dsh-remote-retry-llm-plugin
用于 SSH 远程开发的 DSH 插件（exec/read/write 工具），增加 LLM 重试次数——适用于不稳定或受速率限制的远程 LLM 端点。
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:xiazhicheng/dsh-remote-retry-llm-plugin#6482c77c817a60a8e76d9ac04003bda6dfae8147Remote Retry Llm Plugin 以 dsh-remote-retry-llm-plugin 发布，当前版本为 0.1.0。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
A DeepSeek Harness plugin that provides SSH remote development tools (exec, read, write) and boosts LLM retries to 50+ for unstable endpoints.
When your LLM is flaky — frequent 429/RATE_LIMIT, 5xx/SERVER, timeouts, empty responses — the built-in recovery ends the turn after just 5 retries. This plugin installs an additional listener on the agent loop's agent/request-error recovery waterfall that retries 50 times by default, or forever in always mode, stopping only on success, turn cancellation, or plugin disposal.
Written in TypeScript. Host-only — no client/UI bundle. Safe to mount alongside the built-in
@deepseek-ai/dsh-llm-retry.
The user request, verbatim:
因为我连的 LLM 不稳定，经常被限制，所以我需要你给 dsh 增加一个功能就是重试支持多次，目前默认的是就 5 次，我希望改成更多次，比如 50 次，一直不断重试，一直到 LLM 能正常返回。
Translation: "The LLM I connect to is unstable and frequently gets rate-limited. Add a feature to DSH that retries many times — the default is only 5, I want more, like 50, keep retrying until the LLM returns normally."
This plugin does exactly that, as an installable bundle.
DSH executes provider retry policy on the agent/request-error waterfall — an open-step extension point where each listener receives a failed request's { agent, turn, step, provider, failure, retryPolicy, signal } and returns { kind: 'retry' } to re-run the step, or calls next() to delegate. The built-in @deepseek-ai/dsh-llm-retry already listens there and enforces the provider-owned retryPolicy (normal = 5 retries by default).
This plugin adds a second listener on the same waterfall with its own, more aggressive policy:
always (default) — retry every eligible failure without an attempt limit.normal — retry only retryableCodes up to maxRetries (default 50).Both plugins cooperate by waterfall order: whichever owns a given retry returns { kind: 'retry' }. This plugin keeps its own per-step retry count in memory (keyed by agent object identity + turn + step + provider) and registers no session projection, so it never conflicts with the built-in retry plugin's llmRetry projection. It respects the turn abort signal and disposes cleanly (aborts active waits, drains them), so it never blocks turn quiescence or plugin disposal.
Retries emit llm/retry and llm/retry-started session events so every retry is visible in the UI (matching the built-in retry plugin's format).
This plugin also provides SSH tools that let the agent execute commands, read files, and write files on a remote server — enabling remote development without SSH-ing manually.
| Tool | Description | Parameters |
|---|---|---|
ssh-exec | Execute a shell command on the remote server | command (required), timeoutMs (optional, default 30000) |
ssh-read | Read a file from the remote server | path (required) |
ssh-write | Write content to a file on the remote server | path (required), content (required) |
Agent: ssh-exec { command: "ls -la /home/user/project" }
→ { stdout: "total 48\ndrwxr-xr-x ...", stderr: "", exitCode: 0 }
Agent: ssh-read { path: "/home/user/project/config.yaml" }
→ { content: "host: ...\nport: 22", path: "/home/user/project/config.yaml" }
Agent: ssh-write { path: "/home/user/project/README.md", content: "# My Project" }
→ { success: true, path: "/home/user/project/README.md" }
The SSH connection is established lazily on first tool call and cached for the plugin's lifetime. Connection parameters are configured in the plugin config (see below).
git clone https://github.com/<you>/dsh-remote-retry-llm-plugin.git
cd dsh-remote-retry-llm-plugin
pnpm install # optional: only needed to rebuild lib/ from src/
pnpm build # optional: regenerates lib/ (already committed)
Then in DSH, install the bundle from the absolute package directory:
plugin_manager → install_bundle → target: /absolute/path/to/dsh-remote-retry-llm-plugin
The repo commits lib/, so the bundle loads without a build step and without pnpm build-script approval.
Point plugin_manager install_bundle at the GitHub URL; DSH runs pnpm add and selects the bundle.
