openclaw/crabline

最近记录的来源活动: 2026年9月19日 01:25

SkillsMP 收录数据更新: 2026年11月9日 14:22

已收集 skills: 1

GitHub 星标: 31
GitHub Forks: 11

用 AI 助手安装

复制这段 Prompt，发给你正在使用的 AI 助手。

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/openclaw/crabline.

仓库 README

仓库 README 来源：README.md

查看 README 源文件

展开完整 README

# crabline ![Crabline deterministic messaging-channel QA](assets/crabline-banner.svg) Deterministic local messaging-channel mocks for OpenClaw QA. `crabline` is config-driven, CI-friendly, and deliberately has no `openclaw` dependency. A fixture-level **local mock provider** is a Crabline adapter used directly by fixture commands. A **local provider server** is a separate process that the Crabline channel driver starts so a real OpenClaw channel plugin can exercise its normal provider protocol during deterministic QA. ## What It Provides - local mock providers for `discord`, `feishu`, `googlechat`, `imessage`, `loopback`, `matrix`, `mattermost`, `msteams`, `slack`, `telegram`, `whatsapp`, and `zalo` - a `script` bridge for channels that are still exercised by external commands - per-provider local webhook endpoints for inbound events - local provider servers for Crabline channel-driver tests for Discord, Mattermost, Matrix, Signal, Slack, Telegram, WhatsApp, and Zalo - JSONL recorder files for deterministic wait/watch behavior - nonce-based `send`, `roundtrip`, `agent`, `probe`, `run`, `watch`, and `doctor` commands - text output by default and stable `--json` output for automation Crabline local provider servers are not live-provider coverage and are not the fixture-level local mock provider feature. They let the real OpenClaw channel plugin run against a local provider-shaped API. Release lanes still need the `live` driver and real provider credentials. ## Install ```bash pnpm install pnpm build pnpm verify ``` Dependency updates observe a seven-day release-age policy. Runtime dependency minimums must remain compatible with consumers using the same cooldown. Dependabot applies the same cooldown to version updates, preserving the TypeScript exceptions in `pnpm-workspace.yaml`. Node declarations stay on Node 22 while that runtime remains supported. The project retains pnpm 11 so its single-document lockfile remains readable by GitHub dependency graph and Dependabot tooling. Run locally: ```bash pnpm dev fixtures --config fixtures/examples/crabline.example.yaml pnpm dev roundtrip telegram-dm --config fixtures/examples/crabline.example.yaml ``` The source checkout does not link its own `crabline` bin into `node_modules/.bin`, so use `pnpm dev` there rather than `pnpm exec crabline`. An installed package exposes the `crabline` command used in the examples below. Run `crabline --version` (or `crabline -V`) to print the installed package version. In a source checkout, use `pnpm dev --version`. Node.js consumers can bundle the package with its dependencies, including under pnpm's isolated linker. WhatsApp's crypto dependency uses static imports so it remains available when the bundle runs outside the original installation. ## Quality Gate ```bash pnpm verify ``` That enforces formatting, typecheck, type-aware lint, and Vitest coverage. CI runs the full gate on Node 22 and the runtime tests on Node 24. Workflow validation uses Go 1.26 or newer, with Go 1.27.3 pinned in `tools/go.mod`. The Release workflow supports a manual `dry_run: true` on `main`. Set `tag_name` to the current package version with a `v` prefix (for example, `v0.1.29`). This verifies and packs the selected commit, checks its version and changelog, and uploads the package artifact without publishing to npm or creating a GitHub release. Normal releases require an existing matching stable, SSH-signed annotated tag authorized by `.github/release-allowed-signers`. The workflow verifies the tag and freezes its object and commit before building, then checks both again before npm and GitHub publication. Replacing a tag at the same commit is rejected. ## Config Config file search order: 1. `--config <path>` 2. `./crabline.yaml` 3. `./crabline.yml` 4. `./crabline.json` Top-level shape: ```yaml configVersion: 1 userName: crabline providers: telegram: adapter: telegram telegram: recorder: path: ./.crabline/recorders/telegram.jsonl webhook: host: 127.0.0.1 port: 8790 path: /telegram/webhook fixtures: - id: telegram-dm provider: telegram mode: