lark-remote 0.5.19 → 0.5.21

package/dist/session/claude/sessions.js

@@ -7,6 +7,7 @@ import { samePath, encodeProjectDirName } from '../../platform/path.js';
 import { extractContentBlocks } from '../common/content-blocks.js';
 import { scanJsonlOnce, tailOffsetAfter, scanTailEvents } from '../common/two-pass.js';
 import { getLogger } from '../../logger/index.js';
+import { logNonEnoentDebug } from '../session-store.js';
 import { UsageAccumulator, contextWindowOccupancy, cumulativeUsageFields, } from '../common/usage-accumulator.js';
 import { SessionIndex, extractText, isTaskNotificationText, CLAUDE_MAPPING, } from './session-index.js';
 /**
@@ -19,7 +20,7 @@ import { SessionIndex, extractText, isTaskNotificationText, CLAUDE_MAPPING, } fr
  * Note: this encoding is LOSSY (`my_data/foo` and `my-data/foo` both encode to
  * `my-data-foo`). It is only safe to use the result to **locate** files; never
  * to **decode** a directory name back to a cwd. The cwd must always be read
- * from the JSONL content (regression 2026-06-21 /resume & /cd 路径错乱).
+ * from the JSONL content (regression: /resume & /cd 路径错乱).
  */
 function projectDirForCwd(cwd, projectsDir) {
     return path.join(projectsDir, encodeProjectDirName(cwd));
@@ -32,7 +33,7 @@ function projectDirForCwd(cwd, projectsDir) {
  * MULTIPLE cwd values (pre- and post-relocate) — accept when ANY `cwd` field
  * matches. Early-stops on the first match (findJsonlLine), keeping the
  * single-pass parse-count anchors intact on the fallback path.
- * Regression 2026-08-04 EnterWorktree relocate.
+ * Needed for EnterWorktree relocate.
  */
 function fileContainsCwd(filePath, cwd) {
     return (findJsonlLine(filePath, (l) => {
@@ -50,7 +51,7 @@ function defaultProjectsDir() {
     return path.join(os.homedir(), '.claude', 'projects');
 }
 /**
- * EnterWorktree relocate support (2026-08-04): Claude Code MOVES the
+ * EnterWorktree relocate support: Claude Code MOVES the
  * transcript file to the new cwd's project dir mid-session, so the file may
  * not exist under the requested cwd's encoded dir. Fall back to locating
  * `<sessionId>.jsonl` across all project dirs. Deterministic: entries are
@@ -148,7 +149,7 @@ export function getNewestSession(cwd, opts = {}) {
  *    JSONL output continuously. A file whose mtime is older than
  *    `STALE_MS` cannot be from a running process. This catches the
  *    common Claude CLI 2.x case where the tail is `last-prompt` / `mode`
- *    (no `type:"result"`) and the process exited long ago (regression 2026-06-21).
+ *    (no `type:"result"`) and the process exited long ago.
  *
  * 2. **last-line terminal marker**: if the last JSON line is `result`
  *    (older CLI) or any explicit terminal event, the session is done
@@ -176,7 +177,7 @@ function hasRecentFileActivity(filePath, mtimeMs, now = Date.now()) {
         // Explicit terminal markers — regardless of mtime. Claude CLI 2.x uses
         // `last-prompt` / `mode` (no `result` event) to mark turn end, so a
         // 24-minute-old session with a last-prompt tail is already done
-        // (regression 2026-06-21: previously falsely reported as active).
+        // (previously falsely reported as active).
         if (typeof obj.type === 'string' && CLAUDE_TERMINAL_TYPES.has(obj.type))
             return false;
     }
@@ -190,9 +191,7 @@ function hasRecentFileActivity(filePath, mtimeMs, now = Date.now()) {
  * 否则「文件不可读」会被当成「会话不活跃」。
  */
 function logStatFailure(filePath, err) {
-    if (err.code === 'ENOENT')
-        return;
-    getLogger().debug(`[claude-session-reader] stat ${filePath} failed: ${err?.message ?? String(err)}`);
+    logNonEnoentDebug(err, (m) => `[claude-session-reader] stat ${filePath} failed: ${m}`);
 }
 /**
  * Check if a Claude session (by sessionId + cwd) is still active.
@@ -273,115 +272,42 @@ function isRealUserPrompt(obj) {
     return true;
 }
 /**
- * first-pass scan: stream the file once via `scanJsonlLines`
- * (no `string[]` materialized), parse each line exactly once, and collect
- * all metadata needed by readSessionContent (usage, tailOffset, aiTitle,
- * recap, lastUserMessage). Does NOT retain parsed objects — the events tail
- * is re-parsed separately by `extractEventsFromTail` so memory for both raw
- * line strings and parsed objects is O(tail), not O(whole file).
