提示词缓存

def affinity_enabled(deep_agent: Any) -> bool:
    """Return without inspecting model/binding state when affinity is disabled."""
    deep_config = getattr(deep_agent, "deep_config", None)
    kv_config = getattr(deep_config, "kv_cache_affinity_config", None)
    return getattr(kv_config, "enable_kv_cache_affinity", False) is True


def is_sticky_subagent_type(subagent_type: str) -> bool:
    return str(subagent_type or "").strip() in ("browser_agent", "verification_agent")

# 同文件另有 prefetch_sticky_subagent / finish_subagent / evict_subagent
# 三个生命周期钩子（44-128 行），以及由父会话 id + task_id 推导的稳定子会话 id。

// This is intentionally not a `PartialEq` implementation: request equality includes `input` and
// `client_metadata`, while websocket reuse compares the input separately and ignores metadata.
// Keep the destructuring exhaustive so new request fields require an explicit reuse decision.
fn responses_request_properties_match(
    previous: &ResponsesApiRequest,
    current: &ResponsesApiRequest,
) -> bool {
    // ...解构双方全部字段，逐个比对（318-361 行）：
    //   model / instructions / tools / tool_choice / parallel_tool_calls /
    //   reasoning / store / stream / include / service_tier /
    //   prompt_cache_key / text
    // 只有 input 与 client_metadata 被显式忽略。

/**
 * Token accounting for one model call (cache fields are optional).
 *
 * Counts are DISJOINT: `inputTokens` is uncached input only; cached input is
 * reported separately as `cacheReadTokens`/`cacheWriteTokens` (billed input =
 * sum of the three). Adapters whose providers fold cache hits into a total
 * prompt count (DeepSeek's `prompt_tokens`) subtract them out.
 */
export interface TokenUsage {
  inputTokens: number
  outputTokens: number
  cacheReadTokens?: number
  cacheWriteTokens?: number
  reasoningTokens?: number
}

const AnthropicMessagesCompatSchema = Type.Object({
	supportsEagerToolInputStreaming: Type.Optional(Type.Boolean()),
	supportsLongCacheRetention: Type.Optional(Type.Boolean()),
	sendSessionAffinityHeaders: Type.Optional(Type.Boolean()),
	supportsCacheControlOnTools: Type.Optional(Type.Boolean()),
	supportsTemperature: Type.Optional(Type.Boolean()),
	forceAdaptiveThinking: Type.Optional(Type.Boolean()),
	allowEmptySignature: Type.Optional(Type.Boolean()),

// 同文件另见：cacheControlFormat: Type.Literal("anthropic")（101 行）、
// OpenAI 两张 compat schema 的 supportsLongCacheRetention（111 / 119 行）、
// 以及 cost 中独立的 cacheRead / cacheWrite 费率（147-148 行，支持分档）。