dsh-better-retry
DSH (DeepSeek Harness) Cordis 插件：对任何模型请求失败（任意错误代码）使用指数退避进行重试——可通过 Settings → General 中的滑块调整重试次数上限（0–64，默认 8）。
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:Yaing-Yan/dsh-better-retry#6a799389737d519925b653db02e920610a041d7dBetter Retry 以 dsh-better-retry 发布，当前版本为 1.2.0。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。
A DeepSeek Harness (DSH) Cordis plugin that retries any model-request failure — whatever the error code — with durable, backoff-scheduled retries, plus Settings → General sliders to tune the retry budget (0–64, default 8) and the 429 wait (5–120 s, default 15 s) live. 一个 DSH（DeepSeek Harness）Cordis 插件：无论报什么错都自动重试模型请求（持久化 + 指数退避），并在 设置 → 常规 里提供滑块实时调整重试次数（0–64，默认 8）与 429 等待时长（5–120 秒，默认 15 秒）。
The stock dsh-llm-retry plugin only retries failures whose code appears in the provider route's retryPolicy.retryableCodes (default: EMPTY_RESPONSE, RATE_LIMIT, SERVER, TIMEOUT, TRANSPORT). Everything else — AUTH, QUOTA, INVALID_REQUEST, PI_AI_ERROR, LLM_STREAM_IDLE_TIMEOUT, … — ended the turn immediately with "本轮运行失败".
内置的 dsh-llm-retry 只重试 provider 策略里列出的错误码（默认：空响应、限流、5xx、超时、传输错误）。其它错误（鉴权、配额、无效请求、流空闲超时……）一遇即终，直接「本轮运行失败」。本插件把这张白名单变成"全量重试"。
llm/retry / llm/retry-started events as the stock plugin, so the chat UI shows "正在重试模型请求（n/N）· Xs" and sessions stay replay-safe.settings.yaml, apply to the next failure with no restart.ABORTED (user cancel) and CONTEXT_WINDOW_EXCEEDED (handled by the compaction plugin) are never retried.| Half | Role |
|---|---|
Host (lib/index.js) | Listens on the agent loop's agent/request-error waterfall. For any failure not in the never-retry set and not already recovered downstream, it appends a durable llm/retry event, waits out exponential backoff (500 ms → 10 s cap, 10% jitter) — or, for a failure carrying Retry-After (429 rate-limit/quota), the server-requested wait clamped into the user's 429 window [5 s, retryAfterMs] — appends llm/retry-started, and returns { kind: "retry" } so the loop re-issues the request. The settings live in the dsh-better-retry namespace (maxRetries default 8 max 64, retryAfterMs default 15 s range 5–120 s), hot-reloaded per change. |
Client (lib/client.js) | Registers two settings.general.item rows: range sliders that read/write the Host half's same-origin route /dsh-better-retry/config (the settings wire API only serves apiproxy-allowlisted namespaces, so the Host serves its own route through the settings service) — every release persists dsh-better-retry.maxRetries / retryAfterMs into settings.yaml immediately. |
model request fails (any code)
│
▼
agent/request-error (waterfall)
│ downstream first: stock retry / compaction decide?
▼
never-retry? ── ABORTED / CONTEXT_WINDOW_EXCEEDED ──► pass through
│
▼
append llm/retry (durable) ──► backoff ──► llm/retry-started ──► loop retries
│ (max: slider value)
▼
budget exhausted ──► original failure surfaces ("本轮运行失败")
cd "$DSH_HOME/profiles/web" # DSH_HOME defaults to ~/.dsh
pnpm add https://github.com/Yaing-Yan/dsh-better-retry/archive/refs/tags/v1.0.0.tar.gz
# then add "dsh-better-retry" to the "dsh.profile.bundles" array in package.json
dsh --profile web # restart the app
That is exactly how dsh-at-file / dsh-better-status / dsh-show-picture are installed. The Host half registers the retry listener; the Client half is served as a web module (/plugins/dsh-better-retry/client.js) and adds the slider row to Settings → General.
The plugin can also be created in a single session with cordis_define (kind new):
code.host ← paste the content of plugin/host.jscode.client ← paste the content of plugin/client.jsthen activate with cordis_run (mode run). First activation of the Client half asks for user approval in the UI. Dynamic plugins are process-local: they do not survive a DSH restart and belong to the session that created them. In dynamic mode the budget is fixed at 8 (the sandbox has no settings service); the slider still persists the value for the static install.
Full step-by-step (including upgrades and rollback) is in docs/INSTALL.md.
| Where | What | Default | Range |
|---|---|---|---|
| Settings → General → 失败重试次数 | dsh-better-retry.maxRetries | 8 | 0–64 |
| Settings → General → 429 等待（秒） | dsh-better-retry.retryAfterMs | 15 s | 5–120 s |
~/.dsh/settings.yaml → dsh-better-retry: → maxRetries / retryAfterMs | same, hand-editable | 8 / 15000 ms | 0–64 / 5000–120000 ms |
0 disables the any-code retry entirely (stock policy still applies).policyKey embeds the budget).Retry-After (429 quota/rate-limit) is honored as the wait, clamped into [5 s, retryAfterMs] — never skipped, so quota errors are retried up to the budget (e.g. 60 s Retry-After with the default 15 s cap waits 15 s per retry).AUTH/QUOTA failure only helps for transient blips — a permanently bad key or empty balance will simply fail 9 times (1 + 8 retries) before surfacing.dsh-compaction-basic (compact, then retry from the replacement surface), not by blind repetition.llm-retry invariants (mode, ascending retry numbers, per-chain retryId).dsh-better-retry/
├── lib/
│ ├── index.js # Host half (any-error retry + settings section) — static install
│ └── client.js # Client half (Settings slider) — served as a web module
├── plugin/
│ ├── host.js # Host half — paste as code.host (dynamic; budget fixed at 8)
│ └── client.js # Client half — paste as code.client (dynamic)
├── docs/
│ └── INSTALL.md # Step-by-step install / update / rollback
├── cordis.patch.yml # Bundle patch: inserts the dsh-better-retry row
├── dsh.plugin.json # Plugin metadata (entry + client platform)
├── package.json
└── LICENSE
MIT © 2026 dsh-better-retry contributors