Skip to content

Write HTML. Render video. Built for agents.

23.6 KB
Raw Blame
import { existsSync, readFileSync, statSync } from "node:fs";
import { platform, release } from "node:os";
import { detectWSL } from "./platform.js";

// ---------------------------------------------------------------------------
// Sandbox runtime + agent vendor fingerprinting.
//
// Goal: distinguish "real developer laptop" from "ephemeral managed sandbox
// driving the CLI on someone's behalf" (Codex Cloud, Claude Code Web, Cursor
// Background Agents, etc.) without collecting any PII.
//
// We only read:
//   - well-known kernel strings (release(), /proc/version)
//   - sandbox marker files (/.dockerenv etc.)
//   - the *existence* of vendor environment variables — never the value
//     (some are API keys).
//
// Output is two opaque strings: sandbox_runtime ('gvisor' | 'docker' | ...)
// and agent_runtime ('claude_code' | 'codex' | ...). Both null when unknown.
// ---------------------------------------------------------------------------

export type SandboxRuntime = "gvisor" | "firecracker" | "docker" | "kvm" | "wsl" | null;

export type AgentRuntime =
  | "bob"
  | "claude_code"
  | "codex"
  | "cursor"
  | "copilot_agent"
  | "replit"
  | "hermes"
  | "openclaw"
  | "pi"
  | "gemini_managed_agent"
  | "windsurf"
  | "cline"
  | "gemini_cli"
  | "crush"
  | null;

interface VendorRule {
  name: Exclude<AgentRuntime, null>;
  /** Check returns true when the named agent is driving the CLI. */
  check: (env: NodeJS.ProcessEnv) => boolean;
}

