feat(panel): 分布面板新增推理维度与回合调用统计

README.md

@@ -46,7 +46,7 @@
 - **模型定价**：显示当前模型的输入 / 缓存读 / 缓存写单价（从 provider 配置动态读取）
 - **折叠面板**：主标题默认折叠，点击展开；明细、模型、分布各自独立折叠
 - **颜色自适应**：命中率 ≥85% 绿 · ≥70% 橙 · <70% 红，颜色从主题色自动去饱和
-- **Token 分布**：按角色（系统提示 / 用户 / Agent 指令 / Tool 调用 / Tool 结果）展示估算 Token 占比
+- **Token 分布**：按角色（系统提示 / 用户 / 子代理指令 / Tool 调用 / Tool 结果）展示估算 Token 占比
 - **折叠记忆**：折叠状态持久化，重启后保持
 - **语言适配**：支持 中文 / English / 日本語 / 한국어，自动检测系统语言，`/cache-lang` 运行时切换，偏好持久化优先
 - **多币种**：通过 `/cache-currency` 切换货币，费用和节省同步换算
@@ -150,7 +150,7 @@ npm install -g opencode-visual-cache@latest
 
 通过 `/cache-section` 切换后即时生效，无需重启。此外，该命令还可以开关面板的**外边框**——关闭后内容会顶格显示，释放额外空间。
 
-> **关于 Token 分布数值**：分布面板中"总计"为最后一次 API 调用的精确 token 数，"系统提示"/"用户"等分项为字符级 BPE 估算值。分项之和通常小于总计，差值主要来自 OpenCode 运行时注入的系统提示组成部分，包括环境信息、Skill 目录、工具 Schema 定义等（详见 [`system.ts`](https://github.com/anomalyco/opencode/blob/dev/packages/opencode/src/session/system.ts)、[`tools.ts`](https://github.com/anomalyco/opencode/blob/dev/packages/opencode/src/session/tools.ts)）。这些内容不在 agent 配置的 `prompt` 字段中，因此插件无法估算，属于预期行为。
+> **关于 Token 分布数值**：分布面板中"推理"为 API 返回的**精确值**；"系统提示"/"用户"/"子代理指令"/"Tool 调用"/"Tool 结果"为**估算值**——API 仅返回 token 总量，无法拆分各内容类型，插件按内容类型收集文本后基于字符计数近似估算，数值仅供参考。OpenCode 运行时注入的系统提示内容（环境信息、Skill 目录、工具 Schema 定义等，详见 [`system.ts`](https://github.com/anomalyco/opencode/blob/dev/packages/opencode/src/session/system.ts)、[`tools.ts`](https://github.com/anomalyco/opencode/blob/dev/packages/opencode/src/session/tools.ts)）不在此估算范围内。
 
 ### 4.4 余额查询
 

README_EN.md

@@ -46,7 +46,7 @@ Interested in sub-agent monitoring? Check out [opencode-subagent-magazine](https
 - **Model Pricing**: Input / cache-read / cache-write per-million rates (read from provider config dynamically)
 - **Collapsible**: Main title collapsed by default; click to expand. Detail, model, and distribution sections fold independently
 - **Adaptive Colors**: ≥85% green · ≥70% orange · <70% red, auto-desaturated from current theme
-- **Token Distribution**: Per-role (system / user / agent instr / tool call / tool result) estimated token breakdown
+- **Token Distribution**: Per-role (system / user / sub-agent instr / tool call / tool result) estimated token breakdown
 - **Persistent State**: Fold preferences and config remembered across restarts via api.kv
 - **Language**: Chinese / English / 日本語 / 한국어, auto-detects system locale, with `/cache-lang` for runtime switching — user preference takes priority over auto-detection
 - **Multi-currency**: Switch via `/cache-currency` — costs, savings, and per-million rates convert in real time
@@ -150,6 +150,8 @@ Three sub-sections can be toggled independently to save sidebar space:
 
 Toggled via `/cache-section` — takes effect instantly, no restart required. The same command also toggles the panel **border**; turning it off removes the outline and padding so content fills the full width.
 
+> **About Token Dist. values**: "Reasoning" is an exact value from the API; the other rows (system / user / sub-agent instr / tool call / tool result) are **estimates** — the API only reports total token counts, not how they split across content types, so the plugin collects text per content type and approximates via character counting. Values are indicative only. OpenCode runtime-injected system prompt content (environment info, skill catalog, tool schema definitions — see [`system.ts`](https://github.com/anomalyco/opencode/blob/dev/packages/opencode/src/session/system.ts), [`tools.ts`](https://github.com/anomalyco/opencode/blob/dev/packages/opencode/src/session/tools.ts)) is not covered by these estimates.
+
 ### 4.4 Balance Query
 
 The panel can display account balance from multiple AI providers. With **auto-switch** enabled, the balance query follows the model provider of the current session automatically.