[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M74GPKC7JCRASWRKXZ7CKPE5":3,"skill-same:01M74GPKC7JCRASWRKXZ7CKPE5":74},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":14,"repositoryHtmlUrl":27,"repositoryHomepageUrl":6,"ownerHtmlUrl":28,"ownerBlogUrl":29,"ownerDescription":30,"updatedAt":31,"install":32,"resources":45},"01M74GPKC7JCRASWRKXZ7CKPE5","research-survey",null,"[research-workflow sub-skill] Scoped to research-workflow projects only — do not invoke as a standalone tool. Defines how to survey a sub-topic for a research project. Use when the task is 'investigate this sub-area' rather than 'record this one paper' — e.g., the main agent receives a sub-topic brief, or research-workflow Workflow E discovers it is facing a sub-area sweep, not a single-paper intake. Defines the input contract (so sub-agents can work without follow-up questions), the six survey constitution rules (recency, venue, version, code, content fields, objectivity), the sweep workflow, and the output format. Pairs with research-workflow Workflow E' (sub-topic surveying).","Siyuexi\u002Fcodex-research-kit","Siyuexi","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F50619651?v=4","01M74GNP0DQZKN3088BDHCKCDF","Junjielong Xu","01M74GNP13FNC637NW9K8V98HZ",0,"MIT","https:\u002F\u002Fgithub.com\u002FSiyuexi\u002Fcodex-research-kit\u002Fblob\u002Fmain\u002Fplugins\u002Fcodex-research-kit\u002Fskills\u002Fresearch-survey\u002FSKILL.md",[],[],"2026-11-11 15:11:14",false,"---\nname: research-survey\ndescription: \"[research-workflow sub-skill] Scoped to research-workflow projects only — do not invoke as a standalone tool. Defines how to survey a sub-topic for a research project. Use when the task is 'investigate this sub-area' rather than 'record this one paper' — e.g., the main agent receives a sub-topic brief, or research-workflow Workflow E discovers it is facing a sub-area sweep, not a single-paper intake. Defines the input contract (so sub-agents can work without follow-up questions), the six survey constitution rules (recency, venue, version, code, content fields, objectivity), the sweep workflow, and the output format. Pairs with research-workflow Workflow E' (sub-topic surveying).\"\nlicense: MIT\n---\n\n# research-survey — Surveying a Sub-Topic for a research-workflow Project\n\nThis skill is a **sub-skill of `research-workflow`**. It does not redefine the research constitution — it inherits from research-workflow SKILL.md — and it does not handle single-paper intake (that is research-workflow Workflow E). It handles the case where the task is **\"go investigate sub-area X and bring back a structured survey\"**.\n\nThe skill is designed to be runnable by **isolated sub-agents** (no ability to ask the user back). It therefore enforces an explicit **input contract**: the dispatcher must provide every field the worker needs up-front. If the brief is incomplete, the worker returns a failure receipt rather than guessing.\n\n---\n\n## When to Use\n\n- The main agent or user says: \"调研一下 X 这个子方向 \u002F sweep the literature on X \u002F survey related work on X\"\n- research-workflow Workflow E's step 1 reveals the request is not a single paper but a sub-area\n- A `survey\u002F\u003Cslug>.brief.md` file is handed to a worker (sub-agent or remote routine)\n- Inside an **auto research-workflow** run, the controller dispatches sub-topic sweeps in parallel\n\n## When NOT to Use\n\n- A specific paper is already named and needs recording → research-workflow **Workflow E** (single-paper intake)\n- Early brainstorming with no target sub-topic yet → `brainstorming-research-ideas`\n- The project lives in `~\u002FDocuments\u002Fproject\u002F00-QuickReview\u002F` (no project constitution applies; do whatever is convenient)\n- Citation \u002F BibTeX verification during paper writing → `20-ml-paper-writing`\n\n---\n\n## Section 1 — Input Contract (the brief)\n\nA sub-agent cannot call `ask the user`. To avoid mid-task guessing, every sweep starts from a **brief** that the dispatcher fills out. The brief is a markdown file under `\u003Cproject>\u002Fsurvey\u002F\u003Cslug>.brief.md`, generated from `templates\u002Fsubtopic-brief.md.template`.\n\n### Required fields\n\n| Field | Meaning | Example |\n|---|---|---|\n| `subtopic` | One-line description of the sub-area | \"Advantage normalization techniques in GRPO-style RL fine-tuning\" |\n| `slug` | kebab-case identifier, used in filenames | `grpo-adv-norm` |\n| `parent_proposal_ref` | Relative path (or `proposal.md#section`) the worker reads to learn the main topic | `..\u002Fproposal.md#section-3-method` |\n| `axes_of_interest` | Comma list of dimensions the worker should compare against the parent topic | `task, method, dataset, metric` |\n| `time_window_start` | ISO date, defaults to the project's creation date (NOT today). Fixed across the project's life. | `2026-05-14` (means: papers published on\u002Fafter this date, looking back 5 years → `2021-05-14`) |\n| `venue_whitelist_extra` | Project-specific extras beyond CCF-A (read from project AGENTS.md `accepted_venues`) | `ICLR, COLM, TMLR` |\n| `target_count` | Soft cap on number of papers in the final survey (recommend 8–15; if a sub-topic needs >20, the dispatcher should split it) | `12` |\n| `output_path` | Where to write the survey | `survey\u002Fgrpo-adv-norm.md` |\n\n### Optional fields\n\n| Field | Meaning |\n|---|---|\n| `seed_papers` | Known-relevant papers the dispatcher already has, used as anchors for cite-chain search |\n| `exclude_papers` | DBLP keys \u002F arXiv ids to skip (already in main survey.md, or known irrelevant) |\n| `recency_override` | If set, expands time window beyond 5 years (with reason, e.g., \"include foundational 2017 paper X\") |\n| `notes_for_worker` | Free-form extra context (subtle distinctions, terminology the worker should not confuse) |\n\n### Time window semantics (per the project decision)\n\n- `time_window_start` is the **project's creation date**, recorded once and reused.\n- The accepted publication window is `[time_window_start - 5 years, today]`.\n- Rationale: a survey done in month 1 and a survey done in month 10 of the same project should agree on what \"近五年\" means. Rolling windows break this.\n- If the worker needs to break the 5-year wall, the dispatcher must set `recency_override` in the brief. The worker does NOT decide this on its own.\n\n### If the brief is incomplete\n\nThe worker does **not** guess. It writes a failure receipt (see `templates\u002Ffailure-receipt.md.template`) listing the missing fields and stops. The dispatcher fixes the brief and re-dispatches.\n\n---\n\n## Section 2 — The Survey Constitution (C-S1 ~ C-S6)\n\nThese rules govern every sweep. They sit **below** the research-workflow constitution (C1–C8) and inherit from it; where this skill is silent, research-workflow applies.