Parfii-bot 3d928b41db feat(skills): /sleep-on-it 6-phase wizard + kei-sleep-queue CRUD + incubation prompt

Priority-scaled time budgets (quick/standard/deep/marathon/weekly),
marathon-mode for hard derivations (skips Phase B REM for one task),
checkpointing every N minutes via partial commits.

2026-04-22 02:30:04 +08:00

1.9 KiB

Raw Blame History

Phase 2 — Task type (click)

Map the free-text task to one of five dispatch categories. The remote agent's Phase A uses this to pick the right tool chain.

2a — Click

Emit ONE AskUserQuestion:

{
  "questions": [
    {
      "question": "How should the nightly agent approach this task?",
      "header": "Type",
      "multiSelect": false,
      "options": [
        {
          "label": "Deep research",
          "description": "WebSearch + WebFetch + synthesis — 3+ searches, 2+ page fetches, structured report"
        },
        {
          "label": "Pipeline design",
          "description": "Architect + critic sequence — 5-7 phases with verify-criteria and tradeoffs"
        },
        {
          "label": "Pattern analysis",
          "description": "Query kei-memory + past reports — extract trends across sessions, propose action"
        },
        {
          "label": "Comparative study",
          "description": "Pros/cons matrix across N options the user lists — weighted recommendation"
        },
        {
          "label": "Custom",
          "description": "Follow the task text verbatim — no dispatch tool, free-form response"
        }
      ]
    }
  ]
}

2b — Normalise

Map the clicked label to a compact token the queue file stores:

Label	Token
Deep research	`deep`
Pipeline design	`pipeline`
Pattern analysis	`pattern`
Comparative study	`compare`
Custom	`custom`

Store as TASK_TYPE.

2c — Soft nudge on mismatch

If TASK_TYPE == "custom" AND TASK_TEXT contains any of should I | compare | trade[- ]off | which is better, print a soft hint:

Your task text looks like a comparison — compare or deep usually produce a stronger result than custom. Proceeding anyway.

Do NOT re-ask. One nudge, user keeps control.

Verify-criterion

TASK_TYPE ∈ {deep, pipeline, pattern, compare, custom}.
Exactly ONE AskUserQuestion in this phase.

1.9 KiB Raw Blame History