Releases: yifanes/agent-harness-rs
Releases · yifanes/agent-harness-rs
v0.2.6
Added
- Added
web_fetchas a built-in read-only tool across mock, local, sandbox, and e2b runtimes. - Added hosted-tool abstraction via
HostedTool. - Added Anthropic native hosted
web_searchsupport throughweb_search_20250305. - Added
AgentLoopHarness::with_hosted_tools(...)andAgentLoopHarness::with_web_search(). - Added silent-stop detection for empty/whitespace-only model turns with no tool calls.
Changed
PlanApprovalnow treatsweb_fetchas read-only.- OpenAI-compatible Chat Completions now fails fast when hosted tools are requested; OpenAI web search is reserved for a future Responses API client.
- Anthropic streaming explicitly ignores provider-hosted web-search blocks.
Fixed
- Prevented fetched/cached model-limit entries from reducing known offline fallback limits.
v0.2.2
feat(skills): progressive-disclosure skill catalogue
新增 skills 模块:向系统提示注入紧凑的 name/description/path 技能目录,而非把每个 SKILL.md 正文塞进 prompt。
- 技能文件位置通过
SkillSourcetrait 抽象(默认LocalSkillSource走本地 FS,可由 sandbox 实现覆盖)。 - 模型按需用已有的
read工具拉取完整指令、用bash运行脚本,无需专门的 execute-skill 工具。 - 导出:
SkillsManager/SkillLoader/SkillPromptRenderer/SkillSource/LocalSkillSource/SkillMetadata/parse_skill_frontmatter。
v0.2.1
借鉴 MiMoCode 工具层机制的三项改造:
新增 / 改进
- 错误感知截断:
bounded_preview/clip_overflow在截断超长工具输出时按 head 70% / tail 30% 保留尾部报错段,不再丢失错误信息;sandbox 与 local 运行时共享同一实现,消除重复的 head-only 逻辑。 - 统一执行包装层
BoundedToolRuntime<R>:在AgentLoopHarness::new中包裹所有运行时——tracing span、幂等的 schema 修复(ToolRuntime::repair_invocation作为单一事实源)、轻量校验(失败给教学式错误)、以及对未自限工具(MCP / 自定义)的兜底输出上限。 - 成功静默 / 失败教学:
edit成功路径不再向模型回传repair字段;非法入参时追加由 schema 合成的Expected shape示例(validate_against_schema/example_for_schema)。
修复
- e2b unary 解码路径中既有的
clippy::never_loop。
测试 182 全绿;--all-features clippy 无 error。