@fieldwangai/agentflow 0.1.29 → 0.1.31

package/agents/agentflow-node-executor-code.md

@@ -11,7 +11,8 @@ readonly: true
 
 执行时**只引用本节的变量**，勿自行推导或拼接路径：
 
-- workspaceRoot：${workspaceRoot}（工作区根目录）
+- workspaceRoot：${workspaceRoot}（当前执行工作区根目录，可能由 CD Workspace 节点切换）
+- pipelineWorkspace：${pipelineWorkspace}（流水线所在工作区，写 AgentFlow 结果时使用）
 - flowName：${flowName}
 - uuid：${uuid}
 - instanceId：${instanceId}
@@ -28,5 +29,5 @@ ${taskBody}
 
 以代码实现为核心：理清需求与接口，编写或修改代码，确保可运行、通过类型检查与项目规范，节点中如有写入文件的操作可以执行。任务完成后直接退出，结果由系统自动标记成功。**仅当任务明确失败时**，执行以下命令报告失败（`agentflow` 是可直接在终端运行的 CLI 命令）：
 ```bash
-agentflow apply -ai write-result ${workspaceRoot} ${flowName} ${uuid} ${instanceId} --json '{"status":"failed","message":"失败原因"}'
+agentflow apply -ai write-result ${pipelineWorkspace} ${flowName} ${uuid} ${instanceId} --json '{"status":"failed","message":"失败原因"}'
 ```

package/agents/agentflow-node-executor-planning.md

@@ -11,7 +11,8 @@ readonly: true
 
 执行时**只引用本节的变量**，勿自行推导或拼接路径：
 
-- workspaceRoot：${workspaceRoot}（工作区根目录）
+- workspaceRoot：${workspaceRoot}（当前执行工作区根目录，可能由 CD Workspace 节点切换）
+- pipelineWorkspace：${pipelineWorkspace}（流水线所在工作区，写 AgentFlow 结果时使用）
 - flowName：${flowName}
 - uuid：${uuid}
 - instanceId：${instanceId}
@@ -28,5 +29,5 @@ ${taskBody}
 
 先做规划（目标拆解、步骤与依赖、可选方案），再按规划执行并产出结果，节点中如有写入文件的操作可以执行。任务完成后直接退出，结果由系统自动标记成功。**仅当任务明确失败时**，执行以下命令报告失败（`agentflow` 是可直接在终端运行的 CLI 命令）：
 ```bash
-agentflow apply -ai write-result ${workspaceRoot} ${flowName} ${uuid} ${instanceId} --json '{"status":"failed","message":"失败原因"}'
+agentflow apply -ai write-result ${pipelineWorkspace} ${flowName} ${uuid} ${instanceId} --json '{"status":"failed","message":"失败原因"}'
 ```

package/agents/agentflow-node-executor-requirement.md

@@ -11,7 +11,8 @@ readonly: true
 
 执行时**只引用本节的变量**，勿自行推导或拼接路径：
 
-- workspaceRoot：${workspaceRoot}（工作区根目录）
+- workspaceRoot：${workspaceRoot}（当前执行工作区根目录，可能由 CD Workspace 节点切换）
+- pipelineWorkspace：${pipelineWorkspace}（流水线所在工作区，写 AgentFlow 结果时使用）
 - flowName：${flowName}
 - uuid：${uuid}
 - instanceId：${instanceId}
@@ -28,5 +29,5 @@ ${taskBody}
 
