@ghyper9023/pi-dev-workflow 0.5.1 → 0.6.1

package/extensions/workflow-engine.ts

@@ -389,20 +389,16 @@ function getGitDiffChanges(cwd: string): GitFileChange[] {
 }
 
 /**
- * Convert an internal tool type ("edit"/"new"/"delete") to a git status letter ("M"/"A"/"D").
- * Used to unify all tool entries to git-format display: "M   path", "A   path", "D   path".
+ * 检查路径是否为工作流生成的产物路径（.pi-dev-output/），
+ * 若是则不应出现在 A: M: D: 文件变更列表中。
+ * 该目录已加入 .gitignore，git diff 本身已忽略，
+ * 此函数作为防御性过滤。
  */
-function toGitStatus(toolType: string): string {
-	switch (toolType.toLowerCase()) {
-		case "new": case "write": case "create": case "add": case "created": case "added":
-			return "A";
-		case "delete": case "remove": case "deleted": case "removed":
-			return "D";
-		default:
-			return "M";
-	}
+function isWorkflowArtifactPath(filePath: string): boolean {
+	return filePath.startsWith(".pi-dev-output/");
 }
 
+
 /**
  * Check whether a file's current content hash differs from its baseline.
  * Returns true if the hash changed or the file is no longer accessible.
@@ -447,13 +443,20 @@ function captureBaseline(cwd: string): void {
  *
  * Uses git-format status codes (M, A, D) for display consistency.
  * Deduplicates against existing _workflowFileChanges.
+ *
+ * The dedup key includes stepIndex (filePath:stepIndex) so the same file
+ * can appear across different workflow steps (e.g., created by worker in
+ * loop-group, then reviewed by reviewer) without being collapsed into one.
+ * Without :stepIndex, the Set dedup would wrongly skip a file modified in
+ * a later step just because it was already tracked in an earlier step.
+ *
  */
 function updateToolsFromGit(cwd: string, stepIndex: number, agentName: string): void {
 	const currentChanges = getGitDiffChanges(cwd);
-	const seen = new Set(_workflowFileChanges.map(c => c.filePath));
-
+	const seen = new Set(_workflowFileChanges.map(c => `${c.filePath}:${c.stepIndex}`));
+	
 	for (const change of currentChanges) {
-		if (seen.has(change.path)) continue;
+		if (seen.has(`${change.path}:${stepIndex}`)) continue;
 
 		// ── Baseline filtering ──────────────────────────────
 		// Skip files that were already dirty at workflow start and haven't been touched.
@@ -470,7 +473,13 @@ function updateToolsFromGit(cwd: string, stepIndex: number, agentName: string):
 			}
 		}
 
