@zhongjingwei/dsh-model-retry
English | 中文
DeepSeek Harness Web plugin for configuring active providers' model-request retry budget from Settings → General.
Compatibility: pinned to stock DSH 0.1.0-rc.8. rc.8 moved retry ownership into provider profiles, so this version writes each active normal provider's retryPolicy.maxRetries through the published settings API.
Demo
Set the retry budget to 5, show retries 1/5 through 5/5, and recover when the sixth model request succeeds
The demo uses stock DSH 0.1.0-rc.8 and a plugin tarball packed from this source, all under an isolated DSH_HOME. It changes the retry budget from 2 to 5, then uses DSH's deterministic local LLM mock to return retryable server errors on attempts 1–5 and succeed on attempt 6. Session-title generation is disabled so those six mock requests belong only to the demonstrated model request. DSH executes and reports the retries; the plugin owns the shared retry-budget setting.
Install
For stock rc.8:
dsh plugin --profile web add @zhongjingwei/dsh-model-retry@0.1.0-rc.9
Restart Web after installation. Remove the plugin with:
dsh plugin --profile web remove @zhongjingwei/dsh-model-retry
Behavior
- Adds one numeric retry-count row to Settings → General and shows its provider scope.
- Accepts finite, non-negative integers, including
0.
- Discovers active configurable providers through
llm.providers and uses revision-fenced settings.mutate calls.
- Writes
retryPolicy.maxRetries for normal/default policies while preserving existing backoff and retryable-code fields.
- Leaves
always policies unlimited and leaves retry execution/backoff behavior owned by DSH.
Development
pnpm install
pnpm test
pnpm run build
pnpm run verify:package
pnpm pack
The package is an external Cordis plugin and does not patch DSH core.
License
MIT
