Repository navigation
Releases
v0.1.0-preview — Developer Preview
dsh-llm-verifier v0.1.0-preview
This is the first developer-preview release of dsh-llm-verifier, a validation-first Best-of-3/5 workflow for DeepSeek Harness.
The plugin runs 3 or 5 independent coding candidates in detached Git worktrees, rejects patches that fail project validation, and ranks only the passing candidates with llm-verifier. The original checkout remains unchanged until a separate approval applies the selected patch and reruns validation.
Highlights
- Best-of-3 and Best-of-5 candidate orchestration.
- Detached Git worktrees for candidate edits.
- Deterministic project validation before model-based ranking.
- Failed candidates excluded from ranking.
- Separate approval before applying the winner.
- Repository identity, base
HEAD, clean state, and patch SHA-256 checked again before apply. - Auditable
report.md,winner.patch, and apply-result artifacts. - Credential redaction and process-group cleanup controls.
- Automatic validation detection for supported JavaScript, Python, Rust, Go, and Make-based projects.
Requirements
- DeepSeek Harness
0.1.0-rc.7 - Node.js
24.x - pnpm
11.7.0 - Python
>=3.9,<3.14, managed withuv llm-verifier==0.2.0- macOS or Linux
Install from source
git clone --branch v0.1.0-preview --depth 1 https://github.com/Web0926/dsh-llm-verifier.git
cd dsh-llm-verifier
pnpm install --frozen-lockfile
uv sync --frozen --project python
pnpm run check
dsh plugin --profile web add "$(pwd)"
dsh plugin --profile web listSee the README for the complete workflow.
Important security boundary
This release is for trusted repositories only. Candidate edits are separated with Git worktrees and candidate Harness processes use workspace-write, but validation commands still execute target-repository code on the host. Git worktrees are not a container or security sandbox.
Review every approval prompt, especially the validation commands shown before execution. Real candidate and verifier runs use paid model requests.
See SECURITY.md for the full boundary and reporting process.
Known limitations
- Source installation only; no npm package or prebuilt binary artifact.
- No Windows support.
- Clean, normal Git repository roots only; no dirty worktrees, submodules, sparse checkouts, or linked worktrees.
- Candidate count is fixed to 3 or 5.
- No automatic commit, push, merge, stash, reset, or deployment.
- No container isolation for host-executed validation commands.
- No OpenAI, Vertex, vLLM, or other verifier backends.
- No custom Web UI, ProgressTracker, or early stopping.
Release validation
Validated from the exact main tree at ce06d928d495b865920aa0b24907a3b8d3ead669 with Node.js 24.19.0, pnpm 11.7.0, and the locked Python environment:
- TypeScript typecheck passed.
- 22/22 automated tests passed without live model requests.
- TypeScript build passed.
- Python bridge byte-compilation passed.
- The synthetic retry demo reproduced its expected 2-pass/2-fail baseline and reached 4/4 passing tests after the intended one-line repair.
Feedback wanted
Please share installation friction and sanitized Best-of-3/5 outcomes in the structured feedback issue. Do not post credentials, private source code, raw prompts, complete diffs, absolute local paths, or unredacted reports.
Useful feedback includes cases where multiple candidates pass deterministic tests but the verifier selects a clearly weaker patch.