\n\n### C-S1. Recency: ≥ (project-creation-date − 5 years)\n\nDefault: only papers published in the 5-year window described in Section 1.\n\n- **Why**: surveys for active research projects need current state of the art, not historical foundations.\n- **How to apply**: filter candidates by publication year against `time_window_start − 5y`. Older papers go to a `dropped_for_recency` list at the bottom of the candidates file, not into the survey.\n- **Override**: only when `recency_override` is set in the brief.\n\n### C-S2. Venue: CCF-A by default, plus project-declared extras\n\n- **Default whitelist**: CCF-A conferences and journals across the relevant area (AI\u002FML: NeurIPS, ICML, AAAI, IJCAI, ACL, EMNLP, NAACL, CVPR, ICCV, ECCV, TPAMI, IJCV, JMLR, TOG, SIGGRAPH; SE: ICSE, FSE, ASE, ISSTA, OOPSLA, TSE, TOSEM; Sys: SOSP, OSDI, ASPLOS, ISCA, MICRO, HPCA, SIGCOMM, NSDI, MobiCom, FAST, EuroSys; DB: SIGMOD, VLDB, ICDE; Sec: CCS, S&P, USENIX Security, NDSS; Theory: STOC, FOCS, SODA, LICS, CAV; HCI: CHI, UIST, CSCW). When in doubt about a venue's tier, **do not silently include it** — the worker writes the candidate to the candidates file with `venue_justification: needs-review` and surfaces it in the final receipt for the dispatcher to decide.\n- **Project extras**: the project's `AGENTS.md` may declare an `accepted_venues:` field naming additional venues the project considers acceptable (ICLR, COLM, TMLR are common examples). The worker reads this and merges it with the CCF-A default. The brief's `venue_whitelist_extra` overrides per-sweep.\n- **Workshop papers, arXiv-only papers, technical reports**: not acceptable as primary entries. They may appear as `discovered_via` references for context but do not get rows in the survey.\n- **Why**: keeps the sweep anchored to peer-reviewed venues the project lead trusts, while allowing per-project flexibility.\n\n### C-S3. Version: publication > arXiv\n\n- **Preferred**: the venue's published version (PDF from the publisher's site, the conference proceedings page, or the journal page).\n- **Acceptable fallback**: if the publication version is paywalled and not retrievable, use the arXiv version. Record `version: arxiv-fallback` and `fallback_reason: paywalled` on the entry.\n- **Not acceptable**: using arXiv when the publication version is freely available. The worker must attempt the publication URL first and record the attempt.\n- **Why**: the publication version is the one of record; reviewers will compare to it.\n\n### C-S4. Code: open-source repository required (or proof of absence)\n\nEvery survey entry must carry either:\n\n- `code: \u003Cgithub-url>` — verified to exist and to correspond to the paper, OR\n- `code: not-found` plus a `search_trace` showing the three search steps were performed:\n  1. Paper-side: searched the paper PDF (abstract, footnotes, intro, \"Code available at\", appendix) for any URL\n  2. Author-side: checked the first\u002Flast author's GitHub profile for a repo matching the paper title or method name\n  3. Topic-side: searched GitHub for `\u003Cmethod-name>` and `\u003Cpaper-keywords>` to find third-party reimplementations (if even a third-party repo exists, log it but mark `code: third-party-only`)\n\n`code: not-found` without a `search_trace` is treated as missing data — the entry is rejected from the survey.\n\n### C-S5. Content fields (per entry)\n\nEvery entry must contain these fields, factually stated:\n\n- `what`: what the paper does (1–3 sentences, descriptive only)\n- `goal`: the stated purpose \u002F problem it tries to solve (in the paper's own framing)\n- `relation_to_parent`: how this work connects to the parent project's topic, across the `axes_of_interest`\n- `differences`: in what dimensions this work diverges from the parent topic (task \u002F method \u002F setting \u002F data \u002F metric — be specific)\n- `salience_axes`: which of the brief's `axes_of_interest` this paper **actually touches** (this is factual labeling, not a judgment). Example: `[task, dataset]` if the paper shares the task and dataset but not the method.\n\n### C-S6. Objectivity: no value judgments\n\nThe worker **describes facts**. The worker does **not**:\n\n- Rank papers by importance, relevance, novelty, or \"collision risk\" with the parent topic\n- Use evaluative language: \"the most likely to overlap\", \"the most worth learning from\", \"SOTA\", \"seminal\", \"groundbreaking\", \"weak\", \"strong\"\n- Recommend which paper the user should read first\n\nThe downstream analyst (human or another agent) decides what is salient. The worker's job is to produce a clean, complete, factual record so that downstream judgment is well-supported.\n\nStating that a paper \"shares dataset X with our proposal\" is factual. Stating that it \"is likely to be a direct competitor\" is not. Keep to the first kind.\n\n---\n\n## Section 3 — Workflow (a sweep, end to end)\n\nThis is the worker's procedure. Each step is concrete enough that a sub-agent with no prior context can execute it given only the brief and access to `research-workflow` SKILL.md.\n\n### Step 0 — Read brief and verify completeness\n\n1. Read `\u003Cproject>\u002Fsurvey\u002F\u003Cslug>.brief.md`\n2. Verify all required fields (Section 1) are present and non-empty\n3. Read project `AGENTS.md` to fetch `accepted_venues`\n4. If any required field missing → write failure receipt → stop\n\n### Step 1 — Read the parent proposal\n\nRead the `parent_proposal_ref` target. Extract:\n\n- The parent topic in one sentence\n- The specific dimensions named in `axes_of_interest` (e.g., what task, what method family, what dataset)\n- Vocabulary the parent uses (to align terminology in `relation_to_parent` \u002F `differences` later)\n\n### Step 2 — Discovery (cast a wide net)\n\nGoal: collect 30–50 candidates with metadata only (no PDF download yet).\n\nSources, in order of trust:\n\n1. **Venue ToCs**: for each venue in the whitelist, scan the relevant year's accepted papers list for keyword matches\n2. **DBLP**: author \u002F keyword queries scoped to the venue whitelist and time window\n3. **Google Scholar**: keyword search with `venue:` and year filters; useful for finding the publication version of a paper you only know from arXiv\n4. **Cite-chain from seed_papers**: forward and backward citations of any `seed_papers` in the brief\n5. **arXiv**: only for finding the publication-version URL of a recent paper; never as the primary source\n\nFor each candidate, record metadata in `\u003Cproject>\u002Fsurvey\u002F\u003Cslug>.candidates.md` (template provided):\n\n- Title, authors, venue, year\n- Stable ID (DBLP key preferred, else arXiv id, else DOI)\n- `discovered_via`: which source surfaced it\n- Abstract (copied verbatim)\n- Provisional `venue_status`: in-whitelist \u002F project-extra \u002F needs-review \u002F out-of-scope\n\n### Step 3 — Filter candidates\n\nApply, in order:\n\n1. **Venue filter** (C-S2): drop `out-of-scope`; flag `needs-review` for the dispatcher\n2. **Time filter** (C-S1): drop anything outside the window\n3. **Relevance filter** (abstract-level): drop anything whose abstract clearly does not touch the parent topic. Be permissive at this stage; better to read one extra PDF than to miss a relevant paper\n4. **Deduplication**: a paper that appears in both arXiv form and conference form is one entry, keyed by the conference DBLP key; merge `discovered_via` lists\n5. **Exclude list**: drop anything in the brief's `exclude_papers`\n\nAim to come out of this step with ~15–25 papers that warrant a PDF read.