// Ordering matters: the FIRST rule that matches wins. Put more specific rules
// before more generic ones (e.g. copilot_agent before a hypothetical generic
// 'github_actions' rule).
const VENDOR_RULES: VendorRule[] = [
  // IBM Bob Shell — both shell execution paths stamp BOBSHELL_CLI=1 onto the
  // child environment: childProcessFallback's spawn env and executeWithPty's
  // PTY env. Keep this ahead of other agents because Bob inherits the parent
  // environment, so a Bob session launched from Codex or Claude can carry both
  // vendors' markers; BOBSHELL_CLI identifies the immediate executor.
  // BOBSHELL_API_KEY is deliberately ignored: it is an authentication setting,
  // not proof that Bob launched the current process.
  // Source: IBM Bob Shell 1.0.6 bundle (lb.childProcessFallback / executeWithPty).
  {
    name: "bob",
    check: (env) => typeof env["BOBSHELL_CLI"] === "string",
  },
  // Anthropic Claude Code — sets CLAUDECODE=1 on every Bash/PowerShell tool
  // spawn (Shell.ts:321) and CLAUDE_CODE_ENTRYPOINT at startup, inherited by
  // every child (main.tsx:527). Both propagate to spawned subprocesses.
  // Source: confirmed by @magi from Claude Code internal source.
  {
    name: "claude_code",
    check: (env) =>
      typeof env["CLAUDECODE"] === "string" || typeof env["CLAUDE_CODE_ENTRYPOINT"] === "string",
  },
  // OpenAI Codex (https://github.com/openai/codex).
  // - CODEX_THREAD_ID — set unconditionally on every spawned shell command
  //   (codex-rs/protocol/src/shell_environment.rs:6 constant, set by
  //   codex-rs/core/src/unified_exec/process_manager.rs:1010 and
  //   codex-rs/core/src/tools/runtimes/mod.rs:164).
  // - CODEX_CI — hardcoded in the UNIFIED_EXEC_ENV array, always set on
  //   every unified-exec child (process_manager.rs:70).
  // - CODEX_SANDBOX_NETWORK_DISABLED — set when network sandbox is active
  //   (codex-rs/core/src/sandboxing/mod.rs:135-138, default-on).
  // CODEX_HOME is deliberately NOT used — it's a config override read at
  // Codex startup, not propagated to spawned subprocesses.
  {
    name: "codex",
    check: (env) =>
      typeof env["CODEX_THREAD_ID"] === "string" ||
      typeof env["CODEX_CI"] === "string" ||
      typeof env["CODEX_SANDBOX_NETWORK_DISABLED"] === "string",
  },
  // Cursor Agent marks its shell with CURSOR_AGENT; TERM_PROGRAM identifies
  // the IDE terminal only. Keep this after Claude/Codex for nested agents.
  // Source: https://docs.cursor.com/en/agent/terminal
  {
    name: "cursor",
    check: (env) => Boolean(env["CURSOR_AGENT"]) || env["TERM_PROGRAM"] === "cursor",
  },
  // Windsurf (Codeium) integrated terminal — exports TERM_PROGRAM=windsurf.
  // Attested across many independent detectors (nx
  // packages/nx/src/native/ide/detection.rs, adonisjs/application, ag-grid
  // git-hooks). Compared case-INsensitively (unlike the exact Cursor rule
  // above) because Windsurf sources disagree on casing ("windsurf" vs
  // "Windsurf"); Cursor's do not, so it stays exact. Like Cursor this marks the
  // editor's integrated terminal, not specifically that the Cascade agent is
  // driving; under WSL/remote it can also fall back to TERM_PROGRAM=vscode.
  {
    name: "windsurf",
    check: (env) => env["TERM_PROGRAM"]?.toLowerCase() === "windsurf",
  },
  // GitHub Copilot Coding Agent — runs inside GitHub Actions and the
  // workflow injects an additional marker to distinguish from generic CI.
  // Not yet verified from a public-source citation in this audit; the var
  // names below match GitHub Copilot Coding Agent documentation but
  // should be confirmed before relying on attribution.
  {
    name: "copilot_agent",
    check: (env) =>
      env["GITHUB_ACTIONS"] === "true" &&
      (typeof env["COPILOT_AGENT_ID"] === "string" || env["RUNNER_NAME"] === "Copilot"),
  },
  // Replit — REPL_ID and REPLIT_USER are long-documented environment
  // variables exposed inside every Replit workspace.
  // Source: https://docs.replit.com/replit-workspace/configuring-the-environment
  {
    name: "replit",
    check: (env) => typeof env["REPL_ID"] === "string" || typeof env["REPLIT_USER"] === "string",
  },
  // Nous Research Hermes Agent — cli.py:50 unconditionally executes
  //   os.environ["HERMES_QUIET"] = "1"
  // at module load, so the marker propagates via os.environ to every
  // subprocess spawned by Hermes. Keying on existence (not the literal
  // "1") so we still match if Hermes ever changes the value.
  // Source: https://github.com/NousResearch/hermes-agent (cli.py:50)
  {
    name: "hermes",
    check: (env) => typeof env["HERMES_QUIET"] === "string",
  },
  // openclaw — multi-channel AI gateway. When openclaw spawns a CLI
  // subprocess it builds the child env with OPENCLAW_STATE_DIR /
  // OPENCLAW_CONFIG_PATH / OPENCLAW_DISABLE_AUTO_UPDATE set explicitly
  // (extensions/qa-matrix/src/runners/contract/scenario-runtime-cli.ts:344-351).
  // We key on OPENCLAW_STATE_DIR since it's a path scope-bound to openclaw.
  // Source: https://github.com/openclaw/openclaw
  {
    name: "openclaw",
    check: (env) =>
      typeof env["OPENCLAW_STATE_DIR"] === "string" ||
      typeof env["OPENCLAW_CONFIG_PATH"] === "string",
  },
  // Pi coding agent (https://pi.dev, https://github.com/earendil-works/pi).
  // packages/coding-agent/src/cli.ts:13 unconditionally executes
  //   process.env.PI_CODING_AGENT = "true";
  // at module entry, so every subprocess Pi spawns sees this marker.
  {
    name: "pi",
    check: (env) => typeof env["PI_CODING_AGENT"] === "string",
  },
  // Cline (cline/cline) VS Code extension — injects CLINE_ACTIVE=true into the
  // integrated terminal via vscode.TerminalOptions.env, which the terminal
  // exports to every shell command run in it
  // (apps/vscode/src/hosts/vscode/terminal/VscodeTerminalRegistry.ts:29).
  // Caveat: present only on the default "vscodeTerminal" exec path — the opt-in
  // backgroundExec/YOLO path spawns via child_process without the marker. Same
  // integrated-terminal-only scope as the Cursor/Windsurf rules above.
  // Source: https://github.com/cline/cline (VscodeTerminalRegistry.ts:29)
  {
    name: "cline",
    check: (env) => typeof env["CLINE_ACTIVE"] === "string",
  },
  // Google Gemini CLI (open-source @google/gemini-cli) — DISTINCT from the
  // Gemini managed-agent sandbox. (If a /.agents/ filesystem detector is
  // present in detectAgentRuntime() it runs ahead of this loop and wins for a
  // managed-agent sandbox, leaving this rule to match only the local CLI.)
  // The shell-execution service
  // sets GEMINI_CLI=1 on the child env of every shell command it spawns, so
  // downstream executables can tell they were launched by Gemini CLI
  // (packages/core/src/services/shellExecutionService.ts:56,486-487 — spread
  // onto baseEnv after sanitizeEnvironment, passed as env: to both the
  // child_process and node-pty spawn paths).
  // Caveat: under STRICT sanitization (when GITHUB_SHA is set / the GitHub
  // Action surface) GEMINI_CLI is not allow-listed and gets stripped — reliable
  // for the local CLI, not inside Gemini's GitHub Action runner.
  // Source: https://github.com/google-gemini/gemini-cli (shellExecutionService.ts:56,486-487)
  {
    name: "gemini_cli",
    check: (env) => typeof env["GEMINI_CLI"] === "string",
  },
  // Crush (charmbracelet/crush) — internal/shell/shell.go:43-48,98
  // unconditionally appends CRUSH=1 (plus generic AGENT=crush / AI_AGENT=crush)
  // to the env of every shell it spawns: both the interactive bash tool and the
  // hook runner. We key on CRUSH since AGENT/AI_AGENT are generic and collide.
  // Source: https://github.com/charmbracelet/crush (internal/shell/shell.go:43-48,98)
  {
    name: "crush",
    check: (env) => typeof env["CRUSH"] === "string",
  },
];

