@ynhcj/xiaoyi-channel 0.0.26-beta → 0.0.26-release2026060602

package/dist/index.d.ts

@@ -1,16 +1,3 @@
-import type { OpenClawPluginApi } from "openclaw/plugin-sdk";
-import { xyPlugin } from "./src/channel.js";
-/**
- * Xiaoyi Channel Plugin Entry Point.
- * Exports the plugin for OpenClaw to load.
- * Located at root level following feishu pattern for proper plugin registration.
- */
-declare const plugin: {
-    id: string;
-    name: string;
-    description: string;
-    configSchema: import("openclaw/plugin-sdk").OpenClawPluginConfigSchema;
-    register(api: OpenClawPluginApi): void;
-};
+import { type OpenClawPluginDefinition } from "openclaw/plugin-sdk/core";
+declare const plugin: OpenClawPluginDefinition;
 export default plugin;
-export { xyPlugin };

package/dist/index.js

@@ -1,21 +1,545 @@
-import { emptyPluginConfigSchema } from "openclaw/plugin-sdk";
+import { definePluginEntry } from "openclaw/plugin-sdk/core";
+import { xiaoyiProvider } from "./src/provider.js";
+import { xiaoyiCompactionProvider } from "./src/compaction-provider.js";
 import { xyPlugin } from "./src/channel.js";
+import registerSentinelHook from "./src/cspl/sentinel_hook.js";
+import registerSkillScopeHook from "./src/cspl/skill_scope_hook.js";
 import { setXYRuntime } from "./src/runtime.js";
+import { markCronToolCall, clearCronToolCall, getCurrentSessionContext, setCronToolRunInfo, clearCronToolRunInfo, isCronActive } from "./src/tools/session-manager.js";
+import { configManager } from "./src/utils/config-manager.js";
+import { setJobPushId } from "./src/utils/cron-push-map.js";
+import { getAllPushIds } from "./src/utils/pushid-manager.js";
+import { registerSelfEvolutionToolResultNudge } from "./src/self-evolution-tool-result-nudge.js";
+import { createBeforePromptBuildHandler } from "./src/skill-retriever/hooks.js";
+import { normalizeToolRetrieverConfig } from "./src/skill-retriever/config.js";
+import { registerCLIHook } from "./src/tools/hmos-cli.js";
+import { recoverCronState } from "./src/cron-recovery.js";
+import { writeSkillUsage } from "./src/utils/skills-logger.js";
+import { markSubagentSpawned, markSubagentEnded, deliverSubagentFinalResult, } from "./src/conversation/conversation-manager.js";
+import { notifyCronAgentEnd } from "./src/conversation/cron-buffer.js";
+import { logger } from "./src/utils/logger.js";
 /**
- * Xiaoyi Channel Plugin Entry Point.
- * Exports the plugin for OpenClaw to load.
- * Located at root level following feishu pattern for proper plugin registration.
+ * Parse a file path string to detect if it refers to a SKILL.md file within
+ * a skills directory. Returns the skill name (parent directory) if so.
+ *
+ * Matches paths like:
+ *   ~/.openclaw/workspace/skills/my-skill/SKILL.md
+ *   /home/user/core_skills/my-skill/SKILL.md
+ *   skills/my-skill/SKILL.md
  */
-const plugin = {
+function extractSkillNameFromPath(filePath) {
+    if (typeof filePath !== "string" || !filePath)
+        return null;
+    // Normalize common path prefixes
+    const normalized = filePath.replace(/^~\//, "/home/").replace(/\\/g, "/");
+    // Match: .../skills/<skillName>/SKILL.md  or  .../skills/<skillName>/...
+    // Also match: .../core_skills/<skillName>/SKILL.md
+    const match = normalized.match(/\/(?:core_)?skills\/([^/]+)\/SKILL\.md$/i);
+    return match ? match[1] : null;
+}
+/**
+ * Register the skills diagnostic event listener via after_tool_call hook.
+ *
+ * When openclaw fires a `skill.used` diagnostic event, the skill's SKILL.md
+ * is typically read by the model first.  We detect SKILL.md reads through
+ * the `after_tool_call` hook and write the skill name to the skills log.
