fix: restore bounded agent sandbox execution

.gitea/workflows/hub-check.yml

@@ -93,8 +93,8 @@ jobs:
 
       - name: Prove real Claude SDK Bash sandbox boundary
         run: |
-          sudo install -d -o "$(id -u)" -g "$(id -g)" -m 0700 /var/lib/cph-test
+          sudo install -d -o "$(id -u)" -g "$(id -g)" -m 0700 /w/t
-          CPH_SANDBOX_TEST_ROOT=/var/lib/cph-test \
+          CPH_SANDBOX_TEST_ROOT=/w/t \
             /usr/bin/setpriv --no-new-privs \
             npx vitest run test/integration/agent-sandbox-linux.test.ts
 

docs/adr/0018-agent-execution-surface-bounded-by-workspace.md

@@ -93,23 +93,26 @@ The boundary is enforced by the Claude Code SDK's built-in sandbox
   its Bash subprocesses run sandboxed; if the sandbox can't start, `query()`
   emits an error and exits rather than running unsandboxed.
 - `sandbox.allowUnsandboxedCommands: false` — a tool cannot opt out with the
-  SDK's `dangerouslyDisableSandbox` input.
+  SDK's `dangerouslyDisableSandbox` input. Claude Code 2.1.202 does not enforce
+  that option reliably, so a host-side `PreToolUse` hook also denies every
+  Bash request whose input explicitly sets `dangerouslyDisableSandbox: true`
+  before a process can start.
 - `sandbox.filesystem.denyRead: ["/"]` with `allowRead` for the canonical
   current workspace and a small named system-runtime set — normal reads stay
   in the run's workspace while `/bin`, shared libraries, CA certificates,
   fonts and the configured `cph` executable remain available as the external
   tool exception described above.
-- `sandbox.filesystem.allowWrite` confines deliverable writes to the canonical
+- `sandbox.filesystem.allowWrite: [workspaceDir]` confines persistent host
-  ADR-0007 workspace and opens one service-owned SDK scratch directory under
+  effects to the canonical ADR-0007 workspace. Config/cache/home stay beneath
-  `/tmp/cph-agent-<uid>/<project-hash>`. Config/cache/home stay beneath
   `.cph/agent-runtime/`; `TMPDIR`, `TMP`, `TEMP`, and `CLAUDE_CODE_TMPDIR` all
-  point at the project-keyed scratch directory. Its canonical prefix is capped
+  point at the workspace-local `.cph/t`. The workspace allocator uses stable
-  at 64 bytes so the SDK can append randomized `socat` bridge socket names
+  compact Organization/Project path segments, deployment requires a short
-  without exceeding Linux `sockaddr_un.sun_path`. The per-Organization OS UID,
+  workspace root, and the canonical temp prefix fails fast above 56 bytes so
-  project hash, mode `0700`, symlink rejection, and exact sandbox allowlist keep
+  the SDK can append randomized `socat` bridge socket names without exceeding
-  that scratch isolated; sibling scratch paths and ordinary `/tmp` remain
+  Linux `sockaddr_un.sun_path`. Bubblewrap shadows non-allowlisted host trees
-  denied. Root is denied for writes and only the workspace plus that exact
+  with disposable tmpfs mounts: a shell write there may succeed inside that
-  internal scratch directory are re-opened.
+  private namespace, but it cannot mutate the corresponding host path. The
+  Linux proof checks host state after the sandbox exits.
 - The SDK subprocess environment replaces rather than spreads `process.env`.
   Only provider protocol variables and non-secret runtime variables cross the
   boundary; database, Feishu and Hub session credentials never enter it.

hub/deploy/README.md

@@ -15,6 +15,7 @@ Application code lives in immutable versioned directories under
 sudo BASE=/srv/curriculum-project-hub \
   HUB_DIR=/srv/curriculum-project-hub/releases/<release-id>/hub \
   INSTANCE_ID=org-a \
+  WORKSPACE_ROOT=/w/997 \
   PORT=8788 \
   MEMORY_MAX=16G CPU_QUOTA=400% TASKS_MAX=512 \
   bash /srv/curriculum-project-hub/releases/<release-id>/hub/deploy/install_service.sh
@@ -49,7 +50,7 @@ Default state paths are:
 ```text
 /var/lib/cph-hub/org-a/home
 /var/lib/cph-hub/org-a/state
-/var/lib/cph-hub/org-a/workspaces
+/w/997
 /var/cache/cph-hub/org-a
 ```
 

