STAR

Analyze what a plan's execution run actually produced and judge it against what the plan expected. A PLAN_NAME (slug / numeric prefix / filename) resolves through the plan's exec_runs to its wkdrs/<run>/ directory; a wkdrs/<run>/ path back-resolves to its plan; no argument lists the runs on disk and asks. Inventories the §4 deliverables against disk, corroborates EXEC_LOG's step claims with artifacts, scans training / eval logs for health signals (crashes, NaN, OOM, divergence, overfitting), extracts the metrics the §5 done-criteria name and scores them against those criteria plus root §4 metrics and stated baselines, interprets what the numbers mean for the claim the plan traces to (root kill-criteria, leakage smells, single-seed limits), and appends a lightweight comparison when sibling runs of the same plan exist. Renders curves only when matplotlib is already installed (never installs anything), re-reads every cited number before it enters the report, and writes the analysis under wkdrs/<run>/. Read-only

Summarize what the experiment programme has done lately, on the time axis. No argument resumes from the previous digest's watermark and covers everything since; a PLAN_NAME covers that node's whole family — ancestors for claim context, every descendant for evidence — unbounded in time; `<N>d` or a date covers a window; `all` re-seeds from the beginning. Collects each in-scope run's newest EXPT_ANALYSIS report, tabulates verdicts and headline metrics with their provenance, derives what moved since the previous digest (new runs, changed verdicts, newly analyzed runs), gathers strategy signals and kill-criteria hits, notes which plans were created or revised in the period, and lists the gaps. A run with no analysis report is read raw for a provisional line only, tagged unverified in its own table, never scored and never quoted as a result. Writes one dated digest to wkdrs/digests/. Read-only otherwise: never edits plans, exec_status, logs, or the results ledger, and never re-runs an experiment. Use when the user

Read-only overview of the whole research flow. Scans every metds/plans/*_plan.md, rebuilds the decomposition tree from parent/prefix, reads each node's section status, children, depends_on, and exec_status (plus each run's EXEC_LOG.md for step-level progress), then renders the tree with status, a progress rollup, the single next action, and any staleness. Also checks the surrounding stages — ideas, refs, code reviews, experiment analyses, method documents — for finished work whose follow-up is missing or out of date. Never writes. Use when the user invokes $star-flow-status, or asks for the status / overview / progress of their research or plans, what to work on or execute next, what is left owing, how far a plan or its sub-plans have gotten, or to see the plan tree. Bilingual (en/zh).

Analyze what a plan's execution run actually produced and judge it against what the plan expected. A PLAN_NAME (slug / numeric prefix / filename) resolves through the plan's exec_runs to its wkdrs/<run>/ directory; a wkdrs/<run>/ path back-resolves to its plan; no argument lists the runs on disk and asks. Inventories the §4 deliverables against disk, corroborates EXEC_LOG's step claims with artifacts, scans training / eval logs for health signals (crashes, NaN, OOM, divergence, overfitting), extracts the metrics the §5 done-criteria name and scores them against those criteria plus root §4 metrics and stated baselines, interprets what the numbers mean for the claim the plan traces to (root kill-criteria, leakage smells, single-seed limits), and appends a lightweight comparison when sibling runs of the same plan exist. Renders curves only when matplotlib is already installed (never installs anything), re-reads every cited number before it enters the report, and writes the analysis under wkdrs/<run>/. Read-only

Summarize what the experiment programme has done lately, on the time axis. No argument resumes from the previous digest's watermark and covers everything since; a PLAN_NAME covers that node's whole family — ancestors for claim context, every descendant for evidence — unbounded in time; `<N>d` or a date covers a window; `all` re-seeds from the beginning. Collects each in-scope run's newest EXPT_ANALYSIS report, tabulates verdicts and headline metrics with their provenance, derives what moved since the previous digest (new runs, changed verdicts, newly analyzed runs), gathers strategy signals and kill-criteria hits, notes which plans were created or revised in the period, and lists the gaps. A run with no analysis report is read raw for a provisional line only, tagged unverified in its own table, never scored and never quoted as a result. Writes one dated digest to wkdrs/digests/. Read-only otherwise: never edits plans, exec_status, logs, or the results ledger, and never re-runs an experiment. Use when the user

