wensu 0.4.0__tar.gz

wensu-0.4.0/.gitignore

@@ -0,0 +1,42 @@
+# Python
+__pycache__/
+*.py[cod]
+*.egg-info/
+build/
+dist/
+.venv/
+venv/
+
+# Native host build outputs and per-user registration
+native-host/scholar-host.exe
+native-host/python-path.txt
+native-host/com.scholar.search.json
+native-host/run-host.cmd
+
+# Tooling caches
+.pytest_cache/
+.mypy_cache/
+.ruff_cache/
+.coverage
+htmlcov/
+.pre-commit-cache/
+
+# Local runtime data (must never be committed)
+data/
+fixtures/local/
+*.local.json
+*.local.toml
+.env
+.env.*
+
+# Logs
+*.log
+
+# OS / editor
+.DS_Store
+Thumbs.db
+.idea/
+.vscode/
+
+# Local handoff notes (machine-specific paths; not for public history)
+HANDOFF.md

wensu-0.4.0/AGENTS.md

@@ -0,0 +1,37 @@
+# AGENTS.md · 项目硬约束
+
+本文件是给所有在此仓库工作的 agent 的硬性约束。与代码注释冲突时，以本文件为准。
+
+## 不可协商
+
+1. **只读**：不实现、不调用任何 PDF/CAJ 下载接口。允许 RIS/EndNote 题录文本导出。
+2. **不做第三方题录入库**：不实现把题录写入 Zotero、EndNote、Mendeley 等参考文献库的任何接口、
+   桥接或自动化（包括本地 HTTP 端点、库文件改写、数据库直写）。只输出纯文本 RIS/EndNote，
+   导入由用户自己在对应软件里完成。理由见 ADR 0006：自动入库会越过本站的服务条款边界，
+   并把第三方软件的账号与库文件卷进本项目的责任范围。
+3. **遇风控即停**：检测到验证码、`/verify/home`、403、登录重定向，立即终止当前计划，返回 `blocked`，写审计，冷却 24 小时。禁止自动重试、禁止更换 IP、禁止延时后重试。
+4. **不破解验证码**：任何形式的自动识别、打码、绕过都不允许。
+5. **频率硬下限**：最小间隔 3000ms、并发 1、单次 ≤3 页、每日 ≤300 次（ADR 0005）。这些值原则上只能收紧；放宽必须新增 ADR。
+6. **动作词汇封闭**：`goto` `wait_for` `type` `click` `select` `read_html` `read_text` `scroll` `sleep`。新增动词需 ADR 并同步 host 与扩展。
+7. **域名白名单**：由渠道 manifest 声明，内核与 host 双重校验。
+
+## 开发规范
+
+- Python 3.11+，类型全标注，`mypy --strict` 通过。
+- `ruff` 格式化与检查，提交前跑 `pre-commit`。
+- Conventional Commits：`feat:` `fix:` `docs:` `refactor:` `test:` `chore:`。
+- 新增渠道 = 新增 `channels/<name>/<site_version>/` + `manifest.json`，不改内核。
+- 解析逻辑必须有 fixture 回放测试，禁止只靠线上验证。
+- 运行期数据只写 `SCHOLAR_SEARCH_HOME`（默认 `%LOCALAPPDATA%\scholar-search`），不得写入仓库。
+
+## 隐私
+
+- 不记录 Cookie、账号、完整 HTML。
+- 审计日志只记动作、渠道、耗时、结果数、错误码、风控信号。
+- fixture 必须脱敏；仓库零凭证。
+
+## 发布
+
+- `main` 分支保护，tag 触发发布。
+- 版本号、CHANGELOG、NOTICE 三者必须同步。
+

wensu-0.4.0/CHANGELOG.md

@@ -0,0 +1,470 @@
+# Changelog
+
+All notable changes to this project are documented here.
+Format: [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
+Versioning: [Semantic Versioning](https://semver.org/).
+
+## [0.4.0] - 2026-10-23
+
+### Changed
+
+- Distribution renamed from `scholar-search` to `wensu` (文溯). The old name is taken on PyPI
+  by an unrelated package, so `pip install scholar-search` would install someone else's code.
+- The command is now `wensu`; `scholar-search` remains as a compatibility alias, so MCP client
+  configs written before this release keep working.
