The previous implementation silently skipped any adapter whose import
failed (catch + warn-to-stderr + return []), then printed a successful
"✅ Manifest compiled: N entries". When dist/ was stale (e.g. after
renaming an export the JS adapters re-import) every adapter using that
export would fail to load, get skipped, and the script still exited 0.
An agent reading exit codes to gate work would commit the resulting
manifest and silently delete dozens of unrelated adapter entries.
Three layers of defense:
1. Distinguish skip kinds. Files that don't call `cli(...)` are still
silently dropped (helpers / type modules). Files that look like CLI
modules but fail to import now throw `ManifestImportError`. The
batch scanner aggregates failures and `main()` exits 1 with an
explicit list, leaving the existing manifest on disk untouched.
2. Net-deletion safety net. `main()` diffs the new entries against the
committed manifest and refuses to overwrite when entries would be
removed. `--allow-removals=N` (or bare `--allow-removals` for any)
is the explicit opt-in; the error message tells the caller exactly
what value to pass.
3. Runtime dist guard. `node dist/src/build-manifest.js` now refuses
to run with a clear pointer at `npm run build-manifest` (which uses
tsx). The npm script itself is migrated to `tsx src/build-manifest.ts`
so no project-level command points at the compiled copy anymore.
Release CI gains a manifest-drift gate (build-manifest + git diff
--exit-code) so a tag push can never publish stale or silently-shrunk
manifests. The existing CI check on PRs is preserved.
`ManifestEntry` is split into `src/manifest-types.ts` so runtime code
(discovery.ts) imports the type without pulling the build-time
compiler module.
Tests:
- `loadManifestEntries` throws ManifestImportError on import failure
- helper modules without cli() are still silently skipped
- `scanClisDir` aggregates per-adapter failures
- `diffRemovedEntries` returns expected site/name diff
- `parseBuildManifestArgs` reads --allow-removals[=N]
WAWQAQ feedback: the green left border on the status row looked
disconnected — only on the top half of the card, creating an awkward
stub. Connection state is already conveyed clearly by the colored dot
and the "Connected to daemon" / "Disconnected" text, so the border was
redundant decoration.
Drop the .card.connected/.disconnected/.connecting border-left rules.
No JS or layout changes; cleaner surface, fewer visual variants.
- Merge status row and profile row into a single rounded card with a
brand-colored left border accent indicating connection state
- Render contextId inline next to a "Profile" label with a Copy button,
letting users paste it into `opencli profile rename` without manual
selection (replaces the old full-width code block treatment)
- Show daemon version inline in the status row when connected, and
render the extension version as a tag in the popup header — both
surface version information that helps diagnose stale-daemon issues
- Forward both versions through the existing `getStatus` background
message: extension reads its own version from the manifest, daemon
version is fetched best-effort from `/status` with a 1.5s timeout so
popup never hangs when the daemon is unreachable
Closes#1192. Two changes:
1. New `instagram collection-create <name>` adapter wraps
`POST /api/v1/collections/create/` (multipart `name` +
`module_name=collection_create`, X-IG-App-ID + X-CSRFToken).
2. `instagram saved` gains an optional `--collection <name>` flag.
When set, the adapter resolves the name to a collection id via
`/api/v1/collections/list/` (case-insensitive trim match) and then
fetches `/api/v1/feed/collection/{id}/posts/`. Unknown names throw
with the available list so callers can self-correct.
Both verified end-to-end against a live IG account. Verify fixtures
under ~/.opencli/sites/instagram/verify/ ship the
patterns/notEmpty/mustBeTruthy guards from the latest adapter-author
skill (success-rate-pitfalls §1, §4, §8).
* feat(weibo): add favorites + publish CLI commands
Consolidates #1253 (favorites) and #1254 (publish) into a single PR per maintainer request.
- clis/weibo/favorites.ts: cookie-mode fetch of authenticated user's favorites via weibo.com/u/page/fav/{uid}
- clis/weibo/publish.js: UI-automation post (text up to 2000 chars, up to 9 images jpg/png/gif/webp)
- cli-manifest.json regenerated to include the new commands
Note: favorites.ts uses TypeScript syntax but build-manifest.js scans only *.js — favorites is currently NOT registered in the manifest. Reviewers please check whether to rename to .js or whether the manifest scanner should learn .ts.
Authored-by: hszhsz <heshaoz1990@gmail.com>
* fix(weibo): harden favorites and publish commands
* fix(weibo): publish without execute gate
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
* feat(claude): add Claude adapter
Adds a Claude (claude.ai) browser adapter family with seven commands
modeled on the existing clis/deepseek/ pattern: ask, send, new, status,
read, history, detail.
