dsh-llm-retry-settings

dsh-llm-retry-settings

A settings card for the DSH LLM auto-retry engine (@deepseek-ai/dsh-llm-retry). Tune the retry count and backoff from Settings → General; changes take effect immediately. Since 0.1.7 it can also auto-continue a reply that was cut off by the output-token limit; since 0.1.9 both the extra error codes and the continuation prompt are yours to write.

中文说明

Features

• Includes the settings UI (client bundle lib/client.js): a card in Settings → General — no separate UI package needed.
• Overrides maxRetries, initialDelayMs, maxDelayMs, and jitterRatio on the agent/request-error retry policy.
• New in 0.1.10 — a read-only observation panel at the bottom of the card: retried failures, auto-continues, cap hits and skipped rounds, broken down by error code and provider/model, with the live session models (so you can copy the exact model id into an override) and a log-tail viewer.
• New in 0.1.10 — per-provider/model overrides: first matching rule wins, * wildcards, empty numbers inherit the global values.
• New in 0.1.10 — backoff curve with a wait budget, a "continue after retries are exhausted" switch (transient errors only), prompt templates, and a zh/en UI that follows the kernel locale.
• New in 0.1.9 — add your own error codes straight from the card (the 自定义 group): type a code a provider throws that is not in the known list (e.g. MY_PROVIDER_BUSY) and press 添加. Input is upper-cased and validated (A-Z 0-9 _ - .), duplicates and blanks are rejected with an inline hint, and the code reaches retryableCodes unchanged.
• New in 0.1.9 — the auto-continue prompt is editable (continuationPrompt). Leave it empty to use the built-in wording, or write your own line for relays/models that need different phrasing.
• New in 0.1.8 — auto-continue now actually fires. Two host-side timing traps: session/event is dispatched inside Session.append, so queueing the follow-up straight from the listener died on session append cannot reenter while another append is being published; once that was fixed, waking the agent at that moment turned out to be silently dropped by the driver (wakeDriver only latches in maintenance/abort), so the message sat in the queue forever. The plugin now waits for agent.whenIdle() and re-checks (new turn started / you sent a message / session disposed) before delivering.
• New in 0.1.8 — the host half writes a low-frequency decision log to ~/.dsh/logs/dsh-llm-retry-settings/host.log (capped at 256 KB) and the card points at it. See Troubleshooting.
• New in 0.1.7 — error-code list re-audited against the current host build: added PI_AI_NOT_WARMED (adapter warm-up race; a delayed retry usually succeeds) plus three amber "retrying will not help" codes (UNKNOWN_MODEL, UNSUPPORTED_OPTION, REQUEST_EXTENSION). Also clarified TIMEOUT: a stalled SSE stream (stream-idle watchdog) is reported as TIMEOUT — there is no separate code for it, so the host default retry list already covers hangs.
• New in 0.1.7 — auto-continue on output truncation (autoContinue, off by default). Hitting the output-token ceiling is not a request failure — the call returns successfully with finish = max-tokens — so no retry policy can ever cover it. When enabled, the plugin watches turn/end and queues one follow-up continuation turn per truncation, at most maxContinuations times in a row (the counter resets when the model finishes normally or you send a new message).
• New in 0.1.7 — the error-code chips are grouped into six categories ordered by "will retrying help": transient → rate limit & quota → request & parameters → content & capability → credentials → abort & fallback. Each group shows how many of its codes you selected; picked codes still float to the front within their own group.
• New in 0.1.5 — selected error codes float to the front of the chip list, with the unselected ones after them; the order inside each group stays fixed, so chips never jump around when you toggle them. (Since 0.1.7 the floating happens within each category group.)
• New in 0.1.3 — configurable retryableCodes: extra failure codes to retry on, merged into each provider's own list (never replaces it). Defaults to INVALID_REQUEST + PI_AI_ERROR, so OpenAI-style HTTP 400 errors (thinking-mode reasoning_text) and generic stream failures get retried out of the box.
• Default enabled: false = fully bypassed; nothing changes until you enable the override.

Install

Prerequisite: a DSH Desktop profile (the web profile lives at ~/.dsh/profiles/web).

Option A — GitHub Release package (recommended)

# 1. download the packaged plugin tgz from the v0.1.10 release
gh release download v0.1.10 -R zeng6125-rgb/dsh-llm-retry-settings

# 2. unpack it into the profile's node_modules
mkdir -p ~/.dsh/profiles/web/node_modules
tar -xzf dsh-llm-retry-settings-0.1.10.tgz -C ~/.dsh/profiles/web/node_modules/
mv ~/.dsh/profiles/web/node_modules/package \
   ~/.dsh/profiles/web/node_modules/dsh-llm-retry-settings

# 3. register the bundle in the profile, then restart DSH
#    add "dsh-llm-retry-settings" to dsh.profile.bundles
#    in ~/.dsh/profiles/web/package.json

Option B — dsh CLI / pnpm (requires git + network)

The dsh plugin command forwards its arguments to pnpm in the profile directory:

# from a git repo (pnpm clones it; the committed lib/ means no build needed)
dsh plugin --profile web add github:zeng6125-rgb/dsh-llm-retry-settings

# or from the release tarball URL
dsh plugin --profile web add https://github.com/zeng6125-rgb/dsh-llm-retry-settings/releases/download/v0.1.10/dsh-llm-retry-settings-0.1.10.tgz

Then enable the plugin in the profile: add "dsh-llm-retry-settings" to dsh.profile.bundles (or use the Desktop plugin-inventory UI) and restart DSH.

