fix(agent): stabilize large tool results

Merge PR #1 after CLI, contract, test, and container gates passed.

.gitea/workflows/package.yml

@@ -14,7 +14,7 @@ jobs:
       dockerfile: Dockerfile
       build_context: .
       cache_image: gitea.waternetwork.cn/orgtjwater/tjwateragent:ci-cache
-      test_target: build
+      test_target: test
       deploy_service: agent
       deploy_host: 192.168.1.114
     secrets:

.opencode/agents/instruction.md

@@ -2,10 +2,14 @@
 description: TJWater Agent，用于供水网络分析和操作员工作流
 mode: primary
 model: deepseek/deepseek-v4-flash
-temperature: 0.2
 ---
 你是 TJWater 供水管网分析 Agent，运用水力专业知识，回复用户时使用简体中文，内容要求简洁准确。
 
+## 回复要求
+
+- 工具执行期间不输出过程说明，全部完成后只回复最终结果
+- 直接给出结论、关键数据和可执行建议，默认仅展示最重要的 Top 5；数据不足或任务失败时简要说明影响和下一步
+
 ## 工作流生命周期
 
 Skills 树是**动态生长的**——工作流不是预置的，而是从实际任务中沉淀出来的：

AGENTS.md

@@ -37,3 +37,5 @@ PRs should describe runtime behavior changes, list `bun run check` and any test
 ## Security & Configuration Tips
 
 Do not commit `.env`, logs, session transcripts, generated result references, or `node_modules/`. Keep registry and deploy credentials in Gitea secrets.
+
+Automatic approval for `glob` and `grep` must remain limited to canonical paths inside an explicit safe workspace subtree. Broad workspace-root searches, symlink escapes, external paths, and `.env`, `data/`, or `logs/` targets must stay interactive.

Dockerfile

@@ -65,6 +65,15 @@ COPY cli ./cli
 COPY .opencode ./.opencode
 RUN bun run check
 
+FROM build AS test
+RUN apt-get update && apt-get install -y --no-install-recommends nodejs && \
+    rm -rf /var/lib/apt/lists/*
+COPY contracts ./contracts
+COPY node-tests ./node-tests
+COPY scripts ./scripts
+COPY tests ./tests
+RUN bun run test:ci
+
 FROM build AS runner
 WORKDIR /app
 

README.md

@@ -86,11 +86,13 @@ TJWATER_API_BASE_URL=http://127.0.0.1:8000
 
 `opencode.json` 已启用 `experimental.continue_loop_on_deny`。用户拒绝权限请求后，OpenCode V1 会把拒绝结果交还给 Agent，让其尝试无需该权限的替代方案，而不是直接结束本轮执行。
 
-前端提供三种整体权限模式：“请求批准”只执行 OpenCode 明确允许的白名单，其他权限请求逐次交给用户确认；“自动批准”额外自动放行低风险业务工具，其他请求仍需确认；“始终允许”自动放行当前对话中所有未被 OpenCode 明确禁止的权限请求。自动放行统一使用单次批准，切换整体模式后立即恢复对应策略，不会写入持久授权。
+前端提供三种整体权限模式：“请求批准”只执行 OpenCode 明确允许的白名单，其他权限请求逐次交给用户确认；“自动批准”额外自动放行低风险业务工具、skill，以及真实路径位于工作区安全子树且不涉及 `.env`、`data/`、`logs/` 的 glob/grep；工作区根目录的宽泛搜索仍需确认。“始终允许”自动放行当前对话中所有未被 OpenCode 明确禁止的权限请求。自动放行统一使用单次批准，切换整体模式后立即恢复对应策略，不会写入持久授权。
 
 单次权限请求支持“允许一次”“保存授权”和“拒绝”。“保存授权”使用 OpenCode 的 `always` 回复，仅保存 OpenCode 为本次请求建议的权限范围，并只在当前 OpenCode 会话内生效。外部目录以及 `.env`、`data/`、`logs/` 路径仍由静态配置明确禁止，三种整体模式都不能绕过这些拒绝规则。
 
-`store_render_ref` 只会从 `RESULT_REF_IMPORT_DIR`（默认 `./data/result-imports`）导入包装格式 JSON。文件必须包含 `metadata`、`location.file_path` 和 `data`，且真实路径不能越出导入目录；单文件默认上限为 64 MiB，成功导入后源包装文件会被删除。
+`store_render_ref` 只会从 `RESULT_REF_IMPORT_DIR`（默认 `./data/result-imports`）导入包装格式 JSON。文件必须包含 `metadata`、`location.file_path` 和 `data`，且真实路径不能越出导入目录；单文件默认上限为 128 MiB，成功导入后源包装文件会被删除。
+
+CLI 桥接层对 stdout 设置独立的 128 MiB 硬上限（`MAX_CLI_OUTPUT_BYTES`）；stderr 最多保留 256 KiB（`MAX_CLI_STDERR_BYTES`），超出后截断但不会终止 CLI。`MAX_INLINE_RESULT_BYTES`（默认 12000 字节）只控制 OpenCode 的内联阈值，较大结果由 OpenCode 写入标准 `tool-output` 目录。Agent 启动时及后续定期清理其中超过 `RESULT_REF_TTL_HOURS`（默认 7 天）的 `tool_*` 文件。
 
 ## 配置与安全
 

cli/src/commands/network.ts

@@ -1,5 +1,6 @@
-import { emitApi } from "../core/http.js";
+import { emitApi, requestAllPages } from "../core/http.js";
-import { parseOptions, requiredString } from "../core/options.js";
+import { optionalNumber, parseOptions, requiredString } from "../core/options.js";
+import { success } from "../core/output.js";
 import type { HandlerMap, RuntimeContext } from "../core/types.js";
 
 function apiGet(ctx: RuntimeContext, argv: string[], summary: string, path: string, key: string): Promise<void> {
@@ -7,20 +8,28 @@ function apiGet(ctx: RuntimeContext, argv: string[], summary: string, path: stri
   return emitApi(ctx, summary, { method: "GET", path, params: { [key]: requiredString(values, key) }, requireProject: true });
 }
 
-function apiGetAll(ctx: RuntimeContext, summary: string, path: string): Promise<void> {
+async function apiGetAll(ctx: RuntimeContext, argv: string[], summary: string, path: string): Promise<void> {
-  return emitApi(ctx, summary, { method: "GET", path, requireProject: true });
+  const { values } = parseOptions(argv, { limit: "integer", "page-size": "integer" });
+  const requestedPageSize = optionalNumber(values, "page-size") ?? optionalNumber(values, "limit") ?? 1000;
+  const pageSize = Math.min(1000, Math.max(1, requestedPageSize));
+  const [data, durationMs] = await requestAllPages(
+    ctx,
+    { method: "GET", path, requireProject: true },
+    pageSize,
+  );
+  success(summary, data, ctx, durationMs);
 }
 
 export const networkHandlers: HandlerMap = {
   "network get-junction-properties": (ctx, argv) => apiGet(ctx, argv, "读取节点属性成功", "/junctions/properties", "junction"),
   "network get-pipe-properties": (ctx, argv) => apiGet(ctx, argv, "读取管道属性成功", "/pipes/properties", "pipe"),
-  "network get-all-pipes-properties": (ctx) => apiGetAll(ctx, "读取全部管道属性成功", "/pipes"),
+  "network get-all-pipes-properties": (ctx, argv) => apiGetAll(ctx, argv, "读取全部管道属性成功", "/pipes"),
   "network get-reservoir-properties": (ctx, argv) => apiGet(ctx, argv, "读取水库属性成功", "/reservoirs/properties", "reservoir"),
-  "network get-all-reservoirs-properties": (ctx) => apiGetAll(ctx, "读取全部水库属性成功", "/reservoirs"),
+  "network get-all-reservoirs-properties": (ctx, argv) => apiGetAll(ctx, argv, "读取全部水库属性成功", "/reservoirs"),
   "network get-tank-properties": (ctx, argv) => apiGet(ctx, argv, "读取水箱属性成功", "/tanks/properties", "tank"),
-  "network get-all-tanks-properties": (ctx) => apiGetAll(ctx, "读取全部水箱属性成功", "/tanks"),
+  "network get-all-tanks-properties": (ctx, argv) => apiGetAll(ctx, argv, "读取全部水箱属性成功", "/tanks"),
   "network get-pump-properties": (ctx, argv) => apiGet(ctx, argv, "读取水泵属性成功", "/pumps/properties", "pump"),
-  "network get-all-pumps-properties": (ctx) => apiGetAll(ctx, "读取全部水泵属性成功", "/pumps"),
+  "network get-all-pumps-properties": (ctx, argv) => apiGetAll(ctx, argv, "读取全部水泵属性成功", "/pumps"),
   "network get-valve-properties": (ctx, argv) => apiGet(ctx, argv, "读取阀门属性成功", "/valves/properties", "valve"),
-  "network get-all-valves-properties": (ctx) => apiGetAll(ctx, "读取全部阀门属性成功", "/valves"),
+  "network get-all-valves-properties": (ctx, argv) => apiGetAll(ctx, argv, "读取全部阀门属性成功", "/valves"),
 };

cli/src/core/http.ts

@@ -76,6 +76,100 @@ export async function requestJson(ctx: RuntimeContext, request: RequestOptions):
   return [payload, durationMs];
 }
 
