* fix(doubao/ask): restore Assistant turn detection after 2026-05 DOM refactor
Doubao reworked message-item wrappers and dropped all `receive-message` /
`bg-g-receive-msg-bubble` markers from assistant turns. The legacy 6
`itemSelectors` (`item-kDun2N`, `union_message`, `message-block-container`,
`data-message-id`, `bg-g-send-msg-bubble`, `bg-g-receive-msg-bubble`) match 0
elements on the new DOM, so `getTurnsScript` returned [] and `getDoubaoTurns`
fell through to the whole-page transcript scraper. Assistant text came back as
sidebar labels + history titles + adjacent conversation snippets concatenated
with the real reply — silent SELECTOR failure (no thrown error).
Two minimal changes in `clis/doubao/utils.js` `getTurnsScript`:
1. `itemSelectors`: prepend `[class*="inner-item-"]` and `[class*="top-item-"]`
— the new 2026-05 wrappers. Outer wins via existing ancestor-keep dedup
below, so we get one root per turn (not one per nested chunk).
2. `getRole`: add a third fallback branch — if the root matches
`inner-item-*` / `top-item-*`, contains `.flow-markdown-body`, and has NO
`bg-g-send-msg-bubble` marker (User detection still works), treat it as
Assistant. `.flow-markdown-body` is already in `messageTextSelectors`, so
text extraction kicks in unchanged.
Test added asserting both new wrappers and the `.flow-markdown-body` assistant
fallback are present in the generated script.
Fixes#1478
* test(doubao): cover refactored assistant turns
* fix(youtube): request srv3 format for caption URLs (#1420)
YouTube may return empty responses when caption URLs lack an explicit format
parameter. This adds fmt=srv3 (standard YouTube XML caption format) to the
caption URL when no fmt parameter is already present, with a fallback to the
original URL if srv3 also returns empty.
Also adds HTTP status checking before reading the response body, preventing
silent failures on non-200 responses.
Fixes#1420
* fix(youtube): preserve caption fetch failures
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
* feat(reddit): add reply command for replying to comments
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(reddit/reply): replace silent-sentinel rows with typed errors
reply.js originally mirror-copied comment.js's failure pattern: returning
[{ status: 'failed', message: 'HTTP 403' }] on auth/HTTP/Reddit errors and
relying on the caller to inspect the row instead of throwing. That's the
'silent-sentinel' anti-pattern from typed-errors.md — failures should
surface as typed errors so an agent can actually branch on them.
Round 21 lesson (f) — "grandfathered-not-exempt + helper-refactor boundary
is new" — applies: comment.js / upvote.js / save.js can stay grandfathered,
but a brand-new file does not inherit that exemption.
Changes:
- Throw AuthRequiredError when /api/me.json or /api/comment returns 401/403,
or when /api/me.json returns 200 but data.name is missing (stale anon
session — empty modhash alone isn't a strong enough signal).
- Throw CommandExecutionError for non-2xx HTTP and for non-empty
data.json.errors (e.g. RATELIMIT, NO_TEXT, TOO_OLD).
- Drop the over-defensive `if (!page) throw ...` — registry guarantees a
page object when browser:true.
- Intermediate result object uses `kind` discriminator + `detail` /
`httpStatus` / `where` keys that don't overlap with columns
['status','message'], so the silent-column-drop audit stays quiet
(per PR #1329 sediment).
Verified:
- npx tsc --noEmit clean
- node scripts/check-typed-error-lint.mjs → 189/189, 0 new
- node scripts/check-silent-column-drop.mjs → 103/103, 0 new
- npx vitest run clis/reddit src/convention-audit → 11/11 pass
- node ./dist/src/main.js validate → 0 errors
Success path is unchanged: still returns
[{ status: 'success', message: 'Reply posted on t1_<id>' }].
* fix(reddit): harden reply command contract
* fix(reddit): reject suffixed reply urls
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: jackwener <jakevingoo@gmail.com>
* feat(rednote): add rednote.com adapter mirroring xiaohongshu read commands (#1136)
Implements rednote.com support as discussed in issue #1136. The mainland
xiaohongshu adapter stays in place; international users redirected to
www.rednote.com now have a CLI without a copy-pasted adapter.
Issue #1136 documents that xiaohongshu and rednote share DOM selectors,
URL paths, API paths, response schema, cookies, and the xsec_token auth
mechanism. The only material differences:
Layer xiaohongshu rednote
Web host www.xiaohongshu.com www.rednote.com
API host edith.xiaohongshu.com webapi.rednote.com
Security host fe-static.xhscdn.com as.rednote.com
Cookie root .xiaohongshu.com .rednote.com
Search gate Inline text Full-screen modal + text
## Architecture (minimal)
`clis/xiaohongshu/*` keep all selector / regex / extraction logic. Each
command file is touched minimally to export the IIFE or pipeline so the
sibling adapter can reuse it:
search.js + export const buildSearchExtractJs(webHost)
+ export const command = cli({...})
note.js + export const NOTE_EXTRACT_JS
+ export const command = cli({...})
comments.js + export function buildCommentsExtractJs(withReplies)
+ export parseCommentLimit
+ export const command = cli({...})
download.js + export function buildDownloadExtractJs(noteId)
(CDN allowlist now includes rednote alongside xhscdn)
+ export const command = cli({...})
user.js + export const USER_SNAPSHOT_JS
+ export const command = cli({...})
feed.js + export function buildFeedPipeline(webHost)
+ export const command = cli({...})
notifications.js + export function buildNotificationsPipeline(webHost)
+ export const command = cli({...})
note-helpers.js buildNoteUrl now accepts `cookieRoot` + `signedUrlHint`
options (defaults preserved so xhs callers and tests
are unchanged)
user-helpers.js buildXhsNoteUrl / extractXhsUserNotes accept an
optional `webHost` argument (default xhs)
The `export const command = cli({...})` pattern matches twitter/lists.js
and clis/discord-app/*; without it the build-manifest scanner attributes
xhs's command to whichever rednote sibling triggered the transitive
import first.
## clis/rednote/ — thin shims
Each rednote command file imports the relevant builder / constant from
its xiaohongshu sibling and calls `cli()` with the rednote host triple.
No selectors, regexes, or extraction logic are duplicated.
search.js imports buildSearchExtractJs + noteIdToDate
declares its own WAIT_FOR_CONTENT_JS (modal + text
login-gate variants — the one xhs behaviour that
genuinely differs)
note.js imports NOTE_EXTRACT_JS + buildNoteUrl + parseNoteId
comments.js imports buildCommentsExtractJs + parseCommentLimit
+ buildNoteUrl + parseNoteId
download.js imports buildDownloadExtractJs + buildNoteUrl + parseNoteId
user.js imports USER_SNAPSHOT_JS + extractXhsUserNotes
+ normalizeXhsUserId
## Scope (initial)
Ships the five commands verified live against the user's logged-in
rednote.com session: search / note / comments / user / download.
`feed` and `notifications` are intentionally left out. Both rely on
intercepting the xiaohongshu Pinia store at the `homefeed` / `you`
capture pattern; live verification on rednote returns `tap → dict
(error)` for the feed step, so shipping them would surface a broken
contract. The mainland xiaohongshu commands continue to work. Adding
the rednote-side feed / notifications is straightforward follow-up
work once someone with rednote access maps the network surface.
Creator-center commands (publish, creator-*) have no rednote
counterpart and stay xiaohongshu-only, per the reporter's note in #1136.
## Verification
- clis/xiaohongshu/ + clis/rednote/: 103/103 tests green
- npx tsc --noEmit: clean
- npm run build: 807 manifest entries (xhs 13 + rednote 5 + everything
else preserved)
- silent-column-drop / typed-error-lint: 103 / 189 baseline entries,
no new violations
- Live verify against the user's rednote.com session:
rednote search "travel" --limit 1 → real note row
rednote note <signed-url> → 7 field/value rows
rednote comments <signed-url> --limit 3 → 3 top-level rows
rednote user 5b21f6564eacab3b38f05c39 --limit 2 → 2 profile notes
Spaced 15–30s between runs per the xhs/rednote rate-limit guidance;
no write commands invoked. Regression check: xiaohongshu/feed on
the existing mainland session still returns the standard 6-field
rows after the refactor.
Closes#1136
* fix(rednote): tighten adapter failure boundaries
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
Doctor's job is browser-bridge health diagnosis. The `--no-live` flag
let users skip the connectivity probe (= the core diagnostic), and
`--sessions` listed automation sessions (a separate concern not part of
health). Both flags accreted features that violated the command's
first-principles purpose.
Cleanup chain (removing dead code surfaced by the flag removal):
- `--no-live` / `--sessions` flags removed from `opencli doctor`
- `DoctorOptions.live` / `DoctorOptions.sessions` removed
- `DoctorReport.sessions` removed
- `[SKIP] Connectivity` render branch removed (always-live now)
- `listSessions()` removed (only consumer was doctor)
- `'sessions'` action removed from daemon-client protocol type
- `BrowserSessionInfo` type removed (no remaining consumers)
- extension `handleSessions` action handler removed (1.0.12)
- extension test "reports sessions per session" removed
- `OPENCLI_BROWSER_IDLE_TIMEOUT` test rewired to 'cookies' action
Verification:
- root typecheck + extension typecheck pass
- doctor.test.ts 17/17 pass
- extension/background.test.ts 49/49 pass
- typed-error-lint 189/189 baseline
- silent-column-drop 103/103 baseline
- build + extension build green
Follow-up to feat #1458 (registering tg-cli/discord-cli/wx-cli in
src/external-clis.yaml) — README and README.zh-CN had not been updated
to reflect the new entries.
Updates four spots in each README:
- intro paragraph that names example external CLIs
- "CLI Hub" highlight bullet
- "OpenCLI is not only for websites" bullet list
- the External CLI table itself
Add three local-first messaging CLIs to the External CLI registry so
agents can discover and install them via `opencli external install`:
- `tg-cli` (binary `tg`) — Telegram local sync/search/export via MTProto
- `discord-cli` (binary `discord`) — Discord local sync/search/export
- `wx-cli` (binary `wx`) — WeChat local data CLI
Refresh the External CLI list in skills/opencli-usage/SKILL.md so the
agent-facing skill names stay in sync.
Extends A0 (PR #1404) by dogfooding `installCommanderNamespaceStructuredHelp`
on the four remaining built-in Commander namespaces:
- `opencli daemon --help -f yaml|json`
- `opencli plugin --help -f yaml|json`
- `opencli adapter --help -f yaml|json`
- `opencli profile --help -f yaml|json`
Each emits the same payload shape as `browser`: namespace metadata, every
leaf command's positionals + command_options + description + usage,
namespace_options (empty for these), and program-level global_options.
Agents can fetch every leaf's contract in a single call — no per-leaf
`--help` follow-ups.
Each namespace snapshots its original description at declaration time
because `applyRootSubcommandSummaries(program)` later overwrites
`.description()` with a child-name listing; without the snapshot,
structured help would surface `"restart, status, stop"` instead of
`"Manage the opencli daemon"`. Tests lock the snapshot semantics for
`adapter` explicitly.
Tests: 138/138 (4 new — one per namespace, covering description
preservation, leaf names, positionals, command_options).
Typecheck + build clean.
* perf(reddit): opt 13 browser adapters into shared site-tab lease
Adds `browserSession: { reuse: 'site' }` to every reddit adapter that
already runs `browser: true` on `domain: 'reddit.com'`. Same metadata-only
follow-up to the twitter sweep merged in #1454 — the framework's
`shouldRunPreNav` short-circuit (src/execution.ts:190) skips the redundant
domain-root pre-nav when a sibling adapter already has the tab on
reddit.com, and idle-bound tabs are reused under the `site:reddit` bucket
until expiry.
Scope (13 files, all on `domain: 'reddit.com'` + `Strategy.COOKIE`):
- read (9): frontpage / popular / saved / search / subreddit / upvoted /
user / user-comments / user-posts
- write (4): comment / save / subscribe / upvote
Excluded:
- `hot.js` (no browser:true — public Reddit JSON API, no tab)
- `read.js` (Strategy.COOKIE but no browser:true — non-browser pipeline)
No logic changes; only metadata + manifest regeneration.
Verification:
- npm run check:typed-error-lint → 189/189 unchanged
- npm run check:silent-column-drop → 103/103 unchanged
- npm run test:adapter → 264/264 passed (2146 tests)
- npx vitest run --project unit → 72/72 passed (unrelated EADDRINUSE
flake on daemon.test.ts port 19825, also seen on #1454/#1452)
- tsc --noEmit clean
* fix(reddit): include read in site browser session reuse
The uploadImages function catches errors from page.setFileInput and only
falls back to the legacy base64 DataTransfer method when the message
contains 'Unknown action' or 'not supported'. However, Chrome can also
return 'Not allowed' (code -32000), which was not handled — causing the
publish command to fail instead of using the fallback.
Add 'Not allowed' to the fallback condition so image upload works even
when CDP file injection is blocked by Chrome's security policy.
Co-authored-by: together <together@togetherdeMac-mini.local>
* feat(openreview): add author command for ID-explicit publication lookup
Closes the missing leaf in the openreview adapter. Among the public-strategy
academic adapters, dblp and arxiv both already ship an `author` command for
ID-explicit publication lookup; openreview only had `search` (full-text),
`paper` (detail by note id), `reviews` (thread by forum id) and `venue`
(listing by invitation / venue text). There was no way to ask "give me every
submission this author put on OpenReview, newest first."
`openreview author <profile>`:
- takes a canonical profile id (`~First_LastN`); validated by
`requireProfileId` so a dblp PID or a bare name fails before any
network call,
- hits `/notes?content.authorids=~<id>&limit=<n>&sort=cdate:desc`,
- returns rank-ordered rows with the same shape as `openreview search`
(id / title / authors / venue / pdate / url),
- throws `EmptyResultError` when the profile has no public submissions
instead of returning an empty list,
- inherits the typed-error envelope from `openreviewFetch` so network
failure, non-200, malformed JSON, and in-band error envelopes all
surface as `CommandExecutionError`.
Tests: 6 new `it` blocks plus 1 updated registration test in
`clis/openreview/openreview.test.js`.
- `requireProfileId` (1 block, 9 assertions): accepts canonical
`~First_LastN`, `~Bo_Liu17`, and a multi-segment middle-name id;
rejects empty, whitespace, missing tilde, missing trailing number,
embedded space, and a dblp-style PID.
- 5 author runtime cases covering pre-network ArgumentError, empty
result, non-200, fetch network error, and the happy path with a
request-shape assertion (`content.authorids` filter + `cdate:desc`
sort).
- Registration test extended to expect five commands and lock the new
`columns` contract.
Manifest auto-regenerated to register the new command.
Live-verified end to end against `~Yoshua_Bengio1`: the most recent ICLR
2026 workshop submissions return with the expected fields. A malformed
profile is rejected before any HTTP call. A nonexistent profile yields
`EMPTY_RESULT`.
* fix(openreview): accept real profile id slugs
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
Convert 9 of 18 page.wait(N) calls in clis/claude/ from fixed-duration
sleeps to event-based readiness checks (page.wait({ selector, timeout }),
backed by MutationObserver). Mirrors the deepseek D1 template (PR #1449).
* Page-ready waits (5 converted): utils.js:29 (ensureOnClaude composer),
utils.js:114 (getConversationList recents links), new.js:19 (composer),
detail.js:24 (.font-claude-response message bubble), send.js:26 (composer).
Each resolves as soon as the selector matches, swallowing the timeout so
downstream typed-error helpers (ensureClaudeLogin / ensureClaudeComposer
/ EmptyResultError) still surface the right error when the selector
never mounts (login redirect, empty conversation, etc).
* Dropdown waits (2 converted): utils.js:150 (selectModel post-trigger),
utils.js:178 (setAdaptiveThinking post-trigger). Wait for menuitemradio
/ menuitem to mount instead of a fixed 0.6 s sleep.
* Resume conversation wait (1 converted): ask.js:51 — wait for the resumed
message bubble (MESSAGE_SELECTOR) instead of a fixed 2 s sleep.
