ra-compute

RA Compute

Goal: help an economist finish a research run with durable outputs. Keep the conversation about scripts, runs, logs, artifacts, cost, and next steps.

This public skill is a thin router. Product logic belongs in mecon and the RA Compute server, not in this markdown. Do not expose private platform internals.

Triggers

Use this skill when the user mentions:

• RA Compute, mecon submit, mecon resources, mecon status, mecon logs, mecon watch, mecon download, mecon precompile;
• CPU jobs, GPU jobs, resource profile, cpu-4c-16gb, L4, A100, H100;
• long empirical scripts, structural estimation, JAX/XLA/JIT, BLP, DDC, SMM, bootstrap, placebo, parameter sweep, Monte Carlo;
• large Parquet, DuckDB, too large for laptop, first-pass summaries;
• failed/stalled/timed_out/cancelling compute runs, missing outputs, stale progress, OOM, low GPU utilization, workspace stale, missing API key.

Do not use this skill for dataset discovery/download unless the user is asking to execute a compute job. Use ra-data for catalog/query/download-only work. Do not use it for experiment lifecycle work. Use ra-experiments there.

First Actions

Run the helper first when possible:

skills/ra-compute/scripts/compute-readiness.sh

The helper only calls supported mecon commands. It does not call raw API endpoints and does not submit paid work.

If the helper is unavailable, run these checks manually:

mecon --version
mecon submit --help
mecon resources
mecon doctor

mecon --version must be 0.7.2 or newer for CPU/GPU --resource support.

Server-Owned Flow

Use server-backed commands as the source of truth:

• capability and pricing: mecon resources
• account/workspace readiness: mecon doctor
• submission: mecon submit . --resource <profile>
• live state: mecon status <job_id>, mecon watch <job_id>
• logs/artifacts: mecon logs <job_id>, mecon download <job_id>
• post-run advice: mecon profile <job_id>

Do not maintain pricing tables, raw API payloads, provider details, or hidden scheduling heuristics in the skill.

Load The Right Guide

Read only the guide needed for the task:

• Server/product boundary: references/server-owned-workflows.md
• New job or resource choice: references/resource-selection.md
• Workspace setup and submission: references/submission-flow.md
• Large Parquet / first-pass data work: references/large-parquet.md
• Failed, stalled, timed_out, cancelling, OOM, low utilization: references/diagnosis.md
• Client-facing explanation: references/customer-language.md

Safe Commands

Supported customer-facing path:

mecon doctor
mecon resources
mecon sync
mecon submit . --resource cpu-4c-16gb --timeout 3600
mecon submit . --resource l4 --timeout 3600
mecon watch <job_id>
mecon status <job_id>
mecon logs <job_id>
mecon profile <job_id>
mecon download <job_id>

Use mecon open billing | jobs | keys | dashboard | job <id> when the human must act in the browser.

Hard Rules

• Do not submit CPU work with raw API JSON or helper scripts. Use mecon submit . --resource <profile>.
• Do not copy server-returned pricing or resource policy into the skill as a static table.
• Do not recommend a bigger GPU without logs/profile evidence.
• Do not call a job hung from stale progress text alone. Check status, logs, activity, and artifacts.
• Do not rerun before downloading existing terminal outputs.
• Do not expose Modal/R2/callback internals in client-facing text unless the user explicitly asks for platform details.
• A task is not done until there is a terminal status, a verified blocker, or the user explicitly asked to stop after submission/status.

Completion Output

Report only what matters:

• job id;
• resource profile;
• current or terminal status;
• artifact/download state;
• next concrete command or human action.