-		seen.add(change.path);
+		// Skip files in .pi-dev-output/ — these are workflow-generated artifacts (plans, reviews)
+		// and should NOT be reported as user repo changes. Real git clients (VSCode, Zed) also
+		// ignore this directory (it's in .gitignore). This filter is defense-in-depth.
+		// Note: git always uses forward slashes in paths, even on Windows.
+		if (isWorkflowArtifactPath(change.path)) continue;
+
+		seen.add(`${change.path}:${stepIndex}`);
 		const type: FileChangeEntry["type"] =
 			change.status === "A" ? "new" :
 			change.status === "D" ? "delete" :
@@ -883,47 +892,6 @@ function addWidgetSubStepTool(stepIndex: number, agentName: string, tool: string
 		if (!sub.tools) sub.tools = [];
 		sub.tools.push(tool);
 		if (sub.tools.length > 20) sub.tools = sub.tools.slice(-20); // keep last 20
-
-		// Also track as file change for checkpoint
-		// Support both old format ("edit: path") and new git-format ("M   path", "A   path", "D   path")
-		const oldMatch = tool.match(/^(edit|new|delete|read):\s*(.+)/i);
-		const gitMatch = !oldMatch ? tool.match(/^([MAD])\s{2,}(.+)$/) : null;
-		if (oldMatch) {
-			const changeType = oldMatch[1]!.toLowerCase() as FileChangeEntry["type"];
-			const filePath = oldMatch[2]!.trim();
-			const exists = _workflowFileChanges.some(
-				c => c.filePath === filePath && c.type === changeType && c.stepIndex === stepIndex && c.agent === agentName,
-			);
-			if (!exists && filePath.length > 3) {
-				_workflowFileChanges.push({
-					agent: agentName,
-					stepIndex,
-					type: changeType,
-					filePath,
-					timestamp: new Date().toISOString(),
-				});
-			}
-		} else if (gitMatch) {
-			const gitStatus = gitMatch[1]!;
-			const changeType: FileChangeEntry["type"] =
-				gitStatus === "A" ? "new" :
-				gitStatus === "D" ? "delete" :
-				"edit";
-			const filePath = gitMatch[2]!.trim();
-			const exists = _workflowFileChanges.some(
-				c => c.filePath === filePath && c.type === changeType && c.stepIndex === stepIndex && c.agent === agentName,
-			);
-			if (!exists && filePath.length > 3) {
-				_workflowFileChanges.push({
-					agent: agentName,
-					stepIndex,
-					type: changeType,
-					filePath,
-					timestamp: new Date().toISOString(),
-				});
-			}
-		}
-
 		refreshWidget();
 	}
 }
@@ -1123,33 +1091,81 @@ async function runAgentWithProgress(
 		}
 	}
 
-	// Parse progress messages for tool calls and outputs
-	// NOTE: In-progress parsing is best-effort and intentionally conservative.
-	// Final file change detection relies on git diff --name-status (see updateToolsFromGit below)
-	// which is 100% accurate and free of AI text noise.
-	const result = await spawnSubagent(agent, task, _workflowCwd, signal, timeoutMs, (progress) => {
-		// Try to parse tool calls from progress messages
-		// Only match if it looks like a file path (contains a dot or path separator)
-		const toolMatch = progress.match(/(edit|read|write|new|bash|grep|find|ls|delete|remove)\s*[:：]\s*(\S+)/i);
-		if (toolMatch) {
-			const toolType = toolMatch[1]!.toLowerCase();
-			const target = toolMatch[2]!;
-			// Only classify as file operation if it's a file path-like string
-			if (target.includes(".") || target.includes("/") || target.includes("\\")) {
-				const gitStatus = toGitStatus(toolType);
-				addWidgetSubStepTool(stepIndex, agentName, `${gitStatus}   ${target}`);
-				_widgetExtraToolCount++;
-			}
-		}
-		// Detect output file paths — ONLY match explicit .pi-dev-output paths, never free-form text