* Settle/redundant waits removed (6): ask.js:55 standalone settle (next
ensureClaudeComposer queries composer presence directly via getPageState);
ask.js:83 / ask.js:90 post-toggle settles (next CDP eval flushes React
state between roundtrips); ask.js:103 pre-waitForResponse settle (the
polling loop's first 3 s tick already covers this); read.js:20 post-
ensureOnClaude sleep (ensureOnClaude now waits for the composer selector
itself); send.js:29 post-ensureOnClaude sleep (same).
Three remaining page.wait(N) calls are kept: utils.js:231 post-input
1.2 s React debounce inside sendMessage (the ProseMirror editor needs a
debounce window before the send button enables; reducing this risks
silent send-button-disabled drops), and the 3 s / 1 s polling ticks in
waitForResponse / waitForFilePreview (already polling patterns, out of
scope for D-track wait→event sweep).
Targeted tests: clis/claude + src/browser 389/389 pass; tsc clean;
build clean; typed-error 189/189 baseline (no new); silent-column-drop
103/103 baseline (no new).
D2 in the LLM-adapter wait→event sweep started by deepseek (D1, #1449).
Convert 10 of 18 `page.wait(N)` calls in clis/deepseek/ from fixed-duration
sleeps to event-based readiness checks (`page.wait({ selector, timeout })`,
backed by MutationObserver):
* Page-ready waits (5): utils.js:46, ask.js:38/52, detail.js:28, new.js:19
now wait for the composer textarea (TEXTAREA_SELECTOR) or message bubble
(MESSAGE_SELECTOR) to mount before continuing. Resolves as soon as the
selector matches instead of always sleeping the full duration.
* Settle/redundant waits removed (5): ask.js:56 standalone settle (already
covered by upstream selector waits); ask.js:79/105 post-toggle settles
(next CDP eval gives React time to flush aria-checked updates); ask.js:118
pre-waitForResponse settle (the polling loop's first 3 s tick already
covers this); read.js:19 post-ensureOnDeepSeek sleep (ensureOnDeepSeek
now waits for the textarea selector itself).
* `new.js` now throws CommandExecutionError when the composer fails to
mount within 8 s instead of silently returning "New chat started" on a
half-loaded or logged-out page.
Eight remaining `page.wait(N)` calls are kept: in-loop polling ticks in
waitForResponse / pickResumeUrl / getConversationList / waitForFilePreview
/ send-button-enable polling (these are already polling patterns and
out of scope for D1), and the native-input flush + textarea-mount poll in
send.js.
Targeted tests: clis/deepseek 49/49, src/browser 355/355 pass; build,
typecheck, typed-error and silent-column-drop audits clean.
Proof template for the LLM-adapter wait-cleanup follow-ups.
Read-only Twitter/X adapters now declare `browserSession: { reuse: 'site' }`,
matching the LLM-site adapters (claude/gemini/yuanbao/etc.) and unblocking
the perf wins WAWQAQ called out for the 35s→9s/3.4s thread.js progression
(#OpenCLI:3889b5cf):
- Tab lease shared across calls under `site:twitter` until idle expiry, so
the second-and-later command pays no cold-start tab cost.
- Framework's domain-root pre-nav (`https://x.com`) is skipped on subsequent
calls when the reused tab is already on x.com (`shouldRunPreNav` →
`isDomainRootPreNav` + `urlMatchesDomain` short-circuit at
`src/execution.ts:190`).
Files (17 read-only adapters):
- Strategy.COOKIE × 13: article, bookmark-folder, bookmark-folders,
bookmarks, download, following, likes, list-tweets, lists, profile,
thread, timeline, trending, tweets
- Strategy.UI × 1: followers
- Strategy.INTERCEPT × 2: notifications, search
Insertion point in each file: after `browser: true,` (or after `strategy:`
in download.js which omits the explicit `browser:` field), matching the
convention used by yuanbao/read.js, claude/read.js, etc.
Manifest regenerated (cli-manifest.json: +85/-17 — 17 entries gain the
`browserSession: { reuse: "site" }` block).
Verification:
- npx tsc --noEmit clean
- npx vitest run clis/twitter → 218/218 pass (25 files)
- npx vitest run src/convention-audit.test.ts → 8/8 pass
- typed-error-lint baseline 189/189 (no new violations)
- silent-column-drop baseline 103/103 (no new violations)
Scope notes (intentionally NOT in this PR):
- Write adapters (post/reply/quote/like/retweet/bookmark/follow/list-add/
list-remove/delete/hide-reply/block/accept/follow) are kept as one-shot
by default — `reuse: 'site'` for write paths is a separate decision
about action idempotency under tab reuse.
- The thread.js / timeline.js comments still say "Cookie context
auto-established by framework pre-nav"; the deeper truth (CDP
`getCookies({url})` is origin-independent) was a framing nit on PR C
(#1451) — left as a doc-only follow-up to keep this PR's diff focused
on the perf gain.
Refs: #OpenCLI:3889b5cf (WAWQAQ msg=fa209a2c, msg=35c90460, msg=838128ef
"你们继续做啊… 后面还有那么多其他的东西呢")
* perf(twitter): drop redundant goto+wait — framework auto pre-navs (PR C)
Twelve twitter read adapters did `await page.goto('https://x.com'); await
page.wait(2~3)` purely to establish cookie context for the subsequent
`document.cookie` read. After PR #1450 hoisted those reads to
`page.getCookies({url})` (which queries the CDP cookie store directly,
no navigation needed), the explicit goto+wait became dead.
The framework already pre-navigates to `https://${domain}` for any
adapter declaring `Strategy.COOKIE + domain` (`src/registry.ts:191`),
so the cookie store is populated before `func` runs. The 2-3s
`page.wait` was the slowest part of the redundant call.
Files (all read-only, all ct0/cookie-only):
- bookmark-folder / bookmark-folders / bookmarks
- following / likes / list-add / list-remove / list-tweets / lists
- thread / timeline / tweets
Out of scope (kept as-is): goto calls that navigate to a *specific*
URL needed for content/SPA shell — `trending` (`/explore/tabs/trending`),
`notifications` (`/home`), `article` (`/i/article/{id}`), `profile`
(`/${username}`), and `list-add` line 133 (`/${username}` for UI ops).
Verification:
- npx tsc --noEmit ✓
- npx vitest run clis/twitter → 216/216 ✓
- typed-error-lint 189/189, 0 new ✓
- silent-column-drop 103/103, 0 new ✓
* fix(twitter): keep list UI root navigation
* perf: replace document.cookie reads with page.getCookies({domain}) (Tier 1 cookie API sweep)
Prior pattern in 25 adapter files round-tripped through `page.evaluate(\`document.cookie.split…\`)` to extract a single cookie value (CSRF token, session ID, etc.). CDP's `page.getCookies({domain})` reads the cookie store directly with zero JS-execution overhead.
Files touched (sites: twitter / linkedin / maimai / youtube):
- twitter (15): thread, timeline, list-add, bookmark-folders, following, list-tweets, bookmarks, list-remove, tweets, bookmark-folder, likes, lists, trending — direct 4-line replacement (cookie was outside `page.evaluate`); article, profile — hoisted ct0 read OUT of `page.evaluate` and threw `AuthRequiredError` upfront so unreachable in-evaluate auth branches got cleaned up too.
- linkedin/search.js — JSESSIONID was read inside the per-batch fetch loop's `page.evaluate`; hoisted once before the loop and pass `csrf` value into the template via `JSON.stringify`.
- maimai/search-talents.js — csrftoken cookie hoisted via getCookies; meta-tag fallback preserved inside `page.evaluate` (reached only when no cookie). Also converted the `page.evaluate(async (body) => …, body)` Playwright-style call to OpenCLI's template-string form so the helper actually runs.
- youtube — `SAPISID_HASH_FN` (used by like / unlike / subscribe / unsubscribe) reworked: sapisid is now passed in as a parameter; new `readYoutubeSapisid(page)` helper reads it via CDP. The HMAC-SHA1 compute still happens browser-side (Web Crypto), only the cookie read is hoisted.
Tests updated where mocks specifically referenced `document.cookie` (twitter following / bookmark-folder / bookmark-folders) to mock `getCookies` instead.
Verification:
- `npx tsc --noEmit` clean
- `npx vitest run clis/twitter clis/linkedin clis/youtube` → 264/264 pass
- typed-error-lint 189/189 (no new violations)
- silent-column-drop 103/103 (no new violations)
Scope notes (not in this PR):
- `goto + wait` redundancy and `browserSession: { reuse: 'site' }` rollout are scoped to follow-up PRs B and C per the #OpenCLI:3889b5cf thread plan.
- `document.cookie.match(...)` patterns (instagram 8 / xiaoe / qwen / hupu / tiktok / 1point3acres — ~13 files) are outside the original \`document.cookie.split\` audit scope and will follow as a Tier 1 expansion sweep.
* fix(adapters): read auth cookies by url scope
- bump opencli to 1.7.15 (was 1.7.14)
- extension stays at 1.0.9 (already bumped during the release cycle)
- finalize CHANGELOG: move Unreleased to 1.7.15 with date
Major release: Browser Agent Runtime project (Phase 0/1/2) — alignment
with vercel-labs/agent-browser model. CDP-primary input, AX snapshot/refs
with stale recovery, semantic locators across all primitives, full form
toolbelt (hover/focus/dblclick/check/uncheck/upload/drag/wait-download),
annotated screenshots, and same-origin iframe AX routing.
PR #1399 added auto-restart of stale daemons in BrowserBridge
(daemonVersion ≠ PKG_VERSION → restart). The browser-tabs e2e fake
daemon hard-coded `daemonVersion: 'test'`, so every test reported as
stale and the bridge tried to /shutdown the fake daemon — which has no
shutdown endpoint — causing all 4 tests in the file to exit with code 1.
This has been the failing signal in `e2e-headed (ubuntu-latest)` and
`e2e-headed (macos-latest)` on every main push since #1399.
Read PKG_VERSION from package.json once at module load and feed that to
the fake /status response. The fake daemon now matches the running CLI
so the stale-daemon path is not triggered.
Verification:
- npx tsc --noEmit clean
- npm run build clean
- npx vitest run --project e2e tests/e2e/browser-tabs.test.ts → 4/4 pass
* feat(dianping): resolve unknown cities live from www.dianping.com
The static CITY_ID map in clis/dianping/utils.js only covers ~20 cities,
so passing --city 汕头 (or any other Chinese name / pinyin slug not on
that list) fails with ArgumentError. Adding the missing cityIds by hand
doesn't scale to dianping's full city list and silently goes stale when
the site renumbers cities.
This change adds an async resolver that falls back to dianping.com when
the static map misses:
- Numeric input → pass through unchanged.
- Static map hit → fast path, no network (utils.CITY_ID untouched).
- Pinyin slug (e.g. "shantou") → goto /<slug>, parse cityId out of
any /search/keyword/{id}/ link rendered on the per-city landing page.
- Chinese name (e.g. "汕头") → goto /citylist, walk anchors to build a
Chinese-name → pinyin map, then resolve the slug as above.
Resolved (input → cityId) pairs are memoized per-process so repeat
searches skip both navigations.
Implemented as a new module (clis/dianping/cityResolver.js) so utils.js
stays minimal and the existing synchronous resolveCityId / CITY_ID API
keeps working for direct callers and tests.
Tested:
- Unit tests cover null/numeric/static fast paths, pinyin fallback +
cache, Chinese-name fallback via /citylist + cache for both forms,
rejection of garbage input, rejection of Chinese names not on
/citylist, and CommandExecutionError when the per-city page lacks
a /search/keyword/{id}/ link.
- JSDOM tests cover the pure DOM extractors (buildCitylistMap and
extractCityIdFromPage) against curated HTML fixtures.
- npm test: 3196 passed, 1 skipped (no new failures).
- npx tsc --noEmit: clean.
- opencli validate: 0 errors.
* fix(dianping): require city resolver links to be authoritative
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
* fix(douyin): handle empty response body in browserFetch (#1405)
browserFetch calls res.json() directly, which throws SyntaxError when
the API returns an empty body (content-length: 0). This happens when
the Douyin hashtag search endpoint returns HTTP 200 with no content.
Fix: read response as text first, return null for empty bodies, then
throw a descriptive CommandExecutionError at the caller level.
Fixes#1405
* fix(douyin): wrap browser fetch parse failures
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
Why
- `opencli twitter followers --help` rendered:
Arguments:
user
with a blank trailing column. Both humans and agents could not
recover the parameter's purpose without reading source. WAWQAQ
surfaced this directly: "没有说明当后面的 followers [user] [options]
如果都没填的时候，获取的是什么？"
- This is metadata completeness, not stylistic taste. Failing closed
is the only way to keep the help surface trustworthy as adapters
land.
What
- src/build-manifest.ts: add `findManifestMetadataIssues()` that flags
any positional with empty / whitespace-only / missing `help`. Wired
into `main()` after the import-failures gate; build aborts non-zero
with a per-arg report (`site/cmd positional "name" (sourceFile)`).
- src/build-manifest.test.ts: cover the gate (positives + negatives,
scoped strictly to positionals — named flags are intentionally
out-of-scope).
- 18 adapter offenders (16 required + 2 optional) get explicit help
text:
twitter: followers/following/list-add/list-remove/list-tweets/
search/thread
reddit: search/subreddit/user/user-comments/user-posts
douyin: stats/update
bilibili: subtitle
jike: search
Optional positionals (`twitter followers/following [user]`) now
document the omit semantics — fetches the currently logged-in
account.
- CHANGELOG: document the build gate and the offender list.
Out of scope (planned follow-ups)
- Semantic-quality advisory: optional positional help should also
contain `default / omit / current / logged-in / required unless …`
keywords. That belongs to the planned Arg metadata v2 work
(`when_omitted / when_present / value_format` 3-field schema).
- Named-flag `help` quality. Named flags carry the flag name itself
in help, so a missing `help` is not as opaque; if we want to gate
those too, do it as a separate, intentional decision.
Validation
- `npm run build` → 799 entries, clean.
- `npm run typecheck` → clean.
- `npx vitest run --project unit --project adapter` → 257 + 4 files,
all green (build-manifest 13 tests, manifest gate added).
- Smoke: temporarily reverted `followers.js` help to empty → build
aborts with the exact `twitter/followers positional "user" (...)`
line; restored, build is clean again.
- `npm run check:silent-column-drop` and `check:typed-error-lint`
baselines unchanged.
- bump opencli to 1.7.14 (was 1.7.13)
- extension stays at 1.0.6 (no extension changes since v1.7.13)
- finalize CHANGELOG with the three landed PRs:
* #1399 daemon restart on stale ready state for npm -g upgrade
* #1400 twitter write-action symmetry (unlike/retweet/unretweet/quote)
* #1401 agent-friendly adapter help (drop globally-shared option noise)
Fixes#1376 — YouTube transcript command failed with `No captions available for this video` for all videos.
## Root cause
Transcript adapter used InnerTube `/youtubei/v1/player` API with Android client context (`clientName: 'ANDROID'`, version `20.10.38`) to retrieve caption track URLs. YouTube has restricted/deprecated this approach; the Android client no longer reliably returns captions data.
## Fix
Replace Step 1 (caption track retrieval) with watch page HTML bootstrap parsing — fetch `/watch?v=...` with cookies and extract `ytInitialPlayerResponse.captions.playerCaptionsTracklistRenderer`. This is the same approach used by sibling `clis/youtube/video.js`, so it's an alignment to existing site-local stable pattern, not a new invention.
## 2 head iteration
- `cf77f5e8` initial fix (Step 1 caption retrieval switch + 18/18 unit tests)
- `bb30788c` lead test hardening — source-contract regression test in `transcript.test.js`:
- **positive lock**: must fetch `/watch?v=...`, parse `ytInitialPlayerResponse`, read `playerCaptionsTracklistRenderer`
- **negative lock**: must NOT use `/youtubei/v1/player` or `clientName: 'ANDROID'` (prevents regression)
- stale Android-InnerTube file header comment also updated
## Better-solution evaluation
- Official YouTube Data API captions surface (`developers.google.com/youtube/v3/docs/captions/download`) is owner-authorized API, NOT a public transcript replacement
- yt-dlp also relies on watch-page bootstrap path
- Existing `youtube/video.js` already uses the same `ytInitialPlayerResponse` extraction → this PR aligns transcript with stable site-local pattern instead of inventing a new path
## Typed failure / no-silent-empty boundaries
- watch HTML HTTP failure / missing `ytInitialPlayerResponse` / no `captionTracks` → `CommandExecutionError` (typed fail)
- Empty parsed XML → `EmptyResultError` (existing path, preserved)
- `Strategy.COOKIE` matches YouTube adapter family + `video.js`; cookies/session/consent unavailable → typed fail not silent empty success illusion
## Diff containment
Runtime change limited to Step 1 caption track discovery. XML fetch, segment parsing, chapters, raw/grouped formatting all unchanged.
