dsh-reforge: Traceable Fix Experience for DeepSeek Harness

What Does It Do?

DSH-Reforge is a DSH plugin for DeepSeek Harness that accumulates traceable experience from test-verified Python bug fixes and retrieves only the applicable lessons for later similar tasks. It solves the problem of fixes that work once but leave nothing reusable behind: each repair is replayed, verified and distilled into a candidate lesson with provenance, scope and version constraints. The plugin is independently maintained by wang1299 under the MIT license and is currently an experimental alpha that only commits to DSH 0.2.0-rc.2.

Core Features

• Automatic episode capture: once the repository owner enables learning, direct user requests in normal conversation become execution episodes with bounded summaries and restricted Python code snapshots.
• Verified provenance: the original failure, the post-fix independent acceptance run and regression tests are replayed in a temporary workspace, so an end-of-turn check is never treated as proof of correctness.
• Candidate distillation with review: registered DSH models distill candidates carrying source, applicability conditions and version constraints, followed by an independent provenance review.
• Scoped retrieval: later requests select a small number of lessons by repository, code, dependency conditions and relevance, and may return nothing when evidence is thin or inapplicable.
• Inspect and correct: the plugin detail page shows learning activity, lessons, sources, history and usage records, and supports feedback, revision, disabling or restoring.

How to Use This Plugin?

After enabling it in DSH, the repository owner configures independent acceptance, execution trust, model sharing and distillation call limits, then turns learning on, since learning is off by default. You then file Python bug-fix requests as usual; the plugin checks the original failure, the fix and regression tests in a temporary workspace, and distills candidates once materials and authorization are complete. Open the plugin detail page under Plugins on the left, or the Experience button at the top of an existing session, to review lessons, scope and sources and to give feedback or disable them. The first version focuses on Python plus pytest fixes inside a single repository, and local verification runs repository code without a strong sandbox, so enable it only for repositories you trust.

How to Troubleshoot This Plugin?

If learning produces no lessons, first confirm that execution trust, model sharing, independent acceptance and distillation call limits are all explicitly configured, since a missing item can skip learning. Newly created Python files, symlinks, interleaved requests, version drift or unsupported changes may also be skipped, so re-check the configuration against the official installation guide and retry. Unknown model calls caused by network interruptions are not retried or refunded automatically, and the distillation limit is not a spending cap for the whole session, so watch your call volume.