@rexvane/piweb 0.3.12 → 0.3.14

package/src/lib/session-import/claude.ts

@@ -0,0 +1,229 @@
+/**
+ * Claude Code（~/.claude/projects/<slug>/*.jsonl）
+ *
+ * 每行自带 type，无统一头行；消息行是 {type:user|assistant, message:{role,model,content[],usage,stop_reason}}。
+ * content 块有 text / thinking / tool_use / tool_result / image；工具结果出现在后续 user 行里，
+ * 靠 tool_use_id 与前面的 tool_use 配对（这一步交给 writer 兜底修复）。
+ */
+import path from "node:path";
+import fs from "node:fs/promises";
+import { eachJsonLine, listFiles, readHead, reachedLimit, sortByRecency } from "./files";
+import { claudeRoot } from "./paths";
+import type { ExternalSessionSummary, ImportedBlock, ImportedEntry, ImportedMessage, ImportedSession, ImportSourceModule, ScanOptions } from "./types";
+import { cleanTitle, compactBlocks, outputBlocks, textBlock, toEpochMs } from "./types";
+
+const ROOT = claudeRoot;
+
+/** 合成信封（命令输出、系统提醒等）不是真人输入，导入时跳过 */
+const SYNTHETIC_PREFIXES = ["<command-name>", "<command-message>", "<local-command-stdout>", "<local-command-caveat>", "<system-reminder>", "<user_info>", "<bash-input>", "<bash-stdout>"];
+
+const isSynthetic = (text: string) => SYNTHETIC_PREFIXES.some((prefix) => text.startsWith(prefix));
+
+/**
+ * 子代理的转写不能当作会话：Claude Code 把它们放在
+ * `<项目>/<会话 uuid>/subagents/` 下（含更深的 workflows/wf_<id>/ 层，本机 236 个子代理文件 vs 70 个主会话）。
+ * 那是主代理派出去的活，不是你在用的那条对话。
+ * 这些文件里的记录全是 isSidechain，本来也会被 isMessageLine 挡掉，
+ * 这里再按路径显式排除一次：意图更清楚，也不怕以后格式变化。
+ */
+const isSubagentTranscript = (file: string) => file.split(/[/\\]/).includes("subagents");
+
+function isMessageLine(record: Record<string, unknown>): boolean {
+	return (record.type === "user" || record.type === "assistant") && record.isSidechain !== true && Boolean(record.message);
+}
+
+function projectPathFrom(records: Record<string, unknown>[]): string | undefined {
+	for (const record of records) {
+		if (typeof record.cwd === "string" && record.cwd.trim()) return record.cwd;
+	}
+	return undefined;
+}
+
+function modelFrom(records: Record<string, unknown>[]): string | undefined {
+	for (const record of records) {
+		const message = record.message as { model?: unknown } | undefined;
+		if (typeof message?.model === "string" && message.model.trim()) return message.model;
+	}
+	return undefined;
+}
+
+/** 标题：优先会话摘要行，其次首条真人用户消息（Claude Code 不写标题字段） */
+function titleFrom(records: Record<string, unknown>[]): string | undefined {
+	for (const record of records) {
+		if (record.type === "summary" && typeof record.summary === "string" && record.summary.trim()) return record.summary.trim();
+	}
+	for (const record of records) {
+		if (record.type !== "user") continue;
+		const plain = userTextOf((record.message as { content?: unknown } | undefined)?.content);
+		if (plain && !isSynthetic(plain)) return plain;
+	}
+	return undefined;
+}
+
+function userTextOf(content: unknown): string {
+	const items = Array.isArray(content) ? content : [content];
+	return items
+		.map((item) => (typeof item === "string" ? item : typeof (item as { text?: unknown })?.text === "string" ? ((item as { text: string }).text) : ""))
+		.join("\n")
+		.replace(/\s+/g, " ")
+		.trim();
+}
+
+/** assistant 消息的 content 数组 */
+function assistantBlocks(content: unknown): ImportedBlock[] {
+	if (!Array.isArray(content)) return compactBlocks([textBlock(typeof content === "string" ? content : undefined)]);
+	return compactBlocks(content.flatMap((raw): Array<ImportedBlock | null> => {
+		const block = raw as Record<string, unknown>;
+		if (block?.type === "text") return [textBlock(block.text)];