-		const outputMatch = progress.match(/\.pi-dev-output\/[^\s,;)\]}]{5,}\.\w+/i);
-		if (outputMatch) {
-			const pathCandidate = outputMatch[0]!.trim();
-			if (pathCandidate.length > 15 && pathCandidate.length < 300) {
-				addWidgetSubStepOutput(stepIndex, agentName, pathCandidate);
+	// ── Start git diff polling for real file changes ──
+	// This replaces the unreliable regex-based tool call sniffing that was removed.
+	// git diff --name-status is the same approach used by VSCode, Zed, and every
+	// professional git client — it's deterministic, noise-free, and 100% accurate.
+	let _gitPollTimer: ReturnType<typeof setInterval> | null = null;
+	const _pollSeen = new Set<string>();
+	if (_workflowCwd) {
+		_gitPollTimer = setInterval(() => {
+			try {
+				const currentChanges = getGitDiffChanges(_workflowCwd);
+				for (const change of currentChanges) {
+					const key = `${change.path}:${stepIndex}`;
+					if (_pollSeen.has(key)) continue;
+					_pollSeen.add(key);
+
+					if (isWorkflowArtifactPath(change.path)) continue;
+					if (_workflowBaseline.length > 0) {
+						const baseEntry = _workflowBaseline.find(b => b.path === change.path);
+						if (baseEntry) {
+							if (!hasContentChanged(_workflowCwd, change.path, baseEntry.hash)) continue;
+							// ⭐ 同步删除已变更的基线条目，与 updateToolsFromGit 行为一致
+							_workflowBaseline = _workflowBaseline.filter(b => b.path !== change.path);
+						}
+					}
+
+					// ⭐ Also record in _workflowFileChanges so the completion report file tree
+					// and revertStepChanges have accurate data. The final updateToolsFromGit call
+					// will skip this via seen, so we must record it here.
+					const type: FileChangeEntry["type"] =
+						change.status === "A" ? "new" :
+						change.status === "D" ? "delete" :
+						"edit";
+					_workflowFileChanges.push({
+						agent: agentName,
+						stepIndex,
+						type,
+						filePath: change.path,
+						timestamp: new Date().toISOString(),
+					});
+					addWidgetSubStepTool(stepIndex, agentName, `${change.status}   ${change.path}`);
+					_widgetExtraToolCount++;
+				}
+			} catch { /* 静默忽略 git 错误 */ }
+		}, 5000);
+		_gitPollTimer.unref();
+	}
+
+	// Run the sub-agent with progress reporting
+	// The toolMatch regex sniffing has been removed — it was the primary source of
+	// false positives in the A: M: D: panel (matching AI natural language as tool calls).
+	// File change detection is now handled entirely by git diff polling (above) and
+	// the final updateToolsFromGit call (below), both of which use git's own API.
+	let result;
+	try {
+		result = await spawnSubagent(agent, task, _workflowCwd, signal, timeoutMs, (progress) => {
+			// Detect output file paths — only match .pi-dev-output paths belonging to current workflow
+			if (_workflowId) {
+				const outputMatch = progress.match(/\.pi-dev-output\/[a-zA-Z0-9_\/\.-]+/i);
+				if (outputMatch) {
+					const pathCandidate = outputMatch[0]!.trim();
+					// ⭐ 严格白名单：拒绝中文、引号、括号等非路径字符
+					if (pathCandidate.length > 15 && pathCandidate.length < 300 &&
+						pathCandidate.includes(_workflowId) &&
+						/^[\w.\/-]+$/.test(pathCandidate)) {
+						addWidgetSubStepOutput(stepIndex, agentName, pathCandidate);
+					}
+				}
 			}
+		}, _workflowId ? { workflowId: _workflowId } : undefined);
+	} finally {
+		if (_gitPollTimer) {
+			clearInterval(_gitPollTimer);
+			_gitPollTimer = null;
 		}
-	}, _workflowId ? { workflowId: _workflowId } : undefined);
+	}
 
