Add completely (Claude Code plugin harness with a default-FAIL evaluator in the loop's close path)

[23ag1]

What this adds

One entry to Runtimes, Harnesses & Reference Implementations (placed next to Citadel, the other Claude Code harness):

completely — a Claude Code plugin harness focused on the verification half of harness engineering: done is earned, not asserted.

Why it fits this list's bar

• Specific about how agents are constrained, evaluated, resumed, and orchestrated (the contributing criteria):
  • Constrained — deterministic PreToolUse/PostToolUse hooks: per-edit lint/type gate, dangerous-command block, edit-time write-zone fencing per task, commit-before-close gate, and reviewer findings that block closing while unresolved.
  • Evaluated — an independent, read-only, default-FAIL evaluator sits in the close path of the autonomous loop; every acceptance criterion stays FAIL until evidence is re-run (with path-exercised / exercised-as-a-user / code-read dimensions and an adversarial claim-vs-refute mode).
  • Resumed — claims carry worker-id + heartbeat; an interrupted run's orphaned tasks are auto-reopened at next start (the loop's run-report never reports a half-done run as finished).
  • Orchestrated — parallel fresh-context claude -p workers over disjoint write-zones (a hardened take on the Ralph loop already on this list), with a post-batch integration gate that checks the union composes.
• Original implementation, inspectable — the harness design lives in plain bash/python + markdown contracts; behaviour is pinned by a 77-test contract suite the repo runs on itself.
• Link verified, no duplicate entry, single-line diff per the entry format.

Happy to adjust placement or tighten the description if you prefer.

[23ag1]

Added a primary-source write-up that may help evaluate the entry: Done is earned — a field report on putting a default-FAIL evaluator in the close path. It documents four real incidents from the harness's own build log (a crash that shipped with a fully green self-test, the evaluator rejecting its own verification fix twice, a watchdog false-killing healthy workers, and a safety gate that passed its self-test while dead in production) and the evaluator dimension each one forced. Happy to fold it into the PR as a second entry under Evals & Observability if you think it fits — otherwise the single Runtimes entry stands as-is.