+		if (block?.type === "thinking") return [typeof block.thinking === "string" && block.thinking ? { type: "thinking", thinking: block.thinking } : null];
+		if (block?.type === "tool_use") {
+			const id = typeof block.id === "string" ? block.id : "";
+			const name = typeof block.name === "string" ? block.name : "";
+			return id && name ? [{ type: "toolCall" as const, id, name, arguments: block.input ?? {} }] : [];
+		}
+		if (block?.type === "image") {
+			const source = block.source as Record<string, unknown> | undefined;
+			if (source?.type === "base64" && typeof source.data === "string" && typeof source.media_type === "string") {
+				return [{ type: "image", data: source.data, mimeType: source.media_type }];
+			}
+		}
+		return [];
+	}));
+}
+
+/** user 行可能同时含文本与工具结果：文本进用户消息，工具结果单列 */
+function splitUserContent(content: unknown): { text: ImportedBlock[]; results: Array<{ toolCallId: string; content: ImportedBlock[]; isError: boolean }> } {
+	const text: Array<ImportedBlock | null> = [];
+	const results: Array<{ toolCallId: string; content: ImportedBlock[]; isError: boolean }> = [];
+	const items = Array.isArray(content) ? content : [content];
+	for (const raw of items) {
+		if (typeof raw === "string") {
+			text.push(textBlock(raw));
+			continue;
+		}
+		const block = raw as Record<string, unknown>;
+		if (block?.type === "text") text.push(textBlock(block.text));
+		else if (block?.type === "tool_result" && typeof block.tool_use_id === "string") {
+			results.push({ toolCallId: block.tool_use_id, content: outputBlocks(block.content), isError: block.is_error === true });
+		} else if (block?.type === "image") {
+			const source = block.source as Record<string, unknown> | undefined;
+			if (source?.type === "base64" && typeof source.data === "string" && typeof source.media_type === "string") {
+				text.push({ type: "image", data: source.data, mimeType: source.media_type });
+			}
+		}
+	}
+	return { text: compactBlocks(text), results };
+}
+
+export const claudeSource: ImportSourceModule = {
+	id: "claude",
+
+	async scan(options?: ScanOptions): Promise<ExternalSessionSummary[]> {
+		const limit = options?.limit ?? 0;
+		const candidates = await sortByRecency(await listFiles(ROOT(), { match: (name) => name.endsWith(".jsonl") }));
+		const out: ExternalSessionSummary[] = [];
+		for (const { file, mtime } of candidates) {
+			if (reachedLimit(out.length, limit)) break; // 只读最近的若干条：后面的文件根本不打开
+			if (isSubagentTranscript(file)) continue; // 主代理派给子代理的活不算会话
+			let head: string;
+			try {
+				head = await readHead(file);
+			} catch {
+				continue;
+			}
+			const records = [...eachJsonLine(head)].filter((record) => record.type !== undefined).slice(0, 200);
+			const first = records.find(isMessageLine);
+			if (!first) continue; // 只有元信息、没有消息的文件不列
+			out.push({
+				source: "claude",
+				externalId: path.basename(file, ".jsonl"),
+				title: cleanTitle(titleFrom(records)),
+				projectPath: projectPathFrom(records),
+				model: modelFrom(records),
+				createdAt: toEpochMs(first.timestamp) ?? mtime,
+				updatedAt: mtime,
+				// 消息数需要整读，扫描阶段不统计（界面显示 —）
+				location: file,
+			});
+		}
+		return out;
+	},
+
+	async read(summary: ExternalSessionSummary): Promise<ImportedSession | null> {
+		let text: string;
+		try {
+			text = await fs.readFile(summary.location, "utf8");
+		} catch {
+			return null;
+		}
+		const entries: ImportedEntry[] = [];
+		const skipped: string[] = [];
+		let sidechains = 0;
+		let synthetic = 0;
+		let projectPath = summary.projectPath;
