Tenet: 12+ hour runs with independent critics, not self-certified done

JeiKeiLim’s Tenet (repo Apr 8; npm @jeikeilim/tenet, still pushing into Sep) is the harness for agents that otherwise lose context, drift off-spec, and skip tests across multi-hour sessions. Loop shape: interview → formal spec + quality contract → visuals → DAG decomposition → per-job execute with commits → independent critics → fold learnings into project doctrine. model_tier (frontier | local) reshapes the DAG — local models get finer jobs with explicit acceptance criteria; frontier gets coarser goal-oriented ones — and every worker gets the spec inlined so nobody explores .tenet/ blind.

The review-relevant move is the critic battery after each job: code critic and test critic (grounded, spec inlined) plus interaction-e2e (ungrounded by default — reviews like a user, can catch what the spec missed), plus custom repo critics. Fresh context, no access to the author’s reasoning. The test critic explicitly hunts the oracle problem — AI tests that ratify implementation rather than intent. Failures retry with failure context; crash recovery uses server-ID + heartbeat stall detection for 12+ hour autonomy across Claude Code / OpenCode / Codex. Doctrine in .tenet/project/ is not write-once: jobs flag drift, runs consolidate proposals, an authorized job applies accepted ones so overnight truth doesn’t silently rot.

Why it matters for code review

Continuous factories compound only when each overnight unit pays an independent verification tax before the next job starts — not when the same session self-certifies. Tenet’s grounded/ungrounded critic mix and doctrine self-heal are the merge criteria; the coding agent is the worker behind them. Idle FOMO is only productive after those gates already refuse to call drift “done.”