## Verification
Local: YouTube adapter tests `19/19` (+1 from new test), `npm run build`, typed-error-lint `192/192`, silent-column-drop `103/103`, doc coverage `140/140`, `docs:build`, listing-id advisory unchanged `13`, `git diff --check`, merge-tree clean.
GitHub: build × 3 OS, unit × 2 shards, bun-test, adapter-test, audit, doc-coverage, docs-build all SUCCESS. PR CLEAN/MERGEABLE.
Author: kagura-agent (fork). Lead: codex-mini0. Aux: First-principles-0. Coordination: pr-monitor.
Xiaohongshu image-note publishing reliability fixes for creator center UI (legacy raw-Error write command, not a typed-error migration).
## 3 changes (one publish-path repair)
1. **Open creator publish in image mode**: append `target=image` to the publish URL so it loads directly in image mode instead of default
2. **Exact `图文` tab priority**: prefer exact tab text matching before broad `startsWith/includes`, reducing parent-container misclicks while keeping fallback for UI wording variants
3. **DataTransfer fallback for `Chrome Not allowed`**: when CDP `setFileInput` returns the permission/bridge denial error, fall through to the existing DataTransfer upload path (CDP-first remains primary to avoid base64 bridge/payload limits)
## Lead hardening (`edf8107d`)
Added `clis/xiaohongshu/publish.test.js` regression coverage for all three claimed behaviors:
- `target=image` creator URL locked
- exact tab text matched before broad fallback
- `Chrome Not allowed` falling into DataTransfer path
## Better-solution evaluation (lead + aux 一致)
- **CDP-first kept**: CDP avoids base64 payload/bridge limits; `Not allowed` is a known permission failure class where fallback is appropriate. DataTransfer-first would weaken the common path and reintroduce large-payload fragility.
- **Exact tab text first**: XHS creator markup is private and volatile, selector-only alternative not clearly more stable. Exact text reduces misclicks while broader fallback + post-click `video_surface` check preserve resilience for wording shifts. If exact text disappears, command fails fast with screenshot instead of silent video-mode publish.
- **Scope boundary self-imposed**: not expanding to typed-error migration (publish.js is legacy raw-Error and typed-error-lint already accounts for it).
## Verification
Local: xiaohongshu publish tests `12/12`, typecheck, build/manifest, docs:build, typed-error-lint `189/189`, silent-column-drop `103/103`, doc coverage `140/140`, node --check, git diff --check.
GitHub: build × 3 OS, unit shards, bun-test, adapter-test, audit, docs-build, doc-coverage all SUCCESS. PR CLEAN/MERGEABLE.
Author: E2ern1ty (fork). Lead: codex-mini1. Aux: First-principles-1. Coordination: pr-monitor.
* chore(release): pre-release P0/P1 cleanup
P0 fixes:
- delete src/analysis.ts (179 lines, 0 importers across src/clis/extension)
- remove dead OPENCLI_DIAGNOSTIC negative test assertion
- rename OPENCLI_BROWSER_TIMEOUT to OPENCLI_BROWSER_IDLE_TIMEOUT — the env
controls workspace lease idle release, not command runtime; old name was
misleading and undocumented (no fallback needed)
- add 'fill' to validate.ts KNOWN_STEP_NAMES so adapters using PR #1222's
fill pipeline step do not trip "unknown step name" warnings during validate
P1 fixes:
- BrowserConnect daemon-not-running hint: replace stale "make sure port is
available" with actionable "run opencli doctor / opencli daemon restart"
- TimeoutError hint: lead with --timeout flag, demote env var to secondary
* fix(validate): derive step allowlist from pipeline registry
@pr-monitor flagged the prior "add 'fill' to KNOWN_STEP_NAMES" fix as
treating only the symptom — two parallel hand-maintained lists will keep
drifting whenever a new pipeline step is registered.
Address the root cause: pipeline/registry.ts now exports
`getRegisteredStepNames()` and validate.ts builds KNOWN_STEP_NAMES from
that. Adding a step via `registerStep()` automatically allowlists it.
* test(validate): regression guard for pipeline step allowlist linkage
@pr-monitor follow-up: lock the validate ↔ pipeline registry linkage at
the test layer so future drift is caught immediately.
Changes:
- recompute KNOWN_STEP_NAMES per-call (was const at module load) so
steps registered after validate.ts import (plugins, dynamic registration)
are honoured
- add src/validate.test.ts with 3 cases:
1. every step name from getRegisteredStepNames() exists
2. an adapter using every currently registered step does not warn
3. a step registered at runtime is automatically allowlisted by
validate without any source change to validate.ts
* fix(capabilityRouting): add fill to BROWSER_ONLY_STEPS
Same double-list drift pattern as validate.ts KNOWN_STEP_NAMES (audit
follow-up flagged in this PR's evolution thread). The fill step was
registered in pipeline/registry.ts (PR #1222) but never added to the
browser-only allowlist in capabilityRouting.ts.
Concrete impact:
- shouldUseBrowserSession() didn't recognize a `[{ fill: ... }]` pipeline
as needing a browser, so PUBLIC adapters using fill could end up
without a page and crash inside stepFill at `page!.fillText(...)`
- pipeline/executor.ts's per-step retry policy (BROWSER_ONLY_STEPS gets
2 retries on transient errors, others get 0) skipped fill — losing
retry coverage on a DOM-touching step
Fix:
- add 'fill' to BROWSER_ONLY_STEPS
- add a documenting comment explaining BROWSER_ONLY_STEPS is the
browser-touching subset of registered steps (not the full set)
- export _validateBrowserOnlyStepsAgainstRegistry() so the test layer
catches the inverse drift (browser-only step that no longer exists)
- 3 new tests in capabilityRouting.test.ts:
* pipeline with fill routes to browser session
* BROWSER_ONLY_STEPS subset of registered step names
* fill is in both lists
This addresses @pr-monitor follow-up #3 (audit similar double-list
patterns) for the obvious in-scope candidate. Other candidates outside
this PR's scope: build-manifest serialization vs registry shape, error
code unions vs lint baselines.
* test(validate): use Strategy.PUBLIC enum instead of string cast in regression test
Self-review nit: `strategy: 'public' as never` worked but bypassed the
typed CliOptions union. Use `Strategy.PUBLIC` so the test exercises the
real public API.
Wire up the standard browser-LLM command surface for Yuanbao, matching the
recently shipped chatgpt + claude + qwen baselines:
- status — login + current model + (agentId, convId) + URL
- read — render the visible conversation as User/Assistant rows
- detail — open `<agentId>/<convId>` and read its messages
- history — list sidebar conversations with stable IDs
- send — fire-and-forget, returns once the send button has been clicked
Refactor `ask.js` to share helpers (`sendYuanbaoMessage`, `normalizeBooleanFlag`)
with the new commands via `shared.js`, keeping the public ask behavior intact.
Notable bits:
- `parseYuanbaoSessionId` accepts only full chat URLs or `<agentId>/<convId>`
pairs — Yuanbao chat URLs encode both, and silently opening the wrong agent
on a bare UUID is a worse failure mode than throwing. URL regex anchored
with `(?:[/?#]|$)` so 37+ char tails reject rather than truncate.
- `sendYuanbaoMessage` polls the send button (up to 3s) for the React
re-render that drops `style__send-btn--disabled___*` after composer input —
a fixed wait raced the debounce and produced silent no-op clicks.
- `getYuanbaoMessageBubbles` uses `data-conv-id`/`data-conv-idx`/
`data-conv-speaker` attributes for stable per-turn identity (was relying
on innerHTML alone).
- Status surfaces both human label (`Yuanbao`) and `dt-model-id`
(`hunyuan_gpt_175B_0404`) — sentinel strings would silently look like a
real model name; null is the typed-unknown signal.
Verified: 25 unit tests pass; targeted live smoke for status/read/detail/
history/new/send + ask round-trip on yuanbao.tencent.com.
* feat(qwen): add detail command + fix stale message bubble selector
`getMessageBubbles` was matching `[data-msgid="<id>-question|answer"]` from an
older Qianwen frontend. The reshipped DOM no longer carries that attribute on
chat turns; `[data-message-id]` now lives on citation cards inside assistant
responses, so the old selector silently returned an empty list and `qwen read`
had been silently broken.
Rewire to walk `[data-chat-question-wrap]` and `[data-chat-answers-wrap]` in
DOM order (correct Q/A interleaving) and synthesize stable IDs from the
nearest sibling `data-req-id` so `waitForAnswer.seenAssistantId` and
read/ask/detail dedupe paths keep working. Verified live against an existing
conversation: 3 user turns + 3 assistant turns extracted; old selector
returned 0.
`qwen detail <id|url>`: open a specific conversation by ID or full chat URL,
poll up to 20s for the transcript to render, return Role/Text rows. Adds
`parseQianwenSessionId` (5 unit tests covering ID/URL parsing + ArgumentError
on malformed input). Reuses the same site-level browser session as `read`/
`ask` so consecutive calls continue in the same Qwen tab.
- clis/qwen/detail.js (new)
- clis/qwen/utils.js (parseQianwenSessionId + getMessageBubbles rewire)
- clis/qwen/utils.test.js (new)
- docs/adapters/browser/qwen.md (detail entry + options/columns)
- cli-manifest.json (regenerated)
* fix(qwen): anchor URL regex to reject 33+ hex tail truncation
codex-coder review on PR #1390 caught that
`https://www.qianwen.com/chat/<33+ hex>` would silently truncate to the
first 32 chars and open the wrong conversation. Adds end-of-input /
slash / query / fragment boundary to the URL match group and two new
unit-test cases (digit tail + letters tail) covering the truncation gap.
Add ChatGPT web ask/send/read/history/detail/new/status alongside existing image support. Tighten ChatGPT web helper selectors and typed error contracts, update docs/changelog, regenerate manifest, and seed local ChatGPT verify fixtures for ask/read.
* test(gov-policy): JSDOM-against-frozen-fixture tests for in-browser extractors
Applies the pattern documented in skills/opencli-adapter-author/references/jsdom-fixture-pattern.md
(introduced in #1319 alongside the dianping reference test in #1313) to the
gov-policy adapter.
Refactor: the inline IIFE inside `page.evaluate` template literal is hoisted
to a top-level `extractSearchRows` / `extractRecentRows` function using bare
`document` / `location`. Same code now runs identically in:
- the live browser (injected via `${extractor.toString()}`)
- JSDOM unit tests (with `globalThis.document` / `globalThis.location` swapped)
Tests:
- 6 new cases in clis/gov-policy/gov-policy.test.js (was commands.test.js).
- 3 representative search result cards (1 with real article snippet, 2 with
only publish-time in `.description`) and 5 recent listing rows in the
fixtures.
- ok:false fallback path covered for both extractors.
- Lock-in: `要闻` type-tag prefix fusion in title and empty-source contract
on recent listings (no `.source` / `.from` elements on that page) are
asserted explicitly so a future selector tweak can't silently change them.
Reverse-validated against two buggy variants per the reference doc:
breaking the title selector and stripping the `要闻` prefix both fail the
JSDOM assertions with helpful diffs.
Fixture sanitization follows the reference doc step-by-step: scripts /
styles / iframes / comments / preload links stripped, image srcs replaced
with `placeholder.png`, trimmed to the minimum subtree that exercises the
extractor (3 search items, 5 recent rows), all whitespace-only lines
removed.
* fix(gov-policy): use typed errors for touched commands
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
* enrich(coupang): add product detail cmd + replace silent clamp/sentinel/Error with typed errors
Two enrichment changes plus three silent-failure fixes on top of existing
search / add-to-cart.
New cmd: coupang product
─────────────────────────
Pairs with search as the listing↔detail round-trip target. Reads a logged-in
product page and extracts a single canonical row with price, original_price,
discount_rate, rating, review_count, seller, brand, rocket, delivery_promise,
image_url, url. Three-source extractor (JSON-LD Product schema → bootstrap
globals → DOM) merged in priority order, mirroring the search.js pattern.
The columns use string|null typing — null means "upstream did not provide
this field on this product" (e.g. some items have no original_price).
Failures (login wall / page mismatch / page failed to render) raise typed
errors instead of silently returning empty rows, so callers can treat any
returned row as real data.
Search column shape: added product_id
─────────────────────────────────────
Listing must pair with detail by id. The data was already extracted by
normalizeSearchItem; only the columns array needed updating so the field
projects through to the rendered row. Per the listing-id-pairing convention
(PR #1297) the new column lets agents round-trip rows directly into
`coupang product` without re-scraping URLs.
Silent-failure fixes
────────────────────
1. search --limit silent clamp.
Old: `Math.min(Math.max(Number(kwargs.limit||20),1),50)` silently
rewrote `--limit 999` to 50 and `--limit 0` to 1.
New: `parseLimitArg(raw, 20, 50)` throws ArgumentError on out-of-range
/ non-integer / negative input. Same convention as the typed-fail-fast
memory & PR #1289.
2. search --page silent clamp.
Old: `Math.max(Number(kwargs.page||1),1)` silently lifted negative pages.
New: parsePageArg throws ArgumentError on non-positive input.
3. Generic `throw new Error(...)` → typed errors.
- Empty query, unsupported --filter, missing --product-id/--url
→ ArgumentError
- Login wall detection → AuthRequiredError('coupang.com', ...)
- Empty result / filter-not-rendered → EmptyResultError
- PRODUCT_MISMATCH / OPTION_REQUIRED / button-not-found / unknown
ack failure (add-to-cart) → CommandExecutionError
- The PRODUCT_MISMATCH and `actualProductId || 'unknown'` sentinel were
also fixed (silent-sentinel was the audit hit there).
Coverage
────────
- 21 contract assertions in clis/coupang/coupang.test.js covering
parseLimitArg / parsePageArg (no silent clamp), registry shape (search has
product_id, product is read-class with expected columns, add-to-cart is
write-class), and typed-error pre-flight rejections (empty query / bad
filter / out-of-range limit & page / missing detail args).
- Manifest 763 → 764 (+1 entry: coupang/product).
- Audits: typed-error-lint 196 → 194 (resolved 2 silent-clamp/sentinel
baseline entries; baseline updated). silent-column-drop 103/103 unchanged.
* fix(coupang): tighten product id and browser errors
* fix(coupang): require real product urls
* refactor(linux-do): remove deprecated hot/category/latest compat shims
The three shims have been pure backward-compat wrappers since linux-do/feed
became the unified entrypoint. With no stable release commitment to preserve,
they are pure surface cost: 3 manifest entries, 3 deprecated branches in help
output, and a `buildLinuxDoCompatFooter` helper that exists only to feed them.
- delete clis/linux-do/{hot,category,latest}.js
- drop now-orphaned `buildLinuxDoCompatFooter` from feed.js and unexport
`executeLinuxDoFeed` (no external consumers remain)
- remove the Compatibility section in docs/adapters/browser/linux-do.md
- regenerate cli-manifest.json (-125 lines)
BREAKING CHANGE: `opencli linux-do hot|category|latest` are removed. Use
`opencli linux-do feed --view top --period <period>`,
`opencli linux-do feed --category <id-or-name>`, and
`opencli linux-do feed --view latest` instead.
* fix(linux-do): finish compat shim removal
* refactor(runtime): unify command timeout into a single --timeout arg
Drop the cli-level `timeoutSeconds` build-time ceiling field. A command
now opts into runtime-enforced timeouts purely by declaring an arg named
`timeout`; the user-facing `--timeout` value (its default or override)
is the single authoritative knob, used both by the adapter polling loop
and by the runtime ceiling (with a 30s padding for return + closeWindow
+ trace export).
Behavior:
- Browser commands without a `--timeout` arg fall back to
OPENCLI_BROWSER_COMMAND_TIMEOUT (default 60s, unchanged).