+ * 两条采样逐字段取 max（纯函数；UsageAccumulator 的 per-field max 语义）。
+ *
+ * RED LINE（token 口径单源的一部分）：一条 streaming assistant 消息跨多行
+ * jsonl、共享同一 message id；部分后端（第三方网关，如 DeepSeek）在早期占位
+ * 行（空 thinking 块）写全零 usage，真实用量只在带 stop_reason 的末行。
+ * first-occurrence-wins 会把零记录钉死、丢掉真实用量（regression 2026-08-19：
+ * auto-resume 卡 per-run 统计全 0）。逐字段 max 与行序无关地重建完整用量；
+ * 官方 API 每行重复相同 usage，max 是 no-op。
+ *
+ * 产生：合并后的一条采样。依赖：入参。不改任何外部状态。
  */
-function scalarScan(filePath) {
+export function maxUsageRecord(prev, usage, turnSeq) {
+    return {
+        input: Math.max(prev?.input ?? 0, usage.input_tokens ?? 0),
+        output: Math.max(prev?.output ?? 0, usage.output_tokens ?? 0),
+        cacheRead: Math.max(prev?.cacheRead ?? 0, usage.cache_read_input_tokens ?? 0),
+        cacheCreation: Math.max(prev?.cacheCreation ?? 0, usage.cache_creation_input_tokens ?? 0),
+        turn: prev?.turn ?? turnSeq,
+    };
+}
+/**
+ * usage 装配（纯函数）：每个 message id 一条最优记录、按首次出现顺序喂
+ * UsageAccumulator（Map 保序），acc.last = 末条 message 的完整用量，totals
+ * 让每条 API 响应恰好多计一次。per-run flow 只取末个真人 turn 的记录——
+ * 工具往返的 turn 有多条 assistant 消息（一次模型请求一条），只取 `last`
+ * 会低报。
+ *
+ * 产生：AgentSessionUsage。依赖：入参 records + 标量。不改任何外部状态
+ * （UsageAccumulator 在函数内新建）。
+ */
+export function buildClaudeUsage(records, opts) {
     const acc = new UsageAccumulator();
-    // Per-message-id best usage record (per-field max). A streaming assistant
-    // message spans multiple jsonl lines sharing the same message id; some
-    // backends (third-party gateways, e.g. DeepSeek) write all-zero usage on
-    // the early placeholder lines (empty thinking block) and the real usage
-    // only on the final line carrying stop_reason. First-occurrence-wins dedup
-    // would pin the zero record and discard the real one (regression
-    // 2026-08-19: auto-resume card per-run stats all 0). Per-field max
-    // reconstructs the full usage regardless of line order — on the official
-    // API every line repeats identical usage, so max is a no-op there.
-    const usageByMessageId = new Map();
-    let lastPostTokens = 0;
-    let tailOffset = -1;
-    /** 真人 prompt 序号；末个 turn = 当前值。sidechain/注入/工具结果不递增。 */
-    let turnSeq = 0;
-    let lastAiTitle;
-    let lastRecap;
-    let lastUserMessage;
-    scanJsonlOnce(filePath, (obj, line, offset) => {
-        // --- usage aggregation (was aggregateSessionUsage) ---
-        if (obj.type === 'assistant') {
-            const msg = obj.message;
-            const usage = msg?.usage;
-            if (usage) {
-                const msgId = msg?.id;
-                if (msgId) {
-                    const prev = usageByMessageId.get(msgId);
-                    usageByMessageId.set(msgId, {
-                        input: Math.max(prev?.input ?? 0, usage.input_tokens ?? 0),
-                        output: Math.max(prev?.output ?? 0, usage.output_tokens ?? 0),
-                        cacheRead: Math.max(prev?.cacheRead ?? 0, usage.cache_read_input_tokens ?? 0),
-                        cacheCreation: Math.max(prev?.cacheCreation ?? 0, usage.cache_creation_input_tokens ?? 0),
-                        turn: prev?.turn ?? turnSeq,
-                    });
-                }
-            }
-        }
-        if (obj.type === 'system' && obj.subtype === 'compact_boundary') {
-            acc.bumpCompact();
-            const cm = obj.compactMetadata;
-            if (cm?.postTokens) {
-                lastPostTokens = cm.postTokens;
-            }
-        }
-        // --- last user message offset ---
-        // Record the byte offset where the tail should start: the beginning of
-        // the NEXT line after this user message (see tailOffsetAfter).