 侧重需求理解与拆解，按上述任务完成执行，节点中如有写入文件的操作可以执行。任务完成后直接退出，结果由系统自动标记成功。**仅当任务明确失败时**，执行以下命令报告失败（`agentflow` 是可直接在终端运行的 CLI 命令）：
 ```bash
-agentflow apply -ai write-result ${workspaceRoot} ${flowName} ${uuid} ${instanceId} --json '{"status":"failed","message":"失败原因"}'
+agentflow apply -ai write-result ${pipelineWorkspace} ${flowName} ${uuid} ${instanceId} --json '{"status":"failed","message":"失败原因"}'
 ```

package/agents/agentflow-node-executor-test.md

@@ -11,7 +11,8 @@ readonly: true
 
 执行时**只引用本节的变量**，勿自行推导或拼接路径：
 
-- workspaceRoot：${workspaceRoot}（工作区根目录）
+- workspaceRoot：${workspaceRoot}（当前执行工作区根目录，可能由 CD Workspace 节点切换）
+- pipelineWorkspace：${pipelineWorkspace}（流水线所在工作区，写 AgentFlow 结果时使用）
 - flowName：${flowName}
 - uuid：${uuid}
 - instanceId：${instanceId}
@@ -28,5 +29,5 @@ ${taskBody}
 
 侧重测试与验证，执行并产出结果，节点中如有写入文件的操作可以执行。任务完成后直接退出，结果由系统自动标记成功。**仅当任务明确失败时**，执行以下命令报告失败（`agentflow` 是可直接在终端运行的 CLI 命令）：
 ```bash
-agentflow apply -ai write-result ${workspaceRoot} ${flowName} ${uuid} ${instanceId} --json '{"status":"failed","message":"失败原因"}'
+agentflow apply -ai write-result ${pipelineWorkspace} ${flowName} ${uuid} ${instanceId} --json '{"status":"failed","message":"失败原因"}'
 ```

package/agents/agentflow-node-executor-ui.md

@@ -11,7 +11,8 @@ readonly: true
 
 执行时**只引用本节的变量**，勿自行推导或拼接路径：
 
-- workspaceRoot：${workspaceRoot}（工作区根目录）
+- workspaceRoot：${workspaceRoot}（当前执行工作区根目录，可能由 CD Workspace 节点切换）
+- pipelineWorkspace：${pipelineWorkspace}（流水线所在工作区，写 AgentFlow 结果时使用）
 - flowName：${flowName}
 - uuid：${uuid}
 - instanceId：${instanceId}
@@ -28,5 +29,5 @@ ${taskBody}
 
 理解设计稿或规格（Figma、标注、描述），实现或调整组件与样式，保证布局、间距、层级、断点与 RTL 等与设计一致；必要时做走查与修正，节点中如有写入文件的操作可以执行。任务完成后直接退出，结果由系统自动标记成功。**仅当任务明确失败时**，执行以下命令报告失败（`agentflow` 是可直接在终端运行的 CLI 命令）：
 ```bash
-agentflow apply -ai write-result ${workspaceRoot} ${flowName} ${uuid} ${instanceId} --json '{"status":"failed","message":"失败原因"}'
+agentflow apply -ai write-result ${pipelineWorkspace} ${flowName} ${uuid} ${instanceId} --json '{"status":"failed","message":"失败原因"}'
 ```

package/agents/agentflow-node-executor.md

@@ -11,7 +11,8 @@ readonly: true
 
 执行时**只引用本节的变量**，勿自行推导或拼接路径：
 
-- workspaceRoot：${workspaceRoot}（工作区根目录）
+- workspaceRoot：${workspaceRoot}（当前执行工作区根目录，可能由 CD Workspace 节点切换）
+- pipelineWorkspace：${pipelineWorkspace}（流水线所在工作区，写 AgentFlow 结果时使用）
 - flowName：${flowName}
 - uuid：${uuid}
 - instanceId：${instanceId}
@@ -28,5 +29,5 @@ ${taskBody}
 
 按上述任务完成执行，节点中如有写入文件的操作可以执行。任务完成后直接退出，结果由系统自动标记成功。**仅当任务明确失败时**，执行以下命令报告失败（`agentflow` 是可直接在终端运行的 CLI 命令）：
 ```bash
-agentflow apply -ai write-result ${workspaceRoot} ${flowName} ${uuid} ${instanceId} --json '{"status":"failed","message":"失败原因"}'
+agentflow apply -ai write-result ${pipelineWorkspace} ${flowName} ${uuid} ${instanceId} --json '{"status":"failed","message":"失败原因"}'
 ```

package/agents/en/agentflow-node-executor.md

@@ -11,7 +11,8 @@ You are a flow node executor. Complete the work described in the node context an
 