Read-only overview of the whole research flow. Scans every metds/plans/*_plan.md, rebuilds the decomposition tree from parent/prefix, reads each node's section status, children, depends_on, and exec_status (plus wkdrs/<run>/EXEC_LOG.md for step-level progress), then renders the tree with status, a progress rollup, the single next action, and any staleness. Also checks the surrounding stages — ideas, refs, code reviews, experiment analyses, method documents — for finished work whose follow-up is missing or out of date. Never writes. Use when the user runs /star-flow-status, or asks for the status / overview / progress of their research or plans, what to work on or execute next, what is left owing, how far a plan or its sub-plans have gotten, or to see the plan tree. Bilingual (en/zh).

Analyze what a plan's execution run actually produced and judge it against what the plan expected. A PLAN_NAME (slug / numeric prefix / filename) resolves through the plan's exec_runs to its wkdrs/<run>/ directory; a wkdrs/<run>/ path back-resolves to its plan; no argument lists the runs on disk and asks. Inventories the §4 deliverables against disk, corroborates EXEC_LOG's step claims with artifacts, scans training / eval logs for health signals (crashes, NaN, OOM, divergence, overfitting), extracts the metrics the §5 done-criteria name and scores them against those criteria plus root §4 metrics and stated baselines, interprets what the numbers mean for the claim the plan traces to (root kill-criteria, leakage smells, single-seed limits), and appends a lightweight comparison when sibling runs of the same plan exist. Renders curves only when matplotlib is already installed (never installs anything), re-reads every cited number before it enters the report, and writes the analysis under wkdrs/<run>/. Read-only

Summarize what the experiment programme has done lately, on the time axis. No argument resumes from the previous digest's watermark and covers everything since; a PLAN_NAME covers that node's whole family — ancestors for claim context, every descendant for evidence — unbounded in time; `<N>d` or a date covers a window; `all` re-seeds from the beginning. Collects each in-scope run's newest EXPT_ANALYSIS report, tabulates verdicts and headline metrics with their provenance, derives what moved since the previous digest (new runs, changed verdicts, newly analyzed runs), gathers strategy signals and kill-criteria hits, notes which plans were created or revised in the period, and lists the gaps. A run with no analysis report is read raw for a provisional line only, tagged unverified in its own table, never scored and never quoted as a result. Writes one dated digest to wkdrs/digests/. Read-only otherwise: never edits plans, exec_status, logs, or the results ledger, and never re-runs an experiment. Use when the user

Read-only overview of the whole research flow. Scans every metds/plans/*_plan.md, rebuilds the decomposition tree from parent/prefix, reads each node's section status, children, depends_on, and exec_status (plus wkdrs/<run>/EXEC_LOG.md for step-level progress), then renders the tree with status, a progress rollup, the single next action, and any staleness. Also checks the surrounding stages — ideas, refs, code reviews, experiment analyses, method documents — for finished work whose follow-up is missing or out of date. Never writes. Use when the user runs /star-flow-status, or asks for the status / overview / progress of their research or plans, what to work on or execute next, what is left owing, how far a plan or its sub-plans have gotten, or to see the plan tree. Bilingual (en/zh).

Decompose an existing research plan (written by star-plan-coach and living under metds/plans/) into concrete, executable sub-plans. Reads the parent plan, picks a decomposition axis (milestone / component / claim→experiment), then auto-drafts one execution sub-plan per unit — objective, dependencies, task breakdown, deliverables, and done-criteria — writing each to metds/plans/ under a hierarchical numeric prefix and linking it back to the parent. Supports arbitrary decomposition depth. Use when the user invokes $star-plan-decomposer, or wants to break down / flesh out the concrete execution details of a plan, turn a plan's method or milestones into actionable tasks, or split a plan into sub-plans. Bilingual (en/zh).