- Non-browser commands without a `--timeout` arg now run unbounded
rather than against the previously implicit `timeoutSeconds` cap.
Affected commands keep their old caps via newly added `--timeout` args.
- LLM adapters (gemini/claude/deepseek/doubao/qwen/yuanbao ask) keep
their current `--timeout` defaults; the runtime ceiling is now strictly
more generous (userTimeout + 30s vs. the previous 180s cap), so
`--timeout 600` actually buys 600s of polling rather than dying at 180s.
Closes the design discussion that started from PR #1227, which proposed
a per-site `OPENCLI_GEMINI_ASK_TIMEOUT` env var to work around the same
underlying mismatch.
* fix(timeout): wire --timeout arg into chatgpt/gemini image adapter polling
codex-coder review on PR #1364 caught that the new --timeout arg I added
to chatgpt/image and gemini/image only drove the runtime ceiling — the
adapter still hardcoded `const timeout = 120`, so users passing
--timeout 240/600 saw runtime allow 270s/630s but the adapter stop
polling at 120s. That recreated the same single-knob mismatch this PR
was meant to delete.
Also add the browser-path runWithTimeout assertion codex-coder flagged
as missing: a browser command with --timeout default=5 must call
runWithTimeout with timeout: 35; a browser command without --timeout
arg must fall back to DEFAULT_BROWSER_COMMAND_TIMEOUT.
Image adapters now read kwargs.timeout and reject non-positive-integer
values with ArgumentError (no silent fallback). chatgpt/image.test.js
updated to pass an explicit timeout when calling .func directly (the
test bypasses arg coercion).
* fix(runtime): reject invalid timeout ceilings
* fix(timeout): normalize timeout args to integer values
* fix(timeout): preserve remaining command ceilings
* fix(runtime): validate timeout before browser setup
* enrich(toutiao): hot board + bug fixes (silent column drop, partial render)
Per WAWQAQ "丰富现有 adapter" pivot — Phase 2 site #3.
## New command
- `toutiao hot` (Strategy.PUBLIC, browser:false) — public homepage hot
board via the toutiao.com hot-event/hot-board endpoint. No login required.
Returns 8 stable columns (rank/id/title/query/hot_value/label/url/image).
## Bug fixes for `toutiao articles`
- **Silent column drop fixed**: `parseToutiaoArticlesText` previously
did `if (title && stats) push(...)`, silently dropping any row where
the stats span hadn't finished rendering by the time page.innerText
was read. Slow-render bugs were invisible — adapter looked "complete"
while writers saw extra rows in the dashboard. Partial rows now
surface with `null` stat columns.
- **Silent clamp on `--page` removed**: out-of-range / non-integer
values raise `ArgumentError` with explicit bounds [1, 4]. Same
validation reused by both `articles` and `hot` via `parseArticlesPage`
/ `parseHotLimit` in `utils.js`.
- **Empty result typed**: zero-row scrape now raises `EmptyResultError`
instead of returning `[]` silently (would otherwise look like a
legitimate "no articles" response).
## Refactor
- Parser logic extracted to `clis/toutiao/utils.js` (alongside hot-row
mapping, validators, and the hot-board URL constant).
- `articles.js` switches from declarative `pipeline:` to imperative
`func` form so `parseArticlesPage` validation can run before the
navigation step (declarative pipeline can't pre-validate args).
- Strategy is now explicit: `Strategy.COOKIE, browser: true` for
articles (creator dashboard is logged-in only).
## hot field map
`ClusterIdStr` (or numeric `ClusterId`) → id; `Title` → title;
`QueryWord` → query (falls back to title); `HotValue` → hot_value
(non-negative numeric, else null); `Label`, `Url`, `Image` →
respective columns. `pickImage` walks `Image.url` → first truthy
`Image.url_list[]`. Empty-title rows are dropped (returns null) before
ranks are densely re-assigned 1..N.
## Tests
29 contract assertions across `parseArticlesPage` / `parseHotLimit` /
`parseToutiaoArticlesText` / `mapHotRow` + registry-level shape checks
+ `hot` adapter func behaviour (typed errors / no silent clamp / fetch
failure paths / dense-rank).
## Audits
- typed-error-lint: 196 = 196 (unchanged baseline)
- silent-column-drop: 103 = 103 (unchanged baseline)
- listing-id-pairing: hot has `id` column (round-trippable when a
detail command lands later); advisory list unchanged.
## Manifest
757 → 758 entries (+1 for `hot`).
## Doc
- index.md: toutiao mode 🔐 → 🌐/🔐 (hot is public, articles is logged-in)
- toutiao.md: per-command mode/domain table + column docs + prerequisites
* fix(toutiao): tighten hot and articles contracts
* fix(linkedin): surface detail_error on --details (no silent catch / no silent empty)
The previous --details enrichment path had two indistinguishable failure modes
that both produced `description: '', apply_url: ''`:
1. `if (!job.url)` early return — row had no jobId, so we couldn't navigate.
2. `} catch {}` — page.goto / page.evaluate threw (network, timeout, parse error).
Callers couldn't tell "upstream had no description" from "we failed to fetch",
and the catch swallowed every error without logging. For an enrichment that
costs one page navigation per row, silent failure is especially harmful — users
just see an empty cell with no way to debug.
Fix: replace empty strings with `null` for missing/failed rows, add a new
`detail_error` column (string|null) carrying a short typed reason:
- 'no url' — row had no jobId
- 'fetch failed: <msg>' — page.goto / page.evaluate threw
- 'missing description' — page loaded but body was empty
- null — success
Every failure is also logged to stderr with the offending URL so debugging is
possible. Per-row failures still don't abort the batch (the original intent),
but they're now visible.
Tests: 13 new contract assertions in clis/linkedin/search.test.js covering
parseCsvArg, mapFilterValues (ArgumentError on unknown values), decodeLinkedinRedirect,
and 5 enrichJobDetails paths (no-url / goto-throw / empty-description / success /
multi-row-mixed). Added `export const __test__` for testability.
Audits clean: typed-error-lint 196/196, silent-column-drop 103/103.
* fix(linkedin): fail fast on auth walls
* enrich(ctrip): hotel-suggest + bug fixes (silent clamp, dropped columns, fake URL)
Per WAWQAQ "丰富现有 adapter" pivot — Phase 2 site #1.
## New command
- `ctrip hotel-suggest` — surfaces hotel-context suggestions (cities,
business areas, individual hotels) via the same backing endpoint with
searchType=H. Distinct from `ctrip search` (searchType=D) which returns
destinations / scenic spots / railway stations.
## Bug fixes for `ctrip search`
- **Silent clamp on `--limit` removed**: out-of-range values (≤0, ≥51,
non-integer) now raise `ArgumentError` with explicit bounds rather than
silently snapping to [1, 50].
- **Silent column drop fixed**: previously the adapter discarded `id`,
`cityId`, `cityName`, `provinceName`, `countryName`, `lat`, `lon`, `eName`
and `displayType` from upstream rows. Now all are surfaced as stable
columns.
- **Fake URL fixed**: previously `url` was always `''`. Now constructs
canonical Ctrip URLs by `type` (City / Markland / Hotel / Zone / RailwayStation)
and returns `null` (no silent fabrication) for unknown types.
- **In-band error envelope typed**: `Result: false` payloads now surface
as `COMMAND_EXEC` (was previously not handled — adapter returned empty
rows).
## Doc fix
- `Mode: 🔐 Browser` → `🌐 Public` (search uses public API, no login)
- Add `hotel-suggest` to commands table in both `docs/adapters/index.md`
and `docs/adapters/browser/ctrip.md`.
## Coords picker
Mainland China rows ship `gdLat`/`gdLon` (gaode); international rows ship
`gLat`/`gLon` (wgs84). Adapter picks the first non-zero pair (zero is the
upstream sentinel for "missing"); returns `null` if all variants are zero.
## Tests
25 contract assertions across `parseLimit` / `pickCoords` / `buildUrl` /
`mapSuggestRow` + registry-level checks for both commands (Strategy /
shape parity / typed errors / no silent clamp).
## Audits
- typed-error-lint: 196 = 196 (unchanged baseline)
- silent-column-drop: 103 = 103 (unchanged baseline)
- listing-id-pairing: advisory only (search has `id` round-trip column)
## Manifest
757 → 758 entries (+1 for `hotel-suggest`).
* fix(ctrip): wrap suggest fetch and json failures
Closes#1334.
Exposes viewport overrides for `opencli browser screenshot` so an adapter or
ad-hoc shell user can render a page at a fixed width and capture the full
scrollable height. The ljg-card HTML to PNG pipeline use case.
Behavior:
- `--width W` only overrides device-metrics width; height is left unchanged.
- `--height H` only overrides height (ignored under `--full-page`).
- `--full-page` keeps the existing `captureBeyondViewport` shortcut.
- `--full-page --width W` first reflows at W, then re-overrides to (W, contentH)
so the captured image reflects the layout at the requested width.
- Override is always cleared in `finally`, including on capture failure.
* feat(deepseek): add detail and send commands for explicit conversation control
doubao already ships `detail <id>` and `send` for ID-explicit conversation
read/write; deepseek had only `read` (current page only) plus the
implicit-resume `ask`. Adding both gives users a stable handle when they
know the conversation ID, without going through `ask`'s resume detection
or its full prompt-then-wait pipeline.
`deepseek detail <id>`:
- parses a bare UUID or any URL containing `/a/chat/s/<id>`,
- rejects malformed input via `ArgumentError` before any browser
navigation,
- navigates to `https://chat.deepseek.com/a/chat/s/<id>` and returns
the visible message list,
- throws `EmptyResultError` when the conversation has no rendered
messages.
`deepseek send <id> <prompt>`:
- takes the conversation id as a required positional, because the
framework runs each browser command in an ephemeral per-command
workspace (a fresh tab) and there is no shared "current conversation"
across commands; the navigation must be explicit,
- drives input through CDP `Input.insertText` via `page.nativeType`,
mirroring the doubao adapter (#1278); `execCommand('insertText')` plus
a synthesised input event leaves the React-controlled state desynced
on a freshly-opened tab and the resulting click silently no-ops,
- keeps the verification loop inside the same `page.evaluate` so the
framework cannot close the tab mid-flight; counts user-class bubbles
by text-match (DeepSeek virtualises the message list, so a numeric
bubble-count check is unreliable),
- throws `CommandExecutionError` with a specific reason when the
textarea did not populate, the send button stayed disabled, the
bubble never settled, or the optimistic render rolled back during
a 3s settle window,
- treats "Promise was collected" from the post-click eval as success,
matching the existing pattern in `ask --file`.
Helper `parseDeepSeekConversationId` is exported from utils.js so the
same parser feeds both commands and round-trips the canonical lower-case
ID.
Tests:
- utils.test.js: 5 cases covering bare UUID, upper-case
normalisation, URL extraction with and without query string, empty /
null / whitespace input, and non-UUID rejection.
- detail.test.js: 5 cases covering registration, navigation +
message return, URL normalisation, ArgumentError before browser
navigation, and EmptyResultError on no-messages.
- send.test.js: 7 cases covering registration, ArgumentError on bad
id, full happy-path through nativeType + IIFE verification, the
textarea-mount timeout, missing nativeType helper, focus failure,
IIFE-reason translation to CommandExecutionError, and the
"Promise was collected" success path.
Manifest auto-regenerated to register both commands.
Live-verified end-to-end against my own DeepSeek session:
- `detail` returns the canonical message list for a bare UUID, parses
a full chat URL, and rejects malformed IDs before any browser
navigation,
- `send` lands the prompt as the latest user message in the target
conversation and gets an AI response back; reload of the
conversation page in a separate tab confirms the message persisted
server-side.
* docs(deepseek): document detail and send commands
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
Keep the owned automation container window warm across lease release. Non-final owned leases close their tab; the final owned lease resets its tab to about:blank as a reusable placeholder. Update browser close wording to describe lease release rather than window closure.
Closes#1342.
`opencli deepseek ask` (without --new) used to resume the most recent
conversation by clicking the first `a[href*="/a/chat/s/"]` in DOM order
after a fixed 2-second wait.
Two bugs:
1. Pinned conversations sit in their own DOM section ("置顶") that
renders above "30 天内" and friends. Click-first-anchor lands on the
pinned thread, not the user's most recent. Reproduced live by
pinning a conversation through the sidebar context menu and
observing that the existing logic targets it instead of the most
recent non-pinned thread.
2. The 2s wait is fixed. On a slow network the sidebar has not
populated yet, the click is a no-op, and `ask` silently falls
through to the new-chat path. The user typed "follow up" and
a brand-new conversation gets created.
Replace the click-first-anchor + fixed wait with a new helper
`pickResumeUrl(page)` in utils.js that:
- polls the sidebar for up to 10s (5 attempts × 2s),
- identifies pinned anchors by a text-based check on the section
header (`/^\s*(置\s*顶|Pinned)\s*$/i`); DeepSeek's CSS-module
class names are randomized per build, so the text is the only
stable signal,
- returns the URL of the first non-pinned anchor (or falls back to
the first overall if every visible anchor is pinned),
- returns null if no anchor surfaces in time.
`ask.js` calls the helper and `page.goto`s the returned URL. When the
helper returns null, `ask` now throws a `CommandExecutionError`
instead of silently navigating to a fresh chat. The user gets a clear
"pass --new" hint and their prompt is never sent to a wrong target.
Tests:
- utils.test.js: 4 cases covering happy path, polling-then-success,
timeout returns null, and a structural assertion that the embedded
DOM walker uses text-based pinned detection.
- ask.test.js: replaced the prior "still selects model when no
conversation to resume" test (which exercised the silent
fall-through) with a fail-fast assertion. Updated the resume-success
test to mock the new helper.
* feat: 11 read adapters across 8 sites (dblp / steam / bbc / devto / lobsters / medium / coingecko / hf)
Round 2 of the adapter expansion sweep. All 11 commands hit public APIs (no
browser, no auth), follow the post-#1332 typed-error / no-silent-failure
discipline, and were live-verified against real endpoints.
New adapters:
- dblp/author : recent publications for one author (resolve PID by name, or pass --pid)
- steam/search : storefront name search (storesearch API)
- steam/app : single app detail (appdetails API; HTML entities decoded)
- bbc/topic : per-topic RSS (8 canonical BBC News feeds)
- devto/latest : /api/articles/latest with --page pagination
- lobsters/domain : stories from a specific source domain (/domains/<d>.json)
- medium/tag : tag RSS (description full-length, no silent truncation)
- coingecko/exchanges : trust score + 24h BTC volume leaderboard
- coingecko/categories : sector buckets with 6 sort options
- coingecko/global : aggregate market totals + BTC/ETH dominance
- hf/paper : single-paper detail by arXiv id (summary, ai_summary, ai_keywords, upvotes)
Also adds clis/steam/utils.js + clis/bbc/utils.js as shared helpers (HTML entity
decode, RSS parsing). All listings carry a round-trippable id where a detail
sibling exists; advise:listing-id-pairing reports zero new violations. typed-
error-lint and silent-column-drop gates both unchanged from baseline.
Manifest: 698 → 709 (+11 entries).
* fix: tighten adapter round2 contracts
* Add uisdc news adapter for CLI
Implements a CLI adapter for fetching the latest AI/design news from uisdc.com. Allows specifying the number of news items to return.
* feat(aibase): add aibase daily news adapter
This file implements a news adapter for AIbase that fetches the latest AI industry news and allows for configurable limits on the number of news items returned.
* fix(news): harden uisdc and aibase adapters
---------
Co-authored-by: jackwener <jakevingoo@gmail.com>
Document the pattern for running opencli on a remote machine while keeping
the daemon and Chrome on the local machine. Reverse-tunnel local 19825
back to the remote (via SSH -R or frp) so the remote opencli still talks
to its own loopback and the daemon never leaves localhost.
Captures the rationale we landed on after reviewing #636: native
extension-to-remote-daemon support is deferred until the daemon protocol
gains authentication; in the meantime this is the safe, zero-code path
that achieves the same outcome.
* feat: add tiktok creator-videos command
TikTok Studio creator content list with views/likes/comments/saves/shares.