\n\n### Step 4 — Fetch publication PDFs\n\nFor each surviving candidate:\n\n1. Try the publication URL first (conference proceedings page, journal page, publisher PDF)\n2. If paywalled \u002F 403 \u002F unavailable after a reasonable attempt → fall back to arXiv version, record `version: arxiv-fallback` and the reason\n3. Save the PDF to `\u003Cproject>\u002Fcache\u002Fpapers\u002F\u003Cstable-id>.pdf`\n4. If both publication and arXiv fail → drop from this sweep, log in candidates file as `dropped_for_unavailable`\n\n`cache\u002Fpapers\u002F` is the same cache that research-workflow C4 uses. PDFs landed here become available for subsequent single-paper Workflow E intake.\n\n### Step 5 — Code repository hunt\n\nFor each paper whose PDF landed:\n\n1. **Paper-side search**: open the PDF, search for `github.com`, `gitlab.com`, \"Code available\", \"Implementation\", footnotes on page 1, the abstract page, and the appendix. Capture the URL.\n2. **Author-side search**: if not found, look up the first author and last author on GitHub (try `github.com\u002F\u003Clastname>`, then a GitHub user search). Scan their repos for one matching the paper title or method name.\n3. **Topic-side search**: if still not found, search GitHub for `\u003Cmethod-name>` and 2–3 distinctive keywords from the paper. If only third-party reimplementations exist, record one and mark `code: third-party-only`.\n4. **Verification**: open the candidate repo's README. Confirm it cites this paper or implements its method (don't just match on name). Record the verification.\n\nOutcomes:\n\n- `code: \u003Curl>` (verified official) — record `code_status: official-verified`\n- `code: \u003Curl>` (third-party only) — record `code_status: third-party-only`\n- `code: not-found` — record the full `search_trace` (what was tried, what was found)\n\n### Step 6 — Read and extract\n\nFor each paper, read at minimum: abstract, intro (especially the contributions list), method overview, experiments setup, conclusion. Extract the C-S5 content fields. Stick to descriptive language (C-S6).\n\nCross-reference with the parent proposal vocabulary (from Step 1) so `relation_to_parent` and `differences` are anchored in the project's own terms, not the paper's.\n\n### Step 7 — Write the survey file\n\nWrite to `\u003Cproject>\u002Fsurvey\u002F\u003Cslug>.md` using `templates\u002Fsubtopic-survey.md.template`.\n\nStructure:\n\n1. **YAML frontmatter** with sweep metadata (slug, time window, venue whitelist used, brief path, worker session id if available, total papers in survey, count of `code: not-found`)\n2. **Summary table** listing all papers with the essential columns (title, venue\u002Fyear, code link or `not-found`, salience_axes, verified_by)\n3. **Per-paper sections** with the full C-S5 fields\n4. **Coverage notes** (optional): brief paragraph on which sub-axes are well covered and which are sparse. Factual, not evaluative.\n5. **Dropped candidates appendix**: short list of `needs-review` venue cases and `dropped_for_unavailable` papers, so the dispatcher can decide next steps\n\n### Step 8 — Log the sweep\n\nWrite a log entry to `\u003Cproject>\u002Flog\u002Fentries\u002F\u003Cts>_survey-sweep_\u003Cslug>.md` using `templates\u002Flog-entry-survey-sweep.md.template`. Type is `survey-sweep` (new log type, must be added to research-workflow SKILL.md's log type vocabulary — see research-workflow Section 5).\n\nSummary should state: N papers in survey, M with `code: not-found`, K flagged `needs-review` for venue.\n\n### Step 9 — Final message to dispatcher\n\nThe worker's final message to the dispatcher (the main agent or routine controller) is short and structured:\n\n- Path to `survey\u002F\u003Cslug>.md`\n- Path to `survey\u002F\u003Cslug>.candidates.md`\n- Path to the log entry\n- Counts: total papers, code not-found, venue needs-review\n- Any papers requiring `verified_by` validation before being merged into the project-root `survey.md` via Workflow E\n\nThe dispatcher (not the worker) decides when and how to merge sub-topic surveys into the project-root `survey.md`.\n\n---\n\n## Section 4 — Output Format\n\nThe detailed format lives in `templates\u002Fsubtopic-survey.md.template`. Highlights:\n\n- **YAML frontmatter** carries machine-readable sweep metadata so a downstream merge script can verify counts and constraints\n- **Each paper has both a row in the summary table and a section below** with full fields — the table is for scanning, the sections are for analysis\n- **Every entry has `verified_by: cache\u002Fpapers\u002F\u003Cstable-id>.pdf`** — same convention as research-workflow C4, so Workflow E can promote entries to the root `survey.md` without re-verification\n- **No subjective fields anywhere** in the schema — there is no \"importance score\", \"collision likelihood\", or \"recommended reading order\" column\n\n---\n\n## Section 5 — Dispatch Mode\n\nThe same brief and worker behavior are designed for in-session Codex sub-agents.\n\nThe main agent, during a research-workflow session, hands off a sub-topic sweep:\n\n1. Main agent drafts `\u003Cproject>\u002Fsurvey\u002F\u003Cslug>.brief.md` (using the template; main agent has full context to fill it well)\n2. Main agent spawns a Codex sub-agent with this prompt:\n   > \"You are a research-survey worker. Read `~\u002F.codex\u002Fskills\u002Fresearch-survey\u002FSKILL.md` for the protocol. Your brief is at `\u003Cabsolute-path>\u002Fsurvey\u002F\u003Cslug>.brief.md`. Execute Section 3 end-to-end and return the Section 3 Step 9 final message.\"\n3. Multiple sweeps in parallel → one message with multiple Codex sub-agent tool calls\n4. Worker returns the structured final message; main agent reads `survey\u002F\u003Cslug>.md` and decides on Workflow E promotions\n\n### Sub-agent vs main agent capability differences (operational facts the dispatcher must know)\n\n- **Sub-agent cannot call `ask the user`**. The brief must be complete; if it's not, the worker returns a failure receipt.\n- **Sub-agent cannot launch its own sub-agents**. It does PDF fetch and code search serially. Don't make sweeps too large.