Decompose an existing research plan (written by star-plan-coach and living under metds/plans/) into concrete, executable sub-plans. Reads the parent plan, picks a decomposition axis (milestone / component / claim→experiment), then auto-drafts one execution sub-plan per unit — objective, dependencies, task breakdown, deliverables, and done-criteria — writing each to metds/plans/ under a hierarchical numeric prefix and linking it back to the parent. Supports arbitrary decomposition depth. Use when the user runs /star-plan-decomposer, or wants to break down / flesh out the concrete execution details of a plan, turn a plan's method or milestones into actionable tasks, or split a plan into sub-plans. Bilingual (en/zh).

Decompose an existing research plan (written by star-plan-coach and living under metds/plans/) into concrete, executable sub-plans. Reads the parent plan, picks a decomposition axis (milestone / component / claim→experiment), then auto-drafts one execution sub-plan per unit — objective, dependencies, task breakdown, deliverables, and done-criteria — writing each to metds/plans/ under a hierarchical numeric prefix and linking it back to the parent. Supports arbitrary decomposition depth. Use when the user runs /star-plan-decomposer, or wants to break down / flesh out the concrete execution details of a plan, turn a plan's method or milestones into actionable tasks, or split a plan into sub-plans. Bilingual (en/zh).

Execute a leaf research sub-plan from metds/plans/ against the project code. Orient in the codebase using .env, turn the sub-plan into a concrete checked execution plan, implement it with surgical changes, run light validation, keep intermediate working files under a plan-specific tasks directory, checkpoint progress under a run-specific wkdrs directory, and stop before long or costly experiments by handing the exact command to the user. Syncs user-confirmed deviations and execution-settled values back into the sub-plan with a Revision History trail, keeping the plan file true to what was actually executed. Use when the user invokes $star-plan-executor or asks Codex to execute, implement, carry out, or run a research sub-plan. Supports resume across sessions and bilingual English/Chinese work.

Execute a leaf execution sub-plan (produced by star-plan-decomposer and living under metds/plans/) against the project code. Orients in the codebase (${CODE_NAME}/, read from .env) to build a "current state vs required" gap list, enters plan mode to turn the sub-plan's task breakdown into a concrete executable plan, gates it through ExitPlanMode, then dispatches one subagent per step to modify code and run light validation — stopping before heavy experiments (long/multi-GPU training, costly API calls) and handing those commands back to the user. Keeps intermediate working files under tasks/<plan-name>/ and checkpoints durable execution state plus generated artifacts under wkdrs/<run>/ so runs resume across sessions. Syncs user-confirmed deviations and execution-settled values back into the sub-plan with a Revision History trail, keeping the plan file true to what was actually executed. Use when the user runs /star-plan-executor, or wants to execute / implement / carry out / run a sub-plan, turn an execution p

Execute a leaf execution sub-plan (produced by star-plan-decomposer and living under metds/plans/) against the project code. Orients in the codebase (${CODE_NAME}/, read from .env) to build a "current state vs required" gap list, switches to Cursor plan mode to turn the sub-plan's task breakdown into a concrete executable plan, gates it on user approval, then dispatches one Task subagent per step to modify code and run light validation — stopping before heavy experiments (long/multi-GPU training, costly API calls) and handing those commands back to the user. Keeps intermediate working files under tasks/<plan-name>/ and checkpoints durable execution state plus generated artifacts under wkdrs/<run>/ so runs resume across sessions. Syncs user-confirmed deviations and execution-settled values back into the sub-plan with a Revision History trail, keeping the plan file true to what was actually executed. Use when the user runs /star-plan-executor, or wants to execute / implement / carry out / run a sub-plan, turn a

