dsh-session-resilience
安全的 DSH 重启、令牌感知的重新连接，以及基于策略的会话恢复
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:hasan-aghayev/dsh-session-resilience#f9eda1258b799a9f042dafb98e7cbada8f62313aSession Resilience 以 dsh-session-resilience 发布，当前版本为 0.1.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
English | 中文
DSH Session Resilience keeps long-running web sessions recoverable when the local DSH host restarts, disconnects, reaches a token limit, or begins repeating the same work. It combines loopback restart and stop controls with a host-owned continuation engine, a token-aware reconnect, four recovery policies, idempotency guards, adaptive backoff, loop protection, and a small live recovery panel.
restart_dsh and shutdown_dsh; recovery policy and browser controls stay outside the model prompt.Install the package into the environment used by the DSH profile, add dsh-session-resilience to the profile's bundle list, and run the profile's normal reconcile step.
The bundle patch inserts the Host entry and the web-client entry. The package includes prebuilt lib files so a GitHub installation does not depend on a local TypeScript build.
This repository is intended to be the source of truth for a public release. Pin an exact commit when adding it to a profile. Do not use a moving branch in a production profile.
The repository metadata is already configured for this public GitHub repository. Submit one entry for its public URL to the awesome-dsh-plugin catalog after the local checks pass. The catalog validates the bundle manifest and generates the market data used by dsh-market; npm publication is not required for Git-based installation.
The web sidebar keeps two compact controls beside the status indicator:
The settings card remains available when automatic recovery is disabled, so a user can still perform a deliberate manual restart or stop.
| Policy | Intended use | Default behavior |
|---|---|---|
| Safe | Short or sensitive tasks | 5-minute handoff, 2 consecutive resumes, smaller scan |
| Balanced | Everyday work | 10-minute handoff, 3 consecutive resumes, moderate backoff |
| Long task | Builds, imports, and long agent sessions | 30-minute handoff, 8 consecutive resumes, wider scan |
| Manual | Fine tuning | Uses the individual controls in the card |
The policy is explicit configuration, not a hidden fallback inside the runtime. When Manual is selected, the individual fields are authoritative. The initial Manual values match the Balanced values and preserve the behavior of an existing configuration.
The continuation engine resumes only machine-interrupted, transiently failed, or token-limited turns. User aborts and policy-blocked turns are not resumed automatically. Before a continuation, the idempotency guard can tell the model to verify a tool whose result is unknown or avoid repeating a tool that already succeeded.
Permanent failures such as authentication, balance, missing-model, and context-limit errors are skipped and can be surfaced through browser notifications. Repeated failures use adaptive backoff and stop at the configured limit. The loop guard can cancel and redirect a turn that repeats identical tool calls or assistant text.
Open Settings → Plugins → DSH Session Resilience. The card uses staged edits: values are validated locally and written only after Save. Reset on an individual field removes its user override and returns to the deployment default.
The card includes:
The model can call restart_dsh when the user asks to restart DSH or the current task explicitly requires it. It can call shutdown_dsh only for an explicit stop request. Both tools accept an empty object. Restart results include the instance, previous process, port, captured root sessions, and helper log paths.
The two tool definitions and their result text use the normal tool-call and tool-result tokens for the active model request. Automatic recovery adds the configured continuation message only after a qualifying interruption.
The plugin adds no fixed system-prompt prefix. Its model-visible contribution is limited to the two tool definitions, their results, and a continuation message when recovery is triggered.
@deepseek-ai/cordis and @deepseek-ai/dsh-tools peer packages.0.1.5-rc.2 and Node 22.19+; package metadata accepts DSH >=0.1.0-rc.7 <0.2.0.The repository commits prebuilt lib artifacts because DSH GitHub installation does not assume a local TypeScript toolchain. Run the complete package check:
npm run pack:check
The verification script checks the standalone package identity, bundle patch, required publication files, and JavaScript syntax. The artifact tests cover client registration, the detached restart helper, and the local status-bridge safeguards. The GitHub workflow repeats the package check for every push and pull request. Version tags matching v* run the release workflow, which checks that the tag matches package.json, creates the package archive, and attaches it to a GitHub Release.
For the DSH Plugin Market submission, use a public repository whose root contains this package.json, cordis.patch.yml, README, license, and prebuilt lib files. Add one data/plugins/hasan-aghayev__dsh-session-resilience.yml entry to the awesome-dsh-plugin catalog and follow its validation pull request. After the catalog entry is merged, the market picks it up automatically. For reproducible installation, use dsh plugin add https://github.com/hasan-aghayev/dsh-session-resilience and pin the selected commit in the profile. A GitHub Release archive is also produced for each version tag.
MIT. This project is an independent community plugin and is not affiliated with DeepSeek AI. It uses the documented DSH plugin extension points and does not modify the DSH agent loop.