roundtrip target: id: "100000001" behavior: agent ``` Provider ids are local profile names. Fixtures reference them through `provider`. Built-in adapters infer their platform from `adapter`; `platform` is required only for `adapter: script`. Most built-in provider credentials are optional mock metadata. The WhatsApp webhook requires `appSecret` and `verifyToken` because every verification or delivery request is authenticated. `doctor` checks explicit `env` declarations, script command availability, and config shape without requiring live platform credentials. When configured, Discord `publicKey`, Feishu `encryptKey`, Google Chat endpoint or Pub/Sub identity settings, Microsoft Teams `appId`, Slack `signingSecret`, Telegram `secretToken`, and Zalo `webhookSecret` enforce provider-native webhook authentication. Feishu `verificationToken` remains an additional or loopback-only callback check. Microsoft Teams also requires `appId` for non-loopback or explicitly public webhook endpoints. Authenticated external webhook ingress must advertise an HTTPS `publicUrl`; plain HTTP is limited to loopback-local testing. Feishu text callbacks require valid JSON-encoded `message.content`; malformed content is rejected. ## Built-In Mock Channels All built-in mock providers support: - `probe` - `send` - `roundtrip` - `agent` - `watch` The built-in providers are: - `discord` - `feishu` - `googlechat` - `imessage` - `loopback` - `matrix` - `mattermost` - `msteams` - `slack` - `telegram` - `whatsapp` - `zalo` The `script` adapter can bridge any OpenClaw channel represented by the configured `platform` enum by running local commands for `probe`, `send`, `waitForInbound`, or `watch`. Treat manifests that configure `adapter: script` as executable, trusted code: Crabline runs their declared commands with the configured environment. Load them only from sources you trust and review changes before use. ## Local Provider Servers `serve` starts local provider servers for real OpenClaw channel plugins. This is the preferred Crabline channel-driver path because the plugin still uses its normal provider protocol while the provider endpoint is local and deterministic. It does not use Crabline's fixture-level local mock provider. Commands in this section use the installed-package form. In a source checkout, replace `crabline` with `pnpm dev`. `serve` never accepts credential values in command-line arguments. Set `CRABLINE_ACCESS_TOKEN`, `CRABLINE_ADMIN_TOKEN`, `CRABLINE_BOT_TOKEN`, or `CRABLINE_SIGNING_SECRET`, or pass a bounded JSON object through a file descriptor. Package runners should use stdin: ```bash pnpm dev --json serve slack --credentials-fd 0 < .crabline/serve-credentials.json ``` The installed CLI can also use an inherited descriptor when its launcher preserves descriptors above 2: ```bash crabline --json serve slack --credentials-fd 3 3< .crabline/serve-credentials.json ``` The JSON fields are `accessToken`, `adminToken`, `botToken`, and `signingSecret`. File-descriptor values override environment fallbacks. Do not pass descriptors above 2 through package runners such as `pnpm`; use fd 0 instead. Keep credential files owner-readable or pipe the JSON directly from a secret manager. Ready files contain generated provider and admin credentials. Crabline creates or replaces them with POSIX mode `0600`, but the parent directory still needs to be private. Exclude ready files from version control and CI artifact collection, and delete them after use. On non-POSIX or shared filesystems, verify the effective ACLs before publishing one. Library callers can pass `onEvent` to `startCrablineServer`, an individual provider server, or `startOpenClawCrablineAdapter`. Crabline awaits the callback after appending each API/admin event to `recorderPath`, so callers can react in process while retaining the JSONL artifact as durable evidence. Provider servers initialize recorder ownership before reporting readiness, without creating recorder files or emitting synthetic events. Closing a server stops new recorder admissions and waits for admitted persistence to finish. Shutdown does not wait for event observers, which may themselves call `close()`. Baileys recorder rejections remain handled when shutdown interrupts compressed frame decoding and an event observer subsequently fails. Concurrent and repeated close calls share the same