Give research plans under metds/plans/ a code home. If ${CODE_NAME}/ (from .env) is missing or empty, mine the plan for a search profile, find and score candidate reference implementations on GitHub (plan fit, completeness, license, activity), let the user pick, then clone, strip git history, record provenance, and conservatively rebrand to CODE_NAME. If code already exists, survey it read-only instead. Both paths then design a target architecture plus a migration table, execute only user-approved migrations with per-group verification and git checkpoints, and write the spec to metds/codearc.md with thin pointers in AGENTS.md and .cursor/rules/. Use when the user invokes $star-code-architect or asks Codex to find a reference implementation, scaffold ${CODE_NAME}/, or organize and refactor the existing codebase. Supports bilingual English/Chinese work.

Review code against the project's written conventions and, when scoped to a plan, against what that plan promised. With no argument it reviews all of ${CODE_NAME}/ (read from .env); a PLAN_NAME (slug / numeric prefix / filename) reviews the files that plan touches plus plan conformance (§3 tasks implemented, §4 deliverables on disk, §5 done-criterion supported); an existing path reviews that path; `diff` or a git range reviews only changed files. Gathers cheap static evidence through the .env conda env (never installs tools), collects findings against a six-dimension rubric (docstrings, naming, simplicity, STAR conventions, correctness smells, plan conformance), re-verifies blocker/major findings before reporting, writes the review report under wkdrs/, then offers a per-item-approved fix pass for mechanical, behavior-preserving issues only — feature gaps route to star-plan-executor, plan divergence to star-plan-reviser, structural reorganization to star-code-architect. Use when the user invokes $star-code-rev

Build and verify the project's Python runtime environment so plan execution has a working interpreter. Reads .env: a valid CONDA_HOME creates conda env ENV_NAME (argument, default CODE_NAME); otherwise a .venv under the project root. An existing environment is never deleted — after user confirmation it is renamed to a dated backup (real run date) before rebuilding. Dependencies resolve first-signal-wins: existing CODE_NAME/requirements* → packaging metadata (pyproject / setup.py / environment.yml) → import scan of the code, with generated results written to a two-tier layout (requirements.txt referencing requirements/framework|runtime|optional.txt; conda-only items in requirements/conda.txt). Installs through a uv > pip > conda ladder with CUDA-aware framework wheel selection behind a single install-plan gate, then smoke-tests in three layers (imports → framework/GPU → project entrypoint) and writes ENV_REPORT.md plus a version freeze under wkdrs/. Use when the user invokes $star-env-builder or asks Codex to

Coach a researcher from a vague interest to a defensible research topic through diverge–scan–converge: clarify the seed and its constraints, generate 3–5 genuinely distinct candidate directions, ground the kept ones in an abstract-level literature scan (every named paper transcribed from a record fetched during the run, source URL logged — never from memory), score them on a six-dimension rubric with Pursue / Refine / Park verdicts, then frame the winner into a topic statement with a first validation experiment — written incrementally to metds/ideas/<slug>_idea.md with cross-session resume. The finalized idea file seeds $star-plan-coach. Use when the user runs $star-idea-storm, wants to brainstorm / 头脑风暴 research directions, has an interest area but no committed topic, asks "what should I research", or mentions idea files under metds/ideas. Bilingual (en/zh).