hub/deploy/deploy_platform.sh

@@ -10,6 +10,7 @@
 #   PLATFORM_DEPLOY_BASE     optional, defaults to /srv/curriculum-project-hub
 #   PLATFORM_DEPLOY_RELEASE  optional immutable release id, defaults to git HEAD
 #   PLATFORM_DEPLOY_INSTANCE required, Silo instance id
+#   PLATFORM_DEPLOY_WORKSPACE_ROOT required short per-Silo path (for example /w/997)
 #   PLATFORM_DEPLOY_MEMORY_MAX / CPU_QUOTA / TASKS_MAX required ceilings
 #   PLATFORM_DEPLOY_HEALTH_URL optional, defaults to http://127.0.0.1:8788/api/healthz
 # The target host must have Node.js 24+, npm, rsync, PostgreSQL client tools
@@ -31,6 +32,7 @@ REPO_ROOT="$(cd "$SCRIPT_DIR/../.." && pwd)"
 RELEASE_ID="${PLATFORM_DEPLOY_RELEASE:-$(git -C "$REPO_ROOT" rev-parse --verify HEAD)}"
 [[ "$RELEASE_ID" =~ ^[A-Za-z0-9._-]+$ ]] || { echo "invalid PLATFORM_DEPLOY_RELEASE" >&2; exit 1; }
 INSTANCE_ID="${PLATFORM_DEPLOY_INSTANCE:?PLATFORM_DEPLOY_INSTANCE required}"
+WORKSPACE_ROOT="${PLATFORM_DEPLOY_WORKSPACE_ROOT:?PLATFORM_DEPLOY_WORKSPACE_ROOT required}"
 SERVICE_UNIT="cph-hub-$INSTANCE_ID.service"
 MEMORY_MAX="${PLATFORM_DEPLOY_MEMORY_MAX:?PLATFORM_DEPLOY_MEMORY_MAX required}"
 CPU_QUOTA="${PLATFORM_DEPLOY_CPU_QUOTA:?PLATFORM_DEPLOY_CPU_QUOTA required}"
@@ -67,11 +69,11 @@ fi
 ssh "${SSH_OPTS[@]}" "$DEPLOY_USER@$HOST" "
   set -euo pipefail
   if [ \"\$(id -u)\" = \"0\" ]; then
-    BASE='$BASE' HUB_DIR='$HUB_DIR' INSTANCE_ID='$INSTANCE_ID' PORT='$HUB_PORT' MEMORY_MAX='$MEMORY_MAX' CPU_QUOTA='$CPU_QUOTA' TASKS_MAX='$TASKS_MAX' bash '$HUB_DIR/deploy/install_service.sh'
+    BASE='$BASE' HUB_DIR='$HUB_DIR' INSTANCE_ID='$INSTANCE_ID' WORKSPACE_ROOT='$WORKSPACE_ROOT' PORT='$HUB_PORT' MEMORY_MAX='$MEMORY_MAX' CPU_QUOTA='$CPU_QUOTA' TASKS_MAX='$TASKS_MAX' bash '$HUB_DIR/deploy/install_service.sh'
     systemctl restart '$SERVICE_UNIT'
     systemctl is-active --quiet '$SERVICE_UNIT'
   else
-    sudo -n BASE='$BASE' HUB_DIR='$HUB_DIR' INSTANCE_ID='$INSTANCE_ID' PORT='$HUB_PORT' MEMORY_MAX='$MEMORY_MAX' CPU_QUOTA='$CPU_QUOTA' TASKS_MAX='$TASKS_MAX' bash '$HUB_DIR/deploy/install_service.sh'
+    sudo -n BASE='$BASE' HUB_DIR='$HUB_DIR' INSTANCE_ID='$INSTANCE_ID' WORKSPACE_ROOT='$WORKSPACE_ROOT' PORT='$HUB_PORT' MEMORY_MAX='$MEMORY_MAX' CPU_QUOTA='$CPU_QUOTA' TASKS_MAX='$TASKS_MAX' bash '$HUB_DIR/deploy/install_service.sh'
     sudo -n systemctl restart '$SERVICE_UNIT'
     sudo -n systemctl is-active --quiet '$SERVICE_UNIT'
   fi

