- main.py + config.py: CORS origins read from Settings.cors_origins (comma-separated).
Empty falls back to dev origin localhost:5173; never wildcard. Methods locked
to GET/PATCH/POST; allow_credentials=False.
- scraper/media.py: download_url now passes follow_redirects=False to httpx.
OSS URLs return 200 directly; a 3xx is treated as a download failure (caught
by existing except clauses) instead of opening an SSRF vector to internal hosts.
- routers/voice_clips.py: PATCH /voice-clips/{id}/match accepts an optional
If-Match header carrying the client's last-seen updated_at. On mismatch the
row was modified by another request and the server returns 409. Without the
header the behaviour is unchanged (backward compatible). Response now echoes
the post-update updated_at so the client can fold it into the next request.
- routers/organization.py: /api/accounts now UNION ALLs the wechat_account and
whatsapp_account projections at the SQL level and pushes COUNT + pagination
(offset/limit) down to the DB. Stable ordering on (channel, account_id).
Removes the in-memory materialization that broke at scale.
Tests: 395 passed. Added optimistic-lock coverage (If-Match accepted / stale 409 /
absent header backward-compat). Updated existing PATCH tests for the new
updated_at response field.
Co-Authored-By: Claude <noreply@anthropic.com>
address review Minors: drop redundant filename replace, assert 暂无消息,
tighten downloadBlob return type + try/finally, fix plan group time-picker wording.
Co-Authored-By: Claude <noreply@anthropic.com>
M6 — the scheduling/deployment layer. Makes the system runnable as a
scheduled service + deployable via Docker.
Part A — SyncScheduler (app/sync/scheduler.py):
- APScheduler (AsyncIOScheduler) wraps the orchestrator's sync_* functions.
- Daily full sync at FULL_SYNC_HOUR (default 02:00 local) + incremental
message sync every MESSAGE_INTERVAL_HOURS (default 4).
- ENTITY_FUNCS registry dispatches run_once("<entity>") to the right runner
(departments/accounts/friends/groups/voice_calls/phone_calls/sms/messages/
all_wechat/all).
- Conversations for the message run are discovered from the local customer
mirror (no extra upstream call).
- Failure policy: every run is wrapped — a blown entity is logged + swallowed
so the scheduler keeps its next appointment.
- Per-run JSON-line log (level=info/warning/error, totals, failed entities).
Part B — CLI entry point (app/sync/cli.py):
- `python -m app.sync.cli sync [--entity <name>] [--quiet]` — one-shot sync.
- `python -m app.sync.cli scheduler [--full-sync-hour H] [--message-interval-hours N]`
— long-running scheduler (blocks until SIGINT/SIGTERM).
- Uses _push_settings_into_environ + build_client_from_env (session-cookie
mode preferred; pcLogin deferred).
- Base.metadata.create_all on every command (fresh-checkout friendly).
- Exit codes: 0 success, 1 generic failure, 2 bad CLI / auth misconfig.
Part C — Docker stack:
- yunke-server/Dockerfile: python:3.12-slim, installs deps + app, single
image serves both API (uvicorn) + scheduler (CLI).
- yunkes/Dockerfile: multi-stage node:20 build → nginx serves dist/ +
reverse-proxies /api → backend:8000.
- yunkes/nginx.conf: /api proxy with proxy_buffering off (streaming media),
SPA fallback, asset caching.
- docker-compose.yml: mysql (8.0, healthcheck, volume) + redis (7-alpine) +
backend + scheduler + frontend (port 5173 → 80). Env-driven creds + cadence.
Part D — README.md at repo root: dev mode + docker mode quick-start, config
table, architecture overview, current limitations (pcLogin deferred, WA voice/
media deferred, conservative pacing, conversation discovery is local-only).
Verification:
- pytest: 323 passed (was 294; +29 new for scheduler + CLI).
- npm run build: clean.
- python -m app.sync.cli --help: prints usage.
- docker compose config: VALID.
Co-Authored-By: Claude <noreply@anthropic.com>
Live validation: businessFriendList uses 'page' (not pageNum) and REQUIRES
departmentId (root) + type or returns empty. Same class of bug as wxUser/list.
