Nightshift v0.1.0: public proof run

Nightshift v0.1.0

Initial public release of Nightshift's AgentForge proof run: a bounded agentic build loop where the coding agent writes code, but the harness owns task selection, gates, evaluation, metrics, commits, and retries.

Highlights

• Reusable Ralph-style build loop in ralph-loop.sh
• Structured feature queue in feature_list.json
• Separate evaluator contract and bridge
• Metrics writer with recorded run data
• Next.js dashboard generated by the loop itself
• Public proof run: 29/30 features passed, 41 evaluated attempts, 9.69 average passing score, 109k tokens, about $0.95, 76 minutes elapsed

What this release is

This is the public proof specimen for the Nightshift pattern: scoped task in, build/gate/evaluate/retry loop, committed output only after passing checks.

What this release is not

This is not the final hosted Nightshift supervisor product. The broader system adds queue supervision, GitHub comment triggers, check-run updates, Handoff review, and operational guardrails.