 **Only reference the variables in this section** during execution. Do not derive or concatenate paths on your own:
 
-- workspaceRoot: ${workspaceRoot} (workspace root directory)
+- workspaceRoot: ${workspaceRoot} (current execution workspace root; may be switched by a CD Workspace node)
+- pipelineWorkspace: ${pipelineWorkspace} (pipeline workspace; use this when writing AgentFlow results)
 - flowName: ${flowName}
 - uuid: ${uuid}
 - instanceId: ${instanceId}
@@ -28,5 +29,5 @@ ${taskBody}
 
 Complete the task as described above. If the node involves file writing operations, they can be executed. Exit when done — the system automatically marks the result as success. **Only if the task explicitly fails**, run the following command to report failure (`agentflow` is a CLI command available directly in the terminal):
 ```bash
-agentflow apply -ai write-result ${workspaceRoot} ${flowName} ${uuid} ${instanceId} --json '{"status":"failed","message":"reason for failure"}'
+agentflow apply -ai write-result ${pipelineWorkspace} ${flowName} ${uuid} ${instanceId} --json '{"status":"failed","message":"reason for failure"}'
 ```

package/agents/zh/agentflow-node-executor.md

@@ -11,7 +11,8 @@ readonly: true
 
 执行时**只引用本节的变量**，勿自行推导或拼接路径：
 
-- workspaceRoot：${workspaceRoot}（工作区根目录）
+- workspaceRoot：${workspaceRoot}（当前执行工作区根目录，可能由 CD Workspace 节点切换）
+- pipelineWorkspace：${pipelineWorkspace}（流水线所在工作区，写 AgentFlow 结果时使用）
 - flowName：${flowName}
 - uuid：${uuid}
 - instanceId：${instanceId}
@@ -28,5 +29,5 @@ ${taskBody}
 
 按上述任务完成执行，节点中如有写入文件的操作可以执行。任务完成后直接退出，结果由系统自动标记成功。**仅当任务明确失败时**，执行以下命令报告失败（`agentflow` 是可直接在终端运行的 CLI 命令）：
 ```bash
-agentflow apply -ai write-result ${workspaceRoot} ${flowName} ${uuid} ${instanceId} --json '{"status":"failed","message":"失败原因"}'
+agentflow apply -ai write-result ${pipelineWorkspace} ${flowName} ${uuid} ${instanceId} --json '{"status":"failed","message":"失败原因"}'
 ```

package/bin/lib/agent-runners.mjs

@@ -9,6 +9,30 @@ import { normalizeCursorModelForCli } from "./model-config.mjs";
 import { appendRunLogLine } from "./run-events.mjs";
 import { writeWithPrefix } from "./terminal.mjs";
 import { t } from "./i18n.mjs";
+import { readUserEnvObject } from "./user-env.mjs";
+import { outputNodeBasename } from "../pipeline/get-exec-id.mjs";
+
+function shouldPassCursorModelArg(model) {
+  const text = String(model || "").trim();
+  return text !== "" && !/^auto$/i.test(text);
+}
+
+function childEnv(options = {}, extra = {}) {
+  const optEnv = options && options.env && typeof options.env === "object" ? options.env : {};
+  const userId = optEnv.AGENTFLOW_USER_ID || process.env.AGENTFLOW_USER_ID || "";
+  return { ...process.env, ...readUserEnvObject(userId), ...optEnv, ...extra };
+}
+
+function writeAgentTextArtifacts(absResultPath, absRunDir, instanceId, text) {
+  const body = String(text ?? "").trim();
+  if (!body) return;
+  fs.mkdirSync(path.dirname(absResultPath), { recursive: true });
+  fs.writeFileSync(absResultPath, body + "\n", "utf-8");
+  if (!instanceId) return;
+  const slotPath = path.join(absRunDir, "output", instanceId, outputNodeBasename(instanceId, 1, "result"));
+  fs.mkdirSync(path.dirname(slotPath), { recursive: true });
+  fs.writeFileSync(slotPath, body + "\n", "utf-8");
+}
 