+		let model = summary.model;
+		let provider: string | undefined;
+
+		for (const record of eachJsonLine(text)) {
+			if (record.isSidechain === true) {
+				sidechains += 1;
+				continue;
+			}
+			if (typeof record.cwd === "string" && record.cwd.trim()) projectPath = projectPath ?? record.cwd;
+			const message = record.message as Record<string, unknown> | undefined;
+			if (!message) continue;
+			if (typeof message.model === "string" && message.model.trim()) model = message.model;
+			if (typeof message.provider === "string" && message.provider.trim()) provider = message.provider;
+			const timestamp = toEpochMs(record.timestamp) ?? toEpochMs(message.timestamp) ?? summary.createdAt ?? Date.now();
+
+			if (record.type === "assistant") {
+				const content = assistantBlocks(message.content);
+				if (!content.length) continue;
+				const usage = message.usage as Record<string, unknown> | undefined;
+				entries.push({
+					type: "message",
+					message: {
+						role: "assistant",
+						content,
+						timestamp,
+						...(typeof message.model === "string" ? { model: message.model } : {}),
+						...(typeof message.stop_reason === "string" && message.stop_reason ? { stopReason: message.stop_reason } : {}),
+						...(usage
+							? {
+									usage: {
+										input: Number(usage.input_tokens) || 0,
+										output: Number(usage.output_tokens) || 0,
+										cacheRead: Number(usage.cache_read_input_tokens) || 0,
+										cacheWrite: Number(usage.cache_creation_input_tokens) || 0,
+									},
+								}
+							: {}),
+					},
+				});
+				continue;
+			}
+
+			if (record.type === "user") {
+				const { text: textBlocks, results } = splitUserContent((message as { content?: unknown }).content);
+				const plain = textBlocks.filter((block) => block.type === "text").map((block) => block.text).join("\n");
+				const keepText = textBlocks.filter((block) => block.type !== "text");
+				if (plain && !isSynthetic(plain.trim()) && !/^Caveat:/i.test(plain.trim())) {
+					entries.push({ type: "message", message: { role: "user", content: [...keepText, { type: "text", text: plain }], timestamp } as ImportedMessage });
+				} else if (plain) {
+					synthetic += 1;
+				}
+				for (const result of results) {
+					entries.push({
+						type: "message",
+						message: { role: "toolResult", toolCallId: result.toolCallId, content: result.content, isError: result.isError || undefined, timestamp } as ImportedMessage,
+					});
+				}
+			}
+		}
+
+		if (sidechains) skipped.push(`跳过 ${sidechains} 条侧链（子代理）记录`);
+		if (synthetic) skipped.push(`跳过 ${synthetic} 条合成输入（命令回显/系统提醒）`);
+		if (!entries.some((entry) => entry.type === "message")) return null;
+		return { summary: { ...summary, projectPath, model, provider }, entries, skipped };
+	},
+};

package/src/lib/session-import/codex.ts

@@ -0,0 +1,223 @@
+/**
+ * Codex CLI（~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl）
+ *
+ * 每行统一 `{timestamp, type, payload}`；payload.type = session_meta / response_item / event_msg …
+ * response_item.payload 里才有对话：message / reasoning / function_call / custom_tool_call(+output)。
+ * 注意首行可达十几 KB（含 base_instructions），必须读到换行再 parse。
+ */
+import path from "node:path";
+import fs from "node:fs/promises";
+import { eachJsonLine, listFiles, readHead, reachedLimit, sortByRecency } from "./files";
+import { codexRoot } from "./paths";
+import type { ExternalSessionSummary, ImportedBlock, ImportedEntry, ImportedSession, ImportSourceModule, ScanOptions } from "./types";
