks on real data.
- dsh-auto-dev: best practices, fully tested, reproducible and stable.
- Vibe coding: quality depends on the model's mood that day.
- dsh-auto-dev: debugging, review, release, plus best practices for major languages and platforms — the right method for each situation.
- Vibe coding: long chats forget; every new chat starts from zero.
- dsh-auto-dev: switch chats, come back tomorrow, hand it to a teammate — pick up where you left off.
- Vibe coding: scroll through hundreds of lines to guess how far it got.
- dsh-auto-dev: one board — what's done, what's left, what's waiting on you.
- Vibe coding: every chat is a first meeting; the same pitfall, ten times.
- dsh-auto-dev: pitfalls, environment and business rules are recorded automatically and reused next time.
- Vibe coding: starts on a one-liner, builds the wrong thing.
- dsh-auto-dev: it reads the code before touching anything and registers the task only once the request is clear. Small requests it restates as requirements; product-level or multi-part ones it first breaks down into one professional objective (O — with better approaches and the unknowns you may not have considered) and a few steps (KRs), shows you that plus anything you need to decide, and starts only once you say go; technical details it decides itself. Turn on autopilot and it never waits: it picks the best approach, does the work, and lists every decision it made without you at hand-in.
- Vibe coding: one slip and your code is gone.
- dsh-auto-dev: saves when you start, before a fix and after it passes; roll back anytime. Hard resets, force-pushes, publishing and deploys always need you to click yes first — even on autopilot.
git clone https://github.com/plyflai/dsh-auto-dev.git
dsh plugin --profile <profile> add ./dsh-auto-devRestart DSH afterwards. No runtime dependencies; works out of the box.
In any chat, start with /dsh-auto-dev and say what you want:
/dsh-auto-dev add a search command to the notes CLI. Done when `notes search milk` finds "buy milk".
That's it. Planning, coding, testing and delivery are on it; when a decision is yours, it stops and asks.
Away from the keyboard? Start it on autopilot:
/dsh-auto-dev start --auto add a search command to the notes CLI.
Continuing tomorrow in a new chat:
/dsh-auto-dev adopt
Chats that never use /dsh-auto-dev are left completely alone.
- Receipts: a card in the chat at every milestone — started, unit done, blocked, waiting for you, delivered.
- Dock: one line above the input — how far the task is and whether it needs you.
- Board: the Auto Dev right panel — goal, acceptance criteria, what is still unverified, the latest result, the engineering log.
| Command | |
|---|---|
start <request> |
Start work, or pick up this project's unfinished task (start --auto <request> starts on autopilot) |
new <request> |
Always open a new task |
status |
Current task, progress, what is pending |
tasks |
Recent tasks in this project |
adopt [taskId] |
Continue a task from another chat |
pause · resume · cancel |
Change the task's state |
auto on · auto off |
Autopilot for this task: on, it never waits for you and lists its unconfirmed decisions at hand-in; irreversible steps are still left to you |
organize |
Propose a cleanup of a long plan; nothing changes until you agree |
plan · help |
How to begin |
All are /dsh-auto-dev <command>; plain text after /dsh-auto-dev means start.
Kept in your project's .auto-dev/ as plain files you can read and edit:
| File | What it holds |
|---|---|
incidents.md |
Postmortems: symptom, directions tried, root cause, fix, how it was verified (required after a fix; at most 30 entries) |
project-domain.md |
Business rules: terms and the rulings you made |
memory.json · runtime.json |
Pitfalls and observed environment facts (with source and expiry) |
Their text goes to the model at every start. Commit them to share with your team, or add .auto-dev/ to .gitignore. Task progress lives in DSH's own storage.
Experience across projects lives in ~/.dsh-auto-dev/:
| Path | What it holds |
|---|---|
incidents/<project>.md |
A copy of each project's postmortems at hand-in; another project's start carries the newest entry of the 3 most recent projects (at most 20 projects kept) |
conventions.md |
Team conventions you write yourself (naming, style…), carried verbatim at every start; the project's own conventions win on conflict |
DSH_AUTO_DEV_HOME moves this location.
- Checkpoints are Git-based (under
refs/dsh-auto-dev/) and never move your HEAD, branch or index. Each task keeps its latest 5; the repository keeps the saves of the latest 3 tasks, the rest is left togit gc. - Rollback restores Git-tracked files only, and every rollback needs your click. Untracked files, databases and running services are not covered.
- Only four gates ever stop the model; everything else is a reminder:
- Honest delivery: no passing project check after the last change — the plugin runs it itself and refuses only if it is red;
- Destructive commands:
git reset --hard,git clean -f, force-push,branch -D,stash drop, anypublish, terraform / kubectl / helm,vercel --prod,fly deploy,gh release createrun only after you click yes — on DSH's own approval card when the session asks for approvals, otherwise on a two-option question card. Typed words are never read as a yes; on autopilot these steps are skipped and listed for you; - Your bans: a place you said not to touch is refused; lifting a ban takes the same click;
- Postmortem on disk: a bug was fixed but
incidents.mddid not change — the hand-in is refused once.
- Commands outside that list, such as
rm -rfor SQL, are not stopped; checkpoints and the "deleted · still referenced" card are the safety net.
In chats that opted in, the model gets eight tools:
| Tool | Purpose |
|---|---|
work_start |
Register the task once the request is clear, in the model's own restatement (your words and replies are kept beside it; if the model edits files first, the plugin registers your words for it) |
work_plan |
Optional units, dependencies and grouping |
work_submit |
Hand in a result tied to real executions from this session |
work_status |
Read the task, its executions and the board |
work_control |
Pause, accept, rework, hand off, checkpoints, notes |
work_extend |
For things that need your OK: tells the model what to ask you (it never grants anything itself) |
work_env |
Facts about the environment, with their source |
work_recall |
Postmortem knowledge base: list every incident title (this project and others), or search entries by words |
The engineering playbook lives in skills/auto-dev.
Local development needs Node.js ≥ 20:
npm install
npm run check # unit tests + end-to-end scenarios against the real plugin
npm run build:client # rebuild lib/client.js after editing src/client.js
npm run hooks:install # run the check before every commitHost-side changes need a DSH restart; client changes need a page refresh.