\n- **Sub-agent has fresh context**. It does not see the conversation, MEMORY.md, or any prior decisions. Everything it needs must be in the brief or reachable from disk via the brief.\n- **Parallel sweeps don't share state**. Two workers on adjacent sub-topics will likely re-discover the same paper. Dedup at the dispatcher level using stable IDs after both return.\n\n---\n\n## Section 6 — Anti-patterns\n\n| Anti-pattern | Symptom | Correction |\n|---|---|---|\n| **Abstract-only entries** | Entry exists in `survey\u002F\u003Cslug>.md` but PDF was never fetched | Reject. PDF must be in `cache\u002Fpapers\u002F` with `verified_by` set. Apply C-S3 + research-workflow C4. |\n| **Silent venue inclusion** | A workshop \u002F non-CCF-A paper appears in the survey without justification | Reject. Apply C-S2 — either it matches `accepted_venues` or it goes to `needs-review`, not into the survey. |\n| **`code: not-found` without trace** | Code field marked missing but no `search_trace` recorded | Reject. Apply C-S4 — three-step search is mandatory; the trace proves the search happened. |\n| **arXiv-by-default** | Worker grabs arXiv URL without trying the publication version | Apply C-S3 — try publication first, record the attempt; arXiv is a fallback, not a default. |\n| **Value judgments leaking in** | Entries use \"SOTA\", \"seminal\", \"likely competitor\", \"weak baseline\" | Strip. Apply C-S6 — describe, don't evaluate. The analyst evaluates. |\n| **Rolling time window** | Sweep done in month 10 uses a different \"5 years\" than sweep done in month 1 of the same project | Apply Section 1 — `time_window_start` is the project's creation date, fixed for the project's life. |\n| **Sub-agent guessing missing brief fields** | Worker assumed defaults instead of returning a failure receipt | Reject. Apply Section 1 — if the brief is incomplete, fail loudly, don't paper over. |\n| **Mega-sweep** | A single sub-topic has 40+ papers in the survey | Split the sub-topic into 2–3 narrower briefs. Workers don't read 40 PDFs well. |\n| **Re-discovering papers across parallel sweeps without dedup** | The same paper appears in two sub-topic surveys with different `what` writeups | Dedup at the dispatcher using stable IDs (DBLP > arXiv > DOI). Don't change the worker; fix the merge. |\n| **Drift between modes** | A brief works in Mode A but not Mode B (or vice versa) because someone added a side-channel assumption | Keep the brief format identical. If a mode needs extra context, add it as an optional brief field, not as out-of-band knowledge. |\n\n---\n\n## Section 7 — Integration with research-workflow\n\nThis skill depends on changes to research-workflow SKILL.md. Those changes are:\n\n1. **New log type**: `survey-sweep` added to Section 5's type vocabulary\n2. **New workflow**: Workflow E' (Surveying a sub-topic) added to Section 3, with the entry condition \"the request is a sub-area sweep, not a single-paper intake\"\n3. **Routing table**: Section 6 gains a row \"Surveying a sub-topic → research-survey\"\n4. **Directory layout**: Section 2 gains the `survey\u002F` subdirectory under each project (sibling of `cache\u002F`, `log\u002F`, `doc\u002F`, `src\u002F`). `survey\u002F` is in git; `cache\u002Fpapers\u002F` already isn't.\n5. **Workflow E reference**: Workflow E gains a one-line note: \"If the user mentions a *sub-area* rather than a specific paper, route to research-survey first; Workflow E then promotes individual entries from `survey\u002F\u003Cslug>.md` into the project-root `survey.md`.\"\n\nPromotion path: when the project lead decides a sub-topic survey is ready to fold into the master related-work list, each row of `survey\u002F\u003Cslug>.md` becomes a Workflow E call. The `verified_by` field is already there (same convention), so promotion is a copy, not a re-verification.\n\n---\n\n## Notes for the agent\n\n- This skill is meant to be readable and executable by an isolated worker. Keep procedures concrete; assume no prior conversation.\n- The constitution rules (C-S1 ~ C-S6) are non-negotiable for survey entries. The workflow steps (Section 3) are the default procedure; minor adaptations are fine (e.g., a different discovery order for an unusual sub-area) but the constitution must hold.\n- When in doubt about a venue, mark `needs-review` and surface in the receipt. Do not silently include or exclude.\n- When in doubt about objectivity, ask: \"would this sentence still be true if a hostile reviewer read it?\" If the sentence is a judgment that depends on the reader's priors, rephrase as a fact.\n- The dispatcher (main agent or controller), not the worker, decides when sub-topic surveys are promoted into the project-root `survey.md`. The worker's job ends at Step 9.\n","# research-survey — Surveying a Sub-Topic for a research-workflow Project\n\nThis skill is a **sub-skill of `research-workflow`**. It does not redefine the research constitution — it inherits from research-workflow SKILL.md — and it does not handle single-paper intake (that is research-workflow Workflow E). It handles the case where the task is **\"go investigate sub-area X and bring back a structured survey\"**.\n\nThe skill is designed to be runnable by **isolated sub-agents** (no ability to ask the user back). It therefore enforces an explicit **input contract**: the dispatcher must provide every field the worker needs up-front. If the brief is incomplete, the worker returns a failure receipt rather than guessing.\n\n---\n\n## When to Use\n\n- The main agent or user says: \"调研一下 X 这个子方向 \u002F sweep the literature on X \u002F survey related work on X\"\n- research-workflow Workflow E's step 1 reveals the request is not a single paper but a sub-area\n- A `survey\u002F\u003Cslug>.brief.md` file is handed to a worker (sub-agent or remote routine)\n- Inside an **auto research-workflow** run, the controller dispatches sub-topic sweeps in parallel\n\n## When NOT to Use\n\n- A specific paper is already named and needs recording → research-workflow **Workflow E** (single-paper intake)\n- Early brainstorming with no target sub-topic yet → `brainstorming-research-ideas`\n- The project lives in `~\u002FDocuments\u002Fproject\u002F00-QuickReview\u002F` (no project constitution applies; do whatever is convenient)\n- Citation \u002F BibTeX verification during paper writing → `20-ml-paper-writing`\n\n---\n\n## Section 1 — Input Contract (the brief)\n\nA sub-agent cannot call `ask the user`. To avoid mid-task guessing, every sweep starts from a **brief** that the dispatcher fills out. The brief is a markdown file under `\u003Cproject>\u002Fsurvey\u002F\u003Cslug>.brief.md`, generated from `templates\u002Fsubtopic-brief.md.template`.