 	const agentDuration = Date.now() - agentStartTime;
 
@@ -1163,127 +1179,56 @@ async function runAgentWithProgress(
 		toolCount: _widgetExtraToolCount,
 	});
 
-	// ── Post-completion: parse subagent output for tool calls and file paths ──
-	// Progress messages from spawnSubagent rarely contain tool info,
-	// so we scan the full output after completion.
-	const allOutput = (result.output || "") + "\n" + (result.stderr || "");
+	// ── Post-completion: file change detection via git diff ONLY ──
+	// We do NOT parse agent output text for file paths — that approach is fragile and
+	// produces false positives (e.g., matching JS property names like `.push`, `.length`
+	// from code snippets). Instead, we rely entirely on git diff --name-status, which is
+	// the same approach used by VSCode, Zed, and every other professional git client.
+	// It's deterministic, noise-free, and always produces correct relative paths.
 	const finalOutput = extractFinalOutput(result.output) || result.output;
-	const searchText = allOutput + "\n" + finalOutput;
-
-	// Detect file creation/modification patterns from agent's final output text
-	// The agent's response typically lists files using markdown backticks or bullet points
-	const filePatterns = [
-		// Markdown code blocks with file paths: `src/main.rs`, `path/to/file.ts`
-		/`([^`]+\.[a-zA-Z0-9_]+)`/g,
-		// Bullet points with file operation verbs: - Modify `src/main.rs`, * Created `file.ts`
-		/(?:^|\n)\s*[-*]\s*(?:modified|created|updated|edited|added|deleted|removed|changed|wrote|writes?)\s*[`"']?([^`"'\n,]+\.[a-zA-Z0-9_]+)[`"']?/gim,
-		// Descriptive: "I've modified src/main.rs", "reading config.json"
-		/(?:modified|created|updated|edited|added|deleted|removed|changed|wrote|write|writes|read|reads?)\s+(?:the\s+)?[`"']?([^`"'\n,]+\.[a-zA-Z0-9_]+)[`"']?/gi,
-		// Chinese patterns
-		/(?:编写|创建|修改|删除|读取|写入|更新)\s*(?:了|文件)?\s*[:：]?\s*[`"']?([^`"'\s,，]+\.[a-zA-Z0-9_]+)[`"']?/gi,
-		// File path with action prefix: "edit: src/file.ts", "new: src/file.ts"
-		/(?:^|\n)\s*(?:edit|new|delete|read|modify|create|update|add|remove)\s*[:：]\s*([^\n]+\.[a-zA-Z0-9_]+)/gim,
-	];
-	const seenTools = new Set<string>();
-	for (const pattern of filePatterns) {
+
+	// Find output file paths (plans, reviews) for display in the "output:" section
+	// These are workflow artifacts (.pi-dev-output/) — only show files belonging to
+	// the CURRENT workflow to avoid cross-contamination from previous runs.
+	// Note: file change detection itself relies exclusively on git diff below, not text parsing.
+	// ⭐ IMPORTANT: Only search in CLEAN extracted text (finalOutput), NOT raw JSON output.
+	//   The old approach concatenated allOutput (raw JSON + stderr) into searchText and
+	//   matched against it, producing false positives from JSON stringified content.
+	//   By restricting to extractFinalOutput's clean text, we avoid matching against
+	//   JSON-escaped paths, code snippets, or natural language in raw agent output.
+	const cleanText = finalOutput || result.output || "";
+	const seenOutputs = new Set<string>();
+	if (_workflowId) {
+		// Only match .pi-dev-output paths containing the current workflow UUID
+		const escapedId = _workflowId.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+		const outputPattern = new RegExp(`\\.pi-dev-output\\/[a-zA-Z0-9_\\/.-]*${escapedId}[a-zA-Z0-9_\\/.-]*`, 'g');
 		let m;
-		while ((m = pattern.exec(searchText)) !== null) {
-			const filePath = m[1]!.trim()
-				.replace(/[`'"\)\(\]]+$/, "")
-				.replace(/^[`'"\)\(\[]+/, "")
-				.split(/[\s,;]/)[0]!;
-			// Validate it's a real file path
-			if (filePath.length > 3 && filePath.length < 300 && !seenTools.has(filePath)) {
-				// Skip common non-file matches
-				if (filePath.match(/^(the|a|an|this|that|it|its|my|your|our|their|some|any|all|each|every|both|few|many|several|most|other|another|such|what|which|whose|whom|when|where|why|how|who|being|having|doing|making|taking|giving|getting|setting|using|running|going|coming|looking|finding|keeping|putting)/i)) continue;
-				if (filePath.startsWith("http")) continue;
-				if (filePath.length < 6 && !filePath.includes("/")) continue;
-
-				// 额外过滤器：排除明显不是文件路径的脏数据