// Agents evaluated and deliberately NOT added. Each fails the bar the rules
// above meet — a marker reliably present in the environment of the
// shell/subprocess the agent spawns. Recorded here (not only in the PR) so the
// next person doesn't re-derive it:
//   - OpenHands — OPENHANDS_BUILD_GIT_SHA/_REF exists in the agent-server
//     Dockerfile (base-image-minimal stage, added 2025-11-09 in PR #1100) but
//     is empirically ABSENT from the runtime env of every published
//     ghcr.io/openhands/agent-server image inspected (12+ tags, 2025-10 →
//     2026-01, incl. the introducing PR's merge commit). The declared ENV
//     never reaches the published image. Re-add only if a real published image
//     carries it in `docker inspect .Config.Env`.
//   - Aider — sets no self-identifying env var; both shell-spawn sites
//     (run_cmd.py Popen / pexpect.spawn) pass no env=, so children inherit
//     os.environ verbatim.
//   - Goose — AGENT=goose/GOOSE_TERMINAL=1 are set on the recipe-retry path
//     and the computercontroller MCP extension, but NOT on the default
//     developer `shell` tool (sets only PATH+cwd), so the primary path is
//     undetected.
//   - opencode — OPENCODE_TERMINAL=1 is set only on the interactive PTY panel,
//     not on the model's bash/shell tool.
//   - Roo Code — ROO_ACTIVE is set only on the `vscode` terminal provider; the
//     shipped default is the execa provider (terminalShellIntegrationDisabled
//     defaults true), which sets no marker.
//   - Amp / Devin / Jules / Factory Droid — no verifiable unconditional runtime
//     marker (Amp is closed/minified; Devin/Jules are closed sandboxes;
//     Factory's FACTORY_PROJECT_DIR / DROID_PLUGIN_ROOT are hook-scoped only).

/**
 * Identify the managed sandbox runtime hosting this CLI invocation.
 * Returns null on a normal developer machine. Dispatches to runtime-specific
 * detectors that each return a boolean; the priority order encoded here is
 * deliberate (WSL > gVisor > Docker > Firecracker > KVM).
 */
export function detectSandboxRuntime(): SandboxRuntime {
  if (platform() === "win32") return null;
  if (detectWSL()) return "wsl";
  if (isGVisor()) return "gvisor";
  if (isDocker()) return "docker";
  if (isFirecracker()) return "firecracker";
  if (isKVM()) return "kvm";
  return null;
}

