javadandCursor
b93ecf1691
Keep source, config, and deploy files. Leave runtime data, uploads, and logs out of git. Co-authored-by: Cursor <cursoragent@cursor.com>
664 B
664 B
Skill Orchestration Contract
The main LLM/Agent sees lightweight manifests first. Load full SKILL.md only after selecting a skill. After every skill call inspect needs_more_evidence and next_skill_candidates.
Recommended hard limits: max_skill_hops=6, max_calls_per_skill=2, max_reasoning_iterations=8, max_evidence_items=50, max_rows_to_llm=100. Never re-run the same skill with the same scope unless new evidence or filters change the input.
Suggested actions for the planner: ANSWER, CALL_SKILL, CALL_TOOL, REQUEST_REQUIRED_DATA, VERIFY, STOP.
The LLM orchestrates; deterministic calculations stay in scripts/database engines.