Closes#1251
* feat(claude): align send command columns with doubao
Match the established Status / SubmittedBy / InjectedText shape used by
doubao send so agent loops can rely on a consistent fire-and-forget
output across AI chat adapters.
* fix(claude): preserve DOM order in getVisibleMessages
The previous implementation queried user-message and assistant-message
nodes in two passes, which serialized as [u1, u2, u3, a1, a2, a3] for
multi-turn chats instead of the correct conversation order. Single
combined query preserves DOM order so claude read / detail return
turns in the order the user reads them on the page.
* docs(claude): note --live requirement for read across invocations
* fix(claude): fail fast on auth and empty states
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
* feat(update-check): show extension update notice on exit
The CLI exit hook already prints "Update available" when a newer @jackwener/opencli is on npm. Extension updates were only surfaced inside `opencli doctor`, so users running normal browser commands had no signal that the Chrome extension was out of date.
Solution piggybacks on the existing 24h background fetch:
- Daemon writes the live extensionVersion + lastSeenAt into the shared cache on every hello handshake (rare event, one fs.writeFileSync).
- CLI exit hook reads the cache it already loads and prints an extra extension notice when a newer release is available and the cache is fresh (<7d).
- writeCache becomes a read-merge-write so the daemon's currentExtensionVersion and the CLI's npm latestVersion don't clobber each other.
Net cost on the CLI hot path: zero new I/O, zero new daemon contact. The notice formatter is split into a pure helper (buildUpdateNotices) so the staleness window, equality, and combined-notice cases are unit-tested without touching disk or stderr.
* fix(update-check): tolerate partial cache when daemon writes first
Self-review caught a TypeError path: if the daemon's hello handler runs `recordExtensionVersion` before the CLI's npm fetch ever populated the cache, the resulting cache file has only `currentExtensionVersion` + `extensionLastSeenAt` and no `latestVersion`. The next CLI run then fed `undefined` into `isNewer`, which calls `.replace(...)` on it.
- Mark `lastCheck` and `latestVersion` optional in the cache schema (the merge pattern means either side may write first).
- Guard the CLI notice on `cache.latestVersion` being defined before comparing.
- Guard `checkForUpdateBackground`'s 24h short-circuit on `lastCheck` being defined.
- Add a test for the daemon-only cache case.
* feat(zhihu): add collection command to list favorite items
Add new 'opencli zhihu collection' command that:
- Lists items from a Zhihu collection (requires login)
- Supports pagination with --offset and --limit parameters
- Handles multiple content types: answer, article, pin
- Shows collection statistics: total count, total pages, current page
* feat(zhihu): split collection into collection and collections commands
- Rename zhihu collection list functionality to zhihu collections
- Keep zhihu collection for viewing specific collection contents by ID
- Convert collection.ts to collection.js so build-manifest picks it up
- Add tests for both commands
- Update cli-manifest.json
* fix(zhihu): harden collection read commands
---------
Co-authored-by: Developer <developer@example.com>
Co-authored-by: jackwener <jakevingoo@gmail.com>
* fix(twitter/following): switch from INTERCEPT+autoScroll to COOKIE+cursor pagination
The previous INTERCEPT strategy relied on autoScroll to trigger Twitter's
pagination by scrolling document.body. Twitter's virtual list doesn't grow
document.body.scrollHeight, so scrolls stopped triggering API calls after
the first few pages, capping results at ~50 regardless of limit.
Now uses Strategy.COOKIE with explicit cursor-based GraphQL pagination
(same pattern as twitter/likes), which correctly fetches all pages.
Fixes#1230
* fix(twitter): harden following pagination
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
* feat(boss): add jobType filter and bossOnline output
Add --jobType param (全职/兼职/实习 = 1901/1902/1903) so callers can
exclude internships at the API layer instead of post-filtering by name
keywords. Without this, --experience 应届 returns a mix of 校招 and 实习
because BOSS bundles both under code 108.
Also surface bossOnline (Y/empty) in results so callers can prioritize
HRs currently online — this is the only activity signal exposed by the
web API; 'recently active' / 'newly posted' filters are mobile-only and
not accepted by /wapi/zpgeek/search/joblist.json.
* fix(boss): correct experience codes (应届=102, not 108)
The previous EXP_MAP was off by ~2 across the board. Verified each
code by clicking BOSS web's filter UI and reading the URL:
108 = 在校生 (interns) was: '在校/应届','应届' → 108 (wrong)
102 = 应届生 (校招 full-time) was: '1-3年' → 102 (wrong)
101 = 经验不限 was: '1年以内' → 101 (wrong)
103 = 1年以内 was missing
104 = 1-3年 was: '3-5年' → 103 (wrong)
105 = 3-5年 was: '5-10年' → 104 (wrong)
106 = 5-10年 was: '10年以上' → 105 (wrong)
107 = 10年以上 was missing
This is why --experience 应届 had been returning mostly 实习生 jobs:
it was secretly querying 在校生 (108). The fix makes 应届 actually
mean 应届生 (102 = 校招), and lets users pick 在校生 (108) explicitly
when they do want internships.
