/** * Kimi K2.x-native Sisyphus prompt — rewritten with 8-block architecture. * * Design principles (derived from kimi.com/blog/kimi-k2-6 + arxiv 2602.02276 §4.4.2): * - K2.x was post-trained with Toggle RL (~25-30% token reduction) and a Generative Reward * Model (GRM) that scores: appropriate level of detail, helpfulness, response readiness, * strict instruction following, intent inference. * - The model already has strong intent inference from RL training. Adding Claude-style * "re-verify everything" gates DOUBLE-TAXES the model: external strictness on top of * RL-learned strictness → self-second-guessing, redundant verification loops, and * over-deliberation on already-resolved requests. * - Key fixes over gpt-5-4.ts: * 1. : suppress re-verbalization for already-decided/confirmed turns * 2. : hard stop conditions alongside aggressive parallelism * 3. Tiered (V1/V2/V3): trivial fixes don't trigger full * lsp+tests+build+QA loop — V3 keeps FULL RIGOR with harsh enforcement language * 4. : verbalization explicitly EXCLUDED from trim mandate * * Architecture (8 blocks, same as gpt-5-4.ts): * 1. - Role + K2.x-specific training hint * 2. - Hard blocks + anti-patterns * 3. - Intent gate + verbalization + re_entry_rule * 4. - Codebase assessment + research + tool rules + exploration_budget * 5. - EXPLORE→PLAN→ROUTE→EXECUTE_OR_SUPERVISE→VERIFY→RETRY→DONE * 6. - Category+skills, 6-section prompt, session continuity, oracle * 7. - Task/todo management (scoped threshold for K2.x) * 8. `; return `${agentIdentity} ${identityBlock} ${constraintsBlock} ${intentBlock} ${exploreBlock} ${executionLoopBlock} ${delegationBlock} ${tasksSection} ${styleBlock}`; } export { categorizeTools };