Skip to main content
Version: 1.2.x

v1.2.0

Release Date​

February 3, 2026

SDK​

New Features​

Runtime Environment​

Runtime Environment References Documentation

  • NEW: Executable symlink support via extra_symlink_dir and extra_symlink_executables for exposing runtime executables to system paths

Model Service (Experimental)​

Model Service References Documentation

Handles AI model call communications between agents and LLM inference services:

  • Local mode: File-based IPC between Agent and Roll runtime, implementing request-response via file protocol

  • Proxy mode: Request forwarding to external LLM services with routing and retry support

  • SDK supports anti_call_llm for proxying Agent requests

  • NEW: Complete CLI start command parameters (--config-file, --host, --port, --proxy-base-url, --retryable-status-codes, --request-timeout)

  • NEW: Trajectory (traj) logging - records LLM request/response to JSONL files (ROCK_MODEL_SERVICE_DATA_DIR, ROCK_MODEL_SERVICE_TRAJ_APPEND_MODE)

  • NEW: Server-side ModelServiceConfig configuration (host, port, proxy_base_url, proxy_rules, retryable_status_codes, request_timeout)

Agent Examples​

Examples Directory

Practical examples demonstrating ROCK Agent integration with various AI frameworks:

  • claude_code/: Claude Code Agent integration example using Node.js runtime

    • Command: claude -p ${prompt}
    • Requires: npm install -g @anthropic-ai/claude-code
    • Configuration via rock_agent_config.yaml

  • iflow_cli/: iFlow CLI Agent integration example using Node.js runtime

    • Command: iflow -p ${prompt} --yolo
    • Requires: npm i -g @iflow-ai/iflow-cli@latest

  • swe_agent/: SWE Agent integration example

    • Demonstrates standard ROCK Agent setup pattern

  • iflow_cli/integration_with_model_service/: Model Service integration examples

    • local/: Local mode example with custom LLM backend
      • Shows anti_call_llm usage and model service loop pattern
    • proxy/: Proxy mode example with iFlow CLI
      • Demonstrates rock model-service start --type proxy --proxy-base-url usage

Admin​

New Features​

Task Scheduler​

A flexible task scheduling system for managing periodic maintenance tasks across Ray workers.

Built-in Tasks:

  • ImageCleanupTask: Docker image cleanup using docuum with configurable disk threshold

Configuration Example:

scheduler:
  enabled: true
  worker_cache_ttl: 3600
  tasks:
    - task_class: "rock.admin.scheduler.tasks.image_cleanup_task.ImageCleanupTask"
      enabled: true
      interval_seconds: 3600
      params:
        threshold: "1T"

Extensibility: Create custom tasks by extending BaseTask and implementing the run_action(runtime: RemoteSandboxRuntime) method.

Entrypoints​

  • NEW: Batch sandbox status query API (POST /sandboxes/batch)

    • Efficiently retrieve status information for multiple sandboxes in a single request
    • Accepts BatchSandboxStatusRequest with sandbox_ids list
    • Returns BatchSandboxStatusResponse containing status details for all requested sandboxes
    • Supports up to configurable maximum count (default from batch_get_status_max_count)

  • NEW: Sandbox listing and filtering API (GET /sandboxes)

    • Query and filter sandboxes with flexible query parameters
    • Supports pagination via page and page_size parameters
    • Returns SandboxListResponse with items, total count, and has_more indicator
    • Enables filtering by sandbox attributes (e.g., deployment type, status, custom metadata)
    • Maximum page size configurable via batch_get_status_max_count

Sandbox​

  • Modify Sandbox log format: use iso 8601 format for timestamp, default time zone is Asia/Shanghai
  • Enrich SandboxInfo: add create_time, start_time, stop_time for metrics
  • Add Billing log when sandbox is closed

Enhancements​

Performance Optimizations​

  • Decouple the get_status API logic from Ray, reducing API latency from 1s+ to 100ms+; support dynamically toggling between the new and legacy logic.
  • Sink ray operations from sandbox manager into ray service.