[Bug] Hardcoded Default Values Scattered in Sandbox Configuration

[Timandes]

Bug Description

The Python SDK contains numerous hardcoded default values in sdk/sandbox/config.py and sdk/sandbox/client.py. These values cannot be customized via environment variables, reducing flexibility for different deployment scenarios.

Key hardcoded values include:

In sdk/sandbox/config.py:

Configuration	Hardcoded Value	Line
image	"python:3.11"	L23
auto_clear_seconds	60 * 5 (300)	L24
memory	"8g"	L27
cpus	2	L28
cluster	"zb"	L31
size (SandboxGroup)	2	L35
start_concurrency	2	L36
start_retry_times	3	L37

In sdk/sandbox/client.py:

Configuration	Hardcoded Value	Location
Status check interval	3 seconds	await asyncio.sleep(3) in start()
wait_timeout default	300 seconds	arun() parameter
wait_interval default	10 seconds	arun() parameter
Nohup command timeout	30 seconds	BashAction(..., timeout=30)
Minimum check interval	5 seconds	max(5, wait_interval)

Steps to Reproduce

1. Review the source code at sdk/sandbox/config.py and sdk/sandbox/client.py
2. Observe that default values are directly embedded in the code
3. Try to change cluster default from "zb" to another cluster without modifying code - not possible via environment variable

Expected Behavior

• Default values should be configurable via environment variables (e.g., ROCK_DEFAULT_CLUSTER, ROCK_DEFAULT_IMAGE, ROCK_DEFAULT_MEMORY, ROCK_DEFAULT_CPUS)
• The env_vars.py module should support these configuration environment variables with sensible fallback defaults
• Consistency with existing pattern: ROCK_BASE_URL, ROCK_SANDBOX_STARTUP_TIMEOUT_SECONDS are already configurable via environment variables

Actual Behavior

• cluster defaults to "zb" - a China-specific cluster that may not be appropriate for international users or different deployment environments
• Resource defaults (memory=8g, cpus=2) may not suit all use cases
• No way to change these defaults without modifying the source code

Suggested Solution

1. Add new environment variables to env_vars.py:

"ROCK_DEFAULT_CLUSTER": lambda: os.getenv("ROCK_DEFAULT_CLUSTER", "zb"),
"ROCK_DEFAULT_IMAGE": lambda: os.getenv("ROCK_DEFAULT_IMAGE", "python:3.11"),
"ROCK_DEFAULT_MEMORY": lambda: os.getenv("ROCK_DEFAULT_MEMORY", "8g"),
"ROCK_DEFAULT_CPUS": lambda: float(os.getenv("ROCK_DEFAULT_CPUS", "2")),
"ROCK_DEFAULT_AUTO_CLEAR_SECONDS": lambda: int(os.getenv("ROCK_DEFAULT_AUTO_CLEAR_SECONDS", "300")),

2. Update sdk/sandbox/config.py to use these environment variables:

class SandboxConfig(BaseConfig):
    image: str = env_vars.ROCK_DEFAULT_IMAGE
    auto_clear_seconds: int = env_vars.ROCK_DEFAULT_AUTO_CLEAR_SECONDS
    memory: str = env_vars.ROCK_DEFAULT_MEMORY
    cpus: float = env_vars.ROCK_DEFAULT_CPUS
    cluster: str = env_vars.ROCK_DEFAULT_CLUSTER
    # ...

Error Logs

N/A - This is a configuration flexibility issue, not a runtime error.

Environment Information

• OS: Any
• Python Version: Any
• ROCK Version: Current (as of March 2026)
• Installation Method: pip install rl-rock
• Docker Version: N/A
• Deployment Type: Any

ROCK Configuration

• Runtime Environment Type: Any
• Sandbox Image: Default python:3.11
• Resource Allocation: Default memory=8g, cpus=2

Component Affected

• Sandbox
• Actions
• Deployments
• SDK & API
• Envhub
• CLI
• Performance & Optimization
• Documentation & Examples

---

Additional Context

This issue was identified during the TypeScript SDK code review (PR #492). The TypeScript SDK mirrors the Python SDK's behavior and has the same hardcoded defaults. Fixing this in the Python SDK would also guide the TypeScript SDK implementation.