hub/deploy/install_service.sh

@@ -23,7 +23,7 @@ SERVICE_GROUP="${SERVICE_GROUP:-$SERVICE_USER}"
 SERVICE_HOME="${SERVICE_HOME:-/var/lib/cph-hub/$INSTANCE_ID/home}"
 STATE_DIR="${STATE_DIR:-/var/lib/cph-hub/$INSTANCE_ID/state}"
 CACHE_DIR="${CACHE_DIR:-/var/cache/cph-hub/$INSTANCE_ID}"
-WORKSPACE_ROOT="${WORKSPACE_ROOT:-/var/lib/cph-hub/$INSTANCE_ID/workspaces}"
+WORKSPACE_ROOT="${WORKSPACE_ROOT:?WORKSPACE_ROOT required (use a short per-Silo path such as /w/997)}"
 HOST="${HOST:-127.0.0.1}"
 PORT="${PORT:?PORT is required and must be unique on the host}"
 ENV_FILE="${ENV_FILE:-$BASE/.secrets/$INSTANCE_ID/platform.env}"
@@ -120,6 +120,7 @@ done
 [[ "$MEMORY_MAX" =~ ^[1-9][0-9]*[KMGT]$ ]] || fail "MEMORY_MAX must be a systemd byte size such as 16G"
 [[ "$CPU_QUOTA" =~ ^[1-9][0-9]*%$ ]] || fail "CPU_QUOTA must be a positive percentage such as 400%"
 [[ "$TASKS_MAX" =~ ^[1-9][0-9]*$ ]] || fail "TASKS_MAX must be a positive integer"
+[ "${#WORKSPACE_ROOT}" -le 16 ] || fail "WORKSPACE_ROOT must be at most 16 bytes for Agent sandbox sockets: $WORKSPACE_ROOT"
 
 KEYRING_CREATED=false
 if [ -L "$KEYRING_FILE" ]; then

hub/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "@paradigm/hub",
-  "version": "0.0.5",
+  "version": "0.0.6",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "@paradigm/hub",
-      "version": "0.0.5",
+      "version": "0.0.6",
       "dependencies": {
         "@anthropic-ai/claude-agent-sdk": "^0.3.202",
         "@fastify/cookie": "^11.0.2",

hub/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@paradigm/hub",
-  "version": "0.0.5",
+  "version": "0.0.6",
   "private": true,
   "type": "module",
   "engines": {

hub/src/agent/runner.ts

@@ -29,7 +29,7 @@
  * `workspace.ts` `confine()` path validator as a tool wrapper — the OS sandbox
  * is the mechanism, the contract pins the invariant.
  */
-import { query, type McpServerConfig, type SDKMessage, type SDKAssistantMessage, type SDKUserMessage, type SDKResultMessage, type SDKPartialAssistantMessage } from "@anthropic-ai/claude-agent-sdk";
+import { query, type HookCallback, type McpServerConfig, type SDKMessage, type SDKAssistantMessage, type SDKUserMessage, type SDKResultMessage, type SDKPartialAssistantMessage } from "@anthropic-ai/claude-agent-sdk";
 import type { PrismaClient } from "@prisma/client";
 import { claudeSdkToolConfigForRole } from "./roleTools.js";
 import { createAgentSecurityPolicy } from "./security.js";
@@ -96,6 +96,22 @@ export interface RunResult {
 
 const DEFAULT_MAX_TURNS = 25;
 
+const denyUnsandboxedBash: HookCallback = async (input) => {
+  if (input.hook_event_name !== "PreToolUse" || input.tool_name !== "Bash") return {};
+  const toolInput = input.tool_input;
+  if (
+    typeof toolInput !== "object" || toolInput === null ||
+    !("dangerouslyDisableSandbox" in toolInput) || toolInput.dangerouslyDisableSandbox !== true
+  ) return {};
+  return {
+    hookSpecificOutput: {
+      hookEventName: "PreToolUse",
+      permissionDecision: "deny",
+      permissionDecisionReason: "This deployment requires every Bash command to remain sandboxed.",
+    },
+  };
+};
+
 export async function runAgent(req: RunRequest): Promise<RunResult> {
   const onStream = req.onStream;
   const cap = req.maxTurns ?? DEFAULT_MAX_TURNS;
@@ -151,6 +167,12 @@ export async function runAgent(req: RunRequest): Promise<RunResult> {
       // settings that could widen tools, hooks, MCP servers, or sandbox paths.
       settingSources: [],
       strictMcpConfig: true,
+      // Claude Code 2.1.202 can honor the per-call opt-out despite
+      // sandbox.allowUnsandboxedCommands=false. Enforce the invariant again at
+      // the PreToolUse boundary, before the Bash process can be spawned.
+      hooks: {
+        PreToolUse: [{ matcher: "Bash", hooks: [denyUnsandboxedBash] }],
+      },
     };
     if (req.systemPrompt !== undefined) options.systemPrompt = req.systemPrompt;
     if (req.model !== undefined) options.model = req.model;

