The user has typed star-auto — this file is what that invocation runs. Typing it is the auto grant of conventions §10.7: while this run pursues its goal, the skill each next action names may be started without further naming, the seven explicit-only ones included, and a prepared STOP-line command may be launched…
A command that lets an automated research workflow continue through approved steps toward a stated goal, with an optional stopping rule and involvement level.
Find a reference implementation, scaffold ${CODENAME}/, or reorganize the codebase — giving research plans under metds/plans/ a place for the code to live. If ${CODENAME}/ (from .env) is missing or empty, read the plan for what to search for, find and score candidate implementations on GitHub (plan fit, completeness…
Open-source or publish the project: gather scattered code, polish it, compile README.md. Consolidates into ${CODENAME}/, sweeping tasks/, wkdrs/ and the root, promoting only code passing a three-part evidence test (the README cites it, an executed leaf needs it, or it reproduces a results.md number), placed by…
Review code quality, docstrings, and whether a plan's code matches what it promised. No argument reviews all of ${CODENAME}/; a PLANNAME (slug / prefix / filename) reviews that plan's files plus conformance (§3 tasks, §4 deliverables, §5 done-criterion); a path reviews that path; diff or a git range reviews changed…
Create or rebuild the conda env or venv, install dependencies, verify the environment — so plan execution has a working interpreter. Reads .env: a valid CONDAHOME creates conda env ENVNAME (argument, default CODENAME); otherwise a .venv in the root. An existing environment is never deleted — after confirmation it is…
Analyze experiment results: read the training logs and metrics, score a run against its plan. A PLANNAME resolves via execruns to its wkdrs/ /; a wkdrs/ / path back-resolves to its plan; no argument lists the runs and asks. Inventories the §4 deliverables, checks EXECLOG's claims against artifacts, scans logs for…
Summarize experiment progress since last time — the periodic digest a progress report is written from, in date order. No argument resumes from the last digest; a PLANNAME covers that node's whole family — ancestors for context, descendants for evidence; d or a date covers a window; all covers everything. Collects each…
Show research status: the plan tree, progress, and what to work on or execute next. Read-only overview of the whole flow — scans every metds/plans/plan.md, rebuilds the decomposition tree from parent/prefix, reads each node's section status, children, dependson, and execstatus (plus each run's EXECLOG.md for…
Brainstorm research directions and converge on a topic — for "what should I research" and 头脑风暴. Coaches from a vague interest to a defensible topic through diverge–scan–converge: clarify the starting idea and its constraints, generate 3–5 genuinely distinct candidate directions, ground the kept ones in an…
Compile the finished plans into paper-ready method documents for the write-up. Invoked as star-metd-summarize [OPT] — overview, dataset, framework, training, evaluation; no argument compiles all five, overview last. Extracts what each document needs through a written map, merges passages by what a reader needs, not by…
Write or refine a research plan, proposal, or 开题报告 — one question at a time, stage by stage (problem → related work → method → experiments → risks → milestones), writing each finished section to metds/plans/ and supporting cross-session resume. Use whenever the user wants to flesh out a research idea; grow a finalized…
Break a research plan down into executable sub-plans with concrete, checkable steps — turning its method or milestones into actionable tasks. Reads the parent plan (written by star-plan-coach and living under metds/plans/), picks a decomposition axis (phase / component / evidence), confirms the unit list, then by…
Execute or implement a leaf research sub-plan: turn it into code and run light validation. Orients in the codebase using .env, turns the sub-plan from metds/plans/ into a concrete checked execution plan, implements it with surgical changes, keeps intermediate working files under a plan-specific tasks directory…
Audit a plan against what execution did, then revise it in place with per-item user approval — folding execution results back into the plan. Reads one plan from metds/plans/ plus its wkdrs/ / execution logs and artifacts (a summary of the children for internal nodes), scores completion claim-by-claim against files on…
Bring an existing or half-finished project into STAR without disturbing it — onboarding a repo that did not start from the template. Phase survey inspects the repository read-only (source layout, runtime, data / weights / output locations, entrypoints, git history, prior runs), confirms the mapping, then puts the…
Literature review and related work: per-paper analysis notes plus a verified reference.bib. Builds an auditable base — 5–10 close papers read into notes, plus a classified reference.bib of ≥50 verified entries. No argument finds the method in metds/ and runs the full pass, resuming if metds/refs/ exists; a PLANNAME or…
Runs before the agent uses a tool for Edit, Write, NotebookEdit and Bash tool calls, executing star_involve_gate.sh, star_commit_guard.sh and star_bash_gate.sh (3 commands). From wanghao9610/STAR.
Drive the STAR workflow toward the goal an explicitly invoked $star-auto request states, starting each next skill itself under that invocation's grant. Do not use unless the user typed $star-auto.