Hits the Studio item_list endpoint
(https://www.tiktok.com/tiktok/creator/manage/item_list/v1/?aid=1988) from a
logged-in /tiktokstudio/content session and pages with cursor until limit is
satisfied (server caps size at 50). Username for the resulting video URL is
extracted from the user_text= query param on play_addr / download_info entries,
falling back to scraping a[href*="/video/<id>"] from the Studio page DOM.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(tiktok): regen manifest + replace silent-clamp with ArgumentError
- Regenerate cli-manifest.json (CI gate: must match `npm run build` output)
- Replace `Math.max(1, Number(args.limit) || 20)` and
`Math.min(Math.max(limit, 1), 50)` with an explicit positive-integer
guard + a server-cap-only ternary, per the silent-clamp guidance in
references/typed-errors.md (typed-error-lint baseline is unchanged)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(tiktok): tighten creator videos contract
---------
Co-authored-by: root <root@example.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: jackwener <jakevingoo@gmail.com>
Per WAWQAQ feedback in #OpenCLI thread on the flat "Site adapters (112)" listing:
the bucket conflates real web sites (bilibili, dianping, ...) with desktop apps
(chatgpt-app, chatwise, codex, cursor, discord-app, doubao-app, antigravity, notion).
Group them so agents that fall back to --help can scan by category.
Three buckets, sourced from existing metadata only — no new adapter schema:
- External CLIs: passthrough binaries from loadExternalClis() (docker, gh, vercel, ...)
- App adapters: domain is `localhost` or any non-DNS string (no `.`)
- Site adapters: domain contains `.` (real DNS), or domain is unset (default)
The classifier is one line: `domain.includes('.') ? 'site' : 'app'`. Adapters
without a domain field default to site (most are public web scrapers like
arxiv / wikipedia / spotify / ...).
Verified against the live registry: 7 External CLIs, 8 App adapters
(antigravity, chatgpt-app, chatwise, codex, cursor, discord-app, doubao-app,
notion), 104 Site adapters.
Structured help (-f yaml/json) gains parallel `external_clis` / `app_adapters`
/ `site_adapters` keys; `commands` no longer leaks adapter names.
External CLIs are now hidden from the default Commands listing (mirrors how
site adapters were already filtered) and surfaced in their own section.
- Add clis/test-utils.js with standard createPageMock utility
- Migrate 11 test files to use shared utility (removes ~300 lines of duplication)
- Delete extension/src/cdp.test.ts dead skip test (chrome.scripting.executeScript removed from source)
- Remove clis/pixiv/test-utils.js (superseded by shared utility)
Codify the JSDOM-against-frozen-fixture pattern that PR #1313 introduced
for dianping (and that PR #1318 had to follow up to clean up). The skill
previously had no reference for this category of test, so authors of the
next adapter that hits silent-in-browser-DOM bugs would either reinvent
it or skip it.
Key conventions captured:
- **Mandatory awk 'NF>0' as the final step of fixture creation.** The
blank-line noise that PR #1318 removed (84.6% / 54.8% of file content
in dianping/{shop,search}.html) came from manually stripping
script/style content without collapsing the surrounding newlines.
Skipping this step is the silent quality regression that the next
fixture author would also hit.
- **Trim-to-minimum but never re-flow content.** Some bugs depend on
text-node adjacency without intervening whitespace
(dianping #1312 bug #2: rating "4.8" + reviews "21241条" fused as
"4.821241条"). Pretty-printing the meaningful mega-line would mask
the very condition the test is meant to catch.
- **Reverse-validate the regression guard.** "18/18 tests pass" only
proves agreement with the current implementation, not that the test
would have caught the original bug. Reintroducing the buggy variant
must make the test fail — otherwise the fixture is over-stripped or
the assertion is too loose.
- **__fixtures__/ is the documented exception** to the "no committed
HTML dumps" rule in the skill's "关键约定". Calling that out
explicitly because the rule otherwise reads as "all HTML in repo is
bad," which the dianping fixture pattern intentionally violates for
a real reason.
Background: WAWQAQ in #1313 follow-up thread (`#OpenCLI:36d2f65a`) asked
twice — first about the visible blank-line noise (→ PR #1318 cleanup),
then about the root cause and what should improve in the workflow itself.
This is the workflow improvement.
No new tooling / CI gate / lint introduced (B 1-week gate freeze still
applies). When a fifth fixture site adopts this pattern,
`opencli browser fixture-snapshot` automation can be revisited; until
then, runbook discipline + skill reference is the right scope.
dianping/__fixtures__/search.html and shop.html came out of page.content() with
hundreds of blank lines that JSDOM ignores during parsing — pure visual / disk
noise that bloats reviewer diff and obscures the meaningful DOM subtree the
fixture freezes.
search.html: 372 → 168 lines (-204 lines, -572 bytes)
shop.html: 39 → 6 lines (-33 lines, -64 bytes)
`awk 'NF>0'` keeps every line that has any non-whitespace character, so the
minified mega-line (where the rating-vs-reviews adjacency that triggers #1312
silent fusion lives) is preserved verbatim. dianping.test.js 18 tests still
pass, and reintroducing the buggy `headText.match(/(\d+)条/)` extractor still
makes the regression guard fail with the expected '821241条' (proving the
fusion-bug detection power is intact after the strip).
Per WAWQAQ feedback in #1313 thread; opus independently validated the same
approach before the cleanup.
PR #1312 fixed two silent in-browser DOM bugs that the existing mocked
`page.evaluate` tests could not catch:
1. shop title fallback split on ASCII `[]` while dianping renders
full-width `【】`, so `name` was always empty (or `"undefined"`).
2. headText `\s+` collapse fused rating "4.8" with reviews "21241条",
so a head-wide `/\d+条/` regex captured "4.821241" → 5.
Both bugs only surfaced on live verify; mocked-evaluate unit tests fed
pre-baked results to the func and the real DOM walk never ran.
Make the in-browser extractor logic testable in CI:
- clis/dianping/shop.js, clis/dianping/search.js: extract the IIFE
bodies into top-level `extractShopFields()` / `extractSearchRows()`
using bare `document` / `location`. The live adapters inject these
via `page.evaluate(\`(\${fn.toString()})()\`)` so behavior is
unchanged; both commands re-verified end-to-end against live
dianping (shop returns name=芈重山老火锅(五道口店), reviews=21241,
rating=4.8; search returns 3 result-shaped rows with correct ids).
- clis/dianping/__fixtures__/shop.html (3.4KB), search.html (8.4KB):
sanitized HTML snapshots — scripts/styles/iframes/comments stripped,
img src placeholdered, only structural attributes kept. Trimmed to
the minimum subtree needed to exercise the extractors (search keeps
3 of 15 li cards; shop keeps .shop-head + .desc-info + .review-title
plus full-width 【】 title and headText with the rating/reviews
fusion preserved).
- clis/dianping/dianping.test.js: add a fifth describe block —
"extractors against frozen HTML fixtures" — that loads the fixtures
via JSDOM, swaps `globalThis.document` / `globalThis.location`, and
asserts the post-fix behavior:
* shop: name=芈重山老火锅(五道口店), reviewsRaw=21241条, rating=4.8,
breakdown={口味:4.8,环境:4.8,服务:4.8,食材:4.9}, hours, rank, subway.
* search: 3 rows with correct shop_ids, names, reviewsRaw, priceRaw,
starClass; round-trip through parseReviewCount/parsePrice mappers
to lock in {rating:5.0,reviews:21231,price:109} et al.
* ok:false branches: shop fixture without `.shop-head`, search
fixture with empty `#shop-all-list`.
Manually verified the fixtures would catch the original bugs by running
buggy extractor variants against shop.html — ASCII-bracket fallback
returns `name="undefined"`, and head-wide `/\d+条/` returns `821241条`
(both fail the new assertions).
Test suite grows 14 → 18 passing tests; live verify of both commands
still produces correct output post-refactor.
Pattern intentionally limited to dianping as a reference point. If other
sites with in-browser DOM extraction encounter similar silent bugs, this
JSDOM-against-frozen-fixture pattern can be adopted per-site.
* docs(cases): add three researcher workflow examples
Add use cases under cases/ that exercise the recently-landed
researcher-friendly adapters:
- daily-rl-research-monitor.md uses arxiv recent + openreview venue
+ hf top to compress a morning paper-skim into one shell pipeline.
- find-paper-implementation.md chains arxiv search/paper + dblp
search + hf top + openreview search to map a paper's canonical
record, follow-ups, and community uptake.
- track-conference-papers.md walks openreview venue + reviews to
shortlist accepted papers and digest review threads in batch.
Each file is a real workflow built on commands from #1289 (arxiv
recent), #1294 (openreview), and #1299 (dblp).
* docs(cases): correct venue ids and forum example to ones that return data
The first revision used "ICLR.cc/2026/Conference" and "ICLR 2026 oral"
as venue strings. Both return EMPTY_RESULT today because the venue is
not open. Update each case to use natural-language venue text that
OpenReview currently exposes ("ICLR 2024 oral", "NeurIPS 2025 oral")
and a real forum id (KS8mIvetg2, "Proving Test Set Contamination in
Black-Box Language Models") in the reviews / paper drill-down. Note
the arxiv free-text-search ranking quirk so the worked DPO example
makes sense.
PR #1297 introduced a CI gate that fails when a site has both a listing
and a detail command but the listing rows don't carry an id-shaped column.
The gate came with a 10-entry EXEMPT map (topic-string trending,
profile-attribute rows, UI-only sessions, ...) where each exemption
recorded a "why this listing legitimately doesn't pair" reason.
By the same filter that closed PR #1311 (write-without-delete-pair gate):
Is "listing should pair with detail" a *permanent* anti-pattern, or
case-by-case business judgment?
It's case-by-case. Topic-string listings and profile-attribute rows
genuinely don't pair with a detail command. The fact that we needed an
EXEMPT map with 10 entries and individual reason strings is the smell —
it's not the rule winning, it's the rule failing. Forcing every adapter
PR to either add an id column or file an exemption was a higher cognitive
cost than the silent-loss bugs the rule actually catches.
Changes:
- .github/workflows/ci.yml — drop the "Check listing↔detail id pairing"
step. Other gates (silent-column-drop, typed-error-lint) stay in place.
- package.json — rename the script from `check:listing-id-pairing` to
`advise:listing-id-pairing` to make the advisory nature explicit.
- scripts/check-listing-id-pairing.mjs — drop the `--strict` flag and the
EXEMPT map. The script now always exits 0 and prints an advisory report
of listings that don't carry an id-shaped column. Reviewers/authors use
it as guidance, not a gate.
- docs/conventions/listing-detail-id-pairing.md — rewrite from "MUST" to
"soft convention". Adds an explicit "why advisory, not a gate" section
that lists the legitimate non-pairing categories so future readers know
the rule's boundary.
- docs/developer/ts-adapter.md — match the advisory tone in the
adapter-author guidance.
The doc, the script, and the column patterns table all stay — agents and
adapter authors can still consult them. What's gone is the CI failure and
the per-PR exempt-list maintenance burden.
Net diff: -34 lines (gate + EXEMPT map removed, advisory-tone doc adds
a small "why advisory" section).
Adds a baseline CI gate for convention-audit typed-error lint findings. Also refreshes the silent-column-drop baseline for dianping changes already on main.
The merged adapter had two silent in-browser bugs that the mocked-evaluate
unit tests don't catch — only live verify against www.dianping.com surfaces
them:
1. Shop name returned `undefined`. The fallback parsed `document.title` with
an ASCII-bracket split (`/[\\[\\]]/`) but dianping wraps the name in
full-width brackets `【芈重山老火锅(五道口店)】...`. Switch to a `【...】`
regex so the title fallback actually fires.
2. Reviews returned `5` instead of `21241`. The headText was whitespace-
collapsed to `★★★★★4.821241条...`, fusing the rating and review digits;
a head-wide `/\d+条/` then captured `4.821241` and rounded to `5`. Read
the dedicated `.reviews / .review-num` element ("21241条") instead, with
a `.review-title` "评价(<n>)" fallback.
* feat(dianping): browser adapter — search + shop on www.dianping.com
Adds two browser-mode adapters for the dianping (大众点评) PC site:
- `dianping search "<keyword>" --city <name|id> --limit <n>`: keyword
shop/restaurant search. Returns rank, shop_id, name, rating, reviews,
price, cuisine, district, url. shop_id round-trips into `dianping shop`.
- `dianping shop <shop_id>` (alias `detail`): shop detail sheet
(field/value rows: name, rating, breakdown 口味/环境/服务/食材, reviews,
price, rank, hours, address, subway, features, url).
Both use Strategy.COOKIE on www.dianping.com (the PC site renders search
SSR and does not require JS hydration). m.dianping.com is intentionally
crippled for non-mobile UAs, so it's not used.
Auth detection (utils.detectAuthOrEmpty) inspects both response text and
final URL for the Meituan Yoda captcha redirect (verify.meituan.com) and
the dianping login redirect; raises AuthRequiredError with the captcha
URL embedded so the user can clear it manually in the same profile.
Listing↔detail id pairing: search.shop_id → shop.<id>. Adds 'shop' to
DETAIL_NAMES in scripts/check-listing-id-pairing.mjs so the convention
gate scans this site (35 sites / 78 listings now covered).
* fix(dianping): harden browser failure classification
* fix(dianping): fail on partial missing shop ids
Adds opencli convention-audit for batch convention scanning, with structured output, strict mode, docs, and startup isolation from local user/plugin discovery.
* feat(youtube/xiaohongshu/xiaoe): surface dropped ids/url on listings (sweep)
Round 8 same silent-column-drop class as #1300/#1301/#1302 — row already
emits the id/url field but `columns` array forgot to project it, so table
view drops it and agent loses the chain into detail commands.
- youtube/feed: rename row.videoId → video_id (snake_case convention),
add to columns. youtube/video accepts both URL and id, so url-based
round-trip already worked, but exposing the canonical id removes the
url-parse step for chained calls.
- xiaohongshu/feed: pipeline map already extracts `id` from the homefeed
payload, columns now lists it.
- xiaoe/catalog: pipeline map already projects `url`, columns now lists
it. xiaoe/detail takes a positional url, so this completes the
round-trip explicitly.
Also fixes one camelCase column violation on youtube/feed (videoId vs the
project's snake_case convention as in twitter `is_retweet`/`created_at`,
douban `subject_id`/`photo_id`, hupu `thread_title`).
CI gate `check:listing-id-pairing` ✓ (34 sites, 77 listings, 10 exempt).
typecheck clean. 114 tests pass for youtube + xiaohongshu.
* fix(youtube): keep feed continuation ids after rename
Round 7 — silent-drop sweep. Continues the listing→detail id-pairing
work from #1297. Each row was already extracting these ids/urls
internally; only the `columns` projection was missing, so they showed
up in `-f json` but never on the table view.
| Adapter | Added columns |
|--------------------|-------------------------------------|
| `1688 search` | `item_url`, `member_id` |
| `hupu mentions` | `tid`, `pid`, `url` |
| `douban photos` | `photo_id`, `subject_id` |
| `linux-do tags` | `slug` |
Round-trip wins:
- `1688 search` → `1688 item <item_url>` (item_url is the canonical
detail.1688.com URL); `1688 search` → `1688 store <member_id>`
- `hupu mentions` → `hupu detail <tid>` (and `pid` for the deep link)
- `douban photos` → tied back to the parent movie via `subject_id`
- `linux-do tags` → `linux-do feed --tag <slug>` (slug is the URL form)
No logic change — only the column array. JSON output unchanged.
Tests: 45/45 pass for the four affected sites.
Round 6 — silent-drop audit follow-up. Sibling twitter listings have
been inconsistent about the canonical tweet `id` (rest_id):
- timeline ✓ exposes id
- search ✓ exposes id
- list-tweets ✓ exposes id
- notifications ✓ exposes id
- bookmarks ✗ extracts but drops it from columns
- likes ✗ extracts but drops it from columns
- tweets ✗ extracts but drops it from columns
The `id` is already in the row object — only the `columns` projection
was missing. With the listing↔detail id-pairing CI gate from #1297 now
on main, surfacing `id` makes round-trip into `twitter thread <id>` /
`twitter delete <id>` / `twitter like <id>` work from the table view too
(previously only via `-f json`).
Other field-presentation drift (`name`, `created_at`, `retweets`)
aligned with sibling adapters where those values are already emitted.
