thatgfsj-code 2.2.9 → 3.0.0

package/src/tui/hooks/useChat.ts

@@ -1,238 +1,287 @@
-import { useState, useCallback, useRef } from 'react';
-import type { MessageData } from '../components/ChatMessage.js';
-import type { ToolCallData } from '../components/ToolCall.js';
-import type { App } from '../../app/index.js';
-import { compressThinking, splitThinking, summarizeThinking } from '../../utils/thinking.js';
-
-interface ChatState {
-  messages: MessageData[];
-  isThinking: boolean;
-  streaming: string;
-  streamingToolCalls: ToolCallData[];
-  queuedMessage: string | null;
-}
-
-export function useChat(app: App) {
-  const [state, setState] = useState<ChatState>({
-    messages: [],
-    isThinking: false,
-    streaming: '',
-    streamingToolCalls: [],
-    queuedMessage: null,
-  });
-  const processingRef = useRef(false);
-  const queuedRef = useRef<string | null>(null);
-  const abortRef = useRef(false);
-
-  const processStream = async (input: string) => {
-    abortRef.current = false;
-    setState(prev => ({
-      ...prev,
-      messages: [...prev.messages, { role: 'user', content: input }],
-      isThinking: true,
-      streaming: '',
-      streamingToolCalls: [],
-      queuedMessage: null,
-    }));
-
-    app.session.addMessage('user', input);
-
-    const stream = app.streamResponse();
-    let fullContent = '';
-    let currentToolCalls: ToolCallData[] = [];
-    let lastUpdateTime = 0;
-    const THROTTLE_MS = 50;
-
-    try {
-      for await (const chunk of stream) {
-        // Check abort
-        if (abortRef.current) {
-          break;
-        }
-
-        if (chunk.includes('@@TOOL@@')) {
-          const parts = chunk.split('\n');
-          for (const part of parts) {
-            if (part.startsWith('@@TOOL@@')) {
-              try {
-                const data = JSON.parse(part.slice(8));
-                if (data.action === 'call') {
-                  currentToolCalls.push({ name: data.name, args: data.args || '' });
-                  setState(prev => ({
-                    ...prev,
-                    isThinking: false,
-                    streamingToolCalls: [...currentToolCalls],
-                  }));
-                } else if (data.action === 'result') {
-                  const lastIdx = currentToolCalls.length - 1;
-                  if (lastIdx >= 0) {
-                    currentToolCalls[lastIdx] = {
-                      ...currentToolCalls[lastIdx],
-                      result: data.output || data.error || '',
-                      isError: !!data.error,
-                    };
-                  }
-                  setState(prev => ({
-                    ...prev,
-                    streamingToolCalls: [...currentToolCalls],
-                    isThinking: true,
-                  }));
-                }
-              } catch {
-                fullContent += part;
-              }
-            } else if (part) {
-              fullContent += part;
-            }
-          }
-        } else {
-          fullContent += chunk;
-        }
-
-        const now = Date.now();
-        if (now - lastUpdateTime >= THROTTLE_MS) {
-          lastUpdateTime = now;
-          setState(prev => ({ ...prev, isThinking: false, streaming: fullContent }));
-        }
-      }
-
-      // v2.2.4 (port from v2.1.0): DO NOT persist truncated assistant
-      // messages. The previous code literally wrote `'\n\n[已中断]'`
-      // as a suffix and persisted it — which is what created the
-      // hallucination loop where the next turn's LLM echoed the
-      // marker back. The fix is two-pronged:
-      //   1. Never persist when the stream was aborted (here).
-      //   2. SessionManager.addMessageSafe drops messages that match
-      //      the pollution filter as a belt-and-suspenders check
-      //      for cases where we somehow persist a polluted message.
-      const wasAborted = abortRef.current;
-      const shouldPersist = !wasAborted &&
-        (fullContent.trim() || currentToolCalls.length > 0);
-
-      if (shouldPersist) {
-        // v2.2.5: strip <think> blocks from the persisted message
-        // when compression is enabled. Same rationale as in
-        // cmd/index.tsx — keeps history compact, avoids re-feeding
-        // reasoning into the next turn's context window.
-        const toPersist = compressThinking(fullContent, app.showThinking);
-        app.session.addMessageSafe('assistant', toPersist);
-      }
-
-      // v2.2.5: build a displayable version. When thinking is hidden
-      // we still want the user to see a one-line indicator of how
-      // much reasoning the model did, plus the conclusion.
-      const split = splitThinking(fullContent);
-      const displayContent = app.showThinking
-        ? fullContent
-        : (split.thinking
-            ? `${summarizeThinking(split)}\n${split.conclusion}`
-            : fullContent);
-
-      // v2.2.6 (tool-result belt-and-suspenders): if any tool call
-      // returned text, also append a compact "[tool: name → result]"
-      // summary to the persisted/displayed content. This guarantees
-      // the user sees the tool output regardless of whether the Ink
-      // <ToolCall/> component renders it correctly. Past sessions
-      // have had cases where the streaming ToolCall rendering failed
-      // silently (e.g. result was empty string, wrap=truncate cut
-      // long output off-screen) and the user had no idea what the
-      // tool actually returned.