Compile the research-plan tree under metds/plans/ into paper-ready method documents under metds/. Invoked as $star-metd-summarize [OPT] where OPT is overview, dataset, framework, training, or evaluation; no argument compiles all five in dependency order (dataset → framework → training → evaluation → overview, which links the other four and so compiles last). Rebuilds the plan tree from parent:, then extracts what each document needs through a written map (overview ← root §1 problem, §2 positioning, §3 idea, §6 milestones; dataset ← §4 data choices plus every leaf's §2 datas/ inputs and data-building steps; framework ← §3 route plus modeling leaves and their ${CODE_NAME}/ paths; training ← §3 strategy, §4 budget, inits/ and hyperparameters; evaluation ← §4 benchmarks, baselines, metrics and ablation design plus §5 kill-criteria), merges the passages along the method's axis instead of the plan's, resolves conflicts leaf-over-parent and newer-over-older, marks content from unexecuted leaves as not yet verified,

Coach CS researchers through writing a research plan via staged Socratic questions (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 write or refine a research plan, proposal, or 开题报告; flesh out a research idea; grow a finalized idea file under metds/ideas into a plan; mentions plan files under metds/plans; or has an idea but is unsure how to proceed — even if they never say the word "plan". Bilingual (en/zh).

Review one research plan from metds/plans/ against its execution evidence, then revise it in place with per-item user approval. Read the wkdrs/<run>/ execution logs and artifacts (children rollups for internal nodes), score completion claim-by-claim against files on disk, write a seven-part review report to wkdrs/, walk revision candidates one question at a time, edit the plan file directly, and append a Revision History entry — routing structural re-shaping to star-plan-decomposer and strategy pivots to star-plan-coach. Use when the user invokes $star-plan-reviser or asks Codex to review, audit, or revise a plan after (partial) execution, or to fold execution results back into the plan. Supports bilingual English/Chinese work.

Build an auditable related-work base for the project's method: 5–10 close papers read into per-paper analysis notes, plus a classified reference.bib of ≥50 verified entries, all under metds/refs/. With no argument it reads the method from metds/*.md (falling back to the root plan under metds/plans/, then to a finalized idea file under metds/ideas/, then to a topic the user supplies) and runs the full pass, resuming incrementally when metds/refs/ already exists; a PLAN_NAME or free-text topic scopes the search; `verify` re-fetches every entry and diffs it against the file; `organize` re-classifies the existing bib offline; `synthesize` compiles the existing notes into a related-work narrative under metds/refs/; an arXiv id, DOI, or paper URL appends one paper. Every bib field is transcribed from a record fetched during the run (DBLP → Crossref → Semantic Scholar → arXiv, published version preferred), cached under wkdrs/, and logged with its source URL in metds/refs/refs_index.md — nothing is written from memor

Bootstrap or reorganize the project codebase (${CODE_NAME}/, read from .env) so research plans under metds/plans/ have a code home. When ${CODE_NAME}/ is missing or empty: mine the research plan for a search profile, find and score candidate reference implementations on GitHub (plan fit, completeness, license, activity), let the user pick one, then clone it, strip its git history, record provenance, and conservatively rebrand it to CODE_NAME. When code already exists: survey it with read-only subagents instead. Both paths then design a target architecture plus a migration table, execute only user-approved migrations via orchestrated subagents with per-group verification and git checkpoints, and write the architecture spec to metds/codearc.md with thin pointers in AGENTS.md and .cursor/rules/. Use when the user runs /star-code-architect, wants a reference implementation or starter codebase for a plan, wants to set up / scaffold ${CODE_NAME}/, or wants to organize / refactor the existing codebase and record its

Review code against the project's written conventions and, when scoped to a plan, against what that plan promised. With no argument it reviews all of ${CODE_NAME}/ (read from .env); a PLAN_NAME (slug / numeric prefix / filename) reviews the files that plan touches plus plan conformance (§3 tasks implemented, §4 deliverables on disk, §5 done-criterion supported); an existing path reviews that path; `diff` or a git range reviews only changed files. Gathers cheap static evidence through the .env conda env (never installs tools), collects findings against a six-dimension rubric (docstrings, naming, simplicity, STAR conventions, correctness smells, plan conformance), re-verifies blocker/major findings before reporting, writes the review report under wkdrs/, then offers a per-item-approved fix pass for mechanical, behavior-preserving issues only — feature gaps route to star-plan-executor, plan divergence to star-plan-reviser, structural reorganization to star-code-architect. Use when the user runs /star-code-review