\n\n### Required fields\n\n| Field | Meaning | Example |\n|---|---|---|\n| `subtopic` | One-line description of the sub-area | \"Advantage normalization techniques in GRPO-style RL fine-tuning\" |\n| `slug` | kebab-case identifier, used in filenames | `grpo-adv-norm` |\n| `parent_proposal_ref` | Relative path (or `proposal.md#section`) the worker reads to learn the main topic | `..\u002Fproposal.md#section-3-method` |\n| `axes_of_interest` | Comma list of dimensions the worker should compare against the parent topic | `task, method, dataset, metric` |\n| `time_window_start` | ISO date, defaults to the project's creation date (NOT today). Fixed across the project's life. | `2026-05-14` (means: papers published on\u002Fafter this date, looking back 5 years → `2021-05-14`) |\n| `venue_whitelist_extra` | Project-specific extras beyond CCF-A (read from project AGENTS.md `accepted_venues`) | `ICLR, COLM, TMLR` |\n| `target_count` | Soft cap on number of papers in the final survey (recommend 8–15; if a sub-topic needs >20, the dispatcher should split it) | `12` |\n| `output_path` | Where to write the survey | `survey\u002Fgrpo-adv-norm.md` |\n\n### Optional fields\n\n| Field | Meaning |\n|---|---|\n| `seed_papers` | Known-relevant papers the dispatcher already has, used as anchors for cite-chain search |\n| `exclude_papers` | DBLP keys \u002F arXiv ids to skip (already in main survey.md, or known irrelevant) |\n| `recency_override` | If set, expands time window beyond 5 years (with reason, e.g., \"include foundational 2017 paper X\") |\n| `notes_for_worker` | Free-form extra context (subtle distinctions, terminology the worker should not confuse) |\n\n### Time window semantics (per the project decision)\n\n- `time_window_start` is the **project's creation date**, recorded once and reused.\n- The accepted publication window is `[time_window_start - 5 years, today]`.\n- Rationale: a survey done in month 1 and a survey done in month 10 of the same project should agree on what \"近五年\" means. Rolling windows break this.\n- If the worker needs to break the 5-year wall, the dispatcher must set `recency_override` in the brief. The worker does NOT decide this on its own.\n\n### If the brief is incomplete\n\nThe worker does **not** guess. It writes a failure receipt (see `templates\u002Ffailure-receipt.md.template`) listing the missing fields and stops. The dispatcher fixes the brief and re-dispatches.\n\n---\n\n## Section 2 — The Survey Constitution (C-S1 ~ C-S6)\n\nThese rules govern every sweep. They sit **below** the research-workflow constitution (C1–C8) and inherit from it; where this skill is silent, research-workflow applies.\n\n### C-S1. Recency: ≥ (project-creation-date − 5 years)\n\nDefault: only papers published in the 5-year window described in Section 1.\n\n- **Why**: surveys for active research projects need current state of the art, not historical foundations.\n- **How to apply**: filter candidates by publication year against `time_window_start − 5y`. Older papers go to a `dropped_for_recency` list at the bottom of the candidates file, not into the survey.\n- **Override**: only when `recency_override` is set in the brief.\n\n### C-S2. Venue: CCF-A by default, plus project-declared extras\n\n- **Default whitelist**: CCF-A conferences and journals across the relevant area (AI\u002FML: NeurIPS, ICML, AAAI, IJCAI, ACL, EMNLP, NAACL, CVPR, ICCV, ECCV, TPAMI, IJCV, JMLR, TOG, SIGGRAPH; SE: ICSE, FSE, ASE, ISSTA, OOPSLA, TSE, TOSEM; Sys: SOSP, OSDI, ASPLOS, ISCA, MICRO, HPCA, SIGCOMM, NSDI, MobiCom, FAST, EuroSys; DB: SIGMOD, VLDB, ICDE; Sec: CCS, S&P, USENIX Security, NDSS; Theory: STOC, FOCS, SODA, LICS, CAV; HCI: CHI, UIST, CSCW). When in doubt about a venue's tier, **do not silently include it** — the worker writes the candidate to the candidates file with `venue_justification: needs-review` and surfaces it in the final receipt for the dispatcher to decide.\n- **Project extras**: the project's `AGENTS.md` may declare an `accepted_venues:` field naming additional venues the project considers acceptable (ICLR, COLM, TMLR are common examples). The worker reads this and merges it with the CCF-A default. The brief's `venue_whitelist_extra` overrides per-sweep.\n- **Workshop papers, arXiv-only papers, technical reports**: not acceptable as primary entries. They may appear as `discovered_via` references for context but do not get rows in the survey.\n- **Why**: keeps the sweep anchored to peer-reviewed venues the project lead trusts, while allowing per-project flexibility.\n\n### C-S3. Version: publication > arXiv\n\n- **Preferred**: the venue's published version (PDF from the publisher's site, the conference proceedings page, or the journal page).\n- **Acceptable fallback**: if the publication version is paywalled and not retrievable, use the arXiv version. Record `version: arxiv-fallback` and `fallback_reason: paywalled` on the entry.\n- **Not acceptable**: using arXiv when the publication version is freely available. The worker must attempt the publication URL first and record the attempt.\n- **Why**: the publication version is the one of record; reviewers will compare to it.\n\n### C-S4. Code: open-source repository required (or proof of absence)\n\nEvery survey entry must carry either:\n\n- `code: \u003Cgithub-url>` — verified to exist and to correspond to the paper, OR\n- `code: not-found` plus a `search_trace` showing the three search steps were performed:\n  1. Paper-side: searched the paper PDF (abstract, footnotes, intro, \"Code available at\", appendix) for any URL\n  2. Author-side: checked the first\u002Flast author's GitHub profile for a repo matching the paper title or method name\n  3. Topic-side: searched GitHub for `\u003Cmethod-name>` and `\u003Cpaper-keywords>` to find third-party reimplementations (if even a third-party repo exists, log it but mark `code: third-party-only`)\n\n`code: not-found` without a `search_trace` is treated as missing data — the entry is rejected from the survey.\n\n### C-S5. Content fields (per entry)\n\nEvery entry must contain these fields, factually stated:\n\n- `what`: what the paper does (1–3 sentences, descriptive only)\n- `goal`: the stated purpose \u002F problem it tries to solve (in the paper's own framing)\n- `relation_to_parent`: how this work connects to the parent project's topic, across the `axes_of_interest`\n- `differences`: in what dimensions this work diverges from the parent topic (task \u002F method \u002F setting \u002F data \u002F metric — be specific)\n- `salience_axes`: which of the brief's `axes_of_interest` this paper **actually touches** (this is factual labeling, not a judgment). Example: `[task, dataset]` if the paper shares the task and dataset but not the method.