🔁 Auto-resume on Web restart —— after a dsh web restart, the plugin scans all sessions and picks up the interrupted ones; completed / settled / cancelled sessions are left alone.
🌐 Transient upstream recovery —— rate limits (429), 5xx, network errors (ECONNRESET / timeout), gateway 504, upstream provider faults —— the plugin injects "continue" and the turn retries automatically.
💳 Quota / billing recognition —— 402 / QUOTA / insufficient_balance / insufficient_quota: inject once (covers recovery after top-up), then go settled on repeat failure to prevent infinite loops.
🛡️ Dead-loop guard —— if the previous "continue" produced no content and the same error recurs, the plugin falls back to settled and hands control back to you.
📥 Queue guard —— if the inbox has someone else's pending message (user / client), the plugin does nothing at all (no inject, no resume), avoiding waking the user's message into the session.
⏱️ liveWatch —— keeps polling after boot to catch failures that recur during a run (previously only worked once per restart).
🔢 Two-shot auto-continue —— transient network failures (429 / rate limits) are allowed to inject → fail → inject again → fail → stop, suited to long rate-limit windows; quota failures stay one-shot.
🌍 Cross-platform —— Linux / macOS / Windows behave identically (os.homedir() + node:path, path separators auto-adapt).
Install
dsh plugin --profile web add dsh-autoresume
After installing, restart dsh web. The plugin will scan and inject within the boot grace window (default 30 minutes, configurable via bootGraceMs).
Requires DSH client packages >=0.1.5-rc.2. Depends on @deepseek-ai/dsh-agent (reuses the DSH runtime copy, zero version drift).
You can also install directly from GitHub:
dsh plugin --profile web add github:shengyvself/dsh-autoresume
What you can do with it
After installing, just use DSH normally — the plugin works in the background. Typical scenarios:
"I just changed code and restarted dsh web" —— the plugin scans and picks up the interrupted dev session
"The upstream model hit a rate limit" —— after 429 / rate_limit_exceeded, the plugin injects "continue" and the turn retries
"That 504 gateway timeout earlier" —— gateway HTML error envelopes are recursively unwrapped, correctly judged as network failure, "continue" injected
"Out of balance, just topped up" —— first 402 injects once (covers recovery after top-up), then settles automatically on repeat failure
The plugin never acts on its own; it only judges and injects on web restart or liveWatch poll.
Compatibility
Use case
DSH version
Plugin version
Recommended
0.1.5-rc.2+ (currently maintained)
0.0.21
Minimum compatible
0.1.5-rc.2+
0.0.21
Installing the plugin does not upgrade the host DSH. The peerDependencies declare a minimum client-package version of >=0.1.5-rc.2; verified working on 0.1.5-rc.2.
Configuration
Key
Default
Description
targetSessionId
full scan
When set, only serve that session (compatibility mode)
bootGraceMs
30000 (30 min)
Grace window after web boot for judging
initialDelayMs
3000
Initial check delay
pollIntervalMs
5000
Poll interval when session not ready
promptText
继续(自动)
Injected message body
liveWatch
true
Keep polling after boot to catch mid-run failures
maxResumeAttempts
2
Max consecutive auto-continue attempts for transient failures
skipWhenInputPending
true
Queue guard: skip injection if inbox has someone else's pending message
Security
Reads session event streams only (ctx.sessionPersistence); never writes user files.
Injected messages are sent as plugin identity (source.kind=plugin, form=notice), never disguised as user messages.
No network calls. Except for calling DSH's internal ctx.agents.get() / agent.followup(), there are no outbound requests.
@deepseek-ai/dsh-agent is declared in peerDependencies (provided by the host DSH).
⚠️ Known issue: package.json's dependencies also lists @deepseek-ai/dsh-agent: 0.1.1-rc.2 —— this violates the persona §三.1.3 rule (official @deepseek-ai/* packages go only to peerDependencies). The current install path is a manual symlink reusing the DSH runtime copy, avoiding version drift. Will be cleaned up in the next release.
🌊 The original agent harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, federation, vector RAG integration, and native Claude Code / Codex / Hermes and many more Integrated