Compatibility and resources
This public documentation describes the current source capabilities of envloop 0.1.0, without presenting internal target designs as released features. Interfaces checked on 2026-10-11.
Supported capabilities
| Capability | Status |
|---|---|
el task / client.tasks | Directory / inline two-file tasks, status, waiting, results, snapshots, and stopping |
el job / client.jobs | Dataset selection, attempts, listing, status, waiting, aggregate results, and stopping |
| Runner catalog submissions | runner + Dataset, profile_id, job_id, tag, queue_id, priority, command; single Tasks support task_id; requires a matching APIServer catalog |
| Remote authentication | Bearer HTTPS; trusted loopback HTTP development connections |
Client.local() | Optional EnvPlatform backend; local execution of trusted scripts |
el profile / client.profiles | Local named Profile registry; HTTP management is unavailable |
| Submit by EnvHub Task name / URL | The current user directory submission entrypoint does not support this form |
el train submit/status | Source provides HTTP submission, status, and waiting after submission for independent Train Jobs; requires a remote API and uses train_job_id; this does not verify real GPU training |
client.train, Tinker integration | The current Client has no training entrypoint; envloop has not yet integrated the Tinker SDK directly |
el task logs/cancel/wait | Unavailable; use result, stop, submit --wait, or SDK wait |
CLI executor / provider options do not imply that every backend is available; execution depends on platform configuration and backend support. The local hello example verifies scripts, results, and artifacts, not real Harbor benchmark scores.
Version basis
- envloop source snapshot: CLI parser, Client, and HTTP configuration.
- EnvPlatform source snapshot: matching Task / Job / Profile and dashboard implementations.
- The lite dataset requires the platform to include EnvHub
4bb73a27386ce91b9afc603206b2f2d91b209b7dor a later version. - New Runner catalog parameters require the platform's matching
feat/runner-catalogimplementation; the client forwards through/v1/jobsand/v1/trials; wire verification does not verify Docker / public execution.
Source snapshots document the basis for these pages, not proof that any public deployment has updated to those versions. Public Keys, dataset availability, and cloud execution need separate verification on the target deployment. command / runner.toml requires this round's matching platform implementation. Task / Job / Trial POST request bodies are limited to 512 KiB; other ordinary APIs to 64 KiB; scripts themselves remain limited to 65536 UTF-8 bytes.
Resources
- envloop repository: client source and executable examples.
- EnvPlatform repository: execution platform.
- EnvHub repository: Dataset and Task definitions.
- Report a client issue: usage questions and client feedback.
- Report a platform issue: execution, permissions, and results issues.
The information structure follows the Get started / Core concepts / SDKs and CLI / Resources navigation in the OpenAI Developer quickstart, accessed 2026-10-08; EnvLoop examples and interfaces follow its own implementation.