+ */
+function registerSkillsDiagnosticHook(api) {
+    // Skill name → tool names mapping for direct tool-based skill usage logging
+    const SKILL_TOOLS = {
+        GetCurrentLocation: ["get_user_location"],
+        Schedule: ["create_calendar_event", "search_calendar_event"],
+        memorandum: ["create_note", "search_notes", "modify_note"],
+        gallery: ["search_photo_gallery", "save_media_to_gallery"],
+        contact: ["search_contact"],
+        file: ["search_file", "upload_file", "save_file_to_file_manager"],
+        clock: ["create_alarm", "delete_alarm", "search_alarm", "modify_alarm"],
+        message: ["search_message", "send_message"],
+        phone: ["call_phone"],
+        "xiaoyi-collection": ["add_collection", "query_collection", "delete_collection"],
+        "xiaoyi-image-understanding": ["image_reading"],
+    };
+    // Inverted to tool name → skill name for lookup
+    const TOOL_SKILL_MAP = Object.fromEntries(Object.entries(SKILL_TOOLS).flatMap(([skill, tools]) => tools.map((t) => [t, skill])));
+    // Resolve the actual tool name from call_device_tool wrapper.
+    // The model calls call_device_tool({ toolName: "...", arguments: {...} })
+    // — the real tool name is inside params, not event.toolName.
+    function resolveActualToolName(event) {
+        if (event.toolName === "call_device_tool") {
+            const inner = event.params?.toolName;
+            if (typeof inner === "string" && inner.length > 0)
+                return inner;
+        }
+        return event.toolName;
+    }
+    // Log skill usage for known device tools on before_tool_call
+    api.on("before_tool_call", async (event, _ctx) => {
+        const actualToolName = resolveActualToolName(event);
+        const skillName = TOOL_SKILL_MAP[actualToolName];
+        if (skillName) {
+            writeSkillUsage(skillName);
+        }
+    });
+    // Detect SKILL.md reads on after_tool_call (original behavior)
+    api.on("after_tool_call", async (event, _ctx) => {
+        if (event.toolName === "read") {
+            const skillName = extractSkillNameFromPath(event.params?.path);
+            if (skillName) {
+                writeSkillUsage(skillName);
+            }
+        }
+    });
+}
+/**
+ * 从 openclaw sessionKey 解析 cron jobId。
+ * 格式形如 `agent:<agentId>:cron:<jobId>:run:<startedAtTs>`（隔离 run 为
+ * `cron:<jobId>`），jobId 位于 "cron:" 之后到下一个冒号/空白之前。
+ */
+function parseCronJobIdFromSessionKey(sessionKey) {
+    const match = sessionKey.match(/cron:([^:\s]+)/);
+    return match ? match[1] : undefined;
+}
+/**
+ * Register the cron detection hook.
+ *
+ * When openclaw's cron runner triggers a tool call, the sessionKey has the
+ * format "cron:<jobId>"（或 agent 展开后的 `agent:<id>:cron:<jobId>:run:<ts>`）。
+ * We use this to mark the toolCallId in a global Map
+ * so that sendCommand() can route the command through the push channel
+ * instead of the (non-existent) WebSocket session.
+ */
+function registerCronDetectionHook(api) {
+    api.on("before_tool_call", async (event, ctx) => {
+        // sessionKey 前缀依赖 openclaw 版本行为，不一定可靠；isCronActive(jobId)
+        // 按 jobId 精确判断（provider 已从消息 [cron:...] 标记过该 job），替代
+        // 原先的全局 120s 时间戳，避免把普通对话/其它 job 的工具调用误判为 cron。
+        const sessionKey = ctx.sessionKey ?? "";
+        const sessionJobId = parseCronJobIdFromSessionKey(sessionKey);
+        if ((sessionKey.includes("cron:") || isCronActive(sessionJobId)) && event.toolCallId) {
+            markCronToolCall(event.toolCallId);
+            // 存储 runId，供 call_device_tool 在 ALS 缺失时构造合成 SessionContext
+            if (event.runId) {
+                setCronToolRunInfo(event.toolCallId, event.runId);
+            }
+        }
+    });
+    api.on("after_tool_call", async (event, ctx) => {
+        if (event.toolCallId) {
+            clearCronToolCall(event.toolCallId);
+            clearCronToolRunInfo(event.toolCallId);
+        }
+        // 捕获对话创建的 cron job：agent 调 cron(add) 后，从 result 拿 jobId，
+        // 配合当前会话的 pushId，写入 jobId↔pushId 映射，供 fire 时反查设备。
+        await captureCronAddMapping(event, ctx).catch((err) => {
+            // 捕获失败不影响工具结果
+            console.error("[xy] captureCronAddMapping failed:", err);
+        });
+    });
+}
+/** 从 cron add 工具结果中提取 jobId 并写入 pushId 映射。 */
+async function captureCronAddMapping(event, ctx) {
+    // 两条创建路径都要捕获：
+    //   1) cron agent 工具：toolName==="cron", params.action==="add"
+    //   2) exec 跑 CLI：toolName==="exec", params.command 含 "cron add"
+    //      （agent 实际用的是这条：openclaw cron add --name ... --cron ... --message ...）
+    const isCronAddTool = event.toolName === "cron" &&
