dsh-web
zhu1090093659/dsh-webDeepSeek Harness(DSH)Web 插件聚合生态包 · 一切皆插件,创意工坊分发(dsh-market.com)
统一收集本地插件的加载、失败与任务状态,脱敏写入按日轮转的 JSONL 日志。
简体中文 · DeepSeek Harness · Security
An optional, privacy-restricted JSONL diagnostics service for DSH plugins. It helps troubleshoot plugin startup, supported tool outcomes, scheduled-job changes and bounded failure metadata without collecting conversation text or credentials.
pluginDiagnostics.record(...).It does not log user messages, submitted answers, prompts, model responses, tool arguments, HTTP bodies, tokens, secret values, arbitrary stack traces or user identities. Plugins that do not emit diagnostics remain functional.
Requires a compatible DSH host with the peer services declared in package.json. Install into the profile that hosts your plugins (for example, desktop):
dsh plugin --profile desktop add github:Kerberos255/dsh-plugin-diagnostics
Restart the host after changing plugin code. Other supported plugins detect the pluginDiagnostics service when present; there is no separate dashboard or mandatory settings page.
Log directory: DSH_HOME/plugin-diagnostics/logs/<plugin-name>/YYYY-MM-DD.N.jsonl. Files are limited to two 1 MiB parts per plugin per day; entries older than 14 days are periodically removed. Full or unwritable log destinations silently drop diagnostic entries instead of blocking user tasks.
Check whether the plugin and its dependencies loaded before expecting new logs. A previous-failures count is historical, not evidence that every count was a new error. Diagnostics are deliberately sparse; they cannot reconstruct a session, explain arbitrary model behavior, or replace live host/platform tests.
Review SECURITY.md before sharing logs; even allowlisted metadata may reveal operational timing. Run npm test for portable logger tests. Packaging can be checked with npm pack --dry-run; compatibility with a real DSH host must be tested separately.
MIT licensed. See LICENSE.
For dsh-lossless-context only, the diagnostic allowlist includes bounded numeric stage timings, leaf and merged-summary counts, model-call totals, cache hits, and concurrency. It does not record the model output, original conversation, prompts or credentials.