 /**
  * Run Cursor CLI with stream-json, forward events to stdout, return success/failure.
@@ -25,8 +49,11 @@ export function runCursorAgentForNode(
   const outputDir = instanceId ? path.join(absRunDir, "output", instanceId) : path.join(absRunDir, "output");
   if (instanceId) fs.mkdirSync(outputDir, { recursive: true });
   const absWorkspaceRoot = path.resolve(workspaceRoot);
+  const execWorkspaceRoot = path.resolve(options.execWorkspaceRoot || workspaceRoot);
   const replacements = {
-    workspaceRoot: absWorkspaceRoot,
+    workspaceRoot: execWorkspaceRoot,
+    executionWorkspaceRoot: execWorkspaceRoot,
+    pipelineWorkspace: absWorkspaceRoot,
     promptPath: absPromptPath,
     nodeContext: nodeContext ?? "",
     taskBody: taskBody ?? "",
@@ -61,11 +88,11 @@ export function runCursorAgentForNode(
 
   return new Promise((resolve, reject) => {
     const agentCmd = process.env.CURSOR_AGENT_CMD || "agent";
-    const args = ["--print", "--output-format", "stream-json", "--trust", "--workspace", workspaceRoot];
+    const args = ["--print", "--output-format", "stream-json", "--trust", "--workspace", execWorkspaceRoot];
     const approveMcps = process.env.AGENTFLOW_CURSOR_APPROVE_MCPS !== "0" && process.env.AGENTFLOW_CURSOR_APPROVE_MCPS !== "false";
     if (approveMcps) args.push("--approve-mcps");
     if (options.force) args.push("--force");
-    args.push("--model", model);
+    if (shouldPassCursorModelArg(model)) args.push("--model", model);
     args.push(promptText);
     if (options.flowName && options.uuid) {
       const argvLog = args.slice(0, -1).concat([`(prompt ${args[args.length - 1].length} chars)`]);
@@ -81,13 +108,15 @@ export function runCursorAgentForNode(
     }
     const useStderrInherit = process.env.AGENTFLOW_CURSOR_STDERR_INHERIT === "1" || process.env.AGENTFLOW_CURSOR_STDERR_INHERIT === "true";
     const child = spawn(agentCmd, args, {
-      cwd: workspaceRoot,
+      cwd: execWorkspaceRoot,
       stdio: ["ignore", "pipe", useStderrInherit ? "inherit" : "pipe"],
       shell: false,
+      env: childEnv(options),
     });
 
     let lastResult = null;
     let hadError = false;
+    const assistantTextChunks = [];
     const STDERR_CAP_BYTES = 1024 * 1024;
     const stderrChunks = [];
     let stderrTotalBytes = 0;
@@ -172,6 +201,7 @@ export function runCursorAgentForNode(
               .join("");
             if (text) {
               text = text.replace(/\\n/g, "\n").replace(/\\t/g, "\t");
+              assistantTextChunks.push(text);
               const out = mdStreamer.push(text);
               if (out) writeStdout(out);
             }
@@ -266,6 +296,7 @@ export function runCursorAgentForNode(
         reject(new Error(lastResult?.result || "Agent reported error."));
         return;
       }
+      writeAgentTextArtifacts(absResultPath, absRunDir, instanceId, assistantTextChunks.join("") || lastResult?.result || "");
       resolve();
     });
   });
@@ -286,8 +317,11 @@ export function runOpenCodeAgentForNode(
   const outputDir = instanceId ? path.join(absRunDir, "output", instanceId) : path.join(absRunDir, "output");
   if (instanceId) fs.mkdirSync(outputDir, { recursive: true });
   const absWorkspaceRoot = path.resolve(workspaceRoot);
+  const execWorkspaceRoot = path.resolve(options.execWorkspaceRoot || workspaceRoot);
   const replacements = {
-    workspaceRoot: absWorkspaceRoot,
+    workspaceRoot: execWorkspaceRoot,
+    executionWorkspaceRoot: execWorkspaceRoot,
+    pipelineWorkspace: absWorkspaceRoot,
     promptPath: absPromptPath,
     nodeContext: nodeContext ?? "",
     taskBody: taskBody ?? "",
@@ -325,16 +359,17 @@ export function runOpenCodeAgentForNode(
     if (model) {
       args.push("--model", model);
     }
-    args.push("--dir", workspaceRoot);
+    args.push("--dir", execWorkspaceRoot);
     args.push("--", promptText);
     const spawnOpts = {
-      cwd: workspaceRoot,
+      cwd: execWorkspaceRoot,
       stdio: ["ignore", "pipe", "pipe"],
       shell: false,
+      env: childEnv(options),
     };
     if (options.force) {
       spawnOpts.env = {
-        ...process.env,
+        ...spawnOpts.env,
         OPENCODE_CONFIG_CONTENT: JSON.stringify({
           permission: { external_directory: "allow" },
         }),
@@ -346,6 +381,7 @@ export function runOpenCodeAgentForNode(
 