/**
 * Identify the coding-agent vendor that spawned this process, if any.
 * Returns null on a regular interactive shell. Most rules only check the
 * EXISTENCE of well-known env vars (never their values), but a few agents
 * are best identified by filesystem markers — those run via dedicated
 * detector functions ahead of the env-var rule loop.
 */
export function detectAgentRuntime(): AgentRuntime {
  // Gemini managed agent — keyed on the `/.agents/` platform mount with a
  // gVisor guard against false positives. See `isGeminiManagedAgent` for the
  // uniqueness-anchor-vs-guard split. Env vars alone are insufficient
  // (`GEMINI_API_KEY` is user-settable), so this runs ahead of VENDOR_RULES.
  if (isGeminiManagedAgent()) return "gemini_managed_agent";
  for (const rule of VENDOR_RULES) {
    if (rule.check(process.env)) return rule.name;
  }
  return null;
}

/** Observed harness marker, not proof of a vendor, benchmark, or human user. */
export function detectExecutionHarnessHint(): "harbor" | null {
  // Read presence only. Preserve this context even when the child agent is known.
  return typeof process.env["HARBOR_AGENT"] === "string" ? "harbor" : null;
}

// ---------------------------------------------------------------------------
// New-agent discovery signals.
//
// VENDOR_RULES is a CLOSED allowlist: an agent we haven't written a rule for
// collapses to agent_runtime=null, leaving no trace of what it was. That makes
// the null bucket un-attributable — new agents stay invisible until someone
// reverse-engineers their marker by hand (which is how every rule above was
// derived).
//
// detectAgentHints() adds a self-populating residual signal for exactly that
// null bucket, so an unrecognized agent surfaces on its own in analytics and
// can be promoted to a real VENDOR_RULE later. Callers should only emit these
// when detectAgentRuntime() returns null — a classified event needs no hint.
//
// Privacy — consistent with the "never read secret-shaped values" stance above:
//   - agent_env_hints emits KEY NAMES only, never values.
//   - agent_hint / term_program read the VALUE of three vars whose sole purpose
//     is non-secret self-identification: the emerging AGENT / AI_AGENT
//     agent-name convention (e.g. Crush and Goose set AGENT=<name>) and
//     TERM_PROGRAM's editor name (how the cursor/windsurf rules already work).
//     Each value is passed through a strict short-slug allowlist, so anything
//     long, spaced, or secret-shaped is dropped to null.
// ---------------------------------------------------------------------------

export interface AgentHints {
  /**
   * Best-effort agent name from a self-identifying env-var value: AGENT, else
   * AI_AGENT. null when neither is set or the value isn't a short safe slug.
   */
  agent_hint: string | null;
  /**
   * Raw TERM_PROGRAM value (editor/terminal name) — surfaces IDE-terminal
   * agents not yet covered by the cursor/windsurf rules. Noisy (also set by
   * plain human terminals); read alongside is_tty=false. null when unset/unsafe.
   */
  term_program: string | null;
  /**
   * Sorted, comma-joined list of "agent-ish" env-var KEY names present but
   * matched by no vendor rule — a compact fingerprint that clusters by agent.
   * KEYS only, never values. null when none are present.
   */
  agent_env_hints: string | null;
}

// Self-identifying values are agent/editor NAMES by convention — short slugs.
// Anything longer, spaced, or secret-shaped falls outside this and is dropped.
const SAFE_HINT = /^[a-z0-9_.-]{1,32}$/;

// The slug allowlist alone still accepts SHORT credential-shaped values
// (AGENT=sk-ant-api03, AGENT=AKIAIOSFODNN7EXAMPLE, AGENT=github_pat_abc), so
// two extra guards enforce the "never emit a secret" boundary this PR relies on:
//   1. known credential/token prefixes (compared lowercased), and
//   2. any unbroken alphanumeric run >= 16 chars — the shape of key bodies,
//      hex digests, and base64-ish tokens (agent names segment on _/-/. and
//      keep each run short).
const CREDENTIAL_PREFIXES = [
  "sk-",
  "sk_",
  "pk-",
  "pplx-",
  "ghp_",
  "gho_",
  "ghu_",
  "ghs_",
  "ghr_",
  "github_pat_",
  "glpat-",
  "gsk_",
  "xox",
  "akia",
  "asia",
  "aiza",
  "ya29",
  "hf_",
  "r8_",
];
const LONG_ALNUM_RUN = /[a-z0-9]{16,}/;

function looksLikeCredential(v: string): boolean {
  if (CREDENTIAL_PREFIXES.some((p) => v.startsWith(p))) return true;
  return LONG_ALNUM_RUN.test(v);
}

function sanitizeHint(value: string | undefined): string | null {
  if (typeof value !== "string") return null;
  const v = value.trim().toLowerCase();
  if (!SAFE_HINT.test(v)) return null;
  if (looksLikeCredential(v)) return null;
  return v;
}