+export async function requestAllPages(
+  ctx: RuntimeContext,
+  request: RequestOptions,
+  pageSize: number,
+): Promise<[unknown[], number]> {
+  const items: unknown[] = [];
+  let durationMs = 0;
+  let offset = 0;
+  let expectedTotal: number | null = null;
+
+  while (expectedTotal === null || offset < expectedTotal) {
+    const [payload, pageDurationMs] = await requestJson(ctx, {
+      ...request,
+      params: {
+        ...request.params,
+        limit: pageSize,
+        offset,
+      },
+    });
+    durationMs += pageDurationMs;
+    const page = normalizePage(payload);
+    if (!page) {
+      throw new CliError(
+        "服务端错误",
+        "INVALID_PAGINATION_RESPONSE",
+        "backend collection response must contain items, total, limit, and offset",
+        7,
+        false,
+        payload,
+      );
+    }
+    if (expectedTotal === null) {
+      expectedTotal = page.total;
+    } else if (page.total !== expectedTotal) {
+      throw new CliError(
+        "服务端错误",
+        "PAGINATION_TOTAL_CHANGED",
+        `backend collection total changed from ${expectedTotal} to ${page.total}`,
+        7,
+        true,
+      );
+    }
+    if (page.offset !== offset) {
+      throw new CliError(
+        "服务端错误",
+        "PAGINATION_OFFSET_MISMATCH",
+        `backend collection returned offset ${page.offset}, expected ${offset}`,
+        7,
+        true,
+      );
+    }
+    if (page.items.length === 0 && offset < expectedTotal) {
+      throw new CliError(
+        "服务端错误",
+        "PAGINATION_STALLED",
+        `backend collection returned an empty page at offset ${offset} before total ${expectedTotal}`,
+        7,
+        true,
+      );
+    }
+    items.push(...page.items);
+    offset += page.items.length;
+  }
+
+  return [items.slice(0, expectedTotal ?? 0), durationMs];
+}
+
+function normalizePage(
+  payload: unknown,
+): { items: unknown[]; limit: number; offset: number; total: number } | null {
+  if (!payload || typeof payload !== "object" || Array.isArray(payload)) return null;
+  const page = payload as Record<string, unknown>;
+  if (
+    !Array.isArray(page.items) ||
+    typeof page.limit !== "number" ||
+    !Number.isInteger(page.limit) ||
+    typeof page.offset !== "number" ||
+    !Number.isInteger(page.offset) ||
+    typeof page.total !== "number" ||
+    !Number.isInteger(page.total) ||
+    page.limit <= 0 ||
+    page.offset < 0 ||
+    page.total < 0
+  ) {
+    return null;
+  }
+  return {
+    items: page.items,
+    limit: page.limit,
+    offset: page.offset,
+    total: page.total,
+  };
+}
+
 function mapStatus(status: number): number {
   if (status === 400 || status === 422) return 2;
   if (status === 401) return 3;

cli/src/help/docs.ts

@@ -32,15 +32,15 @@ type CommandSpec = readonly [path: string, summary: string, options: readonly st
 const commandSpecs: readonly CommandSpec[] = [
   ["network get-junction-properties", "读取节点属性", ["--junction <JUNCTION>"], ["tjwater-cli network get-junction-properties --junction J1"]],
   ["network get-pipe-properties", "读取管道属性", ["--pipe <PIPE>"], ["tjwater-cli network get-pipe-properties --pipe P1"]],
-  ["network get-all-pipes-properties", "读取全部管道属性", [], ["tjwater-cli network get-all-pipes-properties"]],
+  ["network get-all-pipes-properties", "读取全部管道属性", ["[--page-size <PAGE_SIZE>]"], ["tjwater-cli network get-all-pipes-properties"]],
   ["network get-reservoir-properties", "读取水库属性", ["--reservoir <RESERVOIR>"], ["tjwater-cli network get-reservoir-properties --reservoir R1"]],
-  ["network get-all-reservoirs-properties", "读取全部水库属性", [], ["tjwater-cli network get-all-reservoirs-properties"]],
+  ["network get-all-reservoirs-properties", "读取全部水库属性", ["[--page-size <PAGE_SIZE>]"], ["tjwater-cli network get-all-reservoirs-properties"]],
   ["network get-tank-properties", "读取水箱属性", ["--tank <TANK>"], ["tjwater-cli network get-tank-properties --tank T1"]],
-  ["network get-all-tanks-properties", "读取全部水箱属性", [], ["tjwater-cli network get-all-tanks-properties"]],
+  ["network get-all-tanks-properties", "读取全部水箱属性", ["[--page-size <PAGE_SIZE>]"], ["tjwater-cli network get-all-tanks-properties"]],
   ["network get-pump-properties", "读取水泵属性", ["--pump <PUMP>"], ["tjwater-cli network get-pump-properties --pump PU1"]],
-  ["network get-all-pumps-properties", "读取全部水泵属性", [], ["tjwater-cli network get-all-pumps-properties"]],
+  ["network get-all-pumps-properties", "读取全部水泵属性", ["[--page-size <PAGE_SIZE>]"], ["tjwater-cli network get-all-pumps-properties"]],
   ["network get-valve-properties", "读取阀门属性", ["--valve <VALVE>"], ["tjwater-cli network get-valve-properties --valve V1"]],
-  ["network get-all-valves-properties", "读取全部阀门属性", [], ["tjwater-cli network get-all-valves-properties"]],
+  ["network get-all-valves-properties", "读取全部阀门属性", ["[--page-size <PAGE_SIZE>]"], ["tjwater-cli network get-all-valves-properties"]],
   ["component option schema", "读取选项 schema", ["--kind <KIND>", "[--pump <PUMP>]"], ["tjwater-cli component option schema --kind time", "tjwater-cli component option schema --kind energy", "tjwater-cli component option schema --kind pump-energy --pump PUMP1", "tjwater-cli component option schema --kind network"]],
   ["component option get", "读取选项属性", ["--kind <KIND>", "[--pump <PUMP>]"], ["tjwater-cli component option get --kind time", "tjwater-cli component option get --kind energy", "tjwater-cli component option get --kind pump-energy --pump PUMP1", "tjwater-cli component option get --kind network"]],
   ["simulation run", "触发指定绝对时间的模拟运行", ["--start-time <START_TIME>", "--duration <DURATION>"], ["tjwater-cli simulation run --start-time 2025-01-02T03:04:05+08:00 --duration 30"], ["tjwater-cli data timeseries realtime links --start-time 2025-01-02T03:04:05+08:00 --end-time 2025-01-02T03:34:05+08:00", "tjwater-cli data timeseries realtime nodes --start-time 2025-01-02T03:04:05+08:00 --end-time 2025-01-02T03:34:05+08:00"]],

node-tests/cli/tjwaterCli.node.mjs

@@ -104,7 +104,11 @@ async function startJsonServer(responseData) {
       url: req.url,
     });
     res.setHeader("content-type", "application/json");
-    res.end(JSON.stringify(responseData));
+    res.end(
+      JSON.stringify(
+        typeof responseData === "function" ? responseData(req) : responseData,
+      ),
+    );
   });
 
   await new Promise((resolveListen, reject) => {
@@ -138,7 +142,20 @@ function normalizeSeenRequest(request) {
   };
 }
 
-async function runAgainstServer(name, runner, args, auth, responseData = { accepted: true }) {
+function defaultContractResponse(req) {
+  const url = new URL(req.url, "http://127.0.0.1");
+  if (["/api/v1/pipes", "/api/v1/reservoirs", "/api/v1/tanks", "/api/v1/pumps", "/api/v1/valves"].includes(url.pathname)) {
+    return {
+      items: [],
+      limit: Number(url.searchParams.get("limit")),
+      offset: Number(url.searchParams.get("offset")),
+      total: 0,
+    };
+  }
+  return { accepted: true };
+}
+
+async function runAgainstServer(name, runner, args, auth, responseData = defaultContractResponse) {
   const server = await startJsonServer(responseData);
   try {
     const result = await runner(["--auth-stdin", ...args], { ...auth, server: server.url });
@@ -232,6 +249,52 @@ test("sends auth headers and simulation body through the backend API contract",
   }
 });
 
