Repository navigation
v0.17.7
Highlights
- Prompt truthfulness.
ReadGroup,MemoryEditandWriteGoalNoteare now enabled for the main agent — the shipped prompt and the goal-mode injection named them while no profile enabled them. The LSP write semantics, skill loading, mermaid rendering and the user-preferences block now match the implementation, and the CONTRACT cost clause is tightened. - Capability-aware prompt gating. Lead-only sections (delegation table,
ReadGroup, WolfPack, the send-half of subagent collaboration, verify delegation) render only for the main agent; spawn-capable sections follow the agent's real spawn list. A non-spawning subagent prompt drops roughly 15k characters per request. The main agent prompt renders unchanged. - Read-only guard for the LSP tool.
renamewithapply: trueis refused underread-onlycapability mode instead of only being documented as read-only; the preview form stays available. - Consistency guard. A new test checks the tool registry against profile tool lists, prompts and runtime injections (with reviewed allowances), so prompt/registry drift fails CI.
- Cleanup. 30 dead tool entries removed from subagent profiles, four unreferenced prompt files deleted, and the four spawning subagents gained
SendSubagentMessage.
Package versions
All packages bumped to 0.17.7.
中文补充
- 提示词求真:3 个"点名却从未启用"的工具完成接线;LSP 写语义、技能加载、mermaid 渲染、用户偏好块均改为如实描述。
- 按能力裁剪:子代理提示词只保留它能执行的段落,非派生子代理每请求约减 15k 字符;主代理渲染不变。
- LSP 只读机械守卫 + 新增 prompt↔工具一致性护栏测试(全 profile 覆盖、带审阅过的豁免清单)。
- 清理 30 条死配置与 4 个零引用提示文件;10 个包版本升至 0.17.7。