-				if (filePath.includes("${") || filePath.includes("\\n") || filePath.includes("\\t")) continue;  // 排除模板字符串和转义字符
-				if (filePath.includes("[]") || filePath.includes("{}")) continue;  // 排除数组/对象字面量
-				if (filePath.match(/^[\s,;)\]}]+$/)) continue;  // 排除纯符号
-				seenTools.add(filePath);
-				const fullMatch = m[0]!.toLowerCase();
-				// Determine operation type and convert to git status
-				let toolType = "edit";
-				if (fullMatch.includes("write") || fullMatch.includes("创建") || fullMatch.includes("new") || fullMatch.includes("add") || fullMatch.includes("created") || fullMatch.includes("added")) {
-					toolType = "new";
-				} else if (fullMatch.includes("delete") || fullMatch.includes("删除") || fullMatch.includes("remove") || fullMatch.includes("deleted") || fullMatch.includes("removed")) {
-					toolType = "delete";
-				} else if (fullMatch.includes("read") || fullMatch.includes("读取")) {
-					toolType = "read";
-				}
-				if (toolType !== "read") {
-					const gitStatus = toGitStatus(toolType);
-					addWidgetSubStepTool(stepIndex, agentName, `${gitStatus}   ${filePath}`);
-					_widgetExtraToolCount++;
-				}
+		while ((m = outputPattern.exec(cleanText)) !== null) {
+			const path_ = m[0]!.trim();
+			if (path_.length > 15 && path_.length < 300 && !seenOutputs.has(path_)) {
+				// ⭐ 严格白名单：仅允许 [a-zA-Z0-9_./-] 字符
+				// 拒绝包含中文、引号、括号、空格等非路径字符的乱码匹配
+				if (!/^[\w.\/-]+$/.test(path_)) continue;
+				seenOutputs.add(path_);
+				addWidgetSubStepOutput(stepIndex, agentName, path_);
 			}
 		}
-	}
-
-	// If we found no file tools from text patterns, try alternative approaches
-	// Look for explicit tool call patterns in the raw JSON output
-	if (seenTools.size === 0) {
-		const jsonLines = (result.output || "").split("\n");
-		for (const line of jsonLines) {
-			try {
-				const event = JSON.parse(line);
-				// Look for tool_use events in the JSON stream
-				if (event.type === "message_update" && event.assistantMessageEvent?.type === "tool_use") {
-					const toolName = event.assistantMessageEvent.name;
-					const args = event.assistantMessageEvent.args || {};
-					// write tool: args contains file_path
-					if (toolName === "write" && args.file_path) {
-						const fp = args.file_path.trim();
-						if (!seenTools.has(fp)) {
-							seenTools.add(fp);
-							addWidgetSubStepTool(stepIndex, agentName, `A   ${fp}`);
-							_widgetExtraToolCount++;
-						}
-					}
-					// edit tool: args contains file_path
-					if (toolName === "edit" && args.file_path) {
-						const fp = args.file_path.trim();
-						if (!seenTools.has(fp)) {
-							seenTools.add(fp);
-							addWidgetSubStepTool(stepIndex, agentName, `M   ${fp}`);
-							_widgetExtraToolCount++;
-						}
-					}
-				}
-			} catch { /* not JSON, skip */ }
-		}
-	}
-
-	// Find output file paths (.pi-dev-output, review reports, plan files)
-	// NOTE: Only match explicit paths in .pi-dev-output/ or known review/plan file patterns.
-	// The old pattern matching loose "output:" text was the root cause of "output::0.00004508" noise.
-	// File change detection now relies on git diff --name-status (updateToolsFromGit below),
-	// which is deterministic and noise-free.
-	const outputPathPatterns = [
-		// Direct reference to .pi-dev-output paths: ".pi-dev-output/pi-plans/xxx.md"
-		/\.pi-dev-output\/[a-zA-Z0-9_\/-]+\.[a-zA-Z0-9]+/g,
-		// Review file patterns: "review-20260520-162800.md"
-		/review-\d{8}-\d{6}\.md/g,
-		// Plan file patterns: "20260520-1628-*.md"
-		/\d{8}-\d{4,6}-[a-zA-Z0-9_-]+\.md/g,
-	];
-	const seenOutputs = new Set<string>();
-	for (const pattern of outputPathPatterns) {
+	} else {
+		// No workflow ID (unlikely): match .pi-dev-output paths conservatively
+		const outputPattern = /\.pi-dev-output\/[a-zA-Z0-9_\/-]+\.[a-zA-Z0-9]+/g;
 		let m;
-		while ((m = pattern.exec(searchText)) !== null) {
+		while ((m = outputPattern.exec(cleanText)) !== null) {
 			const path_ = m[0]!.trim();
-			if (path_.length > 10 && path_.length < 300 && !seenOutputs.has(path_)) {
+			if (path_.length > 15 && path_.length < 300 && !seenOutputs.has(path_)) {
+				// ⭐ 严格白名单
+				if (!/^[\w.\/-]+$/.test(path_)) continue;
 				seenOutputs.add(path_);
 				addWidgetSubStepOutput(stepIndex, agentName, path_);
 			}
 		}
 	}
 
