DSH-Reforge

DEEPSEEK HARNESS · COMMUNITY PLUGIN

Every fix, verified.
Experience, traceable.

Build traceable experience from test-verified Python fixes. Give DeepSeek Harness scoped, source-backed context for later, similar tasks.

0.1.0-alpha.2·Experimental alpha·DSH 0.2.0-rc.2 / 0.2.1-rc.5
dsh plugin --profile web add dsh-reforge@0.1.0-alpha.2
DSH-Reforge — Fix. Verify. Reforge.
EXPERIENCE / SOURCE VERIFIEDSYNTHETIC EXAMPLE
customer-form / submit.py

Keep the draft when saving fails

Independent checks qualify the repair source; applicability and verification records stay with the experience.

Before the fix2 failures
Acceptance2 / 2
Regressions3 / 3
Python + pytest Local structured evidence Learning off by default MIT · Open source

FROM REPAIR TO EXPERIENCE

Give a repair a source
you can return to.

Keep bounded evidence instead of a full conversation. Check what happened, then decide when the experience is applicable.

01 / CAPTURE

Capture what matters

Once authorized, capture bounded task summaries, explicit corrections and limited code snapshots, tied to the repository and revision.

Bounded capture
02 / VERIFY

Verify the repair source

Replay original failures, independent acceptance and regressions in temporary workspaces. Turn completion and model self-reports are not acceptance.

Independent checks
03 / REFORGE

Use it where it applies

Extract a candidate with the selected model, then reverify it independently. Later requests receive scoped context only when conditions and relevance qualify.

Inspect, revise or deactivate

ONE SMALL FIX. A COMPLETE EXAMPLE.

Start with a form
that loses its draft.

Two minutes from a failing form to a verified fix and a stored experience. Then run the same example yourself.

Synthetic data, an explicit patch and real independent tests. This video has no autonomous model coding, automatic extraction or paid calls. Example guide ↗

Keep the draft.
Check the fix with tests.

The example includes the original project, explicit patch, external acceptance and regressions. Its passing results come from real test execution.

  • Original implementation2 failing acceptance tests
  • Fixed independent acceptance2 / 2 passed
  • Regression checks3 / 3 passed
Run the full example
npm run example:customer-form

Clone the repository and install its dependencies first; run from the project root. See the example guide. No key or DSH is needed.

YOUR PROJECT. YOUR DECISION.

Start read-only.
Choose what to enable.

Learning is off by default. The example and website do not need your model key. Understand execution, sharing and costs before enabling learning.

Backups, upgrades and recovery

Authorize execution and sharing

Enable only for trusted Python repositories. Verification executes code; workspace isolation is not a strong security sandbox.

Keep bounded evidence locally

Experience, source records, usage and private snapshots stay local. Enabling model sharing sends bounded repair material to the selected service.

An extraction allowance is not a currency cap

Normal DSH chat may also incur costs. Unknown receipts retain their holds and are not replayed automatically. Upgrades do not reset grants.

A bounded synthetic real-model loop was verified. Both retrieval arms abstained in the later-task pilot, so no effect of using memory was measured. Verified source evidence does not establish future-task benefit. Read the results and limitations

FIX. VERIFY. REFORGE.

Try a small example.
Bring the process to your project.

A complete example. Traceable sources. Permissions you control.