+        (event.params?.action === "add" || event.params?.action === "create");
+    const isExecCronAdd = event.toolName === "exec" && isExecCronAddCommand(event.params?.command);
+    if (!isCronAddTool && !isExecCronAdd)
+        return;
+    console.log(`[CRONMAP] after_tool_call path=${event.toolName}, resultType=${typeof event.result}`);
+    const jobId = readJobIdFromResult(event.result);
+    if (!jobId) {
+        console.log(`[CRONMAP] skip: could not extract jobId. preview=${preview(event.result)}`);
+        return;
+    }
+    console.log(`[CRONMAP] extracted jobId=${jobId}`);
+    const sessionCtx = getCurrentSessionContext();
+    const sessionId = sessionCtx?.sessionId;
+    if (!sessionId) {
+        console.log(`[CRONMAP] skip: no sessionId in ALS scope (ctxFound=${!!sessionCtx})`);
+        return;
+    }
+    const pushId = await resolvePushId(sessionId);
+    if (!pushId) {
+        console.log(`[CRONMAP] skip: no pushId available for sessionId=${sessionId} (no session match, no global, no file)`);
+        return;
+    }
+    console.log(`[CRONMAP] writing map: jobId=${jobId}, sessionId=${sessionId}, pushId=${pushId.substring(0, 16)}...`);
+    await setJobPushId(jobId, {
+        pushId,
+        sessionId,
+        deviceType: sessionCtx?.deviceType,
+        source: event.toolName === "exec" ? "exec-cli" : "conversation",
+    });
+    console.log(`[CRONMAP] map written OK`);
+}
+/** 回退链取 pushId：当前会话 → 全局兜底 → 本地文件首个（保底）。 */
+async function resolvePushId(sessionId) {
+    // 1. 同会话
+    const session = configManager.getPushId(sessionId);
+    if (session)
+        return session;
+    // 2. 全局（任何会话注册过的）
+    const global = configManager.getPushId();
+    if (global)
+        return global;
+    // 3. 文件兜底
+    try {
+        const all = await getAllPushIds();
+        if (all.length > 0)
+            return all[0];
+    }
+    catch {
+        // ignore
+    }
+    return null;
+}
+/** 判断 exec 命令是否为 cron add（匹配 "openclaw cron add" 或裸 "cron add"，排除 list/remove 等）。 */
+function isExecCronAddCommand(command) {
+    if (typeof command !== "string")
+        return false;
+    return /\bcron\s+add\b/.test(command);
+}
+/** 取结果的短预览，用于诊断。 */
+function preview(value) {
+    if (value == null)
+        return String(value);
+    const s = typeof value === "string" ? value : JSON.stringify(value);
+    return s.length > 200 ? s.slice(0, 200) + "…" : s;
+}
+/** 防御性地从 cron add 结果中取 job id。
+ *  覆盖：裸 job 对象、JSON 字符串、exec 输出文本、
+ *  {content:[{text}]} / {stdout} / data/result/job 嵌套。 */
+function readJobIdFromResult(result) {
+    if (!result)
+        return undefined;
+    // {content: [{type:"text", text: "..."}]} — exec 工具的输出信封
+    if (result && typeof result === "object") {
+        const contentArr = result.content;
+        if (Array.isArray(contentArr)) {
+            for (const item of contentArr) {
+                if (item && typeof item === "object") {
+                    const text = item.text;
+                    if (typeof text === "string" && text.trim()) {
+                        const fromContent = readJobIdFromResult(text);
+                        if (fromContent)
+                            return fromContent;
+                    }
+                }
+            }
+        }
+    }
+    // {stdout} — 备选 exec 输出信封
+    if (result && typeof result === "object") {
+        const stdout = result.stdout;
+        if (typeof stdout === "string" && stdout.trim()) {
+            const fromStdout = readJobIdFromResult(stdout);
+            if (fromStdout)
+                return fromStdout;
+        }
+    }
+    let obj = result;
+    if (typeof result === "string") {
+        try {
+            obj = JSON.parse(result);
+        }
+        catch {
+            // 纯文本：可能含 stderr 前缀行 + JSON。用正则抓 "id":"..."。
+            const m = result.match(/"id"\s*:\s*"([^"]+)"/);
+            if (m)
+                return m[1];
+            return undefined;
+        }
+    }
+    if (obj && typeof obj === "object") {
+        const id = obj.id;
+        if (typeof id === "string" && id.trim())
+            return id.trim();
+        for (const k of ["data", "result", "job"]) {
+            const inner = obj[k];
+            if (inner && typeof inner === "object") {
+                const innerId = inner.id;
+                if (typeof innerId === "string" && innerId.trim())
+                    return innerId.trim();
+            }
+        }
+    }
+    return undefined;
+}
+// ── Gateway startup: cron state recovery ────────────────────────────────────
+/**
+ * Register the gateway_start hook for cron state recovery.