\n\n### C-S6. Objectivity: no value judgments\n\nThe worker **describes facts**. The worker does **not**:\n\n- Rank papers by importance, relevance, novelty, or \"collision risk\" with the parent topic\n- Use evaluative language: \"the most likely to overlap\", \"the most worth learning from\", \"SOTA\", \"seminal\", \"groundbreaking\", \"weak\", \"strong\"\n- Recommend which paper the user should read first\n\nThe downstream analyst (human or another agent) decides what is salient. The worker's job is to produce a clean, complete, factual record so that downstream judgment is well-supported.\n\nStating that a paper \"shares dataset X with our proposal\" is factual. Stating that it \"is likely to be a direct competitor\" is not. Keep to the first kind.\n\n---\n\n## Section 3 — Workflow (a sweep, end to end)\n\nThis is the worker's procedure. Each step is concrete enough that a sub-agent with no prior context can execute it given only the brief and access to `research-workflow` SKILL.md.\n\n### Step 0 — Read brief and verify completeness\n\n1. Read `\u003Cproject>\u002Fsurvey\u002F\u003Cslug>.brief.md`\n2. Verify all required fields (Section 1) are present and non-empty\n3. Read project `AGENTS.md` to fetch `accepted_venues`\n4. If any required field missing → write failure receipt → stop\n\n### Step 1 — Read the parent proposal\n\nRead the `parent_proposal_ref` target. Extract:\n\n- The parent topic in one sentence\n- The specific dimensions named in `axes_of_interest` (e.g., what task, what method family, what dataset)\n- Vocabulary the parent uses (to align terminology in `relation_to_parent` \u002F `differences` later)\n\n### Step 2 — Discovery (cast a wide net)\n\nGoal: collect 30–50 candidates with metadata only (no PDF download yet).\n\nSources, in order of trust:\n\n1. **Venue ToCs**: for each venue in the whitelist, scan the relevant year's accepted papers list for keyword matches\n2. **DBLP**: author \u002F keyword queries scoped to the venue whitelist and time window\n3. **Google Scholar**: keyword search with `venue:` and year filters; useful for finding the publication version of a paper you only know from arXiv\n4. **Cite-chain from seed_papers**: forward and backward citations of any `seed_papers` in the brief\n5. **arXiv**: only for finding the publication-version URL of a recent paper; never as the primary source\n\nFor each candidate, record metadata in `\u003Cproject>\u002Fsurvey\u002F\u003Cslug>.candidates.md` (template provided):\n\n- Title, authors, venue, year\n- Stable ID (DBLP key preferred, else arXiv id, else DOI)\n- `discovered_via`: which source surfaced it\n- Abstract (copied verbatim)\n- Provisional `venue_status`: in-whitelist \u002F project-extra \u002F needs-review \u002F out-of-scope\n\n### Step 3 — Filter candidates\n\nApply, in order:\n\n1. **Venue filter** (C-S2): drop `out-of-scope`; flag `needs-review` for the dispatcher\n2. **Time filter** (C-S1): drop anything outside the window\n3. **Relevance filter** (abstract-level): drop anything whose abstract clearly does not touch the parent topic. Be permissive at this stage; better to read one extra PDF than to miss a relevant paper\n4. **Deduplication**: a paper that appears in both arXiv form and conference form is one entry, keyed by the conference DBLP key; merge `discovered_via` lists\n5. **Exclude list**: drop anything in the brief's `exclude_papers`\n\nAim to come out of this step with ~15–25 papers that warrant a PDF read.\n\n### Step 4 — Fetch publication PDFs\n\nFor each surviving candidate:\n\n1. Try the publication URL first (conference proceedings page, journal page, publisher PDF)\n2. If paywalled \u002F 403 \u002F unavailable after a reasonable attempt → fall back to arXiv version, record `version: arxiv-fallback` and the reason\n3. Save the PDF to `\u003Cproject>\u002Fcache\u002Fpapers\u002F\u003Cstable-id>.pdf`\n4. If both publication and arXiv fail → drop from this sweep, log in candidates file as `dropped_for_unavailable`\n\n`cache\u002Fpapers\u002F` is the same cache that research-workflow C4 uses. PDFs landed here become available for subsequent single-paper Workflow E intake.\n\n### Step 5 — Code repository hunt\n\nFor each paper whose PDF landed:\n\n1. **Paper-side search**: open the PDF, search for `github.com`, `gitlab.com`, \"Code available\", \"Implementation\", footnotes on page 1, the abstract page, and the appendix. Capture the URL.\n2. **Author-side search**: if not found, look up the first author and last author on GitHub (try `github.com\u002F\u003Clastname>`, then a GitHub user search). Scan their repos for one matching the paper title or method name.\n3. **Topic-side search**: if still not found, search GitHub for `\u003Cmethod-name>` and 2–3 distinctive keywords from the paper. If only third-party reimplementations exist, record one and mark `code: third-party-only`.\n4. **Verification**: open the candidate repo's README. Confirm it cites this paper or implements its method (don't just match on name). Record the verification.\n\nOutcomes:\n\n- `code: \u003Curl>` (verified official) — record `code_status: official-verified`\n- `code: \u003Curl>` (third-party only) — record `code_status: third-party-only`\n- `code: not-found` — record the full `search_trace` (what was tried, what was found)\n\n### Step 6 — Read and extract\n\nFor each paper, read at minimum: abstract, intro (especially the contributions list), method overview, experiments setup, conclusion. Extract the C-S5 content fields. Stick to descriptive language (C-S6).\n\nCross-reference with the parent proposal vocabulary (from Step 1) so `relation_to_parent` and `differences` are anchored in the project's own terms, not the paper's.\n\n### Step 7 — Write the survey file\n\nWrite to `\u003Cproject>\u002Fsurvey\u002F\u003Cslug>.md` using `templates\u002Fsubtopic-survey.md.template`.\n\nStructure:\n\n1. **YAML frontmatter** with sweep metadata (slug, time window, venue whitelist used, brief path, worker session id if available, total papers in survey, count of `code: not-found`)\n2. **Summary table** listing all papers with the essential columns (title, venue\u002Fyear, code link or `not-found`, salience_axes, verified_by)\n3. **Per-paper sections** with the full C-S5 fields\n4. **Coverage notes** (optional): brief paragraph on which sub-axes are well covered and which are sparse. Factual, not evaluative.\n5. **Dropped candidates appendix**: short list of `needs-review` venue cases and `dropped_for_unavailable` papers, so the dispatcher can decide next steps\n\n### Step 8 — Log the sweep\n\nWrite a log entry to `\u003Cproject>\u002Flog\u002Fentries\u002F\u003Cts>_survey-sweep_\u003Cslug>.md` using `templates\u002Flog-entry-survey-sweep.md.template`. Type is `survey-sweep` (new log type, must be added to research-workflow SKILL.md's log type vocabulary — see research-workflow Section 5).\n\nSummary should state: N papers in survey, M with `code: not-found`, K flagged `needs-review` for venue.