fix(scripts): stop pinning vm=False on sandbox creation

packages/coding-agent/skills/prime-intellect/SKILL.md

@@ -16,7 +16,7 @@ Prime Intellect is an open superintelligence lab building open-source AGI infras
 | Hosted Evaluations | Run evals on Prime-managed infra (`prime eval run --hosted`) | [environments.md](references/environments.md) |
 | Hosted Training | Post-train models against environments (`prime train`, Lab) | [environments.md](references/environments.md) |
 | prime-rl | Large-scale async RL framework for self-managed training | [environments.md](references/environments.md) |
-| Sandboxes | Secure disposable Docker environments for AI-generated code | [sandboxes.md](references/sandboxes.md) |
+| Sandboxes | Secure disposable VM sandboxes for AI-generated code | [sandboxes.md](references/sandboxes.md) |
 | Tunnels | Public HTTPS URLs for local/sandboxed services | [sandboxes.md](references/sandboxes.md) |
 | Inference | OpenAI-compatible API for frontier models | [inference.md](references/inference.md) |
 | Compute | Rent single GPU pods or multi-node clusters | [compute.md](references/compute.md) |

packages/coding-agent/skills/prime-intellect/references/sandboxes.md

@@ -1,6 +1,6 @@
 # Sandboxes & Tunnels
 
-Prime Sandboxes are disposable Docker environments for running AI-generated or untrusted code in the cloud: isolated, fast to create, billed only while running (CPU $0.05/core/hr, memory $0.01/GB/hr, disk $0.001/GB/hr).
+Prime Sandboxes are disposable VM environments for running AI-generated or untrusted code in the cloud: isolated, fast to create, billed only while running (CPU $0.05/core/hr, memory $0.01/GB/hr, disk $0.001/GB/hr).
 
 Live docs: `sandboxes/overview.md`, `sandboxes/cli.md`, `sandboxes/sdk.md`, `sandboxes/images.md`, `sandboxes/tunnel.md` under https://docs.primeintellect.ai/
 
@@ -24,14 +24,13 @@ prime sandbox create python:3.11-slim \
   --idle-timeout-minutes 15 \
   --env APP_ENV=staging \
   --secret API_KEY=sk-abc123 \
-  --start-command "python serve.py --port 8000" \
   --yes
 ```
 
-- `--timeout-minutes` caps total lifetime; `--idle-timeout-minutes` reaps the sandbox early when no exec/upload/download/file-read arrives (1 ≤ idle ≤ timeout ≤ 1440; not supported with `--vm`).
-- `--env` values are plain text; `--secret` values are encrypted at rest and obfuscated in output. Both become environment variables inside the container.
-- Default start command is `tail -f /dev/null` (idle, ready for `prime sandbox run`); `--start-command` replaces the image ENTRYPOINT.
-- Outbound internet is on by default; use `--no-network-access` for isolation when running untrusted code.
+- `--timeout-minutes` caps total lifetime; `--idle-timeout-minutes` reaps the sandbox early when no exec/upload/download/file-read arrives (1 ≤ idle ≤ timeout ≤ 1440; for VM sandboxes, `timeout_minutes` may be negative to disable the lifetime deadline).
+- `--env` values are plain text; `--secret` values are encrypted at rest and obfuscated in output. Both become environment variables inside the sandbox.
+- Default start command keeps the sandbox idle, ready for `prime sandbox run`. To run your own process at boot, pass the command after `--` as separate tokens (no shell): `prime sandbox create python:3.11-slim -- python serve.py --port 8000`.
+- Outbound internet is on by default; restrict egress with network allow/deny lists (`prime sandbox network`, or `network_allowlist`/`network_denylist` in the SDK).
 - `--team-id` bills a team; `--yes` skips confirmation in automation.
 
 Custom images: build and push your own via Prime Images (`sandboxes/images.md`).
@@ -55,7 +54,6 @@ client.wait_for_creation(sandbox.id)
 result = client.execute_command(sandbox.id, "python -c 'print(42)'")
 client.upload_file(sandbox.id, "/workspace/data.csv", "./data.csv")
 client.download_file(sandbox.id, "/workspace/output.csv", "./output.csv")
-exposed = client.expose(sandbox.id, port=8000, name="web")   # public URL
 client.delete(sandbox.id)
 ```
 

scripts/benchmarks/README.md

@@ -36,7 +36,7 @@ The existing Vouch gate controls who can trigger compute usage.
 
 ## Measurements
 
-The default configuration uses two Linux x64 containers, each with 4 vCPU, 8 GB RAM, and 20 GB disk.
+The default configuration uses two Linux x64 VM sandboxes, each with 4 vCPU, 8 GB RAM, and 20 GB disk.
 `config.json` pins the image and sampling policy. Harness dependencies are locked with a seven-day
 release cutoff. Provisioning, harness setup, and source compilation have separate recorded durations
 outside the timed installation interval. Interactive runs use the same small committed Git fixture.

scripts/benchmarks/controller.py

@@ -137,7 +137,6 @@ def start(self, role: str) -> Sandbox:
                 cpu_cores=config.cpu_cores,
                 memory_gb=config.memory_gb,
                 disk_size_gb=config.disk_gb,
-                vm=False,
                 region=config.region,
                 timeout_minutes=config.ttl_minutes,
                 labels=labels(self.report.repository, self.report.run_id, self.report.attempt)

scripts/benchmarks/tests/test_benchmarks.py

@@ -426,7 +426,7 @@ def test_newer_comment_is_not_overwritten(self):
 
 
 class LifecycleTests(unittest.TestCase):
-    def test_sandbox_creation_does_not_inject_credentials(self):
+    def test_sandbox_creation_does_not_inject_credentials_or_pin_vm(self):
         with (
             tempfile.TemporaryDirectory() as directory,
             patch.dict(os.environ, {"PRIME_SANDBOX_API_KEY": "fake", "PINFERENCE_API_KEY": "must-not-use"}),
@@ -441,6 +441,7 @@ def test_sandbox_creation_does_not_inject_credentials(self):
             request = controller.client.create.call_args.args[0]
             self.assertIsNone(request.secrets)
             self.assertIsNone(request.environment_vars)
+            self.assertNotIn("vm", request.model_fields_set)
 
     def test_invalid_results_get_a_failure_notice_and_missing_results_preserve_pending_trust(self):
         with tempfile.TemporaryDirectory() as directory:

scripts/evals/short_swe/build_controller.py

@@ -123,7 +123,6 @@ def build(
             cpu_cores=config["cpu_cores"],
             memory_gb=config["memory_gb"],
             disk_size_gb=config["disk_gb"],
-            vm=False,
             region=config["region"],
             timeout_minutes=BUILD_SANDBOX_TIMEOUT_MINUTES,
             team_id=os.environ.get("PRIME_TEAM_ID") or None,