+ *
+ * On gateway startup, checks .openclaw/cron/ for legacy JSON/JSONL files
+ * and migrates them into the SQLite state database:
+ *  - Reads legacy jobs.json + jobs-state.json → imports into cron_jobs table
+ *  - Reads legacy runs/*.jsonl → imports into cron_run_logs table
+ *  - Archives migrated files with .migrated suffix
+ *
+ * Pattern follows legacy-store-migration.ts and legacy-run-log-migration.ts:
+ * check → load → import into SQLite → archive old files.
+ */
+function registerCronRecoveryHook(api) {
+    api.on("gateway_start", async (_event, _ctx) => {
+        const logTag = "[CRON-RECOVERY-HOOK]";
+        const startTime = Date.now();
+        logger.log(`${logTag} ═══════════════════════════════════════════`);
+        logger.log(`${logTag} gateway_start fired — checking for legacy cron files`);
+        logger.log(`${logTag} Timestamp: ${new Date().toISOString()}`);
+        logger.log(`${logTag} Plugin registration mode: ${api.registrationMode ?? "unknown"}`);
+        let result;
+        try {
+            result = await recoverCronState();
+        }
+        catch (err) {
+            const errMsg = err instanceof Error ? err.message : String(err);
+            const errStack = err instanceof Error ? err.stack : undefined;
+            logger.error(`${logTag} cron state recovery threw: ${errMsg}`);
+            if (errStack)
+                logger.error(`${logTag} Stack: ${errStack}`);
+            // Don't let a recovery failure block gateway startup.
+            logger.log(`${logTag} Recovery failed after ${Date.now() - startTime}ms — gateway startup continues`);
+            return;
+        }
+        const elapsed = Date.now() - startTime;
+        if (result.recovered) {
+            logger.log(`${logTag} ✅ Migration performed successfully in ${elapsed}ms:` +
+                ` storeMigrated=${result.storeMigrated}, ` +
+                ` runLogFilesImported=${result.runLogFilesImported}`);
+        }
+        else {
+            logger.log(`${logTag} ℹ️ No legacy cron files migrated in ${elapsed}ms ` +
+                `(nothing to migrate or database unavailable)`);
+        }
+        // Log diagnostics summary
+        const warnings = result.diagnostics.filter((d) => d.includes("skipping") || d.includes("locked") || d.includes("unavailable"));
+        const errors = result.diagnostics.filter((d) => d.includes("error") || d.includes("failed") || d.includes("Failed"));
+        if (warnings.length > 0) {
+            logger.warn(`${logTag} ${warnings.length} warning(s) from migration:`);
+            for (const w of warnings) {
+                logger.warn(`${logTag}   ⚠ ${w}`);
+            }
+        }
+        if (errors.length > 0) {
+            logger.error(`${logTag} ${errors.length} error(s) from migration:`);
+            for (const e of errors) {
+                logger.error(`${logTag}   ✗ ${e}`);
+            }
+        }
+        if (warnings.length === 0 && errors.length === 0) {
+            logger.log(`${logTag} All diagnostics clean, no warnings or errors`);
+        }
+        logger.log(`${logTag} ═══════════════════════════════════════════`);
+    });
+}
+/**
+ * 从 sessions_spawn 工具结果中提取 ACP child sessionKey。
+ * 兼容三种 result 形态：{status, childSessionKey} 直返、JSON 字符串、
+ * {content:[{text: "<json>"}], details:{...}} 工具信封。