+test("get-all network commands collect every backend page", async () => {
+  const items = Array.from({ length: 2_005 }, (_, index) => ({
+    id: `P${index + 1}`,
+    node1: `N${index + 1}`,
+    node2: `N${index + 2}`,
+  }));
+  const server = await startJsonServer((req) => {
+    const url = new URL(req.url, "http://127.0.0.1");
+    const limit = Number(url.searchParams.get("limit") ?? 100);
+    const offset = Number(url.searchParams.get("offset") ?? 0);
+    return {
+      items: items.slice(offset, offset + limit),
+      limit,
+      offset,
+      total: items.length,
+    };
+  });
+
+  try {
+    const result = await runCli(
+      ["--auth-stdin", "network", "get-all-pipes-properties"],
+      {
+        server: server.url,
+        access_token: "token-1",
+        project_id: "project-1",
+      },
+    );
+
+    assert.equal(result.exitCode, 0, result.stderr);
+    const payload = parseJsonResult(result);
+    assert.equal(payload.data.length, items.length);
+    assert.deepEqual(payload.data[0], items[0]);
+    assert.deepEqual(payload.data.at(-1), items.at(-1));
+    assert.deepEqual(
+      server.seen.map((request) => normalizeSeenRequest(request).query),
+      [
+        { limit: "1000", offset: "0" },
+        { limit: "1000", offset: "1000" },
+        { limit: "1000", offset: "2000" },
+      ],
+    );
+  } finally {
+    await server.close();
+  }
+});
+
 test("uses project scoped headers for realtime data commands", async () => {
   const server = await startJsonServer([{ id: "P1" }]);
   try {

package.json

@@ -8,7 +8,9 @@
     "install:opencode": "bun install --cwd .opencode",
     "typecheck": "tsc --noEmit -p tsconfig.json",
     "typecheck:opencode": "bun run --cwd .opencode typecheck",
+    "test": "bun test tests",
     "test:cli": "node --test node-tests/cli/*.node.mjs",
+    "test:ci": "bun run contract:check && /usr/bin/node --test node-tests/cli/*.node.mjs && bun test tests",
     "dev": "bun --watch src/server.ts",
     "build": "bun run check",
     "check": "bun run typecheck && bun run typecheck:opencode",

src/cli/executeCliCommand.ts

@@ -0,0 +1,215 @@
+import { spawn } from "node:child_process";
+
+import { type RuntimeSessionContext } from "../runtime/sessionContext.js";
+
+type OutputStream = "stdout" | "stderr";
+
+export type CliExecutionResult = {
+  outcome: "completed" | "timeout" | "output_limit";
+  exitCode: number | null;
+  signal: NodeJS.Signals | null;
+  status: number;
+  stderr: string;
+  stderrTruncated: boolean;
+  stdout: string;
+  exceededStream?: OutputStream;
+};
+
+type ExecuteCliCommandOptions = {
+  apiBaseUrl: string;
+  cliPath: string;
+  maxStderrBytes: number;
+  maxStdoutBytes: number;
+  terminationGraceMs?: number;
+};
+
+const getCompletedStatus = (exitCode: number | null, stdout: string) => {
+  let errorCode = "";
+  try {
+    const payload = JSON.parse(stdout) as { error?: { code?: unknown } };
+    errorCode =
+      typeof payload.error?.code === "string" ? payload.error.code : "";
+  } catch {
+    errorCode = "";
+  }
+
+  if (errorCode === "HTTP_401" || errorCode === "UNAUTHENTICATED") {
+    return 401;
+  }
+  if (errorCode === "HTTP_403") {
+    return 403;
+  }
+  return exitCode === 0 ? 200 : 502;
+};
+
+export const executeCliCommand = async (
+  context: RuntimeSessionContext,
+  command: string,
+  timeoutSec: number,
+  options: ExecuteCliCommandOptions,
+): Promise<CliExecutionResult> => {
+  if (!Number.isSafeInteger(options.maxStdoutBytes) || options.maxStdoutBytes <= 0) {
+    throw new Error("maxStdoutBytes must be a positive safe integer");
+  }
+  if (!Number.isSafeInteger(options.maxStderrBytes) || options.maxStderrBytes <= 0) {
+    throw new Error("maxStderrBytes must be a positive safe integer");
+  }
+
+  const child = spawn(
+    options.cliPath,
+    ["--auth-stdin", ...command.split(/\s+/).filter(Boolean)],
+    { stdio: ["pipe", "pipe", "pipe"] },
+  );
+  const stdoutChunks: Buffer[] = [];
+  const stderrChunks: Buffer[] = [];
+  let stdoutBytes = 0;
+  let stderrBytes = 0;
+  let stderrTruncated = false;
+  let terminationReason:
+    | "timeout"
+    | "output_limit"
+    | "execution_error"
+    | null = null;
+  let exceededStream: OutputStream | undefined;
+  let terminationStarted = false;
+  let settled = false;
+  let executionError: Error | null = null;
+  let forceKillTimer: ReturnType<typeof setTimeout> | undefined;
+
+  const result = await new Promise<CliExecutionResult>((resolve, reject) => {
+    const cleanup = () => {
+      clearTimeout(timeoutTimer);
+      if (forceKillTimer) {
+        clearTimeout(forceKillTimer);
+      }
+    };
+
+    const terminate = (
+      reason: "timeout" | "output_limit" | "execution_error",
+    ) => {
+      if (terminationStarted) {
+        return;
+      }
+      terminationStarted = true;
+      terminationReason = reason;
+
+      if (child.exitCode === null && child.signalCode === null) {
+        child.kill("SIGTERM");
+      }
+      forceKillTimer = setTimeout(() => {
+        if (child.exitCode === null && child.signalCode === null) {
+          child.kill("SIGKILL");
+        }
+      }, options.terminationGraceMs ?? 1500);
+    };
+
+    const captureStdout = (data: Buffer) => {
+      if (terminationReason) {
+        return;
+      }
+      if (stdoutBytes + data.length > options.maxStdoutBytes) {
+        exceededStream = "stdout";
+        terminate("output_limit");
+        return;
+      }
+      stdoutChunks.push(data);
+      stdoutBytes += data.length;
+    };
+
+    const captureStderr = (data: Buffer) => {
+      if (terminationReason || stderrTruncated) {
+        return;
+      }
+      const remainingBytes = options.maxStderrBytes - stderrBytes;
+      if (data.length > remainingBytes) {
+        if (remainingBytes > 0) {
+          stderrChunks.push(data.subarray(0, remainingBytes));
+          stderrBytes += remainingBytes;
+        }
+        stderrTruncated = true;
+        return;
+      }
+      stderrChunks.push(data);
+      stderrBytes += data.length;
+    };
+
+    const timeoutTimer = setTimeout(() => {
+      if (child.exitCode === null && child.signalCode === null) {
+        terminate("timeout");
+      }
+    }, timeoutSec * 1000);
+
+    child.stdout.on("data", captureStdout);
+    child.stderr.on("data", captureStderr);
+    child.stdin.on("error", (error) => {
+      if (terminationReason === null) {
+        executionError = error;
+        terminate("execution_error");
+      }
+    });
+    child.on("error", (error) => {
+      if (terminationReason === null) {
+        executionError = error;
+        terminate("execution_error");
+      }
+    });
+    child.on("close", (exitCode, signal) => {
+      if (settled) {
+        return;
+      }
+      settled = true;
+      cleanup();
+      if (terminationReason === "timeout") {
+        resolve({
+          outcome: "timeout",
+          exitCode,
+          signal,
+          status: 504,
+          stderr: "",
+          stderrTruncated,
+          stdout: "",
+        });
+        return;
+      }
+      if (executionError) {
+        reject(executionError);
+        return;
+      }
+      if (terminationReason === "output_limit") {
+        resolve({
+          outcome: "output_limit",
+          exceededStream,
+          exitCode,
+          signal,
+          status: 502,
+          stderr: "",
+          stderrTruncated,
+          stdout: "",
+        });
+        return;
+      }
+
+      const stdout = Buffer.concat(stdoutChunks, stdoutBytes).toString("utf-8");
+      const stderr = Buffer.concat(stderrChunks, stderrBytes).toString("utf-8");
+      resolve({
+        outcome: "completed",
+        exitCode,
+        signal,
+        status: getCompletedStatus(exitCode, stdout),
+        stderr,
+        stderrTruncated,
+        stdout,
+      });
+    });
+
+    child.stdin.end(
+      JSON.stringify({
+        server: options.apiBaseUrl,
+        access_token: context.accessToken,
+        project_id: context.projectId,
+      }),
+    );
+  });
+
+  return result;
+};

src/config.ts

@@ -61,8 +61,20 @@ const envSchema = z
     TJWATER_API_BASE_URL: z.string().default("http://127.0.0.1:8000"),
     // 代理调用 TJWater 后端 API 的超时时间（毫秒）。
     TJWATER_API_TIMEOUT_MS: z.coerce.number().int().positive().default(30000),
-    // 后端结果在直接内联返回给模型前允许的最大字节数。
+    // OpenCode 工具结果以内联形式返回给模型的阈值；更大的结果由 OpenCode 落盘。
     MAX_INLINE_RESULT_BYTES: z.coerce.number().int().positive().default(12000),
+    // 单次 tjwater-cli stdout 的硬上限；超过后终止子进程。
+    MAX_CLI_OUTPUT_BYTES: z.coerce
+      .number()
+      .int()
+      .positive()
+      .default(128 * 1024 * 1024),
+    // 单次 tjwater-cli stderr 最多保留的字节数；超过后截断但不终止进程。
+    MAX_CLI_STDERR_BYTES: z.coerce
+      .number()
+      .int()
+      .positive()
+      .default(256 * 1024),
     // 生成结果 preview 时最多抽样的条目数。
     MAX_PREVIEW_SAMPLE_ITEMS: z.coerce.number().int().positive().default(3),
     // memory 持久化存储目录。
@@ -110,7 +122,7 @@ const envSchema = z
       .number()
       .int()
       .positive()
-      .default(64 * 1024 * 1024),
+      .default(128 * 1024 * 1024),
     // result_ref 保留时长（小时）。
     RESULT_REF_TTL_HOURS: z.coerce.number().int().positive().default(168),
     // 定时清理过期 result_ref 的扫描周期（毫秒）。

src/routes/chatPermissionPolicy.ts

@@ -1,5 +1,14 @@
+import { lstatSync, readdirSync, realpathSync } from "node:fs";
+import { isAbsolute, relative, resolve, sep } from "node:path";
+
 export type ApprovalMode = "request" | "auto" | "always";
 