-
-	// ── Update file changes from git diff (more accurate than text scraping) ──
+	// ── Update file changes from git diff (more accurate than text scraping) ──	
 	updateToolsFromGit(_workflowCwd, stepIndex, agentName);
 	// Update sub-step status based on result
 	const subStatus: WorkflowSubStepWidgetState["status"] =
@@ -1350,41 +1295,19 @@ async function executeSingleStep(
 		throw new Error(`Agent 错误 (exit ${result.exitCode}): ${result.stderr.slice(0, 500)}`);
 	}
 
-	// ── Capture chain context for single-step agents ──
-	const agentChanges = _workflowFileChanges
-		.filter(c => c.stepIndex === stepIndex && c.agent === agentName)
-		.map(c => `${c.type === "new" ? "A" : c.type === "delete" ? "D" : "M"}   ${c.filePath}`);
-
-	let chainKey: string;
-	if (agentName === "planner") chainKey = "计划制定摘要";
-	else if (agentName === "docWriter") chainKey = "文档更新摘要";
-	else chainKey = `${agentName} 执行摘要`;
-
-	if (agentChanges.length > 0) {
-		const editCount = agentChanges.filter(c => c.startsWith("M")).length;
-		const newCount = agentChanges.filter(c => c.startsWith("A")).length;
-		const delCount = agentChanges.filter(c => c.startsWith("D")).length;
-		const statsParts: string[] = [];
-		if (editCount > 0) statsParts.push(`修改 ${editCount} 个`);
-		if (newCount > 0) statsParts.push(`新增 ${newCount} 个`);
-		if (delCount > 0) statsParts.push(`删除 ${delCount} 个`);
-		const statsLine = statsParts.length > 0 ? `改动统计: ${statsParts.join("，")}\n\n` : "";
-
-		updateChainContext(chainKey,
-			`${agentName} 已完成。\n` +
-			statsLine +
-			`变更文件列表:\n${agentChanges.join("\n")}`
-		);
-	} else {
-		updateChainContext(chainKey, `${agentName} 已完成执行，未检测到文件变更。`);
-	}
-
 	// ── Capture AI work summary as supplementary chain context ──
+	// We intentionally do NOT create chain contexts from _workflowFileChanges here.
+	// File-change-based chain contexts ("计划制定摘要"/"文档更新摘要") were removed in v0.6.0
+	// because their M/A/D counts provided negligible value for downstream agents compared
+	// to the agent's own free-text work summary. The file-change approach also created
+	// tight coupling between executeSingleStep and the _workflowFileChanges data structure.
 	const workSummary = extractFinalOutput(result.output);
 	if (workSummary) {
-		updateChainContext(`${agentName} 工作总结`,
-			`${agentName} 已完成工作，以下是其工作总结：\n\n${workSummary}`
-		);
+		updateChainContext(`${agentName} 工作总结`, `${agentName} 已完成工作，以下是其工作总结：\n\n${workSummary}`);
+	} else if (result.output) {
+		// 保底：使用输出前 200 字符作为简略摘要
+		const fallback = result.output.slice(0, 200).trim();
+		if (fallback) updateChainContext(`${agentName} 工作总结`, fallback);
 	}
 }
 
@@ -1448,37 +1371,6 @@ async function executeLoopGroup(
             }
         }
 
