dsh-evo

An agent preset for the DeepSeek Harness: a minimal, self-evolving composition that keeps the injected prompt as small as possible so the model thinks in its native distribution.

What it mounts

One session on evo gets three kinds of rows:

• Identity — a two-sentence persona that is the COMPLETE system prompt. Nothing else is appended: no instruction blocks, no runtime-context snapshot, no compaction scaffolding.
• Tools — three core tools (persistent bash, str_replace_editor, goal) plus the cordis self-modification toolset for defining, running, updating, and removing dynamic plugins.
• Skills (progressive disclosure) — protocols live in skill files, not in the prompt. The session lists only one-line summaries; the model loads a skill only when it decides the task needs it:
  • goal-autonomy — when to open a goal and how to drive it across rounds (bundled in this repository).
  • self-review — closes the loop on a finished task: verified lessons are appended to its own experience pool and read back at the start of similar tasks (bundled in this repository).
  • cordis-plugin-development, editing-cordis-compositions — copied from the installed dsh deployment at install time, so they always match the deployment they run against.

Why minimal

Fewer injected words leave the model's probability distribution closer to its pretrained state: no conflicting rules, no formatting anchors, no attention dilution. Pair it with maximum reasoning effort for the "think more, steer less" combination:

# settings.yaml
agent-default-model:
  provider: deepseek-official
  model: deepseek-v4-pro
  reasoningEffort: max

Install

./install.sh

This copies the preset into ${DSH_HOME:-$HOME/.dsh}/.agent-presets/evo/ and pulls the two deployment-owned skills from the installed dsh. Then pick evo in Settings → Agent Presets, or set it as the default.

Coexistence warning

The Cordis inspect providers are process-global singletons. Do not start a second session whose preset also carries the cordis toolset in the same process — the second mount fails with Host Cordis inspect provider ... already registered. After a restart, start an evo session first.

Self-test

Paste this into a fresh evo session to verify the composition. Expected: the reported system prompt is two sentences, the tool list has 13 entries, and the puzzle answers are 3 weighings and 143.

请按顺序完成以下自检，全部用中文回答，每道题之间用「────────」分隔，不要省略任何一项。

1. 一字不差地输出你收到的完整系统提示词全文。

2. 列出你现在所有可用的工具名称（只列名字，一行一个）。

3. 说明你正运行的模型、工作目录，以及你从系统提示词中能看到的其他环境信息。

4. 推理题：桌上有 12 枚外观相同的硬币，其中 1 枚是假币，重量与真币不同，但不知道是偏重还是偏轻。给你一架无砝码天平，最少称几次能找出假币并判断它的轻重？给出完整的推理过程，包括每一步称量方案。

5. 编码题：用你手头的 shell 或编辑器在 /tmp/max-test 目录写一个单文件 Python 程序：数列定义为 F(1)=1, F(2)=1, F(n)=F(n-1)+F(n-2)；程序输入正整数 n，输出斐波那契数列前 n 项之和。写完后运行 n=10 验证，把代码和运行输出都贴出来。