lark-remote 0.5.18 → 0.5.20

package/dist/runner/common/connection-based-runner.js

@@ -150,7 +150,7 @@ export class ConnectionBasedRunner {
                 // 已被缓存层除名（evictRunnerSlot → markOrphaned）：本次 run 创建的
                 // 连接挂在无人托管的实例上。主动收掉而不是挂 idle timer——否则 agent
                 // server 进程残留至 idle TTL，期间持有 agent 侧会话写锁（codex
-                // "thread already has an active writer"，2026-10-11 线上实测）。
+                // "thread already has an active writer"）。
                 // dispose() 内部自带 try/catch，这里 fire-and-forget 不阻塞事件流收尾。
                 void this.disposeConnections().catch((err) => {
                     getLogger().warn(`[${this.logTag}] orphan dispose error: ${err.message}`);
@@ -209,7 +209,7 @@ export class ConnectionBasedRunner {
     /**
      * Mark this runner as evicted from the bridge's runner cache while a run may
      * still be holding/creating connections on it (queue-advance race during
-     * interruptCurrentRun's await window, 2026-10-11). The active run's finally
+     * interruptCurrentRun's await window). The active run's finally
      * disposes connections eagerly instead of arming the idle TTL timer —
      * otherwise the agent server process lingers for up to 30 minutes holding
      * its session write lock (codex: "thread already has an active writer").
@@ -331,6 +331,36 @@ export class ConnectionBasedRunner {
     isCurrentTurn(generation, client) {
         return this._isRunning && generation === this.turnGeneration && client === this.currentClient;
     }
+    /**
+     * Acquire the per-turn connection and capture the callback anchor
+     * ({@link TurnAnchor}): acquire → activity ping → currentClient → anchor.
+     * Shared preamble of every setupTurn / runCompact; the acquired anchor is
+     * what the turn's hooks (and prompt-settle callbacks) validate against.
+     *
+     * `acquire` is a closure so subclasses keep their own acquire signatures
+     * (pi passes the spawn-bound `{ sessionId }`).
+     */
+    async acquireForTurn(cwd, acquire) {
+        const client = await acquire();
+        this.notifyConnectionActivity(cwd);
+        this.currentClient = client;
+        return { client, generation: this.turnGeneration };
+    }
+    /**
+     * Wrap a per-turn callback with the isCurrentTurn guard: events from a
+     * finished turn are dropped with a debug log and never reach the turn
+     * dispatcher. 6 处守卫原有的日志粒度不一（ACP setupTurn 记 debug，其余静默），
+     * 统一为最详细版本：dropped 一律 debug，便于事后追溯。
+     */
+    forCurrentTurn(anchor, label, handler) {
+        return (...args) => {
+            if (!this.isCurrentTurn(anchor.generation, anchor.client)) {
+                getLogger().debug(`[${this.logTag}] dropped ${label} from a finished turn`);
+                return;
+            }
+            handler(...args);
+        };
+    }
     failTurn(message) {
         this.pushEvents([
             {

package/dist/runner/common/jsonrpc/connection-manager.d.ts

@@ -151,8 +151,8 @@ export declare class ConnectionManager<TClient extends ConnectionClient = JsonRp
      * 定时器回调带 **slot 身份校验**：武装之后 slot 可能已被替换——进程自退走
      * baseOnClose（删 slot 但不经过 disposeSlot）、或同 cwd 已 acquire 了新连接。
      * 这两种情况下本定时器都已过期，不得回收「当时恰好挂在该 cwd 上」的连接：
-     * 2026-10-24 线上事故即一条泄漏的过期 TTL 杀掉继任连接的在途 turn，卡片
-     * 误报 "Codex app-server connection closed"。
+     * 一条泄漏的过期 TTL 会杀掉继任连接的在途 turn，卡片误报
+     * "Codex app-server connection closed"。
      */
     notifyIdle(cwd: string): void;
     private createClient;