+import { cleanTitle, compactBlocks, outputBlocks, parseArguments, textBlock, toEpochMs } from "./types";
+
+const ROOT = codexRoot;
+
+/**
+ * 合成输入前缀（与参考实现同口径：精确前缀，不用 "#" 通配，避免误杀用户粘贴的 markdown）
+ */
+const SYNTHETIC_PREFIXES = [
+	"<",
+	"# AGENTS.md",
+	"# Context from my IDE setup",
+	"# In app browser:",
+	"# Browser comments:",
+	"# Files mentioned by the user:",
+	"# Diff comments:",
+	"# Selected text:",
+	"# Review findings:",
+	"You are Codex",
+];
+
+const isSynthetic = (text: string) => SYNTHETIC_PREFIXES.some((prefix) => text.startsWith(prefix));
+
+interface Meta {
+	cwd?: string;
+	createdAt?: number;
+	model?: string;
+	provider?: string;
+	/** true = 这条 rollout 是别的线程派出来的（子代理），不是主对话 */
+	subagent?: boolean;
+}
+
+/**
+ * 子代理线程的判定，依据 session_meta.payload（本机 51 个 rollout 里有 22 个是子代理）：
+ *  - parent_thread_id：它的存在就说明这条线程由另一条线程派生
+ *  - source 是对象（`{"subagent":{…}}`）而主线程是字符串 "cli" / "vscode"
+ *  - thread_source：主线程恒为 "user"，子代理是 "subagent" / "guardian_review"
+ * 三个条件在本机实测完全一致（22 个子代理全部同时命中，29 个主线程一个都不命中）。
+ */
+function isSubagentMeta(payload: Record<string, unknown>): boolean {
+	if (typeof payload.parent_thread_id === "string" && payload.parent_thread_id) return true;
+	if (payload.source && typeof payload.source === "object") return true;
+	return typeof payload.thread_source === "string" && payload.thread_source !== "user";
+}
+
+function metaFrom(records: Record<string, unknown>[]): Meta {
+	const meta: Meta = {};
+	for (const record of records) {
+		if (record.type !== "session_meta") continue;
+		const payload = record.payload as Record<string, unknown> | undefined;
+		if (!payload) continue;
+		if (typeof payload.cwd === "string") meta.cwd = payload.cwd;
+		meta.createdAt = toEpochMs(payload.timestamp) ?? toEpochMs(record.timestamp);
+		if (typeof payload.model_provider === "string") meta.provider = payload.model_provider;
+		if (typeof payload.model === "string") meta.model = payload.model;
+		if (isSubagentMeta(payload)) meta.subagent = true;
+	}
+	// 模型有时只在 turn_context 里出现
+	for (const record of records) {
+		if (record.type !== "turn_context") continue;
+		const payload = record.payload as Record<string, unknown> | undefined;
+		if (typeof payload?.model === "string") meta.model = payload.model;
+	}
+	return meta;
+}
+
+function messageBlocks(payload: Record<string, unknown>): ImportedBlock[] {
+	const content = payload.content;
+	if (!Array.isArray(content)) return compactBlocks([textBlock(typeof content === "string" ? content : undefined)]);
+	return compactBlocks(content.flatMap((raw): Array<ImportedBlock | null> => {
+		const block = raw as Record<string, unknown>;
+		if (typeof block?.text === "string") return [textBlock(block.text)];
+		return [];
+	}));
+}
+
+function reasoningText(payload: Record<string, unknown>): ImportedBlock | null {
+	const summary = payload.summary;
+	if (typeof summary === "string") return summary.trim() ? { type: "thinking", thinking: summary } : null;
+	if (Array.isArray(summary)) {
+		const text = summary.map((item) => (item as { text?: unknown })?.text).filter((value): value is string => typeof value === "string").join("\n");
+		return text.trim() ? { type: "thinking", thinking: text } : null;
+	}
+	return null;
+}
+
+/** 标题：Codex 不写标题字段，用首条真人用户消息（rollout 头几十行里就有） */
+function titleFrom(records: Record<string, unknown>[]): string | undefined {
+	for (const record of records) {
+		if (record.type !== "response_item") continue;
+		const payload = record.payload as Record<string, unknown> | undefined;
