arvindrk - Load vision.md, categories.json, current horizon (Ruflo "harness:horizon" or agent/harness/horizon.json), feature_list, PROGRESS.
- Use memory_search for past decisions on vision/horizon/balance.
- If needed, propose horizon update (use agent_spawn for sub-agents e.g. vision-align, category-balance, lookahead-decompose when complex):
- Map tasks to vision goals.
- Score proposed next 3-5 for category distribution (weights + no excessive repeats across slice/recent).
- Respect depends_on and unblocked.
- Persist horizon slice + mix summary to Ruflo memory (and optionally file).
- Return the immediate next item + rationale (vision link, category contribution).
- Record any planning notes to PROGRESS (or let main flow do it).
The main orchestrator prompt sequences this before execution. Sub-agents (via Ruflo spawn) can be used for individual steps above when the lookahead is complex.
This skill + vision/categories/horizon make the planner extensible, scalable (sub-agents + memory), and decoupled from repo business logic. Harness core never embeds domain prefs.