package/dist/runner/common/jsonrpc/connection-manager.js

@@ -197,8 +197,8 @@ export class ConnectionManager {
      * 定时器回调带 **slot 身份校验**：武装之后 slot 可能已被替换——进程自退走
      * baseOnClose（删 slot 但不经过 disposeSlot）、或同 cwd 已 acquire 了新连接。
      * 这两种情况下本定时器都已过期，不得回收「当时恰好挂在该 cwd 上」的连接：
-     * 2026-10-24 线上事故即一条泄漏的过期 TTL 杀掉继任连接的在途 turn，卡片
-     * 误报 "Codex app-server connection closed"。
+     * 一条泄漏的过期 TTL 会杀掉继任连接的在途 turn，卡片误报
+     * "Codex app-server connection closed"。
      */
     notifyIdle(cwd) {
         if (this.idleTtlMs <= 0)
@@ -245,8 +245,8 @@ export class ConnectionManager {
                 if (current?.client === client) {
                     // 进程自退（codex app-server 偶发 exit 0，非本程序发起，不经
                     // disposeSlot）也要撤掉挂在旧 slot 上的 idle TTL：定时器不随 slot
-                    // 删除而取消，到期时会按 cwd 误杀继任连接的在途 turn（2026-10-24
-                    // 线上事故，见 notifyIdle 的身份校验与对应回归测试）。
+                    // 删除而取消，到期时会按 cwd 误杀继任连接的在途 turn（见 notifyIdle
+                    // 的身份校验与对应回归测试）。
                     this.clearIdleTimer(current);
                     this.slots.delete(cwd);
                     this.onConnectionLost?.(cwd);

package/dist/runner/common/spawning-runner.d.ts

@@ -4,7 +4,7 @@ import { type Terminator } from '../../platform/terminator.js';
 import { SpawnHeartbeat } from './spawn-heartbeat.js';
 import type { SpawnOptions } from '../types.js';
 /**
- * 进程级退出监听单例分发（修复，2026-08-02）。
+ * 进程级退出监听单例分发。
  *
  * 背景：各 agent runner 的 registerExitHandlers 曾是同构复制，每个实例
  * process.on('exit'|'SIGINT'|'SIGTERM') 注册 3 个永不移除的闭包（捕获整个

package/dist/runner/common/spawning-runner.js

@@ -357,7 +357,7 @@ export class SpawningRunner {
             // matchMode 'agent-invocation'（不是默认的 'executable'）：agent CLI 普遍是
             // `#!/usr/bin/env node` 脚本，ps 看到的是 `node <安装路径>/…/cli.js`，名字只在
             // 路径段里（pi → @earendil-works/pi-coding-agent/…）。用默认档会恒判 mismatch
-            // → 这些 agent 的孤儿永远回收不掉（2026-09-20 实测安装形态）。
+            // → 这些 agent 的孤儿永远回收不掉。
             const verdict = verifyPidIdentityVerdictSync(pid, {
                 expectedBinary,
                 matchMode: 'agent-invocation',

package/dist/runner/dsh/catalog-probe.d.ts

@@ -0,0 +1,17 @@
+/**
+ * DSH 服务端目录探测（/config 卡片动态清单用）。
+ *
+ * 接口只声明卡片预取需要的三个 RPC 方法；具体传输客户端的构造（含 token
+ * 解析路径）收在本模块——router 只依赖接口，不持有 DshClient。
+ */
+import type { DshModelCatalogValue, DshPermissionCatalogValue, DshPresetListValue } from '../../dsh/types.js';
+/** 卡片预取只依赖的 client 方法子集。 */
+export interface DshCatalogClient {
+    listModels(): Promise<DshModelCatalogValue>;
+    listPresets(): Promise<DshPresetListValue>;
+    listPermissionPresets(): Promise<DshPermissionCatalogValue>;
+}
+/** 目录预取客户端工厂类型（测试注入 fake 用）。 */
+export type DshCatalogClientFactory = (baseUrl: string, token?: string, credentialsFile?: string) => DshCatalogClient;
+/** 生产实现：卡片预取是交互路径，DSH 卡死时最多等 5 秒后回退固定清单。 */
+export declare function createDshCatalogClient(baseUrl: string, token?: string, credentialsFile?: string): DshCatalogClient;

package/dist/runner/dsh/catalog-probe.js

@@ -0,0 +1,10 @@
+import { DshClient } from './client.js';
+/** 生产实现：卡片预取是交互路径，DSH 卡死时最多等 5 秒后回退固定清单。 */
+export function createDshCatalogClient(baseUrl, token, credentialsFile) {
+    return new DshClient(baseUrl, {
+        token,
+        credentialsFile,
+        requestTimeoutMs: 5_000,
+    });
+}
+//# sourceMappingURL=catalog-probe.js.map

