JiSeaSuperM/dsh-bestthink-standard

README 摘要

dsh-bestthink-standard English 中文 Best-Thinking + Standard mode — a DeepSeek Harness (DSH) agent preset that enters the measured best-thinking trajectory (the RL-aligned minimal condition) on the very first request, then auto-promotes to the full Standard 25-tool catalog with workspace context restored. Compatible baseline: DeepSeek Harness 0.1.0-rc.5 / commit 47f9438 . License: MIT (ported from xiaobright/dsh-anchored-standard ; agent.cordis.yml derives from the official standard preset — see preset/NOTICE ). Why A frontier reasoning model's behavior is strongly conditioned by the first request's system prompt and visible tool catalog — not by its raw capability. Measured on the Project2 maintenance benchmark (DeepSeek V4 Pro, reasoning effort=max ): Condition Ability : official minimal (1-sentence prompt, 2 tools) 99 / 96 two-phase anchored (minimal first, full catalog after first tool call) 98 / 99 official standard (full 25-tool catalog from the start) 91 The first request commits the session to a trajectory; expanding the tool catalog after the first durable tool call perturbs at most one reasoning block and never flips the mode. So the anchor is free: you get the best-thinkin…

分类