Skip to content

agent-bot

Verified

@eternalnight/agent-bot · v0.1.6 · MIT · Web UI

agent-bot:DeepSeek Harness 的多智能体团队协作与交付流水线(AgentTeams 衍生,MIT)。

Install

dsh plugin add @eternalnight/agent-bot

Confirm the layer applied with dsh --profile default --dump-config — see the install guide.

Source

Tags

Readme

Little fat fish
agent-bot

AGENT-BOT · THE PET

A team hall
that actually ships

Turn one DeepSeek Harness session into a real delivery team: task DAG, independent review, delivery reports. Plus a pet that sleeps, roams and wakes on click — same character in DSH and on Windows.

Download the pet ↓ · GitHub ↗ · npm ↗

GitHub stars  npm  release  license

English (default) · 简体中文(镜像)

25 6 10 min 0
pet animations clickable facilities real sleep cloud dependency
idle / click / sleep / work / drag / roam gate / doctrine / progress / ranks / gear / about stays asleep until clicked, or 10 minutes pass .agent-teams + local saves, MIT

LIVE HALL · CLICKABLE

A real interface, not a mockup

Here is an actual delivery from the plugin: how the goal is written, how the interview is decided, how the artifact runs in place at delivery time, and how the run is summarised. Real machine captures (Windows 11 · DSH desktop, 2026-11-05), copied verbatim into assets/screenshots/ — not beautified, not stitched, not redrawn.

Founding step 1: goal, development mode and lineup Founding step 2: requirements interview, then decide

▲ Goal “make a snake game” · mode and lineup · visual explosion — the interview (runtime / theme / motion) ticked and decided

Delivery: the driver gateway runs the artifact while the chat keeps the evidence

▲ At delivery, the driver gateway runs the freshly shipped artifact while the chat keeps the smoke result, the independent review and the real-render evidence

Delivery result: 2/2 passed, 0 blocked, 13m33s The artifact itself: NEON SERPENT

▲ Result: 2/2 passed · 0 blocked · 13m33s · 165.2K tokens · ¥1.66 — the artifact: NEON SERPENT

One real delivery, five steps:

Step Image What happened
1 任务目标.png Goal set to “make a snake game”, mode and lineup picked; optional visual explosion and remake reference
2 任务拍板.png The AI drafts the requirements interview — runtime, theme, motion — you tick answers and decide
3 Plan review Members, tasks, dependencies, acceptance and verify commands laid out; nothing is created before approval
4 任务交付.png The captain dispatches, members deliver; the gateway runs the artifact with smoke, review and render evidence
5 交付结果.png / 交付demo.png The run reports 2/2 passed, 0 blocked; the artifact is a playable snake game

The numbers in the images are that one run (13m33s · 165.2K tokens · ¥1.66). Your tasks will differ — this is a record, not a promise.


THE PET · LITTLE FAT FISH

Sleeps, roams, and wakes up to work

Windows and DSH share the same character assets; the desktop copy ships as an installer, and the collapsed DSH token is the same pet. Transparent, always-on-top, draggable — the window holds the character and nothing else.

Desktop pet main window First launch Collapsed draggable token

▲ Main window (character only) · first launch · the draggable token after collapsing the sidebar

Behavior sheet: click response / sleep / working / dragging / idle variant

▲ Five frames: click response · sleep · working · dragging · idle variant

→ 25 clips: click responses, a sleep chain, six work poses, drag feedback, idle variants
→ Roams every 18–28s while idle, then pauses and changes clip
→ Sleeps for 10 minutes; click to wake early, otherwise it keeps sleeping
→ Double-click wakes DSH and opens the hall; launches DSH if it is not running
→ A single click only pets it — it never wakes or resizes the DSH window; hover shows a pink Pop-style status card (level · growth · satiety · mood · energy)

IDLE / FEED / SLEEP / WORK / DRAG · REAL FRAMES · MIT


WHY IT IS DIFFERENT

Not another agent dashboard

Most agent UIs hand you a wall of logs. This one gives you a place: who is working, what is blocked, what shipped — and the evidence that it shipped, all in the same spot.

Pillar What it means
A HALL, NOT A DASHBOARD
01 A place, not a log wall
Six clickable facilities, a live scene and an expandable right sidebar. Open a member conversation, expand the hall, review the delivery — in one place.
DELIVERY, NOT VIBES
02 Gates that prove delivery
Task DAG, write scopes, acceptance commands and independent review. Implementers cannot approve their own work; the host re-runs the declared commands.
ONE PET, TWO PLACES
03 The same pet in DSH and on Windows
Real frames: random idle, 18–28s roaming, click responses and a real 10-minute sleep. The collapsed DSH token is the same character.

STILLS

A few stills for the details

The hall and the pet are live in the product, so the stills are trimmed to what they add. Click any one to enlarge.

Founding goal and lineup Delivered artifact Animation frames

▲ The founding wizard inside DSH · the delivered artifact · 25 animation frames

Open the GitHub repo →


UNDER THE PIXELS

A real team runtime under the hall

→ Captain-led task DAG; dispatch only when dependencies are done
→ Durable member sessions, mailbox, attempt ownership
→ Quality contracts: objective / scope / acceptance / verify / verdict
→ Independent review, repair rounds, cold recovery
→ Per-role write scopes; no overlapping edits

.agent-teams/<teamId>/
  team.json
  inbox/
  delivery-report.md
# local-first; no cloud account

WORKSPACE STATE · MIT


QUICK START

One command to found the hall

Install it and the founding wizard, plan review, the right-hand hall and the durable team are ready. The desktop pet installs separately and needs no Python.

