Research & Analysis Workflows

Turn one literature note into an experiment proposal. Reads the note + frontmatter, extracts the core claim or technique, and proposes an experiment that replicates, tests, or extends that claim in the context of the active project (using the NOTES.md tail and project README.md for framing). Output format matches /propose exactly. Appends the new proposal's slug to the literature note's relatedexperiments: frontmatter field. Persists per agency.md's Confirmation principle (confirm under standard, auto-advance under max).

Promote a project echo to global scope with domain tagging. Use when a learning is valuable across multiple projects. Copies selected entries from project .rune/echoes/ to global echoes with domain classification and source provenance.

/fetch-paper . Downloads the full text into raw/papers/ .pdf (or .md for HTML; GitHub repos land in raw/repos/ .md), derives a Better-BibTeX-style citekey from authors + year, and chains into /ingest to produce the literature note. Handles arxiv abs/pdf URLs, direct PDFs, and GitHub repo READMEs. Does not re-download if the target already exists.

Run final-scoring pass on an MLE-bench experiment. Inferences the trained model on the held-out test/ directory (the only skill allowed to read test/), produces submission.csv in MLE-bench format, invokes mlebench grade-sample, and writes finalmetrics.json with grader output + provenance (task ID, grader version, timestamp, MLE-bench git SHA). Does not mutate metrics.json. Appends one FINALSCORE line to meta/iterationlog.md for /lint to grep.

Scaffold an MLE-bench task into the active project. Resolves the task's prepared data from the local MLE-bench cache, creates a dated experiment folder with train/ and test/ symlinks, writes splits.yaml (stratified / temporal / random strategy matched to the task), drafts a seed proposal under experiments/proposals/, and writes an experiment-level CLAUDE.md with task-specific context. Does not run /implement or /iterate — hands off to the user.

Scaffold an experiment folder at experiments/YYYY-MM-DD- / with README.md, notes.qmd, config.yaml, results/, log.md, metrics.json. Pre-fills frontmatter from meta/templates/experiment.md. If the project opts into HCE (has a splits.yaml or test/ directory), also copies the split spec — otherwise splits.yaml is skipped.

Promote a ripe cluster of related concepts into a Map of Content at mocs/ .md. Detects clusters of >=5 related concepts on a theme that are not already substantially covered by an existing MoC, then synthesizes the MoC (framing + layered organization + concept wikilinks) from the concept files. In standard repos it proposes and waits; in agency: max repos it creates, commits, and pushes autonomously. This is the "act on it" counterpart to /lint's MoC-candidate detection.

Scan the active project for @import directives pointing at agentic-research/concepts/ and append idempotent usedby: back-references to each target concept file. Standalone skill factored out of /ingest so it can run independent of any raw/ file. /ingest and /new-project both call it.