+		if (payload?.type !== "message" || payload.role !== "user") continue;
+		const blocks = messageBlocks(payload);
+		const plain = blocks.filter((block) => block.type === "text").map((block) => block.text).join("\n").replace(/\s+/g, " ").trim();
+		if (plain && !isSynthetic(plain)) return plain;
+	}
+	return undefined;
+}
+
+export const codexSource: ImportSourceModule = {
+	id: "codex",
+
+	async scan(options?: ScanOptions): Promise<ExternalSessionSummary[]> {
+		const limit = options?.limit ?? 0;
+		const candidates = await sortByRecency(await listFiles(ROOT(), { match: (name) => name.startsWith("rollout-") && name.endsWith(".jsonl") }));
+		const out: ExternalSessionSummary[] = [];
+		for (const { file, mtime } of candidates) {
+			if (reachedLimit(out.length, limit)) break; // 只读最近的若干条
+			let head: string;
+			try {
+				head = await readHead(file);
+			} catch {
+				continue;
+			}
+			const records = [...eachJsonLine(head)].slice(0, 300);
+			const meta = metaFrom(records);
+			if (meta.subagent) continue; // 主代理派给子代理的活，不算会话
+			const hasMessage = records.some((record) => record.type === "response_item");
+			if (!hasMessage) continue;
+			out.push({
+				source: "codex",
+				externalId: path.basename(file, ".jsonl"),
+				title: cleanTitle(titleFrom(records)),
+				projectPath: meta.cwd,
+				model: meta.model,
+				provider: meta.provider,
+				createdAt: meta.createdAt ?? mtime,
+				updatedAt: mtime,
+				location: file,
+			});
+		}
+		return out;
+	},
+
+	async read(summary: ExternalSessionSummary): Promise<ImportedSession | null> {
+		let text: string;
+		try {
+			text = await fs.readFile(summary.location, "utf8");
+		} catch {
+			return null;
+		}
+		const entries: ImportedEntry[] = [];
+		const skipped: string[] = [];
+		let synthetic = 0;
+		let reasoningSkipped = 0;
+		let meta: Meta = { cwd: summary.projectPath, createdAt: summary.createdAt, model: summary.model, provider: summary.provider };
+
+		for (const record of eachJsonLine(text)) {
+			if (record.type === "session_meta") {
+				meta = { ...meta, ...metaFrom([record]) };
+				continue;
+			}
+			if (record.type !== "response_item") continue;
+			const payload = record.payload as Record<string, unknown> | undefined;
+			if (!payload) continue;
+			const timestamp = toEpochMs(record.timestamp) ?? toEpochMs(payload.timestamp) ?? meta.createdAt ?? Date.now();
+
+			switch (payload.type) {
+				case "message": {
+					const role = payload.role;
+					if (role === "developer") continue; // 系统提示，不进对话
+					const blocks = messageBlocks(payload);
+					const plain = blocks.filter((block) => block.type === "text").map((block) => block.text).join("\n").trim();
+					if (role === "user") {
+						if (!plain || isSynthetic(plain)) {
+							synthetic += 1;
+							continue;
+						}
+						entries.push({ type: "message", message: { role: "user", content: blocks, timestamp } });
+					} else if (role === "assistant") {
+						if (!blocks.length) continue;
+						entries.push({ type: "message", message: { role: "assistant", content: blocks, timestamp, ...(meta.model ? { model: meta.model } : {}) } });
+					}
+					continue;
+				}
+				case "reasoning": {
+					// 有摘要就保留成 thinking；只有加密内容时跳过
+					const thinking = reasoningText(payload);
+					if (!thinking) {
+						reasoningSkipped += 1;
+						continue;
+					}
+					entries.push({ type: "message", message: { role: "assistant", content: [thinking], timestamp, ...(meta.model ? { model: meta.model } : {}) } });
+					continue;
+				}
+				case "function_call":
