datarobot-skills-utils

Runtime utilities for running LLM calls through the DataRobot CLI's dr opencode runtime, used by the DataRobot agent skills in datarobot-agent-skills.

Beta. This package exists to support the DataRobot agent skills. Its API may change between minor versions without notice, and it carries no official support. Pin an exact version if you depend on it.

Prerequisites

• Python 3.11 or later. The package has no third-party dependencies.
• The DataRobot CLI (dr) on your PATH, logged in with dr auth login. Every LLM call goes through dr opencode, so the CLI's login and model access apply.

Installation

pip install datarobot-skills-utils

Example

Start one private dr opencode serve for a batch of calls, send a prompt, and read the token usage:

from datarobot_skills_utils.opencode import (
    OpenCodeServer,
    UsageMeter,
    dr_available,
    run_worker,
)

model = "datarobot/anthropic/claude-sonnet-4-6"  # any model listed by `dr opencode models`

if not dr_available():
    raise SystemExit("The DataRobot CLI (dr) is not on PATH.")

usage = UsageMeter(model)
with OpenCodeServer(models=[model], reasoning="max") as server:
    text, meta = run_worker(
        "Name three risks of hard-coding API keys.",
        model,
        server_url=server.url,
        cwd=server.workdir,
    )
    usage.record(meta)

print(text)
print(usage.snapshot()["total"])  # calls, input/output/reasoning/cache tokens, cost

API

Everything is exported from datarobot_skills_utils.opencode.

• OpenCodeServer(models=(), reasoning=None): a private dr opencode serve on a free localhost port, in an isolated temporary working directory. Use it as a context manager; url and workdir are set while it runs. Passing models and reasoning ("max", "off", or a specific effort) sets the reasoning effort and output-token limits for those models without editing your opencode.json.
• run_worker(message, model, server_url=None, cwd=None, timeout=120, attempts=2, ...): one dr opencode run completion, attached to a server or run standalone. Returns (text, token_meta), and retries an empty response.
• UsageMeter(model, reasoning_effort=None): thread-safe token and cost accounting across a run. Set phase before a batch, call record(meta) after each call, and read totals with snapshot().
• parse_events(stdout) and strip_code_fences(text): parse dr opencode run --format json output into text plus token accounting, and strip Markdown fences from JSON replies.
• dr_available(), worker_env(...), resolve_effort(...), max_reasoning_effort(...), token_limits(...): helpers for checking the CLI and resolving per-model reasoning effort and output limits.

License

Apache License 2.0. See LICENSE.