hello_world

A long-lived AI companion layer for DeepSeek Harness (dsh). It turns a stateless chat model into something that keeps an identity, remembers across weeks, has a mood that shifts on its own, reaches out first, and can be talked to from QQ or WeChat.

Written as a Cordis plugin bundle: one main entry plus five sibling packages that each own a section of the console or a route.

hello_world
├── index.js              main entry — config, storage, dialogue loop, memory tools
├── client.js             console shell injected into the dsh web UI
├── cordis.patch.yml      bundle manifest (patch format)
├── lib/                  ~80 modules, see "Module map" below
├── plugins/              five sibling packages + shared helpers
│   ├── social/           relationship attitude, silence log, per-peer diary
│   ├── growth/           commitments, principles, chapters, self-narrative
│   ├── channels/         QQ / WeChat admission and dispatch
│   ├── console/          web routes, 3D constellation, growth & attitude panels
│   └── discovery/        plugin manifest discovery and sandbox planning
├── client/               four pages + vendored 3d-force-graph
├── skills/image-prompt/  prompt-writing skill used before image/video generation
├── config/               providers.example.json
└── locale/               zh / en

What it actually does

Area	Behaviour
Identity	persona.md, self.md, world-prompt.md on disk, editable at any time. The model reads them every turn; nothing is baked into the prompt string.
Memory	SQLite (node:sqlite) with FTS5. Four layers — working set, episodic rows, consolidated facts, principles. Retrieval is BM25 + link-graph expansion, fused with RRF, deduped with MMR. Importance decays on an Ebbinghaus curve, so unused memories sink instead of piling up.
Affect	PAD (pleasure / arousal / dominance) plus an energy axis. Each turn nudges the state; between turns it relaxes toward baseline on a half-life. Extreme values are compressed with tanh so it never saturates at 0 or 1. Derived labels (开心 / 委屈 / 恼火 / 心软 …) are computed, not stored.
Life line	Time-of-day rhythm, quiet hours, nightly review at a configured time, diaries, and proactive messages driven by a gate that requires a real reason (unanswered thread, a due commitment, a day change) rather than a timer alone.
Relationships	Per-peer attitude vectors (warmth / trust / interest / irritation / avoidance / formality), a silence log, and a diary written from the companion's own point of view.
Channels	QQ and WeChat bridges. Inbound text and media, outbound text, files and voice. Per-peer and per-group allowlists; unknown peers are dropped before they reach the model.
Media	see_image tries a vision chain (deepseek → qwen → zhipu) and steps down when a provider returns an unusable answer. whale_generate_image / whale_generate_video use DashScope's async API behind a two-stage confirmation gate.
Console	WebGL memory constellation with bloom halos, growth timeline, attitude panel, book shelf.
Host access	whale_host exposes shell and file operations in two risk tiers. Anything above the low tier needs a short-lived grant tied to the requester's own recent message.

The confirmation gate

Generation and high-risk host actions never fire on the model's own initiative. The sequence is:

1. The model calls the tool. It returns a ticket describing exactly what will run, and does nothing else.
2. The user replies in chat. The reply is matched against a deny-first regex — 不 / 别 / 算了 / cancel wins over 好 / 可以 / go, so "不用了" never reads as approval.
3. Only then does the ticket execute. Tickets expire; they are also invalidated if no new message has arrived from that peer since the ticket was issued.

Grants work the same way: 30-minute TTL, and the requester must have sent a message after the grant was created. Talking to a bot that is holding a stale grant gets you nothing.

Install

Requires dsh ≥ 0.2.0 and Node ≥ 20.

npx @deepseek-ai/dsh plugin --profile web add github:fangqiing/hello_world

That is the whole install. The bundle ships its own cordis.patch.yml (dsh.bundle.patch), so dsh inserts the rows for you on add.

The rows are inserted disabled on purpose — a plugin that starts talking to people the moment it is installed is a bad neighbour. Open Settings → Plugins, enable hello_world, and restart. First boot creates <DSH_HOME>/hello_world/<worldId>/ and writes the three identity templates there.

To enable the sibling packages as well:

npx @deepseek-ai/dsh plugin --profile web add github:fangqiing/hello_world   # same package, siblings are entry files

# then in your profile's cordis.patch.yml
- insert:
    - id: hello_world
      name: hello_world
      config:
        companionName: 助手
        timeZone: Asia/Shanghai
        life: true
    - id: hello_world-console
      name: hello_world/plugins/console/entry.js
    - id: hello_world-social
      name: hello_world/plugins/social/entry.js

Prefer a working copy over a package install? Clone and point at the directory instead:

git clone https://github.com/fangqiing/hello_world plugins/hello_world
# then: name: plugins/hello_world

Not on npm — github: is the supported path. If you want a pinned install, append a tag or commit: github:fangqiing/hello_world#e320550.