+				case "custom_tool_call": {
+					const callId = typeof payload.call_id === "string" ? payload.call_id : "";
+					const name = typeof payload.name === "string" ? payload.name : "";
+					if (!callId || !name) continue;
+					const args = parseArguments(payload.arguments ?? payload.input ?? {});
+					entries.push({ type: "message", message: { role: "assistant", content: [{ type: "toolCall", id: callId, name, arguments: args }], timestamp, stopReason: "toolUse", ...(meta.model ? { model: meta.model } : {}) } });
+					continue;
+				}
+				case "function_call_output":
+				case "custom_tool_call_output": {
+					const callId = typeof payload.call_id === "string" ? payload.call_id : "";
+					if (!callId) continue;
+					entries.push({ type: "message", message: { role: "toolResult", toolCallId: callId, content: outputBlocks(payload.output ?? payload.content), timestamp } });
+					continue;
+				}
+				default:
+					continue;
+			}
+		}
+
+		if (synthetic) skipped.push(`跳过 ${synthetic} 条合成输入`);
+		if (reasoningSkipped) skipped.push(`跳过 ${reasoningSkipped} 条仅含加密内容的推理记录`);
+		if (!entries.some((entry) => entry.type === "message")) return null;
+		return { summary: { ...summary, ...meta, projectPath: meta.cwd, model: meta.model, provider: meta.provider }, entries, skipped };
+	},
+};

package/src/lib/session-import/dsh-zstd.ts

@@ -0,0 +1,69 @@
+/**
+ * dsh 的会话是 **多帧拼接的 zstd**（`session.jsonl.zstd`）：`zstdDecompressSync` 一次只解出第一帧，
+ * 必须按 zstd magic（28 B5 2F FD）切帧后逐帧解压再拼起来，否则只能拿到第一行。
+ */
+import fs from "node:fs/promises";
+import { zstdDecompressSync } from "node:zlib";
+
+const MAGIC = Buffer.from([0x28, 0xb5, 0x2f, 0xfd]);
+
+/** 解压多帧 zstd；帧边界只是"看起来像 magic"的位置，解压失败就把这片并入下一帧 */
+export function decodeZstdFrames(raw: Buffer): Buffer {
+	if (raw.length < MAGIC.length || !raw.subarray(0, MAGIC.length).equals(MAGIC)) {
+		// 不是多帧流（可能单帧或未压缩）：交给单帧解压，失败则原样当作文本
+		try {
+			return zstdDecompressSync(raw);
+		} catch {
+			return raw;
+		}
+	}
+	const cuts: number[] = [];
+	for (let i = 0; (i = raw.indexOf(MAGIC, i)) !== -1; i += MAGIC.length) cuts.push(i);
+	cuts.push(raw.length);
+	const parts: Buffer[] = [];
+	let start = 0;
+	for (let i = 1; i < cuts.length; i += 1) {
+		try {
+			parts.push(zstdDecompressSync(raw.subarray(cuts[start], cuts[i])));
+			start = i;
+		} catch {
+			// 误判的帧边界：与下一帧合并后重试
+		}
+	}
+	return Buffer.concat(parts);
+}
+
+export async function readZstdJsonl(file: string): Promise<string> {
+	return decodeZstdFrames(await fs.readFile(file)).toString("utf8");
+}
+
+/**
+ * 只解出前面的行即可判断会话是否值得导入（扫描用）：逐帧解压，凑够若干帧就停，
+ * 避免为了列个表把 39MB 全解出来。
+ *
+ * `complete` 表示"解出来的就是全部内容"——真实文件能到 7559 帧，前 40 帧只覆盖开头一小段，
+ * 所以只有 complete 为真时才能断言"这个会话里面没有消息"。
+ */
+export async function readZstdHead(file: string, maxFrames = 40): Promise<{ text: string; complete: boolean }> {
+	const raw = await fs.readFile(file);
+	const cuts: number[] = [];
+	for (let i = 0; (i = raw.indexOf(MAGIC, i)) !== -1; i += MAGIC.length) cuts.push(i);
+	if (!cuts.length) {
+		try {
+			return { text: zstdDecompressSync(raw).toString("utf8").slice(0, 64 * 1024), complete: true };
+		} catch {
+			return { text: "", complete: false };
+		}
+	}
+	cuts.push(raw.length);
+	const total = cuts.length - 1;
+	const out: string[] = [];
+	for (let i = 1; i < cuts.length && i <= maxFrames; i += 1) {
+		try {
+			out.push(zstdDecompressSync(raw.subarray(cuts[i - 1], cuts[i])).toString("utf8"));
+		} catch {
+			// 边界误判：跳过这片
+		}
+	}
+	return { text: out.join(""), complete: total <= maxFrames };
+}