Files
WebHooker/AGENTS.md
T
翅膀 5fcb129170 docs(worker): align guides with runtime behavior
Synchronize both language guides, API references and examples with
the implemented authorization, storage and delivery behavior.
Document migration requirements and accepted operational limits.
2026-10-23 23:24:05 +08:00

47 KiB
Raw Blame History

AGENTS.md — WebHooker

Project Purpose

Nuxt 4 (Nitro) app deployed as a Cloudflare Worker that receives GitHub webhooks and dispatches processed events to Discord channels/threads, Telegram chats/topics and Feishu group chats, and receives Discord interactions (slash commands, buttons, modals) via the Interactions Endpoint plus Telegram bot /gh commands and Feishu bot /gh commands + card buttons via their respective webhooks.

Core pipeline: GitHub Webhook → Worker (verify + filter + format) → Discord (REST) / Telegram (Bot API)

Key Decisions

  • Runtime: Cloudflare Workers via the Nitro cloudflare_module preset (entry .output/server/index.mjs), H3 event handlers in server/routes/
  • UI: Vue 3 + Tailwind CSS v4 (@tailwindcss/vite) + @elysola/ui (Nuxt module, Rc prefix); admin console is a client-side SPA (routeRules: "/admin/**": { ssr: false }), home/legal pages render server-side; app-level composites (FilterCombobox, ConfirmDialog) are built on the reka-ui primitives that @elysola/ui ships (same copy, pre-bundled via vite.optimizeDeps.include so dev keeps a single instance and dialogs/popovers stack correctly) and styled like their Rc counterparts
  • Styling: Elysola UI supplies semantic components and blue theme tokens; app/assets/css/main.css retains --wh-* RGB-triplet aliases for app-specific Tailwind utilities and bespoke layouts (avoiding collision with Elysola's --surface/--text/--border hex tokens); theme switching supports auto/light/dark via useTheme / ThemeToggle, toggling <html>.dark with inline anti-flash head script
  • Discord interactions: HTTPS Interactions Endpoint (POST /discord/interactions, Ed25519-signed) — no Discord Gateway / Durable Object; bot stays offline, messages always sent via REST
  • Storage: D1 is the source of truth for config (routes/groups via d1_routes/d1_groups, see server/lib/storage/config-store.ts; each route's per-group order lives in d1_routes.position — migration 0012 — and routes load ordered by group_id, position, id, falling back to id order on an unmigrated database). Config writes are per group — one group row (createGroup insert-if-absent, updateGroup compare-and-swap on d1_groups.version), one group's routes or fragments in one transaction (saveGroupRoutes, saveGroupFragments), renameGroup/deleteGroup as single batches — and never fall back to KV: a failed D1 write is an error. Access checks, read-modify-write paths, webhook ingress and the queue consumer read with { fresh: true } (no cache, so revoked routes/targets stop at once); other reads are cached in memory ≤60 s per isolate and loadConfig returns a fresh object per call. A stored route or group row that cannot be read (corrupt JSON, wrong shape, a pattern RE2 cannot compile) is reported as unreadable for its group only: that group's deliveries fail visibly and retry, and its routes cannot be replaced without discardUnreadable. Legacy KV config:* is seeded into D1 once (marker config:d1-seeded, only while both tables are empty); server/middleware/config-store.ts binds the D1 store on every request. The admin API returns a content revision for every group and per-group route/fragment list and answers 409 when a write's revision is stale; the check runs inside the write's transaction (groupVersionGuard aborts the D1 batch unless d1_groups.version is unchanged, and every route/fragment write bumps it), so two writes from the same revision can never both succeed. Rename/delete commit history moves/detachment and the retired-id marker in the same batch. Every group carries an immutable server-assigned instance (never sent to clients; existing groups got one in a one-time identity migration, which also sets legacyId): secrets, invites, delivery/dedup scopes (@i:{instance}), target ledgers, message-tracking chains and late send/audit rows follow the instance, so a group re-created under a reused id inherits nothing of the former one. Dispatch never treats unreadable configuration as empty (no KV copy → error → retry). Also in D1: send/audit logs, dedup (dedup_keys), delivery state (delivery_state) and message tracking (message_tracking via server/lib/storage/d1.ts canUseD1 gate). KV keeps only short-lived/ephemeral state (tokens, OAuth state, admin sessions, invites, Feishu token / Discord command-sync / Telegram webhook records, i18n overrides; legacy tenant:* secrets are only read as a fallback) with explicit TTLs. R2 parks oversized queue payloads (webhooks/YYYY/MM/DD/*.json, server/lib/storage/payload.ts) instead of KV. A storage-prune scheduled task cleans up expired dedup/delivery/message-tracking rows
  • Signature verification: Web Crypto API (HMAC-SHA256 for GitHub/Gitea, Ed25519 for Discord, timing-safe secret-token compare for Telegram)
  • Webhook providers: pluggable forge adapters under server/lib/providers/ (github, gitea) — each verifies its own signature format and normalizes its payload to a GitHub-shaped WebhookEvent; a custom provider accepts arbitrary signed JSON posts (X-WebHooker-Signature) as custom events; GitLab etc. can be added later