+- The Python module (`scholar_search`), the runtime directory
+  (`%LOCALAPPDATA%\scholar-search`) and `SCHOLAR_SEARCH_HOME` are unchanged, so upgrades keep
+  quota, cooldown and audit state.
+- Repository and documentation links point at `github.com/hui-xu-ai/wensu`.
+
+## [0.3.19] - 2026-10-21
+
+### Changed
+
+- `NOTICE` now carries a full provenance statement: which public MIT projects informed this
+  one, exactly what was taken (selector strings, step order, known pitfalls) and what was
+  not (every file, module and function here is written for this project), plus the
+  licence of each runtime and development dependency.
+- `README` gained a 来源与致谢 section; `docs/reference-parity.md` no longer calls the
+  selectors "照搬" and points at the provenance statement.
+
+### Fixed
+
+- Test fixtures are now fully desensitised: recorded abstracts, author names,
+  affiliations and keyword labels are replaced with synthetic placeholders, and author
+  mail addresses are removed. Only markup, class names and public bibliographic facts
+  (title, journal, volume/issue/pages, DOI) remain.
+
+### Docs
+
+- Version numbers in examples are placeholders now (`<版本>`) instead of a pinned release,
+  so they cannot go stale; only the status section names a real version.
+- The proxy trap (TLS interception making CNKI show a certificate error) is documented in
+  the README, the Codex guide, the install skill and the incident ledger, including the
+  detail that a command-line fetch bypasses the system proxy and therefore proves nothing.
+
+## [0.3.18] - 2026-10-21
+
+### Fixed
+
+- `core/locking.py` branches on `sys.platform` instead of `os.name`. mypy only narrows
+  the former, so on Linux the Windows branch was analysed too: `msvcrt` came back
+  untyped and the `fcntl` ignores were reported as unused. That made the
+  `ubuntu-latest` jobs fail at the type-check step while Windows passed. Verified with
+  both `mypy src` and `mypy --platform linux src`.
+
+## [0.3.17] - 2026-10-21
+
+### Fixed
+
+- The sort-correction plan's pause now uses the shared reply margin. It still carried the
+  old `sleep 1500 / timeout 2000` pair that 0.3.15 proved too tight for a throttled tab;
+  the plan only runs when the grid comes back ascending, so it had not failed yet.
+
+### Changed
+
+- `settle()` in `channels/cnki/v2026_10/plan_search.py` is public so the sort-correction
+  plan reuses it instead of hand-rolling its own pause budget.
+
+### Docs
+
+- User guide, architecture, upgrade/operations, CNKI-maintenance, browser setup, Zotero
+  (manual import only) and publishing notes; plus a `scholar-install` skill that walks an
+  agent through a fresh install.
+
+## [0.3.16] - 2026-10-21
+
+### Fixed
+
+- Document types are normalised in one place. The result grid labels some rows
+  「中文图书」/「外文期刊」, and the old fallback (`token.casefold()`) pushed that raw
+  Chinese string into `doc_type`, where the export mapping could not recognise it and
+  RIS fell back to `GEN` — a book exported as generic. `parse_search` and
+  `parse_detail` now share `channels/cnki/v2026_10/doc_types.py`, which strips language
+  prefixes, wrappers （《期刊》, （网络首发）） and the 文献类型 label, and returns
+  `None` for a type it cannot map so the export renders generic with a note instead of a
+  meaningless value.
+
+## [0.3.15] - 2026-10-21
+
+### Fixed
+
+- Pause deadlines now carry a 2 s reply margin. In live use the 2.5 s settle after a
+  navigation answered at 3004 ms against a 3000 ms deadline and the plan died with
+  `bridge.timeout`. The cause is timer alignment: while the CNKI window is not the
+  active window the browser buckets timers to whole seconds, so any pause can answer up
+  to ~1 s late. Every `sleep` action now uses `_settle()`, which adds the margin.
+- Worst-case plan budget rebalanced to stay inside the 120 s cap that the extra margins
+  would have broken: `read_html` 6 s → 4 s, grid/filter waits 8 s → 6 s, query-box and
+  result-grid waits 12 s → 10 s. Observed live latency is unchanged (navigation ~1.9 s,
+  grid ~2.1 s), so each deadline still sits 3–5× above it. ADR 0005 limits, the verb
+  set and the plan shape are untouched.