shutdown. All transports are closed even if one fails, and the persistence drain includes asynchronous cleanup after a failed recorder lock acquisition. Discord and Mattermost reject malformed WebSocket upgrade URLs with HTTP 400 and close the connection while continuing to serve other clients. Recorder files should normally have one filesystem name. If multiple processes cannot share the same OS account home, the home is read-only, or they write through hardlinks to the same recorder inode, set `CRABLINE_RECORDER_LOCK_DIR` to the same absolute writable directory for every writer. Pre-create that directory with the ownership, group, or ACLs required by those writers. Crabline refuses hardlinked server-recorder writes without this shared lock namespace. The configured path must be canonical and contain no symlink components. Scope each lock directory to one recorder filesystem; lock identities omit device numbers so containers that mount the same inode under different device IDs still coordinate. Otherwise, Unix writers coordinate through the OS account's `~/.cache/crabline/locks/server-recorder` namespace. Feishu (programmatic): ```ts import { readFile } from "node:fs/promises"; import { startFeishuServer } from "@openclaw/crabline"; const server = await startFeishuServer({ tls: { key: await readFile("fixtures/server-key.pem"), cert: await readFile("fixtures/server-cert.pem"), }, recorderPath: ".crabline/feishu-events.jsonl", }); // Configure the client with server.manifest and its fixture CA before startup. // Close the client first, then await server.close(). ``` The native subset supports tenant tokens, bot identity, authenticated WebSocket discovery, protobuf events and ping/pong, fragmented UTF-8 events, SDK ACKs, and REST message creation/replies for text, localized posts, and static cards. Each event and all its fragments go to one randomly selected connected client, reserving one ACK slot on that client. Credentials are generated per server unless supplied. The manifest contains credentials; keep it private. TLS is opt-in and the default bind is loopback. See [Feishu setup](docs/channel-setup.md#feishu-programmatic) for admin ingress, resource limits, and receipt semantics. This starter is a library API; the CLI server catalog and OpenClaw bridge do not register Feishu. Discord: ```bash crabline --json serve discord --ready-file .crabline/discord-server.json ``` The manifest contains both bot identities and tokens, fixture guild/text/voice IDs, the REST API root, Gateway metadata and WebSocket endpoints, the voice endpoint and test-only CA, authenticated admin ingress, and recorder path. The server implements the Discord REST v10 and Gateway subset used by the real OpenClaw Discord plugin: bot identity, `/gateway/bot`, JSON Gateway HELLO/IDENTIFY/READY/heartbeat/resume, DM and guild channels, message lifecycle and replies, reactions, threads, retrievable multipart attachments, typing, basic guild/member lookups, application command registration, and authenticated voice-state plus TLS WebSocket/UDP voice negotiation. Admin ingress turns an injected user message into a normal `MESSAGE_CREATE` Gateway dispatch. Trust the voice CA only in the isolated test client process. Replies include the current immediate parent message, without recursively embedding earlier replies. Each parent retains its `message_reference` for explicit lookup; deleted parents return `referenced_message: null`. This local provider server is separate from the fixture-level Discord local mock provider listed under Built-In Mock Channels. The fixture adapter models Discord interactions directly for Crabline commands; the server lets a real Discord client exercise provider-native REST and Gateway boundaries. The supported Gateway is v10 JSON with one shard. Transport compression, voice media encryption/streaming, interactions, audit logs, and Discord's distributed permission and rate-limit systems are outside the current subset. Mattermost: ```bash crabline --json serve mattermost --ready-file .crabline/mattermost-server.json ``` The JSON manifest contains: - `baseUrl`: OpenClaw `channels.mattermost.baseUrl` / `MATTERMOST_URL` - `botToken`: OpenClaw `channels.mattermost.botToken` / `MATTERMOST_BOT_TOKEN` - `webhookToken` / `MATTERMOST_TOKEN`: verifies native outgoing webhook posts - `endpoints.websocketUrl`: native