Build and verify the project's Python runtime environment so plan execution has a working interpreter. Reads .env: a valid CONDA_HOME creates conda env ENV_NAME (argument, default CODE_NAME); otherwise a .venv under the project root. An existing environment is never deleted — after user confirmation it is renamed to a dated backup (real run date) before rebuilding. Dependencies resolve first-signal-wins: existing CODE_NAME/requirements* → packaging metadata (pyproject / setup.py / environment.yml) → import scan of the code, with generated results written to a two-tier layout (requirements.txt referencing requirements/framework|runtime|optional.txt; conda-only items in requirements/conda.txt). Installs through a uv > pip > conda ladder with CUDA-aware framework wheel selection behind a single install-plan gate, then smoke-tests in three layers (imports → framework/GPU → project entrypoint) and writes ENV_REPORT.md plus a version freeze under wkdrs/. Use when the user runs /star-env-builder, wants the project's

Coach a researcher from a vague interest to a defensible research topic through diverge–scan–converge: clarify the seed and its constraints, generate 3–5 genuinely distinct candidate directions, ground the kept ones in an abstract-level literature scan (every named paper transcribed from a record fetched during the run, source URL logged — never from memory), score them on a six-dimension rubric with Pursue / Refine / Park verdicts, then frame the winner into a topic statement with a first validation experiment — written incrementally to metds/ideas/<slug>_idea.md with cross-session resume. The finalized idea file seeds /star-plan-coach. Use when the user runs /star-idea-storm, wants to brainstorm / 头脑风暴 research directions, has an interest area but no committed topic, asks "what should I research", or mentions idea files under metds/ideas. Bilingual (en/zh).

Compile the research-plan tree under metds/plans/ into paper-ready method documents under metds/. Invoked as /star-metd-summarize [OPT] where OPT is overview, dataset, framework, training, or evaluation; no argument compiles all five in dependency order (dataset → framework → training → evaluation → overview, which links the other four and so compiles last). Rebuilds the plan tree from parent:, then extracts what each document needs through a written map (overview ← root §1 problem, §2 positioning, §3 idea, §6 milestones; dataset ← §4 data choices plus every leaf's §2 datas/ inputs and data-building steps; framework ← §3 route plus modeling leaves and their ${CODE_NAME}/ paths; training ← §3 strategy, §4 budget, inits/ and hyperparameters; evaluation ← §4 benchmarks, baselines, metrics and ablation design plus §5 kill-criteria), merges the passages along the method's axis instead of the plan's, resolves conflicts leaf-over-parent and newer-over-older, marks content from unexecuted leaves as not yet verified,

Coach CS researchers through writing a research plan via staged Socratic questions (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 write or refine a research plan, proposal, or 开题报告; flesh out a research idea; grow a finalized idea file under metds/ideas into a plan; mentions plan files under metds/plans; or has an idea but is unsure how to proceed — even if they never say the word "plan". Bilingual (en/zh).

Review one research plan (any node under metds/plans/) against its execution evidence, then revise it in place with per-item user approval. Dispatches read-only subagents to inspect wkdrs/<run>/ execution logs and artifacts (children rollups for internal nodes), scores completion claim-by-claim against files on disk, writes a seven-part review report to wkdrs/, walks revision candidates one question at a time, edits the plan file directly and appends a Revision History entry — routing structural re-shaping to star-plan-decomposer and strategy pivots to star-plan-coach. Use when the user runs /star-plan-reviser, or wants to review / audit / revise a plan after (partial) execution, check what a plan actually did versus what it promised, or fold execution results back into the plan. Bilingual (en/zh).