+export type PermissionApprovalContext = {
+  metadata?: Record<string, unknown>;
+  patterns?: readonly string[];
+  workspaceRoot?: string;
+};
+
 const lowRiskToolPermissions = new Set([
   "apply_layer_style",
   "geocode",
@@ -12,10 +21,31 @@ const lowRiskToolPermissions = new Set([
   "zoom_to_map",
 ]);
 
+const lowRiskSearchRootNames = new Set([
+  ".opencode",
+  "cli",
+  "contracts",
+  "node-tests",
+  "scripts",
+  "src",
+  "tests",
+]);
+
 const normalizePermission = (permission: string) => permission.trim().toLowerCase();
 
-export const canAutoApprovePermission = (permission: string): boolean => {
+export const canAutoApprovePermission = (
+  permission: string,
+  context: PermissionApprovalContext = {},
+): boolean => {
   const normalized = normalizePermission(permission);
+  if (normalized === "skill") {
+    return true;
+  }
+
+  if (normalized === "glob" || normalized === "grep") {
+    return isSafeWorkspaceSearch(normalized, context);
+  }
+
   if (lowRiskToolPermissions.has(normalized)) {
     return true;
   }
@@ -30,6 +60,7 @@ export const canAutoApprovePermission = (permission: string): boolean => {
 export const resolvePermissionApproval = (
   approvalMode: ApprovalMode,
   permission: string,
+  context: PermissionApprovalContext = {},
 ) => {
   if (approvalMode === "always") {
     return {
@@ -40,7 +71,7 @@ export const resolvePermissionApproval = (
     } as const;
   }
 
-  if (approvalMode === "auto" && canAutoApprovePermission(permission)) {
+  if (approvalMode === "auto" && canAutoApprovePermission(permission, context)) {
     return {
       autoApprove: true,
       title: "已自动批准低风险权限",
@@ -54,3 +85,138 @@ export const resolvePermissionApproval = (
     detail: undefined,
   } as const;
 };
+
+const isSafeWorkspaceSearch = (
+  permission: "glob" | "grep",
+  context: PermissionApprovalContext,
+): boolean => {
+  const workspaceRoot = context.workspaceRoot?.trim();
+  if (!workspaceRoot) {
+    return false;
+  }
+
+  const requestedPath =
+    typeof context.metadata?.path === "string" && context.metadata.path.trim()
+      ? context.metadata.path
+      : workspaceRoot;
+  let root: string;
+  let searchRoot: string;
+  try {
+    root = realpathSync.native(resolve(workspaceRoot));
+    searchRoot = realpathSync.native(resolve(root, requestedPath));
+  } catch {
+    return false;
+  }
+
+  let relativePath = relative(root, searchRoot);
+  if (
+    relativePath === ".." ||
+    relativePath.startsWith(`..${sep}`) ||
+    isAbsolute(relativePath)
+  ) {
+    return false;
+  }
+
+  const expressions: string[] = [];
+  if (permission === "glob" && typeof context.metadata?.pattern === "string") {
+    expressions.push(context.metadata.pattern);
+  }
+  if (permission === "glob") {
+    expressions.push(...(context.patterns ?? []));
+  }
+  if (typeof context.metadata?.include === "string") {
+    expressions.push(context.metadata.include);
+  }
+  if (
+    expressions.some(
+      (expression) =>
+        isAbsolute(expression) ||
+        containsParentTraversal(expression) ||
+        containsAmbiguousGlobSyntax(expression) ||
+        containsProtectedPath(expression),
+    )
+  ) {
+    return false;
+  }
+
+  if (!relativePath) {
+    if (permission !== "glob") {
+      return false;
+    }
+    const literalPrefix = getLiteralGlobPrefix(expressions[0]);
+    if (!literalPrefix) {
+      return false;
+    }
+    try {
+      searchRoot = realpathSync.native(resolve(root, literalPrefix));
+    } catch {
+      return false;
+    }
+    relativePath = relative(root, searchRoot);
+  }
+
+  return (
+    relativePath !== "" &&
+    relativePath !== ".." &&
+    !relativePath.startsWith(`..${sep}`) &&
+    !isAbsolute(relativePath) &&
+    !containsProtectedPath(relativePath) &&
+    isSafeSearchTarget(searchRoot, relativePath)
+  );
+};
+
+const isSafeSearchTarget = (searchRoot: string, relativePath: string): boolean => {
+  try {
+    const target = lstatSync(searchRoot);
+    if (target.isFile()) {
+      return true;
+    }
+    if (!target.isDirectory()) {
+      return false;
+    }
+    const topLevelName = relativePath.split(sep)[0];
+    if (!topLevelName || !lowRiskSearchRootNames.has(topLevelName)) {
+      return false;
+    }
+
+    const pending = [searchRoot];
+    while (pending.length > 0) {
+      const directory = pending.pop()!;
+      for (const entry of readdirSync(directory, { withFileTypes: true })) {
+        if (entry.isSymbolicLink() || containsProtectedPath(entry.name)) {
+          return false;
+        }
+        if (entry.isDirectory()) {
+          pending.push(resolve(directory, entry.name));
+        }
+      }
+    }
+    return true;
+  } catch {
+    return false;
+  }
+};
+
+const getLiteralGlobPrefix = (expression: string | undefined): string | null => {
+  const firstSegment = expression
+    ?.replaceAll("\\", "/")
+    .replace(/^\.\//, "")
+    .split("/")[0];
+  return firstSegment && !/[*?[\]{}()!+@]/.test(firstSegment)
+    ? firstSegment
+    : null;
+};
+
+const containsParentTraversal = (value: string): boolean =>
+  value.replaceAll("\\", "/").split("/").includes("..");
+
+const containsAmbiguousGlobSyntax = (value: string): boolean =>
+  /[?[\]{}()!+@\\]/.test(value);
+
+const containsProtectedPath = (value: string): boolean => {
+  const normalized = value.replaceAll("\\", "/").toLowerCase();
+  return (
+    normalized.includes(".env") ||
+    /(?:^|[^a-z0-9_-])(?:data|logs)(?:$|[^a-z0-9_-])/.test(normalized)
+  );
+};

src/routes/chatStream.ts

@@ -111,24 +111,40 @@ const toRuntimeModel = (model?: SupportedModel) => {
   };
 };
 
-const emitFallbackMessage = async (
+const emitFinalMessage = async (
   runtime: OpencodeRuntimeAdapter,
   sessionId: string,
   clientSessionId: string,
+  currentAssistantMessageIds: Set<string>,
+  assistantTextParts: Map<string, Map<string, string>>,
   write: (event: string, data: Record<string, unknown>) => void,
 ) => {
-  const messages = await runtime.messages(sessionId);
+  let text = [...currentAssistantMessageIds]
-  const assistantMessage = [...messages]
     .reverse()
-    .find((message) => message.info.role === "assistant");
+    .map((messageId) => [...(assistantTextParts.get(messageId)?.values() ?? [])].join(""))
-  const parts = assistantMessage?.parts ?? [];
+    .find((content) => content.length > 0) ?? "";
-  const text = collectTextContent(parts);
+
+  if (!text) {
+    const messages = await runtime.messages(sessionId);
+    const assistantMessage = [...messages]
+      .reverse()
+      .find(
+        (message) =>
+          message.info.role === "assistant" &&
+          (currentAssistantMessageIds.size === 0 ||
+            currentAssistantMessageIds.has(message.info.id)),
+      );
+    text = collectTextContent(assistantMessage?.parts ?? []);
+  }
+
   if (text) {
     write("token", {
       session_id: clientSessionId,
       content: text,
     });
+    return true;
   }
+  return false;
 };
 
 export const streamPromptResponse = async ({
@@ -156,15 +172,14 @@ export const streamPromptResponse = async ({
   const emittedToolParts = new Set<string>();
   const emittedQuestionToolParts = new Set<string>();
   const emittedQuestionRequestIds = new Set<string>();
+  const currentAssistantMessageIds = new Set<string>();
+  const assistantTextParts = new Map<string, Map<string, string>>();
   const partTypes = new Map<string, Part["type"]>();
-  const pendingPartTextDeltas = new Map<string, string[]>();
+  const pendingTextDeltas = new Map<string, string[]>();
-  const reasoningDeltas = new Map<string, string[]>();
   const reasoningStatuses = new Map<string, "running" | "completed">();
   const toolStatuses = new Map<string, string>();
   let firstSessionEventLogged = false;
   let firstNonStatusEventLogged = false;
-  let firstTokenLogged = false;
-  let firstReasoningLogged = false;
   let firstToolEventLogged = false;
   let lastSessionStatus: string | null = null;
   let lastSessionStatusMessage: string | null = null;
@@ -379,6 +394,11 @@ export const streamPromptResponse = async ({
         const permissionApproval = resolvePermissionApproval(
           approvalMode,
           event.properties.permission,
+          {
+            metadata: event.properties.metadata,
+            patterns: event.properties.patterns,
+            workspaceRoot: process.cwd(),
+          },
         );
         logDevelopmentDebug("permission request received", {
           ...debugContext,
@@ -425,6 +445,11 @@ export const streamPromptResponse = async ({
         const permissionApproval = resolvePermissionApproval(
           approvalMode,
           event.properties.action,
+          {
+            metadata: event.properties.metadata,
+            patterns: event.properties.resources,
+            workspaceRoot: process.cwd(),
+          },
         );
         logDevelopmentDebug("permission v2 request received", {
           ...debugContext,
@@ -626,45 +651,26 @@ export const streamPromptResponse = async ({
       if (event.type === "message.updated") {
         if (event.properties.info.role === "assistant") {
           sawResponseActivity = true;
+          currentAssistantMessageIds.add(event.properties.info.id);
         }
         continue;
       }
 