  • Per-group webhook ingress: optional POST /webhook/{groupId} with a per-group secret stored in the group's D1 row (rotated/removed by CAS with an owner re-check; legacy KV tenant:* read as a fallback) — Gitea/classic-GitHub/custom webhooks are verified against the group's secret instead of the operator's global ones; only that group's routes fire. Tenant-signed payloads lose their installationId (no App-token enrichment, no installation provisioning; the installation check is skipped for the group's own ingress). The legacy POST /webhook (global secrets, all routes except isolated groups without a bound source) stays as before
  • GitHub App tenant isolation: Group.installationId binds a group to one GitHub App installation; events whose payload.installation.id differs are rejected at dispatch (hard isolation on top of the optional owners list). The App's Setup URL points at GET /auth/github/install, which renders a choice page (create inst-{installationId} or bind to a group the signed-in user owns, verified by role); POST /auth/github/install/bind performs the provisioning. Both first verify with the user's own OAuth token (scopes repo read:org) that they manage the installation (userManagesInstallation: the installation's User account, or an active admin of its organization — mere access is not enough; needs GITHUB_APP_ID/GITHUB_PRIVATE_KEY; super admins bypass); self-signup only makes the installer owner of the dedicated inst-{id} group while it is bound to that installation and has no members. Ids inst-* / u-* are reserved for super admins, and an unbound existing inst-* group is never auto-bound. installation.created webhook events (global endpoint only) auto-provision as a fallback (create inst-{installationId} or bind existing groups whose owners match the installing account)
  • GitHub OAuth: octokit (token is stored hashed for reverse lookup)
  • Admin WebUI: /admin config console, OAuth-session protected via ADMIN_USER_IDS whitelist (local-only DEBUG_MODE=true bypasses authentication); left sidebar navigation (Overview / Groups / Send Logs / Audit / Stats) + topbar, deep-linkable routes (/admin overview · /admin/groups · /admin/groups/{groupId} a group's routes, members and webhook · /admin/logs · /admin/audit · /admin/metrics, parsed by app/utils/admin-path.ts; the admin page's validate 404s any other path) — the page has a fixed key, so one console instance (and its in-flight saves) persists across these URLs while each view refreshes its data when shown; every page is a history entry, so the browser's back button returns from a group to the list (no in-page back button) and leaving with an open editor or an unsaved route order asks first; route switches/actions lock while a route write is in flight; sign-out asks for confirmation (sidebar, or a top-right button on narrow screens); editor dialogs open without focusing a field (useDialogFocus); the overview dashboard (AdminHome.vue) aggregates delivery KPIs, a routes table and recent send logs; the stats page (MetricsPanel.vue) renders KPI cards plus per-platform/per-event ok/failed bar legends and a per-status breakdown, and filters by group (?groupId=); the route/group editors (RouteEditor.vue / GroupEditor.vue) are centered, scrollable RcDialog forms with an eyebrow header and sectioned bodies
  • Access control: every group has role-based members (owner / admin / viewer); super admins bypass; legacy adminIds are read as owners (backward compatible); owners manage members + invites; the source-isolation fields owners, installationId and isolated stay super-only (an owner's write keeps the stored values). Self-signup personal groups are isolated: true — the global endpoint never dispatches to them until a super admin binds a source (groupAcceptsGlobalIngress); routes whose group is missing never fire
  • Invites: single-use 7-day links (invite:{token}, KV TTL derived from expiresAt) for joining a group as admin/viewer; an invite belongs to the issuing group's instance; acceptance holds a short claim and commits the invite-used record in the same D1 transaction as the CAS member update, guarded so a used or revoked invite (revocations are D1 records) or one whose issuer is no longer an owner never admits anyone; ALLOW_SELF_SIGNUP=1 gives access-less users an isolated personal group on first login (self-service SaaS entry)
  • Group lifecycle (server/lib/web/group-lifecycle.ts): delete detaches send/audit history (group_id = NULL), drops message tracking, revokes invites + webhook secret and retires the id (group-retired:{id} in the idempotency store, 4 days); rename moves routes + fragments and send/audit history atomically and records a group-alias:* row (D1, 4 days; KV group:renamed:* without D1) in the same transaction so queued tenant deliveries of pre-instance messages follow; secrets, invites and message tracking stay with the instance; the old id is retired
  • Audit log: every admin operation (logins, group/route/member/invite changes) recorded in D1 audit_logs; pruned by the scheduled trigger after AUDIT_RETENTION_DAYS (default 90)