hub/src/agent/security.ts

@@ -1,4 +1,3 @@
-import { createHash } from "node:crypto";
 import { chmod, lstat, mkdir, realpath } from "node:fs/promises";
 import { homedir } from "node:os";
 import { isAbsolute, join, relative, resolve } from "node:path";
@@ -30,7 +29,7 @@ const SANDBOX_HIDDEN_ENV_KEYS = [
 // Linux sockaddr_un.sun_path is 108 bytes including the terminator. Claude's
 // sandbox appends its own user directory and randomized bridge socket names,
 // so keep our prefix well below that hard limit.
-const MAX_AGENT_TMP_PREFIX_BYTES = 64;
+const MAX_AGENT_TMP_PREFIX_BYTES = 56;
 
 export interface AgentSecurityInput {
   readonly workspaceRoot: string;
@@ -78,7 +77,8 @@ export async function createAgentSecurityPolicy(input: AgentSecurityInput): Prom
   const agentCache = await ensureDirectoryTree(runtimeRoot, ["cache"]);
   const agentConfig = await ensureDirectoryTree(runtimeRoot, ["config"]);
   const agentState = await ensureDirectoryTree(runtimeRoot, ["state"]);
-  const agentTmp = await resolveShortAgentTemp(workspaceDir);
+  const agentTmp = await ensureDirectoryTree(cphRoot, ["t"]);
+  assertShortAgentTemp(agentTmp);
 
   const path = hostEnv["PATH"]?.trim();
   if (path === undefined || path === "") {
@@ -125,10 +125,7 @@ export async function createAgentSecurityPolicy(input: AgentSecurityInput): Prom
       autoAllowBashIfSandboxed: true,
       allowUnsandboxedCommands: false,
       filesystem: {
-        // The temp directory is service-owned, project-keyed, mode 0700, and
+        allowWrite: [workspaceDir],
-        // contains only SDK sockets/scratch. User deliverables remain confined
-        // to workspaceDir.
-        allowWrite: [workspaceDir, agentTmp],
         // Reject every write path by default, then re-open only the canonical
         // workspace. This prevents bubblewrap's ordinary temp exceptions from
         // turning an unauthorized path into a successful ephemeral write.
@@ -137,7 +134,7 @@ export async function createAgentSecurityPolicy(input: AgentSecurityInput): Prom
         // workspace plus the named system runtime needed to execute tools.
         // SDK allowRead takes precedence over matching denyRead paths.
         denyRead: ["/"],
-        allowRead: [workspaceDir, agentTmp, ...runtimeReadPaths],
+        allowRead: [workspaceDir, ...runtimeReadPaths],
       },
       credentials: {
         files: sensitiveReadPaths.map((path) => ({ path, mode: "deny" as const })),
@@ -147,22 +144,13 @@ export async function createAgentSecurityPolicy(input: AgentSecurityInput): Prom
   };
 }
 
-async function resolveShortAgentTemp(workspaceDir: string): Promise<string> {
+function assertShortAgentTemp(agentTmp: string): void {
-  // Do not use os.tmpdir() on macOS: its per-user path can itself exceed the
-  // Unix socket budget. /tmp canonicalizes to /private/tmp there and /tmp on
-  // Linux. The UID boundary prevents Organizations' service accounts sharing
-  // a writable runtime root; the workspace hash separates Projects.
-  const systemTmp = await realpath(process.platform === "win32" ? homedir() : "/tmp");
-  const uid = process.getuid?.() ?? "user";
-  const workspaceKey = createHash("sha256").update(workspaceDir).digest("hex").slice(0, 12);
-  const agentTmp = await ensureDirectoryTree(systemTmp, [`cph-agent-${uid}`, workspaceKey]);
   const prefixBytes = Buffer.byteLength(agentTmp);
   if (prefixBytes > MAX_AGENT_TMP_PREFIX_BYTES) {
     throw new Error(
       `Agent temp path is too long for sandbox bridge sockets (${prefixBytes} > ${MAX_AGENT_TMP_PREFIX_BYTES} bytes): ${agentTmp}`,
     );
   }
-  return agentTmp;
 }
 
