marin-community/marin

scrub-experiment-issue-tldrs

Scheduled scrub: TL;DR blocks on experiment issues.

manage-hero-run

Launch, monitor, hand off, resume, rollback, or babysit expensive Marin production. Typically >=1e22 model flops.

commit

Lint, run the pre-PR checks, commit, push, and author or update the branch's pull request in the required plain-text format. Use when committing, pushing, or creating/updating a PR.

refresh-fork

Rebase one Marin fork onto newer upstream per its config/external/migration.toml descriptor, run the fork's e2e, and open the Marin PR or file a blocker issue.

noslop

Clean a branch with strict gates for low-value tests and agentic prose. Use when asked to deslop, simplify, clean up, make a diff minimal, review test quality, tighten comments/docs/PR text, or remove AI-writing patterns such as rhetorical "X, not Y"…

writing-style

Marin house writing style. Use when drafting or revising Marin-authored prose, including commit messages and GitHub PR, issue, or comment text.

write-tests

Write or revise Marin tests with an emphasis on behavior, regression coverage, pytest style, and avoiding "slop tests." Use when adding tests, fixing failing tests, reviewing test quality, or deciding what test would catch a bug.

ab-test-zephyr

Run a Zephyr control and treatment on pre-normalized data and compare per-stage Finelog CPU, elapsed-time, and memory stats. Use for ad hoc comparisons and PR performance gates; add named treatments only when requested.

consult-echo

Search and cite Marin's Echo activity and wiki, then capture incident records or reusable shared knowledge. Use when looking for prior discussions, decisions, workflows, incident patterns, exact errors, tags, or identifiers; when prior context could shorten…

add-pallas-kernel

Add, modify, or autotune a TPU/GPU Pallas kernel.

write-ops-log

Publish a tagged postmortem incident record to Echo. Use after an infrastructure or durable debugging session, then link the canonical Echo URL from the associated PR or issue.

write-design-doc

Produce a 1-page design proposal for an explicit design task or a design-level change identified by another change workflow. Do not use to answer or evaluate an idea inline.

fix-issue

End-to-end workflow to fix a GitHub issue in marin-community/marin.

run-ferries

Launch, monitor, and seal Marin canary and daily ferry runs.

deploy-iris-controllers

Deploy the Iris controller to one cluster or across the fleet, with automatic progress through passed gates. Use when restarting, rolling out, or rolling back a controller.

add-grant

Add a GCP IAM or IAP user grant to the marin-iac Pulumi stacks, encrypting human principals. Use when someone needs access to a project role, bucket/secret/service account, or an IAP-gated web service (evaldash, grafana), whether requested locally or on a…

review-grant

Review a marin-iac user-grant PR by decrypting the changed principals into the real emails and grants, confirm with the user, then approve, merge, and drive pulumi up. Use when reviewing a PR that edits iam_data.yaml or a service's IAP viewers.

reserve-gpu

Reserve Iris-backed CoreWeave H100 or GB200 nodes with dev_gpu.py. Use for interactive GPU debugging, multi-node tests, or reconnecting to a dev GPU session.

query-inference-metrics

Investigate a vLLM serve's performance (throughput, TTFT/TPOT latency, queue depth, KV-cache usage) from durable Finelog telemetry with SQL. Use when asked how an inference or eval serve performed, or why it was slow. For kernel-level JAX profiling, use…

debug

Debug code bugs or Iris/Zephyr/TPU infrastructure faults with a structured incident record.

recover-stuck-k8s-pod

Diagnose and safely recover stuck terminating Kubernetes pods on Marin CoreWeave clusters, especially node-bound GPU pods. Use for deletion hangs, suspected uninterruptible GPU/NCCL waits, node cordoning or reboot decisions, and force-deletion requests.

archive-experiments

Retire legacy experiment scripts behind a dated archive tag.

fix-docs

De-rot markdown docs in lib/iris, lib/zephyr, and lib/fray.

organize-experiments

Curate the experiment report index at docs/reports/index.md.

profile-training

Profile JAX training and analyze hotspots. Use when profiling or optimizing training throughput.

run-research

Multi-session research workflow: compose logbooks, experiment issues, documentation updates, and snapshot discipline for long-running investigations.

task-snapshot

Create stable commit or tag snapshots for task milestones, experiment results, and reproducible artifact links.

file-issue

File a GitHub issue for a bug or improvement found this session.

review-pr

Multi-agent correctness review of a pull request.

babysit-job

Monitor an Iris job and recover it on failure. Use when asked to babysit or watch a job or run.

triage-canary

Triage a failed canary ferry run (CI-invoked).

lint-review

In CI, run the infra/lint catalog review over a PR.

add-dataset

Register or inspect a Hugging Face dataset for Marin pipelines.

background-research

Forage prior work before or during Marin research threads: search internal Marin artifacts and external literature/code; produce a cited brief with negative results and ranked experiment hypotheses.

task-logbook

Maintain append-only task/research logbooks and publish the important parts upward into coordinating GitHub issues.

update-docs

Update documentation, runbooks, and reusable task guidance when implementation work or experiments change behavior or operational practice.

wandb-reporting

Use W&B runs, reports, and artifacts consistently for experiments, benchmarks, and task results with dense numeric output.

change-grug

Modify or upstream a Grug/Grugformer experiment variant.

scrub-docs-code-parity

Scheduled scrub: docs and code parity.

scrub-reflection-self-improvement

Scheduled scrub: repository self-improvement.