ScoreBench Docs¶

Harness is a local-first middleware layer between solving agents and challenge platforms. It owns connector credentials, scoped run tokens, submissions, candidate bundles, raw venue evidence, scores, logs, and comparison reports.

Use these docs when operating the Harness web UI, launching agents, maintaining connectors, or deploying the daemon behind HTTPS.

Web UI Shortcuts¶

• Chart dashboard: interactive strategy comparison charts. This opens the featured VLIW dashboard when available and otherwise falls back to the merged strategy comparison.
• Export Studio: label and focus selected runs, add a narrative, share a composed view, or export HTML, SVG, and PNG snapshots.
• Runs: search personal runs, open isolated dashboards, hide a run from your own comparisons, explicitly publish it, or permanently delete it. Runs are private by default.
• Account: profile, exercise API keys, credentials, sessions, users, and logout.

Start Here¶

• Install the Scorebench Skill: required setup for every Codex or Claude Code worker before a run handoff.
• Harness Guide: end-to-end operator workflow for the web UI, CLI authentication, credential profiles, run tokens, dashboards, logs, and security boundaries.
• Connector Reference: credential shapes, config keys, submission behavior, refresh support, solution fetch support, and raw result payloads for each connector.
• Middleware Protocol: HTTP schema, trust model, and exact agent submission contract.
• Architecture: data model, service boundaries, and report pipeline.
• Export Studio: focused run comparisons and shareable report snapshots.
• API Cost Accounting: hardcoded model prices, cost-axis semantics, estimate markers, and the runner usage contract.
• Public Deployment: Nginx, HTTPS, and systemd deployment notes.