@co0ontty/wand 1.36.0 → 1.39.0

package/dist/structured-session-manager.d.ts

@@ -12,6 +12,13 @@ interface CreateStructuredSessionOptions {
     model?: string;
     /** 用户预设的思考深度。留空 / null 视为 off。 */
     thinkingEffort?: SessionSnapshot["thinkingEffort"];
+    /**
+     * 恢复用的初始会话 id：
+     *   - Codex：历史 thread id，首条消息即 `codex exec ... resume <id>` 续接。
+     *   - Claude：历史 session id，首条消息即 `--resume` / SDK resume 续接。
+     * 留空表示新建会话。
+     */
+    claudeSessionId?: string;
 }
 /**
  * 把任意外部输入收敛到合法的 thinkingEffort 枚举值。`null` / 非法值都视为
@@ -150,6 +157,47 @@ export declare class StructuredSessionManager {
     private normalizeToolInput;
     private normalizeToolResultContent;
     private extractCodexText;
+    /**
+     * Merge one codex `item.*` event into `turnState.blocks`.
+     *
+     * 三种 phase 行为：
+     *   - "started":   首次出现的 item，块直接 push（tool_result 走 upsert 配对）。
+     *                  text/thinking/TodoWrite 这种"靠 id 替换"的块记录到
+     *                  codexBlockIndex 里，方便后续 updated/completed 找回原位。
+     *   - "updated":   codex 重发完整 ThreadItem（不是 delta）。已记录过的块就
+     *                  替换；新块按 started 路径处理。
+     *   - "completed": 把"in_progress"卡片定型——text 同时更新 turnState.result
+     *                  以便 result fallback 不为空；tool_use ↔ tool_result 通过
+     *                  共享 id 配对到一起（包括 file_change 子项的 `${id}#i`）。
+     */
+    private applyCodexItem;
+    /**
+     * Map a codex `item.{started,updated,completed}` payload into wand's
+     * `ContentBlock[]` so the chat UI's existing tool/diff/todo cards just work.
+     *
+     * Codex `exec --json` emits 8 item.type values (see
+     * `codex-rs/exec/src/exec_events.rs`); below they're routed to whatever wand
+     * tool name reuses an existing renderer:
+     *
+     *   agent_message     → text
+     *   reasoning         → thinking
+     *   command_execution → tool_use "Bash" + tool_result
+     *   file_change       → one tool_use per file, named Edit/Write/Bash by `kind`
+     *                       (codex does NOT carry old_string/new_string in the
+     *                       exec stream, only the path list; diff card body is
+     *                       empty but the file row + status still render)
+     *   mcp_tool_call     → tool_use named "<server>__<tool>" + tool_result
+     *   web_search        → tool_use "WebSearch" + tool_result (results not in stream)
+     *   todo_list         → tool_use "TodoWrite" (replaced in place on each update)
+     *   error             → text block prefixed with ❌
+     *
+     * Returns [] when there is nothing to emit yet (e.g. agent_message at
+     * `item.started` before any text has been produced).
+     *
+     * Callers handle in-place replacement for `item.updated` via
+     * `turnState.codexBlockIndex`; tool_use ↔ tool_result pairing still goes
+     * through `upsertCodexBlock` by matching ids.