Co-Authored-By: Claude <noreply@anthropic.com>
Replicates the Customer all-fields pattern across the remaining three
entities. Backend now exposes `raw_json` on MessageOut / StaffOut /
AccountOut so the frontend can render every 云客 source field without a
second endpoint. Each entity gains a Chinese-label field utility + an
el-drawer surfacing all fields (structured-first, raw_json fallback,
unknown keys under 其他).
Backend:
- schemas.py: add `raw_json: dict | None` to MessageOut / StaffOut /
AccountOut. MessageOut/StaffOut pick the column up via from_attributes;
AccountOut is built explicitly in the router, which now threads
`raw_json` through from both WechatAccount and WhatsappAccount rows.
- tests: new raw_json serialization tests for messages, staff, accounts
(assert populated dict round-trips verbatim + NULL serializes as null).
Frontend (yunkes/):
- New utils: messageFields.ts, staffFields.ts, accountFields.ts — each
with a 中文 label map, ordered groups, structured-column preference,
and an 其他 fallback for unknown raw_json keys.
- Message: clicking a chat bubble opens an all-fields drawer (referMsg
renders as pretty-printed JSON). Image preview + file download links
stopPropagation so their own interactions still work.
- Staff: row click → drawer; columns relabeled (员工ID, 销售手机).
- Accounts: new AccountList.vue at /accounts with a 账号 nav item,
unified WeChat/WhatsApp table (渠道/账号ID/昵称/微信号/归属员工/好友数/
群数/收发/在职/最后更新), channel filter, pagination, row-click drawer.
- api: add Account type + getAccounts(); add raw_json to Message/Staff.
Co-Authored-By: Claude <noreply@anthropic.com>
Backend:
- CustomerOut now exposes raw_json (verbatim upstream friend record) so the
frontend can render every 云客 field without per-field structured columns.
- The verbatim blob round-trips via from_attributes; no model change needed.
Frontend:
- CustomerList table mirrors the 云客 客户列表 column set (头像/微信昵称/微信号/
备注/真实姓名/手机/性别/地区/描述/销售员工/部门/来源/添加方式/添加时间/最近
聊天/消息数/渠道). A readField helper reads the structured column first and
falls back to raw_json.<camelCase>.
- Row click opens an all-fields drawer (right side) rendering EVERY field of
the record with Chinese labels, grouped (基本信息/联系方式/归属员工与部门/时间
与来源/消息统计/好友状态/其他). Timestamps format to local datetime, flags
render 是/否, image URLs render as el-image thumbnails, empty values show —.
- New src/utils/customerFields.ts holds the camelCase→中文 label map, the
structured-column alias map, group definitions, and value-type formatters;
unknown raw_json keys auto-flow into an 其他 section so new upstream fields
surface without code changes.
- Drawer 查看聊天记录 action navigates to /chat/:channel/:ownerId/:talkerId
(owner_account_id + friend_id + channel), preserving the existing path.
Co-Authored-By: Claude <noreply@anthropic.com>
Backend (scrape_wechat_accounts):
- Send wxUser/list as form-urlencoded (not JSON) with the REQUIRED filters
isOnJob / departmentId / startdate / enddate + empty userId/salesname/
weChatNickName, matching the live request (empty fails with code -90000).
- scrape_wechat_accounts now takes department_id (the company-root dept id)
plus optional start/end defaulting to a recent 7-day UTC window.
- paginate_by_page gains a form=True mode (data= + form content-type).
- Orchestrator sync_wechat_accounts resolves the root dept from the
Department table (or accepts one) and skips with a warning when unknown.
- seed_accounts.py resolves the root via scrape_departments.
Frontend (yunkes):
- Add StaffList.vue (/staff) with is_on_job filter + pagination.
- Add /staff route and a 员工 nav item between 客户 and 群.
Co-Authored-By: Claude <noreply@anthropic.com>
Live validation found two real bugs:
- YunkeClient sent 'source: yunkepc' on every request; msg/chats binds
source as Integer -> 500. Removed (cookie auth suffices, per Phase 0).
- WeChat message talker_id used the masked 'talker' field; the timeline
joins on the full peer id. Now: 1:1 -> oriTalker (full), group -> talker
(member). ori_talker keeps the raw oriTalker (peer/chatroom).
Co-Authored-By: Claude <noreply@anthropic.com>
Backend (Part A):
- MessageOut: add file_th_url, file_hd_url, duration, transcription_text, md5
(room_id + sender_head_url already existed) so the chat timeline can render
image thumbnails, voice duration/transcription, and avatars from a single
response.