  • Group webhook log channel: optional Group.logTarget (Discord channel/thread or Telegram chat/topic) receives one summary message per webhook the group's routes dispatched (event, repo, delivery id, per route×target ✅/❌ outcome, green/red color); best-effort, not recorded in send_logs
  • Per-group forge branding: optional Group.forgeSources (a list of { host, type: "github" | "gitea", name? } the group defines itself) labels each message's footer with the first entry whose type matches the event's provider and whose host matches the repository URL's hostname (GitHub matches github.com, so two Gitea instances can be git1.example.com/git2.example.com); the label is the entry's optional name (fallback: host); links are derived from the repo URL and footer icons use raster PNGs Discord renders (GitHub's fluidicon.png, Gitea's /assets/img/favicon.png — .ico favicons are silently ignored); Discord shows the footer icon + name, Telegram a linked name
  • Delivery queue: when the QUEUE binding is present, webhook ingress enqueues one message per event (not per target) to a Cloudflare Queue (webhooker-delivery) and answers 200 only after the send succeeded (503 + dedup claim released otherwise); a Nitro cloudflare:queue plugin consumes batches, takes each delivery with a 5-minute lease (acquireDelivery, conditional UPSERT on delivery_state), dispatches, and retries when any target failed transiently (5xx/429/network/rate-limit/FEISHU_TOKEN, a crashed route, an unreadable ledger, a busy update lock) with exponential backoff (5s/30s/2m/10m) up to the queue max_retries, after which the DLQ (webhooker-delivery-dlq) marks the delivery dead. Targets that succeeded or failed permanently are recorded (delivery-target:* in the idempotency store, keyed by the group's instance, written right after each target and read with one range query) and skipped on retries, so retries never duplicate and one bad target never cancels the others; a permanent failure is remembered (delivery-failed:*) so the delivery ends failed, not delivered. A message whose storage calls fail is retried on its own (the batch is never failed whole); pre-upgrade delivery-state:{provider}:{groupId}:* rows are honored only for the group that owned the id (legacyId), never for the global endpoint. The consumer runs max_batch_size: 3 with one per-batch budget (12 minutes, 20 sends — the D1 query / subrequest limits are per invocation): a message left unstarted is retried (copied instead on its last retry if it was never started), and one that made progress continues in a new message. Oversized payloads (queue message over 120,000 UTF-8 bytes) are parked in R2 (webhooks/YYYY/MM/DD/*.json, falling back to KV queue:payload:* without the R2 binding, 4-day TTL; kept when the queue send fails, since the queue may have accepted it; orphans pruned by storage-prune) and resolved by the consumer — a missing/corrupt parked payload never dispatches {}: the delivery stays retrying with a PAYLOAD_MISSING send-log entry asking for a redelivery; delivery state is tracked as the D1 delivery_state table (KV delivery-state:* fallback). Without the QUEUE binding, dispatch stays inline (existing behavior)
  • Storage quotas: the design avoids KV write pressure (Workers Free KV allows ~1,000 writes/day) by keeping high-frequency ephemeral writes in D1 instead — webhook dedup and every other one-time claim (dedup_keys, atomic ON CONFLICT UPSERT via idempotencyStore with has/claim/release), delivery state and message tracking all use D1 rows via canUseD1 (prepare+batch probe) with automatic KV fallback when D1 is unavailable or a table is missing (isMissingTable, unmigrated); other D1 errors propagate so one claim never splits across D1 and KV. D1 Free allows 100,000 rows written/day, so the per-event KV write cost drops to ~0; config is cached in memory only, with D1 authoritative. In-flight work shared across concurrent requests (setup, fallback marker, Feishu / App tokens, translations) goes through sharedFlight, which waits on another request's run only for a bounded time (a canceled request's promise may never settle)
  • Filter DSL: a route's flat filters can be replaced by a nested ast (all/any/not nodes, server/lib/events/filter-ast.ts) that takes precedence when present; the field filter type reads any payload value by JSONPath (path, arrays expand so any element matches) and 12 comparison operators (op: eq/ne/contains/startsWith/endsWith/regex/gt/gte/lt/lte/in/exists) extend the legacy glob/regex/case-insensitive-exact semantics; named filter fragments (D1 d1_fragments via server/lib/fragments.ts, per-group) are editor-side templates inlined into the route's ast on insert — the matcher never resolves fragment references; the route editor (RouteEditor.vue) exposes a recursive AST builder (FilterNodeEditor.vue: plain selects for filter type/operator, an all/any segmented toggle, NOT toggle + delete pinned per condition), searchable multi-value / path inputs (FilterCombobox.vue) fed by a bilingual suggestion catalog (app/utils/filter-catalog.ts: GitHub events/actions from @octokit/webhooks plus Gitea extras, common payload paths and their known values, English keys searchable in either language) where custom values are allowed but rejected when the matcher could never match them (validateFilterValue), a live explanation and a stateless POST /admin/api/test-match dry-run
  • Local dev: wrangler + Miniflare

