Cloudflare Adapters¶

Two adapters let you run agents on Cloudflare infrastructure instead of locally: the Cloudflare Agents SDK adapter and the Codex-on-Cloudflare adapter.

Cloudflare Agents SDK Adapter¶

Module: bernstein.adapters.cloudflare_agents Class: CloudflareAgentsAdapter

Spawns agents via a Cloudflare Workers backend using npx wrangler dev locally. The adapter launches a local wrangler dev server that hosts a Cloudflare Agents SDK worker, passing the task prompt and model as Worker variables.

Prerequisites¶

• Node.js 18+ and npm
• wrangler installed globally: npm install -g wrangler
• wrangler login completed
• A Cloudflare account with Workers enabled

Environment variables¶

How it works¶

1. The adapter builds a npx wrangler dev command with --var flags injecting the prompt, model, and session ID.
2. The command is wrapped with build_worker_cmd() for process visibility in bernstein ps.
3. Environment variables are filtered to only forward the Cloudflare-specific keys listed above (via build_filtered_env()).
4. The wrangler dev server runs as a subprocess with stdout/stderr captured to .sdd/runtime/<session>.log.
5. A timeout watchdog monitors the process.

Configuration in bernstein.yaml¶

cli: cloudflare_agents

Spawn parameters¶

The adapter's spawn() method accepts the standard CLIAdapter parameters:

Codex-on-Cloudflare Adapter¶

Module: bernstein.adapters.codex_cloudflare Class: CodexCloudflareAdapter

Runs OpenAI Codex agents inside Cloudflare sandboxes rather than locally. Combines Codex CLI capabilities with Cloudflare's isolated sandbox infrastructure for secure, scalable code execution.

Configuration¶

CodexSandboxConfig dataclass fields:

Usage¶

from bernstein.adapters.codex_cloudflare import (
    CodexCloudflareAdapter,
    CodexSandboxConfig,
)

adapter = CodexCloudflareAdapter(CodexSandboxConfig(
    cloudflare_account_id="abc123",
    cloudflare_api_token="cf_token_...",
    openai_api_key="sk-...",
    memory_mb=1024,
    max_execution_minutes=60,
))

# Execute a task
result = await adapter.execute(
    prompt="Add input validation to all API endpoints",
    workspace_id="task-123",
    model="codex-mini",
    timeout_minutes=45,
)

print(result.status)                # "completed", "failed", "timeout"
print(result.files_changed)         # ["src/api/validation.py", ...]
print(result.execution_time_seconds)
print(result.stdout)

Execution lifecycle¶

1. Create sandbox -- provisions a Cloudflare sandbox container with the specified image, memory, CPU, and network settings. Injects OPENAI_API_KEY, WORKSPACE_R2_BUCKET, and WORKSPACE_ID as environment variables.
2. Sync workspace -- the sandbox pulls workspace files from the configured R2 bucket.
3. Inject Codex command -- sends codex exec --full-auto -m <model> <prompt> to the sandbox.
4. Poll for completion -- checks sandbox status every 5 seconds until completed, failed, or timeout.
5. Collect results -- fetches stdout/stderr logs from the sandbox.
6. Cleanup -- terminates the sandbox on timeout or error.

Result type¶

CodexSandboxResult fields:

Management methods¶

# Check status of a running sandbox
status = await adapter.get_status("sandbox-id")

# Cancel execution
await adapter.cancel("sandbox-id")

# Get logs
logs = await adapter.get_logs("sandbox-id")

Choosing between adapters¶