- PhoneCallOut: add record_file so the Calls view can offer recording playback.
- SmsOut: add content (full body) so the SMS detail dialog shows the whole text.
- Tests: red->green for each new field (media fields default None + populated,
record_file round-trip, content round-trip). Suite 267 -> 271.
Frontend (Part B):
- api/index.ts: wrappers getGroups/getGroupMembers/getCalls/getSms/getSyncStatus
+ Group/GroupMember/PhoneCall/Sms/SyncState/SyncRun/SyncStatus types in
types.ts (mirror the *Out pydantic schemas).
- views/GroupList.vue (/groups): table + channel filter + pagination; row click
opens the group timeline.
- views/CallList.vue (/calls): table + call-type filter (phone_call/voice_call/
all); "播放录音" action opens an el-dialog with <audio controls>.
- views/SmsList.vue (/sms): table; row click opens a detail dialog with the full
content.
- views/SyncStatus.vue (/sync): two tables (states + recent_runs) with a manual
refresh; run status color-coded.
- ChatTimeline.vue: reused for both 1:1 DM (talkerId) and group (roomId) by
branching the query; group title + per-member sender already handled by
MessageBubble.
- router: add /groups /calls /sms /sync + /group/:channel/:ownerId/:roomId.
- layouts/MainLayout.vue: replace the single 客户列表 button with an el-menu
(客户 / 群 / 通话 / 短信 / 同步); org tree kept on the left.
Co-Authored-By: Claude <noreply@anthropic.com>
Mirrors the WeChat scraper patterns (M3) using the shared paginator +
parser + repo building blocks. Three new async generators yield
snake_case dicts + verbatim raw_json, ready for the channel-generic
upsert_customer / upsert_messages repos.
- app/scraper/whatsapp.py (new): scrape_whatsapp_accounts /
scrape_whatsapp_friends (page-number pagination over the
data.overviews / data.totalCount envelope) and scrape_whatsapp_messages
(backward-cursor pagination, JSON body {uid,friendUid,start,end} —
NOT form-encoded like WeChat msg/chats). Drops type==-2 dividers;
maps mediaName -> file_url as a deferred-download key.
- app/scraper/parsers.py: add normalize_whatsapp_type (0->text,
1->image, -2->divider, else unknown; never raises) and a shared
coerce_int_or_none scraper helper (mirrors repos._coerce_int so the
scraper emits type_raw ints without a DB dependency).
- app/storage/repos.py: add upsert_whatsapp_account (PK uid) which
also upserts the embedded Staff(user_id, name) via the existing
upsert_staff_from_account helper; WhatsappAccount imported.
Tests (respx mocks, no network): 3 cassette fixtures + 3 scraper test
modules (accounts/friends/messages), parsers + repos extended with WA
cases. Divider drop, backward-cursor advance (end == E1-1), JSON body
asserts, embedded-staff upsert, and channel-generic Customer/Message
upserts all covered.
Suite: 182 -> 211 passing. ruff clean.
Co-Authored-By: Claude <noreply@anthropic.com>
Cleanup after successful live validation (10 real customers scraped+stored):
- app/config.py: add yk_userid / yk_company / yk_cookie Settings fields so
existing-session creds can be supplied via .env (bypasses programmatic
login for validation with a live browser session).
- app/storage/repos.py: coerce list values to ';'-joined strings in
_build_column_values. Fixes a real bug: the live friends-list API returns
Customer.deptRoute as a LIST, but dept_route is a scalar String(255)
column. raw_json still preserves the original list (no data loss).
- scripts/smoke_wechat.py:
* Revert build_client_from_env to read os.environ ONLY (no get_settings
fallbacks), so the monkeypatch.setenv unit tests stay hermetic and the
real .env stops leaking in. 3 previously-red tests green again.
* Add _push_settings_into_environ() (Settings -> os.environ, real env
wins, only fills gaps) and call it as the first line of main(); so the
smoke run still honors .env without the pure builder depending on
Settings. .env -> Settings -> os.environ (once) -> build_client_from_env.
* Keep the await _print_summary fix from validation.
- tests/test_repos.py: regression test for list dept_route coercion
(stored joined 'A;B', raw_json keeps the list).
