Open-Athena

Evaluate a MarinFold contacts-v1 checkpoint with the fixed exp89 contact benchmark. Use for checkpoint scoring, R-precision/AUC requests, comparisons with structure baselines, or reproducing contact metrics on local, CUDA, or Iris TPU execution.

Write a Zephyr/Iris data-generation pipeline that finishes in minutes (not hours) and ships a clean corpus (not a silently-corrupted one). Use when authoring or reviewing an `exp<N>_data_*/cli.py` (or any `map_shard`-based job that fetches per-row inputs and emits parquet). Covers the per-row / per-shard / per-worker / per-job decisions for wall-clock plus the fail-loud default for data quality — including **neural / accelerator (TPU) tokenizers** when the model is the cost (batched bucketed inference, `ResourceConfig.with_tpu`, and parallel scans for sequential-recurrence models like SSMs/RNNs).

Monitor/babysit a job continuously and recover on failure. Use when asked to babysit, monitor, or watch a job, pipeline, workflow, or training run.

Start, monitor, and babysit Zephyr pipeline jobs on Iris. Use when launching a zephyr job, watching it run, or restarting after failures.

Post a results-analysis comment to an experiment issue after a completed run.

Compute $ cost estimate for an experiment notebook without dispatching to Modal.

Execute an experiment's notebook end-to-end and report results back to its GitHub issue. Use when the user points you at an experiment issue ("run

Answer a single "where is X implemented / how does Y work" question about the Marin codebase by searching the embeddings-backed code lane first (main + in-flight branches), and escalating to the broader mumwelt research skill only if a clear, citable answer doesn't emerge. Use for one-locus code lookups; for "what happened / why was X chosen" or anything spanning subsystems, use mumwelt directly.

Answer a broad, ambiguous, or multi-part question about Marin by decomposing it into parallel searches (across subagents) over the local marinmirror corpus and weekly summaries, then synthesizing a cited answer. Use for "give me the full picture of X", literature-review-style asks, retros, or anything one search query won't cover. For a single "where is X / how is Y implemented" code lookup, reach for the lighter mumwelt-code first.

Publish a finished research answer (Markdown with inline URL citations) to a private GitHub gist, rendered as a LaTeX-styled HTML document, and return a one-click htmlpreview.github.io link for easy sharing. Use after mumwelt to hand someone a clean, readable, link-preserving writeup.

Set up and launch a marin-launched DNA training experiment on the shared iris cluster. Each experiment is a self-contained directory with its OWN pyproject (marin packages in BASE deps, local lib as a git dep) — the pattern that makes iris/zephyr workers resolve the runtime. Use when creating a new exp<N>, launching training on the marin/iris cluster, packaging an experiment for marin, or debugging launch failures like "No module named 'zephyr'", "marin-iris client is too old", or a TPU job that won't start. NOT for snakemake pipelines (each snakemake/<pipeline>/ has its own README) or one-off analysis scripts (those go in scripts/).

Long-running, iterative investigation tracked in its own GitHub issue — the issue body is the living results doc, comments are the append-only iteration log, and every code reference is a commit-pinned permalink. Use when the user asks to investigate a question over multiple sessions and track the work in a dedicated issue. Common triggers include "let's investigate X and track it in an issue", "open an issue for this and we'll iterate", "dig into Y across sessions", "why is Z slow — track it", "explore this dataset and log findings as we go". Do NOT use for: running a pipeline by itself (even a snakemake/analysis run), single-session debugging or fixes, one-off questions, or regular feature PRs.

Upload a local file (screenshot, plot, data file) to a user-owned GitHub gist and get back a stable raw URL that renders inline in PR/issue comments and markdown documents. Use whenever you need to share an image or artifact in a GitHub comment, an analysis issue, a README, or any markdown doc and the file is NOT already committed to a repo. Common triggers include "attach this plot to the PR", "include the screenshot in the issue", "embed this figure in the analysis", or "I need a link for this image".

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

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.

Debug code bugs or Iris/Zephyr/TPU infrastructure faults with a structured debug log.

Rearrange `.discord-agent` submodule commits so that general/reusable work lands on upstream `discord-agent@main`, and Marin-specific customizations sit as a single squashed commit on top of that on `marin`. Use after adding commits directly to `marin` that should have gone upstream.

Edit the active plan in a live plantt tab (hosted or localhost) by natural-language request. Starts a localhost relay, drives the open plantt tab through it, and shuts the relay down when finished. Use when the user wants to change a plan they have open in plantt — move tasks, add workstreams/milestones, adjust capacity, etc.