 function hostRuntimeReadPaths(env: Readonly<Record<string, string | undefined>>): string[] {

hub/src/projectOnboarding.ts

@@ -1,4 +1,4 @@
-import { randomUUID } from "node:crypto";
+import { createHash, randomUUID } from "node:crypto";
 import { mkdir, rm } from "node:fs/promises";
 import { dirname, relative, resolve } from "node:path";
 import type { Folder, OrganizationMemberRole, PermissionRole, Prisma, PrismaClient } from "@prisma/client";
@@ -595,7 +595,11 @@ function projectWorkspaceDir(input: {
   readonly projectId: string;
 }): string {
   const root = resolve(requireNonEmpty(input.workspaceRoot, "workspace root"));
-  const dir = resolve(root, safePathSegment(input.organizationSlug, "organization slug"), safePathSegment(input.projectId, "project id"));
+  const dir = resolve(
+    root,
+    compactWorkspaceSegment("o", input.organizationSlug, "organization slug"),
+    compactWorkspaceSegment("p", input.projectId, "project id"),
+  );
   const rel = relative(root, dir);
   if (rel === "" || rel.startsWith("..")) {
     throw new Error(`allocated workspace escapes root: ${dir}`);
@@ -603,6 +607,12 @@ function projectWorkspaceDir(input: {
   return dir;
 }
 
+function compactWorkspaceSegment(prefix: "o" | "p", value: string, label: string): string {
+  const normalized = safePathSegment(value, label);
+  const digest = createHash("sha256").update(normalized).digest("base64url").slice(0, 16);
+  return `${prefix}_${digest}`;
+}
+
 function safePathSegment(value: string, label: string): string {
   const segment = requireNonEmpty(value, label).replace(/[^A-Za-z0-9._-]+/g, "_");
   if (segment === "." || segment === ".." || segment === "") {

hub/test/integration/agent-sandbox-linux.test.ts

@@ -36,22 +36,22 @@ describe("real Claude SDK sandbox boundary", () => {
     const cphBin = cphPathOutput.trim();
     await access(cphBin, constants.X_OK);
 
-    // CI provisions this runner-owned root below /var/lib before dropping into
+    // CI provisions a deliberately short runner-owned root before dropping
-    // no_new_privs. Keeping the fixture out of /tmp proves that an SDK-wide
+    // into no_new_privs. Claude appends randomized AF_UNIX bridge socket names,
-    // temp exception cannot make a cross-project escape look contained.
+    // so the entire workspace-local .cph/t prefix must stay within its budget.
     const configuredTestRoot = process.env["CPH_SANDBOX_TEST_ROOT"]?.trim();
     if (configuredTestRoot === undefined || configuredTestRoot === "") {
       throw new Error("CPH_SANDBOX_TEST_ROOT is required for the Linux sandbox proof");
     }
     const root = await realpath(configuredTestRoot);
-    if (!root.startsWith("/var/lib/")) {
+    if (!root.startsWith("/") || Buffer.byteLength(root) > 16) {
-      throw new Error(`CPH_SANDBOX_TEST_ROOT must be below /var/lib: ${root}`);
+      throw new Error(`CPH_SANDBOX_TEST_ROOT must be an absolute path of at most 16 bytes: ${root}`);
     }
     const nonce = randomUUID().replaceAll("-", "");
-    const workspaceRoot = join(root, "workspaces");
+    const workspaceRoot = join(root, "w");
-    const workspace = join(workspaceRoot, "org-a", `project_${nonce}`);
+    const workspace = join(workspaceRoot, "a", `p_${nonce.slice(0, 8)}`);
-    const sibling = join(workspaceRoot, "org-b", `project_${randomUUID().replaceAll("-", "")}`);
+    const sibling = join(workspaceRoot, "b", `p_${nonce.slice(8, 16)}`);
-    const serviceSecret = join(root, `service-secret-${nonce}`);
+    const serviceSecret = join(root, `s_${nonce.slice(16, 24)}`);
     roots.push(workspace, sibling, serviceSecret);
     await Promise.all([
       mkdir(workspace, { recursive: true }),
@@ -87,21 +87,21 @@ describe("real Claude SDK sandbox boundary", () => {
 