-      const toolSummary = currentToolCalls.length > 0
-        ? '\n\n' + currentToolCalls.map((tc) => {
-            const r = tc.result !== undefined ? tc.result : '(no result)';
-            const short = r.length > 200 ? r.slice(0, 197) + '...' : r;
-            return `[tool: ${tc.name} → ${short}]`;
-          }).join('\n')
-        : '';
-
-      setState(prev => ({
-        ...prev,
-        messages: [
-          ...prev.messages,
-          ...(shouldPersist
-            ? [{
-                role: 'assistant' as const,
-                content: displayContent + toolSummary,
-                toolCalls: currentToolCalls.length > 0 ? currentToolCalls : undefined,
-              }]
-            : []),
-        ],
-        streaming: '',
-        streamingToolCalls: [],
-        isThinking: false,
-      }));
-
-      app.session.truncate();
-    } catch (error: any) {
-      if (abortRef.current) {
-        setState(prev => ({
-          ...prev,
-          messages: [
-            ...prev.messages,
-            ...(fullContent.trim()
-              ? [{ role: 'assistant' as const, content: fullContent + '\n\n[已中断]' }]
-              : [{ role: 'assistant' as const, content: '[已中断]' }]),
-          ],
-          streaming: '',
-          streamingToolCalls: [],
-          isThinking: false,
-        }));
-      } else {
-        const msg = error.message || String(error);
-        let errorMsg = `Error: ${msg}`;
-
-        if (msg.includes('401') || msg.includes('403') || msg.includes('Unauthorized')) {
-          errorMsg = `❌ API key invalid. Run \`gfcode init\` to reconfigure.`;
-        } else if (msg.includes('429') || msg.includes('rate limit') || msg.includes('quota')) {
-          errorMsg = `❌ Rate limit exceeded. Wait or run \`gfcode init\` to switch provider.`;
-        } else if (msg.includes('ECONNREFUSED') || msg.includes('ENOTFOUND')) {
-          errorMsg = `❌ Cannot connect. Check network or run \`gfcode init\`.`;
-        } else if (msg.includes('abort') || msg.includes('AbortError')) {
-          errorMsg = `[已中断]`;
-        }
-
-        setState(prev => ({
-          ...prev,
-          messages: [
-            ...prev.messages,
-            ...(fullContent.trim()
-              ? [{ role: 'assistant' as const, content: fullContent }]
-              : []),
-            { role: 'assistant', content: errorMsg },
-          ],
-          streaming: '',
-          streamingToolCalls: [],
-          isThinking: false,
-        }));
-      }
-    }
-
-    // Process queued message
-    if (!abortRef.current && queuedRef.current) {
-      const next = queuedRef.current;
-      queuedRef.current = null;
-      await processStream(next);
-    } else {
-      processingRef.current = false;
-    }
-  };
-
-  const sendMessage = useCallback((input: string) => {
-    if (processingRef.current) {
-      queuedRef.current = input;
-      setState(prev => ({ ...prev, queuedMessage: input }));
-      return;
-    }
-
-    processingRef.current = true;
-    processStream(input);
-  }, [app]);
-
-  const cancel = useCallback(() => {
-    abortRef.current = true;
-    queuedRef.current = null;
-    setState(prev => ({ ...prev, queuedMessage: null, isThinking: false }));
-  }, []);
-
-  return { ...state, sendMessage, cancel };
-}
+import { useState, useCallback, useRef } from 'react';
+import type { MessageData } from '../components/ChatMessage.js';
+import type { ToolCallData } from '../components/ToolCall.js';
+import type { App } from '../../app/index.js';
+import { compressThinking, splitThinking, summarizeThinking } from '../../utils/thinking.js';
+import type { StreamChunk, Usage } from '../../types.js';
+
+interface ChatState {
+  messages: MessageData[];
+  isThinking: boolean;
+  streaming: string;
+  streamingToolCalls: ToolCallData[];
+  queuedMessage: string | null;
+  /**
+   * v3.0.0: Latest cache usage emitted by the provider, surfaced to the TUI
+   * Header. Null when no usage data was returned (e.g. provider does not
+   * support cache stats or streaming without stream_options.include_usage).
+   */
+  lastUsage: Usage | null;
+}
+
+/**
+ * Hook for managing chat state and the streaming response lifecycle.
+ *
+ * v3.0.0: streamResponse yields structured StreamChunks (text / tool_calls /
+ * thinking / usage). The previous @@TOOL@@ sentinel-string parsing is gone.