Tests: tweets.test.js asserts `toEqual` on the columns array — updated
that assertion. Other twitter tests use `toMatchObject` and pass
unchanged. 81/81 in `clis/twitter/`.
While auditing instagram/facebook/pixiv coverage gaps, found that pixiv
listings already extract `user_id` and construct `url` per row but drop
both fields from the table view (`columns` doesn't list them). The data
is in the row object — only the column projection was missing.
Per the listing↔detail id pairing convention (#1297), surface them so:
- `user_id` round-trips from `ranking` / `search` → `user` / `illusts`
- `url` is the canonical share link for every illust / user record
Changes:
- `ranking`: + user_id, + url
- `search`: + user_id, + url
- `illusts`: + url (user_id is the arg, no need to repeat per row)
- `user`: + url
No behavior change beyond the table view — JSON output already had these
fields, so existing scripts that consume `-f json` keep working.
* feat(dblp): public bibliography adapter — search + paper
Wraps the dblp.org public API:
- `dblp search <query>` → /search/publ/api JSON, projected into one row per hit
- `dblp paper <key>` → /rec/<key>.xml, parsed into a one-row record
Why dblp on top of arxiv/openreview: dblp is the largest, oldest CS
bibliography (7M+ entries) and the only one of the three that consistently
indexes pre-arXiv literature, journal articles, books, and theses. The
canonical record key (e.g. `conf/nips/VaswaniSPUJGKP17`) round-trips
cleanly between the two commands per the listing↔detail convention.
Implementation notes:
- No deps beyond the registry — XML parsed with conservative regexes,
same approach as the arxiv adapter.
- Polite User-Agent per dblp's API guidance; HTTP 429 mapped to a
CommandExecutionError with a "lower --limit" hint.
- Author homonym suffixes (`"Smith 0001"`) trimmed for clean output.
- 39 unit tests cover validators, XML extraction, both commands.
* fix(dblp): fail fast on API status envelopes
* feat(convention): listing↔detail id pairing rule + CI gate
Adds a hard convention: when a site exposes both a listing-class command
(search / hot / top / recent / ...) and a detail-class command (read /
paper / article / view / ...), every listing row MUST surface an id-shaped
column whose value round-trips into the detail command. Without that, an
agent has no way to follow up on a listing row except re-searching by
title or scraping URLs out of band — both of which break the agent-native
contract.
What's in this PR
- docs/conventions/listing-detail-id-pairing.md — full rule, examples
table, why-it-matters, what counts as id-shaped, exemption taxonomy,
how to add an id column to a listing.
- scripts/check-listing-id-pairing.mjs — validator that reads
cli-manifest.json, classifies each entry as listing / detail / other,
and fails when a listing on a site that also has a read-detail command
is missing an id-shaped column. Exemption allowlist records WHY each
pair is exempt so future maintainers know what to verify.
- npm run check:listing-id-pairing — strict-mode wrapper.
- CI: new step in build job runs the validator after the manifest
freshness check on Linux.
- docs/developer/ts-adapter.md — cross-link from the adapter authoring
guide.
- docs/.vitepress/config.mts — sidebar entries for the new conventions
section.
Fixes brought to zero violations
- 1688/search: add offer_id (already extracted, just surfaced)
- bluesky/user: add uri (AT URI round-trips into bluesky/thread)
- tieba/search: add id + url (thread_id already extracted)
- tieba/hot: add url (rows are topics, not threads — url is the
best-effort round-trip handle, doc'd as such)
Exemptions (intentional, doc'd in EXEMPT map with rationale)
- nowcoder/hot, bluesky/trending, twitter/trending — listing rows are
topic strings, not posts.
- lesswrong/user, reddit/user — rows are profile-attribute key/value
pairs, addressed by the username arg.
- discord-app/search — desktop UI session, message ids not extractable.
- notion/search — Strategy.UI Quick Find, page ids not exposed in DOM.
Validator output after this PR: 32 sites scanned, 75 listings checked,
7 exempted, 0 violations.
* fix(convention): tighten listing id gate
* fix(convention): close url-derived id loophole
* feat(indeed): add `search` and `job` adapters (US site)
Adds an Indeed adapter that fills the US job-search gap (alongside
existing 51job / boss-zhipin / linkedin coverage). Both commands run
through a real browser session because Indeed sits behind Cloudflare
and answers bare HTTP fetches with `403` + `cf-mitigated: challenge`.
## Commands
- `indeed search <query>` — keyword job search
- args: `query`, `--location`, `--fromage`, `--sort`, `--start`, `--limit`
- columns: `rank, id, title, company, location, salary, tags, url`
- `indeed job <jk>` (alias `detail`, `view`) — full job posting
- args: `id` (positional, the 16-char hex `jk` from `search`)
- columns: `id, title, company, location, salary, job_type, description, url`
## Listing↔detail id pairing
`search.id` is the Indeed `jk` (job key, 16-char lowercase hex). It feeds
directly into `indeed job <jk>`. Conforms to the listing↔detail id
pairing convention proposed in #1297.
## CF challenge handling
The adapter polls the result selectors for up to 15s after navigation,
giving the browser time to clear the Cloudflare interstitial. If the
challenge is still up after the wait, the adapter throws a
`CommandExecutionError` with a hint pointing the user at the connected
browser to clear it once. Subsequent calls reuse the warmed cookies via
`Strategy.COOKIE`, mirroring the v2ex / boss / linkedin patterns.
## Validation
`utils.js` keeps argument validation pure and unit-testable:
- `requireJobKey` rejects anything that isn't a 16-char lowercase hex
- `requireFromage` only accepts `1` / `3` / `7` / `14` (Indeed's enum)
- `requireSort` only accepts `relevance` / `date`
- `requireBoundedInt(limit, default=15, max=25)` — Indeed serves at most
one page (10 jobs/page); ArgumentError on out-of-range, no silent
clamping, per the typed-error feedback in #1289.
## Tests
18 unit tests in `clis/indeed/indeed.test.js` cover registration,
validators, URL builders, and DOM-card normalizers. Browser-driven
verification stays out of CI by design (CF challenge is interactive).
## Docs
- `docs/adapters/browser/indeed.md` — full adapter doc with prerequisite
CF-challenge notes and listing↔detail id pairing callout.
- Sidebar entry + adapter index row.
* fix(indeed): tighten timeout fail-fast and runtime tests
* fix(indeed): align readiness with search parser
* feat(openreview): add public adapter — search/venue/paper/reviews
OpenReview is the open peer-review platform used by ICLR / TMLR / COLM
and ML workshops. Its v2 API exposes everyone-readable submissions,
reviews, and decisions without auth, so all four commands run with
`browser: false`.
Commands:
- `openreview search <query>` — full-text search
- `openreview venue <venue>` — list submissions; accepts either a venue
display name (matched against `content.venue`, e.g. "ICLR 2024 oral")
or a full invitation id (e.g. "ICLR.cc/2025/Conference/-/Submission")
via `/-/` heuristic; supports offset pagination
- `openreview paper <id>` — single-paper detail with full abstract
- `openreview reviews <forum>` — paper + threaded reviews/decisions/
comments, ordered chronologically with paper lifted to row 0;
classifies notes via invitation tail (REVIEW / DECISION / REBUTTAL /
COMMENT / META_REVIEW / WITHDRAWAL); per-row truncation via
`--max-length` (min 200)
Listing IDs round-trip into `paper`/`reviews`. PDF URLs normalized to
absolute `https://openreview.net/pdf/...`. `pdate` falls back to
`cdate` when missing, formatted as `YYYY-MM-DD`.
All limits/offsets/ids fail-fast with typed errors (`ArgumentError`,
`EmptyResultError`, `CommandExecutionError`) — no silent clamping, no
empty-array fallbacks. fetch + json + non-2xx + 404 are wrapped so
network/API failures never look like empty results.
Tests: 23 unit tests covering the column contract, content extraction,
date/PDF normalization, invitation-vs-venue dispatch, error paths
(network/JSON/HTTP), pagination offset accounting, and the reviews
classifier + section joiner + truncation.
Live-verified against api2.openreview.net for search ("diffusion
model"), venue ("ICLR 2024 oral"), paper (KS8mIvetg2), and reviews on
that paper's full thread.
* fix(openreview): tighten error and review typing
* fix(openreview): stabilize review contracts
Follow-up from PR #1293 review: 'all answers' was misleading because
the implementation is limit-bounded (default 10, max 100) rather than
unbounded pagination. Spell out the actual contract — including the
accepted-answer-outside-page fallback path — so users don't expect
infinite-scroll behaviour.
Non-blocking docs-only change flagged by codex-mini1 + First-principles-1
during #1293 review.
* feat(lobsters): surface short_id + created_at on listings, add `read <short_id>`
Same agent-native gap as the just-merged hackernews PR (#1288):
1. The 4 listings (hot / newest / active / tag) didn't surface each story's
`short_id`. Agents could see the title and a comments URL but couldn't
pass the id back into a follow-up command. Add `id` (= `short_id`) and
`created_at` columns; `created_at` is cheap signal for "how stale is this".
2. There was no way to read a story + comment tree from the CLI. Lobsters
makes this nicer than HN: `https://lobste.rs/s/<short_id>.json` returns
the story plus a flat `comments[]` array where each entry already carries
`parent_comment` and `depth`, so we get the full thread in one HTTP call
and just DFS using the parent map.
`read` mirrors the `hackernews read` shape (POST row + L0/L1/… indented
comments, `[+N more replies]` stubs at depth/limit cutoffs) so the two
adapters feel the same to agents that already learned one. Same typed
fail-fast envelope: `ArgumentError` on bad short_id / non-positive limit /
depth / replies, `EmptyResultError` on 404 or empty body, `CommandExecutionError`
on other HTTP failures.
Tests cover all 4 listings (column shape + map step), `read` registration,
positional arg shape, ArgumentError fail-fast (no fetch on bad input),
EmptyResultError on 404, threaded-tree assembly from a flat `comments[]`,
and the `+N more replies` depth-cutoff path.
* test(lobsters): lock read fail-fast coverage
* docs(lobsters): list read command
* feat(devto): surface article id + published_at on listings, add `read <id>`
Agent-native gap: devto listings (`top`/`tag`/`user`) didn't include the
article `id`, so an agent couldn't round-trip from a listing into a body
read. They also dropped `reading_time` and `published_at`, which are cheap
signals the API gives you for free.
Changes:
- `top` / `tag` / `user`: add `id`, `reading_time`, `published_at` columns
alongside existing rank/title/etc. `user` keeps its no-author shape since
it's already user-scoped.
- New `devto read <id>`: hits `dev.to/api/articles/<id>` and returns one
row with the article body (truncated by `--max-length`, default 20000,
min 100). DEV.to's public API does not expose comments yet, so this is
intentionally a single-row reader rather than a HN/lobsters-style
threaded tree — if/when comments become public we can extend to
POST + L0/L1.
- Typed fail-fast: `ArgumentError` for non-numeric id and for `--max-length`
below 100; `EmptyResultError` on 404; `CommandExecutionError` for other
non-2xx HTTP statuses. No silent clamps.
- Defensive tag normalization: the `/api/articles/<id>` endpoint returns
`tag_list` as a comma-string and `tags` as an array (the opposite shape
from listing endpoints). Caught this on live verification — both shapes
now collapse to a comma-joined string.
Tests: 12 vitest assertions covering listing column shape (all 3) +
register/args/strategy + typed-error fail-fast paths + happy-path body
extraction + truncation marker + alternate tag_list shape.
Live verification: `devto top --limit 3` and `devto read 3602287` both
return the expected agent-native shape.
* fix(devto): harden article read contract
X removed the post-count caption from each cell on `/explore/tabs/trending`.
The adapter still iterated `divs[2..]` looking for a numeric text node and
fell back to the literal string "N/A" when it found none — which was every
row, on every call. We were emitting a silent-wrong column for every result.
Drop the column and the no-longer-relevant scan loop. Add a regression test
on the columns shape so the column doesn't slip back in.
Live runs of `opencli twitter trending` previously returned rows like
`{rank: 1, topic: "...", tweets: "N/A", category: "..."}` — the `N/A` was
not a transient outage, it was structural.
* feat(arxiv): full abstract/authors, surface pdf+categories+comment, add `recent <category>`
`paper` was silently truncating the abstract to 200 chars and dropping all but
the first 3 authors — agents calling it for a paper summary lost data. Stop
truncating, return all authors, and surface the rest of what the Atom feed
already gives us: pdf url (`<link rel="related">`), all `categories`,
`primary_category`, and the author `comment` (page count, conference, etc.).
`search` keeps a compact list shape (no abstract column, but adds
`primary_category`).
New `arxiv recent <category>` lists newest submissions in a category sorted by
`submittedDate desc` — fills a gap (previously you had to know a search term
to surface anything). Validates the category string and rejects malformed
input via `ArgumentError`.
`search` also switches its no-results path from `CliError('NOT_FOUND', ...)`
to `EmptyResultError` to match the convention other public-API adapters use.
Tests cover: command registration, full-abstract / all-authors parsing, XML
entity decoding in titles, pdf/categories/comment extraction, and category
validation.
* fix(arxiv): harden category and limit validation
* feat(hackernews): add `read <id>` and surface item id on every listing
Two related agent-flow gaps in the HN adapters:
1. `top`/`best`/`ask`/`new`/`show`/`jobs`/`search` all carry the HN item
id internally (firebase items are fetched by id; algolia hits include
`objectID`) but drop it before output. Without an id column the agent
can see the title but has no handle to follow up with.
2. There was no way to read a story's discussion. The whole reason an
agent looks at HN is the comments — and that capability was missing.
This PR adds:
- `id` column on every listing adapter (firebase items: numeric id;
algolia search hits: `objectID` string). Existing column order is
preserved otherwise.
- `hackernews read <id>` — public/non-browser adapter that fetches the
story plus a tree of top-level comments + inline replies via
`https://hacker-news.firebaseio.com/v0/item/<id>.json`. Mirrors the
`reddit read` shape (`type/author/score/text`) so agents can use both
with one mental model. HTML-only fields (comment text) are converted
to plain text with anchor URLs preserved.
- Column-contract tests covering all listings + the new read adapter.
- Doc entry under `docs/adapters/browser/hackernews.md`.
Tested locally via `~/.opencli/clis/hackernews/` overrides:
opencli hackernews top --limit 3 # id present
opencli hackernews search rust --limit 2 # id (objectID) present
opencli hackernews read 47999636 --limit 5 # threaded output
* fix(hackernews): typed fail-fast for read
* fix(douban): drop unparseable fields from movie-hot, add id/votes
The chart page (movie.douban.com/chart) only exposes a single comma-joined
text dump in `.pl2 p`, of the shape:
<release_dates...> / <actors...> / <regions...> / <director_zh> /
<runtime>分钟 / <other_titles> / <genres> / <director with English> /
<languages>
The previous `loadDoubanMovieHot` tried to anchor on the release-date
regex and take `parts[releaseIndex - 1]` as director and
`parts[releaseIndex - 2]` as region. That breaks in two ways:
1. Most entries have multiple release dates back-to-back, so the
"anchor minus one" position is itself a date. Director output becomes
`'2025-09-07(多伦多电影节)'` and region is empty — silent wrong data.
2. For entries with a single release date, the offsets land on actor
names, not director / region.
The page does not actually carry a clean director or region per row —
that's only available on the subject detail page. Trying to reconstruct
either from the chart string is the canonical "verify passes but data is
wrong" failure (success-rate-pitfalls §2 sibling DOM contamination).
Fix: drop `director`, `region`, `quote` from the listing. Surface what
the chart page actually provides reliably:
- `id` — extracted from the subject URL, ready for `douban subject`
- `votes` — from `.star .pl` (`(62484人评价)`), useful as popularity signal
- existing `rank`, `title`, `rating`, `year`, `url`
Agents that need director / region should follow up with
`opencli douban subject <id>`, which is already wired for that data.
* fix(douban): fail fast on empty movie hot
* fix(bilibili,reddit): add identifier and url columns to hot lists
Both `bilibili hot` and `reddit hot` previously dropped their per-row
identifier and URL on the way out, breaking the typical agent flow where
the next call needs a `bvid` / `postId` to fetch detail or comments.