Architecture

app/                     # Vue 3 UI (Nuxt app dir)
├── app.vue              # root component (NuxtPage)
├── assets/css/main.css  # Tailwind v4 entry: Elysola-compatible RGB aliases + bespoke layout classes + Vue transition glue
├── pages/               # index (landing), terms, privacy, admin/[...slug] (console SPA)
├── components/          # ConsolePage (sidebar shell + topbar), AdminHome (overview dashboard),
│                        # RouteList (drag/keyboard reorder mode) + RouteCard, RouteEditor (FilterNodeEditor
│                        # AST builder + FilterCombobox searchable inputs), GroupEditor, MembersPanel, WebhookPanel,
│                        # ConfirmDialog (app-wide destructive-action confirm), SendLogs, AuditLog, MetricsPanel,
│                        # AppToasts, LegalLayout
├── composables/         # useApi (ApiError + isConflict for 409s), useI18n, useToasts, useConfirm (promise-based confirm), useBackdropGuard (editor dialogs ignore
│                        # backdrop clicks), useDialogFocus (editor dialogs open without focusing a field), useGroups (per-group create/update/remove),
│                        # useGroupRoutes + useFragments (send the loaded revision, reload on 409), useFilterNode (AST helpers), useLogs,
│                        # useAudit, useInvites, useWebhook
├── types.ts             # shared client types (Route, Group, Filter, ...)
└── utils/              # legal.ts (terms/privacy HTML bodies, zh/en), format.ts, admin-path.ts (console URL parsing),
                         # filter-catalog.ts (bilingual filter suggestions + value/path validators)
server/                  # Nitro server
├── routes/              # H3 handlers: /health, /webhook[/:groupId], /discord/interactions, /telegram/webhook,
│                        # /auth/github* (incl. POST /auth/github/link — bot account-link confirmation),
│                        # /admin/{login,logout,invite,api/**}, /api/{comment,merge,close,react,richheader}
├── middleware/          # config-store (binds the D1 config store on every request)
├── tasks/               # scheduled (cron */5): discord-sync, telegram-sync, audit-prune, storage-prune
├── plugins/             # Nitro plugins: queue-consumer (hooks cloudflare:queue → handleQueueBatch)
├── error-handler.ts     # JSON error handler
└── lib/
    ├── types.ts         # Env, Config, Route, Filter, Group, WebhookEvent, NeutralMessage
    ├── config.ts        # getConfigStore/initConfigStore (D1 store, else KV-only), loadRoutes/saveRoutes/saveGroupRoutes, loadConfig (fresh object per call)
    ├── fragments.ts     # named filter fragments (loadFragments(groupId?)/saveGroupFragments — D1 d1_fragments, no KV; read errors propagate)
    ├── config/          # config schema + migration + validation
    │   └── schema.ts    # CONFIG_SCHEMA_VERSION, valibot route/group/filter schemas, migrateRoutes/Groups, validateRoutes/Groups (non-destructive), explainRoute
    ├── cf.ts            # cfEnv(event) — env bindings from event.context.cloudflare
    ├── http.ts          # shared HTTP helpers
    ├── webhook.ts       # processWebhook/handleWebhook: tenant lookup, provider detect/verify/parse, tenant installation strip, dedup (D1/KV), awaited enqueue (503 + claim release on failure) or inline dispatch
    ├── queue/           # Cloudflare Queue delivery pipeline
    │   ├── delivery.ts  # DeliveryMessage, enqueueWebhook (byte-size check, parked-payload cleanup), classifyDelivery/isRetryableFailure, retry backoff,
    │   │                # acquireDelivery lease, delivery-state (D1 w/ KV fallback), per-target done ledger, payload parking (R2 → KV), PayloadUnavailableError
    │   └── consumer.ts  # handleQueueBatch: lease → resolve payload → dispatch (tenant alias, skipTarget on retries) → classify → ack/retry/DLQ
    ├── core/
    │   └── dispatch.ts  # Platform-neutral dispatch: source checks (tenant/isolation/installation/owners/providers) → match routes → formatEvent (FORMAT
    │                    # failures recorded) → driver.send or deliverUpdate (locked, versioned, forge-origin-keyed edit chains) → batched recordSend + group log
    ├── events/
    │   ├── match.ts     # matchRoute, eventOwners; unified pattern syntax (*/? globs + //-wrapped regex)
    │   └── filter-ast.ts # FilterNode evaluator (all/any/not + field JSONPath + 12 ops), containsKeyword, explainFilter/explainFilterNode; pattern helpers
    │                    # (globs and user regexes compiled with RE2 via `re2js` — linear-time matching, no ReDoS; bounded cache with separate
    │                    # key spaces; regexSafetyIssue rejects patterns RE2 cannot compile, e.g. backreferences/lookaround)
    ├── providers/       # Forge webhook providers (verify + parse/normalize to GitHub-shaped events)
    │   ├── types.ts     # Provider interface (matches/verify/parse)
    │   ├── hmac.ts      # HMAC-SHA256 + timing-safe compare helpers
    │   ├── index.ts     # detectProvider() registry (gitea, github, custom)
    │   ├── github/      # X-GitHub-Event + X-Hub-Signature-256 ("sha256=" prefix); extracts installation.id
    │   │   ├── verify.ts
    │   │   └── parse.ts
    │   ├── gitea/       # X-Gitea-Event + X-Gitea-Signature (plain hex HMAC)
    │   │   ├── verify.ts
    │   │   └── parse.ts # parse + normalize Gitea payloads to GitHub shape
    │   └── custom/      # X-WebHooker-Signature (sha256= HMAC) + arbitrary JSON → `custom` events
    ├── formatters/      # Platform-neutral message formatters (was formatter.ts)
    │   ├── index.ts     # formatEvent: builds FormatContext → findFormatter → .format (falls back to formatGeneric)
    │   ├── types.ts     # FormatContext + EventFormatter (formatter plugin interface)
    │   ├── registry.ts  # eventFormatters[] + findFormatter() — 29-event formatter plugin registry
    │   ├── colors.ts    # GITHUB_COLORS + WORKFLOW_CONCLUSION_EMOJI
    │   ├── helpers.ts   # emojiPrefix, T, buildMessage, commitLink/branchLink/tagLink
    │   └── *.ts         # push, pull-request, issues, comments, workflow, release, create,
    │                    # repo, check, review, commit-comment, deployment, member, label,
    │                    # milestone, discussion, repository, security, generic, ping, custom
    ├── drivers/         # Platform drivers (pluggable push targets)
    │   ├── types.ts     # PlatformDriver interface + SendResult (send + edit)
    │   ├── index.ts     # getDriver() registry (discord default + telegram + feishu)
    │   ├── discord/