+     */
     private extractCodexItemBlock;
     private upsertCodexBlock;
     /**

package/dist/structured-session-manager.js

@@ -8,6 +8,7 @@ import { query as sdkQuery } from "@anthropic-ai/claude-agent-sdk";
 import { prepareSessionWorktree } from "./git-worktree.js";
 import { truncateMessagesForTransport } from "./message-truncator.js";
 import { buildChildEnv } from "./env-utils.js";
+import { buildLanguageDirective } from "./language-prompt.js";
 function defaultStructuredRunner(provider) {
     return provider === "codex" ? "codex-cli-exec" : "claude-cli-print";
 }
@@ -389,9 +390,9 @@ function buildAppendSystemPromptParts(language, mode) {
             : "You are running in a fully managed, autonomous mode. The user may not be available to respond to questions or confirmations in a timely manner. You MUST make all decisions independently — choose the best approach yourself instead of asking the user for preferences, confirmations, or clarifications. If multiple approaches are viable, pick the one you judge most appropriate and proceed. Never block on user input unless the task is fundamentally ambiguous and cannot be reasonably inferred. Be decisive and self-directed.");
     }
     if (trimmedLanguage) {
-        parts.push(isChinese
-            ? "请使用中文回复。所有解释、注释和对话文本都使用中文。当你通过 Task 工具派发子代理（subagent）时，必须在传给子代理的 prompt 里明确要求它也使用中文回复，确保所有子代理的输出同样遵循该语言偏好。"
-            : `Please respond in ${trimmedLanguage}. Use ${trimmedLanguage} for all your explanations, comments, and conversational text. When you dispatch a subagent via the Task tool, you MUST explicitly instruct the subagent in its prompt to also respond in ${trimmedLanguage}, so every subagent's output follows the same language preference.`);
+        const directive = buildLanguageDirective(trimmedLanguage);
+        if (directive)
+            parts.push(directive);
     }
     return parts;
 }
@@ -579,7 +580,7 @@ export class StructuredSessionManager {
             output: "",
             archived: false,
             archivedAt: null,
-            claudeSessionId: null,
+            claudeSessionId: options.claudeSessionId?.trim() || null,
             messages: [],
             queuedMessages: [],
             structuredState: {
@@ -1160,6 +1161,7 @@ export class StructuredSessionManager {
                 sessionId: session.claudeSessionId,
                 model: session.selectedModel ?? session.structuredState?.model,
                 usage: undefined,
+                codexBlockIndex: new Map(),
             };
             let lineBuf = "";
             let stderr = "";
@@ -1231,23 +1233,24 @@ export class StructuredSessionManager {
                     return;
                 }
                 if (parsed?.type === "item.started" && parsed.item) {
-                    const block = this.extractCodexItemBlock(parsed.item, false);
-                    if (block) {
-                        turnState.blocks.push(block);
-                        syncSnapshot();
-                        scheduleEmit();
-                    }
+                    this.applyCodexItem(turnState, parsed.item, "started");
+                    syncSnapshot();
+                    scheduleEmit();
+                    return;
+                }
+                if (parsed?.type === "item.updated" && parsed.item) {
+                    // codex `item.updated` 重新发送完整 ThreadItem（不是 delta）。
+                    // 对 text/thinking/TodoWrite 走 codexBlockIndex 替换；对 tool_use
+                    // 仍然按现有 id 复用，避免重复卡片。
+                    this.applyCodexItem(turnState, parsed.item, "updated");
+                    syncSnapshot();
+                    scheduleEmit();
                     return;
                 }
                 if (parsed?.type === "item.completed" && parsed.item) {
-                    const block = this.extractCodexItemBlock(parsed.item, true);
-                    if (block) {
-                        if (block.type === "text")
-                            turnState.result = block.text;
-                        this.upsertCodexBlock(turnState.blocks, block);
-                        syncSnapshot();
-                        scheduleEmit();
-                    }
+                    this.applyCodexItem(turnState, parsed.item, "completed");
+                    syncSnapshot();
+                    scheduleEmit();
                     return;
                 }
                 if (parsed?.type === "turn.completed") {
@@ -2562,16 +2565,88 @@ export class StructuredSessionManager {
         }
         return "";
     }
+    /**
+     * Merge one codex `item.*` event into `turnState.blocks`.
+     *
+     * 三种 phase 行为：
+     *   - "started":   首次出现的 item，块直接 push（tool_result 走 upsert 配对）。
+     *                  text/thinking/TodoWrite 这种"靠 id 替换"的块记录到
+     *                  codexBlockIndex 里，方便后续 updated/completed 找回原位。
+     *   - "updated":   codex 重发完整 ThreadItem（不是 delta）。已记录过的块就
+     *                  替换；新块按 started 路径处理。
+     *   - "completed": 把"in_progress"卡片定型——text 同时更新 turnState.result
+     *                  以便 result fallback 不为空；tool_use ↔ tool_result 通过
+     *                  共享 id 配对到一起（包括 file_change 子项的 `${id}#i`）。
+     */
+    applyCodexItem(turnState, item, phase) {
+        const completed = phase === "completed";
+        const itemId = typeof item.id === "string" ? item.id : "";
+        const blocks = this.extractCodexItemBlock(item, completed);
+        if (blocks.length === 0)
+            return;
+        const index = turnState.codexBlockIndex ??= new Map();
+        for (const block of blocks) {
+            // text / thinking / TodoWrite tool_use 的卡片是"按 item id 整体替换"语义，
+            // 否则一个 agent_message 在 updated/completed 时会被重复 push 多次。
+            const replaceable = block.type === "text"
+                || block.type === "thinking"
+                || (block.type === "tool_use" && block.name === "TodoWrite");
+            if (replaceable && itemId) {
+                const existing = index.get(itemId);
+                if (existing !== undefined && existing < turnState.blocks.length) {
+                    turnState.blocks[existing] = block;
+                }
+                else {
+                    index.set(itemId, turnState.blocks.length);
+                    turnState.blocks.push(block);
+                }
+                if (block.type === "text" && completed) {
+                    turnState.result = block.text;
+                }
+                continue;
+            }
+            // 其它块（tool_use 非 Todo / tool_result / 文件改动的多 sub-id 块）
+            // 仍然走原有 upsert：tool_result 按 tool_use_id 配对，其余直接 push。
+            this.upsertCodexBlock(turnState.blocks, block);
+        }
+    }
+    /**
+     * Map a codex `item.{started,updated,completed}` payload into wand's
+     * `ContentBlock[]` so the chat UI's existing tool/diff/todo cards just work.