+
+## [0.3.14] - 2026-10-21
+
+### Fixed
+
+- Search and expert-search plans wait 2.5 s after the opening navigation before looking
+  for the query box. CNKI's first load performs a second hop that replaced the content
+  script, so `type` could fail with `bridge.timeout` even though `goto` and `wait_for`
+  had succeeded. Behaviour, limits and verbs are unchanged; the plan now has one extra
+  `sleep` action.
+
+## [0.3.13] - 2026-10-21
+
+### Changed
+
+- Daily network-action ceiling raised from 200 to 300 per ADR 0005. All other limits
+  are unchanged: 3 s floor plus jitter, concurrency 1, page 1–3, `limit` 1–50,
+  40 actions / 120 s per plan, 24 h cooldown.
+- Documented captcha handling: the plan stops and returns `blocked`; the user solves
+  the challenge in the browser, then explicitly clears the cooldown with
+  `scholar-search reset --confirm`. No automatic retry.
+
+## [0.3.12] - 2026-10-21
+
+### Fixed
+
+- Returned `page` now reflects the page that was actually requested, even when the
+  site's own pager marker lags behind the AJAX turn.
+- The 50-per-page switch waits 2.5 s for the grid to rebuild; a fast read returned the
+  previous 20-row page.
+- CNKI dissertation DOIs (`/d.cnki.`) mark a detail record as `thesis` when no explicit
+  type field is present; live verification now exports `THES`/`Thesis`.
+
+## [0.3.11] - 2026-10-21
+
+### Changed
+
+- **Human-like pacing** The kernel now adds 0–1.2 s of random jitter on top of the
+  3 s floor, with a ~10% chance of a 2–4 s pause. Clicks scroll the target into view and
+  fire mousedown/mouseup/click with random 40–90 ms gaps; navigation gets a 0.5–1.5 s
+  dwell before the first action. Typing stays paste-like, as a person would paste a query.
+- **Sort correction** The grid is read after sorting; if the active column is `ASC`, the
+  channel returns one corrective plan and the service runs it, instead of blind
+  double-clicks. This keeps the requested descending order without relying on timing.
+
+## [0.3.10] - 2026-10-21
+
+### Fixed
+
+- **Date sort direction** The CNKI grid already opens on publication-time DESC;
+  the product clicked the tab once, which toggled it to ASC. `sort=date` now sends no
+  sort clicks. Citation sort clicks twice (first activation ASC, second DESC).
+- **Pagination settle** A first page turn now waits 800 ms after the grid renders; the
+  turn was previously clicked while the grid was still binding and was ignored.
+- **Content-script readiness** The background retries `sendMessage` for up to ~1.8 s
+  when a navigation outruns `document_idle` injection, instead of failing with
+  "Receiving end does not exist".
+- Read timeout trimmed to 7 s so every query/page/sort/page-size combination stays
+  within the 120 s plan budget.
+
+## [0.3.9] - 2026-10-21
+
+### Fixed
+
+- **Click delivery** 0.3.8 deferred the click with setTimeout, which Chrome throttles
+  in background tabs; pagination then executed after the read and stayed on page 1.
+  The content script clicks synchronously again, and the background now treats a
+  click whose reply is lost to navigation as started (the plan's wait_for verifies
+  the page). Non-click verbs still require a real reply.
+
+## [0.3.8] - 2026-10-21
+
+### Fixed
+
+- **Click reply loss** A submit/search click can start a navigation that destroys the
+  content-script context before its reply is sent, producing `bridge.timeout` at the
+  kernel. The content script now replies first and activates the element on the next
+  task; the plan still verifies the result with its following `wait_for`.
+
+## [0.3.7] - 2026-10-21
+
+### Fixed
+
+- **Live detail type** Online-first journal detail pages have no volume/issue, so the
+  previous metadata-only rule left `doc_type` empty and exported `GEN`. The journal
+  header link text (`查看该刊数据库收录来源`) is now used as article-level structural
+  evidence, so real online-first details export `JOUR`. Verified live on a CNKI
+  journal article (abstract, DOI, keywords, 0 risk).
+
+## [0.3.6] - 2026-10-21
+
+### Fixed
+
+- Read detail document types from complete type values or explicitly labelled
+  type fields. Funding, subjects and source names cannot supply a document type;
+  explicit online-first journal metadata is retained.
+- Reserve two seconds between CNKI search content waits and their outer action
+  deadlines, allowing a timeout response and its risk snapshot to reach the kernel.