-		// ── After loop agent completes: capture changes + stats for chain context ──
-		const loopAgentChanges = _workflowFileChanges
-			.filter(c => c.stepIndex === stepIndex && c.agent === step.loopAgentName)
-			.map(c => `${c.type === "new" ? "A" : c.type === "delete" ? "D" : "M"}   ${c.filePath}`);
-
-		// Use agent-specific keys so worker and trimmer changes coexist without overwriting
-		const chainKey = step.loopAgentName === "worker" ? "代码实施摘要" : "代码精简摘要";
-
-		if (loopAgentChanges.length > 0) {
-			const editCount = loopAgentChanges.filter(c => c.startsWith("M")).length;
-			const newCount = loopAgentChanges.filter(c => c.startsWith("A")).length;
-			const delCount = loopAgentChanges.filter(c => c.startsWith("D")).length;
-			const statsParts: string[] = [];
-			if (editCount > 0) statsParts.push(`修改 ${editCount} 个`);
-			if (newCount > 0) statsParts.push(`新增 ${newCount} 个`);
-			if (delCount > 0) statsParts.push(`删除 ${delCount} 个`);
-			const statsLine = statsParts.length > 0 ? `改动统计: ${statsParts.join("，")}\n\n` : "";
-
-			updateChainContext(chainKey,
-				`${step.loopAgentName === "worker" ? "代码实施" : "代码精简"}已完成。\n` +
-				statsLine +
-				`变更文件列表:\n${loopAgentChanges.join("\n")}` +
-				(planFileRelPath ? `\n\n实施计划: ${planFileRelPath}\n(可在 .pi-dev-output/pi-plans/ 中 grep UUID ${_workflowId} 找到)` : "")
-			);
-		} else {
-			updateChainContext(chainKey,
-				`${step.loopAgentName} 已完成执行，未检测到文件变更。\n` +
-				(planFileRelPath ? `实施计划: ${planFileRelPath}\n(可在 .pi-dev-output/pi-plans/ 中 grep UUID ${_workflowId} 找到)` : "")
-			);
-		}
-
 		// ── Capture loop agent work summary as supplementary chain context ──
 		const loopFinalText = extractFinalOutput(agentResult.output);
 		if (loopFinalText) {
@@ -1525,25 +1417,11 @@ async function executeLoopGroup(
 			}
 		}
 
-		// ── After reviewer: capture review context for next loop ──
-		// Use agent-specific key so worker-reviewer and trimmer-reviewer feedback don't interfere
-		const reviewChainKey = step.loopAgentName === "worker" ? "代码审查反馈" : "精简审查反馈";
-		if (reviewSummary) {
-			const reviewCountParts: string[] = [];
-			if (reviewSummary.critical > 0) reviewCountParts.push(`严重 ${reviewSummary.critical} 个`);
-			if (reviewSummary.medium > 0) reviewCountParts.push(`中等 ${reviewSummary.medium} 个`);
-			if (reviewSummary.low > 0) reviewCountParts.push(`低 ${reviewSummary.low} 个`);
-			const reviewStats = reviewCountParts.length > 0 ? `发现 ${reviewCountParts.join("，")} 问题。` : "未发现问题。";
-
-			updateChainContext(reviewChainKey,
-				`${reviewStats}\n` +
-				`完整审查报告在 .pi-dev-output/pi-review/md/ 目录中，\n` +
-				`请在工作流输出目录中 grep UUID "${_workflowId}" 查找最新报告。`
-			);
-		}
-
 		// ── Capture reviewer work summary as supplementary chain context ──
-		// Reuse already-parsed extractedOutput to avoid double parsing
+		// The old review-chain-context entries ("代码审查反馈"/"精简审查反馈") based on
+		// _workflowFileChanges were removed in v0.6.0 because they duplicated information
+		// from the review report file and polluted chain context with file-path-heavy data.
+		// Reuse already-parsed extractedOutput to avoid double parsing.
 		if (extractedOutput) {
 			updateChainContext(`审查工作总结`,
 				`审查者已完成审查，以下是其审查总结：\n\n${extractedOutput}`

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ghyper9023/pi-dev-workflow",
-  "version": "0.5.1",
+  "version": "0.6.1",
   "keywords": [
     "pi-package"
   ],