package/dist/runner/dsh/client.d.ts

@@ -19,7 +19,7 @@
  * `sessionEvents()` adds per-session filtering, reconnect + follow-snapshot
  * replay, and seq-based dedup.
  */
-import type { DshApprovalStreamItem, DshCommandResult, DshModelCatalogValue, DshModelSelection, DshOpenTurnState, DshPermissionCatalogValue, DshJobView, DshPresetListValue, DshProjectionItem, DshProjectionSnapshot, DshQuestion, DshRemoteEventOutcome, DshSessionPageValue, DshStreamItem, DshTokenUsage } from '../../dsh/types.js';
+import type { DshApprovalStreamItem, DshCommandResult, DshModelCatalogValue, DshModelSelection, DshOpenTurnState, DshPermissionCatalogValue, DshJobView, DshPresetListValue, DshProjectionItem, DshProjectionSnapshot, DshQuestion, DshRemoteEventOutcome, DshSessionPageValue, DshStreamItem } from '../../dsh/types.js';
 import type { GoalView } from '../types.js';
 export { DshError } from './errors.js';
 interface WebSocketMessageEvent {
@@ -325,13 +325,20 @@ export declare class DshClient {
      */
     controlFrames(signal: AbortSignal): AsyncGenerator<DshProjectionItem>;
     /**
-     * One session's current background-job roster.
+     * Subscribe to one session's background-job roster as a live stream.
      *
-     * `job/list` 是流（生命周期变化时推整份名单），但"此刻有哪些任务在跑"只需
-     * 要第一帧：文档明确首帧就是当前真相（whole-set replacement）。取到即断开，
-     * 不长期持有连接——任务列表只用于收尾时的一句话与停止按钮。
+     * `job/list` 是整份替换流（whole-set replacement）：首帧就是当前真相，之后
+     * 任务生命周期一变就再推一整份。**run 结束之后**必须用这个长订阅——后台任务
+     * 跨 turn 存活，任务结束这件事只会在这条流上出现；收尾时取首帧即断开
+     * （{@link jobRows}）只够画那一瞬间的卡片。
+     */
+    watchJobs(sessionId: string, signal: AbortSignal): AsyncGenerator<DshJobView[]>;
+    /**
+     * One session's current background-job roster (snapshot).
      *
-     * 失败返回 undefined（与"没有任务"区分）：拿不到就别在卡片上写"0 个后台任务"。
+     * "此刻有哪些任务在跑"只需要第一帧：文档明确首帧就是当前真相，取到即断开，
+     * 不长期持有连接。失败返回 undefined（与"没有任务"区分）：拿不到就别在卡片
+     * 上写"0 个后台任务"。
      */
     jobRows(sessionId: string, opts?: {
         signal?: AbortSignal;
@@ -341,8 +348,6 @@ export declare class DshClient {
     jobKill(sessionId: string, jobId: string, opts?: {
         signal?: AbortSignal;
     }): Promise<boolean>;
-    /** Extract TokenUsage from an assistant/message data payload (absent → undefined). */
-    static usageOf(data: Record<string, unknown>): DshTokenUsage | undefined;
 }
 /** Stable client-side prompt identity persisted on the accepted user message. */
 export declare function newPromptRequestId(): string;