    │   │   ├── index.ts       # DiscordDriver: send/edit → renderNeutralMessage + rest.sendMessage/editMessage
    │   │   ├── render.ts      # renderNeutralMessage: NeutralMessage → Discord FormattedMessage
    │   │   ├── rest.ts        # Discord REST sendMessage/editMessage with retry + rate-limit handling
    │   │   ├── interactions.ts # Ed25519 verify + interaction handlers (/gh, buttons, modals)
    │   │   └── commands.ts    # APP_COMMANDS + registerGlobalCommands/syncGuildCommands/syncCommands
    │   └── telegram/
    │       ├── index.ts      # TelegramDriver: send/edit → renderNeutralMessage + rest.sendMessage/editMessageText (always text messages;
    │       │                 # avatar as link preview: rich-header card, or the avatar image without a host)
    │       ├── render.ts     # renderNeutralMessage: NeutralMessage → Telegram HTML (parse_mode HTML)
    │       ├── rest.ts       # Telegram Bot API sendMessage/sendPhoto/editMessage* (chat_id + message_thread_id), retry
    │       ├── updates.ts    # POST /telegram/webhook: secret-token verify (always) + handleTelegramUpdate
    │       ├── secret.ts     # telegramWebhookSecret: TELEGRAM_WEBHOOK_SECRET or a secret derived from TELEGRAM_TOKEN
    │       └── commands.ts   # Telegram /gh login (private chat only)|logout|comment|merge|close + reply-message parsing + update_id dedup + syncTelegramWebhook
│   └── feishu/
│       ├── index.ts      # FeishuDriver: send/edit → renderNeutralMessage + rest.sendMessage/updateMessage
│       ├── render.ts     # renderNeutralMessage: NeutralMessage → Feishu interactive card
│       ├── rest.ts       # getTenantAccessToken (memory + KV cache per app, refused tokens dropped) + sendMessage/updateMessage/sendText/updateCard, retry
│       └── updates.ts    # FEISHU_ENCRYPT_KEY signature verify + decrypt / FEISHU_VERIFICATION_TOKEN check (refuses all without either) + url_verification
│                         # + event_id dedup + /gh commands (login p2p only) + card.action.trigger buttons
    ├── github/
    │   ├── oauth.ts     # OAuth URL, callback token exchange, getUserOctokit, getInstallationDetails/userManagesInstallation, comment/getComment/editComment/deleteComment/merge/close actions
    │   ├── app-token.ts # createAppJwt-backed getInstallationToken + GITHUB_API_VERSION (shared App auth) + githubApiUrl (only https://api.github.com/repos/* is fetched, no redirects)
    │   ├── check-run.ts # getCheckSuiteBuildLogUrl: Check-suite build-log enrichment (installation token)
    │   ├── workflow-jobs.ts # getWorkflowRunJobs: fetch a workflow_run's jobs via jobs_url (App token or unauthenticated)
    │   └── store.ts     # KV token CRUD (rotation drops the old reverse key; findUserIdByToken checks the current stored token) + D1 discord/telegram/feishu links
    ├── web/             # HTTP UI/API logic (called from server/routes)
    │   ├── oauth.ts     # handleOAuthStart/Callback (state cookie for browser logins), bot-link confirmation page + handleLinkConfirm, install page + bind
    │   │                # (installation access verified), isolated personal-group self-signup, token delete (own token or super admin)
    │   ├── actions.ts   # POST /api/comment|merge|close|react (Bearer token auth via shared middleware)
    │   ├── admin.ts     # adminLogin/Logout, adminApi* (routes|groups|me|logs|invites|audit|webhook|metrics|delivery), per-group create/update/delete,
    │   │                # revisions + 409 conflicts, super-only isolation fields, SQL-scoped logs/audit/metrics
    │   ├── auth.ts      # Shared auth middleware + guards: requireAnyAccess, requireGroup(Role), bearerUserId, clientIp
    │   ├── invites.ts   # Invite CRUD (KV invite:{token}, 7d TTL) + acceptInvite (join group as admin/viewer)
    │   ├── session.ts   # Session CRUD (KV session:{id}), isAdminUser, cookie helpers
    │   ├── groups.ts    # Group CRUD wrappers (create/update CAS/delete/rename via the config store), member roles (normalizeGroupMembers/memberRole),
    │   │                # source checks (groupAccepts*), ensureInstallationGroup (lock-free), resolveScope + role helpers (roleAt/canEditRoutes/canEditGroup)
    │   ├── group-lifecycle.ts # retire/alias group ids, move/detach group history on rename/delete
    │   ├── tenants.ts   # Per-group webhook secret (32-byte random hex in the D1 group row, CAS + owner re-check; legacy KV tenant:* read fallback)
    │   └── richheader.ts # GET /api/richheader: Open Graph page for Telegram avatar link-preview card
    ├── observability/  # delivery metrics aggregation
    │   └── metrics.ts  # DeliveryMetrics + getDeliveryMetrics (SQL GROUP BY over send_logs in one D1 batch: totals, per platform/event/status, duration, attempts, recent failures)
    ├── storage/        # storage primitives: canUseD1 probe, D1 config store (memory cache ≤60 s → D1), R2 payload store
    │   ├── d1.ts       # canUseD1(db) — prepare+batch probe for D1 availability (falls back to KV)
    │   ├── config-store.ts # ConfigStore: d1ConfigStore (per-group writes, CAS updateGroup, atomic rename/delete, one-time KV seed, one-time identity migration, unreadable-row isolation)
    │   │                   # and kvConfigStore (no D1); `position` order when migrated; ConfigConflictError
    │   └── payload.ts  # PayloadStore backed by R2 (webhooks/YYYY/MM/DD/*.json), no-binding stub throws
    └── lib/             # shared infra
        ├── i18n.ts      # loadTranslations (KV i18n:{lang} overrides), t() with param interpolation
        ├── idempotency.ts # IdempotencyStore (has/claim/release) + kv/d1 stores (D1 atomic UPSERT, KV fallback only for a missing table) + tenantScope (`@global`) + deliveryKey
        ├── correlation.ts # newCorrelationId() — per-request/delivery correlation id for logs + responses
        ├── shared-flight.ts # sharedFlight: in-flight work shared by concurrent requests, waited on for a bounded time
        ├── message-tracker.ts # MessageTracker (versioned { messageId, version } entries, legacy bare ids read as version 0) + kv/d1 trackers with KV fallback
        ├── send-log.ts  # SendRecord, recordSend/getSendLog/getSendLogById/getSendLogByDelivery/getFailedSendLog (D1 send_logs)
        ├── send-log-batch.ts # recordSendBatch (D1 batch INSERT for send_logs)
        ├── audit.ts     # recordAudit/getAuditLog/pruneAuditLogs (D1 audit_logs, best-effort writes)
        ├── log.ts       # JSON console logger (info/warn/error/fatal)
        └── locales/     # en.ts, zh.ts translation dictionaries