     let stdoutLogBuf = "";
     let stderrLogBuf = "";
+    let stdoutCaptured = "";
 
     function drainLogBuf(buf, tag) {
       let idx;
@@ -372,7 +408,9 @@ export function runOpenCodeAgentForNode(
     child.stdout.on("data", (chunk) => {
       if (coloredPrefix) writeWithPrefix(process.stdout, chunk, coloredPrefix, agentContentColor);
       else process.stdout.write(agentContentColor(chunk));
-      stdoutLogBuf += String(chunk).replace(/\r\n/g, "\n").replace(/\r/g, "\n");
+      const normalizedChunk = String(chunk).replace(/\r\n/g, "\n").replace(/\r/g, "\n");
+      stdoutCaptured += normalizedChunk;
+      stdoutLogBuf += normalizedChunk;
       stdoutLogBuf = drainLogBuf(stdoutLogBuf, "opencode-stdout");
     });
 
@@ -404,6 +442,7 @@ export function runOpenCodeAgentForNode(
         reject(new Error(`OpenCode CLI exited ${code}.`));
         return;
       }
+      writeAgentTextArtifacts(absResultPath, absRunDir, instanceId, stripAnsi(stdoutCaptured));
       resolve();
     });
   });
@@ -426,8 +465,11 @@ export function runClaudeCodeAgentForNode(
   const outputDir = instanceId ? path.join(absRunDir, "output", instanceId) : path.join(absRunDir, "output");
   if (instanceId) fs.mkdirSync(outputDir, { recursive: true });
   const absWorkspaceRoot = path.resolve(workspaceRoot);
+  const execWorkspaceRoot = path.resolve(options.execWorkspaceRoot || workspaceRoot);
   const replacements = {
-    workspaceRoot: absWorkspaceRoot,
+    workspaceRoot: execWorkspaceRoot,
+    executionWorkspaceRoot: execWorkspaceRoot,
+    pipelineWorkspace: absWorkspaceRoot,
     promptPath: absPromptPath,
     nodeContext: nodeContext ?? "",
     taskBody: taskBody ?? "",
@@ -464,7 +506,7 @@ export function runClaudeCodeAgentForNode(
     const bypassPermissions =
       process.env.AGENTFLOW_CLAUDE_CODE_BYPASS_PERMISSIONS !== "0" &&
       process.env.AGENTFLOW_CLAUDE_CODE_BYPASS_PERMISSIONS !== "false";
-    const args = ["-p", "--output-format", "stream-json", "--verbose", "--add-dir", workspaceRoot];
+    const args = ["-p", "--output-format", "stream-json", "--verbose", "--add-dir", execWorkspaceRoot, "--add-dir", absWorkspaceRoot];
     if (bypassPermissions) args.push("--dangerously-skip-permissions");
     if (model) args.push("--model", model);
     args.push(promptText);
@@ -490,14 +532,16 @@ export function runClaudeCodeAgentForNode(
       process.env.AGENTFLOW_CLAUDE_CODE_STDERR_INHERIT === "1" ||
       process.env.AGENTFLOW_CLAUDE_CODE_STDERR_INHERIT === "true";
     const child = spawn(claudeCmd, args, {
-      cwd: workspaceRoot,
+      cwd: execWorkspaceRoot,
       stdio: ["ignore", "pipe", useStderrInherit ? "inherit" : "pipe"],
       shell: false,
+      env: childEnv(options),
     });
 
