Prioritize

Cross-system recommendation engine. A deterministic collector reads local mirrors, skips when nothing changed, otherwise makes one OpenAI call and writes grounded recommendations. The Pulse page is the primary surface; this skill is for running it on demand and acting on what it says.

Ideal State

• What to work on comes back as a short ranked list, each item naming its next action and the real inputs behind it, with nothing invented.
• The recommendations are current: a stale latest.json is refreshed first, unchanged inputs cost no model call, and source errors are named rather than hidden.
• Acting on a pick goes to the skill that owns it under normal permissions; the engine itself never creates, edits or sends anything.

Intent Routing

Intent	Trigger	File
Run	"run prioritize", "refresh priorities", "what should I work on" (when latest is stale)	Intents/Run.md
Review	"show priorities", "why is X recommended", "let's do the first one", "dismiss that"	Intents/Review.md

File	Read when
Setup.md	first-time install: Config.json, the OpenAI key, the 120-minute schedule

Data & Config

• Config: LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Prioritize/Config.json (lane, window, budget, limits, principal context, category definitions, and routes: which skill owns the action behind each category). Without it every command exits 1 and points at the setup guide: Setup.md.
• State: LIFEOS/MEMORY/STATE/prioritize/ (latest.json, status.json, dismissed.json, history/); override with LIFEOS_PRIORITIZE_DIR.
• Inputs (read-only): LIFEOS/PULSE/state/work-cache.json (optional; skipped with a recorded error when absent), Achilles.ts open-crit --json (rides in the work source), Pulse /api/synapse, KnowledgeQuery.ts, Upgrades.ts list, USER/TELOS/PRINCIPAL_TELOS.md, TELOS.md freshness, LIFEOS_STATE.json, MEMORY/STATE/work-attention.json.
• Pulse endpoints: GET /api/prioritize (page data), POST /api/prioritize/{dismiss,undismiss,run} (dismissal, detached forced run), at PULSE_URL (default http://localhost:31337).
• Schedule: a job that runs Prioritize.ts run every 120 minutes, set up with whatever scheduler your install uses; the model call goes through Inference.ts with an OpenAI key (Setup.md).

Gotchas

• Synapse is read through Pulse /api/synapse; when Pulse is down the capture source errors and the run continues on the rest (recorded in errors).
• KnowledgeQuery.ts --created-after includes _archive notes; the loader filters them out.
• PRINCIPAL_TELOS goal text is truncated with "..."; goal ids (G0…) match WorkSweep's goal:G# labels.

Examples

User: "what should I be working on?"
→ Review: read latest.json; if older than the job cadence, Run first
→ Answer with the focus summary and top three, each with its next action and evidence

User: "do the second one"
→ Review: route the recommendation's next action to the skill Config.json `routes` names for its category (a work issue, an Upgrade record, a Telos edit, a draft), under the normal permission rules