tests/                   # bun test unit tests (webhook, formatter, discord, telegram, admin, groups, invites, audit, send-log, token-store,
                         # data-integrity, ingress-dispatch, discord-interactions, telegram-webhook, ...)
tests/d1-sqlite.ts       # in-memory SQLite behind the D1 API (all migrations applied, FKs on) + a Map-backed KV, for real-SQL tests
tests/fixtures/          # provider payload fixtures (github/gitea/custom) feeding provider + formatter tests
tests/__snapshots__/     # formatter snapshot golden files (toMatchSnapshot)

Responsibilities

  • Verify GitHub webhook signatures (Web Crypto HMAC-SHA256, X-Hub-Signature-256)
  • Verify Gitea webhook signatures (Web Crypto HMAC-SHA256, plain hex X-Gitea-Signature)
  • Verify custom webhook signatures (Web Crypto HMAC-SHA256, GitHub-style sha256= via X-WebHooker-Signature; optional replay protection via X-WebHooker-Timestamp + X-WebHooker-Nonce — signature over {timestamp}.{nonce}.{body}, ±5 min window, nonces claimed in the idempotency store; only JSON objects are accepted)
  • Normalize Gitea webhook payloads to a GitHub-shaped WebhookEvent (push compare_url → compare, pull_request_comment → pull_request_review_comment, ...)
  • Verify Discord interactions (Web Crypto Ed25519, X-Signature-Ed25519 over timestamp + body)
  • Verify Telegram webhook calls (X-Telegram-Bot-Api-Secret-Token always: TELEGRAM_WEBHOOK_SECRET or a secret derived from TELEGRAM_TOKEN; closed without a token)
  • Verify Feishu callbacks (X-Lark-Signature = SHA-256 of timestamp + nonce + FEISHU_ENCRYPT_KEY + body, AES-256-CBC decryption of encrypted bodies, and/or the FEISHU_VERIFICATION_TOKEN in the payload; all callbacks are refused when neither is configured)
  • Filter events by: event type, repo name, actor, action, branch, keyword, or any payload field (field with a JSONPath path — arrays expand so any element matches) — every filter type supports */? glob matching and //-wrapped regular expressions (case-insensitive) plus 12 comparison operators (op: eq/ne/contains/startsWith/endsWith/regex/gt/gte/lt/lte/in/exists)
  • Combine filters into an AST (Route.ast: all/any/not nodes) that overrides the flat filters AND-list; explainRoute/explainFilterNode render a human-readable description of the tree
  • Store named filter fragments (D1 d1_fragments, per group, GET/PUT /admin/api/groups/:groupId/fragments) as editor-side templates; inserting a fragment inlines its node into the route's ast — dispatch never resolves fragment references
  • Offer a stateless filter dry-run (POST /admin/api/test-match, body { node | filters, event?, payload }) that evaluates in memory and returns { matched, explanation } — no event payload is stored
  • Validate route/group config against valibot schemas (validateRoutes/validateGroups only log warnings; a D1 row that fails the shape checks on load is not loaded but reported as unreadable for its group); CONFIG_SCHEMA_VERSION marks the schema version and migrateRoutes/migrateGroups migrate legacy shapes (e.g. target → targets)
  • Filter routes by group owner restriction (Group.owners), group source-platform restriction (Group.providers: github/gitea), GitHub App installation restriction (Group.installationId), isolation (Group.isolated groups take no global-endpoint events without a bound source) and group existence (routes of a missing group never fire) — all before any filter is evaluated — and skip a group's fallback routes whenever a regular route of that group matched; stop evaluating the group's further routes when a matched route has stop: true (both scoped per group, so one group's routes never silence another's)
  • Auto-provision GitHub App installs: the App's Setup URL flow (/auth/github/install choice page + POST /auth/github/install/bind, owner-role verified for existing groups) and the installation.created webhook fallback both create inst-{installationId} groups or bind existing groups
  • Enforce role-based access on every admin API: super admins bypass, owner manages the group (routes/members/invites/settings), admin edits routes, viewer is read-only; legacy adminIds groups resolve to owner members
  • Issue single-use 7-day group invite links (invite:{token}); accepting joins as admin/viewer (never owner); ALLOW_SELF_SIGNUP=1 creates a deterministic personal group (u-{userId}) on first login
  • Record every admin operation (login/logout, group/route/member/invite changes) to D1 audit_logs; the scheduled trigger prunes entries past AUDIT_RETENTION_DAYS
  • Mention Discord roles on route trigger: route-level discordRoleIds are rendered as <@&id> into the Discord message content (Telegram targets ignore the field)
  • Format 28 GitHub/Gitea event types plus custom webhooks as platform-neutral messages (Discord embeds + Telegram HTML)
  • Show the forge source (named per Group.forgeSources host entries) in the message footer when the group defines a host matching the event's repository
  • Route messages to Discord channels/threads and Telegram chats/topics via REST
  • Edit already-sent messages in place for workflow_run / check_run progress (stable updateKey, chain keyed by group + route + forge origin (provider@host) + update key and target platform, message_tracking via D1 with KV msg:* fallback; older events never overwrite newer state, a transient edit failure is reported instead of posting a duplicate, a new message is sent only when the old one is gone); workflow_run keys on the run id while check_run keys on check suite (or app) + check name + commit SHA (falling back to the run id) so providers that emit a fresh run id per phase — e.g. Cloudflare Pages — still edit the same message
  • Record every dispatch attempt to D1 send_logs (route id, event, target, ok/error, duration, error code)
  • Aggregate delivery metrics (server/lib/observability/metrics.ts) from send_logs — totals, ok/failed counts + failure rate, per-platform/per-event/per-status breakdowns, average duration and attempts, recent failures; getDeliveryMetrics(db, groupId?) sends all five queries as one D1 batch (totals, average duration and attempts share one scan) (a single round trip and a consistent snapshot) and scopes every query by group_id when a group is passed
  • Expose admin observability endpoints — GET /admin/api/metrics?groupId= (delivery metrics; every aggregate is limited in SQL to the reader's groups, getDeliveryMetrics(db, groupId | groupIds)) and GET /admin/api/delivery/:deliveryId (the delivery's attempts that belong to the reader's groups); logs and audit are scoped in SQL before the limit — through the /admin/api/[...slug] catch-all route (server/routes/admin/api/[...slug].ts) that wires every admin API handler to its method+path
  • Serve a per-group webhook ingress (POST /webhook/{groupId}, per-group secret in the group's D1 row) for Gitea/classic-GitHub/custom senders; only that group's routes fire; dedup keys are provider- and tenant-scoped (delivery:{provider}:@i:{instance}:{id} via idempotencyStore, D1 dedup_keys with KV fallback)
  • Issue a per-request correlation id (requestId) in webhook responses and dispatch logs
  • When the QUEUE binding is present, enqueue each verified webhook as a single Queue message (webhooker-delivery) instead of dispatching inline; the consumer resolves the payload, re-scopes routes to the tenant group, and dispatches; retryable failures (5xx/network/429-exhaustion) are retried with exponential backoff (5s/30s/2m/10m) up to the queue max_retries, then the DLQ marks the delivery dead
  • Track delivery state in the D1 delivery_state table (delivery-state:* KV fallback: pending/processing/delivered/retrying/failed/dead) so redelivered messages are skipped idempotently; oversized payloads are parked in R2 (webhooks/YYYY/MM/DD/*.json, KV queue:payload:* fallback) and deleted after dispatch
  • Prune expired D1 rows via the scheduled storage-prune task (dedup keys past expiry, delivery state older than 7 days, message tracking older than 30 days; audits via audit-prune)
  • Send a per-event summary (event, repo, delivery id, per route×target ✅/❌ outcome) to the group's logTarget when configured
  • Serve /gh slash commands + message context-menu commands + PR merge/close buttons (github.com repositories only; deferred replies edited via /webhooks/{application_id}/{token}/messages/@original) + comment modals (comments over 4000 characters are edited on GitHub instead); every interaction id runs once
  • Link chat accounts to GitHub only after the GitHub user confirms on a page naming both accounts (POST /auth/github/link); Telegram/Feishu /gh login answers only in private chats
  • Serve Telegram /gh commands (login/logout/comment/merge/close) via reply-message parsing
  • Serve Feishu /gh commands (login/logout/comment/merge/close) and card Merge/Close buttons via the /feishu/webhook endpoint (X-Lark-Signature verify + url_verification challenge)
  • Sync application commands from the scheduled trigger (global ~1h propagation + per-guild instant)
  • Sync the Telegram webhook URL from the scheduled trigger (setWebhook)