+ * 仅接受 status="accepted" 且 key 形如 agent:<id>:acp:<uuid> 的结果。
+ */
+function extractAcceptedAcpChildSessionKey(result) {
+    const pick = (obj) => {
+        if (!obj || typeof obj !== "object")
+            return null;
+        const record = obj;
+        if (record.status !== "accepted")
+            return null;
+        const key = record.childSessionKey;
+        return typeof key === "string" && key.includes(":acp:") ? key : null;
+    };
+    const direct = pick(result);
+    if (direct)
+        return direct;
+    const envelope = result;
+    const fromDetails = pick(envelope?.details);
+    if (fromDetails)
+        return fromDetails;
+    if (Array.isArray(envelope?.content)) {
+        for (const item of envelope.content) {
+            if (typeof item?.text !== "string")
+                continue;
+            try {
+                const fromText = pick(JSON.parse(item.text));
+                if (fromText)
+                    return fromText;
+            }
+            catch {
+                // 非 JSON 文本，忽略
+            }
+        }
+    }
+    if (typeof result === "string") {
+        try {
+            return pick(JSON.parse(result));
+        }
+        catch {
+            return null;
+        }
+    }
+    return null;
+}
+function registerSubagentHooks(api) {
+    // subagent_spawned: fires after a subagent run is successfully registered.
+    // We increment the expected completion count so that onIdle knows to wait.
+    api.on("subagent_spawned", async (_event, ctx) => {
+        const requesterSessionKey = ctx?.requesterSessionKey;
+        if (!requesterSessionKey)
+            return;
+        const count = markSubagentSpawned(requesterSessionKey);
+        if (count > 0) {
+            logger.log(`[XY-SUBAGENT] spawned, requesterSessionKey=${requesterSessionKey.slice(0, 30)}, expected=${count}`);
+        }
+    });
+    // ACP SPAWN TRACKING: openclaw 的 subagent_spawned hook 仅由内嵌 subagent
+    // 路径（subagent-spawn.ts）发射；acp-spawn.ts（runtime="acp"，如 Claude Code）
+    // 不发射 —— 实测 ACP run 仅在结束时经 run registry 触发 subagent_ended，
+    // 导致等待态从未建立、父 turn onIdle 直接发 final 结束对话。
+    // 因此在 after_tool_call 补记：sessions_spawn 成功且 childSessionKey 为 ACP
+    // 形态时递增等待计数。内嵌 subagent 的 childKey 形如 agent:<id>:subagent:<id>，
+    // 不命中此分支，仍由 subagent_spawned hook 负责，不会重复计数。
+    api.on("after_tool_call", async (event, ctx) => {
+        try {
+            if (event?.toolName !== "sessions_spawn" || event?.error)
+                return;
+            const childKey = extractAcceptedAcpChildSessionKey(event?.result);
+            if (!childKey)
+                return;
+            const sessionKey = ctx?.sessionKey;
+            if (!sessionKey)
+                return;
+            const count = markSubagentSpawned(sessionKey);
+            if (count > 0) {
+                logger.log(`[XY-SUBAGENT] ACP spawned, childSessionKey=${childKey.slice(0, 40)}, expected=${count}`);
+            }
+        }
+        catch (err) {
+            logger.error(`[XY-SUBAGENT] Error tracking ACP spawn:`, err);
+        }
+    });
+    // subagent_ended: fires when a subagent run terminates (complete/error/killed).
+    // This is the PRIMARY delivery tracking mechanism. When all expected
+    // subagents have ended and parent has settled, we finalize the A2A session.