* fix(boss): validate job type filter
* fix(boss): keep legacy campus experience alias
---------
Co-authored-by: youhh <youhh@1051233107@qq.com>
Co-authored-by: jackwener <jakevingoo@gmail.com>
* feat(deepseek): add vision mode support
DeepSeek added a third model "识图模式" (Vision Mode) that accepts
image uploads for visual understanding. Add vision to the --model
choices, update selectModel to use explicit index mapping for all
three models, skip the search toggle in vision mode (not available),
and extend waitForFilePreview to detect image thumbnails via send
button state since vision mode shows a preview image instead of a
filename label.
Also catch "Not allowed" errors from setFileInput (Cloudflare may
block CDP file operations) so the DataTransfer fallback can run.
Closes#1215
* fix(deepseek): harden vision upload mode
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
* fix(chatgpt): fix image generation detection and output path
Three fixes for chatgpt image command:
1. Page navigation: ChatGPT redirects away from the conversation
after sending. Poll for the /c/ URL after send, then periodically
reload the conversation page during image wait to pick up
asynchronously rendered images.
2. Composer selector: add fallback selectors for the chat input
since ChatGPT uses different aria-labels across UI versions.
3. Output path: the default '~/Pictures/chatgpt' was passed as a
literal string without tilde expansion, creating a directory
named '~' in the working directory. Removed the string default
and use os.homedir() fallback instead.
Fixes#1206
* fix(chatgpt): fail fast on image export failures
* fix(chatgpt): avoid reloads during image generation
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
* fix(chatgpt-app): support Traditional Chinese UI labels
The send button and Options button matchers only included Simplified
Chinese ("发送", "选项"). On macOS systems with Traditional Chinese as
the system language, the ChatGPT desktop app exposes "傳送" and "選項"
via the Accessibility API, causing `chatgpt-app send` to fail with
"Could not find send button" and `chatgpt-app model` to fail with
"Could not find Options button" for zh-TW / zh-HK users.
Verified via AXUIElement walk on ChatGPT 1.2026.104 / macOS 26 with
system language set to Traditional Chinese.
The "Stop generating" detection at line 314 already handles Traditional
Chinese because 停止生成 uses identical glyphs in both writing systems.
"Legacy models" at line 261 still lacks any Chinese variant but is not
addressed here since the Traditional Chinese translation has not been
verified on a live UI.
* test(chatgpt-app): cover traditional chinese ax labels
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
* fix(zhihu): fix identity detection, comment, answer, and search
Identity detection: Zhihu removed __INITIAL_STATE__ and moved the
user avatar from a profile link into a button. Added fallback that
extracts the user slug from the header avatar alt text.
Comment and answer: Zhihu moved the comment editor into a Modal
and changed the submit button behavior, breaking the UI-based
write flow. Replaced with direct API calls (POST /api/v4/answers/
{id}/comments and POST /api/v4/questions/{id}/answers) which are
reliable and much simpler.
Search: Zhihu's search API now returns mixed result types (ads,
education, hot_timing) alongside search_result. Updated the filter
to select by object.type (answer/article/question) and increased
fetch size to compensate for non-content results.
Fixes#1198
* fix(zhihu): rewrite like, follow, favorite to use API
Same DOM breakage as comment/answer. Replaced UI-based click
flows with direct Zhihu API calls for all write commands.
* fix(zhihu): harden api write regressions
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
* feat(zlibrary): add search and info commands
Add Z-Library adapter with two browser-based commands:
- `search` — Search books by title, author, or ISBN.
Navigates to /s/<query> and extracts results from
<z-bookcard> shadow DOM custom elements.
- `info` — Get book details and available download formats
from a book page URL.
Uses Strategy.COOKIE with browser automation to bypass
Cloudflare protection. The adapter reuses the user's existing
Z-Library login cookies from system Chrome.
Known limitation: actual file downloading requires Playwright's
download event handling (page.on('download')). OpenCLI's browser
automation does not currently intercept file downloads. Users
needing to download files should use Playwright to navigate to
the book URLs discovered by this adapter.