+- Add synthetic online-first metadata fixtures and regression coverage for all
+  supported query/page/sort/page-size combinations. Live verification is pending.
+- Accept wrapped (`《期刊》`) and status-suffixed (`期刊（网络首发）`) type values, and
+  scan `.top-first` with the same exact-value rule, so explicit types are not lost
+  to page decoration. Body/recommendation text still cannot supply a type.
+
+## [0.3.5] - 2026-10-21
+
+### Fixed
+
+- **R034-01** Detail document type is read only from the article metadata header
+  (`.top-first`, `.doc-top`, `.top-tip`, `li.top-space`). Body text, abstracts and
+  related recommendations can no longer re-type the current article, and a concrete
+  volume/issue/pages citation always wins as journal evidence. Fixtures added for
+  "related thesis recommendation" and "conference notice in body".
+
+## [0.3.4] - 2026-10-21
+
+### Fixed
+
+- **CSP** Clicks on CNKI's `<a href="javascript:void(0);">` pagination/sort links are
+  cancelled at the default-action stage, so the page CSP no longer blocks a
+  `javascript:` navigation while the site's own click handlers still run. Plain
+  links and inputs are clicked as before.
+
+## [0.3.3] - 2026-10-21
+
+### Fixed
+
+- **F01** The result table's `td.data` document type is parsed into `Record.doc_type`;
+  RIS/EndNote now map journal/thesis/conference/newspaper/book/patent/standard/yearbook,
+  and an unknown or missing type renders as `GEN` / `Generic` (plus a RIS note) instead of
+  being silently exported as a journal article. `schemas/record.json` and the export
+  validator include the new field; mixed-type fixture added.
+- **F02** Detail readiness now waits for the article-level selector
+  `.wx-tit h1, #ChDivSummary` instead of a bare `h1` (the login modal), and the inner wait
+  is 8s so it finishes inside the 15s action budget and can still return the final page
+  or risk snapshot.
+
+## [0.3.2] - 2026-10-21
+
+### Fixed
+
+- **V01** A quota reservation that cannot be persisted now blocks that same action;
+  an existing but corrupt/unreadable quota file is a hard error instead of being read
+  as zero usage.
+- **V02** If the cross-process plan lock cannot be created, the plan is refused; the
+  process-local fallback was removed.
+- **V03** The shared cooldown is re-checked after the interval sleep and immediately
+  before the browser action.
+- **V04** A cross-domain login redirect is still refused, but the reply keeps the final
+  URL and risk metadata so the kernel can persist the 24h cooldown.
+- **V05** Extension permissions and content injection are scoped to the enabled
+  channel (`https://*.cnki.net/*`) and top frame only.
+
+### Changed
+
+- Skills now match the real contract: one page per call, no fabricated export
+  `count`, and a single, consistent rule against batch detail loops.
+- `install.ps1` falls back to `python`/`python3` on PATH and records the runtime HOME
+  in `runtime-home.txt`; the host reads it at startup.
+- Launcher logs move under `SCHOLAR_SEARCH_HOME/host/logs`.
+- `scholar-search assets` refuses to overwrite existing directories without `--force`.
+- Added `docs/integration/codex.md`.
+
+## [0.3.1] - 2026-10-21
+
+Risk follow-up fixes from `independent-20261021-02-risk-followup`.
+
+### Fixed
+
+- **R01** Every content-script reply (success, timeout, missing element, exception)
+  carries the risk snapshot; `click`/`type`/`select` refuse to run while a captcha or
+  login gate is already visible; the background no longer drops `data` on failures.
+- **R02** A whole plan holds a cross-process lock; the quota slot is reserved
+  atomically in `wait_turn` (interval check + increment + persist), so two processes
+  cannot both spend the last request.
+- **R03** The shared cooldown is re-checked before every action, not only at plan start.
+- **R04** Actions stay bound to the tab opened by the plan's `goto`; a plan cannot
+  drift to whatever tab the user activates. The extension keeps a configured host
+  scope, and the host only lets a request narrow its trusted startup scope, never
+  replace it.
+- **R05** If quota persistence fails, the next `wait_turn` fails closed instead of
+  trusting the stale on-disk counter.
+- **R06** `HTTP 429` (status or error text) is now a terminal `rate_limited` signal.
