Setup
setupInput: the current research directory. The agent checks uv and installs or locates the harness. frontier setup creates the active frontier.yaml and output directory without replacing an existing config. The agent asks for a path only when the workspace is unclear. Output: the config path, output directory, installation method, and any missing optional dependencies. No paper search occurs.
Config
config
config Track 2026 papers on agent execution harnessesInput: a natural-language change request, or no argument. With no argument, the agent reads the current YAML and asks what to change. It resolves ambiguous author or venue IDs rather than guessing. It writes a candidate file under .frontier/config-drafts/, validates it, shows a field-level diff, and applies it to frontier.yaml only after the proposed settings are clear. Output: the active YAML path and a summary of changed fields. Config does not run Sync.
Discover
discoverInput: the active YAML; normally no extra text. The agent explicitly runs Sync and Rank and opens the resulting Discovery page. For stable topic IDs, the page compares the current query scope with a previous run: newly published, newly observed, newly included by a changed scope, Gate changes, and rank changes are distinct. Provider result caps and alphaXiv top-k coverage are marked as incomplete. Select papers in the local page; its server writes 02-rank/briefing-selection.json in the same Run. Static HTML cannot write to itself and offers a downloadable selection instead. Output: a Run path, Discovery page, counts, and coverage caveats.
Briefing
briefingInput: the selected candidate IDs from the current Run, plus the existing YAML. If there is no selection and several eligible papers, the agent asks which to cover. It can enrich a paper with alphaXiv page evidence when a PDF URL and API key are available. It then checks and revises the deterministic draft against the source paper, runs Review, and renders only an approved briefing. Each selected paper receives its own HTML and OpenSlide project. Review checks references, numbers, and structural policy, but cannot replace a human reading of the method or results. Output: per-paper paths and unresolved source limitations.