Message Format Spec

  • Every message title must start with the repo, then optional #number, then : subject: {repo}{#number}: {subject} (e.g. acme/widget#7: Add feature). Repo comes from payload.repository.full_name; fall back to t("common.repository") when missing.
  • Only the repo head is hyperlinked (never the whole title). Drivers split the title via splitMessageTitle (server/lib/formatters/helpers.ts): the Discord embed title is {repo}{#number} linked to the event's object URL (message.url, e.g. the issue/PR/comment html_url) and : {subject} renders as the first description line; Telegram keeps the one-line title with an inline repo link and a plain subject. Messages without a colon separator (a : followed by a space) keep the legacy whole-title link — use colon-free wording for such titles.
  • Do NOT use "Comment on org/repo" / "Review on org/repo" prefixes. Comments, reviews and inline comments use the same {repo}{#number}: {title} title as their parent object.
  • All event-specific emoji live in server/lib/formatters/ (via the emojiPrefix helper), never in the locale files. Emoji is controlled per group through the Group.emoji toggle (default true); showEmoji=false must leave out every emoji the formatters add to titles, descriptions, fields and links (text taken from the payload keeps its own emoji).
  • Milestone progress bars (🟢🟡🟠⬜) are data visualization and are exempt from the emoji toggle.
  • workflow_run messages render a Jobs field as a Discord diff fenced code block (server/lib/formatters/workflow.ts): one line per job, ${marker} ${name} ${emoji}${status} plus · {duration} for completed jobs. + (green) for success, - (red) for other non-pending conclusions, a plain space for running/queued/pending; job names clamp to 200 chars and the field to the platform field limit. Drivers without diff support (Telegram, Feishu) collapse it to inline code.
  • Commit hashes, branches and tags render as inline code wrapped in a hyperlink (commitLink/branchLink/tagLink helpers in server/lib/formatters/helpers.ts, e.g. [`abc123d`](https://.../commit/abc123def456), [`main`](https://.../tree/main)), falling back to plain inline code when the repo base URL is unavailable.
  • Content is clamped to the Discord embed limits (title 256, description 4096, field value 1024, 25 fields) both in the formatters and as a final safety net in the Discord render; the Telegram render caps the whole message at 4096 chars with tag-safe truncation. Commit subjects render only the first line, truncated to 200 chars.
  • Locale templates use a {emoji} placeholder immediately followed by the text (no space); the formatter injects em(...) which carries the trailing space.