-        if (obj.type === 'user') {
-            const msg = obj.message;
-            if (msg?.role === 'user') {
-                tailOffset = tailOffsetAfter(line, offset);
-            }
-        }
-        // --- ai-title extraction ---
-        if (obj.type === 'ai-title' && typeof obj.aiTitle === 'string') {
-            lastAiTitle = obj.aiTitle;
-        }
-        // --- recap extraction (last isCompactSummary wins) ---
-        if (obj.type === 'user' && obj.isCompactSummary === true) {
-            const msg = obj.message;
-            if (msg?.content) {
-                const recapContent = extractText(msg.content);
-                if (recapContent) {
-                    lastRecap = recapContent;
-                }
-            }
-        }
-        // --- last user message for displayTitle ---
-        // Skip: 1) tool_result content, 2) isCompactSummary (recap),
-        //       3) isMeta (skill injection / command echo / project rules)
-        if (obj.type === 'user' && !obj.isCompactSummary && !obj.isMeta) {
-            const msg = obj.message;
-            if (msg?.role === 'user' && msg.content) {
-                const content = msg.content;
-                const isToolResult = Array.isArray(content) &&
-                    content.length > 0 &&
-                    content[0]?.type === 'tool_result';
-                if (!isToolResult) {
-                    const userContent = extractText(content);
-                    if (userContent && !isTaskNotificationText(userContent)) {
-                        lastUserMessage = userContent;
-                    }
-                }
-            }
-        }
-        // --- turn boundary for per-run usage ---
-        // 真人 prompt 开启新 turn；注入/工具结果/sidechain 不递增。usage 记录在
-        // 解析到 assistant 行时已经带上当前 turnSeq，末尾只汇总末个 turn。
-        if (isRealUserPrompt(obj))
-            turnSeq++;
-    });
-    // Feed one best record per message id, in first-appearance order (Map
-    // preserves insertion order), so acc.last is the LAST message's full usage
-    // and totals sum each API response exactly once. Per-run flow is only the
-    // records belonging to the last real-user turn — a tool-looping turn has
-    // several assistant messages (one per model request), so `last` alone
-    // under-reports it.
     const turn = { input: 0, output: 0, cacheRead: 0, cacheCreation: 0 };
-    for (const record of usageByMessageId.values()) {
+    for (const record of records) {
         acc.add(record);
-        if (turnSeq > 0 && record.turn === turnSeq) {
+        if (opts.turnSeq > 0 && record.turn === opts.turnSeq) {
             turn.input += record.input;
             turn.output += record.output;
             turn.cacheRead += record.cacheRead;
@@ -393,7 +319,7 @@ function scalarScan(filePath) {
     const l = acc.last;
     // 没有真人 user prompt（旧数据/截断文件）→ 无法划分 turn，回退末条采样，
     // 与 dsh/codex「无 turn 标记退回末条」一致；不能把整份文件当成一轮。
-    if (turnSeq === 0 && l) {
+    if (opts.turnSeq === 0 && l) {
         turn.input = l.input;
         turn.output = l.output;
         turn.cacheRead = l.cacheRead;
@@ -406,7 +332,7 @@ function scalarScan(filePath) {
     // the input window). Take the max so a pre-compact turn larger than the
     // post-compact size is not under-reported.
     const lastWindow = l ? contextWindowOccupancy(l) : 0;
-    let contextLength = lastPostTokens;
+    let contextLength = opts.lastPostTokens;
     if (contextLength === 0 || lastWindow > contextLength) {
         contextLength = lastWindow;
     }
@@ -415,22 +341,134 @@ function scalarScan(filePath) {
     // `total_token_usage`=session 累计）——否则 jsonl 兜底路径会把 session 累计
     // 当成"本 run"显示，与卡片"本 run ≤ 累计"不变量冲突（累计反而更小）。
     const lastTotal = turn.input + turn.output + turn.cacheRead + turn.cacheCreation;
-    const usage = {
+    return {
         inputTokens: turn.input,
         outputTokens: turn.output,
         contextLength,
-        compactCount: acc.compactCount,
+        compactCount: opts.compactCount,
         cacheReadTokens: turn.cacheRead > 0 ? turn.cacheRead : undefined,
         cacheCreationTokens: turn.cacheCreation > 0 ? turn.cacheCreation : undefined,
         totalTokens: lastTotal,
         ...cumulativeUsageFields(t),
     };
+}
+function newScalarScanState() {
+    // Per-message-id best usage record (per-field max). A streaming assistant
+    // message spans multiple jsonl lines sharing the same message id; some
+    // backends (third-party gateways, e.g. DeepSeek) write all-zero usage on
+    // the early placeholder lines (empty thinking block) and the real usage
+    // only on the final line carrying stop_reason. First-occurrence-wins dedup
+    // would pin the zero record and discard the real one (auto-resume card
+    // per-run stats all 0). Per-field max reconstructs the full usage
+    // regardless of line order — on the official API every line repeats
+    // identical usage, so max is a no-op there.