// A key (uppercased) looks like a coding-agent marker. Excludes the two
// value-captured generics (read via agent_hint) and the SSH/GPG "agent" false
// friends, which are credential agents, not coding agents. No bare `CODING`
// token — it substring-matches `ENCODING` (e.g. PYTHONIOENCODING) and real
// coding-agent keys already match via `AGENT` (e.g. PI_CODING_AGENT).
export const HINT_KEY_PATTERN = /AGENT|ASSISTANT|COPILOT|CODEX|CLAUDE|LLM|_THREAD_ID$|_SESSION_ID$/;

function isDiscoveryHintKey(upperKey: string): boolean {
  if (upperKey === "AGENT" || upperKey === "AI_AGENT") return false;
  if (upperKey.startsWith("SSH_") || upperKey.startsWith("GPG_")) return false;
  if (["NPM_CONFIG_USER_AGENT", "TERM_SESSION_ID", "XDG_SESSION_ID"].includes(upperKey))
    return false;
  return HINT_KEY_PATTERN.test(upperKey);
}

function discoveryHintPriority(key: string): number {
  // Kubernetes service expansion can otherwise consume the entire hint cap.
  if (/_SERVICE_(HOST|PORT)(_|$)|_PORT(_\d+_(TCP|UDP)(_(ADDR|PORT|PROTO))?)?$/.test(key)) return 2;
  if (/^(HARBOR_AGENT|CURSOR_AGENT|ANTIGRAVITY_AGENT)$|_(SESSION_ID|THREAD_ID)$/.test(key))
    return 0;
  return 1;
}

/**
 * Residual discovery signals for the agent_runtime=null bucket. See the section
 * comment above for intent and the privacy contract. Pure over process.env.
 */
export function detectAgentHints(): AgentHints {
  const env = process.env;
  const agent_hint = sanitizeHint(env["AGENT"]) ?? sanitizeHint(env["AI_AGENT"]);
  const term_program = sanitizeHint(env["TERM_PROGRAM"]);

  const keys = new Set<string>();
  for (const key of Object.keys(env)) {
    if (key.length > 64) continue;
    const upper = key.toUpperCase();
    if (isDiscoveryHintKey(upper)) keys.add(upper);
  }
  const selected = [...keys]
    .sort((a, b) => discoveryHintPriority(a) - discoveryHintPriority(b) || a.localeCompare(b))
    .slice(0, 16)
    .sort();
  const agent_env_hints = selected.length ? selected.join(",") : null;

  return { agent_hint, term_program, agent_env_hints };
}

// ---------------------------------------------------------------------------
// Sandbox runtime detectors — one per runtime, kept small and side-effect-free.
// ---------------------------------------------------------------------------

/**
 * gVisor reports kernel string `4.19.0-gvisor` (current) or `4.4.0` (legacy
 * Sentry kernel).
 *
 * `4.19.0-gvisor` is unambiguous — no real Linux box reports that string.
 * `4.4.0` collides with Ubuntu 16.04 LTS / older real kernels, so we only
 * accept it as a gVisor signal when /proc/version ALSO contains "gVisor".
 */
function isGVisor(): boolean {
  const kernel = release();
  if (kernel.includes("gvisor")) return true;
  if (platform() !== "linux") return false;
  try {
    const procVersion = readFileSync("/proc/version", "utf-8");
    return procVersion.includes("gVisor");
  } catch {
    return false;
  }
}

function isDocker(): boolean {
  if (existsSync("/.dockerenv")) return true;
  if (platform() !== "linux") return false;
  try {
    const cgroup = readFileSync("/proc/1/cgroup", "utf-8");
    return cgroup.includes("docker") || cgroup.includes("containerd");
  } catch {
    return false;
  }
}

