fix(web): add prompt cache affinity to canvas agent (#74)

Co-authored-by: Tiantianr <260026737+Tiantianr@users.noreply.github.com>

web/src/components/canvas/canvas-assistant-panel.tsx

@@ -23,6 +23,7 @@ import { NODE_DEFAULT_SIZE } from "@/constant/canvas";
 import { CanvasNodeType, type CanvasAssistantMessage, type CanvasAssistantPendingBackendSession, type CanvasAssistantReference, type CanvasAssistantSession, type CanvasNodeData } from "@/types/canvas";
 import { useCanvasAgentStore } from "@/stores/canvas/use-canvas-agent-store";
 import { previewCanvasAgentOps, summarizeCanvasAgentOps, type CanvasAgentOp, type CanvasAgentOperationImpact, type CanvasAgentSnapshot } from "@/lib/canvas/canvas-agent-ops";
+import { canvasAgentPromptCacheKey } from "@/lib/openai-prompt-cache";
 
 export const CANVAS_AGENT_PANEL_MOTION_MS = 500;
 const PANEL_MOTION_SECONDS = CANVAS_AGENT_PANEL_MOTION_MS / 1000;
@@ -412,7 +413,7 @@ export function CanvasAssistantPanel({ nodes, selectedNodeIds, snapshot, project
             const result = await requestToolResponse({ ...requestConfig, systemPrompt: "" }, messages, ONLINE_AGENT_TOOLS, "required", (text) => {
                 streamed = text;
                 if (text.trim()) upsertMessage(sessionId, { id: assistantId, role: "assistant", text });
-            });
+            }, { promptCacheKey: canvasAgentPromptCacheKey(sessionId) });
             addOnlineLog("模型工具回复", result);
             if (result.toolCalls.length) {
                 const writableCalls = result.toolCalls.filter(isWritableToolCall);
@@ -468,7 +469,7 @@ export function CanvasAssistantPanel({ nodes, selectedNodeIds, snapshot, project
         const next = await requestToolResponse({ ...requestConfig, systemPrompt: "" }, nextMessages, ONLINE_AGENT_TOOLS, "auto", (text) => {
             streamed = text;
             if (text.trim()) upsertMessage(sessionId, { id: assistantId, role: "assistant", text });
-        });
+        }, { promptCacheKey: canvasAgentPromptCacheKey(sessionId) });
         addOnlineLog(`Agent Tool Loop ${step + 1} 回复`, next);
         if (next.toolCalls.length) {
             const writableCalls = next.toolCalls.filter(isWritableToolCall);

web/src/lib/openai-prompt-cache.ts

@@ -0,0 +1,17 @@
+export const OPENAI_PROMPT_CACHE_KEY_MAX_LENGTH = 64;
+
+export function clampOpenAIPromptCacheKey(value?: string) {
+    const key = value?.trim();
+    if (!key) return undefined;
+    return Array.from(key).slice(0, OPENAI_PROMPT_CACHE_KEY_MAX_LENGTH).join("");
+}
+
+export function canvasAgentPromptCacheKey(sessionId: string) {
+    const value = sessionId.trim();
+    return value ? clampOpenAIPromptCacheKey(`canvas-agent:${value}`) : undefined;
+}
+
+export function withOpenAIPromptCacheKey<T extends Record<string, unknown>>(body: T, value?: string): T & { prompt_cache_key?: string } {
+    const key = clampOpenAIPromptCacheKey(value);
+    return key ? { ...body, prompt_cache_key: key } : body;
+}

web/src/services/api/image.ts

@@ -7,6 +7,7 @@ import { buildImageReferencePromptText } from "@/lib/image-reference-prompt";
 import { channelRequest } from "@/services/api/custom-channel-relay";
 import { imageToDataUrl } from "@/services/image-storage";
 import type { ReferenceImage } from "@/types/image";
+import { withOpenAIPromptCacheKey } from "@/lib/openai-prompt-cache";
 
 export type AiTextMessage = {
     role: "system" | "user" | "assistant";
@@ -100,7 +101,7 @@ type GeminiPayload = {
     promptFeedback?: { blockReason?: string };
 };
 type GeminiStreamState = { buffer: string; text: string; toolCalls: ResponseToolCall[]; error?: string };
-type RequestOptions = { signal?: AbortSignal };
+type RequestOptions = { signal?: AbortSignal; promptCacheKey?: string };
 
 const QUALITY_BASE: Record<string, number> = {
     low: 1024,
@@ -945,13 +946,21 @@ export async function requestToolResponse(config: AiConfig, messages: ResponseIn
                 parallel_tool_calls: false,
             }, onDelta, options);
         }
-        return await requestStreamingResponse(requestConfig, {
-            model: requestConfig.model,
-            input: toResponseInput(withSystemMessage(requestConfig, messages)),
-            tools: tools.map(toResponseTool),
-            tool_choice: toolChoice,
-            parallel_tool_calls: false,
-        }, onDelta, options);
+        return await requestStreamingResponse(
+            requestConfig,
+            withOpenAIPromptCacheKey(
+                {
+                    model: requestConfig.model,
+                    input: toResponseInput(withSystemMessage(requestConfig, messages)),
+                    tools: tools.map(toResponseTool),
+                    tool_choice: toolChoice,
+                    parallel_tool_calls: false,
+                },
+                options?.promptCacheKey,
+            ),
+            onDelta,
+            options,
+        );
     } catch (error) {
         throw new Error(readAxiosError(error, "请求失败"));
     }