dsh-error-improvement
可配置的用户确认经验，帮助 DeepSeek Harness 避免重复犯错。
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:wbushihenshuai-design/dsh-error-improvement#bb32d50e281fbf6863ce1e99e3b708b866068fb2Error Improvement 以 dsh-error-improvement 发布，当前版本为 0.1.1。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 standalone DeepSeek Harness plugin that turns user-confirmed mistakes into bounded pre-task checks, helping an agent avoid repeating the same error.
It is intentionally independent from EverOS, memory services, databases, browsers, and network APIs. The only persistent state is the normal DSH settings namespace error-improvement.
Tested with DSH 0.1.2-rc.1 (Node 22.19+ or 24+).
confirmed checkbox is enabled.form: instructions), never direct-user provenance.< and > are escaped), count-limited, and character-limited.error-improvement settings section through DSH's settings provider.agent/pre-step waterfall and inspects the complete downstream message batch. Unlike a one-shot injection at the first step only, this ensures prevention rules are visible before every LLM call — including follow-up tool calls that the model makes within the same turn.assist mode injects only lessons whose explicit scope/keywords match the current direct-user message. Latin/digit keywords require whole-token equality (no substring false positives); CJK uses bigram matching. If scope/keywords are blank, it falls back to conservative multi-term matching.strict mode injects all enabled, confirmed, complete lessons up to the configured limits.strict is a stronger prompt policy, not hard enforcement. It never expands model or tool capabilities.
| Field | Default | Meaning |
|---|---|---|
enabled | true | Enables lesson injection. |
mode | assist | assist matches relevant rules; strict injects all eligible rules. |
maxLessons | 5 | Maximum injected rules per turn, clamped to 1–50. |
maxChars | 6000 | Maximum complete lesson block, clamped to 500–50000 characters. |
lessons[] | [] | User-managed lesson records. |
| Field | Default | Meaning |
|---|---|---|
compaction.enabled | true | Enables automatic pre-step pressure compaction and context-overflow recovery. |
compaction.thresholdRatio | 0.8 | Start normal compaction once the active conversation route reaches this fraction of its advertised context window. |
compaction.retainRatio | 0.16 | Recent conversation fraction retained verbatim; it must be lower than thresholdRatio. |
compaction.summarizationProvider / summarizationModel | empty | Primary summary route. Leave both empty to use the active conversation route. The pair must be complete or both blank. |
compaction.fallbackSummarizationProvider / fallbackSummarizationModel | empty | Explicit fallback summary route when the primary call fails (for example, exhausted balance or authentication/provider failure). The pair must be complete or both blank. |
compaction.maxTokens | 8192 | Maximum summary output tokens. |
When the primary route is configured but its summary call fails, the plugin tries the explicit fallback route once. If no explicit fallback is set, it retries using the current conversation route when that is different. It never chooses an arbitrary provider/model: a route must already be known to DSH, so recovery stays predictable and reproducible. The upstream DSH engine records the normal durable transaction (compaction/start → summary → checkpoint replacement → compaction/end) and then retries a context-overflow request only after the replacement was committed.
⚠️ Threshold and summarization model context window. If your summarization model has the same advertised context window as your conversation model, the default threshold of
0.8(80 %) can cause a chicken-and-egg problem: the compaction engine fires once the conversation reaches 80 % of the context window, but the summarization call then sends the full conversation to the same model and overflows its limit. Setcompaction.thresholdRatioto0.5or lower so compaction starts well before the conversation fills the window, leaving enough headroom for the summarization request. Alternatively, configure a different summarization model with a larger context window.
A lesson contains:
{
"id": "stable-unique-id",
"title": "Verify the active profile",
"mistake": "Edited a different profile from the one used by Desktop",
"prevention": "Confirm the active profile before editing any profile files",
"scope": "DSH Desktop profile changes",
"keywords": "desktop profile package.json",
"confirmed": true,
"enabled": true
}
Do not edit DSH Desktop security settings (mode, openBrowser, networkExposure, sandbox presets). Install through the normal DSH profile plugin mechanism, then restart DSH Desktop once.
Add this to your DSH Desktop profile's package.json (the file at C:\Users\<you>\.dsh\profiles\desktop\package.json):
{
"dependencies": {
"dsh-error-improvement": "https://github.com/wbushihenshuai-design/dsh-error-improvement"
}
}
Also add "dsh-error-improvement" to the dsh.profile.bundles array in the same file.
Then run in the profile directory:
pnpm install
Restart DSH Desktop. The settings section appears under Settings → 错误改进.
{
"dependencies": {
"dsh-error-improvement": "link:D:/work/DS/dsh-error-improvement"
}
}
Preserve all existing bundles and dependencies when merging. The plugin's cordis.patch.yml mounts exactly one Host row; its dsh.client metadata loads the Settings UI.
npm run ci in this repository.package.json, lockfile, and cordis.patch.yml.Requires Node.js 22.19+ or 24+.
npm install
npm run ci
npm run pack:check
The CI command performs formatting/lint checks, TypeScript checks, 24 Host/client/matching/integration tests, production build, safety preflight scan (all published JS files), and a clean-consumer pack smoke test that installs the tarball into a fresh node_modules and imports the real artifact.
src/index.ts — Host registration and pre-step middleware.src/lessons.ts — schema, matching, sanitization, atomic rendering, message provenance.src/client.js — immediately loaded Settings UI with conflict-aware save controller.test/ — Host, matching, safety, client controller, and integration tests.scripts/preflight.mjs — scans all published JS for forbidden coupling.scripts/pack-smoke.mjs — clean-consumer tarball install + import verification.No telemetry and no network requests. Lessons remain in the user's DSH settings document and are sent only as part of the local agent context when selected.
MIT © 2026 wbushihenshuai-design