Evidence Arena
shengshifantang/dsh-evidence-arena · skills
Evidence Arena: parallel isolated DSH coding runs with deterministic judging and guarded promotion — standalone workbench plugin for DeepSeek Harness.
★ 2GitHub stars
1Forks
2026-09-02Last updated
TypeScriptLanguage
MITLicense
Key features
- SpeedPer-node and whole-run wall time: Network, provider queues, and local gates affect it
- /arena-read returns runs, Setup, on-demand file diffs, candidate-preview state, and the privacy-bounded portable report over a trusted-host channel.
- repeat runs and compare distributions
- TokensProvider-reported input, output, reasoning, cache, and total tokens: Missing telemetry is not guessed
- BehaviorModel-call count, tool-call count, and recent activity: Exposes differences such as direct editing versus evidence-first discovery
Requirements
- The standalone repository passes both TypeScript faces, 18 test files / 83 tests, Host ESM build, browser loader build, and package-closure verification on macOS.
- A complete SDK E2E uses real DSH JSON-RPC child processes and the Agent loop: two Builders edit isolated worktrees through Harness tools, then project tests, logic/security Reviewers, Session JSONL and token metering, candidate web preview, and two-phase promotion all run.
Install command
dsh plugin --profile web add github:shengshifantang/dsh-evidence-arena