dsh-channel-gateway
A channel gateway plugin for DeepSeek Harness.
A channel adapter normalizes provider traffic into one message contract, the
gateway authorizes it and announces a single channel/inbound event, and a
consumer answers through ctx.channels.send with the route the inbound message
carried.
The package covers the transport half of a chat integration: it fixes the
message contract and the seam between a provider and a consumer. What a message
means, whether it was seen before, and where it goes next are decided above it.
Contract
Five surfaces, and they are the public API — keep them small, stable and
versionable. Field-by-field reference in docs/architecture.md.
| Surface | Shape |
|---|
| Inbound | InboundMessage { channel, providerMessageId, actor, place, visibility, text, attachments?, timestamp, replyTo?, raw? } |
| Outbound | OutboundMessage { channel, route, text, attachments?, replyTo?, format? } |
| A transport | Channel { name, capabilities, start(inbox), stop(), send(message), resolveAttachment? } |
| The gateway | ctx.channels.register(channel) · ctx.channels.send(message) · ctx.channels.resolveAttachment(channel, attachment) · event channel/inbound |
| Authorization | ChannelAuth { authorize(request) }, defaulting to an allowlist |
Three parts of the contract exist because a consumer needs them and a provider
does not supply them in a usable form:
actor / place / visibility — who spoke, where, and in front of whom.
providerMessageId — the provider's stable id, so a consumer can
deduplicate and accept durably.
place.route — an opaque, channel-minted destination. A consumer stores
the route it saw and hands it back as OutboundMessage.route; no layer above
the adapter composes a provider address.
Install
The package is a DSH bundle: adding it inserts the gateway service row.
dsh plugin add @wowyuarm/dsh-channel-gateway --profile <profile>
Channel adapters are opt-in rows, because each needs its own credentials. A
profile adds the ones it wants:
- insert:
- id: channel-telegram
name: '@wowyuarm/dsh-channel-gateway/telegram'
config: { token: '<bot token>' }
- id: channel-weixin
name: '@wowyuarm/dsh-channel-gateway/weixin'
config: { token: '<iLink bot token>' }
Who may speak is the gateway row's own configuration, not the adapter's:
- id: channel-gateway
config:
allow:
- telegram:123456789 # one actor of one channel
- weixin:* # every actor of one channel
# - '*' # everyone
An empty allow list admits nobody.
Consuming it
import type { Context } from '@deepseek-ai/cordis'
import type {} from '@wowyuarm/dsh-channel-gateway'
export const name = 'my-ingress'
export const inject = ['channels']
export function apply(ctx: Context) {
ctx.on('channel/inbound', (message) => {
// Deduplicate on (message.channel, message.providerMessageId), accept it
// durably, route it.
void handle(message)
})
}
async function handle(message: InboundMessage) {
// ... later, answer it:
await ctx.channels.send({ channel: message.channel, route: message.place.route, text: 'ok' })
}
Adapters
| Adapter | Transport | Entry | Notes |
|---|
| Telegram | Bot API long poll (getUpdates) | @wowyuarm/dsh-channel-gateway/telegram | Text and media. An attachment sends from local bytes (data), else a provider ref or url; resolveAttachment reads an inbound attachment's bytes by download. format: 'markdown' renders as Telegram HTML. DSH_TELEGRAM_TOKEN is the fallback for config.token. |
| Weixin | WeChat iLink long poll (ilink/bot/getupdates) | @wowyuarm/dsh-channel-gateway/weixin | Text and media. A reply quotes the inbound context_token, which WeChat expires after roughly two minutes; the adapter refreshes a stale one before sending. format: 'markdown' renders as sanitized WeChat Markdown. The media download/upload path follows the reference iLink protocol and is not yet verified against a live account. Starts from a token it is given — the QR login flow is not implemented yet. |
Both adapters honour HTTPS_PROXY/NO_PROXY (Node's own fetch does not, unless
NODE_USE_ENV_PROXY is set).
Compatibility
A DSH release can reach this package only through the two published packages it
imports: @deepseek-ai/cordis (the plugin and service API) and
@deepseek-ai/schemastery (row config). No @deepseek-ai/dsh-* package is
imported — npm run check:boundaries fails the build if one appears — so the
host coupling is exactly those two versions.
| Runtime | cordis | schemastery | Verified by |
|---|
DSH 0.1.5-rc.3 (npm latest) | 4.0.2 | 3.18.2 | the 0.1.0 release: npm run typecheck, npm test, npm run build |
DSH 0.1.7-rc.1 · 0.1.7-rc.2 (npm next) | 4.0.4 | 3.18.4 | the same three checks, plus a real row mount in a booted profile |
| Declared peer floor | 4.0.1 | 3.18.1 | the same three checks |
The devDependencies pin the second row, so a checkout typechecks, tests and
builds against what the current DSH ships. The declared peers stay at
^4.0.1 / ^3.18.1: the floor admits the older DSH line, the ceiling admits
the current one, and the two ranges overlap, so one published version installs
against either without a second copy of cordis in the profile.
Re-verify against a running DSH:
npm run build
cat > /tmp/channel-gateway.yml <<'EOF'
- insert:
- id: channel-gateway
name: 'file:///absolute/path/to/dsh-channel-gateway/lib/index.js'
EOF
dsh --profile <profile> --patch /tmp/channel-gateway.yml --help
A row that fails to import is reported on stderr as N entry did not activate
with its reason; a clean run prints no such line. To assert the service seam
rather than the import alone, add a second row to the overlay that declares
inject: ['channels'] and reads ctx.channels.send from it.
Development
npm install
npm run typecheck # tsc, strict
npm test # boundary guard + vitest
npm run build # emits lib/
src/ may import only its own relative modules, Node builtins, and
@deepseek-ai/cordis / @deepseek-ai/schemastery; adapters may add undici.
npm run check:boundaries enforces exactly that, so the neutral seam cannot
quietly acquire a host dependency.
License
MIT. The Telegram transport is adapted from
dsh-telegram-channel (MIT);
the WeChat protocol shape follows
openclaw-weixin (MIT) and
nanobot (MIT).