@ruqinhu/dsh-weixin-link

English | 中文

Weixin (WeChat) iLink channel for DeepSeek Harness — send, receive, and deliver turn results automatically.

DSH has no built-in IM delivery. This plugin adds two things:

1. an agent can push a report to Weixin (weixin_push tool)
2. when a turn ends, its text is sent to Weixin by the plugin — the model does not have to remember to call a tool

The second one is the point. See "Why delivery is not the model's job".

The constraint that shapes everything

iLink requires a context_token on every send, and that token can only be obtained from an inbound message. There is no way to mint one.

So a quiet night means the 07:00 report cannot be delivered until the owner sends something. Rather than pretend otherwise, the plugin parks such messages and replays them the moment a real inbound message refreshes the token. No fallback channel, by design — a silent fallback hides the failure instead of surfacing it.

Why delivery is not the model's job

An earlier version relied on a sentence in every scheduled prompt: "when done you must call weixin_push". It failed quietly:

• The model wrote its answer into the session, the turn ended completed, and the owner received nothing. From the model's point of view, answering in-session was the correct thing to do — it had no way to know the turn was addressed to a phone.
• A different failure: the model called weixin_push with the body, then added "已通过微信回复你了" in the session. The owner got the meta line, not the content.
• With the task failing, the prompt still said "you must send", so the error was delivered as if it were the report.

Delivery is a transport concern, not a matter of model discipline.

The plugin now subscribes to DSH's global session/event bus and, on turn/end, takes the last assistant/message text of that turn and sends it itself. DSH's dsh-schedule only awaits sessions.flush() after queueing a message; it neither knows nor cares what the agent answered — so the channel has to do this.

Which turns are delivered

Keyed on the source.kind of the turn's first user/message:

source.kind	Origin	Delivered
schedule	cron task	✅
weixin-link	owner messaged from Weixin	✅
user	typed in terminal / web UI	❌ the answer is already on screen
runtime-context / skill-catalog / time-context	injected context	❌ not speech
compact-checkpoint / provider / fallback	compaction, session-title generation	❌ internal maintenance; sending it is noise

Exactly once

One delivery per turn. If the model called weixin_push itself during that turn, auto-delivery stands down for that turn — otherwise the owner gets the same content twice.

Empty turns

A scheduled task that finishes without output sends 本次无产出（timestamp）. Silence would make "ran and failed" indistinguishable from "never ran", and those two need different responses.

A prompt can opt out of that notification: when the entire turn text is exactly [SILENT], delivery is skipped. Only whole-text equality counts; mentioning the marker inside a real report does not suppress it.

Length cap

2200 characters per message, enforced. iLink's failure for over-long payloads is not a clean rejection — a send can return a message id and still never arrive. Full reports go to the HTML page; the message carries the summary and, when truncated, a line saying so plus the link on the first line.

Enforced in code rather than left to the model, because delivery no longer consults the model at all.

Install

dsh plugin add @ruqinhu/dsh-weixin-link

Credentials

The token is read through the DSH credential store — via ctx.credentials.resolve() — not from process.env directly. That is deliberate: the store already layers $DSH_HOME/.credentials.yaml, inherited environment, and project/user .env files. Reading the environment directly would shadow the managed store.

Configure the credential under the name in tokenEnv (default WEIXIN_BOT_TOKEN).

Configuration

Key	Default	Meaning
tokenEnv	WEIXIN_BOT_TOKEN	Credential name holding the bot token
accountIdEnv	WEIXIN_BOT_ACCOUNT_ID	Credential name for the account/bot id
baseUrl	https://ilinkai.weixin.qq.com	iLink endpoint
dataDir	~/.dsh/weixin-link	Tokens, outbox, receipts, delivery trace
poll	true	Run the inbound long-poll loop
pollTimeoutMs	40000	Client budget; must exceed the server's 35 s hold
defaultUserId	""	Recipient; empty means "most recent inbound sender"
echoOwnMessages	false	Surface the bot's own messages as inbound

Send policy

Auto-delivery and weixin_push share one send path:

1. Try once.
2. Classify the failure:
   • stale token → park, do not retry (only a new inbound message helps)
   • rate limited → back off 30 s, 60 s, 120 s; if still limited, park
   • anything else → one immediate retry, then give up
3. Write one JSONL receipt per attempt.

Parking on exhausted rate limiting is not incidental. A report that exhausts the backoff ladder and is then dropped looks identical to one that was never scheduled — the failure mode that once swallowed a whole morning report.

The one exception: auto-delivery that finds no context_token at all (no inbound message ever received) does not park. Replaying a week-old report is worse than dropping it.

Tools

Tool	Purpose
weixin_push	Send text on demand; parks automatically when it cannot deliver
weixin_status	Login state, token age, last delivery, outbox depth

weixin_push remains for "send this now". Routine delivery does not need it — the plugin does it, and calling the tool stands auto-delivery down for that turn to avoid a duplicate.

Receipts and trace

Receipts dataDir/receipts/YYYY-MM-DD.jsonl: one line per send attempt, keyed by the operator's +08:00 day rather than UTC.

Delivery trace dataDir/delivery-trace.log: one line per turn decision, including skipped turns and why.

Keeping them apart matters: a receipt only exists for a send that was attempted, and a turn skipped by policy leaves no receipt at all. When a report does not arrive, the first question is whether delivery was considered — only the trace answers that.

send session=... turn=12 len=63
ok   session=... turn=12 msgid=7521135910923441000
skip session=... turn=13 reason=source-not-deliverable

Inbound delivery

Inbound text is delivered to the most recently active root agent as an ordinary follow-up turn, tagged with source weixin-link. Without an agent in the profile the token cache is still refreshed — which is the part that unblocks delivery.

On a cold start the session is resumed through sessionController.resolveAgent(), which returns one of three shapes — a bare Agent, { agent }, or { error } — and all three must be handled.

sessionController must be listed in static inject. cordis throws on an undeclared service rather than returning undefined; swallowed by a try/catch, that presents as "messages arrive but nothing ever answers".

License

MIT