+
+### Compatibility
+
+Two assertions in the earlier audit conflict with the newer risk scope and now fail
+by design: the host no longer accepts an arbitrary `example.org` manifest without a
+trusted scope, and the extension no longer navigates to unconfigured hosts.
+
+## [0.3.0] - 2026-10-21
+
+Security and contract hardening from the independent offline audit
+(`independent-20261021-01`). Breaking: several invalid inputs that were quietly
+accepted now fail loudly.
+
+### Fixed
+
+- **F01** Navigation replies carry the final URL and current title; the adapter
+  surfaces a nested `data.url`; HTTP status is collected from the content script
+  and `403` reaches the risk guard.
+- **F02** Captcha visibility requires an on-screen, non-zero-size element that
+  intersects the viewport; every action result carries the risk snapshot.
+- **F03** Quota and interval are read and written under an inter-process lock;
+  plans are serialized process-wide, so two servers cannot double-spend.
+- **F04** The adapter never replays an action after a connection failure.
+- **F05** A failed action stops the plan; the service raises `ActionError` and
+  never parses or caches the page as a success.
+- **F06** The cooldown deadline is latched in memory, so a failed state write
+  cannot unblock the process that just detected a risk.
+- **F07** Cache hits still respect an active cooldown.
+- **F08** Replies are correlated with request ids; a handshake precedes the
+  first action; the bridge protocol doc separates TCP JSON from native framing.
+- **F09** One closed parameter schema (`core/params.py`) is enforced by kernel
+  and host; deadlines above 15s are rejected; frames are capped at 2 MiB on
+  write and `read_html` truncates by UTF-8 bytes with a `truncated` flag;
+  `wait_for` honours `state=hidden`.
+- **F10** Export validates records against the public schema and refuses
+  newline/type injection in any field.
+- **F11** Malformed search pages and non-article detail pages raise `ParseError`
+  instead of being cached as empty/valid results.
+- **F12** Detail URL authorisation uses the parsed path prefix.
+- **F13** Old site-version directories stay discoverable; selecting a version
+  no longer requires every directory to match the channel manifest.
+- **F14** Domain authority moves to the selected manifest (sent per request);
+  the host no longer hard-codes CNKI and the extension accepts any https host.
+- **F15** Host runtime files live under `SCHOLAR_SEARCH_HOME/host`; the
+  installer labels that directory low-integrity writable.
+- **F16** HTTP Host check compares the whole hostname, not a prefix.
+- **F17** Wheel ships native-host, extension and schemas as package assets;
+  `scholar-search assets --target DIR` extracts them; sdist includes browser
+  sources and fixtures.
+- **F18** `schemas/tools.json` matches the four live tools; README, NOTICE and
+  package metadata are on 0.3.0.
+- **F19** `status.bridge_ready` probes a live host instead of checking for a file.
+- **F20** `ruff format --check` passes and `.pre-commit-config.yaml` is present.
+
+
+## [0.2.6] - 2026-10-21
+
+### Added
+
+- `sort=date` and `sort=citation` drive the result grid's sort tabs (`#orderList`).
+  A relevance reset precedes the target click, because the grid remembers the previous
+  direction and a bare click would toggle unpredictably.
+- Live-verified: `limit=50, sort=date` returns 50 rows in strict descending date order
+  (2026 → 2024), 12 pages, no verification prompt.
+
+## [0.2.5] - 2026-10-21
+
+### Added
+
+- Field detection for expert search now covers every documented code: `SU`, `TKA`, `KY`,
+  `TI`, `FT`, `AU`, `FI`, `RP`, `AF`, `FU`, `AB`, `CO`, `RF`, `CLC`, `LY`, `DOI`, `CF`.
+  Unknown `XX=` / `XX%` codes also route to the expert page instead of the one-box search.
+- Skill documentation now lists the official field table and operators (`and`/`*`,
+  `or`/`+`, `not`/`-`, `%` for author matching).
+- Live-verified: `TI='IPMC' and AU % '王'` returns 43 hits across 2014–2026.
+
+## [0.2.4] - 2026-10-21
+
+### Added
+
+- `limit > 20` switches the result grid to 50 rows per page (12 pages for the 595-hit
+  `IPMC` query); with the 3-page cap a single call returns up to ~150 records.
+- `year_from` / `year_to` are applied client-side to the fetched records. The site's own
+  year control is never touched; `total` still reports the site's full result count.