       if (event.type === "message.part.delta" && event.properties.field === "text") {
         sawResponseActivity = true;
+        currentAssistantMessageIds.add(event.properties.messageID);
         const partType = partTypes.get(event.properties.partID);
         if (partType === "text") {
-          if (!firstTokenLogged) {
+          const messageParts = assistantTextParts.get(event.properties.messageID) ?? new Map();
-            firstTokenLogged = true;
+          messageParts.set(
-            logDevelopmentDebug("first response token emitted", {
+            event.properties.partID,
-              ...debugContext,
+            `${messageParts.get(event.properties.partID) ?? ""}${event.properties.delta}`,
-              partId: event.properties.partID,
+          );
-              elapsedMs: Math.max(0, Date.now() - requestStartedAt),
+          assistantTextParts.set(event.properties.messageID, messageParts);
-              sincePromptDispatchMs: Math.max(0, Date.now() - promptStartedAt),
-            });
-          }
-          emittedText = true;
-          write("token", {
-            session_id: clientSessionId,
-            content: event.properties.delta,
-          });
-        } else if (partType === "reasoning") {
-          if (!firstReasoningLogged) {
-            firstReasoningLogged = true;
-            logDevelopmentDebug("first reasoning delta received", {
-              ...debugContext,
-              partId: event.properties.partID,
-              elapsedMs: Math.max(0, Date.now() - requestStartedAt),
-              sincePromptDispatchMs: Math.max(0, Date.now() - promptStartedAt),
-            });
-          }
-          const pending = reasoningDeltas.get(event.properties.partID) ?? [];
-          pending.push(event.properties.delta);
-          reasoningDeltas.set(event.properties.partID, pending);
         } else if (!partType) {
-          const pending = pendingPartTextDeltas.get(event.properties.partID) ?? [];
+          const pending = pendingTextDeltas.get(event.properties.partID) ?? [];
           pending.push(event.properties.delta);
-          pendingPartTextDeltas.set(event.properties.partID, pending);
+          pendingTextDeltas.set(event.properties.partID, pending);
         }
         continue;
       }
@@ -673,23 +679,19 @@ export const streamPromptResponse = async ({
         sawResponseActivity = true;
         const part = event.properties.part;
         partTypes.set(part.id, part.type);
+        if (part.type === "text" || part.type === "reasoning" || part.type === "tool") {
+          currentAssistantMessageIds.add(part.messageID);
+        }
         if (part.type === "text") {
-          const pending = pendingPartTextDeltas.get(part.id) ?? [];
+          const pendingText = (pendingTextDeltas.get(part.id) ?? []).join("");
-          pendingPartTextDeltas.delete(part.id);
+          pendingTextDeltas.delete(part.id);
-          for (const content of pending) {
+          const messageParts = assistantTextParts.get(part.messageID) ?? new Map();
-            emittedText = true;
+          messageParts.set(part.id, part.text || pendingText);
-            write("token", {
+          assistantTextParts.set(part.messageID, messageParts);
-              session_id: clientSessionId,
+        } else {
-              content,
+          pendingTextDeltas.delete(part.id);
-            });
+        }
-          }
+        if (part.type === "reasoning") {
-        } else if (part.type === "reasoning") {
-          const pending = pendingPartTextDeltas.get(part.id) ?? [];
-          if (pending.length > 0) {
-            const existing = reasoningDeltas.get(part.id) ?? [];
-            reasoningDeltas.set(part.id, existing.concat(pending));
-          }
-          pendingPartTextDeltas.delete(part.id);
           const reasoningStatus = part.time.end ? "completed" : "running";
           if (reasoningStatuses.get(part.id) !== reasoningStatus) {
             reasoningStatuses.set(part.id, reasoningStatus);
@@ -697,14 +699,10 @@ export const streamPromptResponse = async ({
               ...debugContext,
               partId: part.id,
               status: reasoningStatus,
-              chunkCount: (reasoningDeltas.get(part.id) ?? []).length,
               elapsedMs: Math.max(0, Date.now() - requestStartedAt),
             });
           }
-          const reasoningDetail = buildReasoningProgressDetail(
+          const reasoningDetail = buildReasoningProgressDetail(part.time.end);
-            reasoningDeltas.get(part.id) ?? [],
-            part.time.end,
-          );
           emitProgress({
             id: part.id,
             phase: "planning",
@@ -932,13 +930,14 @@ export const streamPromptResponse = async ({
     }
 
     await promptPromise;
-    if (!emittedText) {
+    emittedText = await emitFinalMessage(
-      logDevelopmentDebug("no streamed text emitted, falling back to messages()", {
+      runtime,
-        ...debugContext,
+      sessionId,
-        elapsedMs: Math.max(0, Date.now() - requestStartedAt),
+      clientSessionId,
-      });
+      currentAssistantMessageIds,
-      await emitFallbackMessage(runtime, sessionId, clientSessionId, write);
+      assistantTextParts,
-    }
+      write,
+    );
     emitProgress({
       id: "request-received",
       phase: "start",

src/routes/chatStreamEvents.ts

@@ -376,12 +376,6 @@ const formatProgressValue = (value: unknown): string => {
   }
 };
 
-const normalizeProgressText = (chunks: string[]) =>
-  chunks.join("").replace(/\s+/g, " ").trim();
-
-const truncateProgressText = (text: string, maxLength: number) =>
-  text.length > maxLength ? `${text.slice(0, maxLength - 3)}...` : text;
-
 const summarizeToolParams = (params: Record<string, unknown>) => {
   const ignoredKeys = new Set(["reason", "request_reason", "why", "purpose", "rationale"]);
   const summary = Object.entries(params)
@@ -413,19 +407,8 @@ export const buildSessionStatusDetail = (status: { type: string; message?: strin
 };
 
 export const buildReasoningProgressDetail = (
-  chunks: string[],
   ended?: string | number | Date | null,
-) => {
+) => ended ? "分析步骤已整理完成。" : "Agent 正在分析问题。";
-  const reasoningText = truncateProgressText(normalizeProgressText(chunks), 800);
-  if (ended) {
-    return reasoningText
-      ? `推理过程：${reasoningText}`
-      : "当前推理阶段已完成，Agent 将继续输出答案或进入工具执行。";
-  }
-  return reasoningText
-    ? `正在推理：${reasoningText}`
-    : "Agent 正在拆解问题、梳理执行步骤并判断是否需要调用工具。";
-};
 
 export const buildToolProgressDetail = (
   tool: string,

src/runtime/opencode.ts

@@ -7,6 +7,10 @@ import { resolve } from "node:path";
 
 import { config } from "../config.js";
 import { logger } from "../logger.js";
+import {
+  cleanupExpiredToolOutputs,
+  resolveOpencodeToolOutputDirectory,
+} from "./opencodeToolOutputCleanup.js";
 
 const isDevelopmentDebugLoggingEnabled = process.env.NODE_ENV === "development";
 
@@ -46,6 +50,7 @@ const getRuntimeMessageId = (message: RuntimeMessage) => message.info.id;
 export class OpencodeRuntimeAdapter {
   private clientPromise: Promise<OpencodeClient> | null = null;
   private closeServer: (() => void) | null = null;
+  private toolOutputCleanupTimer: ReturnType<typeof setInterval> | null = null;
 
   async ensureClient(): Promise<OpencodeClient> {
     if (!this.clientPromise) {
@@ -380,12 +385,18 @@ export class OpencodeRuntimeAdapter {
   }
 
   async dispose(): Promise<void> {
+    if (this.toolOutputCleanupTimer) {
+      clearInterval(this.toolOutputCleanupTimer);
+      this.toolOutputCleanupTimer = null;
+    }
     this.closeServer?.();
     this.closeServer = null;
     this.clientPromise = null;
   }
 
   private async bootstrapClient(): Promise<OpencodeClient> {
+    await this.cleanupToolOutputs();
+
     // embedded 模式下，把服务内工具桥地址注入到 opencode 进程环境里，
     // 这样 .opencode/tools 下的自定义工具可以回调本服务。
     process.env.TJWATER_AGENT_INTERNAL_BASE_URL = `http://127.0.0.1:${config.PORT}`;
@@ -436,9 +447,39 @@ export class OpencodeRuntimeAdapter {
     this.closeServer = () => {
       runtime.server.close();
     };
+    this.startToolOutputCleanupLoop();
 