+     *
+     * Codex `exec --json` emits 8 item.type values (see
+     * `codex-rs/exec/src/exec_events.rs`); below they're routed to whatever wand
+     * tool name reuses an existing renderer:
+     *
+     *   agent_message     → text
+     *   reasoning         → thinking
+     *   command_execution → tool_use "Bash" + tool_result
+     *   file_change       → one tool_use per file, named Edit/Write/Bash by `kind`
+     *                       (codex does NOT carry old_string/new_string in the
+     *                       exec stream, only the path list; diff card body is
+     *                       empty but the file row + status still render)
+     *   mcp_tool_call     → tool_use named "<server>__<tool>" + tool_result
+     *   web_search        → tool_use "WebSearch" + tool_result (results not in stream)
+     *   todo_list         → tool_use "TodoWrite" (replaced in place on each update)
+     *   error             → text block prefixed with ❌
+     *
+     * Returns [] when there is nothing to emit yet (e.g. agent_message at
+     * `item.started` before any text has been produced).
+     *
+     * Callers handle in-place replacement for `item.updated` via
+     * `turnState.codexBlockIndex`; tool_use ↔ tool_result pairing still goes
+     * through `upsertCodexBlock` by matching ids.
+     */
     extractCodexItemBlock(item, completed) {
         const id = typeof item.id === "string" ? item.id : randomUUID();
         const type = typeof item.type === "string" ? item.type : "unknown";
         if (type === "agent_message") {
             const text = this.extractCodexText(item);
-            return text ? { type: "text", text } : null;
+            return text ? [{ type: "text", text }] : [];
         }
         if (type === "reasoning") {
             const text = this.extractCodexText(item);
-            return text ? { type: "thinking", thinking: text } : null;
+            return text ? [{ type: "thinking", thinking: text }] : [];
         }
         if (type === "command_execution") {
             const command = typeof item.command === "string" ? item.command : "";
@@ -2579,28 +2654,213 @@ export class StructuredSessionManager {
             const exitCode = typeof item.exit_code === "number" ? item.exit_code : null;
             const status = typeof item.status === "string" ? item.status : completed ? "completed" : "in_progress";
             if (!completed) {
+                return [{
+                        type: "tool_use",
+                        id,
+                        name: "Bash",
+                        input: { command, status },
+                    }];
+            }
+            // codex 的 status 可能是 declined（sandbox 拒了命令）/ failed（执行失败）—
+            // 这时 exit_code 经常是 null，光靠 exitCode !== 0 判 is_error 会漏。
+            const isError = status === "failed" || status === "declined"
+                || (typeof exitCode === "number" && exitCode !== 0);
+            const fallbackText = status === "declined"
+                ? "command declined by sandbox"
+                : (exitCode === null ? "" : `exit_code: ${exitCode}`);
+            return [{
+                    type: "tool_result",
+                    tool_use_id: id,
+                    content: aggregatedOutput || fallbackText,
+                    is_error: isError,
+                }];
+        }
+        if (type === "file_change") {
+            // 注意：codex exec stream 没有 old_string/new_string——只给 path + kind。
+            // 这里每个 file 一个 sub-id（`${item.id}#${i}`），这样如果 codex 一次给多
+            // 个文件，每个文件能独立成卡片 + 独立 tool_result 状态。
+            const rawChanges = Array.isArray(item.changes) ? item.changes : [];
+            const status = typeof item.status === "string" ? item.status : completed ? "completed" : "in_progress";
+            const isError = status === "failed";
+            const blocks = [];
+            rawChanges.forEach((entry, idx) => {
+                if (!entry || typeof entry !== "object")
+                    return;
+                const change = entry;
+                const path = typeof change.path === "string" ? change.path : "";
+                const kind = typeof change.kind === "string" ? change.kind : "update";
+                const subId = `${id}#${idx}`;
+                let toolName;
+                let input;
+                if (kind === "add") {
+                    toolName = "Write";
+                    input = { file_path: path, content: "" };
+                }
+                else if (kind === "delete") {
+                    // 复用 Bash 终端卡，rm 语义直观
+                    toolName = "Bash";
+                    input = { command: `rm ${path}`, description: `delete ${path}`, status };
+                }
+                else {
