marin-community
- 46 skills
- 0 followers
- 6 hours ago last updated
- ▌ Task Logbook · marin-communityMaintain a cross-project task or research log in Echo. Use for multi-step investigations, experiments, long-running implementation work, or any task whose milestones must survive the current agent session without adding a repository logbook file.
- ▌ Write Design Doc · marin-communityPublish or revise a concise cross-project design proposal in Echo. Use for an explicit design task or a design-level change whose context, decision, interfaces, tradeoffs, and rollout need review without adding a repository design document.
- ▌ Rl Supervisor Init · marin-communityInitialize and run a focused RL operations session across Iris and explicitly selected supported backends: inventory jobs, dispatch diagnosis, perform authorized lifecycle actions, and maintain a reproducible handoff for an explicit list of experiment directories. This is a no-commit role that never edits repository code or configuration and never opens or updates pull requests.
- ▌ Rl Agentic Job Cleanup · marin-communityPreserve and, when authorized, publish a terminal agentic Iris RL run: stop pending retries, select and validate a checkpoint, export model weights, preserve metrics and traces, verify destinations, and only then reclaim storage. Use for runs with an agent harness or per-trial artifacts; use rl-standard-job-cleanup for parquet-only RL.
- ▌ Rl Agentic Launch Iris · marin-communityValidate, submit, and observe agentic MarinSkyRL training on Iris when the configuration uses Harbor, Daytona, terminal-bench, or another sandboxed agent harness.
- ▌ Build Gpu Rl Image Iris · marin-communityBuild and validate compatibility MarinSkyRL GPU-RL containers when the user explicitly requests one. Maintained Iris training and utilities use frozen environments instead.
- ▌ Rl Job Health Deep Dive · marin-communityDiagnose one live or suspect RL job on Iris or an explicitly selected supported backend and return an evidence-backed KILL, NO-KILL, or ERROR recommendation. Use when ordinary state and metric checks cannot distinguish progress from a silent stall, or when validating a new model, geometry, configuration, or image. This is a read-only repository role: it never edits code, commits, pushes, opens PRs, or kills jobs.
- ▌ Rl Standard Job Cleanup · marin-communityPreserve, validate, and hand off a terminal standard MarinSkyRL run backed by dataset rows and programmatic rewards, without an agent harness or per-trial artifacts. Use after completion or failure, or before an authorized cancellation. Use rl-agentic-job-cleanup for agent trajectories.
- ▌ Rl Standard Launch Iris · marin-communityValidate, submit, and observe standard MarinSkyRL training on Iris for dataset-backed rewards without Harbor, Daytona, terminal-bench, or another agent harness.
- ▌ Experiment Status Artifact · marin-communityRender a fleet of experiment runs as a compact, operational status artifact. Use when a user needs a dashboard or overview across multiple runs, arms, datasets, or configurations rather than a prose report.
- ▌ Debug · marin-communityDiagnose a stated code, JAX, Marin, Iris, Zephyr, or TPU fault or startup/performance regression; do not activate for ordinary implementation or optimization without a symptom.
- ▌ Commit · marin-communityUse only when a change is ready to commit, push, or open or update its pull request; do not activate during implementation.
- ▌ Noslop · marin-community bundleDeslop, simplify, or review low-value tests and prose only when explicitly requested for a branch or diff.
- ▌ Research · marin-community bundleResearch an explicitly requested Marin question or coordinate a multi-session experiment; find prior work, assess evidence, and record results in the task's existing durable surface.
- ▌ Use Iris · marin-community bundleUse Iris to submit, inspect, debug, monitor, or recover jobs and tasks; diagnose scheduling and federation; deploy controllers; or reserve dev GPUs and TPUs. Use for ordinary Iris operations, including babysitting a job, controller rollouts, accelerator sessions, and stuck CoreWeave pods. Use a narrower production-run or Zephyr skill when that workflow is explicitly requested.
- ▌ Add Grant · marin-communityImplement a fully specified request for a GCP IAM resource grant or Cloud Run IAP viewer in marin-iac.
- ▌ Fix Issue · marin-communityImplement and land a fix for an explicitly identified marin-community/marin GitHub issue.
- ▌ Review Pr · marin-communityReview an explicitly identified Marin pull request against the repository's correctness and maintainability expectations.
- ▌ Scan Logs · marin-communityUse scripts/logscan.py only when explicitly requested to scan an oversized log or to perform Gemini-backed log analysis.
- ▌ File Issue · marin-communityFile a GitHub issue only when explicitly requested or delegated by another selected workflow; do not create issues merely because a problem was found.
