docs: 交付文档、维护指引与端到端回归

补齐 README、开发计划、维护交接与 E2E 用例，执行最终验收。

e2e/services/fake-llm/protocol.ts

@@ -0,0 +1,77 @@
+// Wire types shared by the fake LLM server and the test-side client (src/fakes/llm.ts).
+// Type-only: the server imports this with `import type`, so Node's type stripping
+// needs nothing at runtime.
+
+/** A serialisable RegExp. */
+export interface Pattern {
+  source: string
+  flags: string
+}
+
+/**
+ * What the model "says" for one matched turn.
+ *
+ * - `message` / `fields` fill a structured (json_schema) reply — e.g. ChatAgent's
+ *   ChatResponse. `fields` overlays any schema field: transfer_to_human, create_ticket, …
+ * - `text` is used for plain-text requests (the transfer agent, the key check).
+ *   Falls back to `message`.
+ */
+export interface Reply {
+  message?: string
+  fields?: Record<string, unknown>
+  text?: string
+  /** Token counts to report, for metering. Default 1/1. */
+  usage?: { prompt_tokens: number; completion_tokens: number }
+}
+
+/** How the turn goes wrong instead of (or before) answering. */
+export interface Fault {
+  /** Answer with this HTTP status and an OpenAI-style error instead of a completion. */
+  status?: number
+  message?: string
+  /** Wait this long first — past AGENT_RUN_TIMEOUT to exercise the backend's timeout. */
+  delayMs?: number
+  /** Return content that is not the JSON the request's schema asks for. */
+  malformed?: boolean
+}
+
+export interface ToolCall {
+  name: string
+  arguments: Record<string, unknown>
+}
+
+export interface ScenarioRule {
+  /** Tested against the last user message. Omitted = matches anything (otherwise). */
+  match?: Pattern
+  /** Ask the backend to run a tool first; `reply` is sent after the tool result comes back. */
+  toolCall?: ToolCall
+  reply: Reply
+  fault?: Fault
+}
+
+export interface ScenarioSpec {
+  /**
+   * Some backend calls never see the agent's instructions (the transfer agent's prompt
+   * has no `[[e2e:<id>]]` marker), so a scenario can also match on the unique agent name.
+   */
+  agentName?: string
+  rules: ScenarioRule[]
+}
+
+export interface RecordedRequest {
+  at: string
+  system: string
+  lastUser: string
+  lastRole: string
+  /** 'json_schema:<name>' | 'json_object' | null */
+  responseFormat: string | null
+  tools: string[]
+  /** Contents of the tool-result messages in this request — e.g. what knowledge search returned. */
+  toolResults: string[]
+  /** What the fake answered: 'tool_call:<name>' | 'structured' | 'text' | 'error:<status>' | 'malformed' */
+  answeredWith: string
+  /** The model the backend asked for. */
+  model: string
+  /** The last 6 characters of the API key it sent — which key the org's config used. */
+  keySuffix: string
+}