zeng6125-rgb/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 0.1.12 — auto-continue failed on session format v4: message sources must now be producer-owned, so the continuation carries source.kind = "plugin:dsh-llm-retry" instead of the retired {kind:"plugin"} wrapper; the "continuation landed" line in host.log is detected again.
• New in 0.1.11 — compatible with the new kernel 0.1.7-alpha.1 (the settings API moved to SettingsForms: volatile form fields, live config sync via loader/volatile-update, the card reads/writes through remote.settings).
• New in 0.1.11 — smoother settings-card scrolling; bug fixes.
• 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.