Khaos-Brain: Skill for Codex

KB Sleep Maintenance

Sleep is the sole knowledge-decision and canonical model-generation owner. The scheduled local task is one serialized cycle: Sleep freezes and publishes first; only after an unblocked completed Sleep terminal does the same task run the existing immutable Dream simulation. Dream never publishes authority. Readable cards and the active index are deterministic projections of Sleep's authority.

Authority and entrypoint

Work from the repository root. Read PROJECT_SPEC.md, docs/maintenance_agent_worldview.md, docs/maintenance_runbook.md, and .agents/skills/local-kb-retrieve/MAINTENANCE_PROMPT.md. Current user instructions override repository defaults.

Run:

python .agents/skills/local-kb-retrieve/scripts/kb_sleep.py --json

The native lifecycle implementation owns all mutation, terminal validation, and its immutable run receipt. Do not create a parallel Sleep implementation or a second model writer.

Required behavior

1. Acquire the independent local-cycle task lease and preserve the same run id through closure. The wrapper owns the complete local cycle; do not launch a separate Dream automation. Acquire the sole global KB writer only around mutation phases, delegate its exact token to the child phase, and release it promptly so the independent organization task can continue.
2. Resume the exact open frozen batch before admitting later work. If no batch is open, freeze one finite ordered batch_plan with immutable item identities, input watermark and digest, current-generation identity, prior convergence streak, and tested batch-size bounds. The persistent batch_head binds the exact plan and checkpoint digests. Later arrivals never expand that batch.
3. Give every settled frozen item exactly one verified completed disposition or one explicit blocked disposition with a named owner and executable reopen condition. Persist each item result and the batch checkpoint before continuing so a later Sleep reuses verified work and reprocesses at most the incomplete item.
4. Create or reuse one stable candidate only when a bounded scenario-action-result relation and sufficient evidence exist. Represent it immediately as a LogicGuard model revision with a root Claim, explicit Context and Method, typed support or challenge nodes, and explicit gaps. Never invent Evidence, Warrant, Assumption, Rebuttal, or Limitation merely to fill the model.
5. Keep trusted promotion evidence-dependent. Require current independent validation; weak or duplicated evidence never satisfies promotion.
6. Park unresolved candidates with a machine-evaluable reopen condition and a seven-day decision boundary. Reopen exactly once only after a material qualifying evidence delta.
7. Immediately exclude strong contradictory trusted knowledge and complete its downgrade review in this Sleep pass.
8. Consume each typed Dream model-gap handoff exactly once, record one acknowledgement, and let Dream remain an immutable simulation owner only. Dream runs as the second phase of this task only after Sleep publishes cleanly; blocked or progress-saved Sleep leaves Dream not_run.
9. Assemble exact model revisions into physically separated public, private, and candidate ModelMeshes. Admit a canonical cross-model relation only when qualifying non-AI provenance supports it. Co-use, lexical similarity, and retired related_cards values remain unresolved grounding proposals, never edges.
10. Audit each important model for missing context, action, evidence, warrant, assumption, opposition/rebuttal, and boundary conditions. Give every absence one stable open disposition, required grounded input, and machine-readable reopen condition; never invent the missing content.
11. Stage models, meshes, deterministic readable projections, the exact active index, and generation manifests away from the current generation. Validate the complete staged generation before one atomic pointer switch written last. Ordinary planning, progress_saved, timeout, or pre-activation failure keeps the prior validated generation readable and never treats pending work as current-generation corruption.
12. Exclude rejected, merged, superseded, parked, retired, deprecated, history-only, provenance-incomplete, and contradicted records from retrieval projection.
13. Commit the new watermark only after every frozen item is settled and dispositions, model/mesh publication, lifecycle review, handoff acknowledgements, exact index validation, and the final pointer switch are durable. progress_saved and failure keep the previous committed watermark unchanged.
14. Do not require a human to read files, choose cards, or approve ordinary maintenance decisions. Escalate only a real safety or authority boundary.
15. Stop starting new items at the 660-second native soft deadline, durably seal the checkpoint, release every delegated/global writer token, and return progress_saved before the outer hard timeout. The complete Sleep plus Dream cycle uses a 2400-second native timeout inside a 2700-second owner timeout, with larger aggregate and installer owners; timeout never means success and requires confirmed zero-descendant cleanup. One malformed item may settle as blocked only with a named owner and executable reopen condition; publish completed siblings as completed_with_blocks, never repeat them, and record only the Dream phase as not_run. The independent organization task is never cancelled or rewritten by local failure. The target-owned wrapper owns same-run terminalization; do not invoke a status helper or retry the child.