\n\n### Step 9 — Final message to dispatcher\n\nThe worker's final message to the dispatcher (the main agent or routine controller) is short and structured:\n\n- Path to `survey\u002F\u003Cslug>.md`\n- Path to `survey\u002F\u003Cslug>.candidates.md`\n- Path to the log entry\n- Counts: total papers, code not-found, venue needs-review\n- Any papers requiring `verified_by` validation before being merged into the project-root `survey.md` via Workflow E\n\nThe dispatcher (not the worker) decides when and how to merge sub-topic surveys into the project-root `survey.md`.\n\n---\n\n## Section 4 — Output Format\n\nThe detailed format lives in `templates\u002Fsubtopic-survey.md.template`. Highlights:\n\n- **YAML frontmatter** carries machine-readable sweep metadata so a downstream merge script can verify counts and constraints\n- **Each paper has both a row in the summary table and a section below** with full fields — the table is for scanning, the sections are for analysis\n- **Every entry has `verified_by: cache\u002Fpapers\u002F\u003Cstable-id>.pdf`** — same convention as research-workflow C4, so Workflow E can promote entries to the root `survey.md` without re-verification\n- **No subjective fields anywhere** in the schema — there is no \"importance score\", \"collision likelihood\", or \"recommended reading order\" column\n\n---\n\n## Section 5 — Dispatch Mode\n\nThe same brief and worker behavior are designed for in-session Codex sub-agents.\n\nThe main agent, during a research-workflow session, hands off a sub-topic sweep:\n\n1. Main agent drafts `\u003Cproject>\u002Fsurvey\u002F\u003Cslug>.brief.md` (using the template; main agent has full context to fill it well)\n2. Main agent spawns a Codex sub-agent with this prompt:\n   > \"You are a research-survey worker. Read `~\u002F.codex\u002Fskills\u002Fresearch-survey\u002FSKILL.md` for the protocol. Your brief is at `\u003Cabsolute-path>\u002Fsurvey\u002F\u003Cslug>.brief.md`. Execute Section 3 end-to-end and return the Section 3 Step 9 final message.\"\n3. Multiple sweeps in parallel → one message with multiple Codex sub-agent tool calls\n4. Worker returns the structured final message; main agent reads `survey\u002F\u003Cslug>.md` and decides on Workflow E promotions\n\n### Sub-agent vs main agent capability differences (operational facts the dispatcher must know)\n\n- **Sub-agent cannot call `ask the user`**. The brief must be complete; if it's not, the worker returns a failure receipt.\n- **Sub-agent cannot launch its own sub-agents**. It does PDF fetch and code search serially. Don't make sweeps too large.\n- **Sub-agent has fresh context**. It does not see the conversation, MEMORY.md, or any prior decisions. Everything it needs must be in the brief or reachable from disk via the brief.\n- **Parallel sweeps don't share state**. Two workers on adjacent sub-topics will likely re-discover the same paper. Dedup at the dispatcher level using stable IDs after both return.\n\n---\n\n## Section 6 — Anti-patterns\n\n| Anti-pattern | Symptom | Correction |\n|---|---|---|\n| **Abstract-only entries** | Entry exists in `survey\u002F\u003Cslug>.md` but PDF was never fetched | Reject. PDF must be in `cache\u002Fpapers\u002F` with `verified_by` set. Apply C-S3 + research-workflow C4. |\n| **Silent venue inclusion** | A workshop \u002F non-CCF-A paper appears in the survey without justification | Reject. Apply C-S2 — either it matches `accepted_venues` or it goes to `needs-review`, not into the survey. |\n| **`code: not-found` without trace** | Code field marked missing but no `search_trace` recorded | Reject. Apply C-S4 — three-step search is mandatory; the trace proves the search happened. |\n| **arXiv-by-default** | Worker grabs arXiv URL without trying the publication version | Apply C-S3 — try publication first, record the attempt; arXiv is a fallback, not a default. |\n| **Value judgments leaking in** | Entries use \"SOTA\", \"seminal\", \"likely competitor\", \"weak baseline\" | Strip. Apply C-S6 — describe, don't evaluate. The analyst evaluates. |\n| **Rolling time window** | Sweep done in month 10 uses a different \"5 years\" than sweep done in month 1 of the same project | Apply Section 1 — `time_window_start` is the project's creation date, fixed for the project's life. |\n| **Sub-agent guessing missing brief fields** | Worker assumed defaults instead of returning a failure receipt | Reject. Apply Section 1 — if the brief is incomplete, fail loudly, don't paper over. |\n| **Mega-sweep** | A single sub-topic has 40+ papers in the survey | Split the sub-topic into 2–3 narrower briefs. Workers don't read 40 PDFs well. |\n| **Re-discovering papers across parallel sweeps without dedup** | The same paper appears in two sub-topic surveys with different `what` writeups | Dedup at the dispatcher using stable IDs (DBLP > arXiv > DOI). Don't change the worker; fix the merge. |\n| **Drift between modes** | A brief works in Mode A but not Mode B (or vice versa) because someone added a side-channel assumption | Keep the brief format identical. If a mode needs extra context, add it as an optional brief field, not as out-of-band knowledge. |\n\n---\n\n## Section 7 — Integration with research-workflow\n\nThis skill depends on changes to research-workflow SKILL.md. Those changes are:\n\n1. **New log type**: `survey-sweep` added to Section 5's type vocabulary\n2. **New workflow**: Workflow E' (Surveying a sub-topic) added to Section 3, with the entry condition \"the request is a sub-area sweep, not a single-paper intake\"\n3. **Routing table**: Section 6 gains a row \"Surveying a sub-topic → research-survey\"\n4. **Directory layout**: Section 2 gains the `survey\u002F` subdirectory under each project (sibling of `cache\u002F`, `log\u002F`, `doc\u002F`, `src\u002F`). `survey\u002F` is in git; `cache\u002Fpapers\u002F` already isn't.\n5. **Workflow E reference**: Workflow E gains a one-line note: \"If the user mentions a *sub-area* rather than a specific paper, route to research-survey first; Workflow E then promotes individual entries from `survey\u002F\u003Cslug>.md` into the project-root `survey.md`.\"\n\nPromotion path: when the project lead decides a sub-topic survey is ready to fold into the master related-work list, each row of `survey\u002F\u003Cslug>.md` becomes a Workflow E call. The `verified_by` field is already there (same convention), so promotion is a copy, not a re-verification.\n\n---\n\n## Notes for the agent\n\n- This skill is meant to be readable and executable by an isolated worker. Keep procedures concrete; assume no prior conversation.\n- The constitution rules (C-S1 ~ C-S6) are non-negotiable for survey entries. The workflow steps (Section 3) are the default procedure; minor adaptations are fine (e.g., a different discovery order for an unusual sub-area) but the constitution must hold.\n- When in doubt about a venue, mark `needs-review` and surface in the receipt. Do not silently include or exclude.