Prompts¶

.openharness/prompts.yaml 用于存放 workspace 级 prompt 配置。

它承载原先 settings.yaml 里的 system_prompt 内容，但现在独立出来，便于把“运行设置”和“提示词文本”分开维护。

完整示例¶

base:
  inline: |-
    You are Open Agent Harness running inside the current workspace.
    Prefer workspace-local configuration and tools.

llm_optimized:
  providers:
    openai:
      inline: |-
        Be concise, tool-oriented, and explicit about assumptions.
  models:
    default:
      inline: |-
        Prefer short, direct tool call arguments.

compose:
  order:
    - base
    - llm_optimized
    - agent
    - actions
    - project_agents_md
    - skills
  include_environment: false

顶层字段¶

字段	必填	说明
`base`	否	Workspace 级基础提示词
`llm_optimized`	否	面向 provider 或模型别名的优化提示词
`compose`	否	静态 system prompt 段的拼装顺序

`base`¶

支持 inline 或 file 二选一：

base:
  inline: |-
    You are Open Agent Harness.

base:
  file: ./.openharness/prompts/base.md

file 路径相对 workspace 根目录解析。

`llm_optimized`¶

可按 provider 或模型别名注入补充提示词：

llm_optimized:
  providers:
    openai:
      inline: Be concise and tool-oriented.
  models:
    default:
      file: ./.openharness/prompts/openai-default.md

规则	说明
优先级	`models` 精确匹配 > `providers`
Provider key	AI SDK provider 标识
Model key	`settings.yaml -> models` 中声明的别名；加载时会解析成具体 `model_ref`

`compose`¶

compose:
  order:
    - base
    - llm_optimized
    - agent
    - agent_switches
    - subagents
    - project_agents_md
    - skills
    - actions
    - environment
  include_environment: true

可用段名：

base
llm_optimized
agent
actions
project_agents_md
skills
agent_switches
subagents
environment

规则	说明
`system_reminder`	不在这里配置，由运行时动态注入
`actions`	当前 agent 无可见 actions 时自动跳过
`project_agents_md`	根目录无 `AGENTS.md` 时自动跳过
`skills`	当前 agent 无可见 skills 时自动跳过
`include_environment`	是否追加运行环境摘要，默认 `false`

Note

为兼容旧配置，运行时仍可读取 settings.yaml 里的旧 system_prompt 字段；但新配置推荐统一使用 .openharness/prompts.yaml。