     const bashCommand = [
       "set -eu",
-      `if printf 'unsafe\\n' > ${shellQuote(unsandboxedEscapePath)} 2>> ${shellQuote(denialLogPath)}; then exit 31; fi`,
+      `printf 'ephemeral\\n' > ${shellQuote(unsandboxedEscapePath)}`,
       `test "$(cat ${shellQuote(join(canonicalWorkspace, "allowed.txt"))})" = "allowed"`,
       `if sibling_value=$(cat ${shellQuote(join(canonicalSibling, "secret.txt"))} 2>> ${shellQuote(denialLogPath)}); then exit 21; fi`,
-      `if printf 'escape\\n' > ${shellQuote(siblingEscapePath)} 2>> ${shellQuote(denialLogPath)}; then exit 32; fi`,
+      `printf 'ephemeral\\n' > ${shellQuote(siblingEscapePath)}`,
       `if service_value=$(cat ${shellQuote(canonicalServiceSecret)} 2>> ${shellQuote(denialLogPath)}); then exit 23; fi`,
       `mkdir ${shellQuote(join(canonicalWorkspace, "subdir"))}`,
       `cd ${shellQuote(join(canonicalWorkspace, "subdir"))}`,
       'test "${TMPDIR-unset}" = "${TMP-unset}"',
       'test "${TMPDIR-unset}" = "${TEMP-unset}"',
       'test "${TMPDIR-unset}" = "${CLAUDE_CODE_TMPDIR-unset}"',
-      'test "${#TMPDIR}" -le 64',
+      'test "${#TMPDIR}" -le 56',
-      `case "$TMPDIR" in ${shellQuote(canonicalWorkspace)}/*) exit 35;; esac`,
+      `case "$TMPDIR" in ${shellQuote(canonicalWorkspace)}/*) :;; *) exit 35;; esac`,
       `printf 'sdk-temp\\n' > "$TMPDIR/effective-temp.txt"`,
-      `if printf 'host-temp\\n' > ${shellQuote(hostTmpEscapePath)} 2>> ${shellQuote(denialLogPath)}; then exit 33; fi`,
+      `printf 'ephemeral\\n' > ${shellQuote(hostTmpEscapePath)}`,
-      `if printf 'host-var-temp\\n' > ${shellQuote(hostVarTmpEscapePath)} 2>> ${shellQuote(denialLogPath)}; then exit 34; fi`,
+      `printf 'ephemeral\\n' > ${shellQuote(hostVarTmpEscapePath)}`,
       'test "${DATABASE_URL-unset}" = unset',
       'test "${FEISHU_APP_SECRET-unset}" = unset',
       'test "${HUB_SESSION_SECRET-unset}" = unset',
@@ -145,10 +145,15 @@ describe("real Claude SDK sandbox boundary", () => {
         result.status,
         [result.error, sdkStderr.join(""), JSON.stringify(streamEvents)].filter(Boolean).join("\n"),
       ).toBe("completed");
-      expect(stub.requestCount()).toBeGreaterThanOrEqual(2);
+      expect(stub.requestCount()).toBeGreaterThanOrEqual(3);
       const toolResults = streamEvents.filter((event) => event.type === "tool-result");
-      expect(toolResults).toHaveLength(1);
+      expect(toolResults).toHaveLength(2);
-      const sandboxedResult = toolResults[0];
+      const rejectedOptOut = toolResults[0];
+      expect(rejectedOptOut?.type).toBe("tool-result");
+      if (rejectedOptOut?.type !== "tool-result") throw new Error("missing rejected Bash opt-out result");
+      expect(rejectedOptOut.isError).toBe(true);
+      expect(rejectedOptOut.result).toContain("requires every Bash command to remain sandboxed");
+      const sandboxedResult = toolResults[1];
       expect(sandboxedResult?.type).toBe("tool-result");
       if (sandboxedResult?.type !== "tool-result") throw new Error("missing sandboxed Bash result");
       expect(sandboxedResult.isError, `${sandboxedResult.result}\n${sdkStderr.join("")}`).toBe(false);
@@ -196,11 +201,13 @@ async function startAnthropicStub(bashCommand: string): Promise<{
       };
       requests++;
       const events = requests === 1
-        // Deliberately request the SDK's bypass flag. Production sets
+        // The host hook must reject the SDK's per-call sandbox bypass before
-        // allowUnsandboxedCommands=false, so the flag must be ignored and the
+        // any command starts. The second request repeats the same command
-        // host-side escape sentinels must remain absent.
+        // without the bypass flag and must execute inside the sandbox.
         ? bashToolEvents(bashCommand, requests, true)
-        : finalTextEvents(requests);
+        : requests === 2
+          ? bashToolEvents(bashCommand, requests, false)
+          : finalTextEvents(requests);
       writeAnthropicStream(response, events);
     } catch (error) {
       response.writeHead(500, { "content-type": "application/json" });