Build an auditable related-work base for the project's method: 5–10 close papers read into per-paper analysis notes, plus a classified reference.bib of ≥50 verified entries, all under metds/refs/. With no argument it reads the method from metds/*.md (falling back to the root plan under metds/plans/, then to a finalized idea file under metds/ideas/, then to a topic the user supplies) and runs the full pass, resuming incrementally when metds/refs/ already exists; a PLAN_NAME or free-text topic scopes the search; `verify` re-fetches every entry and diffs it against the file; `organize` re-classifies the existing bib offline; `synthesize` compiles the existing notes into a related-work narrative under metds/refs/; an arXiv id, DOI, or paper URL appends one paper. Every bib field is transcribed from a record fetched during the run (DBLP → Crossref → Semantic Scholar → arXiv, published version preferred), cached under wkdrs/, and logged with its source URL in metds/refs/refs_index.md — nothing is written from memor

Bootstrap or reorganize the project codebase (${CODE_NAME}/, read from .env) so research plans under metds/plans/ have a code home. When ${CODE_NAME}/ is missing or empty: mine the research plan for a search profile, find and score candidate reference implementations on GitHub (plan fit, completeness, license, activity), let the user pick one, then clone it, strip its git history, record provenance, and conservatively rebrand it to CODE_NAME. When code already exists: survey it with read-only Task subagents instead. Both paths then design a target architecture plus a migration table, execute only user-approved migrations via orchestrated Task subagents with per-group verification and git checkpoints, and write the architecture spec to metds/codearc.md with thin pointers in AGENTS.md and .cursor/rules/. Use when the user runs /star-code-architect, wants a reference implementation or starter codebase for a plan, wants to set up / scaffold ${CODE_NAME}/, or wants to organize / refactor the existing codebase and

Review code against the project's written conventions and, when scoped to a plan, against what that plan promised. With no argument it reviews all of ${CODE_NAME}/ (read from .env); a PLAN_NAME (slug / numeric prefix / filename) reviews the files that plan touches plus plan conformance (§3 tasks implemented, §4 deliverables on disk, §5 done-criterion supported); an existing path reviews that path; `diff` or a git range reviews only changed files. Gathers cheap static evidence through the .env conda env (never installs tools), collects findings against a six-dimension rubric (docstrings, naming, simplicity, STAR conventions, correctness smells, plan conformance), re-verifies blocker/major findings before reporting, writes the review report under wkdrs/, then offers a per-item-approved fix pass for mechanical, behavior-preserving issues only — feature gaps route to star-plan-executor, plan divergence to star-plan-reviser, structural reorganization to star-code-architect. Use when the user runs /star-code-review

Build and verify the project's Python runtime environment so plan execution has a working interpreter. Reads .env: a valid CONDA_HOME creates conda env ENV_NAME (argument, default CODE_NAME); otherwise a .venv under the project root. An existing environment is never deleted — after user confirmation it is renamed to a dated backup (real run date) before rebuilding. Dependencies resolve first-signal-wins: existing CODE_NAME/requirements* → packaging metadata (pyproject / setup.py / environment.yml) → import scan of the code, with generated results written to a two-tier layout (requirements.txt referencing requirements/framework|runtime|optional.txt; conda-only items in requirements/conda.txt). Installs through a uv > pip > conda ladder with CUDA-aware framework wheel selection behind a single install-plan gate, then smoke-tests in three layers (imports → framework/GPU → project entrypoint) and writes ENV_REPORT.md plus a version freeze under wkdrs/. Use when the user runs /star-env-builder, wants the project's

Coach a researcher from a vague interest to a defensible research topic through diverge–scan–converge: clarify the seed and its constraints, generate 3–5 genuinely distinct candidate directions, ground the kept ones in an abstract-level literature scan (every named paper transcribed from a record fetched during the run, source URL logged — never from memory), score them on a six-dimension rubric with Pursue / Refine / Park verdicts, then frame the winner into a topic statement with a first validation experiment — written incrementally to metds/ideas/<slug>_idea.md with cross-session resume. The finalized idea file seeds /star-plan-coach. Use when the user runs /star-idea-storm, wants to brainstorm / 头脑风暴 research directions, has an interest area but no committed topic, asks "what should I research", or mentions idea files under metds/ideas. Bilingual (en/zh).