+                    toolName = "Edit";
+                    input = { file_path: path, old_string: "", new_string: "" };
+                }
+                if (!completed) {
+                    blocks.push({ type: "tool_use", id: subId, name: toolName, input });
+                }
+                else {
+                    blocks.push({ type: "tool_use", id: subId, name: toolName, input });
+                    blocks.push({
+                        type: "tool_result",
+                        tool_use_id: subId,
+                        content: isError ? `file change failed: ${path}` : "",
+                        is_error: isError,
+                    });
+                }
+            });
+            return blocks;
+        }
+        if (type === "mcp_tool_call") {
+            const server = typeof item.server === "string" ? item.server : "mcp";
+            const tool = typeof item.tool === "string" ? item.tool : "tool";
+            const args = item.arguments && typeof item.arguments === "object" ? item.arguments : {};
+            const errObj = item.error && typeof item.error === "object" ? item.error : null;
+            const status = typeof item.status === "string" ? item.status : completed ? "completed" : "in_progress";
+            const isError = !!errObj || status === "failed";
+            if (!completed) {
+                return [{
+                        type: "tool_use",
+                        id,
+                        name: `${server}__${tool}`,
+                        input: args,
+                    }];
+            }
+            let resultText = "";
+            if (errObj && typeof errObj.message === "string") {
+                resultText = errObj.message;
+            }
+            else if (item.result && typeof item.result === "object") {
+                const resultRec = item.result;
+                const inner = this.extractCodexText(resultRec.content);
+                resultText = inner || JSON.stringify(resultRec).slice(0, 4096);
+            }
+            return [{
+                    type: "tool_result",
+                    tool_use_id: id,
+                    content: resultText,
+                    is_error: isError,
+                }];
+        }
+        if (type === "web_search") {
+            const query = typeof item.query === "string" ? item.query : "";
+            if (!completed) {
+                return [{
+                        type: "tool_use",
+                        id,
+                        name: "WebSearch",
+                        input: { query },
+                    }];
+            }
+            return [{
+                    type: "tool_result",
+                    tool_use_id: id,
+                    // codex 不在 exec 流里回 search 结果，这里给个占位让 UI 卡片完成态。
+                    content: query ? `query: ${query}` : "",
+                }];
+        }
+        if (type === "collab_tool_call") {
+            // codex 的子-agent 编排（spawn_agent / send_input / wait / close_agent）。
+            // 没有对应 Claude tool，所以名称用 "Codex/<op>" 让 UI 默认 tool 卡渲染时
+            // 一眼能看出来是 codex 多 agent 操作。
+            const tool = typeof item.tool === "string" ? item.tool : "collab";
+            const prompt = typeof item.prompt === "string" ? item.prompt : "";
+            const senderId = typeof item.sender_thread_id === "string" ? item.sender_thread_id : "";
+            const receiverIds = Array.isArray(item.receiver_thread_ids)
+                ? item.receiver_thread_ids.filter((v) => typeof v === "string")
+                : [];
+            const agentsStates = item.agents_states && typeof item.agents_states === "object"
+                ? item.agents_states
+                : {};
+            const status = typeof item.status === "string" ? item.status : completed ? "completed" : "in_progress";
+            const toolName = `Codex/${tool}`;
+            const input = { tool };
+            if (prompt)
+                input.prompt = prompt;
+            if (senderId)
+                input.sender_thread_id = senderId;
+            if (receiverIds.length > 0)
+                input.receiver_thread_ids = receiverIds;
+            if (Object.keys(agentsStates).length > 0)
+                input.agents_states = agentsStates;
+            if (!completed) {
+                return [{ type: "tool_use", id, name: toolName, input }];
+            }
+            // 完成态：把每个 receiver agent 的最终状态汇总成可读 result。
+            const summaryLines = [];
+            for (const [tid, state] of Object.entries(agentsStates)) {
+                if (!state || typeof state !== "object")
+                    continue;
+                const rec = state;
+                const s = typeof rec.status === "string" ? rec.status : "?";