Development

Package manager is bun — never use npm/npx (no package-lock.json, lockfile is bun.lock; CI runs bun install --frozen-lockfile).

bun install           # Install dependencies (updates bun.lock)
bun run dev           # Nuxt dev (HMR + Nitro dev server)
bun run build         # Production build (nuxt build, cloudflare_module preset)
bunx wrangler dev     # Miniflare preview of a built worker (bun run build first)
bun run typecheck     # Type checking (nuxt typecheck)
bun run lint          # ESLint
bun test              # Unit tests (under tests/)

Test suites beyond the per-module unit tests: provider fixtures (tests/fixtures/ + tests/provider-fixtures.test.ts), formatter snapshots (tests/formatter-snapshot.test.ts → tests/__snapshots__/), and platform contract tests (tests/platform-contract.test.ts, which assert the Discord/Telegram renderers clamp to their platform limits). CI (.github/workflows/ci.yml) runs bun install --frozen-lockfile + bun test + bun run lint; CodeQL (codeql.yml) and Dependabot (dependabot.yml) are configured under .github/.

Documentation

Keep every functional change in sync with the docs. After implementing a feature, fix, or refactor, update all of the following that are affected:

  • AGENTS.md (this file) — architecture tree, responsibilities, key decisions, config
  • README.md / README.zh.md — features, setup, configuration, supported events
  • docs/ (VitePress) — both docs/ (English) and docs/zh/ (Chinese) mirrors
  • config.example.yaml / .env.example — example config/secret files

Rule: no functional change ships without its documentation; docs and code must not drift.

Configuration

  • Local dev: .dev.vars (wrangler reads this for env bindings)
  • Production: wrangler secret put <NAME> for each secret
  • Routes: config in D1 d1_routes/d1_groups (authoritative), seeded from legacy KV config:routes/config:groups on first load; cache-only KV keys
  • KV namespace: Required binding for token/state/session/cache storage (dedup/delivery-state/message-tracking fall back to KV when D1 is unavailable)
  • D1 database: Binding DB (database webhooker; its id is set in wrangler.jsonc) for send_logs + audit_logs + discord_links + telegram_links + d1_groups + d1_routes + d1_fragments + dedup_keys + delivery_state + message_tracking tables
  • Queue: optional QUEUE producer binding plus consumers webhooker-delivery and its DLQ webhooker-delivery-dlq (declared in wrangler.jsonc); when absent, webhook dispatch stays inline
  • R2: optional PAYLOAD binding (bucket webhooker-payloads) for oversized queue payloads; without it, oversized payloads fall back to KV queue:payload:*
  • Access control: ADMIN_USER_IDS (super admins), DEBUG_MODE=true (local-only authentication bypass), ALLOW_SELF_SIGNUP (optional personal group on first login), AUDIT_RETENTION_DAYS (default 90) — all plain env vars, not secrets
  • Discord: DISCORD_PUBLIC_KEY (Interactions Endpoint signature verification, from Discord Developer Portal) and DISCORD_APPLICATION_ID (optional, auto-resolved via GET /oauth2/applications/@me when omitted) are required for interactions
  • Telegram: TELEGRAM_TOKEN (Bot API token from BotFather) required for Telegram routes; TELEGRAM_WEBHOOK_SECRET (optional secret token for POST /telegram/webhook verification — derived from TELEGRAM_TOKEN when unset, never disabled); avatars are sent as a link-preview card via the built-in GET /api/richheader (overridable with TELEGRAM_RICH_HEADER_HOST)
  • Feishu: FEISHU_APP_ID / FEISHU_APP_SECRET for sending; inbound commands/buttons need FEISHU_ENCRYPT_KEY and/or FEISHU_VERIFICATION_TOKEN
  • Webhook providers: GITEA_WEBHOOK_SECRET (required to receive Gitea webhooks; Gitea signs X-Gitea-Signature with the hex HMAC-SHA256 of the body)

Deployment

bunx wrangler secret put GITHUB_WEBHOOK_SECRET
bunx wrangler secret put DISCORD_TOKEN
bunx wrangler secret put DISCORD_PUBLIC_KEY
bunx wrangler kv namespace create KV
# Update wrangler.jsonc with KV ID
bunx wrangler d1 create webhooker
# Update wrangler.jsonc d1_databases with the database ID
bunx wrangler queues create webhooker-delivery
bunx wrangler queues create webhooker-delivery-dlq
# Queues are declared in wrangler.jsonc (QUEUE binding); no env var needed
bun run db:migrate:prod   # wrangler d1 migrations apply webhooker --remote (migrations/0001..0013)
# R2 bucket for oversized payloads (P0): bunx wrangler r2 bucket create webhooker-payloads
bun run deploy            # db:migrate:prod (pending migrations only) && wrangler deploy

Full list of secrets used: GITHUB_WEBHOOK_SECRET, GITEA_WEBHOOK_SECRET, GITHUB_APP_ID, GITHUB_PRIVATE_KEY (PKCS#8 PEM), GITHUB_CLIENT_ID, GITHUB_CLIENT_SECRET, DISCORD_TOKEN, DISCORD_PUBLIC_KEY, TELEGRAM_TOKEN, TELEGRAM_WEBHOOK_SECRET, FEISHU_APP_ID, FEISHU_APP_SECRET, FEISHU_ENCRYPT_KEY, FEISHU_VERIFICATION_TOKEN, ADMIN_USER_IDS, DEBUG_MODE, plus optional BASE_URL, DISCORD_APPLICATION_ID, TELEGRAM_RICH_HEADER_HOST, NUXT_PUBLIC_DOCS_URL, NUXT_PUBLIC_REPO_URL, NUXT_PUBLIC_LEGAL_CONTACT. See .env.example and docs/guide/configuration.md.

Notes

  • Commands sync from the scheduled trigger (*/5 * * * *): registered per-guild for instant availability (at most 30 API calls per run, resuming from a saved cursor) and globally (~1h propagation); both re-register when the command hash changes, the global set at least daily.
  • The bot is always offline (no Discord Gateway); interactions arrive via the HTTP endpoint.