\n- When in doubt about objectivity, ask: \"would this sentence still be true if a hostile reviewer read it?\" If the sentence is a judgment that depends on the reader's priors, rephrase as a fact.\n- The dispatcher (main agent or controller), not the worker, decides when sub-topic surveys are promoted into the project-root `survey.md`. The worker's job ends at Step 9.\n",{"name":5,"license":15,"description":7},"main","7d4c66b07159d74c8fa9cb32475f008f8d36fe0e","plugins\u002Fcodex-research-kit\u002Fskills\u002Fresearch-survey","https:\u002F\u002Fgithub.com\u002FSiyuexi\u002Fcodex-research-kit","https:\u002F\u002Fgithub.com\u002FSiyuexi","siyuexi.github.io","Ph.D., Chinese University of Hongkong, Shenzhen.","2026-06-04 14:03:14",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add Siyuexi\u002Fcodex-research-kit --skill research-survey",{"tool":38,"command":39},"bunx","bunx skills add Siyuexi\u002Fcodex-research-kit --skill research-survey",{"tool":41,"command":42},"pnpm","pnpm dlx skills add Siyuexi\u002Fcodex-research-kit --skill research-survey","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FSiyuexi\u002Fcodex-research-kit\u002Fblob\u002Fmain\u002Fplugins\u002Fcodex-research-kit\u002Fskills\u002Fresearch-survey\u002FSKILL.md",true,[46,52,58,62,66,70],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"agents\u002Fopenai.yaml","REFERENCE","参考文档",241,"https:\u002F\u002Fgithub.com\u002FSiyuexi\u002Fcodex-research-kit\u002Fblob\u002Fmain\u002Fplugins\u002Fcodex-research-kit\u002Fskills\u002Fresearch-survey\u002Fagents\u002Fopenai.yaml",{"path":53,"resourceType":54,"resourceTypeDesc":55,"sizeBytes":56,"sourceHtmlUrl":57},"templates\u002Fcandidates.md.template","OTHER","其他",1674,"https:\u002F\u002Fgithub.com\u002FSiyuexi\u002Fcodex-research-kit\u002Fblob\u002Fmain\u002Fplugins\u002Fcodex-research-kit\u002Fskills\u002Fresearch-survey\u002Ftemplates\u002Fcandidates.md.template",{"path":59,"resourceType":54,"resourceTypeDesc":55,"sizeBytes":60,"sourceHtmlUrl":61},"templates\u002Ffailure-receipt.md.template",1512,"https:\u002F\u002Fgithub.com\u002FSiyuexi\u002Fcodex-research-kit\u002Fblob\u002Fmain\u002Fplugins\u002Fcodex-research-kit\u002Fskills\u002Fresearch-survey\u002Ftemplates\u002Ffailure-receipt.md.template",{"path":63,"resourceType":54,"resourceTypeDesc":55,"sizeBytes":64,"sourceHtmlUrl":65},"templates\u002Flog-entry-survey-sweep.md.template",1412,"https:\u002F\u002Fgithub.com\u002FSiyuexi\u002Fcodex-research-kit\u002Fblob\u002Fmain\u002Fplugins\u002Fcodex-research-kit\u002Fskills\u002Fresearch-survey\u002Ftemplates\u002Flog-entry-survey-sweep.md.template",{"path":67,"resourceType":54,"resourceTypeDesc":55,"sizeBytes":68,"sourceHtmlUrl":69},"templates\u002Fsubtopic-brief.md.template",2337,"https:\u002F\u002Fgithub.com\u002FSiyuexi\u002Fcodex-research-kit\u002Fblob\u002Fmain\u002Fplugins\u002Fcodex-research-kit\u002Fskills\u002Fresearch-survey\u002Ftemplates\u002Fsubtopic-brief.md.template",{"path":71,"resourceType":54,"resourceTypeDesc":55,"sizeBytes":72,"sourceHtmlUrl":73},"templates\u002Fsubtopic-survey.md.template",3807,"https:\u002F\u002Fgithub.com\u002FSiyuexi\u002Fcodex-research-kit\u002Fblob\u002Fmain\u002Fplugins\u002Fcodex-research-kit\u002Fskills\u002Fresearch-survey\u002Ftemplates\u002Fsubtopic-survey.md.template",[75,82,89,96,103,110],{"skillNo":76,"name":77,"nameZh":6,"description":78,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":79,"categories":80,"keywords":81,"publishedAt":19},"01M74GQZRN00Z8X286V57MRQJZ","vibe-survey","Compatibility alias for research-survey. Use when the user says vibe-survey or asks for the old \u002Fvibe-survey behavior.","https:\u002F\u002Fgithub.com\u002FSiyuexi\u002Fcodex-research-kit\u002Fblob\u002Fmain\u002Fplugins\u002Fcodex-research-kit\u002Fskills\u002Fvibe-survey\u002FSKILL.md",[],[],{"skillNo":83,"name":84,"nameZh":6,"description":85,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":86,"categories":87,"keywords":88,"publishedAt":19},"01M74GQY8M79RG5N901D6YQ4F9","vibe-review","Compatibility alias for research-review. Use when the user says vibe-review or asks for the old \u002Fvibe-review behavior.","https:\u002F\u002Fgithub.com\u002FSiyuexi\u002Fcodex-research-kit\u002Fblob\u002Fmain\u002Fplugins\u002Fcodex-research-kit\u002Fskills\u002Fvibe-review\u002FSKILL.md",[],[],{"skillNo":90,"name":91,"nameZh":6,"description":92,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":93,"categories":94,"keywords":95,"publishedAt":19},"01M74GQVHX1BBTCC4SRS31G18H","vibe-research","Compatibility alias for research-workflow. Use when the user says vibe-research or asks for the old \u002Fvibe-research behavior.","https:\u002F\u002Fgithub.com\u002FSiyuexi\u002Fcodex-research-kit\u002Fblob\u002Fmain\u002Fplugins\u002Fcodex-research-kit\u002Fskills\u002Fvibe-research\u002FSKILL.md",[],[],{"skillNo":97,"name":98,"nameZh":6,"description":99,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":100,"categories":101,"keywords":102,"publishedAt":19},"01M74GQQRT2EF984V3VEV373NN","vibe-report","VRC report support and deployment skill. Use when the controller asks for a visual web report, deployed report, Feishu-readable report page, report template, or Bridge \u002Freport workflow. Codex defaults to data\u002Fbackend validation and deployment; Claude defaults to HTML\u002Ffront-end authoring.","https:\u002F\u002Fgithub.com\u002FSiyuexi\u002Fcodex-research-kit\u002Fblob\u002Fmain\u002Fplugins\u002Fcodex-research-kit\u002Fskills\u002Fvibe-report\u002FSKILL.md",[],[],{"skillNo":104,"name":105,"nameZh":6,"description":106,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":107,"categories":108,"keywords":109,"publishedAt":19},"01M74GQMHST4D5QD2EBKBZSK83","vibe-discuss","Feishu-mediated peer collaboration protocol for Claude and Codex on research-workflow projects. Use when the user asks to set up or operate a topic\u002Fproject group, review or implement the protocol in protocol.md, coordinate Claude\u002FCodex discussion, log consensus\u002Fdebate\u002Fhandoff outcomes, manage shared workspace write discipline, run independent cross-agent reviews, or start\u002Finspect the optional full-auto worktree sandbox.","https:\u002F\u002Fgithub.com\u002FSiyuexi\u002Fcodex-research-kit\u002Fblob\u002Fmain\u002Fplugins\u002Fcodex-research-kit\u002Fskills\u002Fvibe-discuss\u002FSKILL.md",[],[],{"skillNo":111,"name":112,"nameZh":6,"description":113,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":114,"categories":115,"keywords":116,"publishedAt":19},"01M74GQHT45KQT8GE88ZPBQXKA","vibe-cowork","Compatibility alias for research-cowork. Use when the user says vibe-cowork or asks for the old \u002Fvibe-cowork behavior.","https:\u002F\u002Fgithub.com\u002FSiyuexi\u002Fcodex-research-kit\u002Fblob\u002Fmain\u002Fplugins\u002Fcodex-research-kit\u002Fskills\u002Fvibe-cowork\u002FSKILL.md",[],[]]