- ▌ Add Dataset · marin-communityRegister a named Hugging Face dataset for Marin by inspecting its schema and adding the appropriate experiments/datasets module.
- ▌ Change Grug · marin-communityModify or upstream behavior in a named Grug/Grugformer experiment variant.
- ▌ Lint Review · marin-communityRun the read-only infra/lint PR reporter only when invoked by CI or explicitly requested; do not select it for the commit workflow's fix-and-respond review.
- ▌ Run Ferries · marin-communityLaunch, monitor, or seal a Marin canary or daily ferry only when explicitly requested.
- ▌ Update Docs · marin-communityUpdate durable docs when explicitly requested or when current work changes behavior, exposes stale instructions, establishes an operational pattern, or yields reusable research guidance.
- ▌ Write Tests · marin-communityAdd, revise, or review Marin tests for an explicit behavior change, regression, or test-quality request.
- ▌ Consult Echo · marin-community bundleSearch or cite Echo when Marin's prior-work policy applies, another selected workflow requires it, or the user explicitly asks; do not substitute it for local code search.
- ▌ Refresh Fork · marin-community bundleRefresh a named Marin external fork pin onto a newer upstream base using its configured descriptor and required end-to-end test.
- ▌ Review Grant · marin-communityReview an explicitly identified marin-iac grant PR that edits IAM data or a deploy-target module, confirm its decrypted principals and roles, then apply only the confirmed grant.
- ▌ Marina Applet · marin-community bundleBuild, validate, publish, update, inspect, query, roll back, or archive a dynamic Marina applet. Use for Artifact+-style web apps that should run behind Marina authentication, carry small bundled assets or data, use an applet-owned Postgres schema, or add a trusted Python ASGI backend without a Marina image deployment.
- ▌ Query Finelog · marin-communityQuery Finelog logs and telemetry for Iris tasks, workers, profiles, training, vLLM, and cross-cluster forwarding. Use for schema discovery, SQL, memory or CPU summaries, hub-versus-regional comparisons, counter semantics, and query-performance diagnosis.
- ▌ Triage Canary · marin-communityTriage a failed canary ferry run only when invoked by CI with the required CANARY context.
- ▌ Write Ops Log · marin-communityPublish a tagged Echo record after a live infrastructure incident, then link its canonical URL from the associated pull request or issue.
- ▌ Writing Style · marin-community bundleApply Marin house style when producing or revising commit, pull-request, issue, comment, documentation, report, or blog prose.
- ▌ Ab Test Zephyr · marin-communityRun an explicitly requested Zephyr control/treatment benchmark on the same pre-normalized sample and compare Finelog stage metrics.
- ▌ Babysit Zephyr · marin-communityLaunch or continuously monitor a specified Zephyr pipeline on Iris only when asked to babysit it; restart only after explicit approval.
- ▌ Manage Hero Run · marin-communityLaunch, resume, monitor, hand off, or seal an explicitly requested production-critical Marin run, normally at least 1e22 FLOPs.
- ▌ Profile Training · marin-communityProfile a named JAX, Levanter, or Marin run, or investigate a measured startup, compilation, initialization, or throughput bottleneck.
- ▌ Add Pallas Kernel · marin-community bundleAdd or change a named Pallas/Mosaic kernel, including its reference implementation, correctness tests, wrapper, or requested tuning.
- ▌ Trace Pulumi Diff · marin-community bundleRun a read-only preview for a specified Marin infra/pulumi stack and trace each pending resource change to merged pull requests since its latest successful update when that update records a clean Git checkout. Use when a stack has unapplied changes or drift and someone needs to know which PRs explain the preview.
- ▌ Deploy Hero Change · marin-communityDeploy a significant code change (backend, kernel, optimizer, data path) to the live hero run: relaunch it under a new run id from a permanent checkpoint, compare against the old run over a trial window, and roll back if the gate fails; use only when the user asks to deploy a change to the hero.
- ▌ Archive Experiments · marin-communityArchive explicitly named legacy experiment files by publishing a dated tag and canonical issue breadcrumbs.
- ▌ Organize Experiments · marin-communityHarvest experiment issue reports and curate docs/reports/index.md only when explicitly requested.
- ▌ Scrub Docs Code Parity · marin-communityRun the docs/code-parity scrub only from its scheduler or an explicit request for that scrub.
- ▌ Scrub Experiment Issue Tldrs · marin-communityRun the experiment-issue TL;DR scrub only from its scheduler or an explicit request for that scrub.
- ▌ Scrub Reflection Self Improvement · marin-communityRun the repository self-improvement scrub only from its scheduler or an explicit request for that scrub.