Chevet Subagent Actor (chevet-subagent)

A low-cost, ultra-resilient LLM subagent actor designed for agent-to-agent MCP calling and autonomous workflows.

Features

• OpenAI-Compatible Chat Completions: Accepts standard OpenAI messages array, temperature, max_tokens, response_format, etc.
• Dual-Provider Architecture:
  • Primary: Chevet AI Qwen 3.5 9B (qwen3.5:9b), high speed and cost efficiency.
  • Secondary / Fallback: OpenRouter DeepSeek Flash (deepseek/deepseek-v4-flash-0731), resilient fallback on provider outage or rate limiting.
• Fail-closed usage pricing: A completed response costs $0.0025 on Free, $0.0023 on Starter, $0.0022 on Scale, or $0.0020 on Business; each started 5,000 generated-token block costs $0.0009. All tools are included in the normal tool allowance. No PPE event is charged when required AI/tool work, output validation, or full charge preflight fails.
• MCP-Ready: Can be called directly via Apify's MCP tools by parent agents (Cursor, Claude Desktop, Antigravity, AutoGen, CrewAI, etc.).
• Bounded Tooling: Optional web, data, document, calculation, restricted-WASM Python, and per-run workspace-file tools. Python has no shell, network, package installation, host filesystem, or environment-variable access.

Input Specification

Field	Type	Required	Description
messages	ChatMessage[]	Optional*	OpenAI-compatible messages: [{ role: 'user', content: '...' }]
prompt	string	Optional*	Single prompt string (used if messages is omitted).
system_prompt	string	No	System message to set persona/instructions.
enable_tools	boolean	No	Enables the documented bounded tools for this run (default false).
tool_profile	string	No	full (default) preserves every existing tool. Choose research, data, web_api, or workspace to send and permit only the focused tool set.
temperature	number	No	Sampling temperature (default 0.7).
max_tokens	integer	No	Per-turn maximum generated tokens (default 16,000, maximum 16,384). It is an allowance, not a guaranteed output length; unused tokens are not generated or billed.
max_total_output_tokens	integer	No	Aggregate generated-token ceiling across all model calls, 1–64,000 (default 64,000). The actor fails without PPE charging if this is reached before completion.
max_tool_uses	integer	No	Total tool-use limit, 0–4 (default 4), including restricted Python.
top_p	number	No	Nucleus sampling probability (default 1.0).
json_mode	boolean	No	Enforces JSON output (default false).
response_format	object	No	OpenAI format, e.g. {"type": "json_object"}.
model	string	No	Optional model override.
debug	boolean	No	Output telemetry including latency and fallback tracking.

* Either messages or prompt must be provided.

Pricing and limits

The completed-request event includes up to 24,000 initial input-context tokens and costs $0.0025 on Free, $0.0023 on Starter, $0.0022 on Scale, or $0.0020 on Business. Across a tool run, provider-reported input is guarded at 40,000 aggregate tokens. Input/prefill tokens are not separately billed. The actor totals only provider-reported completion_tokens across every model turn; reasoning-token telemetry is never billed separately. It charges $0.0009 for each started 5,000-token output block: 0 output tokens adds $0, 1–5,000 adds $0.0009, and 5,001–10,000 adds $0.0018. Normal tools and bridge orchestration have no separate PPE price; downstream Apify runs use the customer's configured Apify spend cap.

For example, a Free-tier completed request with 7,200 generated tokens costs $0.0025 + $0.0018 = $0.0043. The run preflights the complete event total against the user's Apify maximum using Apify's effective price for that run, then charges only after a valid final response exists. A failed provider path, aggregate input/output limit, or incomplete charge prevents publication of a success output.

The 64,000-token aggregate budget is spread over multiple calls. The default per-turn allowance is 16,000 tokens, so a normal response can use the full first three 5,000-token billing blocks when needed. It does not make every response 16,000 tokens long: providers report and bill only the tokens actually generated. Before every request, the runner dynamically reserves prompt and tool-schema space inside Qwen's 32,000-token context window. If a model reaches that shaped per-turn limit before finishing, the actor fails without publishing a paid success.

Restricted Python tool

When tools are enabled, the agent may call run_python for bounded computation, data transformation, and code tests. Python runs in a fresh WebAssembly worker with a 30-second maximum deadline and a memory limit. It accepts a restricted Python subset: no packages, shell commands, network, JavaScript interop, environment access, subprocesses, native filesystem access, or private-object introspection. Only math may be imported. Use assert statements for tests and print() for results. read_text(path) / write_text(path, content) and text-only open(path, 'r'|'w'|'a') only access the run workspace.

Tool profiles and bounded web retrieval

full keeps the complete tool surface for backwards compatibility. Focused profiles reduce prompt overhead and deny tools outside the selected set: research offers web research, PDFs, time, calculation, and workspace reading; data offers workspace/data processing and restricted Python; web_api offers safe HTTP and web retrieval; workspace offers file, search, structured-data, calculation, and restricted Python tools.

There is no deep-research meta-tool. The agent has a hard maximum of four total tool calls, so it must select sources deliberately. fetch_web_page returns a single clean excerpt of at most 3,000 characters, does not support offsets or pagination, and tool observations are compacted again before entering model history. This prevents a sequence of webpage calls from importing whole sites into the context window.

Example Input: Standard Chat

{
  "messages": [
    { "role": "system", "content": "You are a concise research subagent." },
    { "role": "user", "content": "Summarize the 3 key advantages of event-driven architectures." }
  ],
  "temperature": 0.3,
  "max_tokens": 500,
  "enable_tools": false
}

Example Input: Fast Single Prompt

{
  "prompt": "Classify this feedback as positive, neutral, or negative: 'The onboarding was fast but documentation lacks examples.'",
  "json_mode": true
}

Output Specification

The actor returns standard OpenAI-compatible completion objects to both the Dataset and Key-Value Store (OUTPUT record):

{
  "id": "chatcmpl-m7x9...",
  "object": "chat.completion",
  "created": 1725622400,
  "model": "qwen3.5:9b",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "..."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 42,
    "completion_tokens": 128,
    "total_tokens": 170
  },
  "provider_used": "chevet",
  "model_used": "qwen3.5:9b",
  "latency_ms": 384,
  "tools_enabled": false,
  "billing_summary": {
    "completed_request_count": 1,
    "llm_output_5k_count": 1,
    "total_completion_tokens": 128,
    "charges_completed": true
  }
}

Environment Variables

• CHEVET_API_KEY: API key for Chevet AI.
• CHEVET_BASE_URL: Base URL for Chevet AI (default: https://api.chevetai.com/v1).
• CHEVET_MODEL: Primary model ID (default: qwen3.5:9b).
• OPENROUTER_API_KEY: API key for OpenRouter backup.
• OPENROUTER_BASE_URL: Base URL for OpenRouter (default: https://openrouter.ai/api/v1).
• OPENROUTER_MODEL: Backup model ID (default: deepseek/deepseek-v4-flash-0731).
• STANDBY_API_KEY: Required bearer key for the optional Apify Standby HTTP API. Normal batch runs never use it.