factor-mining-demo

Factor Mining Demo

Use this skill for the Factor Mining Demo product flow. The agent writes or locates one local plugin.py. Factor Mining validates, stores, backtests, and returns workflow, job, artifact, and factor result data through the bundled Factor Mining Demo MCP tools.

If the MCP tools are unavailable, stop and tell the user the Factor Mining Demo plugin tools are not loaded. Do not use raw HTTP calls or local script execution as a fallback.

Some hosts display bundled MCP tool names with a provider prefix, such as fm-demo__factor_mining_demo_status. Treat those as the same Factor Mining Demo MCP tools.

MCP Workflow

Use these tools for product actions:

• factor_mining_demo_status
• factor_mining_demo_setup_browser
• factor_mining_demo_list_public_tasks
• factor_mining_demo_create_task_session
• factor_mining_demo_create_custom_session
• factor_mining_demo_parse_plugin_metadata
• factor_mining_demo_request_dedup_context
• factor_mining_demo_upload_backtest_wait
• factor_mining_demo_resume_run
• factor_mining_demo_get_workflow
• factor_mining_demo_get_job
• factor_mining_demo_get_artifact
• factor_mining_demo_clear_config

Start with factor_mining_demo_status. Continue only when the local direct Agent API Key validates through Factor Mining /health and /agent/status. If status reports setup is required, call factor_mining_demo_setup_browser and tell the user to enter the key in the local browser page. Never ask the user to paste the key into chat.

Determine whether the user is starting from a public task or a custom idea. For a public task flow, call factor_mining_demo_list_public_tasks, show concise choices, and ask the user to pick one unless they explicitly ask the agent to choose. Create the session with factor_mining_demo_create_task_session.

For a custom idea, create a direct task_payload and call factor_mining_demo_create_custom_session. The payload must include task_id, title, category, description, non-empty allowed_data, and fwd_period. allowed_data must include every input column the generated plugin.py needs, such as close, volume, funding_rate_close, or open_interest_close. Include useful hints, economic mechanisms, regime considerations, risk sources, and target behavior when they are available.

When you have a draft description and formula, call factor_mining_demo_request_dedup_context and use the returned similar-factor guidance to avoid near-duplicates. This context informs local revision only.

Write or locate one plugin.py, then call factor_mining_demo_parse_plugin_metadata. The parser is static and must not import or execute generated code. When the metadata is valid and the user is ready to submit, call factor_mining_demo_upload_backtest_wait. Use fwd_period=7 when neither the task nor user specifies a horizon. The tool saves default_factor_card.json and image artifacts referenced by the factor card into factor_mining_demo_artifacts/ beside plugin.py unless an explicit output_dir is provided. Use the returned artifact paths and original saved images; do not recreate backend plots from JSON when those image artifacts are available. The MCP tool also tries the standard backtest image artifacts default_cs_nav_curves.png, default_cs_profile_4panel.png, and default_group_return_plot.png even when the factor card does not list them.

After upload, inspect the returned ok, status, terminal_status, failures, sanitized job statuses, artifact availability, and factor-card metrics. Fetch the default factor card through factor_mining_demo_get_artifact when it is not already present and the job state says an artifact should be available. Use factor_mining_demo_resume_run after an interrupted session.

When artifact.image_artifacts is present, show those saved original images in the conversation when the host supports image display. Do not print local absolute paths as text; use concise filenames in text summaries. If no images are available, say whether the standard image artifacts were unavailable.

Never show backend job IDs, presigned URLs, raw credentials, bearer tokens, or plugin.py source in user-facing summaries.

plugin.py Contract

Use this minimum shape when the user has not supplied an existing plugin. The metadata values must be static top-level literals so Factor Mining can parse them without executing source code.

from typing import Any, Dict

import pandas as pd

FACTOR_TYPE = "snake_case_unique_factor_type"
FACTOR_NAME = "human_readable_factor_name"
FACTOR_DEFAULT_PARAMS = {"window": 7}

FACTOR_SECTIONS = {
    "__FACTOR_DESCRIPTION__": "Trailing close-to-close momentum.",
    "__FACTOR_FORMULA__": "close / close[window bars ago] - 1",
    "__FACTOR_TYPE__": FACTOR_TYPE,
    "__FACTOR_PARAM_FIELDS__": "        private int _window;\n",
    "__FACTOR_INIT__": '            _window = GetIntParameter("window", 7);\n',
    "__FACTOR_LOG__": '            Log($"[INIT] window={_window}");\n',
    "__PRICE_WINDOW_EXPR__": "_window + 1",
    "__EXTRA_BUF_FIELDS__": "",
    "__EXTRA_BUF_ENQUEUE__": "",
    "__EXTRA_BUF_DEQUEUE__": "",
    "__EXTRA_BUF_TOARRAY__": "",
    "__FACTOR_COMPUTE_BODY__": """
            var n = prices.Length;
            if (n < _window + 1) return false;
            var past = prices[n - _window - 1];
            if (past == 0) return false;
            rawSignal = prices[n - 1] / past - 1.0;
            return true;
""",
}


def build_signal(close: pd.DataFrame, params: Dict[str, Any], **data: Any) -> pd.DataFrame:
    window = int(params.get("window", FACTOR_DEFAULT_PARAMS["window"]))
    signal = close.pct_change(window)
    return signal.reindex_like(close)

Keep Python build_signal and FACTOR_SECTIONS compute logic aligned. Return a pd.DataFrame aligned with close, use only current and historical data, and keep all data columns within the session allowed_data contract.

Security

• Use only the bundled Factor Mining Demo MCP tools for formal product actions.
• Never ask for OpenAI API keys, Codex auth files, BYOK secrets, frontend user credentials, or raw Factor Mining credentials.
• Never print, persist in logs, or summarize full credential values.
• Do not call hosted generation endpoints; local-agent mode generates code inside the active coding agent.
• Do not call frontend-user credential management, BYOK, Codex profile, task publishing, or generic URL/API surfaces.
• Do not import, exec, eval, or otherwise execute generated plugin.py.
• Do not print generated plugin.py source in summaries.
• Treat downstream IDs, presigned URLs, and service metadata as internal.
• Artifact 404 and 410 responses mean unavailable. Authentication, authorization, network, malformed response, and server errors must fail clearly with redacted messages.