Menú
Run numeric or ML code through isanagent execution tools without flooding context—artifacts, incremental runs, sandbox paths.
Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.
Basado en la clasificación ocupacional SOC
MaxEvolve Triton/PyTorch to JAX/Pallas porting with GpuToJax 12-step pipeline, interpret=True validation, MAP-Elites evolution, and Colab/SSH hardware profiling.
Manage Google Colab sessions and execute code on remote Colab VMs via the colab-cli.
Use when you need to generate a synthetic dataset. You can use Afterimage Python package to create, conversational, tool-calling, multi-choice questions, or other types of structured output datasets for SFT, or preference data for DPO.
Schedule deferred or recurring work via the isanagent cron tool—add/remove jobs with correct runtime context fields.
Configure long wall clocks and use execution_run_background + job polling for ML or hours-long runs without blocking the agent loop.
Turn papers and docs into a concrete training or evaluation recipe (datasets, metrics, hyperparameters, fine-tuning code).
| name | execution-research |
| description | Run numeric or ML code through isanagent execution tools without flooding context—artifacts, incremental runs, sandbox paths. |
| requires | {"bins":[],"env":[]} |
| always | false |
Use this playbook whenever you are writing or running analysis code, plots, tables, or benchmarks inside the workspace execution harness.
execution_env_info once. Note default_provider, max_output_bytes, max_wall_secs, default_run_timeout_secs, timeout_policy, and artifact caps. If execution is disabled in config.toml ([harness.execution] enabled = false), stop and tell the user to turn it back on—do not invent a workaround.max_output_bytes (with a floor). Large prints truncate; design outputs accordingly.execution_session_create — set language (python, shell, or omit for default Python). Keep one active run or background job per session; wait for each to finish before starting another on the same session_id.execution_run — pass session_id, source code, and timeout_secs within max_wall_secs. Add description for any non-trivial run so the terminal strip and JSONL audits show intent. Prefer small, verifiable steps (import → load → transform → plot) instead of one giant cell. For long wall clocks (many minutes or hours), prefer execution_run_background (with description) and poll execution_job_status / execution_job_result (see skill long-running-execution) so the agent thread is not blocked for the entire duration.execution_artifact_list — when you need filenames or sizes under .execution_artifacts/<session_id>/ (Jupyter materialized payloads, or files your code wrote there on local runs).glob_files / search_text — inspect text logs, CSV snippets, or small previews by path; never pull multi‑MB content into the reply.execution_session_close with the same session_id when the analysis branch is done so slots are freed.display_data / similar payloads become files under sandbox_dir/.execution_artifacts/<session_id>/<run_uuid>/. The tool response includes attachments (sandbox-relative paths, MIME). Treat those paths as the source of truth; summarize, do not inline blobs.execution_run until timeout, cancel, cwd change, or close). If you need a clean interpreter each time, that is a host config choice (local_python_mode = subprocess in config.toml)—you cannot toggle it from a tool. For plots/tables, still savefig / write files under .execution_artifacts/... or another sandbox path you resolve with normal file tools so results stay on disk.execution_run payloads in a tight loop when results do not change—isanagent may inject a doom-loop corrective user message; change code, parameters, or diagnostics instead.execution_session_close after a long exploratory branch.Use execution_cancel on that session_id, then decide whether to retry with a shorter timeout, smaller data, or different approach.