The row ships with sensible defaults in cordis.patch.yml. Edit config there (HMR applies changes live in YAML-enabled profiles), or change it in Settings → Plugins:
- id: retry-llm-plugin
name: dsh-remote-retry-llm-plugin
config:
mode: always # 'always' (default, no limit) | 'normal' (honor maxRetries)
maxRetries: 50 # normal-mode budget after the first request (default 50)
retryableCodes: # normal-mode eligible codes (default transient set)
- EMPTY_RESPONSE
- RATE_LIMIT
- SERVER
- TIMEOUT
- TRANSPORT
excludeCodes: # never retried, even in always mode
- AUTH
- MISSING_CREDENTIAL
- NO_ADAPTER
- INVALID_REQUEST
- PROTOCOL
- CONTEXT_OVERFLOW
- IMAGE_OFFLOAD_REQUIRED
- REGISTRATION_DISPOSED
backoff:
initialDelayMs: 500 # first local delay (default 500)
maxDelayMs: 30000 # cap (default 30000; built-in uses 10000)
jitterRatio: 0.2 # symmetric jitter range (default 0.2)
respectProviderRetryAfter: true # honor a valid provider Retry-After within maxDelayMs
# ── SSH remote development ──
ssh:
host: myserver.com # SSH hostname (default localhost)
port: 22 # SSH port (default 22)
username: myuser # SSH username
identityFile: ~/.ssh/id_rsa # SSH private key path
password: '' # password auth (alternative to key)
remoteDir: /home/myuser # default remote working directory
| Option | Default | Notes |
|---|---|---|
mode | always | Retry every eligible failure until success / cancel / dispose |
maxRetries | 50 | Used only in normal mode |
retryableCodes | EMPTY_RESPONSE, RATE_LIMIT, SERVER, TIMEOUT, TRANSPORT | Normal-mode eligible set |
excludeCodes | AUTH, MISSING_CREDENTIAL, NO_ADAPTER, INVALID_REQUEST, PROTOCOL, CONTEXT_OVERFLOW, IMAGE_OFFLOAD_REQUIRED, REGISTRATION_DISPOSED | Permanent failures + codes owned by other recovery policies |
backoff.initialDelayMs | 500 | |
backoff.maxDelayMs | 30000 | Wider than the built-in 10s, for slow-recovering remote endpoints |
backoff.jitterRatio | 0.2 | |
respectProviderRetryAfter | true | |
ssh.host | localhost | SSH hostname |
ssh.port | 22 | SSH port |
ssh.username | '' | SSH username |
ssh.identityFile | '' | SSH private key path |
ssh.password | '' | SSH password (alternative to key) |
ssh.remoteDir | '' | Default remote working directory |
Backoff is bounded exponential with symmetric jitter, matching the built-in policy's formula.
You can also get unlimited retries without this plugin by setting retryPolicy.mode: always on your provider route directly — this is the native DSH knob:
- name: '@deepseek-ai/dsh-llm-deepseek'
config:
apiKeyEnv: DEEPSEEK_API_KEY
retryPolicy:
mode: always
backoff:
initialDelayMs: 1000
maxDelayMs: 30000
jitterRatio: 0.2
- name: '@deepseek-ai/dsh-llm-retry'
Use this plugin when you want the boost regardless of each provider's own policy, or a finite 50-retry budget that's larger than the default 5.
dsh-remote-retry-llm-plugin/
├── cordis.patch.yml # Loader patch: installs the retry-llm-plugin plugin row + defaults
├── icon.svg # Plugin Manager card icon
├── lib/ # Committed build output (loads without a build step)
│ ├── index.js # Host plugin: the agent/request-error recovery listener
│ └── types/
│ └── index.d.ts # Type declarations
├── locale/
│ ├── en.json # Plugin Manager display title + description (English)
│ └── zh.json # (中文)
├── src/
│ └── index.ts # TypeScript source — the plugin (source of truth)
├── tsdown.config.ts # Build config: src/ -> lib/
├── tsconfig.json # Type-check config (pnpm typecheck)
├── package.json # Bundle manifest: dsh.bundle.patch, exports, icon, devDeps
├── README.md # This file (English)
├── README.zh.md # 中文说明
├── LICENSE # Apache-2.0
└── .gitignore
pnpm install
pnpm build # tsdown: src/index.ts -> lib/index.js + lib/types/index.d.ts
pnpm typecheck # tsc --noEmit
lib/ is committed so installations that skip the build still work. Rebuild after editing src/.
Built and tested against DeepSeek Harness 0.1.7-rc.1 (Cordis 4.0.4, schemastery 3.18.4). Runtime dependencies: ssh2 (SSH client), @deepseek-ai/schemastery (config schema), @deepseek-ai/dsh-tools (tool definitions). The @deepseek-ai/* entries in devDependencies are for type-checking and building only.
Apache-2.0 — see LICENSE.