package/dist/runner/dsh/client.js

@@ -239,6 +239,7 @@ export class DshClient {
                     err instanceof DshError &&
                     err.code === 'http' &&
                     err.details.status === 401) {
+                    getLogger().warn(`[dsh] 401 while authenticating ${method}, re-minting cookie once`);
                     continue;
                 }
                 throw err;
@@ -271,6 +272,7 @@ export class DshClient {
             }
             if (res.status === 401 && attempt === 0) {
                 this.cookiePromise = undefined;
+                getLogger().warn(`[dsh] 401 for ${method}, re-minting cookie once`);
                 continue;
             }
             if (res.status === 401) {
@@ -927,25 +929,35 @@ export class DshClient {
         }
     }
     /**
-     * One session's current background-job roster.
+     * Subscribe to one session's background-job roster as a live stream.
      *
-     * `job/list` 是流（生命周期变化时推整份名单），但"此刻有哪些任务在跑"只需
-     * 要第一帧：文档明确首帧就是当前真相（whole-set replacement）。取到即断开，
-     * 不长期持有连接——任务列表只用于收尾时的一句话与停止按钮。
+     * `job/list` 是整份替换流（whole-set replacement）：首帧就是当前真相，之后
+     * 任务生命周期一变就再推一整份。**run 结束之后**必须用这个长订阅——后台任务
+     * 跨 turn 存活，任务结束这件事只会在这条流上出现；收尾时取首帧即断开
+     * （{@link jobRows}）只够画那一瞬间的卡片。
+     */
+    async *watchJobs(sessionId, signal) {
+        for await (const frame of this.muxStream('job/list', { args: dshArgs('job/list', { sessionId }) }, signal)) {
+            const rows = frame;
+            if (rows?.type !== 'rows')
+                continue;
+            yield Array.isArray(rows.jobs) ? rows.jobs : [];
+        }
+    }
+    /**
+     * One session's current background-job roster (snapshot).
      *
-     * 失败返回 undefined（与"没有任务"区分）：拿不到就别在卡片上写"0 个后台任务"。
+     * "此刻有哪些任务在跑"只需要第一帧：文档明确首帧就是当前真相，取到即断开，
+     * 不长期持有连接。失败返回 undefined（与"没有任务"区分）：拿不到就别在卡片
+     * 上写"0 个后台任务"。
      */
     async jobRows(sessionId, opts = {}) {
         const timer = new AbortController();
         const timeout = setTimeout(() => timer.abort(), opts.timeoutMs ?? 2_000);
         const signal = opts.signal ? AbortSignal.any([opts.signal, timer.signal]) : timer.signal;
         try {
-            for await (const frame of this.muxStream('job/list', { args: dshArgs('job/list', { sessionId }) }, signal)) {
-                const rows = frame;
-                if (rows?.type === 'rows') {
-                    return Array.isArray(rows.jobs) ? rows.jobs : [];
-                }
-            }
+            for await (const rows of this.watchJobs(sessionId, signal))
+                return rows;
             return undefined;
         }
         catch (err) {
@@ -968,10 +980,6 @@ export class DshClient {
             return false;
         }
     }
-    /** Extract TokenUsage from an assistant/message data payload (absent → undefined). */
-    static usageOf(data) {
-        return data.usage ?? undefined;
-    }
 }
 /** Stable client-side prompt identity persisted on the accepted user message. */
 export function newPromptRequestId() {

package/dist/runner/dsh/exit-cancel.d.ts

@@ -5,8 +5,8 @@
  * `dsh web` host (`DshDaemon`). So the usual "kill my spawned child" exit
  * cleanup does nothing for it, and a bridge that exits mid-turn leaves the
  * turn running with no subscriber — an orphan turn that keeps consuming tokens
- * and mutating the workspace invisibly (2026-10-19: one ran 2h21m / 345 steps
- * after its bridge died).
+ * and mutating the workspace invisibly, potentially for hours after the bridge
+ * died.
  *
  * `process.on('exit')` cannot await, so this path uses the same synchronous
  * curl transport the session reader already uses: bounded by `timeoutMs` and

package/dist/runner/dsh/exit-cancel.js

@@ -5,8 +5,8 @@
  * `dsh web` host (`DshDaemon`). So the usual "kill my spawned child" exit
  * cleanup does nothing for it, and a bridge that exits mid-turn leaves the
  * turn running with no subscriber — an orphan turn that keeps consuming tokens
- * and mutating the workspace invisibly (2026-10-19: one ran 2h21m / 345 steps
- * after its bridge died).
+ * and mutating the workspace invisibly, potentially for hours after the bridge
+ * died.
  *
  * `process.on('exit')` cannot await, so this path uses the same synchronous
  * curl transport the session reader already uses: bounded by `timeoutMs` and

package/dist/runner/dsh/inflight.d.ts

@@ -3,7 +3,7 @@
  *
  * A DSH turn runs inside the shared `dsh web` host, not in a child process. If
  * the bridge dies (SIGKILL, crash) the turn keeps running with no subscriber —
- * the 2026-10-19 incident ran 2h21m / 345 steps after its bridge exited. The
+ * one orphaned turn kept going for hours after its bridge exited. The
  * runner writes one entry when it submits a prompt and removes it once the run
  * (or a verified stop) accounts for that prompt; the next startup reconciles
  * anything left behind:

package/dist/runner/dsh/inflight.js

@@ -3,7 +3,7 @@
  *
  * A DSH turn runs inside the shared `dsh web` host, not in a child process. If
  * the bridge dies (SIGKILL, crash) the turn keeps running with no subscriber —
- * the 2026-10-19 incident ran 2h21m / 345 steps after its bridge exited. The
+ * one orphaned turn kept going for hours after its bridge exited. The
  * runner writes one entry when it submits a prompt and removes it once the run
  * (or a verified stop) accounts for that prompt; the next startup reconciles
  * anything left behind:

package/dist/runner/dsh/runner.d.ts

@@ -18,7 +18,7 @@
  * back to the Web UI prompt.
  */
 import type { AgentSessionReader } from '../../session-contract.js';
-import type { AgentEvent, AgentStatusInfo, SpawnOptions, StopOutcome } from '../types.js';
+import type { AgentEvent, AgentStatusInfo, BackgroundJobRoster, SpawnOptions, StopOutcome } from '../types.js';
 import type { AgentRunner } from '../types.js';
 import { DshClient, type WebSocketFactory } from './client.js';
 import type { DshDaemon } from './daemon.js';
@@ -155,12 +155,51 @@ export declare class DshRunner implements AgentRunner {
      */
     private submitPrompt;
     run(message: string, opts: SpawnOptions): AsyncGenerator<AgentEvent>;
+    /**
+     * 准备一个 turn（run() 的准备阶段）：模型对齐 + 权限预设对齐 + goal seed 探测。
+     *
+     * 红线（逐条保留）：
+     * - 对齐失败 fail closed：显式配置 provider 时 selectModel 失败必须 throw
+     *   （静默回退到其他 provider 会让用户以为切换成功，实际模型与账单都错）；
+     *   权限预设失败由 alignPermissionPreset throw（用户明确选了安全策略，不能在
+     *   错误的权限下继续跑 prompt）。
+     * - 准备期间被 stop（abort）→ 返回 { ok:false }，由调用方 yield interrupted
+     *   result 收尾。
+     * - goal seed 只探测不推断：读不到按 unknown 上报，不冒充「没有 goal」。
+     */
+    private prepareTurn;
     /**
      * Consume one turn from both transports. `session/follow` owns transcript
      * events; `$events` owns the answerable approval waterfall. The merged
      * stream keeps either transport from starving the other.
+     *
+     * 事件源分派：projection（goal + 状态条）→ handleProjectionItem；
+     * $events waterfall（审批/提问）→ handleApprovalWaterfallItem；
+     * assistant-stream 直转发（translator 持有本 turn 的 usage 累计状态）；
+     * session/follow 事件 → handleSessionEvent（含 result 收尾的 jobs + goal 复核）。
      */
     private runTurn;
+    /**
+     * 事件源 handler：Host-wide 投影（goal + 状态条）。
+     * 返回要回放的事件（可能为空——未变化/非本 session 的帧不产事件）。
+     */
+    private handleProjectionItem;
+    /**
+     * 事件源 handler：$events 审批/提问 waterfall（channel 状态、审批 waterfall
+     * 登记、提问卡、cancel 回流、审批卡）。
+     * 返回要回放的事件；channel 状态类帧（ready/unavailable/request）只改内部
+     * 状态、不产事件。
+     */
+    private handleApprovalWaterfallItem;
+    /**
+     * 事件源 handler：`session/follow` 会话事件。
+     * - goal/change：权威变更，不等投影直接更新；
+     * - tool/call：先记账（审批卡命令展示用）再落入 translator（卡片工具行来源）；
+     * - approval/decided：审批回流，清审批卡；
+     * - 其余事件经 translator 翻译；result 事件先做收尾（后台任务名单 + goal
+     *   复核）再随 result 一起返回。
+     */
+    private handleSessionEvent;
     /**
      * 收尾时仍在跑的后台任务。
      *
@@ -169,6 +208,13 @@ export declare class DshRunner implements AgentRunner {
      * 任务列表不等于有任务在跑。
      */
     private collectBackgroundJobs;
+    /**
+     * 订阅本 session 的后台任务名单（跨 turn 生命周期）。
+     *
+     * 卡片收尾后仍要靠它把「任务结束了」推给用户：run 的事件流在 result 就关了，
+     * 而后台任务可能再跑半小时。名单是整份替换，消费方自己做 diff。
+     */
+    watchJobs(sessionId: string, signal: AbortSignal): AsyncGenerator<ReadonlyArray<BackgroundJobRoster>>;
     /** Kill one background job of this runner's session (card button path). */
     killJob(sessionId: string, jobId: string): Promise<boolean>;
     /** Apply the configured permission preset once per session, before prompt. */
@@ -272,7 +318,7 @@ export declare class DshRunner implements AgentRunner {
      *
      * turn 跑在常驻共享的 `dsh web` host 里，不在本进程的子进程树上；所以
      * "杀自己的子进程"这套退出清理对它无效，桥接一退出，turn 继续无人接收地跑
-     * （2026-10-19 实测：一个孤儿 turn 在桥接死后又跑了 2h21m / 345 个 step）。
+     * （孤儿 turn 在桥接死后可能再跑数小时）。
      * 这里复用进程级退出分发器，用同步 curl 发一次 `session/cancel`。
      */
     registerExitHandlers(): void;