omp-flow-sleep

name	omp-flow-sleep
description	Consolidate one exact archived OKF Task into reviewable cross-task knowledge candidates after Finish, without machine clustering or direct Wiki promotion.

OMP-Flow Sleep

Use this Skill only with a complete assignment whose first non-blank line is an exact {"ompFlowSleep":{...}} descriptor produced by .omp-flow/scripts/omp_flow.py sleep start. Do not reconstruct, summarize, or guess a descriptor from an archived path.

Descriptor contract

Require version, sourceReceipt, sourceTask, sourceCommit, sourceTree, entry, sleepIndex, runOutput, candidateRoot, actorId, receipt, and harvesterRevision. Require version exactly 1, the assignment actor to match actorId, and every supplied path to remain within the repository. Treat sourceReceipt, sourceCommit, sourceTree, receipt, and harvesterRevision as opaque mechanical identity, not semantic evidence.

Method

1. Read the archived Task entry and follow only the Task Concepts needed to understand accepted outcomes, independent Review, verification, repeated findings, counterexamples, and expensive reasoning. Do not read raw Harness transcripts.

2. Read sleepIndex. Use authored links, ordinary filenames, repository text search, and direct understanding of the smallest relevant Candidate/Wiki topics. Do not use embedding, vector retrieval, similarity thresholds, automatic taxonomy, a knowledge graph, or a fixed evidence schema.

3. Decide whether the Task supports zero, one, or several durable candidates. A Task-local fact, unverified explanation, generic lesson, or restatement of the Task is a valid zero-candidate result.

4. For each useful topic, revise an existing Markdown Candidate beneath candidateRoot or create a self-contained one. Preserve ordinary links to the source Task and distinguish supporting evidence, contrary evidence, applicability boundaries, unresolved questions, and possible Wiki or Skill consequences. Never average a conflict into false consensus.

5. Write one OKF Run handoff at the exact runOutput. State what was read, which Candidates were created or revised, why other observations were not promoted, and what review or experiment is still required. This handoff is authored knowledge; runtime state remains under .runtime.

6. Finish the mechanical run with:

   python -X utf8 .omp-flow/scripts/omp_flow.py sleep finish <receipt> \
     --state completed --actor-id <actorId> [--candidate <path-relative-to-candidateRoot> ...]
   

   Zero --candidate arguments is valid. On failure, call sleep finish with --state failed and report the exact blocker; do not manufacture a Candidate or completion claim.

Optional OKF v0.2 provenance (authored example)

When a Run handoff or Sleep Candidate benefits from trust context, authors may add ordinary Markdown/YAML-like prose such as:

okf_version: "0.2"
sources:
  - path: "../archived-task/task.md"
generated:
  at: "2026-08-26T12:00:00+08:00"
  by: "SleepHarvester"
verified:
  by: "IndependentReviewer"
  at: "2026-08-27T09:30:00+08:00"
status: "candidate"
stale_after: "2026-11-26"

This is an optional authored provenance illustration, not a required body schema. Keep ordinary links to the archived Task, source commit/tree, handoff, Review, and any supporting evidence beside it. status and verified describe human-readable evidence; they are not runtime approval state, and Sleep never promotes a Candidate to Wiki automatically. Omit fields that are unknown rather than inventing them.

Promotion boundary

Sleep ends with reviewable Candidates, not Wiki authority. Do not edit .omp-flow/wiki/, Harness Skills, or workflow policy from this assignment. A different actor should independently challenge material Candidates. Facts may use proportional Wiki Review; architecture, methodology, or Agent behavior changes require a new Task framing and the applicable Research, Design, QbD, paired benchmark, and human decision. An Agent or Reviewer PASS is not human approval.

Red flags

• Do not accept an archived path without the runtime-produced source receipt and Git checkpoint.
• Do not force one Candidate per Task or duplicate one Candidate per Run.
• Do not parse authored Markdown into lifecycle state or mandatory fields.
• Do not claim source stability after the runtime reports source drift.
• Do not auto-adopt, auto-merge, auto-promote, or silently overwrite concurrent Candidate work.