Note: the command is dsh plugin (a subcommand of the dsh CLI), not dsh-plugin. The dsh CLI is bundled with the Desktop app; if it is not on your PATH, invoke it via the app's node_modules bin, e.g. node "<app>/node_modules/@deepseek-ai/dsh/lib/bin.js" plugin --profile web ....

Option C — from source

git clone https://github.com/zeng6125-rgb/dsh-llm-retry-settings.git
cd dsh-llm-retry-settings
npm install
npm run build        # node scripts/build.mjs → lib/index.js + lib/client.js (no DSH checkout needed)

Link the local build into the profile and register the bundle:

dsh plugin --profile web link "$PWD"
# then add "dsh-llm-retry-settings" to dsh.profile.bundles and restart DSH

Usage

1. Open DSH Settings → General.
2. Find the LLM 自动重试 card.
3. Toggle 开启覆盖 (enabled) to apply the override.
4. Set maxRetries / initialDelayMs / maxDelayMs / jitterRatio, pick extra retryable codes (chips) — or type a code the list does not know into the 自定义 group — and click 保存.
5. Optionally toggle 输出截断自动续写 (autoContinue) and set its maxContinuations cap — independent of the retry override above. The line it sends can be rewritten in 续写提示词; leave it empty for the built-in wording.

Changes are written to the dsh-llm-retry settings namespace and picked up live by the retry engine.

Troubleshooting

A truncated reply did not get continued. The host half logs every decision it makes to ~/.dsh/logs/dsh-llm-retry-settings/host.log — one line per event, capped at 256 KB. ctx.logger output is not persisted anywhere in desktop builds, so this file is the only way to see what happened. Read the last lines:

• no activate v… line at boot → the plugin is not loaded (check dsh.profile.bundles in the profile).
• activate v… names the running build — if it is not the version you installed, the agent was not restarted.
• settings registered … autoContinue=false → the switch is off.
• a turn/end … line with nothing after it → autoContinue was off at that moment, or maxContinuations is 0.
• bail … → the reason is in the line (no agent instance for this session, followup unavailable, …).
• 放弃投递 … → the agent was still busy when the follow-up was ready, and something changed in the meantime (a new turn started, you typed a message, the session was disposed), so the stale continuation was dropped on purpose.
• 续写已投递 … chain=N/M → the continuation was queued. 连续续写触顶 means the maxContinuations cap is hit; the chain resets as soon as the model finishes a turn normally or you send a message.

Host-half changes (lib/index.js) need a DSH restart. The client card reloads with the page; the agent does not.

Configuration

How it works

The host half registers the dsh-llm-retry settings namespace (schema validation + persistence + live sync) and prepends an agent/request-error listener that rewrites retryPolicy before the official @deepseek-ai/dsh-llm-retry recover runs:

agent/request-error  →  [this plugin: override count/backoff]  →  dsh-llm-retry recover

The auto-continue half listens to the session log instead, because a truncated reply is a successful request:

adapter finish="max-tokens"  →  agent-loop turn/end{reason:"max-tokens"}  →  [this plugin]  →  agent.followup(continuationPrompt || built-in wording)

turn-stopping cannot be used for this: its payload carries no end reason, so it cannot tell "the model finished" from "the model was cut off". The continuation is sent as a plugin-sourced user message, so the chat renders it as an injected-context row labelled dsh-llm-retry rather than pretending you typed it. Constructor-seeded events (resume / fork / replay) never reach session/event, so reopening an old truncated session does not trigger a continuation.

Which wording works is provider-dependent: some relays do not echo the previous reasoning back on the next request, so the model cannot see where its thinking stopped and may restart the task. If that happens, rewrite continuationPrompt (e.g. ask it to answer directly without re-thinking) — sending more "continue" messages cannot fix it.

UI

Settings → General → LLM 自动重试 card. Edits are draft-based: click 保存 to commit or 放弃 to discard. After saving, the card verifies the write against the settings snapshot and shows 已保存 ✓ / 保存失败 ✗.

The card holds two independent sections: the retry override (count / backoff / jitter plus the grouped error-code chips, with a 自定义 group that accepts free-form codes) and 输出截断自动续写 (its own switch, the maxContinuations cap and the continuationPrompt textarea). A switched-off section is dimmed and its inputs are disabled, but the switch itself stays clickable — turning auto-continue on does not require the retry override to be on.

Requirements

• Host plugin: @deepseek-ai/dsh-llm-retry
• Client runtime: @deepseek-ai/dsh-client-web-react
• DSH settings service

Build

npm run build        # node scripts/build.mjs
npm run typecheck    # tsc --noEmit

License

MIT