     return {
-        tailOffset,
-        usage,
-        aiTitle: lastAiTitle,
-        recap: lastRecap,
-        lastUserMessage,
+        usageByMessageId: new Map(),
+        lastPostTokens: 0,
+        compactCount: 0,
+        tailOffset: -1,
+        turnSeq: 0,
+    };
+}
+/** assistant 行 → 按 message id 逐字段 max 合并 usage（was aggregateSessionUsage）。 */
+function collectAssistantUsage(state, obj) {
+    if (obj.type !== 'assistant')
+        return;
+    const msg = obj.message;
+    const usage = msg?.usage;
+    if (!usage)
+        return;
+    const msgId = msg?.id;
+    if (!msgId)
+        return;
+    state.usageByMessageId.set(msgId, maxUsageRecord(state.usageByMessageId.get(msgId), usage, state.turnSeq));
+}
+/** compact_boundary → 计数 + 压缩后 token 水位。 */
+function collectCompactBoundary(state, obj) {
+    if (obj.type !== 'system' || obj.subtype !== 'compact_boundary')
+        return;
+    state.compactCount++;
+    const cm = obj.compactMetadata;
+    if (cm?.postTokens) {
+        state.lastPostTokens = cm.postTokens;
+    }
+}
+/** 记录 tail 起点：本 user 行之后的下一行字节偏移（见 tailOffsetAfter）。 */
+function collectTailOffset(state, obj, line, offset) {
+    if (obj.type !== 'user')
+        return;
+    const msg = obj.message;
+    if (msg?.role === 'user') {
+        state.tailOffset = tailOffsetAfter(line, offset);
+    }
+}
+/** ai-title：最后一条生效。 */
+function collectAiTitle(state, obj) {
+    if (obj.type === 'ai-title' && typeof obj.aiTitle === 'string') {
+        state.lastAiTitle = obj.aiTitle;
+    }
+}
+/** recap：最后一条 isCompactSummary 生效。 */
+function collectRecap(state, obj) {
+    if (obj.type !== 'user' || obj.isCompactSummary !== true)
+        return;
+    const msg = obj.message;
+    if (!msg?.content)
+        return;
+    const recapContent = extractText(msg.content);
+    if (recapContent) {
+        state.lastRecap = recapContent;
+    }
+}
+/** displayTitle 用的最近真人输入：跳过 tool_result / recap / isMeta / 任务通知。 */
+function collectLastUserMessage(state, obj) {
+    if (obj.type !== 'user' || obj.isCompactSummary || obj.isMeta)
+        return;
+    const msg = obj.message;
+    if (msg?.role !== 'user' || !msg.content)
+        return;
+    const content = msg.content;
+    const isToolResult = Array.isArray(content) &&
+        content.length > 0 &&
+        content[0]?.type === 'tool_result';
+    if (isToolResult)
+        return;
+    const userContent = extractText(content);
+    if (userContent && !isTaskNotificationText(userContent)) {
+        state.lastUserMessage = userContent;
+    }
+}
+/** 真人 prompt 开启新 turn；注入/工具结果/sidechain 不递增。 */
+function countRealUserPrompt(state, obj) {
+    if (isRealUserPrompt(obj))
+        state.turnSeq++;
+}
+/**
+ * 单行分派：收集器调用顺序与原回调逐字一致（尤其 turnSeq 在最后递增，
+ * 保证 assistant usage 记录归属的是「本行之前」的 turn）。
+ */
+function scanScalarLine(state, obj, line, offset) {
+    collectAssistantUsage(state, obj);
+    collectCompactBoundary(state, obj);
+    collectTailOffset(state, obj, line, offset);
+    collectAiTitle(state, obj);
+    collectRecap(state, obj);
+    collectLastUserMessage(state, obj);
+    countRealUserPrompt(state, obj);
+}
+function scalarScan(filePath) {
+    const state = newScalarScanState();
+    scanJsonlOnce(filePath, scanScalarLine.bind(null, state));
+    return {
+        tailOffset: state.tailOffset,
+        usage: buildClaudeUsage(state.usageByMessageId.values(), {
+            turnSeq: state.turnSeq,
+            lastPostTokens: state.lastPostTokens,
+            compactCount: state.compactCount,
+        }),
+        aiTitle: state.lastAiTitle,
+        recap: state.lastRecap,
+        lastUserMessage: state.lastUserMessage,
     };
 }
 /**