VV Harness: sort every rule by enforcement tier — exit codes beat prose
Jul 4 Eftimie: harness = list of things you stopped trusting the model to do; mechanical/prompted/structural/instructional ladder; scope hooks, completion gates, fixture-tested rails.
Source: The VV Harness
Ovidiu Eftimie (4 Jul 2026; plugin at oeftimie/vv-claude-harness) reframes harness design as a trust ledger, not a feature list. An agent harness is a list of things you stopped trusting the model to do. Every rule lives in one of four enforcement tiers, and the tier — not the wording — predicts whether it survives a three-hour session: mechanical (shell hooks with exit codes; no negotiation), prompted (hooks that deliver a message the agent may ignore), structural (file formats like features.json with passes: false — agents respect structure more than prose), and instructional (CLAUDE.md / skills — “most of the time,” decaying as context grows). The history of the harness is promoting rules up that ladder: TDD from polite request to a hook that rejects non-passing completions; scope from prose to a PreToolUse path check; reviewer honesty by granting the reviewer no Edit/Write tools.
What the mechanical and structural tiers actually guarantee: session death and compaction survive via features.json + context_summary.md and a SessionStart reinjection; unfinished work is refused until tests pass (fast smoke under ~15s before the full suite); parallel agents get physical lanes — each feature owns paths, and a hook blocks out-of-scope edits before they land in the tree; metrics route money (rework thrice in-scope → escalate to Opus); always-on CLAUDE.md overhead cut from ~14.7K to ~4.2K tokens. Known-challenges ledger is load-bearing: lead death kills in-process teammates; Agent Teams is experimental and delivery can lie; idle hooks only prompt; the completion gate cannot judge whether tests are worth passing; small work is cheaper sequential. The hidden rung: a mechanical rule is only mechanical if the mechanism is tested — a fail-open hook manufactures confidence worse than no hook. v3.4’s audit found four hooks silently broken; the fix was fixtures, a test suite, and CI.
Why it matters for code review
Always-on factories only beat bursty Claude Code heroics when “done” is an exit code, not a CLAUDE.md hope. Read any harness README by sorting claims into those four tiers and discounting instructional to “usually” — then ask whether the mechanical tier is itself under test. Continuous verified momentum looks like scope hooks and completion gates that cannot be skipped; FOMO that nothing ran overnight is cheaper than waking to green prose that never enforced the boundary it named.