+
+### Fixed
+
+- Search URLs carry a per-call nonce so every call starts from the grid's default state;
+  previously a repeat navigation could inherit the previous page size and break the switch.
+
+## [0.2.3] - 2026-10-21
+
+### Fixed
+
+- Account-gate detection now requires a **visible** login modal; the previous body-text match
+  flagged CNKI's always-present login panel and produced a false cooldown.
+- Native hosts register themselves under `hosts/` and answer a ping; the adapter picks a live
+  host instead of trusting a single stale `bridge.json`.
+- Removed host “retirement”: it caused reconnect churn where each new host killed the previous one.
+- Verified after the fix: one read-only search returned 595 hits, no cooldown.
+
+## [0.2.2] - 2026-10-21
+
+### Changed
+
+- Year filters are disabled: the result-page year control (`#btnFilterYear`) triggered an
+  account verification gate (“手机号认证”) during a live test. `year_from` / `year_to`
+  now fail with an explicit error instead of touching that control.
+- Risk guard detects account verification gates via the content script (`auth_required`)
+  and reports `login_redirect`, so a gate cannot be mistaken for an empty result page.
+- Incident recorded in `docs/incidents.md`; 62 tests.
+
+## [0.2.1] - 2026-10-21
+
+### Fixed
+
+- Detail parser reads volume/issue/pages from the `.top-first` citation line
+  (`力学学报 . 2022 ,54 (01) : 184-195`); verified by fixture replay and a live MCP call.
+- Added `tests/fixtures/cnki_detail_print_2026_10.html` and a regression test (59 tests).
+
+## [0.2.0] - 2026-10-20
+
+### Added
+
+- Kernel pipeline: registry, policy, risk guard, planner, audit, cache, service.
+- Channel registry with site-version selection and manifest/version consistency checks.
+- Policy engine: verb/domain allowlists, 3s hard floor, daily quota, plan size and duration limits.
+- Risk guard: captcha/verify/403/login detection, 24h persisted cooldown, manual reset only.
+- JSONL audit log (structural facts only, no page content or credentials).
+- SQLite response cache with TTL.
+- RIS and EndNote text export.
+- Fixture adapter for offline replay.
+- 40 unit and contract tests covering policy, risk, planner, cache, registry, service, export.
+- Native bridge: 4-byte-length JSON framing, request correlation, timeout handling.
+- Bridge policy: nine-verb allowlist, https-only, CNKI host allowlist, sleep bound.
+- Loopback TCP channel between host and MCP server, per-run random token.
+- MV3 extension skeleton (background relay + nine generic DOM verbs).
+- `native-host/install.ps1` registers the host for Edge and Chrome under HKCU.
+- 47 tests including bridge framing, policy and TCP round-trip.
+- CNKI channel v2026.10: search plan, result-table parser, detail plan and parser.
+- MCP server with four tools: search, detail, export, status.
+- stdio transport and loopback HTTP transport with bearer token and Host check.
+- `scholar-search status` reports channels, cooldown, quota and bridge readiness.
+- 51 tests; ruff and mypy --strict clean.
+- Verified selectors against a live CNKI page (595 hits, 20 rows); trimmed fixture in `tests/fixtures`.
+- Search plan now drives the search box (goto → type → click → wait → read) instead of a static URL.
+- Native host launcher is a compiled .exe: Edge refuses to start `.cmd` hosts.
+- Host runtime files (endpoint, log) moved to the per-user temp directory, because browsers launch native hosts at low integrity and `AppData\Local` is not writable there.
+- Single-instance check and per-connection replies fix duplicate-host races.
+- 52 tests.
+- Expert search: field expressions (`SU=`, `TI=`, `KY=`, `AU=` …) go through `AdvSearch?type=expert`.
+- Pagination: `page=2..3` clicks `#PageNext` and waits for the next result batch.
+- Search limit is honoured (`limit` was previously ignored).
+- Cache keys include the package version so parser changes do not serve stale hits.
+- Plan budget stays under 120s via per-action timeouts; page capped at 3.
+- 56 tests; live-verified: keyword 595 hits, expert expression 57 hits, page 2 distinct.
+- Detail parser rewritten for the 2026 layout: `.wx-tit h1`, `h3.author`, `li.top-space` rows; strips author affiliation marks and “附视频” suffixes.