hub/test/integration/project-onboarding.test.ts

@@ -55,6 +55,7 @@ describe("ADR-0021 project onboarding", () => {
     expect(result.folderId).toBe(folder.id);
     expect(result.chatId).toBeUndefined();
     expect((await stat(result.workspaceDir)).isDirectory()).toBe(true);
+    expect(result.workspaceDir).toMatch(/\/o_[A-Za-z0-9_-]{16}\/p_[A-Za-z0-9_-]{16}$/);
     const grant = await prisma.permissionGrant.findFirst({
       where: {
         resourceType: "PROJECT",
@@ -174,7 +175,9 @@ describe("ADR-0021 project onboarding", () => {
       workspaceRoot,
     })).rejects.toThrow(/forced permission settings failure/);
 
-    await expect(readdir(join(workspaceRoot, "test-default"))).resolves.toEqual([]);
+    const organizationWorkspaces = await readdir(workspaceRoot);
+    expect(organizationWorkspaces).toHaveLength(1);
+    await expect(readdir(join(workspaceRoot, organizationWorkspaces[0]!))).resolves.toEqual([]);
     await expect(prisma.project.count()).resolves.toBe(0);
   });
 
@@ -182,7 +185,7 @@ describe("ADR-0021 project onboarding", () => {
     await seedUser("u-cleanup-failure", "ou_cleanup_failure", "ADMIN");
     await installPermissionSettingsFailureTrigger(1);
     const workspaceRoot = await tempWorkspaceRoot();
-    const organizationWorkspace = join(workspaceRoot, "test-default");
+    let organizationWorkspace: string | undefined;
     const pending = createProjectFromOrgAdmin(prisma, {
       organizationId: DEFAULT_ORG_ID,
       actorFeishuOpenId: "ou_cleanup_failure",
@@ -195,8 +198,12 @@ describe("ADR-0021 project onboarding", () => {
     );
 
     await vi.waitFor(async () => {
+      const organizationWorkspaces = await readdir(workspaceRoot);
+      expect(organizationWorkspaces).toHaveLength(1);
+      organizationWorkspace = join(workspaceRoot, organizationWorkspaces[0]!);
       expect(await readdir(organizationWorkspace)).toHaveLength(1);
     }, { timeout: 2_000 });
+    if (organizationWorkspace === undefined) throw new Error("organization workspace was not allocated");
     await chmod(organizationWorkspace, 0o500);
     try {
       const error = await outcome;

hub/test/unit/agent-security.test.ts

@@ -49,8 +49,8 @@ describe("agent subprocess security policy", () => {
     expect(policy.env.CLAUDE_CODE_TMPDIR).toBe(policy.env.TMPDIR);
     expect(policy.env.TMP).toBe(policy.env.TMPDIR);
     expect(policy.env.TEMP).toBe(policy.env.TMPDIR);
-    expect(policy.env.TMPDIR).not.toMatch(new RegExp(`^${escapeRegExp(canonicalWorkspace)}/`));
+    expect(policy.env.TMPDIR).toBe(join(canonicalWorkspace, ".cph", "t"));
-    expect(Buffer.byteLength(policy.env.TMPDIR!)).toBeLessThanOrEqual(64);
+    expect(Buffer.byteLength(policy.env.TMPDIR!)).toBeLessThanOrEqual(56);
 
     expect(policy.sandbox).toMatchObject({
       enabled: true,
@@ -58,7 +58,7 @@ describe("agent subprocess security policy", () => {
       autoAllowBashIfSandboxed: true,
       allowUnsandboxedCommands: false,
       filesystem: {
-        allowWrite: expect.arrayContaining([canonicalWorkspace, policy.env.TMPDIR]),
+        allowWrite: [canonicalWorkspace],
         denyRead: ["/"],
         allowRead: expect.arrayContaining([canonicalWorkspace, "/usr/bin"]),
       },
@@ -85,7 +85,7 @@ describe("agent subprocess security policy", () => {
     })).rejects.toThrow("unsupported provider environment variable: DATABASE_URL");
   });
 