Configuration

Everything lives under the plugin's config: block. Defaults are in lib/config.js.

Key	Default	Notes
companionName	助手	Shown in the console and used as the persona fallback name.
worldId / worldName	default / 深海	worldId selects the data subdirectory. Changing it gives you a clean slate without deleting anything.
timeZone	Asia/Shanghai	Drives rhythm, quiet hours and the nightly review. Validated on load.
scope	all	all = every new session in the profile; workspace = only sessions whose cwd is under workspaceDir.
dataDir	(empty)	Empty means <DSH_HOME>/hello_world/<worldId>.
recallTopK	8	Episodic rows pulled per turn before fusion.
workingMemoryTokens	320	Budget for the always-present working set.
dailyTokenBudget	0	0 disables the cap.
reviewTime	23:00	Nightly consolidation trigger.
quietHours	['23:30','08:00']	No proactive messages inside this window, except a day-change greeting.
proactiveDailyLimit	6	Hard ceiling per day, on top of the per-period quota.
allowExternalTools	false	Lets channel-originated agents run shell/file tools. Leave off unless you have thought about it.

Providers

Copy config/providers.example.json to config/providers.json and fill in what you have. The file is gitignored. Nothing is required — with no providers configured the plugin runs entirely on the host model's own text ability, and only see_image / generation degrade.

The vision chain is tried in order and each entry can be disabled independently. A provider returning a refusal-flavoured answer (rather than an error) is treated as a failure and the chain steps to the next one.

Tools

Tool	Tier	Purpose
memory_search	read	FTS5 + graph-expanded recall over episodes and facts.
memory_read	read	Fetch a specific memory or a time range.
memory_remember	write	Store an episode with importance and optional links.
memory_update	write	Revise or re-weight an existing row.
see_image	read	Vision chain over a local path or URL.
weather_now	read	Current conditions for a city.
whale_status	read	Mood, drives, current peer profile, memory counts.
whale_peer	read	Relationship state for one peer.
whale_read	read	Read recent conversation from a channel.
whale_send	write	Send text to a peer or group.
whale_send_file / whale_send_text_file	write	Send an attachment, or write text to a file and send it.
whale_job_status / whale_job_cancel	read	Track a long-running generation or shell job.
whale_generate_image	gated	DashScope async text-to-image (also used for stickers).
whale_generate_video	gated	DashScope async text-to-video.
whale_host	gated	Shell / file operations, low and high risk tiers.
whale_disclose	gated	Reveal that the companion is an AI, with a configured disclosure policy.

gated means the two-stage confirmation above. write means it acts immediately but only within the peer's own conversation.

Module map

lib/ is grouped by concern rather than alphabetically:

• Storage & recall — store.js, memory-hygiene.js, relevance.js, temporal-memory.js, dormant-traces.js
• Affect & drives — cognition.js, mood-engine.js, life.js, proactive-gate.js, rhythm-policy.js
• Voice — voice-style.js, voice-rules.js, voice-grice.js, voice-cleaner.js, voice-repair.js, voice-variants.js, voice-poetic.js, expression.js
• Relationships — social.js, peer-identity.js, peer-registry.js, per-peer-diary.js, growth.js, principle-engine.js
• Channels — channels.js, channel-admission.js, media-inbound.js, bridge-cursor.js, silk.js, asr-pipeline.js
• Media — vision-tool.js, generation.js, read-jobs.js
• Host & safety — host-tools.js, host-control.js, grant.js, auth-guard.js, disclosure-gate.js, mode-confirm.js
• Console — constellation-3d.js, constellation-layout.js, render-core.js, render-glow.js, web.js

Known limits

• Vision is only as good as the provider you configure. The chain degrades gracefully but cannot make a small model see better.
• Video generation parameters (wan2.7-t2v) are wired from the documented schema but have not been exercised end to end as much as image generation.
• Constellation wants roughly 2 MB of vendored 3d-force-graph on first load. The vendored copy is served from the plugin, not a CDN, so it works offline.
• Memory quality depends on consolidation running. If the process is killed before the nightly review for many days, episodes accumulate and recall gets noisier — the decay curve keeps it bounded but not tidy.
• Channel bridges are a separate package. This bundle speaks to them through a local HTTP API; it does not implement the WeChat or QQ protocol itself.

License

MIT. See LICENSE.

The image-prompt skill under skills/image-prompt/ is adapted from MIT-licensed sources, credited in its front matter.