Mattermost WebSocket endpoint - `adminToken`: send this as the `X-Crabline-Admin-Token` header when posting test user messages - `endpoints.adminInboundUrl`: authenticated POST endpoint for inbound messages - `recorderPath`: JSONL file of local provider API/admin traffic The server implements a Mattermost API subset for text DM and channel roundtrips, including REST authentication/status codes, WebSocket authentication and `hello`, sequenced events, typing, and post create/edit/delete events. Admin ingress is only the test control plane; injected messages require native 26-character lowercase channel/user IDs and are delivered to clients as channel-scoped native `posted` events. Optional `channelName` and `channelDisplayName` values populate the native event fields. `POST /api/v4/posts` requires a JSON media type. Matrix: ```bash crabline --json serve matrix --ready-file .crabline/matrix-server.json ``` The JSON manifest contains: - `baseUrl`: OpenClaw `channels.matrix.homeserver` - `accessToken`: OpenClaw `channels.matrix.accessToken` - `botUserId`: OpenClaw `channels.matrix.userId` - `endpoints.clientApiRoot`: Matrix Client-Server API root - `adminToken`: send this as the `X-Crabline-Admin-Token` header when posting test user messages - `endpoints.adminInboundUrl`: authenticated POST control endpoint for inbound messages - `recorderPath`: JSONL file of local provider API/admin traffic The server implements an unencrypted Matrix Client-Server API subset including `whoami`, filters, push rules, joined rooms and members, room state, `/sync`, room event sends, typing, and read receipts. Admin ingress is only the test control plane; injected messages are delivered to clients as native `m.room.message` events through `/sync`. Optional `roomName` sets newly created room state, `direct: true` publishes `m.direct` account data, and `threadId` accepts the raw Matrix root event ID used in the native thread relation. State-event reads return the same stored content advertised through `/sync`, including `m.room.create`, room names, and membership. Slack: ```bash crabline --json serve slack --ready-file .crabline/slack-server.json ``` The JSON manifest contains: - `endpoints.apiRoot`: set OpenClaw's Slack API override / `SLACK_API_URL` to this value - `botToken`: set OpenClaw `channels.slack.botToken` to this value - `signingSecret`: set OpenClaw `channels.slack.signingSecret` to this value - `adminToken`: send this as the `X-Crabline-Admin-Token` header when posting test user messages - `endpoints.adminInboundUrl`: authenticated POST endpoint for Events API-shaped test user messages - `endpoints.eventsUrl`: local Slack Events API endpoint - `recorderPath`: JSONL file of local provider API/admin traffic The admin token is generated randomly unless `adminToken` is supplied through the credential ingress above. Implemented Slack Web API endpoints include `auth.test`, `chat.postMessage`, `conversations.open`, `conversations.info`, `conversations.history`, and `conversations.replies`. Signal: ```bash crabline --json serve signal --ready-file .crabline/signal-server.json ``` The JSON manifest contains: - `endpoints.apiRoot`: `signal-cli daemon --http`-compatible API root - `account`: mock Signal account served by the daemon - `adminToken`: send this as the `X-Crabline-Admin-Token` header when posting test user messages - `endpoints.adminInboundUrl`: authenticated POST endpoint for inbound messages - `endpoints.eventsUrl`: `signal-cli` SSE receive endpoint - `endpoints.rpcUrl`: `signal-cli` JSON-RPC endpoint - `recorderPath`: JSONL file of local provider API/admin traffic The server implements the HTTP surface exposed by `signal-cli daemon --http`: `check`, `events`, and the JSON-RPC methods needed for text sends, typing, receipts, and reactions. Admin ingress injects text-only receive events. It does not replace the `signal-cli` client or pretend to be Signal's public service. Malformed HTTP request targets receive HTTP 400, and the server continues accepting subsequent requests. Telegram: ```bash crabline --json serve telegram --ready-file .crabline/telegram-server.json ``` The JSON manifest contains: - `endpoints.apiRoot`: set OpenClaw `channels.telegram.apiRoot` to this value - `botToken`: set OpenClaw `channels.telegram.botToken` to this value - `adminToken`: send this as the `X-Crabline-Admin-Token` header