     return runtime.client;
   }
+
+  private async cleanupToolOutputs(): Promise<void> {
+    const directory = resolveOpencodeToolOutputDirectory();
+    const ttlMs = config.RESULT_REF_TTL_HOURS * 60 * 60 * 1000;
+    try {
+      const result = await cleanupExpiredToolOutputs(directory, ttlMs);
+      if (result.removed > 0) {
+        logger.info(
+          { directory, ...result },
+          "removed expired opencode tool output files",
+        );
+      }
+    } catch (error) {
+      logger.warn(
+        { err: error, directory },
+        "failed to clean expired opencode tool output files",
+      );
+    }
+  }
+
+  private startToolOutputCleanupLoop(): void {
+    if (this.toolOutputCleanupTimer) {
+      return;
+    }
+    this.toolOutputCleanupTimer = setInterval(() => {
+      void this.cleanupToolOutputs();
+    }, config.RESULT_REF_CLEANUP_INTERVAL_MS);
+    this.toolOutputCleanupTimer.unref();
+  }
 }
 
 export const opencodeRuntime = new OpencodeRuntimeAdapter();
@@ -448,6 +489,10 @@ function buildOpencodeConfig(): Record<string, unknown> {
     deepMerge(readProjectOpencodeConfig(), readEnvOpencodeConfig()),
     {
       model: config.OPENCODE_MODEL,
+      tool_output: {
+        max_bytes: config.MAX_INLINE_RESULT_BYTES,
+        max_lines: 2000,
+      },
     },
   );
 }

src/runtime/opencodeToolOutputCleanup.ts

@@ -0,0 +1,59 @@
+import { readdir, rm, stat } from "node:fs/promises";
+import { homedir } from "node:os";
+import { join } from "node:path";
+
+export type ToolOutputCleanupResult = {
+  removed: number;
+  scanned: number;
+};
+
+export const resolveOpencodeToolOutputDirectory = (): string => {
+  const dataRoot = process.env.XDG_DATA_HOME?.trim() || join(homedir(), ".local", "share");
+  return join(dataRoot, "opencode", "tool-output");
+};
+
+export const cleanupExpiredToolOutputs = async (
+  directory: string,
+  ttlMs: number,
+  now = Date.now(),
+): Promise<ToolOutputCleanupResult> => {
+  if (!Number.isFinite(ttlMs) || ttlMs <= 0) {
+    throw new Error("tool output cleanup ttlMs must be a positive number");
+  }
+
+  let entries;
+  try {
+    entries = await readdir(directory, { withFileTypes: true });
+  } catch (error) {
+    if (isNodeError(error, "ENOENT")) {
+      return { removed: 0, scanned: 0 };
+    }
+    throw error;
+  }
+
+  let removed = 0;
+  for (const entry of entries) {
+    if (!entry.isFile() || !entry.name.startsWith("tool_")) {
+      continue;
+    }
+
+    const path = join(directory, entry.name);
+    try {
+      const file = await stat(path);
+      if (now - file.mtimeMs <= ttlMs) {
+        continue;
+      }
+      await rm(path);
+      removed += 1;
+    } catch (error) {
+      if (!isNodeError(error, "ENOENT")) {
+        throw error;
+      }
+    }
+  }
+
+  return { removed, scanned: entries.length };
+};
+
+const isNodeError = (error: unknown, code: string): error is NodeJS.ErrnoException =>
+  error instanceof Error && "code" in error && error.code === code;

src/server.ts

@@ -1,5 +1,4 @@
 import { randomUUID } from "node:crypto";
-import { spawn } from "node:child_process";
 import cors from "cors";
 import express from "express";
 
@@ -11,6 +10,7 @@ import {
   runWithCredentialRefresh,
 } from "./auth/credentialRefresh.js";
 import { SessionTranscriptStore } from "./sessions/transcriptStore.js";
+import { executeCliCommand } from "./cli/executeCliCommand.js";
 import { ChatSessionBridge } from "./chat/sessionBridge.js";
 import { config } from "./config.js";
 import { SessionUiStateStore } from "./sessions/uiStateStore.js";
@@ -229,7 +229,13 @@ app.post("/internal/tools/tjwater-cli-call", async (req, res) => {
     result = await runWithCredentialRefresh(
       credentialRefreshCoordinator,
       context,
-      (activeContext) => executeCliCommand(activeContext, command, timeoutSec),
+      (activeContext) =>
+        executeCliCommand(activeContext, command, timeoutSec, {
+          apiBaseUrl: config.TJWATER_API_BASE_URL,
+          cliPath: config.TJWATER_CLI_PATH,
+          maxStderrBytes: config.MAX_CLI_STDERR_BYTES,
+          maxStdoutBytes: config.MAX_CLI_OUTPUT_BYTES,
+        }),
     );
   } catch (error) {
     if (!(error instanceof CredentialRefreshError)) {
@@ -266,6 +272,20 @@ app.post("/internal/tools/tjwater-cli-call", async (req, res) => {
     return;
   }
 
+  if (result.outcome === "output_limit") {
+    res.status(502).json({
+      ok: false,
+      schema_version: "tjwater-cli/v1",
+      summary: "CLI 输出超过安全限制",
+      error: {
+        code: "OUTPUT_LIMIT_EXCEEDED",
+        message: `${result.exceededStream ?? "output"} exceeded ${config.MAX_CLI_OUTPUT_BYTES} bytes`,
+        retryable: false,
+      },
+    });
+    return;
+  }
+
   if (result.status === 401) {
     markAuthExpired(
       getRuntimeSessionContext(sessionId) ?? context,
@@ -288,80 +308,19 @@ app.post("/internal/tools/tjwater-cli-call", async (req, res) => {
     return;
   }
 
-  try {
+  if (result.stdout.trim()) {
-    res.json(JSON.parse(result.stdout));
+    res.status(200).type("application/json").send(result.stdout);
-  } catch {
+    return;
-    res.json({
-      ok: true,
-      schema_version: "tjwater-cli/v1",
-      raw: result.stdout,
-      stderr: result.stderr || undefined,
-    });
   }
+  res.json({
+    ok: true,
+    schema_version: "tjwater-cli/v1",
+    raw: "",
+    stderr: result.stderr || undefined,
+    stderr_truncated: result.stderrTruncated || undefined,
+  });
 });
 
-const executeCliCommand = async (
-  context: RuntimeSessionContext,
-  command: string,
-  timeoutSec: number,
-) => {
-  const child = spawn(
-    config.TJWATER_CLI_PATH,
-    ["--auth-stdin", ...command.split(/\s+/).filter(Boolean)],
-    { stdio: ["pipe", "pipe", "pipe"] },
-  );
-  let stdout = "";
-  let stderr = "";
-  child.stdout.on("data", (data: Buffer) => {
-    stdout += data.toString("utf-8");
-  });
-  child.stderr.on("data", (data: Buffer) => {
-    stderr += data.toString("utf-8");
-  });
-  child.stdin.write(
-    JSON.stringify({
-      server: config.TJWATER_API_BASE_URL,
-      access_token: context.accessToken,
-      project_id: context.projectId,
-    }),
-  );
-  child.stdin.end();
-
-  const exitCode = await new Promise<number | null>((resolve, reject) => {
-    const timer = setTimeout(() => {
-      child.kill("SIGTERM");
-      resolve(-1);
-    }, timeoutSec * 1000);
-    child.on("close", (code) => {
-      clearTimeout(timer);
-      resolve(code);
-    });
-    child.on("error", (error) => {
-      clearTimeout(timer);
-      reject(error);
-    });
-  });
-  let errorCode = "";
-  try {
-    const payload = JSON.parse(stdout) as { error?: { code?: unknown } };
-    errorCode =
-      typeof payload.error?.code === "string" ? payload.error.code : "";
-  } catch {
-    errorCode = "";
-  }
-  const status =
-    exitCode === -1
-      ? 504
-      : errorCode === "HTTP_401" || errorCode === "UNAUTHENTICATED"
-        ? 401
-        : errorCode === "HTTP_403"
-          ? 403
-          : exitCode === 0
-            ? 200
-            : 502;
-  return { exitCode, status, stderr, stdout };
-};
-
 app.post("/internal/tools/store-render-ref", async (req, res) => {
   if (req.header("x-agent-internal-token") !== internalToken) {
     res.status(403).json({ message: "forbidden" });