+    api.on("subagent_ended", async (event, ctx) => {
+        try {
+            const requesterSessionKey = ctx?.requesterSessionKey;
+            if (!requesterSessionKey) {
+                logger.log(`[XY-SUBAGENT-END] no requesterSessionKey in ctx`);
+                return;
+            }
+            const transition = markSubagentEnded(requesterSessionKey);
+            logger.log(`[XY-SUBAGENT-END] ended, targetSessionKey=${event?.targetSessionKey?.slice(0, 30)}, outcome=${event?.outcome}, complete=${transition?.isComplete ?? false}, shouldFinalize=${transition?.shouldFinalize ?? false}, transition=${!!transition}`);
+            if (transition?.shouldFinalize) {
+                logger.log(`[XY-SUBAGENT-END] Starting finalization...`);
+                // Grace: announce 投递的 sendText 与 subagent_ended hook 并发到达
+                // （实测 sendText 可能先于 hook 约 1s，也可能晚于 hook；长文本还会
+                // 被 textChunkLimit 分片成多次 sendText）。捕获窗口直到
+                // finalDeliveryStarted 才关闭，此处的等待直接扩大迟到来片的入窗机会。
+                await new Promise((r) => setTimeout(r, 3000));
+                // 最终交付由对话管理层统一负责（含 config 解析、final 帧、清理）
+                await deliverSubagentFinalResult({
+                    state: transition.state,
+                    reason: "all-subagents-ended-after-parent-settled",
+                });
+                logger.log(`[XY-SUBAGENT-END] Finalized A2A session after all subagents ended`);
+            }
+        }
+        catch (err) {
+            logger.error(`[XY-SUBAGENT-END] Error in subagent_ended hook:`, err);
+        }
+    });
+}
+function registerFullHooks(api) {
+    // SUBAGENT HOOKS: track subagent spawn/end lifecycle for session keep-alive
+    registerSubagentHooks(api);
+    // CRON AGGREGATION: 通过 agent 事件总线感知 cron isolated run 结束。
+    // 注意：agent_end 插件 hook 只在 dispatch-from-config 路径触发，cron
+    // isolated run 不会触发它；embedded run 的 lifecycle 事件才是可靠信号。
+    // phase=end/error 后到达的 sendText 即为 announce 最终结果（放行）。
+    api.agent.events.registerAgentEventSubscription({
+        id: "xy-cron-turn-tracker",
+        description: "Track cron isolated run completion for sendText aggregation",
+        streams: ["lifecycle"],
+        handle: (event) => {
+            try {
+                const phase = event?.data?.phase;
+                // 仅 phase=end 是 run 级终态信号；phase=error 可能是单次模型调用
+                // 失败（后续还有重试/start），不能据此放行 announce。
+                // 真正失败的 run（无 end）由 120s 安全超时兜底 flush 缓冲文本。
+                if (phase !== "end")
+                    return;
+                const sessionKey = event?.sessionKey ?? "";
+                // 从 sessionKey（agent:<id>:cron:<jobId>:run:<ts>）解析 jobId，只结束
+                // 该 job 的 turn；isCronActive(jobId) 作为无 cron 前缀版本的兜底（按
+                // jobId 精确判断，普通对话的 lifecycle end 不会命中）。
+                const cronJobId = parseCronJobIdFromSessionKey(sessionKey);
+                if (sessionKey.includes("cron:") || sessionKey.startsWith("cron:") || isCronActive(cronJobId)) {
+                    logger.log(`[XY-CRON] Cron run lifecycle end, sessionKey=${sessionKey || "unknown"}, jobId=${cronJobId ?? "-"}`);
+                    notifyCronAgentEnd(cronJobId);
+                }
+            }
+            catch (err) {
+                logger.error(`[XY-CRON] Error in cron lifecycle subscription:`, err);
+            }
+        },
+    });
+    // SKILL RETRIEVER HOOK: before_prompt_build hook
+    const pluginConfig = api.pluginConfig || {};
+    const skillRetrieverConfig = normalizeToolRetrieverConfig({
+        enabled: pluginConfig.skillRetrieverEnabled ?? true,
+        maxTools: pluginConfig.skillRetrieverMaxTools ?? 6,
+        includeUninstalledOnly: true,
+        envFilePath: "~/.openclaw/.xiaoyienv",
+        timeoutMs: pluginConfig.skillRetrieverTimeoutMs ?? 10000,
+    });
+    const beforePromptBuildHandler = createBeforePromptBuildHandler(skillRetrieverConfig);
+    api.on("before_prompt_build", async (event, ctx) => {
+        logger.log(`[BEFORE_PROMPT_BUILD] hook fired, sessionKey=${ctx.sessionKey || "undefined"}, sessionId=${ctx.sessionId || "undefined"}`);
+        return beforePromptBuildHandler(event, ctx);
+    });
+    registerSelfEvolutionToolResultNudge(api);
+}
+const plugin = definePluginEntry({
     id: "xiaoyi-channel",
     name: "Xiaoyi Channel",
     description: "Xiaoyi channel plugin - Xiaoyi A2A protocol integration",
-    configSchema: emptyPluginConfigSchema(),
     register(api) {
-        setXYRuntime(api.runtime);
+        // Always register the provider so wrapStreamFn/prepareExtraParams work
+        // in ALL registration modes (not just "full").