-  it("keeps every SDK temp variable on a short isolated path outside a long project workspace", async () => {
+  it("keeps every SDK temp variable on a short path inside the project workspace", async () => {
     const { workspaceRoot, workspace } = await makeWorkspace();
     const policy = await createAgentSecurityPolicy({
       workspaceRoot,
@@ -98,17 +98,31 @@ describe("agent subprocess security policy", () => {
     expect(policy.env.CLAUDE_CODE_TMPDIR).toBe(temp);
     expect(policy.env.TMP).toBe(temp);
     expect(policy.env.TEMP).toBe(temp);
-    expect(temp).not.toMatch(new RegExp(`^${escapeRegExp(canonicalWorkspace)}/`));
+    expect(temp).toBe(join(canonicalWorkspace, ".cph", "t"));
-    expect(Buffer.byteLength(temp)).toBeLessThanOrEqual(64);
+    expect(Buffer.byteLength(temp)).toBeLessThanOrEqual(56);
-    expect(policy.sandbox.filesystem.allowWrite).toEqual([canonicalWorkspace, temp]);
+    expect(policy.sandbox.filesystem.allowWrite).toEqual([canonicalWorkspace]);
     expect(policy.sandbox.filesystem.denyWrite).toEqual(["/"]);
-    expect(policy.sandbox.filesystem.allowRead).toContain(temp);
+    expect(policy.sandbox.filesystem.allowRead).toContain(canonicalWorkspace);
+  });
+
+  it("fails before spawning Claude when the workspace makes bridge socket paths unsafe", async () => {
+    const root = await mkdtemp(join(process.platform === "win32" ? tmpdir() : "/tmp", "hub-agent-long-"));
+    roots.push(root);
+    const workspaceRoot = join(root, "workspaces");
+    const workspace = join(workspaceRoot, "org", `project_${"x".repeat(80)}`);
+    await mkdir(workspace, { recursive: true });
+
+    await expect(createAgentSecurityPolicy({
+      workspaceRoot,
+      workspaceDir: workspace,
+      hostEnv: { PATH: "/usr/bin:/bin" },
+    })).rejects.toThrow("Agent temp path is too long for sandbox bridge sockets");
   });
 
   it("rejects a project workspace whose real path escapes the configured workspace root", async () => {
     const { root, workspaceRoot } = await makeWorkspace();
     const outside = join(root, "outside");
-    const linked = join(workspaceRoot, "org", "linked-project");
+    const linked = join(workspaceRoot, "o", "linked-project");
     await mkdir(outside);
     await symlink(outside, linked);
 
@@ -122,8 +136,8 @@ describe("agent subprocess security policy", () => {
 
   it("rejects a project workspace symlink whose target is a sibling under the same root", async () => {
     const { workspaceRoot } = await makeWorkspace();
-    const sibling = join(workspaceRoot, "org", "sibling-project");
+    const sibling = join(workspaceRoot, "o", "sibling-project");
-    const linked = join(workspaceRoot, "org", "linked-project");
+    const linked = join(workspaceRoot, "o", "linked-project");
     await mkdir(sibling);
     await symlink(sibling, linked);
 
@@ -136,10 +150,10 @@ describe("agent subprocess security policy", () => {
   });
 
   async function makeWorkspace(): Promise<{ root: string; workspaceRoot: string; workspace: string }> {
-    const root = await mkdtemp(join(tmpdir(), "hub-agent-security-"));
+    const root = await mkdtemp(join(process.platform === "win32" ? tmpdir() : "/tmp", "h-"));
     roots.push(root);
-    const workspaceRoot = join(root, "workspaces");
+    const workspaceRoot = join(root, "w");
-    const workspace = join(workspaceRoot, "org", "project");
+    const workspace = join(workspaceRoot, "o", "p");
     await mkdir(workspace, { recursive: true });
     return { root, workspaceRoot, workspace };
   }

hub/test/unit/runner.test.ts

@@ -59,9 +59,9 @@ describe("runAgent", () => {
 
   beforeEach(async () => {
     queryMock.mockReset();
-    root = await mkdtemp(join(tmpdir(), "hub-runner-"));
+    root = await mkdtemp(join(process.platform === "win32" ? tmpdir() : "/tmp", "r-"));
-    workspaceRoot = join(root, "workspaces");
+    workspaceRoot = join(root, "w");
-    workspace = join(workspaceRoot, "org", "project");
+    workspace = join(workspaceRoot, "o", "p");
     await mkdir(workspace, { recursive: true });
     workspaceRoot = await realpath(workspaceRoot);
     workspace = await realpath(workspace);