- bilibili/hot: add `bvid` and `url` columns (constructed from bvid)
- reddit/hot: surface `postId`, `author`, `url` (already in evaluate but
dropped in map)
Tested via local `~/.opencli/clis/<site>/hot.js` overrides.
* test(bilibili,reddit): lock hot list identifier columns
Move the Browser Bridge daemon WebSocket out of the MV3 service worker and into an offscreen document. Remove the popup/action UI and obsolete extension log forwarding now that doctor is the diagnostic surface.
Pairs with the new collection-create adapter so users (and future
fixture-teardown logic) can clean up saved-post collections from CLI.
- POST /api/v1/collections/{id}/delete/ with multipart module_name=collection_settings
- Accepts collection name (case-insensitive) or numeric collection_id; resolves
via /collections/list/ first so unknown / duplicate names error explicitly
instead of bubbling up a 404 or silently deleting the wrong one.
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>
Hotfix release for the 1.7.17 doctor regression: `opencli doctor` failed connectivity probe with `Browser session is required` because the doctor probe didn't pass a session to the new strict-session browser bridge. Also adds new adapters and adapter fixes that were ready immediately after 1.7.17.
### Bug Fixes
* **doctor** — pass an internal `__doctor__` browser session to the live connectivity probe so `opencli doctor` works again under the explicit-session browser model introduced in 1.7.17. ([#1485](https://github.com/jackwener/opencli/issues/1485))
* **browser** — `--session <name>` is now declared as a `requiredOption` so Commander itself rejects calls missing the flag before runtime, and the help line is marked `(required)` instead of being hidden under `Options:`. ([#1485](https://github.com/jackwener/opencli/issues/1485))
* **doubao/ask** — restore Assistant detection after the 2026-05 DOM refactor. ([#1484](https://github.com/jackwener/opencli/issues/1484))
* **youtube** — request `srv3` format for caption URLs. ([#1422](https://github.com/jackwener/opencli/issues/1422))
Extension bumped to 1.0.12 (workspace → session lease routing, drop `handleSessions` handler). Major simplification pass: browser/adapter session model rewrite, `--workspace` removed, doctor surface trimmed to its core job.
### ⚠ BREAKING CHANGES
* **browser session model** — replace the browser-facing `--workspace` model with explicit `--session <name>` on `opencli browser *`. Browser commands now require a session name, `browser bind`/`unbind` use `--session`, and bind no longer accepts `--domain`, `--path-prefix`, or `--allow-navigate-bound`. Browser primitives keep their session tab by design; the browser namespace no longer exposes `--keep-tab`. ([#1461](https://github.com/jackwener/opencli/issues/1461))
* **adapter site sessions** — replace adapter metadata `browserSession: { reuse: 'site' }` with `siteSession: 'persistent'`, and replace the user override `--reuse <none|site>` / `OPENCLI_BROWSER_REUSE` with `--site-session <ephemeral|persistent>`. Persistent site sessions keep a stable site tab open without idle expiry. ([#1462](https://github.com/jackwener/opencli/issues/1462))
* **doctor** — remove `--no-live` and `--sessions` flags from `opencli doctor`. Doctor always runs the live browser connectivity probe (that's its core job); session enumeration was never part of health diagnosis. The underlying `'sessions'` daemon protocol action and the `BrowserSessionInfo` public type are removed as dead code. ([#1470](https://github.com/jackwener/opencli/issues/1470))
### Features
* **chatgpt** — `ask` and `send` now accept local image paths and upload them through the composer before submitting the prompt. ([#1476](https://github.com/jackwener/opencli/issues/1476))
### Internal
* **extension 1.0.12** — drop `handleSessions` action handler (no remaining consumers after doctor cleanup).
* **extension 1.0.11** — switch Browser Bridge lease routing from user-facing workspaces to explicit browser sessions.
* **external** — register `tg-cli`, `discord-cli`, and `wx-cli` as external CLI integrations. ([#1458](https://github.com/jackwener/opencli/issues/1458))
### Bug Fixes
* **xiaohongshu** — fall back to base64 upload when CDP `DOM.setFileInputFiles` returns `Not allowed` on creator center. ([#1374](https://github.com/jackwener/opencli/issues/1374))
* **chatgpt** — switch to locale-stable send button selector so non-English UIs don't break send. ([#1354](https://github.com/jackwener/opencli/issues/1354))
### Performance
* **adapters** — hoist cookie reads to `page.getCookies` across Tier 1 (25 files), eliminating per-call CDP round trips. ([#1450](https://github.com/jackwener/opencli/issues/1450))
* **twitter** — drop redundant `goto + wait` in adapter steps; framework auto pre-navigates. ([#1451](https://github.com/jackwener/opencli/issues/1451))
* **twitter** — enable `browserSession.reuse: 'site'` on 17 read-only adapters so repeated reads share one tab. ([#1454](https://github.com/jackwener/opencli/issues/1454))
* **browser** — split interactive and automation windows so `opencli browser *` and adapter-driven background commands no longer share one Chrome window; tab groups are isolated by role.
### Internal
* **extension 1.0.10** — rename the adapter-owned Chrome tab group from `OpenCLI Automation` to `OpenCLI Adapter`. ([#1457](https://github.com/jackwener/opencli/issues/1457))
* **docs** — list `tg-cli`, `discord-cli`, `wx-cli` in External CLI README sections. ([#1459](https://github.com/jackwener/opencli/issues/1459))
Extension bumped to 1.0.9 (Accessibility.enable allowlist + downloads permission + cross-origin frame target attach for AX). Major Browser Agent Runtime release: full Phase 0/1/2 alignment with `vercel-labs/agent-browser` model — CDP-primary input, AX snapshot/refs with stale recovery, semantic locators across all primitives, full form toolbelt (hover/focus/dblclick/check/uncheck/upload/drag/wait-download), annotated screenshots, and same-origin iframe AX routing. Cross-origin OOPIF AX is best-effort (Chrome extension API limitation).
### ⚠ BREAKING CHANGES
* **browser lifecycle** — replace `--focus` / `OPENCLI_WINDOW_FOCUSED` with `--window foreground|background` / `OPENCLI_WINDOW`, and replace `--live` / `OPENCLI_LIVE` with `--keep-tab true|false` / `OPENCLI_KEEP_TAB`. `opencli browser *` defaults to a foreground window and keeps its tab; browser-backed adapter commands default to a background automation window and release their tab unless the adapter uses site-level reuse.
### Features
* **help / browser** — `opencli browser --help -f yaml|json` now emits a structured, agent-ready index of all browser leaf commands (including nested `tab`, `get`, and `dialog` commands), their positionals, command options, namespace options, and root global options. Individual browser commands also support structured help, backed by a shared Commander option/argument spec extractor.
* **help / built-in namespaces** — `opencli daemon|plugin|adapter|profile --help -f yaml|json` now emit the same structured payload as `browser`. One agent call returns every leaf's positionals, options, descriptions, and global options — no per-leaf `--help` follow-ups needed. Original namespace descriptions are preserved through `applyRootSubcommandSummaries()` via a snapshot at namespace declaration time.
* **browser state** — add opt-in AX snapshot refs via `browser state --source ax`, including backend-node click resolution and role/name stale-ref recovery for the Phase 0 browser-agent runtime prototype.
* **browser state** — AX snapshots now include same-origin iframe refs, and `browser state --compare-sources` prints DOM-vs-AX observation metrics for the Phase 1 default-source decision without dumping page contents.
* **browser locators** — `browser find`, `browser click`, and `browser get text|value|attributes` now accept semantic locator flags (`--role`, `--name`, `--label`, `--text`, `--testid`) so agents can act on common controls without a separate state-ref lookup.
* **browser locators** — semantic locator flags now work across input/action primitives (`type`, `fill`, `select`, `hover`, `focus`, `dblclick`, `check`, `uncheck`, `upload`) plus prefixed `--from-*` / `--to-*` locators for `drag`.
* **browser actions** — add `browser hover`, `browser focus`, and `browser dblclick` primitives backed by the same target resolver and CDP input path as `browser click`.
* **browser actions** — add `browser check` and `browser uncheck` primitives that ensure checkbox / radio / aria-checked controls reach the requested state instead of blindly toggling.
* **browser upload** — add `browser upload <target> <file...>` to attach local files to `input[type=file]` targets through CDP `DOM.setFileInputFiles`, with local path validation and file-input verification.
* **browser actions** — add `browser drag <source> <target>` for CDP mouse drag sequences between two resolved element centers.
* **browser wait / extension 1.0.8** — add `browser wait download [pattern]` backed by Chrome's downloads lifecycle API, so agents can wait for file downloads by filename/URL pattern and receive completed/failed download metadata.
* **browser state / extension 1.0.9** — AX snapshots can now route same-origin iframe refs through `frameId`. Cross-origin OOPIF AX routing is best-effort because real Chrome extension smoke tests show `chrome.debugger` may not expose attachable iframe targets to extensions.
* **browser screenshot** — add `browser screenshot --annotate`, which refreshes DOM refs and overlays visible `[N]` labels on the screenshot so visual inspection maps back to `browser click <ref>` targets.
### Bug Fixes
* **browser click** — `browser click` now prefers CDP `Input.dispatchMouseEvent` over DOM `el.click()`, so custom dropdowns that depend on pointer/mouse events (Radix, shadcn, Material UI, Mercury-style category pickers) open and select reliably while retaining JS click as a fallback for older backends or zero-rect targets.
* **browser state / extension 1.0.7** — `browser state --source ax` now enables the CDP Accessibility domain before reading the AX tree, fixing real-Chrome snapshots that previously returned only `RootWebArea` with zero refs.
* **help / build** — every positional arg must now declare a non-empty `help` string. The build-manifest step fails closed when a positional has empty / whitespace-only / missing `help`, so `opencli <site> <cmd> --help` always shows callers what each parameter is for. Pre-existing offenders (`twitter followers/following/list-add/list-remove/list-tweets/search/thread`, `reddit search/subreddit/user/user-comments/user-posts`, `douyin stats/update`, `bilibili subtitle`, `jike search`) now have explicit help text — most notably `twitter followers [user]` and `following [user]` now document that omitting the user fetches the currently logged-in account.
* **help** — adapter help is now agent-friendly: per-command listings drop the `[options]` noise from globally-shared options (`--format`, `--trace`, `-v`, `-h`, etc.) and only mention them at the site level, so `opencli twitter` etc. read like a flat command index. ([#1401](https://github.com/jackwener/opencli/issues/1401))
* **twitter** — write-action symmetry P0: add `unlike`, `retweet`, `unretweet`, and `quote` to round out the read/write coverage. ([#1400](https://github.com/jackwener/opencli/issues/1400))
### Bug Fixes
* **browser daemon** — `npm install -g @jackwener/opencli@latest` now correctly auto-restarts a stale ready-state daemon so users pick up the new version without a manual `opencli daemon restart`. ([#1399](https://github.com/jackwener/opencli/issues/1399))
Extension bumped to 1.0.6 (screenshot `--width` / `--height` / `--full-page` flags, automation tab group color marker, automation container reuse fix).
### ⚠ BREAKING CHANGES
* **linux-do** — remove deprecated compatibility shims `linux-do hot`, `linux-do category`, `linux-do latest`. Use `linux-do feed --view top --period <period>`, `linux-do feed --category <id-or-name>`, and `linux-do feed --view latest` instead.
* **grok ask** — drop the `--web` flag and the legacy `<textarea>` composer path. The default flow is now the only path and uses the current ProseMirror+TipTap composer (the path that used to require `--web true`). Existing scripts passing `--web` will get an "unknown option" error from commander; remove the flag.
* **env** — rename `OPENCLI_BROWSER_TIMEOUT` to `OPENCLI_BROWSER_IDLE_TIMEOUT`. The variable controls workspace lease idle release time, not per-command runtime; the new name reflects that. Old name was undocumented and removed without a fallback.
* **registry** — remove the unused `Strategy.HEADER`; adapter authors should use `Strategy.COOKIE` and set headers explicitly inside browser-side fetches.
### Features
* **observation** — add trace artifact primitives, `browser console`, `browser network --since/--follow/--failed`, and adapter `--trace=retain-on-failure` for failure-retained browser evidence.
* **autofix** — retire `OPENCLI_DIAGNOSTIC`; adapter repair now uses `--trace retain-on-failure`, trace `summary.md`, and error-envelope trace metadata.
* **browser** — `bind` attaches `bound:*` workspaces to user-owned Chrome tabs without taking over window lifecycle; `sessions` reports `idleMsRemaining: null` for bound workspaces because they do not schedule idle close timers. ([#1169](https://github.com/jackwener/opencli/issues/1169), [#929](https://github.com/jackwener/opencli/issues/929))
* **browser lifecycle** — owned browser workspaces now lease tabs inside a shared dedicated automation container instead of owning one Chrome window per workspace; lease state is persisted for MV3 service-worker reconciliation and idle cleanup is backed by alarms.
* **browser session** — adapter commands can opt into site-level tab reuse with `browserSession.reuse = 'site'`; Grok and other browser-backed LLM adapters now keep a shared site tab by default, and users can override with `--reuse <none|site>`.
* **grok** — add browser-web baseline commands: `read`, `history`, `detail`, `new`, `send`, and `status` (existing `ask` and `image` unchanged).
* **yuanbao** — add browser-web baseline commands: `send`, `status`, `read`, `history`, and `detail` (joining the existing `ask` and `new`).
* **qwen** — add `detail` command for opening a specific historical conversation by id.
* **web read** — make page extraction render-aware: same-origin iframe content is merged into the Markdown source, `--wait-for` can wait inside main/iframe documents, `--wait-until networkidle` waits for captured requests to settle, and `--diagnose` reports frames, empty containers, and API-like XHRs for shell/AJAX pages.
### Bug Fixes
* **pipeline / capabilityRouting** — the `fill` pipeline step (introduced in [#1222](https://github.com/jackwener/opencli/issues/1222)) now correctly triggers a browser session and gets transient retry coverage; previously a pipeline using only `fill` could crash on a missing page object. ([#1393](https://github.com/jackwener/opencli/issues/1393))
* **xiaohongshu publish** — improve image publishing reliability via creator-center URL routing, tab priority handling, and DataTransfer fallback.
* **youtube** — use watch-page HTML for transcript captions to recover when the public transcript API is unavailable.
* **desktop adapters** — restore 11 desktop adapter commands that were lost from the manifest due to a factory-pattern regression.
### Internal
* **cleanup** — remove dead `src/analysis.ts` (179 lines, 0 importers), retire `OPENCLI_DIAGNOSTIC` test residue, derive validator step allowlist from the live pipeline registry to prevent future drift.
OpenCLI gives you one surface for three different kinds of automation:
- **Use built-in adapters** for sites like Bilibili, Zhihu, Xiaohongshu, Reddit, HackerNews, Twitter/X, and [many more](#built-in-commands).
- **Let AI Agents operate any website** — install the `opencli-adapter-author` skill in your AI agent (Claude Code, Cursor, etc.), and it can navigate, click, type, extract, and inspect any page through your logged-in browser via `opencli browser` primitives.
- **Let AI Agents operate any website** — install the `opencli-adapter-author` skill in your AI agent (Claude Code, Cursor, etc.), and it can navigate, click, type/fill, extract, and inspect any page through your logged-in browser via `opencli browser` primitives.
- **Write new adapters** end-to-end with `opencli browser` + the `opencli-adapter-author` skill, which guides from first recon through field decoding, code, and `opencli browser verify`.
It also works as a **CLI hub** for local tools such as `gh`, `docker`, and other binaries you register yourself, plus **desktop app adapters** for Electron apps like Cursor, Codex, Antigravity, ChatGPT, and Notion.
It also works as a **CLI hub** for local tools such as `gh`, `docker`,`tg-cli`, `discord-cli`, `wx-cli`, and other binaries you register yourself, plus **desktop app adapters** for Electron apps like Cursor, Codex, Antigravity, ChatGPT, and Notion.
## Highlights
- **Desktop App Control** — Drive Electron apps (Cursor, Codex, ChatGPT, Notion, etc.) directly from the terminal via CDP.
- **Browser Automation for AI Agents** — Install the `opencli-adapter-author` skill, and your AI agent can operate any website: navigate, click, type, extract, screenshot — all through your logged-in Chrome session.