tests/cli/executeCliCommand.test.ts

@@ -0,0 +1,137 @@
+import { describe, expect, test } from "bun:test";
+import { fileURLToPath } from "node:url";
+
+import { executeCliCommand } from "../../src/cli/executeCliCommand.js";
+import { type RuntimeSessionContext } from "../../src/runtime/sessionContext.js";
+
+const cliPath = fileURLToPath(
+  new URL("../fixtures/fakeCli.mjs", import.meta.url),
+);
+
+const context: RuntimeSessionContext = {
+  accessToken: "test-token",
+  actorKey: "actor-1",
+  clientSessionId: "client-1",
+  projectId: "project-1",
+  projectKey: "project-1",
+  sessionId: "session-1",
+  traceId: "trace-1",
+};
+
+const run = (
+  command: string,
+  options: {
+    maxStderrBytes?: number;
+    maxStdoutBytes?: number;
+    terminationGraceMs?: number;
+    timeoutSec?: number;
+  } = {},
+) =>
+  executeCliCommand(context, command, options.timeoutSec ?? 1, {
+    apiBaseUrl: "http://127.0.0.1:8000",
+    cliPath,
+    maxStderrBytes: options.maxStderrBytes ?? 8,
+    maxStdoutBytes: options.maxStdoutBytes ?? 64,
+    terminationGraceMs: options.terminationGraceMs ?? 20,
+  });
+
+describe("executeCliCommand", () => {
+  test("accepts output at the byte limit", async () => {
+    await expect(run("stdout 123456", { maxStdoutBytes: 6 })).resolves.toMatchObject({
+      outcome: "completed",
+      exitCode: 0,
+      status: 200,
+      stdout: "123456",
+    });
+  });
+
+  test("does not treat the OpenCode 12000-byte inline threshold as a CLI limit", async () => {
+    const stdout = "x".repeat(12_001);
+    const result = await run(`stdout ${stdout}`, { maxStdoutBytes: 128 * 1024 * 1024 });
+
+    expect(result.outcome).toBe("completed");
+    expect(result.stdout).toBe(stdout);
+  });
+
+  test("rejects multibyte output above the byte limit without returning a partial body", async () => {
+    await expect(run("stdout 水水", { maxStdoutBytes: 5 })).resolves.toMatchObject({
+      outcome: "output_limit",
+      exceededStream: "stdout",
+      status: 502,
+      stderr: "",
+      stdout: "",
+    });
+  });
+
+  test("truncates stderr independently without terminating a successful command", async () => {
+    await expect(
+      run("stderr-success 123456789", { maxStderrBytes: 6 }),
+    ).resolves.toMatchObject({
+      outcome: "completed",
+      exitCode: 0,
+      status: 200,
+      stderr: "123456",
+      stderrTruncated: true,
+      stdout: '{"ok":true}',
+    });
+  });
+
+  test("waits for a SIGTERM-aware process to close after timeout", async () => {
+    const startedAt = Date.now();
+    const result = await run("term", {
+      terminationGraceMs: 100,
+      timeoutSec: 0.25,
+    });
+
+    expect(result).toMatchObject({ outcome: "timeout", status: 504 });
+    expect(Date.now() - startedAt).toBeGreaterThanOrEqual(270);
+  });
+
+  test("uses SIGKILL when a timed-out process ignores SIGTERM", async () => {
+    const result = await run("ignore-term", { timeoutSec: 0.25 });
+
+    expect(result).toMatchObject({
+      outcome: "timeout",
+      signal: "SIGKILL",
+      status: 504,
+    });
+  });
+
+  test("keeps the timeout outcome when closing stdin also errors", async () => {
+    const largeContext = {
+      ...context,
+      accessToken: "x".repeat(1024 * 1024),
+    };
+
+    await expect(
+      executeCliCommand(largeContext, "ignore-term", 0.25, {
+        apiBaseUrl: "http://127.0.0.1:8000",
+        cliPath,
+        maxStderrBytes: 8,
+        maxStdoutBytes: 64,
+        terminationGraceMs: 20,
+      }),
+    ).resolves.toMatchObject({
+      outcome: "timeout",
+      signal: "SIGKILL",
+      status: 504,
+    });
+  });
+
+  test("rejects a deterministic stdin pipe error without crashing", async () => {
+    const largeContext = {
+      ...context,
+      accessToken: "x".repeat(1024 * 1024),
+    };
+
+    await expect(
+      executeCliCommand(largeContext, "closed-stdin", 1, {
+        apiBaseUrl: "http://127.0.0.1:8000",
+        cliPath,
+        maxStderrBytes: 8,
+        maxStdoutBytes: 64,
+        terminationGraceMs: 20,
+      }),
+    ).rejects.toBeInstanceOf(Error);
+  });
+});

tests/fixtures/fakeCli.mjs

@@ -0,0 +1,39 @@
+#!/usr/bin/env node
+
+import { closeSync } from "node:fs";
+
+const command = process.argv[3];
+const value = process.argv[4] ?? "";
+
+if (command === "stdout") {
+  process.stdout.write(value);
+  process.exit(0);
+}
+
+if (command === "stderr") {
+  process.stderr.write(value);
+  process.exit(1);
+}
+
+if (command === "stderr-success") {
+  process.stderr.write(value);
+  process.stdout.write('{"ok":true}');
+  process.exit(0);
+}
+
+if (command === "term") {
+  process.on("SIGTERM", () => {
+    setTimeout(() => process.exit(0), 30);
+  });
+  setInterval(() => undefined, 1000);
+}
+
+if (command === "ignore-term") {
+  process.on("SIGTERM", () => undefined);
+  setInterval(() => undefined, 1000);
+}
+
+if (command === "closed-stdin") {
+  closeSync(0);
+  setInterval(() => undefined, 1000);
+}

tests/routes/chatPermissionPolicy.test.ts

@@ -1,4 +1,7 @@
 import { describe, expect, it } from "bun:test";
+import { mkdtemp, mkdir, rm, symlink } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { dirname, join } from "node:path";
 
 import {
   canAutoApprovePermission,
@@ -9,10 +12,100 @@ describe("permission approval policy", () => {
   it.each([
     "show_chart",
     "web_search",
+    "skill",
   ])("allows low-risk permission %s", (permission) => {
     expect(canAutoApprovePermission(permission)).toBe(true);
   });
 
+  it("allows structured searches within the workspace", () => {
+    const workspaceRoot = process.cwd();
+    const context = {
+      workspaceRoot,
+      metadata: { path: join(workspaceRoot, "src"), include: "*.ts" },
+      patterns: ["*.ts"],
+    };
+    expect(canAutoApprovePermission("glob", context)).toBe(true);
+    expect(canAutoApprovePermission("grep", context)).toBe(true);
+  });
+
+  it("allows a glob with an explicit safe prefix from the workspace root", () => {
+    expect(
+      canAutoApprovePermission("glob", {
+        workspaceRoot: process.cwd(),
+        metadata: { path: process.cwd(), pattern: "src/**/*.ts" },
+        patterns: ["src/**/*.ts"],
+      }),
+    ).toBe(true);
+  });
+
+  it.each([
+    { metadata: { path: dirname(process.cwd()) }, patterns: ["*"] },
+    { metadata: { path: join(process.cwd(), ".local.env") }, patterns: ["*"] },
+    { metadata: { path: join(process.cwd(), "data") }, patterns: ["*"] },
+    { metadata: { path: join(process.cwd(), "src") }, patterns: ["../logs/**"] },
+    { metadata: { path: process.cwd() }, patterns: ["**/*.env"] },
+    { metadata: { path: process.cwd() }, patterns: ["**/*"] },
+    { metadata: { path: join(process.cwd(), "src") }, patterns: [".[e]nv"] },
+  ])("keeps protected or external searches interactive", (request) => {
+    expect(
+      canAutoApprovePermission("glob", {
+        workspaceRoot: process.cwd(),
+        ...request,
+      }),
+    ).toBe(false);
+  });
+
+  it("keeps grep from the workspace root interactive", () => {
+    expect(
+      canAutoApprovePermission("grep", {
+        workspaceRoot: process.cwd(),
+        metadata: { path: process.cwd(), include: "*.ts" },
+        patterns: ["secret"],
+      }),
+    ).toBe(false);
+  });
+
+  it("rejects a workspace symlink that resolves outside the workspace", async () => {
+    const workspaceRoot = await mkdtemp(join(tmpdir(), "permission-workspace-"));
+    const externalRoot = await mkdtemp(join(tmpdir(), "permission-external-"));
+    try {
+      await mkdir(join(externalRoot, "src"));
+      const linkedPath = join(workspaceRoot, "linked");
+      await symlink(join(externalRoot, "src"), linkedPath, "dir");
+
+      expect(
+        canAutoApprovePermission("grep", {
+          workspaceRoot,
+          metadata: { path: linkedPath, include: "*.ts" },
+          patterns: ["secret"],
+        }),
+      ).toBe(false);
+    } finally {
+      await Promise.all([
+        rm(workspaceRoot, { force: true, recursive: true }),
+        rm(externalRoot, { force: true, recursive: true }),
+      ]);
+    }
+  });
+
+  it("rejects protected descendants below an otherwise safe search root", async () => {
+    const workspaceRoot = await mkdtemp(join(tmpdir(), "permission-descendant-"));
+    try {
+      const sourceRoot = join(workspaceRoot, "src");
+      await mkdir(join(sourceRoot, "data"), { recursive: true });
+
+      expect(
+        canAutoApprovePermission("grep", {
+          workspaceRoot,
+          metadata: { path: sourceRoot, include: "*.ts" },
+          patterns: ["secret"],
+        }),
+      ).toBe(false);
+    } finally {
+      await rm(workspaceRoot, { force: true, recursive: true });
+    }
+  });
+
   it.each([
     "bash",
     "edit",

tests/routes/chatStream.test.ts

@@ -15,6 +15,242 @@ const createEventStream = (events: unknown[]) => ({
 });
 