+- Risk guard now flags verification interstitials by document title (“安全验证”), after an expired detail URL served one.
+- Native host runs one process per browser connection: the previous single-instance check deadlocked reconnects.
+- Skills: `scholar-search`, `scholar-detail`, `scholar-export`.
+- Docs: reference parity table and incident log.
+- 58 tests; release workflow on version tags.
+
+## [0.1.0] - 2026-10-20
+
+### Added
+
+- Project skeleton: microkernel layout, channel plugin directory, adapters, transports.
+- Core contracts (`core/contracts.py`): action vocabulary, plan, record, channel and adapter protocols.
+- Hard limits in `core/config.py` (3s floor, 1 concurrency, 3 pages per call, 200 per day, 24h cooldown).
+- Error hierarchy with risk-blocked payload.
+- ADRs 0001-0004: microkernel, read-only, native bridge, risk stop.
+- JSON schemas for records and MCP tools.
+- CNKI channel manifest (site version 2026.10), implementation pending.
+
+
+
+
+
+
+
+
+
+
+
+

wensu-0.4.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 scholar-meta contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

wensu-0.4.0/NOTICE

@@ -0,0 +1,89 @@
+wensu (formerly scholar-search)
+===============================
+
+Version: 0.4.0 (2026-10-23)
+
+Academic search: a read-only, microkernel-based scholarly metadata service.
+
+This project is an independent, non-official tool. It is not affiliated with,
+endorsed by, or sponsored by any academic database operator.
+
+
+Provenance and third-party notices
+----------------------------------
+
+This codebase was written for this project. No source files, modules or functions
+were copied from another repository. What it does reuse is narrower than "code":
+
+  1. CSS selector strings that identify elements on CNKI pages. These are facts
+     about a public web page (class names, ids), not creative expression; they
+     were checked against live pages and are covered by fixture tests.
+  2. The order of the interaction steps (open search page, fill box, submit,
+     read grid), which is dictated by how the site works.
+  3. Knowledge of known pitfalls, such as the hidden captcha widget that stays in
+     the DOM, which would otherwise produce false positives.
+
+Those were informed by three public MIT-licensed projects:
+
+  cookjohn/cnki-skills        https://github.com/cookjohn/cnki-skills
+  SepineTam/cnki-mcp          https://github.com/SepineTam/cnki-mcp
+  wuruiqi/cnki-mcp            https://github.com/wuruiqi/cnki-mcp
+
+All three are MIT-licensed. This project is MIT-licensed as well, so the terms are
+compatible either way; the attribution above is given because it is the right
+thing to do, not because a licence compels it for the three items listed.
+
+What was NOT taken from them: the browser transport (they drive a debug port; this
+project uses a browser extension plus a native messaging host), the microkernel
+and channel layout, the policy/quota/risk engine, the parsers, the export layer,
+and every line of Python, JavaScript, C# and PowerShell in this repository.
+
+If you believe a fragment here is derived from your work in a way this notice does
+not describe, open an issue and it will be corrected or removed.
+
+See docs/reference-parity.md for a feature-by-feature comparison, including where
+this project deliberately diverges.
+
+
+Dependencies
+------------
+
+Runtime (installed by pip, not vendored into this repository):
+
+  mcp              MIT
+  beautifulsoup4   MIT
+  lxml             BSD-3-Clause
+  uvicorn          BSD-3-Clause
+  packaging        Apache-2.0 or BSD-2-Clause
+
+Development only:
+
+  pytest, pytest-cov, ruff, mypy, pre-commit    MIT
+
+Licence texts are distributed with each package; the statements above follow the
+metadata those packages publish.
+
+
+Trademarks
+----------
+
+"CNKI", "知网", and other names are trademarks of their respective owners.
+They are used here only to describe interoperability targets, not as branding.
+
+
+Data and fixtures
+-----------------
+
+No bibliographic data, credentials, cookies, or session material is distributed
+with this project.
+
+The HTML files under tests/fixtures/ are page **structure** recorded from public
+pages, then desensitised:
+
+  - abstracts, author names, affiliations and keyword labels were replaced with
+    synthetic placeholders;
+  - mail addresses and any session-scoped parameters were removed;
+  - what remains is markup, class names, and public bibliographic facts (article
+    titles, journal names, volume/issue/page numbers, DOIs).
+
+They exist so the parsers can be tested offline without touching the site.