when posting test user messages - `endpoints.adminInboundUrl`: authenticated POST endpoint for test user messages; OpenClaw reads them through Telegram `getUpdates` - `recorderPath`: JSONL file of local provider API/admin traffic The admin token is generated randomly unless `adminToken` is supplied through the credential ingress above. The inbound endpoint rejects requests without the matching admin header (or `Authorization: Bearer <token>`). Implemented Telegram Bot API endpoints include `getMe`, `sendMessage`, `sendPhoto`, `sendDocument`, `sendVideo`, `sendAudio`, `sendAnimation`, `editMessageText`, `deleteMessage`, `setMessageReaction`, `createForumTopic`, `editForumTopic`, `pinChatMessage`, `unpinChatMessage`, `getUpdates`, `deleteWebhook`, `setWebhook`, `setMyCommands`, `deleteMyCommands`, `sendChatAction`, and `answerCallbackQuery`. Webhook secrets use Telegram's 1-256 character letter/digit/underscore/hyphen alphabet. Text and caption fields enforce native string types and 4,096/1,024 UTF-16 limits, entity metadata is retained, and `@username` method targets resolve to numeric `Chat.id` values. WhatsApp: ```bash crabline --json serve whatsapp --ready-file .crabline/whatsapp-server.json ``` The JSON manifest contains: - `endpoints.apiRoot`: versioned WhatsApp Graph API root - `endpoints.phoneNumberUrl`: provider-native phone-number resource - `endpoints.baileysWebSocketUrl`: Baileys-compatible WebSocket URL for `waWebSocketUrl`, including the local provider access token query parameter - `accessToken`: bearer token for local provider requests - `adminToken`: send this as the `X-Crabline-Admin-Token` header when posting test user messages - `selfJid`: canonical local authenticated WhatsApp user JID; legacy `@c.us` input is normalized to `@s.whatsapp.net` - `env.CLOUD_API_ACCESS_TOKEN`: same value as `accessToken` - `env.CLOUD_API_VERSION`: Graph API version used by the local server - `env.WA_BASE_URL`: local Graph API origin - `env.WA_PHONE_NUMBER_ID`: local sender phone-number resource ID - `endpoints.adminInboundUrl`: authenticated POST endpoint for test user messages using the WhatsApp Business webhook payload shape - `endpoints.messagesUrl`: provider-native Cloud API message and status endpoint - `recorderPath`: JSONL file of local provider API/admin traffic and Baileys WebSocket stanzas Pass `endpoints.baileysWebSocketUrl` to Baileys as `waWebSocketUrl` when a runtime needs to connect through the local provider. Cloud API-compatible clients can send text messages through `POST /v25.0/<phone-number-id>/messages` with a bearer token. The local server completes the Baileys Noise handshake, serves bootstrap/device/prekey queries, and records encrypted outbound WebSocket stanzas. Direct `msg` and `pkmsg` text sends are also recorded as normalized accepted-send evidence for the OpenClaw bridge. Group outbound uses sender-key `skmsg` encryption and is outside this supported subset, so OpenClaw Crabline outbound targets are direct users only. Group inbound injection remains supported. The WebSocket endpoint rejects clients that do not present the access token embedded in the manifest URL. The admin token is generated randomly unless `adminToken` is supplied through the credential ingress above. OpenClaw bridge callers should post injected user messages with the `providerUrl`, `providerHeaders`, and `providerBody` returned by `createOpenClawCrablineInbound()`. Zalo: ```bash crabline --json serve zalo --ready-file .crabline/zalo-server.json ``` The JSON manifest contains: - `endpoints.apiRoot`: trusted `ZALO_API_URL` - `botToken`: OpenClaw `channels.zalo.botToken` / `ZALO_BOT_TOKEN` - `botId`: local Zalo bot identity - `adminToken`: send this as the `X-Crabline-Admin-Token` header when posting test user messages - `endpoints.adminInboundUrl`: authenticated POST control endpoint for inbound messages - `recorderPath`: JSONL file of local provider API/admin traffic The server implements the Zalo Bot API subset used by OpenClaw: `getMe`, single-update long polling through `getUpdates`, `sendMessage`, `sendPhoto`, `sendChatAction`, and webhook registration, inspection, and deletion. These Bot API methods accept GET query parameters or POST requests. Admin ingress injects a native Zalo update into the active polling or webhook transport. Webhook delivery posts the native `{ event_name, message }` update directly. OpenClaw-specific endpoint, config, and target mapping remain in the isolated bridge. Programmatic `startZaloServer()` callers can tune the retained polling queue with `maxPendingInboundEvents` and `maxPendingInboundBytes`; the byte limit defaults to 64 MiB. The admin ingress accepts JSON like: ```json { "chatId": "6ede9afa66b88fe6d6a9", "chatType": "PRIVATE", "senderId": "6ede9afa66b88fe6d6a9", "senderName": "Alice", "text": "user nonce-123" } ``` ## Target IDs Built-in providers accept native channel identifiers. Crabline does not add `telegram:`, `discord:`, `slack:`, or other local prefixes. ```yaml target: id: "C1234567890" ``` Thread targets use the platform's native thread identifier: ```yaml target: id: "C1234567890" channelId: "C1234567890" threadId: "1700000000.000100" ``` Examples: - Slack conversations: `C1234567890`, `G1234567890`, or `D1234567890` - Slack direct sends may also target user IDs such as `U1234567890` or `W1234567890` - Slack threads: `1700000000.000100` - Telegram chats: `-1001234567890` or `@channelusername` - Telegram topics: `42` - Telegram username targets require `@`, contain 5-32 letters, digits, or underscores, and normalize to lowercase. Numeric chat IDs must be nonzero and have an absolute value no greater than `2^52 - 1`. - Built-in WhatsApp Cloud API users: digits-only `wa_id` values such as `15551234567` - Built-in WhatsApp Cloud API fixtures do not support `threadId` targets. - OpenClaw WhatsApp bridge users: `15551234567@s.whatsapp.net` (legacy `15551234567@c.us` inputs are accepted and canonicalized) - OpenClaw WhatsApp bridge groups: `120363001234567890@g.us` - Discord channels and threads: Discord snowflake ids such as `123456789012345678` - iMessage recipients: E.164 phone numbers such as `+15551234567`, email addresses such as `user@example.com`, or chat GUIDs such as `iMessage;-;chat-guid` and `SMS;+;chat-guid` - Matrix rooms: scoped IDs such as `!abcdef:matrix.org` or Matrix v12 domainless room IDs; thread IDs are raw root event IDs such as `$eventid:matrix.org` - Mattermost channels, users, and root posts: exactly 26 lowercase alphanumeric characters, such as `abcdefghijklmnopqrstuvwx12` - Microsoft Teams: non-empty opaque Bot Connector `conversation.id` values such as `a:opaque-conversation-id` or `19:conversation@thread.v2` - Zalo users, OAs, and chats: non-whitespace provider IDs such as `user-1` or `group-1` Scoped Matrix room, event, and historical user IDs must contain well-formed Unicode. Unpaired UTF-16 surrogates are rejected; valid astral characters are preserved. ## Webhooks Each built-in provider starts a local webhook during `probe`, `waitForInbound`, or `watch`. Webhook requests can use the provider's native event shape, or this simple JSON shape with native thread ids: ```json { "id": "slack-inbound-1", "threadId": "C1234567890", "text": "reply nonce-123", "author": "assistant" } ``` Nested message payloads are also accepted: ```json { "message": { "id": "slack-inbound-1", "threadId": "C1234567890", "text": "reply nonce-123" } } ``` Malformed webhooks return `400`, and non-JSON requests return `415`. Configured webhook paths must be canonical URL pathnames without normalization, query, or fragment ambiguity. ## Evidence Flow `send` records an outbound user event in the provider recorder. For `roundtrip` and `agent` modes, the local mock also records a deterministic assistant reply: ```text [telegram mock] hello nonce-123 ``` `waitForInbound` reads the recorder until it finds a matching non-user event. `watch` streams matching recorder events. This gives CI channel coverage without live service latency, external credentials, webhooks exposed to the internet, or provider SDK state. Recorder files remain JSONL. Individual events use one object per line; an accepted multi-message webhook is committed as one versioned batch object so readers never observe a partial batch. Crabline's recorder APIs flatten those batch objects back into their individual events. ## More Setup Detail See [Channel Setup](docs/channel-setup.md) for the provider matrix, webhook paths, and OpenClaw live-vs-mock guidance.

正在加载仓库 README…

这个仓库中的 skills

软件质量保证分析师与测试员

1 个职业分类 · 已分类 100%

skill

职业分类

描述

更新

autoreview

软件质量保证分析师与测试员

Pre-commit/ship code review: Codex default; optional Claude or Pi.

原文语言：英语

2026-09-19

已展示 1 / 1 个已收集 Skill。