I build small, evidence-driven reliability tools for the emerging DeepSeek Harness ecosystem.

我专注于从真实故障和可复现实验出发，为 DeepSeek Harness 打磨小而可靠的社区工具。

Preflight → Classify → Minimize → Verify → Monitor

Read the case study · 阅读中文案例

A review-first Web plugin that probes the request dialect of explicitly configured OpenAI-compatible providers before proposing the smallest supported Harness compatibility change. It verifies the real runtime after an approved write and keeps one-click rollback available.

• Status: public v0.1.0-preview.2; intentionally not promoted to stable
• Best for: custom gateways that reject otherwise valid requests because their accepted request shape differs from Harness defaults
• npm preview · Release · Official community discussion · Public tester guide

A deterministic Web client plugin that explains one high-confidence Windows sandbox spawn EPERM signature in context. It does not rerun commands, change the session, send telemetry, or infer beyond the evidence in one tool-result block.

• Status: stable v0.2.2, intentionally limited to one reproduced signature
• Best for: distinguishing a sandbox boundary from a test assertion failure
• npm · Release · Official community discussion · Field-report gate

An external CLI that finds a 1-minimal set of out-of-tree plugins that still reproduces a broken Harness profile. It works in disposable shadow profiles, preserves the real profile, and produces a redacted machine-readable report.

• Status: stable v0.3.0; v0.3.1 is available on the npm next channel
• Best for: plugin-interaction failures and small, shareable reproductions
• Validation gate: 0/3 independent real-profile reports; no speculative stable promotion
• npm · v0.3.1 prerelease · Official community discussion · Field-report intake

A deterministic post-run verifier for relationships that aggregate evaluation metrics lose: parent/child lineage, admission cardinality, continuation identity, FIFO delivery, settlement provenance, and explicit report relays.

• Status: research v0.1.0-preview.1; the offline verifier works, while the live runner remains an evidence-producing experiment rather than a supported npm path
• Best for: checking durable multi-agent behavior after a Harness run
• Release · Official community discussion · Research evidence

An external compatibility canary that detects Codex App Server protocol drift the pinned Harness adapter cannot safely interpret. The first live experiment found a concrete error-union value that degrades to unknown.

• Status: public v0.1.0; weekly live monitoring is active
• Best for: turning dependency drift into a reproducible report before it is mistaken for a model or user failure
• npm · Release · Official compatibility report · Current canary finding

• Start with reproduced behavior or a controlled experiment.
• Search for maintained solutions before creating another project.
• Prefer a narrow, testable contract over a broad diagnostic guess.
• Keep diagnostics reversible, privacy-conscious, and easy to remove.
• Validate public artifacts from a clean install, not only from source.
• Expand only when new evidence demonstrates that the existing boundary is too narrow.

These are independent, unofficial community projects. Feedback and privacy-reviewed field reports are welcome in each repository.