+        api.registerProvider(xiaoyiProvider);
+        // Register the compaction provider so openclaw's safeguard hook uses
+        // our summarization path (which injects x-hag-trace-id) instead of the
+        // built-in LLM path that bypasses wrapStreamFn.
+        api.registerCompactionProvider(xiaoyiCompactionProvider);
+        if (api.registrationMode === "cli-metadata") {
+            return;
+        }
+        if (api.registrationMode === "tool-discovery") {
+            registerFullHooks(api);
+            return;
+        }
+        // Register channel plugin and set runtime
         api.registerChannel({ plugin: xyPlugin });
+        setXYRuntime(api.runtime);
+        if (api.registrationMode === "discovery") {
+            return;
+        }
+        if (api.registrationMode === "full") {
+            registerFullHooks(api);
+            // CSPL sentinel hook: before_tool_call + after_tool_call security scanning
+            registerSentinelHook(api);
+            // CSPL skill scope hook: before_install security scanning
+            registerSkillScopeHook(api);
+            // Cron detection hook: marks toolCallIds from cron sessions
+            registerCronDetectionHook(api);
+            // CLI exec hook: intercepts built-in exec for HarmonyOS CLI skill tools
+            registerCLIHook(api);
+            // Cron recovery hook: prunes stale cron-push-map and pushData on gateway startup
+            registerCronRecoveryHook(api);
+            // Skills diagnostic hook: log skill usage (detected via SKILL.md reads)
+            registerSkillsDiagnosticHook(api);
+        }
     },
-};
+});
 export default plugin;
-// Also export the plugin directly for testing
-export { xyPlugin };

package/dist/provider-discovery.d.ts

@@ -0,0 +1,2 @@
+import { xiaoyiProvider } from "./src/provider.js";
+export default xiaoyiProvider;

package/dist/provider-discovery.js

@@ -0,0 +1,4 @@
+// Provider discovery entry for fast-path provider resolution.
+// Exported as default so normalizeDiscoveryModule can unwrap it via .default.
+import { xiaoyiProvider } from "./src/provider.js";
+export default xiaoyiProvider;

package/dist/src/acp-session-binding.d.ts

@@ -0,0 +1,37 @@
+import { type BindingTargetKind } from "openclaw/plugin-sdk/conversation-runtime";
+type XyBindingTargetKind = "subagent" | "acp";
+type XyAcpBindingRecord = {
+    accountId: string;
+    conversationId: string;
+    parentConversationId?: string;
+    deliveryTo?: string;
+    targetKind: XyBindingTargetKind;
+    targetSessionKey: string;
+    agentId?: string;
+    label?: string;
+    boundBy?: string;
+    boundAt: number;
+    lastActivityAt: number;
+};
+type XyAcpBindingManager = {
+    accountId: string;
+    getByConversationId: (conversationId: string) => XyAcpBindingRecord | undefined;
+    listBySessionKey: (targetSessionKey: string) => XyAcpBindingRecord[];
+    bindConversation: (params: {
+        conversationId: string;
+        parentConversationId?: string;
+        targetKind: BindingTargetKind;
+        targetSessionKey: string;
+        metadata?: Record<string, unknown>;
+    }) => XyAcpBindingRecord | null;
+    touchConversation: (conversationId: string, at?: number) => XyAcpBindingRecord | null;
+    unbindConversation: (conversationId: string) => XyAcpBindingRecord | null;
+    unbindBySessionKey: (targetSessionKey: string) => XyAcpBindingRecord[];
+    stop: () => void;
+};
+export declare function createXyAcpBindingManager(params: {
+    accountId?: string;
+    cfg: any;
+}): XyAcpBindingManager;
+export declare function getXyAcpBindingManager(accountId?: string): XyAcpBindingManager | null;
+export {};