Co-Authored-By: Claude <noreply@anthropic.com>
Centralize scraping stealth in YunkeClient.request() so automated traffic
resembles the normal CRM web app and is paced conservatively to avoid
triggering 云客 risk control (pulls M6 rate-limit work forward):
- app/config.py: add 6 tunable settings with conservative defaults
(request_min_interval=1.0, request_jitter=0.8, max_requests_per_run=0,
risk_backoff_{base=2.0, max=60.0, attempts=4}).
- app/scraper/base.py: RateLimiter gains optional jitter + a new wait()
method (fixed min+jitter per-call sleep); gap-based acquire() is unchanged
for backward compat (default jitter=0.0).
- app/session/yunke_client.py: every request now carries browser-identical
headers (UA/sec-ch-ua/accept/accept-language/referer) merged with the auth
headers (per-request overrides win); awaits the shared jittered limiter
before sending; enforces a per-run request cap (RequestCapReached); and
backs off exponentially on risk signals (HTTP 429 or 200 body carrying
频繁/风控/限制/稍后), distinct from the existing 401 relogin-once path.
- tests/conftest.py: zero the throttle env in the test suite so the
conservative prod defaults don't slow unit tests.
- tests/test_yunke_client_stealth.py: 9 new tests (browser headers,
override, limiter spy, 429 backoff, 200-marker backoff, persistent-429
raise, volume cap, unlimited, 401-relogin preserved). TDD red->green.
Suite: 171 -> 181 passing. ruff clean.
Co-Authored-By: Claude <noreply@anthropic.com>
Adds a shared-token auth dependency for the viewing API (Task 2.9). Trust
model: when API_TOKEN is empty (the default) auth is disabled — we trust the
network boundary (local/intranet tool). When set, every request must present
the token via `Authorization: Bearer <token>` or `?token=<token>`, else 401.
`/api/health` is exempt so liveness probes work without the token.
- app/config.py: add `api_token: str = ""` to Settings.
- app/api/auth.py: `require_auth` dependency (secrets.compare_digest, reads
get_settings() per-request so token rotation via cache_clear() works).
- app/api/main.py: apply the dependency globally; document the trust model.
- .env.example: document API_TOKEN.
- tests/test_api_auth.py: 8 tests covering disabled/correct/missing/wrong/
malformed-header/query-param/health-exempt paths.
Co-Authored-By: Claude <noreply@anthropic.com>
Task 2.7 — the coordination layer tying scrapers + repos together with
resumable, incremental sync.
app/sync/watermarks.py — SyncState CRUD keyed on (entity_type, channel,
scope_id): get_watermark / get_last_timestamp / get_last_version /
set_watermark (upsert). A partial advance (timestamp-only) does not clobber
the other cursor with NULL.
app/sync/orchestrator.py — the sync brain:
* sync_wechat_messages: time-window incremental. Reads last_timestamp for
(message, wechat, "<wechat_id>:<friend_id>"), queries [last_ts, now] via
scrape_wechat_messages, upserts each, best-effort downloads media, advances
the watermark to max(timestamp). A re-run with no new messages queries an
empty window and fetches nothing — the watermark advanced past.
* sync_{departments,wechat_accounts,wechat_friends,wechat_groups,
wechat_voice_calls}: full re-scrape + idempotent upsert (these list APIs
have no since-param; 'incremental' = re-scrape all, rely on PK idempotency).
Watermark records last_synced_at only.
* sync_all_wechat: coordinator running departments -> accounts -> friends ->
groups -> voice_calls, then per-conversation messages for an explicit
(wechat_id, friend_id) list. Auto-discovery deferred.
Each run logs a SyncRun row (started -> success/failed, counts, error).
Failure policy: a scraper that raises is caught, recorded as failed, and a
failed SyncResult is returned — the orchestrator does not re-raise, so one
broken entity does not abort the batch. Media-download failures are
best-effort. The orchestrator never commits; callers own the transaction.
Tests (TDD red -> green; in-memory SQLite; scrapers/repos monkeypatched):
- watermarks: set/get, upsert (same key twice -> 1 row), partial-advance
preserves the other cursor, defaults when absent.
- orchestrator messages: first run upserts 2 + advances watermark to T2;
second run scrape called with start==T2 and upserts 0 (incremental proof);
media failure does not abort; SyncRun success recorded.