 describe("streamPromptResponse", () => {
+  it("emits only the final assistant text after tool-driven intermediate messages", async () => {
+    const runtime = {
+      subscribeEvents: async () =>
+        createEventStream([
+          {
+            type: "message.part.delta",
+            properties: {
+              sessionID: "runtime-session-1",
+              messageID: "assistant-intermediate",
+              partID: "text-part-intermediate",
+              field: "text",
+              delta: "正在加载工作流并尝试分页参数。",
+            },
+          },
+          {
+            type: "message.part.updated",
+            properties: {
+              sessionID: "runtime-session-1",
+              part: {
+                id: "text-part-intermediate",
+                sessionID: "runtime-session-1",
+                messageID: "assistant-intermediate",
+                type: "text",
+                text: "正在加载工作流并尝试分页参数。",
+                time: { start: 1, end: 2 },
+              },
+              time: 2,
+            },
+          },
+          {
+            type: "message.part.delta",
+            properties: {
+              sessionID: "runtime-session-1",
+              messageID: "assistant-final",
+              partID: "text-part-final",
+              field: "text",
+              delta: "共识别 56 条瓶颈管段，建议优先改造 Top 5。",
+            },
+          },
+          {
+            type: "message.part.updated",
+            properties: {
+              sessionID: "runtime-session-1",
+              part: {
+                id: "text-part-final",
+                sessionID: "runtime-session-1",
+                messageID: "assistant-final",
+                type: "text",
+                text: "共识别 56 条瓶颈管段，建议优先改造 Top 5。",
+                time: { start: 3, end: 4 },
+              },
+              time: 4,
+            },
+          },
+          {
+            type: "session.idle",
+            properties: { sessionID: "runtime-session-1" },
+          },
+        ]),
+      prompt: async () => undefined,
+      messages: async () => [
+        {
+          info: { id: "assistant-intermediate", role: "assistant" },
+          parts: [
+            {
+              id: "text-part-intermediate",
+              sessionID: "runtime-session-1",
+              messageID: "assistant-intermediate",
+              type: "text",
+              text: "正在加载工作流并尝试分页参数。",
+            },
+          ],
+        },
+        {
+          info: { id: "assistant-final", role: "assistant" },
+          parts: [
+            {
+              id: "text-part-final",
+              sessionID: "runtime-session-1",
+              messageID: "assistant-final",
+              type: "text",
+              text: "共识别 56 条瓶颈管段，建议优先改造 Top 5。",
+            },
+          ],
+        },
+      ],
+    } as unknown as OpencodeRuntimeAdapter;
+    const events: Array<{ event: string; data: Record<string, unknown> }> = [];
+
+    await streamPromptResponse({
+      runtime,
+      sessionId: "runtime-session-1",
+      clientSessionId: "client-session-1",
+      message: "分析管网瓶颈",
+      write: (event, data) => events.push({ event, data }),
+    });
+
+    expect(
+      events
+        .filter((item) => item.event === "token")
+        .map((item) => item.data.content)
+        .join(""),
+    ).toBe("共识别 56 条瓶颈管段，建议优先改造 Top 5。");
+  });
+
+  it("uses the final text event cache when the messages lookup fails", async () => {
+    const runtime = {
+      subscribeEvents: async () =>
+        createEventStream([
+          {
+            type: "message.part.updated",
+            properties: {
+              sessionID: "runtime-session-1",
+              part: {
+                id: "text-part-final",
+                sessionID: "runtime-session-1",
+                messageID: "assistant-final",
+                type: "text",
+                text: "最终分析结果。",
+                time: { start: 1, end: 2 },
+              },
+              time: 2,
+            },
+          },
+          {
+            type: "session.idle",
+            properties: { sessionID: "runtime-session-1" },
+          },
+        ]),
+      prompt: async () => undefined,
+      messages: async () => {
+        throw new Error("transient messages lookup failure");
+      },
+    } as unknown as OpencodeRuntimeAdapter;
+    const events: Array<{ event: string; data: Record<string, unknown> }> = [];
+
+    const result = await streamPromptResponse({
+      runtime,
+      sessionId: "runtime-session-1",
+      clientSessionId: "client-session-1",
+      message: "分析管网瓶颈",
+      write: (event, data) => events.push({ event, data }),
+    });
+
+    expect(result.failed).toBe(false);
+    expect(
+      events
+        .filter((item) => item.event === "token")
+        .map((item) => item.data.content)
+        .join(""),
+    ).toBe("最终分析结果。");
+  });
+
+  it("keeps reasoning generic while preserving tool execution details", async () => {
+    const runtime = {
+      subscribeEvents: async () =>
+        createEventStream([
+          {
+            type: "message.part.updated",
+            properties: {
+              sessionID: "runtime-session-1",
+              part: {
+                id: "reasoning-part-1",
+                sessionID: "runtime-session-1",
+                messageID: "assistant-1",
+                type: "reasoning",
+                text: "内部推理：尝试 limit=5000 并读取临时路径。",
+                time: { start: 1, end: 2 },
+              },
+              time: 2,
+            },
+          },
+          {
+            type: "message.part.delta",
+            properties: {
+              sessionID: "runtime-session-1",
+              messageID: "assistant-1",
+              partID: "reasoning-part-1",
+              field: "text",
+              delta: "内部推理：尝试 limit=5000 并读取临时路径。",
+            },
+          },
+          {
+            type: "message.part.updated",
+            properties: {
+              sessionID: "runtime-session-1",
+              part: {
+                id: "tool-part-1",
+                sessionID: "runtime-session-1",
+                messageID: "assistant-1",
+                type: "tool",
+                callID: "call-1",
+                tool: "tjwater_cli",
+                state: {
+                  status: "error",
+                  input: {
+                    command: "network get-all-pipes-properties --limit 5000",
+                    reason: "尝试突破分页限制",
+                  },
+                  error: "HTTP_422 raw backend payload with trace_id=secret-trace",
+                  time: { start: 1, end: 2 },
+                },
+              },
+              time: 2,
+            },
+          },
+          {
+            type: "session.idle",
+            properties: { sessionID: "runtime-session-1" },
+          },
+        ]),
+      prompt: async () => undefined,
+      messages: async () => [],
+    } as unknown as OpencodeRuntimeAdapter;
+    const events: Array<{ event: string; data: Record<string, unknown> }> = [];
+
+    await streamPromptResponse({
+      runtime,
+      sessionId: "runtime-session-1",
+      clientSessionId: "client-session-1",
+      message: "分析管网瓶颈",
+      write: (event, data) => events.push({ event, data }),
+    });
+
+    const reasoningProgress = events.find(
+      (item) => item.event === "progress" && item.data.id === "reasoning-part-1",
+    );
+    const toolProgress = events.find(
+      (item) => item.event === "progress" && item.data.id === "tool-part-1",
+    );
+    expect(reasoningProgress?.data.detail).toBe("分析步骤已整理完成。");
+    expect(toolProgress?.data.detail).toBe(
+      "tjwater_cli 调用失败；调用原因：尝试突破分页限制；关键参数：command=network get-all-pipes-properties --limit 5000；错误：HTTP_422 raw backend payload with trace_id=secret-trace",
+    );
+  });
+
   it("forwards opencode permission requests as SSE payloads", async () => {
     const runtime = {
       subscribeEvents: async () =>

tests/runtime/opencodeToolOutputCleanup.test.ts

@@ -0,0 +1,62 @@
+import { describe, expect, test } from "bun:test";
+import {
+  lstat,
+  mkdir,
+  mkdtemp,
+  readFile,
+  rm,
+  stat,
+  symlink,
+  utimes,
+  writeFile,
+} from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+
+import { cleanupExpiredToolOutputs } from "../../src/runtime/opencodeToolOutputCleanup.js";
+
+describe("cleanupExpiredToolOutputs", () => {
+  test("removes only expired regular tool output files", async () => {
+    const directory = await mkdtemp(join(tmpdir(), "opencode-tool-output-cleanup-"));
+    const expired = join(directory, "tool_expired");
+    const current = join(directory, "tool_current");
+    const unrelated = join(directory, "keep.txt");
+    const toolDirectory = join(directory, "tool_directory");
+    const toolSymlink = join(directory, "tool_symlink");
+    try {
+      await Promise.all([
+        writeFile(expired, "expired"),
+        writeFile(current, "current"),
+        writeFile(unrelated, "unrelated"),
+        mkdir(toolDirectory),
+      ]);
+      await symlink(unrelated, toolSymlink);
+      const now = Date.now();
+      const old = new Date(now - 8 * 24 * 60 * 60 * 1000);
+      await utimes(expired, old, old);
+
+      const result = await cleanupExpiredToolOutputs(
+        directory,
+        7 * 24 * 60 * 60 * 1000,
+        now,
+      );
+
+      expect(result).toEqual({ removed: 1, scanned: 5 });
+      await expect(stat(expired)).rejects.toMatchObject({ code: "ENOENT" });
+      expect(await readFile(current, "utf8")).toBe("current");
+      expect(await readFile(unrelated, "utf8")).toBe("unrelated");
+      expect((await stat(toolDirectory)).isDirectory()).toBe(true);
+      expect((await lstat(toolSymlink)).isSymbolicLink()).toBe(true);
+    } finally {
+      await rm(directory, { force: true, recursive: true });
+    }
+  });
+
+  test("treats a missing tool output directory as empty", async () => {
+    const directory = join(tmpdir(), `missing-tool-output-${crypto.randomUUID()}`);
+    await expect(cleanupExpiredToolOutputs(directory, 1_000)).resolves.toEqual({
+      removed: 0,
+      scanned: 0,
+    });
+  });
+});