     let lastResult = null;
     let hadError = false;
     let sessionId = null;
+    const assistantTextChunks = [];
     const STDERR_CAP_BYTES = 1024 * 1024;
     const stderrChunks = [];
     let stderrTotalBytes = 0;
@@ -574,6 +618,7 @@ export function runClaudeCodeAgentForNode(
               if (!block || typeof block !== "object") continue;
               if (block.type === "text" && block.text) {
                 const text = normalizeStreamTextChunk(block.text);
+                assistantTextChunks.push(text);
                 const out = mdStreamer.push(text);
                 if (out) writeStdout(out);
               } else if (block.type === "thinking") {
@@ -642,6 +687,7 @@ export function runClaudeCodeAgentForNode(
         reject(new Error(String(msg)));
         return;
       }
+      writeAgentTextArtifacts(absResultPath, absRunDir, instanceId, assistantTextChunks.join("") || lastResult?.result || "");
       resolve();
     });
   });
@@ -743,7 +789,7 @@ export function runCursorAgentWithPrompt(cliWorkspace, promptText, options = {})
   const approveMcps = process.env.AGENTFLOW_CURSOR_APPROVE_MCPS !== "0" && process.env.AGENTFLOW_CURSOR_APPROVE_MCPS !== "false";
   if (approveMcps) args.push("--approve-mcps");
   args.push("--force");
-  args.push("--model", model);
+  if (shouldPassCursorModelArg(model)) args.push("--model", model);
   args.push(promptText);
 
   const useStderrInherit = process.env.AGENTFLOW_CURSOR_STDERR_INHERIT === "1" || process.env.AGENTFLOW_CURSOR_STDERR_INHERIT === "true";
@@ -751,6 +797,7 @@ export function runCursorAgentWithPrompt(cliWorkspace, promptText, options = {})
     cwd: ws,
     stdio: ["ignore", "pipe", useStderrInherit ? "inherit" : "pipe"],
     shell: false,
+    env: childEnv(options),
   });
 
   let lastResult = null;
@@ -939,10 +986,11 @@ export function runOpenCodeAgentWithPrompt(cliWorkspace, promptText, options = {
     cwd: ws,
     stdio: ["ignore", "pipe", "pipe"],
     shell: false,
+    env: childEnv(options),
   };
   if (options.force) {
     spawnOpts.env = {
-      ...process.env,
+      ...spawnOpts.env,
       OPENCODE_CONFIG_CONTENT: JSON.stringify({
         permission: { external_directory: "allow" },
       }),
@@ -1047,6 +1095,7 @@ export function runClaudeCodeAgentWithPrompt(cliWorkspace, promptText, options =
     cwd: ws,
     stdio: ["ignore", "pipe", useStderrInherit ? "inherit" : "pipe"],
     shell: false,
+    env: childEnv(options),
   });
 
   let lastResult = null;

package/bin/lib/api-runner.mjs

@@ -19,11 +19,23 @@ import { spawnSync } from "child_process";
 
 import { loadAgentPromptWithReplacements, stripYamlFrontmatter } from "./agents-path.mjs";
 import { appendRunLogLine } from "./run-events.mjs";
+import { outputNodeBasename } from "../pipeline/get-exec-id.mjs";
 
 const DEFAULT_OPENAI_BASE = "https://api.openai.com/v1";
 const MAX_TOOL_ROUNDS = parseInt(process.env.AGENTFLOW_API_MAX_ROUNDS ?? "30", 10) || 30;
 const MAX_TOKENS = parseInt(process.env.AGENTFLOW_API_MAX_TOKENS ?? "8192", 10) || 8192;
 