DeepSeek Harness

dsh plugin --profile desktop add @eternalnight/agent-bot
# reopen DSH → create a team

Desktop pet installer

DSH-Pet-Setup-<version>.exe
# per-user install into %LOCALAPPDATA%\Programs\DSH Pet, desktop shortcut
# most recent pet build · GitHub Releases

From GitHub

git clone https://github.com/EternalNight996/agent-bot
dsh plugin --profile desktop add ./agent-bot

Smallest loop: Create team → fill goal and lineup → tick the interview and decide → review the plan before approving → watch members work, and see the artifact run in the driver gateway on delivery.

Run the pet from source:

.\desktop-pet\start-pet.cmd
& "$env:USERPROFILE\.dsh\dsh-runtimes\dsh-primary-runtime\dependencies\python\python.exe" desktop-pet\pet.py --workspace .
python desktop-pet\pet.py --check --workspace .

Runs for real (2026-11-05): pet.py --check --workspace . exits 0 with {"ok": true, "level": 2, "stage": "fry", "frames": 16, ...} (completedTasks follows finished work, so it is not pinned).


DELIVERY GATES

Delivery gates

Signal Evidence
Full suite pnpm verify exits 0
Browser-side contract pnpm verify:web-routes passes
Release artifact Package smoke test checks runtime and support QR codes, excludes source / tests / docs
A real team run Captain dispatches → members write → independent review → synthesis (covered by the full suite)
Isolation An implementer cannot approve its own implementation

This is a tool for parallel wall-clock time plus independent verification, not a promise that every task gets cheaper; for small tasks a single agent is usually the better deal.

Quality gate configuration

The plugin config field quality controls how gates are audited:

Option Default Effect
autoVerify true The host re-runs every declared verify command before accepting a finished quality task
autoReview false Completes ready review tasks automatically once gates, acceptance and host commands pass
requireAcceptance true Automatic review requires every acceptance item of the reviewed task to pass
autoDetect true Detects the generated project's own package.json test / typecheck scripts and runs them too
reviewMode "auto" auto: the host closes it; human: stops at “awaiting human review”
commands ["git diff --check"] Fallback verify commands for generated projects
commandsByProject {} Verify commands per project root
verifyTimeoutMs 120000 Host timeout per verify command
quality:
  autoReview: true
  reviewMode: "human"   # stop at “awaiting human review” after the host passes
  commandsByProject:
    "projects/snake": ["node projects/snake/smoke.mjs"]

ROLES & TEMPLATES

Roles and templates

Lineups are picked by “what happens next”; the seat cap is 5, and idle roles spend no tokens. Six task lineups ship today (defined in src/client/lineup-store.ts, display names from src/agency-role-seeds.ts):

Lineup Roles (id → display name) Good for
Direct delivery product-manager PM · engineering-senior-developer senior developer · testing-evidence-collector evidence collector · engineering-code-reviewer code reviewer Fewest detours, straight to the task
Bug fix engineering-codebase-onboarding-engineer onboarding engineer · engineering-senior-developer senior developer · testing-evidence-collector evidence collector Understand the code, then locate and fix
Full-stack feature product-manager PM · engineering-frontend-developer frontend developer · engineering-backend-architect backend architect · testing-evidence-collector evidence collector · engineering-code-reviewer code reviewer Contract first, front and back in parallel (all 5 seats)
Read-only audit engineering-codebase-onboarding-engineer onboarding engineer · engineering-code-reviewer code reviewer · testing-evidence-collector evidence collector · engineering-technical-writer technical writer Touch nothing, only gather evidence
Performance product-manager PM · engineering-senior-developer senior developer · testing-performance-benchmarker performance benchmarker · engineering-devops-automator DevOps automator Metrics first, conclusions backed by data
Security security-appsec-engineer AppSec engineer · security-cloud-security-architect cloud security architect · engineering-devops-automator DevOps automator · testing-evidence-collector evidence collector Reproduce first, minimal diff

Six mode presets are generated from development modes as well, plus twelve founding templates: direct delivery / bug fix / new feature / research / audit / full-stack / hardening / build / refactor / docs / design review / short drama.

Every role gets a fixed contract: ownership, must / must-not, done criteria, a five-point report and a budget. The full-stack lineup keeps all 5 seats on purpose: one less handoff.


BOUNDARIES

Boundaries

  • No hidden cloud sync: growth and feedback diagnostics stay local.
  • Approval is explicit: a plan draft starts nothing before you approve it.
  • Deployment needs a clear confirmation: agents never quietly replace a running session.
  • Review is independent: an implementer cannot approve its own implementation.

DEVELOPMENT

Development

pnpm install
pnpm typecheck
pnpm build
pnpm verify
pnpm verify:web-routes

After a fresh clone run pnpm build before pnpm verify: lib/client/ intermediates and *.map files are generated and stay out of git. DSH lifecycle skills are resolved live from the shared memory vault and are not vendored here.


If the hall makes your agents ship better
Star the repo so more people stuck with log walls and unproven delivery can find it.

⭐ Star on GitHub → · Download the pet · npm install


COMMUNITY

Community

QQ group WeChat group
QQ group QR WeChat group QR

Buy the author a coffee (optional):

WeChat Pay WeChat Reward
WeChat Pay QR WeChat Reward QR

Sponsorship is optional and changes nothing; the plugin stays MIT and free. See SPONSOR.md.

License & links

MIT © 2026 Mengyou Technology Co., Ltd. Author: 梦游 (Mengyou) · Company: Mengyou Technology Co., Ltd. · Email: [email protected]. Terms and attribution in NOTICE.md and LICENSE. The sidebar panel UI, the delivery governance layer and the desk pet are original to this project.