Compile the research-plan tree under metds/plans/ into paper-ready method documents under metds/. Invoked as /star-metd-summarize [OPT] where OPT is overview, dataset, framework, training, or evaluation; no argument compiles all five in dependency order (dataset → framework → training → evaluation → overview, which links the other four and so compiles last). Rebuilds the plan tree from parent:, then extracts what each document needs through a written map (overview ← root §1 problem, §2 positioning, §3 idea, §6 milestones; dataset ← §4 data choices plus every leaf's §2 datas/ inputs and data-building steps; framework ← §3 route plus modeling leaves and their ${CODE_NAME}/ paths; training ← §3 strategy, §4 budget, inits/ and hyperparameters; evaluation ← §4 benchmarks, baselines, metrics and ablation design plus §5 kill-criteria), merges the passages along the method's axis instead of the plan's, resolves conflicts leaf-over-parent and newer-over-older, marks content from unexecuted leaves as not yet verified,

Coach CS researchers through writing a research plan via staged Socratic questions (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 write or refine a research plan, proposal, or 开题报告; flesh out a research idea; grow a finalized idea file under metds/ideas into a plan; mentions plan files under metds/plans; or has an idea but is unsure how to proceed — even if they never say the word "plan". Bilingual (en/zh).

Review one research plan (any node under metds/plans/) against its execution evidence, then revise it in place with per-item user approval. Dispatches read-only Task subagents to inspect wkdrs/<run>/ execution logs and artifacts (children rollups for internal nodes), scores completion claim-by-claim against files on disk, writes a seven-part review report to wkdrs/, walks revision candidates one question at a time, edits the plan file directly and appends a Revision History entry — routing structural re-shaping to star-plan-decomposer and strategy pivots to star-plan-coach. Use when the user runs /star-plan-reviser, or wants to review / audit / revise a plan after (partial) execution, check what a plan actually did versus what it promised, or fold execution results back into the plan. Bilingual (en/zh).

Build an auditable related-work base for the project's method: 5–10 close papers read into per-paper analysis notes, plus a classified reference.bib of ≥50 verified entries, all under metds/refs/. With no argument it reads the method from metds/*.md (falling back to the root plan under metds/plans/, then to a finalized idea file under metds/ideas/, then to a topic the user supplies) and runs the full pass, resuming incrementally when metds/refs/ already exists; a PLAN_NAME or free-text topic scopes the search; `verify` re-fetches every entry and diffs it against the file; `organize` re-classifies the existing bib offline; `synthesize` compiles the existing notes into a related-work narrative under metds/refs/; an arXiv id, DOI, or paper URL appends one paper. Every bib field is transcribed from a record fetched during the run (DBLP → Crossref → Semantic Scholar → arXiv, published version preferred), cached under wkdrs/, and logged with its source URL in metds/refs/refs_index.md — nothing is written from memor

Adopt an already-started project into STAR without disturbing it. Phase `survey` probes the existing repository read-only (source layout, runtime, data / weights / output locations, launch entrypoints, git history, prior runs), confirms the mapping with the user, then lands the mechanical setup — writes .env, reaches large existing directories by symlink rather than moving them, wraps existing launch commands into execs/scpts/ — and records a work inventory of what is already built, already run, and already concluded in metds/adopt.md, with the user's chosen historical runs ledgered into wkdrs/. Phase `backfill` runs after the plan tree exists: it matches that inventory to the leaves and, per leaf and only on the user's confirmation, records exec_status / exec_runs so the tree reflects real progress instead of reading as 0%. Use when the user runs $star-proj-adopt, wants to bring an existing / partially finished project into STAR, asks how to onboard a repo that did not start from the template, or needs alrea