- orchestrator friends: 3 records upserted, SyncRun success, watermark
last_synced_at set.
- failure policy: scraper raises -> SyncRun failed + failed result, no raise
(both friends and message paths).
- sync_all_wechat: 5 entities + 2 message runs in order, all success.
Suite: 149 -> 163 passing. ruff clean.
Co-Authored-By: Claude <noreply@anthropic.com>
M2 Task 2.6 (media downloader) + Task 2.8 (media serve API).
Downloader (app/scraper/media.py):
- download_url(client, url): fetches bytes via the YunkeClient's underlying
httpx client (carries session cookies; respx-mockable; follows redirects).
- download_message_media(client, msg, store, session): URL-driven fetcher that
picks the best http(s) URL (file_url -> file_hd_url -> file_th_url), dedups
by md5 (skips the network when a Media row with local file already exists),
falls back to sha256(bytes) when no md5, stores via ObjectStore.put_by_hash,
and upserts a Media row. Skips type-10 system messages (non-media per
phase0-findings §2) and type-3 opaque-id / text (no http url). Resilient:
download errors are logged and return None (no raise).
- media_type_from_message / guess_mime_type helpers.
Repo (app/storage/repos.py): upsert_media — idempotent Core upsert into Media by
media_hash. Volatile download artifacts (local_path/size_bytes/mime_type) use
COALESCE(incoming, existing) on conflict so a re-scrape that did not re-download
does not clobber previously stored values. Exported from app.storage.
Serve API (app/api/routers/media.py): GET /api/media/{media_hash} streams the
local file via FileResponse (handles HTTP Range / 206). 404 when the row is
absent, has no local file, or the file is missing on disk. Content-type from the
row's mime_type, falling back to ext-derived guess.
Tests (+19, suite 130 -> 149, all green; ruff clean):
- test_media_downloader.py: image/video download + upsert; md5 dedup skips
second network hit; sha256 fallback when no md5; type-3/text/system -> None;
download failure -> None (no raise); upsert_media idempotency +
null-local_path preservation + 3x-same-key-one-row.
- test_api_media.py: 200 + content-type + bytes; 404 unknown hash; 404 missing
file on disk; Range -> 206 partial; ext-derived content-type when mime NULL.
Co-Authored-By: Claude <noreply@anthropic.com>
- scrape_wechat_voice_calls: wraps POST /wechatweb/callRecord/queryWeChatVoiceList
with page-number pagination (reuses paginate_by_page); normalizes records to
the PhoneCall column shape with channel='voice_call'. Field mapping per
phase0-findings §4: id->call_id, duration(str s)->call_duration(int),
startTime(YYYY-MM-DD HH:MM:SS)->start_call_time(ms epoch), callStatus->
call_status, talker->call_number, talkerNickName->plan_customer_name,
ossFileName->record_file (may be null). isSend (0=in/1=out) is inverted to
incoming_call (1=in/0=out) so the column's meaning matches the phone_call
channel.
- upsert_voice_call: idempotent Core upsert into PhoneCall by call_id
(sqlite/mysql dialect-aware). Adds call_duration/start_call_time to the
_INT_FIELDS coercion net (scraper already emits ints; defensive for
hand-built records).
- Group-message test (Task 2.3): locks in that scrape_wechat_messages already
handles chatroom conversations (friendId=@chatroom, roomid set, talker=member
wxid) — yields room_id populated and talker_id = the sending member. No
scraper change required.
Suite: 123 -> 130 passing. ruff clean.
Co-Authored-By: Claude <noreply@anthropic.com>
Add scrape_wechat_groups (POST /wechatweb/chatroom/getChatroomList) and
scrape_wechat_talkers (POST /wechatweb/wxFriends/talkerList), both reusing
the page-number paginator. Groups parse the semicolon-joined memberIds into
a member_ids_list so upsert_chat_group can mirror GroupMember rows
(composite PK channel/room_id/member_id) idempotently alongside the
ChatGroup row. Talkers are a sync-planning input (member_count>0 marks a
group conversation); no dedicated table.
Co-Authored-By: Claude <noreply@anthropic.com>
M2 organization scrapers for yunke-server:
- scrape_departments: POST /usercenter/dept/tree. Single request (no
pagination). Defensively unwraps the envelope (data is node list OR an
object wrapping nodes under children), flattens BOTH nested and flat
shapes, and reconstructs the materialized path by walking the pId chain
(ancestor names joined with /). Maps id/pId/name -> dept_id/parent_id/name.