+function writeAgentTextArtifacts(absResultPath, absRunDir, instanceId, text) {
+  const body = String(text ?? "").trim();
+  if (!body || !absResultPath || !absRunDir) return;
+  fs.mkdirSync(path.dirname(absResultPath), { recursive: true });
+  fs.writeFileSync(absResultPath, body + "\n", "utf-8");
+  if (!instanceId) return;
+  const slotPath = path.join(absRunDir, "output", instanceId, outputNodeBasename(instanceId, 1, "result"));
+  fs.mkdirSync(path.dirname(slotPath), { recursive: true });
+  fs.writeFileSync(slotPath, body + "\n", "utf-8");
+}
+
 // ─── 工具定义 ────────────────────────────────────────────────────────────────
 
 const TOOL_DEFS = [
@@ -203,6 +215,7 @@ async function runOpenAiLoop(apiKey, baseUrl, model, systemPrompt, userContent,
     ...(systemPrompt ? [{ role: "system", content: systemPrompt }] : []),
     { role: "user", content: userContent },
   ];
+  let finalText = "";
 
   for (let round = 0; round < MAX_TOOL_ROUNDS; round++) {
     log(`[api/openai] round ${round + 1}`);
@@ -218,6 +231,7 @@ async function runOpenAiLoop(apiKey, baseUrl, model, systemPrompt, userContent,
       const txt = typeof msg.content === "string" ? msg.content : "";
       if (txt.trim()) options.onToolCall("assistant", txt.slice(0, 200));
     }
+    if (typeof msg.content === "string" && msg.content.trim()) finalText = msg.content;
 
     if (choice.finish_reason === "stop" || choice.finish_reason === "end_turn" || !msg.tool_calls?.length) {
       log(`[api/openai] finished (${choice.finish_reason ?? "no-tool-calls"})`);
@@ -240,10 +254,12 @@ async function runOpenAiLoop(apiKey, baseUrl, model, systemPrompt, userContent,
     }
     messages.push(...toolResults);
   }
+  return finalText;
 }
 
 async function runAnthropicLoop(apiKey, model, systemPrompt, userContent, workspaceRoot, log, options) {
   const messages = [{ role: "user", content: userContent }];
+  let finalText = "";
 
   for (let round = 0; round < MAX_TOOL_ROUNDS; round++) {
     log(`[api/anthropic] round ${round + 1}`);
@@ -256,6 +272,8 @@ async function runAnthropicLoop(apiKey, model, systemPrompt, userContent, worksp
       const textBlock = resp.content?.find((b) => b.type === "text");
       if (textBlock?.text) options.onToolCall("assistant", textBlock.text.slice(0, 200));
     }
+    const textBlock = resp.content?.find((b) => b.type === "text");
+    if (textBlock?.text?.trim()) finalText = textBlock.text;
 
     if (resp.stop_reason === "end_turn" || resp.stop_reason === "stop_sequence") {
       log(`[api/anthropic] finished (${resp.stop_reason})`);
@@ -280,6 +298,7 @@ async function runAnthropicLoop(apiKey, model, systemPrompt, userContent, worksp
     }
     messages.push({ role: "user", content: toolResults });
   }
+  return finalText;
 }
 
 // ─── 公共解析函数 ─────────────────────────────────────────────────────────────
@@ -313,8 +332,11 @@ export function parseApiModel(str) {
  *   uuid       — 用于日志
  *   onToolCall — (subtype: string, name: string) => void  供 spinner 展示
  */
-export async function runApiAgentForNode(workspaceRoot, { promptPath, nodeContext, taskBody, subagent, instanceId }, options = {}) {
+export async function runApiAgentForNode(workspaceRoot, { promptPath, nodeContext, taskBody, intermediatePath, resultPathRel, subagent, instanceId }, options = {}) {
   const absRoot = path.resolve(workspaceRoot);
+  const execRoot = path.resolve(options.execWorkspaceRoot || workspaceRoot);
+  const absRunDir = intermediatePath ? path.resolve(workspaceRoot, intermediatePath) : "";
+  const absResultPath = absRunDir && resultPathRel ? path.join(absRunDir, resultPathRel) : "";
   const flowName = options.flowName ?? null;
   const uuid = options.uuid ?? null;
 