+                const msg = typeof rec.message === "string" && rec.message ? ` — ${rec.message}` : "";
+                summaryLines.push(`${tid.slice(0, 8)}: ${s}${msg}`);
+            }
+            const isError = status === "failed"
+                || summaryLines.some((l) => /errored|not_found|interrupted/.test(l));
+            const content = summaryLines.length > 0
+                ? summaryLines.join("\n")
+                : (status === "completed" ? "ok" : status);
+            return [
+                { type: "tool_use", id, name: toolName, input },
+                { type: "tool_result", tool_use_id: id, content, is_error: isError },
+            ];
+        }
+        if (type === "todo_list") {
+            // codex 的 todo: { items: [{ text, completed: bool }] }
+            // wand UI（renderTodoWrite）读的是 block.input.todos = [{content, status, activeForm}]
+            // 这里做形状翻译；in_progress 状态 codex 不区分，全部 pending → completed 二值。
+            const rawItems = Array.isArray(item.items) ? item.items : [];
+            const todos = rawItems.map((entry) => {
+                const rec = (entry && typeof entry === "object") ? entry : {};
+                const text = typeof rec.text === "string" ? rec.text : "";
+                const done = rec.completed === true;
                 return {
+                    content: text,
+                    status: done ? "completed" : "pending",
+                    activeForm: text,
+                };
+            });
+            return [{
                     type: "tool_use",
                     id,
-                    name: "Bash",
-                    input: { command, status },
-                };
-            }
-            return {
-                type: "tool_result",
-                tool_use_id: id,
-                content: aggregatedOutput || (exitCode === null ? "" : `exit_code: ${exitCode}`),
-                is_error: typeof exitCode === "number" && exitCode !== 0,
-            };
+                    name: "TodoWrite",
+                    input: { todos },
+                }];
+        }
+        if (type === "error") {
+            // item-level error（不是 top-level error 事件，那个走 codexErrors / 退出报错路径）
+            const message = this.extractCodexText(item) || "codex item error";
+            return [{ type: "text", text: `❌ ${message}` }];
         }
+        // unknown / 兜底：completed 时尝试取 text 字段免得 silently 丢
         if (completed) {
             const text = this.extractCodexText(item);
             if (text)
-                return { type: "text", text };
+                return [{ type: "text", text }];
         }
-        return null;
+        return [];
     }
     upsertCodexBlock(blocks, block) {
+        // tool_use 按 id 去重——file_change 在 item.started 已经 push 过一份 tool_use，
+        // 到 item.completed 还会再发一份相同 id 的（带 status 更新），不去重就出现
+        // 两张同名卡片。command_execution 不受影响（它在 completed 只 emit tool_result）。
+        if (block.type === "tool_use") {
+            const existingIndex = blocks.findIndex((existing) => existing.type === "tool_use" && existing.id === block.id);
+            if (existingIndex >= 0) {
+                blocks[existingIndex] = block;
+                return;
+            }
+        }
         if (block.type === "tool_result") {
             const toolUseIndex = blocks.findIndex((existing) => existing.type === "tool_use" && existing.id === block.tool_use_id);
             if (toolUseIndex >= 0) {
@@ -2724,8 +2984,12 @@ export class StructuredSessionManager {
             inputTokens: typeof source.input_tokens === "number" ? source.input_tokens : undefined,
             outputTokens: typeof source.output_tokens === "number" ? source.output_tokens : undefined,
             cacheReadInputTokens: typeof source.cached_input_tokens === "number" ? source.cached_input_tokens : undefined,
+            reasoningOutputTokens: typeof source.reasoning_output_tokens === "number" ? source.reasoning_output_tokens : undefined,
         };
-        if (value.inputTokens === undefined && value.outputTokens === undefined && value.cacheReadInputTokens === undefined) {
+        if (value.inputTokens === undefined
+            && value.outputTokens === undefined
+            && value.cacheReadInputTokens === undefined
+            && value.reasoningOutputTokens === undefined) {
             return undefined;
         }
         return value;

package/dist/types.d.ts

@@ -365,6 +365,8 @@ export interface ConversationTurn {
         outputTokens?: number;
         cacheReadInputTokens?: number;
         cacheCreationInputTokens?: number;
+        /** codex 专属：reasoning_output_tokens（GPT-5 等带思考模型，per-turn 计费）。 */
+        reasoningOutputTokens?: number;
         totalCostUsd?: number;
     };
 }