/**
 * AWS Firecracker microVMs expose /dev/vsock and report sys_vendor='Amazon EC2'
 * with product_name containing 'Firecracker'. Full EC2 reports a real instance
 * type like 't3.large', so the product_name check distinguishes them.
 */
function isFirecracker(): boolean {
  if (platform() !== "linux") return false;
  if (!existsSync("/dev/vsock")) return false;
  try {
    const sysVendor = readFileSync("/sys/class/dmi/id/sys_vendor", "utf-8").trim();
    if (sysVendor !== "Amazon EC2") return false;
    const productName = readFileSync("/sys/class/dmi/id/product_name", "utf-8").trim();
    return productName.toLowerCase().includes("firecracker");
  } catch {
    return false;
  }
}

function isKVM(): boolean {
  if (platform() !== "linux") return false;
  try {
    const sysVendor = readFileSync("/sys/class/dmi/id/sys_vendor", "utf-8").trim();
    return sysVendor === "QEMU" || sysVendor.includes("KVM");
  } catch {
    return false;
  }
}

/**
 * Gemini managed-agent sandbox — Google's Managed Agents runtime (the
 * Antigravity base agent). The platform auto-discovers the agent definition
 * under `/.agents/` and runs it inside a gVisor kernel.
 *
 * Signal hierarchy (the two checks are NOT co-equal):
 *   - `/.agents/` is the *uniqueness anchor*: the platform's agent-definition
 *     mount root. Per Google's Managed Agents docs the runtime scans `/.agents/`
 *     for the agent's instructions (`/.agents/AGENTS.md`) and skills
 *     (`/.agents/skills/<name>/SKILL.md`). Nothing in the generic
 *     Google-Cloud-on-gVisor universe (Cloud Run gen2, GKE Sandbox, Fly.io)
 *     mounts `/.agents/` at the filesystem root.
 *
 *     We key on the `/.agents/` DIRECTORY, not `/.agents/AGENTS.md`: AGENTS.md
 *     is OPTIONAL. Google's docs state "AGENTS.md is optional ... the
 *     system_instruction and AGENTS.md are additive; both apply when present",
 *     so an agent may declare its instructions inline via `system_instruction`
 *     in agent.yaml and ship no AGENTS.md file. Keying on the file would
 *     silently miss every managed agent that uses inline instructions or a
 *     skills-only definition; the directory mount generalizes across all
 *     managed agents that ship any definition.
 *   - `isGVisor()` is a *guard*, not a co-uniqueness signal. gVisor itself
 *     is shared with GKE Sandbox + Cloud Run gen2 — it does not discriminate
 *     the managed-agent surface from those. Its job here is to rule out the
 *     unlikely case of a human creating `/.agents/` on a non-sandbox host.
 *
 * Known coverage gap: an agent defined with ONLY inline `system_instruction`
 * (no skills, no AGENTS.md) may not materialize a `/.agents/` mount — that
 * tail can't be closed from the docs and needs an empirical spin to confirm.
 * The common case (skills and/or AGENTS.md present) is covered.
 *
 * Things deliberately NOT keyed on (each fails the uniqueness test —
 * shared across the broader Google-Cloud-on-gVisor universe or trivially
 * user-settable on any host):
 *   - gVisor alone
 *   - `Google Compute Engine` DMI (entire GCP reports this)
 *   - `job` cgroup (Google-internal but broadly present)
 *   - egress-proxy env / CA-cert env cluster (any MITM container sets these)
 *   - `/workspace/` (the agent's data mount — generic, not unique)
 *   - `GEMINI_API_KEY` (user-settable on any host)
 *
 * Source: Google Managed Agents docs (ai.google.dev/gemini-api/docs/custom-agents
 * + managed-agents-quickstart) for the `/.agents/` mount contract and AGENTS.md
 * optionality; empirical introspection of live managed-agent sandboxes by
 * gemini-agent (2026-06-09) for the gVisor pairing — present across 3
 * independent fresh sandbox spins (spike + `b9db4e56` + `d59d6361`).
 */
function isGeminiManagedAgent(): boolean {
  if (platform() !== "linux") return false;
  // The uniqueness anchor: the managed-agent definition mount root. We key on
  // the `/.agents/` directory (not the optional AGENTS.md file) so skills-only
  // and inline-instruction agents are still detected. Nothing else on gVisor
  // (Cloud Run, GKE Sandbox, Fly.io) creates this path. Require an actual
  // directory — a stray file or symlink named `/.agents` must not match.
  try {
    if (!statSync("/.agents").isDirectory()) return false;
  } catch {
    return false; // ENOENT / EACCES — no mount, not a managed agent.
  }
  // The guard: rule out a stray user-created `/.agents/` on a non-sandbox
  // host. Not a second uniqueness signal — gVisor isn't unique on its own.
  return isGVisor();
}