@@ -329,7 +351,9 @@ export async function runApiAgentForNode(workspaceRoot, { promptPath, nodeContex
 
   // ── 读取 Agent 角色提示，注入 nodeContext/taskBody ────────────────────────
   const replacements = {
-    workspaceRoot: absRoot,
+    workspaceRoot: execRoot,
+    executionWorkspaceRoot: execRoot,
+    pipelineWorkspace: absRoot,
     nodeContext: nodeContext ?? "",
     taskBody: taskBody ?? "",
     flowName: flowName ?? "",
@@ -349,12 +373,14 @@ export async function runApiAgentForNode(workspaceRoot, { promptPath, nodeContex
   if (provider === "anthropic") {
     const key = process.env.ANTHROPIC_API_KEY;
     if (!key) throw new Error("[api-runner] ANTHROPIC_API_KEY is required for api:anthropic/* models");
-    await runAnthropicLoop(key, model, systemPrompt, userContent, absRoot, log, options);
+    const finalText = await runAnthropicLoop(key, model, systemPrompt, userContent, execRoot, log, options);
+    writeAgentTextArtifacts(absResultPath, absRunDir, instanceId, finalText);
   } else {
     const key = process.env.OPENAI_API_KEY;
     if (!key) throw new Error("[api-runner] OPENAI_API_KEY is required for api:openai/* models");
     const baseUrl = (process.env.OPENAI_BASE_URL ?? DEFAULT_OPENAI_BASE).trim();
-    await runOpenAiLoop(key, baseUrl, model, systemPrompt, userContent, absRoot, log, options);
+    const finalText = await runOpenAiLoop(key, baseUrl, model, systemPrompt, userContent, execRoot, log, options);
+    writeAgentTextArtifacts(absResultPath, absRunDir, instanceId, finalText);
   }
 
   log(`done instanceId=${instanceId ?? "-"}`);

package/bin/lib/apply.mjs

@@ -23,6 +23,7 @@ import { formatDuration } from "./terminal.mjs";
 import { printEntryAndFlowFiles, printNodeStatusTable, runValidateFlowAndExitIfInvalid } from "./ui-print.mjs";
 import { clearApplyActiveLock, writeApplyActiveLock } from "./run-apply-active-lock.mjs";
 import { ensureReference, findFlowNameByUuid, getFlowDir, getRunDir } from "./workspace.mjs";
+import { readUserEnvObject } from "./user-env.mjs";
 
 const PARALLEL_PREFIX_COLORS = [
   (s) => chalk.cyan(s),
@@ -344,11 +345,11 @@ ${currentContent}
                     fs.mkdirSync(path.dirname(tmpPromptFile), { recursive: true });
                     fs.writeFileSync(tmpPromptFile, fullPrompt, "utf-8");
                     
-                    const result = spawnSync(opencodeCmd, ["--prompt-file", tmpPromptFile, "--print"], {
-                      cwd: workspaceRoot,
-                      env: { ...process.env, OPENCODE_NON_INTERACTIVE: "1" },
-                      stdio: ["ignore", "pipe", "pipe"],
-                    });
+	                    const result = spawnSync(opencodeCmd, ["--prompt-file", tmpPromptFile, "--print"], {
+	                      cwd: workspaceRoot,
+	                      env: { ...process.env, ...readUserEnvObject(process.env.AGENTFLOW_USER_ID || ""), OPENCODE_NON_INTERACTIVE: "1" },
+	                      stdio: ["ignore", "pipe", "pipe"],
+	                    });
                     
                     try { fs.unlinkSync(tmpPromptFile); } catch (_) {}