HAR: open harness — isolate, verify, hand off proof (not narration)

os-factory/har (HAR; npm @osfactory/har, docs at harproject.dev; v1.14.2 Sep 14) is the open, agent-agnostic harness that turns “trust the agent’s self-report” into isolate → build → verify → hand off evidence. It ships as CLI and MCP server plus a local Mission Control dashboard. Works with Claude Code, Cursor, Codex, or any MCP agent against the same machine-readable .har/ contract — stack, scripts, and checks in one place instead of drifting across README / CLAUDE.md / Cursor rules / CI yaml.

The loop is explicit. har onboard scaffolds the contract; har env launch N gives slot N its own worktree, branch, ports, and (when needed) database so a fleet can run without colliding on shared git state or ports; har env verify N --full runs the project’s real checks and records what passed, including a validated tree hash so a pass is tied to the exact code that was checked. Hand-off keeps the branch and the proof artifacts after teardown — reviewers get code plus evidence, not a transcript of confidence. har env maintain diffs the installed harness against current templates to catch silent drift before an overnight run fails for a confusing reason. Plugins register framework-specific verification stages (e.g. Playwright). Honest framing: HAR coordinates work around the model; it does not claim to replace human judgment on high-blast-radius merges.

Why it matters for code review

Always-on factories need a portable verify contract that survives swapping Claude for Codex tomorrow. HAR relocates the review surface to deterministic gates and tree-bound evidence — the morning inbox is branches that already paid the verification tax, not a pile of unverified agent narration. Continuous modest throughput compounds past peak-session heroics only when every slot leaves proof a human (or a merge policy) can audit without re-running the night by hand.