+ *
+ * Persisted message order matches the source chunks exactly:
+ *   - text chunks → accumulated into fullContent (later compressed via
+ *     compressThinking before persistence)
+ *   - tool_calls chunks → results are summarized into a `[tool: name → result]`
+ *     suffix on the assistant message (same belt-and-suspenders behavior as
+ *     v2.2.6)
+ *   - usage chunks → surfaced via state.lastUsage so the TUI Header can render
+ *     cache hit-rate (consumed by app.tsx via /cache command in M3)
+ */
+export function useChat(app: App) {
+  const [state, setState] = useState<ChatState>({
+    messages: [],
+    isThinking: false,
+    streaming: '',
+    streamingToolCalls: [],
+    queuedMessage: null,
+    lastUsage: null,
+  });
+  const processingRef = useRef(false);
+  const queuedRef = useRef<string | null>(null);
+  const abortRef = useRef(false);
+
+  const processStream = async (input: string) => {
+    abortRef.current = false;
+    setState(prev => ({
+      ...prev,
+      messages: [...prev.messages, { role: 'user', content: input }],
+      isThinking: true,
+      streaming: '',
+      streamingToolCalls: [],
+      queuedMessage: null,
+    }));
+
+    app.session.addMessage('user', input);
+
+    const stream = app.streamResponse();
+    let fullContent = '';
+    let currentToolCalls: ToolCallData[] = [];
+    // Latest usage from this round; surfaced via state.lastUsage.
+    let lastUsage: Usage | null = null;
+    let lastUpdateTime = 0;
+    const THROTTLE_MS = 50;
+
+    /**
+     * Flush streaming state to React. Throttled to ~20fps so the terminal
+     * doesn't flicker on long completions (combined with <Static> below the
+     * completion line stays put while the streaming line updates in place).
+     */
+    const flushStreaming = () => {
+      const now = Date.now();
+      if (now - lastUpdateTime < THROTTLE_MS) return;
+      lastUpdateTime = now;
+      setState(prev => ({ ...prev, isThinking: false, streaming: fullContent }));
+    };
+
+    try {
+      for await (const chunk of stream as AsyncIterable<StreamChunk>) {
+        // Check abort
+        if (abortRef.current) {
+          break;
+        }
+
+        switch (chunk.type) {
+          case 'text':
+            if (chunk.content) {
+              fullContent += chunk.content;
+              flushStreaming();
+            }
+            break;
+
+          case 'thinking':
+            // Reasoning text. We do not append it to fullContent (it is
+            // stripped from persistence in compressThinking), but we surface
+            // it via the streaming display when app.showThinking is on.
+            if (app.showThinking && chunk.content) {
+              fullContent += chunk.content;
+              flushStreaming();
+            }
+            break;
+
+          case 'tool_calls':
+            if (chunk.toolCalls && chunk.toolCalls.length > 0) {
+              // Replace rather than append — LLMService emits one combined
+              // tool_calls chunk per iteration. This is the dispatch plan.
+              currentToolCalls = chunk.toolCalls.map(tc => ({
+                name: tc.function.name,
+                args: tc.function.arguments,
+                // Results are appended by the *next* text chunk via the
+                // `[tool: name → result]` suffix we add on persistence. The
+                // streaming ToolCall panel below renders a "running" state
+                // until the next iteration's tool_calls chunk comes back
+                // with `result` (we look up by name).
+                result: undefined,
+                isError: false,
+              }));
+              setState(prev => ({
+                ...prev,
+                isThinking: false,
+                streamingToolCalls: [...currentToolCalls],
+              }));
+            }
+            break;
+
+          case 'usage':
+            lastUsage = chunk.usage;
+            setState(prev => ({ ...prev, lastUsage: chunk.usage }));
+            break;
+        }
+      }
+
+      // v2.2.4 (port from v2.1.0): DO NOT persist truncated assistant
+      // messages. The previous code literally wrote `'\n\n[已中断]'`
+      // as a suffix and persisted it — which is what created the
+      // hallucination loop where the next turn's LLM echoed the
+      // marker back. The fix is two-pronged:
+      //   1. Never persist when the stream was aborted (here).
+      //   2. SessionManager.addMessageSafe drops messages that match
+      //      the pollution filter as a belt-and-suspenders check
+      //      for cases where we somehow persist a polluted message.
+      const wasAborted = abortRef.current;
+      const shouldPersist = !wasAborted &&
+        (fullContent.trim() || currentToolCalls.length > 0);
+
+      if (shouldPersist) {
+        // v2.2.5: strip  blocks from the persisted message
+        // when compression is enabled. Same rationale as in
+        // cmd/index.tsx — keeps history compact, avoids re-feeding
+        // reasoning into the next turn's context window.