- scrape_wechat_accounts: POST /wechatweb/wxUser/list. Page-number
pagination via the existing paginate_by_page helper. Envelope differs
from friend_list (no basicResult wrapper): data.data/data.total. Maps
wechatId/nickName/alias/phone/userId/userName/contactsNum/groupNum/
sendNum/receiveNum/isOnJob/dbSize/onlineStatus/lastUpdate/version to
WechatAccount columns; carries embedded staff pair for the repo layer.
Repos:
- upsert_department: idempotent upsert into Department (PK dept_id).
- upsert_staff: primitive idempotent staff upsert (PK user_id, is_on_job
coerced to bool).
- upsert_wechat_account: now ALSO upserts the embedded Staff row via
upsert_staff_from_account (best-effort; skips when account has no user_id).
- upsert_staff_from_account: refactored to delegate to upsert_staff and
forward phone/sale_phone/is_on_job when present.
Tests (TDD red->green): test_scrape_departments (nested + flat + wrapped-data
+ empty), test_scrape_wechat_accounts (2-page pagination + body advance +
empty first page), test_repos extended (dept idempotency/root parent, direct
staff upsert + bool coercion, account-also-upserts-staff, re-upsert updates
both, no-user_id skips staff). Cassettes: dept_tree_fixture (same 4 nodes in
both nested + flat shapes), wxUser_fixture (4 capture-shaped accounts incl.
departed + orphan).
Suite: 97 -> 111 passing. ruff check clean.
Co-Authored-By: Claude <noreply@anthropic.com>
Replace the Phase-0 placeholder password encryption with real RSA
(PKCS#1 v1.5). encrypt_password now builds an RSA public key from the
getSecureKey modulus/exponent (RSAPublicNumbers, e=int(hex,16),
n=int(hex,16)), encrypts the UTF-8 password, and returns hex (documented
assumption -- base64 is the one-line fallback noted in the docstring).
get_secure_key now returns the inner data dict to match the new contract;
pc_login keeps {account, password} with a comment on the username/password
and account/pwd fallbacks if pcLogin rejects.
Extend normalize_wechat_type with the Phase-0 confirmed codes:
2 -> image, 3 -> image_legacy, 4 -> video, 10 -> system (1 -> text and
-2 -> divider unchanged; 6/9/13/15/21/22 stay unknown until confirmed).
Tests:
- encrypt_password round-trip: generate an RSA keypair, build secure_key
from its public numbers, encrypt, then decrypt with the private key and
assert the original password comes back. Pins 512-hex output for a
2048-bit key. No network.
- Parser parametrize table extended for the new codes (and a downstream
message-fixture assertion updated from unknown -> image_legacy).
Suite: 87 -> 97 passing. ruff check clean.
Co-Authored-By: Claude <noreply@anthropic.com>
Wires the existing pieces together into a runnable integration check:
scrape ~10 WeChat friends + ~50 of their messages (last 7 days) from the
real yunke CRM (read-only), upsert into a local SQLite smoke DB, then
print a summary that verifies rows landed.
Two auth modes:
* Existing-session (preferred): YK_USERID/YK_COMPANY[/YK_COOKIE] -> YunkeClient
constructed from a captured session, login bypassed.
* Login mode (Phase-0 blocked): YUNKE_USERNAME/YUNKE_PASSWORD -> attempts
client.login(), which the real CRM will reject until the placeholder
password encryption is replaced. Prints a clear message and exits 3.
build_client_from_env() is split out as a pure, network-free helper so
auth-mode selection is unit-testable. main() is async (asyncio.run in
__main__). HTTPStatusError from a 401/expired session is wrapped into a
helpful message + exit 4 (no traceback). client.aclose() + engine.dispose()
always run in finally.
Defaults to SMOKE_DB_URL=sqlite+aiosqlite:///./smoke.db so no MySQL is
required. Run from yunke-server/ with the venv active:
python -m scripts.smoke_wechat
Tests cover the entrypoint shape and auth-mode selection (no network).
Suite: 77 -> 87 passing.
Co-Authored-By: Claude <noreply@anthropic.com>
浅念