bstack P12 — Persistent Loop Discipline. Fresh-context restart per iteration. State in the filesystem.

npx skills add broomva/persist

Then in your agent session:

persist iterate PROMPT.md \
 --max-iterations 50 \
 --max-wall-clock 14400 \
 --success-condition "grep:DONE:STATUS"

Each iteration spawns a fresh agent context; state survives in PROMPT.md + git tree. Loop exits when success_condition fires or budget exhausted.

When the conversation context window rots past ~100K tokens, model coherence degrades silently. METR's Time Horizon 1.1 puts the deployable 80%-reliability horizon at ~1h on Opus 4.6. Above that, in-context loops (ReAct/TAO) fail. persist solves this by restarting the context every iteration while keeping state in the filesystem.

• P5 Parallel Agents — N persist loops per worktree
• P7 CI Watcher — each iteration's PR uses p9 watch
• P10 Worktree Hygiene — clean tree before each iteration
• P11 Empirical Feedback — multi-modal success conditions
• P6 Bookkeeping — bookkeeping replay between loops

See SKILL.md for the full primitive contract.

MIT.