- **Browser Automation for AI Agents** — Install the `opencli-adapter-author` skill, and your AI agent can operate any website: navigate, click, type/fill, extract, screenshot — all through your logged-in Chrome session.
- **Multi-profile Browser Bridge** — Install the extension in each Chrome profile you want to use, then route commands with `--profile`, `OPENCLI_PROFILE`, or `opencli profile use`.
- **Website → CLI** — Turn any website into a deterministic CLI: 90+ pre-built adapters, or write your own with the `opencli-adapter-author` skill + `opencli browser verify`.
- **Website → CLI** — Turn any website into a deterministic CLI: 100+ site surfaces are already registered, or write your own with the `opencli-adapter-author` skill + `opencli browser verify`.
- **Account-safe** — Reuses Chrome/Chromium logged-in state; your credentials never leave the browser.
- **AI Agent ready** — One skill takes you from site recon through API discovery, field decoding, adapter writing, and verification.
- **CLI Hub** — Discover, auto-install, and passthrough commands to any external CLI (gh, docker, obsidian, etc).
- **CLI Hub** — Discover, auto-install, and passthrough commands to any external CLI (gh, docker, obsidian, tg-cli, discord-cli, wx-cli, etc).
- **Zero LLM cost** — No tokens consumed at runtime. Run 10,000 times and pay nothing.
- **Deterministic** — Same command, same output schema, every time. Pipeable, scriptable, CI-friendly.
@@ -89,6 +89,18 @@ Use OpenCLI directly when you want a reliable command instead of a live browser
-`opencli external register mycli` exposes a local CLI through the same discovery surface.
If you want to add your own commands, start with the [Extending OpenCLI guide](./docs/guide/extending-opencli.md). README keeps this short; the guide covers the directory layout, source-control model, and install commands.
| Need | Recommended path |
|------|------------------|
| Keep personal website commands in your own Git repo | `opencli plugin create` + `opencli plugin install file://...` |
| Quickly draft a private local adapter | `opencli browser init <site>/<command>` in `~/.opencli/clis/` |
| Modify an official adapter locally | `opencli adapter eject <site>` + `opencli adapter reset <site>` |
| Wrap an existing local binary | `opencli external register <name>` |
## For AI Agents
OpenCLI's browser commands are designed to be used by AI Agents — not run manually. Install skills into your AI agent (Claude Code, Cursor, etc.), and the agent operates websites on your behalf using your logged-in Chrome session.
@@ -138,9 +150,9 @@ The agent handles all the `opencli browser` commands internally — you just des
- [`skills/opencli-usage/SKILL.md`](./skills/opencli-usage/SKILL.md) — command and site reference
`opencli browser open <url>` and `opencli browser tab new [url]` both return a target ID. Use `opencli browser tab list` to inspect the target IDs of tabs that already exist, then pass `--tab <targetId>` to route a command to a specific tab. `tab new` creates a new tab without changing the default browser target; only `tab select <targetId>` promotes that tab to the default target for later untargeted `opencli browser ...` commands.
`opencli browser` commands require `--session <name>`. `opencli browser --session work open <url>` and `opencli browser --session work tab new [url]` both return a target ID. Use `opencli browser --session work tab list` to inspect target IDs, then pass `--tab <targetId>` to route a command to a specific tab. `tab new` creates a new tab without changing the default browser target; only `tab select <targetId>` promotes that tab to the default target for later untargeted commands in the same session.
## Core Concepts
@@ -148,7 +160,7 @@ Available browser commands include `open`, `state`, `click`, `type`, `select`, `
`opencli browser` commands are the low-level primitives that AI Agents use to operate websites. You don't run these manually — instead, install the `opencli-adapter-author` skill into your AI agent, describe what you want in natural language, and the agent handles the browser operations.
For example, tell your agent: *"Help me check my Xiaohongshu notifications"* — the agent will use `opencli browser open`, `state`, `click`, etc. under the hood.
For example, tell your agent: *"Help me check my Xiaohongshu notifications"* — the agent will use `opencli browser --session <name> open`, `state`, `click`, etc. under the hood.
### Built-in adapters: stable commands
@@ -160,16 +172,16 @@ When the site you need is not yet covered, use the `opencli-adapter-author` skil
1. Recon the site and classify its pattern (SPA / SSR / JSONP / Token / Streaming).
2. Discover the right endpoint — network inspection, initial state, bundle search, token trace, or interceptor fallback.
6. Persist site knowledge to `~/.opencli/sites/<site>/` so the next adapter for the same site is faster.
### CLI Hub and desktop adapters
OpenCLI is not only for websites. It can also:
- expose local binaries like `gh`, `docker`, `obsidian`, or custom tools through `opencli <tool> ...`
- expose local binaries like `gh`, `docker`, `obsidian`,`tg-cli`, `discord-cli`, `wx-cli`, or custom tools through `opencli <tool> ...`
- control Electron desktop apps through dedicated adapters and CDP-backed integrations
## Prerequisites
@@ -186,17 +198,16 @@ OpenCLI is not only for websites. It can also:
|----------|---------|-------------|
| `OPENCLI_DAEMON_PORT` | `19825` | HTTP port for the daemon-extension bridge |
| `OPENCLI_PROFILE` | — | Browser Bridge profile alias/contextId to use when multiple Chrome profiles are connected |
| `OPENCLI_WINDOW_FOCUSED` | `false` | Set to `1` to open the automation container in the foreground (useful for debugging). The `--focus` flag sets this. |
| `OPENCLI_LIVE` | `false` | Set to `1` to keep the automation lease open after an adapter command finishes (useful for inspection). The `--live` flag sets this. |
| `OPENCLI_WINDOW` | command default | Set to `foreground` or `background` to override Browser Bridge window placement. Browser-backed commands also accept `--window <foreground\|background>`. |
| `OPENCLI_KEEP_TAB` | command default | Set to `true` or `false` to keep or release the browser tab lease after a browser-backed adapter command. Browser-backed adapter commands also accept `--keep-tab <true\|false>`. |
| `OPENCLI_BROWSER_CONNECT_TIMEOUT` | `30` | Seconds to wait for browser connection |
| `OPENCLI_BROWSER_COMMAND_TIMEOUT` | `60` | Seconds to wait for a single browser command |
| `OPENCLI_CDP_ENDPOINT` | — | Chrome DevTools Protocol endpoint for remote browser or Electron apps |
| `OPENCLI_VERBOSE` | `false` | Enable verbose logging (`-v` flag also works) |
| `OPENCLI_DIAGNOSTIC` | `false` | Set to `1` to capture structured diagnostic context on failures |
| `DEBUG_SNAPSHOT` | — | Set to `1` for DOM snapshot debug output |
`--focus` works for both `opencli browser *` and browser-backed adapter commands. `--live` is mainly for adapter commands: browser subcommands already keep the automation lease open until you run `opencli browser close` or the idle timeout expires.
`opencli browser *` requires an explicit `--session <name>`, uses a foreground browser window by default, and keeps that session's tab lease until `browser --session <name> close` or idle cleanup. Browser-backed adapters use a background adapter window and release one-shot tab leases by default. Interactive adapters can declare `siteSession: 'persistent'` to keep a stable site tab for continuity; pass `--site-session ephemeral`for a one-shot tab.
## Update
@@ -239,6 +250,7 @@ To load the source Browser Bridge extension:
90+ adapters in total — **[→ see all supported sites & commands](./docs/adapters/index.md)**
100+ site surfaces in total — **[→ see all supported sites & commands](./docs/adapters/index.md)**
`*``opencli xiaoyuzhou podcast`, `podcast-episodes`, `episode`, `download`, and `transcript` require local Xiaoyuzhou credentials in `~/.opencli/xiaoyuzhou.json`.
@@ -281,6 +294,9 @@ OpenCLI acts as a universal hub for your existing command-line tools — unified
- Verify with `opencli browser --session recon verify <site>/<name>` before shipping.
Adapters you write outside the repo live at `~/.opencli/clis/<site>/<name>.js`. Site knowledge (endpoints, field maps, fixtures) accumulates in `~/.opencli/sites/<site>/` so the next adapter for the same site starts from context instead of zero.
For long-lived personal commands that should live in your own Git repo, use a local plugin instead; see [Extending OpenCLI](./docs/guide/extending-opencli.md). Quick private adapters can still live at `~/.opencli/clis/<site>/<name>.js`. Site knowledge (endpoints, field maps, fixtures) accumulates in `~/.opencli/sites/<site>/` so the next adapter for the same site starts from context instead of zero.
A 30-second morning routine that surfaces what changed overnight in reinforcement-learning and large-model research, without opening a browser.
## What I wanted
Before reading anything, decide where to spend my 20 minutes of paper time:
- which `cs.LG` and `cs.AI` papers landed in the last 24 hours
- which OpenReview submissions at recent venues (NeurIPS 2025 right now, ICLR 2024 / NeurIPS 2024 as historical reference) carry titles and primary areas relevant to my work
- which papers the Hugging Face Daily Papers community is talking about today
Skim signals, then drill in. The point is to filter, not to read everything.
## Commands
```bash
# 1. arxiv recent in the two relevant categories (newest 30 each)
That is the entire collection step. The four files together are the whole signal surface for one morning.
## What I do with the output
Pipe the four JSON files into a one-shot LLM digest with a fixed prompt:
```
Here are four JSON arrays of papers from the last 24 hours.
Group them into:
1. Direct hits on RLHF / preference optimization / reasoning RL.
2. Adjacent (offline RL, world models, agent benchmarks).
3. Notable infra (training, evaluation, data).
For each, give me title + arxiv id + one-sentence why-it-matters.
Skip everything that is review / survey / position paper.
```
The LLM compresses ~120 entries into a 10-line shortlist in seconds. I then open whichever 2 to 3 papers actually clear the bar.
## Why CLI beats the browser version
- Four pages of clicking and scrolling collapses into four `opencli` calls.
- The output is structured JSON, so the digest prompt can reason about it deterministically. No copy-paste, no "I missed paper 14".
- Works inside any agent loop. A scheduled task can run the four commands, push them to an LLM, and message the digest somewhere. No browser kept open.
- Zero token cost on the OpenCLI side. The only paid step is the digest call at the end.
The arxiv adapter's `recent <category>` (added in #1289) is the lever here. Without it I would have to fall back to the arxiv listings page, which means scraping HTML in agent code instead of consuming a structured listing.
# Find a paper's implementation and follow-up work
Given a single paper title or arxiv id, walk three sources in one chain to find the canonical reference, follow-up citations, and any community-fine-tuned models or Spaces that already build on it.
## What I wanted
I read a paper abstract, decide it is interesting, and want to answer three questions before deciding to actually re-read the paper or reproduce it:
1. Has anyone already implemented or fine-tuned on top of it (Hugging Face)?
2. Who has cited or extended it (dblp / OpenReview)?
3. What is the canonical bibliographic record (dblp key for citation, full arxiv metadata for reading)?
Doing this in a browser means three tabs and two minutes of context-switching. The point is to compress that into one shell pipeline.
## Commands
Worked example: "Direct Preference Optimization" (DPO).
```bash
# 1. Canonical arxiv record (full abstract, authors, pdf url, categories).
# Note: arxiv free-text search ranks by recency, so the original DPO
# paper does not always come back first. When the canonical id is
Three of the four are public-strategy adapters, no browser session needed. The OpenReview call also lands without auth for public venues.
## What I do with the output
For DPO the chain produces:
- arxiv record: paper id `2305.18290`, full abstract, pdf link.
- dblp record: canonical key `conf/nips/RafailovSMMEF23`, NeurIPS 2023, co-author list (useful to find related work by same lab).
- HF Daily Papers (last 30 days): every paper whose title mentions DPO or preference. Each one is a candidate "follow-up work I should know about".
- OpenReview: the original submission's review thread, if posted (lets me see what reviewers actually pushed back on, which is more useful than the published abstract).
I dump all four JSON outputs into a single LLM call with the prompt: *"Build a one-paragraph 'state of the field' summary for this paper as of today. Cite each follow-up by arxiv id."* That gives me a research-debt brief in 30 seconds.
## Why this is worth a CLI chain
- Each adapter alone is just "search a website". The value is the chain. Four `opencli` calls feed into one LLM call. No browser, no copy-paste.
- Output is identifier-rich (arxiv id, dblp key, venue id, HF paper id). I can re-feed any of those into the next call, e.g. once I find a follow-up arxiv id from HF Daily Papers I run `opencli arxiv paper <new-id>` immediately.
- Survives use inside an agent loop. Same chain runs unattended for a batch of 20 papers from a reading list.
- Zero token cost for the discovery half. Only the final summary step pays for inference.
Without `opencli dblp search` (added in #1299) and `opencli openreview search` (added in #1294), this whole pipeline used to require either web scraping in agent code or paying for a research-paper API. Both adapters being public-strategy means they slot in cleanly.
# Track a conference's accepted papers and reviews from the terminal
Once an OpenReview venue opens its decisions (or releases reviews publicly during the discussion phase), I want a one-shot way to pull the full venue listing and dive into individual review threads, without clicking through 200+ submission pages.
## What I wanted
For each major venue I follow (ICLR, NeurIPS, ICML), the same three things every time decisions are visible:
1. The full list of accepted papers at the venue, with titles and forum ids.
2. For any paper I flagged interesting from the list: the full review thread, including reviewer scores, rebuttals, and the AC's decision rationale.
3. A way to pipe both into LLM-driven shortlisting ("which of these 100 oral papers actually intersect with my research direction").
The OpenReview UI is fine for one paper at a time, but unusable for batch reasoning across the whole acceptance list.
## Commands
Worked example: ICLR 2024 oral track, then drill into one paper's reviews using a real forum id.
```bash
# 1. Full list of papers at a venue (natural-language venue text;
# if the venue is not yet open OpenReview returns EMPTY_RESULT
`venue` returns each entry with a forum id you can hand straight back into `reviews` and `paper`. No id lookup gymnastics. `reviews` returns the full thread as a JSON array: a `PAPER` row with the abstract, then one `REVIEW` row per reviewer (with `rating`, `confidence`, summary, weaknesses, questions), followed by author rebuttals and the AC's decision rationale.
## What I do with the output
Two distinct workflows depending on the phase of the venue:
### Phase A: filtering the acceptance list
After `venue` returns 200 entries, dump the JSON into an LLM with the prompt:
```
Here is the full acceptance list at <venue>. Filter to papers that intersect
with my research interests:
- reinforcement learning from preference / reward feedback
- reasoning training (process reward, RLVR, RLHF variants)
- long-horizon agent benchmarks
For each match: title + forum_id + one-sentence why-it-matters.
```
This collapses 200 papers to a 10-paper shortlist in seconds. The forum ids are the keys I will use in Phase B.
### Phase B: depth-reading the shortlist
For each shortlisted forum id, run `opencli openreview reviews <forum-id>` and feed the JSON to an LLM with the prompt:
```
Summarize the review thread:
- reviewer scores
- the strongest critique
- whether the rebuttal addressed it
- final decision and AC rationale
```
This is faster than reading three reviews + rebuttal + meta-review per paper. For 10 papers this turns 60 minutes of OpenReview clicking into 10 minutes of summary reading, then I open the actual reviews only for papers where the summary flagged something worth knowing.
## Why this beats opening OpenReview
- One `venue` call replaces scrolling a paginated UI for 200+ papers.
-`reviews` returns the entire thread as JSON, so an LLM can reason over the whole review-rebuttal-decision arc at once. The web view forces you to scroll three reviews + N rebuttals + meta separately.
- Forum ids returned from `venue` are stable and reusable across calls. Easy to keep a personal reading list as `forum-ids.txt` and run `for id in $(cat forum-ids.txt); do opencli openreview reviews $id; done`.
- The whole loop is public-strategy. No login required for venues with public reviewing.
`opencli openreview` (added in #1294) is the lever. Before this adapter existed, the same workflow needed either OpenReview's Python client or HTML scraping inside agent code. Both have higher friction than `opencli openreview reviews <forum-id>` returning structured JSON in one shot.
<summary>The dominant sequence transduction models are based on complex recurrent or convolutional neural networks. We propose a new simple network architecture, the Transformer, based solely on attention.</summary>
description:'24h ticker statistics for top trading pairs by volume',
domain:'data-api.binance.vision',
strategy:Strategy.PUBLIC,
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
jakevin