fix: agent 不再继承操作者的 Claude Code 配置; 关闭 agent stdin

docs/reports/e2e-smoke-2026-10-25.md

@@ -77,3 +77,11 @@ The spawned `claude` inherits the user's global configuration. In the rerun, the
 ### Still untested
 
 Tier-2 auto-resolution on a real conflict, three or more conflicting tasks, an agent that fails or times out, `gemini`/`codex`.
+
+## Third run: agents isolated from the operator's Claude Code setup (2026-10-26)
+
+Change: `claude` is started with `--setting-sources project,local` (opt back in with `FSC_CLAUDE_USER_SETTINGS=1`) and with stdin closed.
+
+- Trivial prompt, default vs isolated: 22 s vs 9 s. Default: the agent could recite the operator's tracker rules and the Stop hook; isolated: it knew neither.
+- Same two-task repo as the first run, real `claude` agents: exec phase **20 s** (first run: 77 s). Both tasks merged at tier 1; `main` has the two implementations; agent logs contain no tracker or hook talk.
+- **Correction:** this run took 15 min 28 s wall clock, not 20 s: analysis and execution finished within about a minute (file timestamps: 20 s + 27 s) and the process then idled for ~14 minutes. So the slow runs are not explained by a user-level hook (this run had them isolated). The cause is a different defect, found right after this report and fixed in a follow-up PR: the agent timeout timer was never cleared, so a successful run kept the process alive until the 15-minute timeout fired. Failing runs exited through `process.exit(1)` and never showed it. That also explains why only successful runs were slow.

packages/core/src/analyzerNative.ts

@@ -6,6 +6,7 @@ import { analyzerPrompt } from './prompts/analyzerPrompt';
 import { getCodingStyle } from './codingStyle';
 import { getWorkStyleDescription, WorkStyle } from './workStyle';
 import { trimJSONObjectArray } from './utils/trimJSON';
+import { claudeIsolationArgs } from './execution/types';
 
 const ALLOWED_TOOLS = 'Bash,ReadEdit,Glob,Grep,WebFetch,WebSearch,Write,TodoWrite,SlashCommand';
 const DEFAULT_TIMEOUT_SECONDS = 1200;
@@ -91,9 +92,11 @@ export async function analyzeCodebaseNative(config: Config, repoPath: string): P
         ALLOWED_TOOLS,
         '--permission-mode',
         'bypassPermissions',
+        ...claudeIsolationArgs(),
       ],
       cwd: scratch,
       env: { ...agentEnv(config), IS_SANDBOX: '1' },
+      stdin: 'ignore',
       stdout: 'pipe',
       stderr: 'pipe',
     });

packages/core/src/execution/local.ts

@@ -64,6 +64,8 @@ export class LocalExecution implements ExecutionStrategy {
           FSC_WORKTREE: worktreePath,
           ...(unit.context?.memorixUri ? { MEMORIX_URL: unit.context.memorixUri } : {}),
         },
+        // No stdin: an inherited, silent stdin makes `claude -p` wait ~3 s before it starts.
+        stdin: "ignore",
         stdout: "pipe",
         stderr: "pipe",
       });

packages/core/src/execution/types.ts

@@ -63,9 +63,24 @@ export interface WorktreeInfo {
 }
 
 // Pre-built agent arg builders
+/**
+ * Agents must not inherit the operator's personal Claude Code setup: user-level hooks (a Stop hook
+ * can block or slow a headless run) and the user's global CLAUDE.md both apply by default. Measured
+ * on a trivial prompt: default 22 s and the agent recited the operator's tracker rules; with
+ * `--setting-sources project,local` 9 s and none of it. Project and local settings still apply.
+ * Set FSC_CLAUDE_USER_SETTINGS=1 to opt back in.
+ */
+export function claudeIsolationArgs(): string[] {
+  return process.env.FSC_CLAUDE_USER_SETTINGS === "1" ? [] : ["--setting-sources", "project,local"];
+}
+
 export const AGENT_BUILDERS: Record<string, AgentConfig["buildArgs"]> = {
-  claude: (prompt: string, cwd: string) =>
-    ["-p", prompt, "--allowedTools", "Edit,Write,Bash,Read,Glob,Grep", "--permission-mode", "bypassPermissions"],
+  claude: (prompt: string, _cwd: string) => [
+    "-p", prompt,
+    "--allowedTools", "Edit,Write,Bash,Read,Glob,Grep",
+    "--permission-mode", "bypassPermissions",
+    ...claudeIsolationArgs(),
+  ],
   codex: (prompt: string, cwd: string) =>
     ["exec", "--approval-mode", "full-auto", "-w", cwd, prompt],
   gemini: (prompt: string, cwd: string) =>

packages/core/test/execution/mapping.test.ts

@@ -83,3 +83,33 @@ describe("normalizeAgentType", () => {
     expect(args).toContain("--permission-mode");
   });
 });
+
+describe("claude isolation", () => {
+  const { toWorkUnitForAgent } = require("../../src/execution/mapping");
+  const task = { ID: "t", title: "x", description: "y", priority: 1 };
+
+  test("agents ignore the operator's user-level settings by default", () => {
+    const prev = process.env.FSC_CLAUDE_USER_SETTINGS;
+    delete process.env.FSC_CLAUDE_USER_SETTINGS;
+    try {
+      const args = toWorkUnitForAgent(task, "claude-code", ".").agentConfig.buildArgs("p", "/wt");
+      const i = args.indexOf("--setting-sources");
+      expect(i).toBeGreaterThan(-1);
+      expect(args[i + 1]).toBe("project,local");
+    } finally {
+      if (prev !== undefined) process.env.FSC_CLAUDE_USER_SETTINGS = prev;
+    }
+  });
+
+  test("FSC_CLAUDE_USER_SETTINGS=1 opts back in", () => {
+    const prev = process.env.FSC_CLAUDE_USER_SETTINGS;
+    process.env.FSC_CLAUDE_USER_SETTINGS = "1";
+    try {
+      const args = toWorkUnitForAgent(task, "claude-code", ".").agentConfig.buildArgs("p", "/wt");
+      expect(args).not.toContain("--setting-sources");
+    } finally {
+      if (prev === undefined) delete process.env.FSC_CLAUDE_USER_SETTINGS;
+      else process.env.FSC_CLAUDE_USER_SETTINGS = prev;
+    }
+  });
+});