feat(subagent): 实验功能设置新增子代理模型配置

- 子代理设置页支持选择次主力模型，仅保存模型引用（不再写补丁字段），
  子代理直接继承所选模型的上下文长度与思考模式
- 界面只读展示继承的上下文长度与思考模式，移除补丁编辑项
- 仪表盘统计识别子代理请求（__secondary__）：归为「子代理模型（估算）」
  并缓存次主力模型定价；自定义网关模型价格按官方同名模型跨 provider 匹配
- bump v0.7.1

src/lib/subagent-settings.ts

@@ -0,0 +1,155 @@
+import type { ExperimentalEnvStatus, SecondaryModelConfig } from "../types";
+
+/**
+ * Helpers for the `[experimental]` and `[secondary_model]` sections of the
+ * Kimi Code config.toml. Both live in `config.raw_other` (the pass-through
+ * bucket for unknown top-level TOML sections), so these helpers read/write
+ * that JSON object directly.
+ */
+
+function asRecord(value: unknown): Record<string, unknown> {
+  if (value && typeof value === "object" && !Array.isArray(value)) {
+    return value as Record<string, unknown>;
+  }
+  return {};
+}
+
+function isTruthyEnv(value: string | undefined): boolean {
+  if (!value) return false;
+  return ["1", "true", "yes", "on"].includes(value.trim().toLowerCase());
+}
+
+// ---------------------------------------------------------------------------
+// [experimental] — experimental feature flags
+// ---------------------------------------------------------------------------
+
+export interface ExperimentalFlagDef {
+  /** Flag id as stored in config.toml `[experimental]` (e.g. "secondary-model"). */
+  id: string;
+  /** Single-feature env var that can force the flag (locks the UI toggle). */
+  envVar: string;
+}
+
+/** Known experimental flags (Kimi Code v1 + v2 registries, merged). */
+export const EXPERIMENTAL_FLAGS: ExperimentalFlagDef[] = [
+  { id: "secondary-model", envVar: "KIMI_CODE_EXPERIMENTAL_SECONDARY_MODEL" },
+  { id: "tool-select", envVar: "KIMI_CODE_EXPERIMENTAL_TOOL_SELECT" },
+  { id: "acp-v2", envVar: "KIMI_CODE_EXPERIMENTAL_ACP_V2" },
+  {
+    id: "persistence_minidb_readmodel",
+    envVar: "KIMI_CODE_EXPERIMENTAL_PERSISTENCE_MINIDB_READMODEL",
+  },
+];
+
+export const EXPERIMENTAL_MASTER_ENV = "KIMI_CODE_EXPERIMENTAL_FLAG";
+
+/** Whether the master env var is set to a truthy value (locks every flag on). */
+export function isMasterEnvOn(env: ExperimentalEnvStatus | null): boolean {
+  return isTruthyEnv(env?.[EXPERIMENTAL_MASTER_ENV]);
+}
+
+/** Whether a flag's value is overridden by its env var (set & non-empty),
+ *  locking the config toggle. Kimi Code lets the env var force the flag on
+ *  *or* off, so a set env var locks regardless of its truthiness. */
+export function isFlagLockedByEnv(
+  env: ExperimentalEnvStatus | null,
+  flag: ExperimentalFlagDef
+): boolean {
+  return env !== null && flag.envVar in env;
+}
+
+/** The value a locked flag is forced to by its env var. */
+export function forcedEnvValue(
+  env: ExperimentalEnvStatus | null,
+  flag: ExperimentalFlagDef
+): boolean {
+  return isTruthyEnv(env?.[flag.envVar]);
+}
+
+/** Read `[experimental]` from raw_other: `{ flagId: boolean }`. */
+export function getExperimentalFlags(rawOther: unknown): Record<string, boolean> {
+  const section = asRecord(rawOther)["experimental"];
+  const flags = asRecord(section);
+  const out: Record<string, boolean> = {};
+  for (const [k, v] of Object.entries(flags)) {
+    out[k] = v === true;
+  }
+  return out;
+}
+
+/** Set one flag in `[experimental]`. Removing all flags drops the section. */
+export function setExperimentalFlag(
+  rawOther: unknown,
+  id: string,
+  enabled: boolean
+): unknown {
+  const root = { ...asRecord(rawOther) };
+  const flags = { ...getExperimentalFlags(rawOther) };
+  if (enabled) {
+    flags[id] = true;
+  } else {
+    delete flags[id];
+  }
+  if (Object.keys(flags).length > 0) {
+    root.experimental = flags;
+  } else {
+    delete root.experimental;
+  }
+  return root;
+}
+
+// ---------------------------------------------------------------------------
+// [secondary_model] — subagent secondary model
+// ---------------------------------------------------------------------------
+
+/** Read `[secondary_model]` from raw_other; `{}` when absent. */
+export function getSecondaryModel(rawOther: unknown): SecondaryModelConfig {
+  const section = asRecord(rawOther)["secondary_model"];
+  const cfg = asRecord(section);
+  const out: SecondaryModelConfig = {};
+  if (typeof cfg.model === "string" && cfg.model !== "") out.model = cfg.model;
+  if (typeof cfg.default_effort === "string" && cfg.default_effort !== "")
+    out.default_effort = cfg.default_effort;
+  if (typeof cfg.max_output_size === "number" && cfg.max_output_size > 0)
+    out.max_output_size = cfg.max_output_size;
+  if (typeof cfg.max_context_size === "number" && cfg.max_context_size > 0)
+    out.max_context_size = cfg.max_context_size;
+  return out;
+}
+
+/**
+ * Write `[secondary_model]` with only the `model` field — no patch fields.
+ * Subagents then bind the pointed model entry directly and inherit its
+ * settings (context size, capabilities, effort fallback), which also keeps
+ * usage records keyed on the real model alias instead of the synthesized
+ * `__secondary__` derived entry. Unknown fields already present in the
+ * section are preserved; the known patch fields (`default_effort`,
+ * `max_output_size`, `max_context_size`) are dropped so stale patches are
+ * cleaned up when the model is re-picked.
+ */
+export function setSecondaryModelOnly(
+  rawOther: unknown,
+  alias: string
+): unknown {
+  const root = { ...asRecord(rawOther) };
+  const section = asRecord(asRecord(rawOther)["secondary_model"]);
+  const next: Record<string, unknown> = { model: alias };
+  const knownFields: ReadonlySet<string> = new Set([
+    "model",
+    "default_effort",
+    "max_output_size",
+    "max_context_size",
+  ]);
+  for (const [k, v] of Object.entries(section)) {
+    if (!knownFields.has(k)) next[k] = v;
+  }
+  root.secondary_model = next;
+  return root;
+}
+
+/** Remove `[secondary_model]` entirely (the "not configured" option). */
+export function clearSecondaryModel(rawOther: unknown): unknown {
+  const root = { ...asRecord(rawOther) };
+  delete root.secondary_model;
+  return root;
+}