* fix(zlibrary): harden input and empty extraction
---------
Co-authored-by: jean <jean@jeandeMacBook-Pro.local>
Co-authored-by: jackwener <jakevingoo@gmail.com>
* fix(deepseek): fix send button detection in sendMessage
The previous selector `btn.closest('div')?.querySelector('textarea')`
always returned null because the button itself is a div, so
closest('div') returns the button, which has no textarea inside.
This caused every send to fall through to the Enter key fallback.
Walk up from the textarea to find the input container, then select
the last enabled non-toggle button with an SVG icon (the send
button). Excludes `.ds-toggle-button` elements (DeepThink / Search
toggles) so only the actual send button is clicked.
* fix(deepseek): fail closed when upload never enables send
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
* feat(browser): bind current tab to bound workspace
* docs(browser): document bound session idle semantics
* test(extension): cover bind-current owned-overwrite refusal
Adds regression for the second guard in handleBindCurrent that refuses
binding when the bound:* workspace already has an owned automation
window. Previously only the non-bound prefix path was tested.
* refactor(browser): rename bind command
* fix(browser): bind only current window tabs
* fix(browser): fail unbind when detach command fails
* feat(google-scholar): add cite and profile commands, fix search dedup
- cite: get BibTeX/EndNote/RefMan/RefWorks citation for a paper.
Clicks the cite button in search results and fetches the citation
content from Google's citation endpoint.
- profile: view an author's Google Scholar profile (h-index,
i10-index, citation count, top papers). Accepts author name
or Scholar user ID.
- search: fix duplicate results caused by CSS selector matching
both outer container (.gs_r.gs_or.gs_scl) and inner child
(.gs_ri) for each paper.
Closes#1174, closes#1175
* fix(google-scholar): fail fast on cite and profile misses
* fix(google-scholar): document new commands and lock dedup test
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
Pre-navigate Uiverse commands and retry detached browser bridge failures so code and preview flows stop falling back to about:blank. Broaden preview element matching for input-root components and cover the new navigation contract in tests.
* separate author name from date text in search results
* fix(xiaohongshu): constrain author date stripping
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
* fix(youtube): channel videos-tab fallback reads wrong tab from InnerTube response
After PR #1109, `opencli youtube channel <id>` still returns empty
`recent_videos` for channels whose Home tab is empty AND whose InnerTube
`/youtubei/v1/browse` response includes multiple tabs.
Root cause: the fallback fetch sends a browse request with the Videos
tab's `params`. The response, however, includes ALL tabs (Home, Videos,
Shorts, ...), with only the requested tab marked `selected: true`. The
existing code reads `tabs?.[0]?.tabRenderer?.content?.richGridRenderer?.contents`
— for multi-tab responses `tabs[0]` is Home (empty), so `richGrid` ends
up `[]` and `recentVideos` stays empty. PR #1109's test channels happened
to return single-tab lists with Videos at index 0, masking the bug.
Fix: find the tab with `selected: true` instead of assuming `tabs[0]`.
Reproducer: `opencli youtube channel UC44DSuDgw7_qccvZzIK3Jpg`
(杀鱼伟-Vi, ~3.1K subs, posts daily). Returns 0 videos pre-patch, 30+
videos post-patch.
`npm run typecheck` clean, `npm test` passes (1952/1952).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(youtube): preserve videos tab fallback
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: jackwener <jakevingoo@gmail.com>
* fix(doubao): use ID selector for send button
The clickSendButtonScript was searching for the send button by walking up
the DOM tree only 2 levels from the textarea, but the actual send button
#flow-end-msg-send is at level 5. This caused message sending to fail.
Fix by directly selecting the button via its ID.
* test(doubao): update send button selector assertions
* fix(doubao): keep send-button fallback contract
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
* fix(toutiao): move NON_TITLE_LINES inside function scope
NON_TITLE_LINES was defined outside parseToutiaoArticlesText() as a
module-level const. When the function is serialized via .toString()
and injected into browser evaluate context, outer scope variables
are not available, causing 'NON_TITLE_LINES is not defined' error.
Fix: move NON_TITLE_LINES inside the function so it's included in
the serialized string.
* test(toutiao): cover serialized articles parser
---------
Co-authored-by: sontjer <sontjer@github.com>
Co-authored-by: jackwener <jakevingoo@gmail.com>
* docs: update extension install to Chrome Web Store link
Extension is now published on Chrome Web Store. Replace manual
download/unpack instructions with the store link across READMEs
and skill docs.
* docs: restore manual install as Option B alongside Chrome Web Store
The shared article-download pipeline strips all <button> elements
via STRIPPED_TAGS, which is correct for article adapters (zhihu,
weixin) but causes web read to silently lose meaningful button
content like "Download All" on generic pages.
Override the button stripping in web read's configureTurndown
callback so button text is preserved as inline content.
Fixes#1184
jakevin