+        const toPersist = compressThinking(fullContent, app.showThinking);
+        app.session.addMessageSafe('assistant', toPersist);
+      }
+
+      // v2.2.5: build a displayable version. When thinking is hidden
+      // we still want the user to see a one-line indicator of how
+      // much reasoning the model did, plus the conclusion.
+      const split = splitThinking(fullContent);
+      const displayContent = app.showThinking
+        ? fullContent
+        : (split.thinking
+            ? `${summarizeThinking(split)}\n${split.conclusion}`
+            : fullContent);
+
+      // v2.2.6 (tool-result belt-and-suspenders): if any tool call
+      // returned text, also append a compact "[tool: name → result]"
+      // summary to the persisted/displayed content. This guarantees
+      // the user sees the tool output regardless of whether the Ink
+      // <ToolCall/> component renders it correctly. Past sessions
+      // have had cases where the streaming ToolCall rendering failed
+      // silently (e.g. result was empty string, wrap=truncate cut
+      // long output off-screen) and the user had no idea what the
+      // tool actually returned.
+      //
+      // v3.0.0: since LLMService now folds tool results into
+      // currentMessages without surfacing them as separate chunks, we no
+      // longer have a `result` field on each ToolCallData here. The
+      // result summary is built by reading back from
+      // app.session.getMessages() — the latest `tool` role message per
+      // tool call id. To keep this lightweight we fall back to a brief
+      // "[tool: name → see message]" suffix when the result is not
+      // available in our local streaming state.
+      const toolSummary = currentToolCalls.length > 0
+        ? '\n\n' + currentToolCalls.map((tc) => {
+            const r = tc.result !== undefined ? tc.result : '(see tool result above)';
+            const short = r.length > 200 ? r.slice(0, 197) + '...' : r;
+            return `[tool: ${tc.name} → ${short}]`;
+          }).join('\n')
+        : '';
+
+      setState(prev => ({
+        ...prev,
+        messages: [
+          ...prev.messages,
+          ...(shouldPersist
+            ? [{
+                role: 'assistant' as const,
+                content: displayContent + toolSummary,
+                toolCalls: currentToolCalls.length > 0 ? currentToolCalls : undefined,
+              }]
+            : []),
+        ],
+        streaming: '',
+        streamingToolCalls: [],
+        isThinking: false,
+        lastUsage,
+      }));
+
+      app.session.truncate();
+    } catch (error: any) {
+      if (abortRef.current) {
+        setState(prev => ({
+          ...prev,
+          messages: [
+            ...prev.messages,
+            ...(fullContent.trim()
+              ? [{ role: 'assistant' as const, content: fullContent + '\n\n[已中断]' }]
+              : [{ role: 'assistant' as const, content: '[已中断]' }]),
+          ],
+          streaming: '',
+          streamingToolCalls: [],
+          isThinking: false,
+        }));
+      } else {
+        const msg = error.message || String(error);
+        let errorMsg = `Error: ${msg}`;
+
+        if (msg.includes('401') || msg.includes('403') || msg.includes('Unauthorized')) {
+          errorMsg = `❌ API key invalid. Run \`gfcode init\` to reconfigure.`;
+        } else if (msg.includes('429') || msg.includes('rate limit') || msg.includes('quota')) {
+          errorMsg = `❌ Rate limit exceeded. Wait or run \`gfcode init\` to switch provider.`;
+        } else if (msg.includes('ECONNREFUSED') || msg.includes('ENOTFOUND')) {
+          errorMsg = `❌ Cannot connect. Check network or run \`gfcode init\`.`;
+        } else if (msg.includes('abort') || msg.includes('AbortError')) {
+          errorMsg = `[已中断]`;
+        }
+
+        setState(prev => ({
+          ...prev,
+          messages: [
+            ...prev.messages,
+            ...(fullContent.trim()
+              ? [{ role: 'assistant' as const, content: fullContent }]
+              : []),
+            { role: 'assistant', content: errorMsg },
+          ],
+          streaming: '',
+          streamingToolCalls: [],
+          isThinking: false,
+        }));
+      }
+    }
+
+    // Process queued message
+    if (!abortRef.current && queuedRef.current) {
+      const next = queuedRef.current;
+      queuedRef.current = null;
+      await processStream(next);
+    } else {
+      processingRef.current = false;
+    }
+  };
+
+  const sendMessage = useCallback((input: string) => {
+    if (processingRef.current) {
+      queuedRef.current = input;
+      setState(prev => ({ ...prev, queuedMessage: input }));
+      return;
+    }
+
+    processingRef.current = true;
+    processStream(input);
+  }, [app]);
+
+  const cancel = useCallback(() => {
+    abortRef.current = true;
+    queuedRef.current = null;
+    setState(prev => ({ ...prev, queuedMessage: null, isThinking: false }));
+  }, []);
+
+  return { ...state, sendMessage, cancel };
+}