Makes the AI auto-reply bot visible and controllable from the inbox,
completes the handoff, and records token spend. Builds on the existing
lib/ai engine rather than replacing it.
Track A — user-facing
- messages.ai_generated flag → small "AI" badge on bot-generated bubbles
- "Take over" / "Resume AI" banner per conversation
(POST /api/ai/autoreply/[id]) — take-over pauses + assigns to the agent
- handoff now assigns to a configurable target (agent or unassigned
queue) and writes an internal note summarizing the exchange; the
existing on_conversation_assigned trigger notifies the assignee
- Settings → AI: "Hand off to" target picker
Track B — observability + safety
- providers return normalized token usage; logged per run to ai_usage_log
(admin-readable) from both auto-reply and draft
- account-wide auto-reply rate limit (30/min) on top of the
per-conversation cap
Migration: 033_ai_reply_polish.sql (messages.ai_generated,
ai_configs.handoff_agent_id, conversations.ai_handoff_summary,
ai_usage_log). Tests + typecheck + lint pass (607 tests).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds an AI reply assistant powered by the account's own OpenAI or
Anthropic key (stored AES-256-GCM-encrypted; no per-seat AI fee).
- AI-drafted replies in the inbox composer (✨ button, agent+), via
read-only POST /api/ai/draft.
- Optional auto-reply bot on inbound messages that no Flow consumed and
no agent owns, with a per-conversation cap and clean [[HANDOFF]] to a
human. Stands down when a message-level automation would also reply.
- Settings → AI Assistant (admin+): provider/model, key, business
context, toggles, and a Test-key button that verifies before saving.
- Provider layer (src/lib/ai) with OpenAI + Anthropic adapters behind
one interface; migration 029 (ai_configs + per-conversation columns +
atomic claim_ai_reply_slot RPC).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds the machine-to-machine auth layer the public REST API needs so
external scripts and automations can drive wacrm (issue #245). This is
the groundwork release; data endpoints land one at a time on top of it.
- Migration 026: account-scoped `api_keys` table. SHA-256 hash at rest
(plaintext shown once), display `key_prefix`, `scopes[]`, RLS via
`is_account_member` (members read, admin+ manage).
- `src/lib/api-keys/`: key gen/hash (wacrm_live_ prefix, timing-safe),
scope vocabulary, service-role auth-path store.
- `requireApiKey(request, scope?)` (src/lib/auth/api-context.ts):
bearer parse -> hash lookup -> liveness -> per-key rate limit ->
scopes-only authz. Returns a service-role-scoped account context.
- `/api/v1` envelope ({data} / {error:{code,message}}) + GET /api/v1/me
reference endpoint that exercises the whole stack.
- Dashboard: admin+ key management routes + Settings -> API keys tab.
- Docs (docs/public-api.md), CHANGELOG, README, .env note.
Checks: typecheck clean, lint 0 errors, full suite green (+27 tests).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Follow-up to #259. Voice notes are encoded to Ogg/Opus in the browser via
opus-recorder (small WASM encoder, lazy-loaded on record), so there's no
server-side transcode and nothing to provision on the host. The server
stores the .ogg like any other attachment; WhatsApp renders it as a voice
note.
- Composer: MediaRecorder + /api/whatsapp/transcode-audio replaced with
opus-recorder. Encoder worker vendored at public/opus/encoderWorker.min.js
(no install-time download to fail), ignored by eslint + prettier.
- Removed the transcode route, the audio-transcode lib + tests, the
ffmpeg-static dep + ambient types, the FFMPEG_PATH env doc, and the
unused transcodeAudio rate limit.
- Added opus-recorder ambient types.
Still behind the 'chat_media' beta flag with per-kind size caps.
typecheck/lint/build green; 414 tests pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Agents can now attach a photo/video/document or record a voice note in
the inbox composer and send it live in a 1:1 thread — previously media
could only be received from customers or sent via the Flows builder.
- 023_chat_media.sql: new account-scoped `chat-media` Storage bucket
(mirrors flow-media RLS from 020) plus Meta-accepted audio MIME types.
- meta-api: sendMediaMessage gains an `audio` kind (link-only — Meta
rejects caption/filename on audio; WhatsApp renders OGG/Opus as a
voice note).
- /api/whatsapp/send: media branch reusing the phone-variant retry loop.
- /api/whatsapp/transcode-audio: Node-runtime route remuxing the
WebM/Opus that Chromium's MediaRecorder produces to OGG/Opus via
ffmpeg-static (Firefox/Safari record accepted formats and skip it).
- composer: attach menu, file pickers, voice recorder + auto-transcode,
preview with caption; thread: optimistic handleSendMedia.
- shared uploadAccountMedia helper (flows form refactored onto it).
- next.config: allow microphone=(self) (was fully blocked) + CSP
media-src; document optional FFMPEG_PATH override.
Tests: meta-api audio payload shape, buildMediaPath, audio-transcode
helpers (14 new). typecheck/lint/build green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Addresses the L1 finding from the final security review. The six
admin-class write endpoints under /api/account had no rate
limits, so a compromised admin session (or a script run in a
loop) could spam invite creation / role flips / member removes
fast enough to flood the DB or audit logs.
Adds a new `RATE_LIMITS.adminAction` bucket — 30/min/user — and
wires it into:
- POST /api/account/invitations → admin:inviteCreate
- DELETE /api/account/invitations/[id] → admin:inviteRevoke
- PATCH /api/account/members/[userId] → admin:memberRole
- DELETE /api/account/members/[userId] → admin:memberRemove
- POST /api/account/transfer-ownership → admin:transferOwnership
- PATCH /api/account → admin:rename
Each endpoint keys its own bucket so spamming one doesn't starve
another. 30/min is comfortably above any realistic human pace
(the Members tab is a clicks-only UI; a legit admin would
struggle to hit ten of these per minute by hand) while still
bounding accidental abuse.
The bucket joins the existing send / broadcast / react /
invitationPeek / invitationRedeem buckets in `RATE_LIMITS` —
same in-memory fixed-window limiter, same `rateLimitResponse`
helper for the 429 shape.
Severity in the final review: LOW. The admin-trust model means
this isn't blocking — but the cost of adding it is six
near-identical 4-line additions and zero net risk, so worth
landing for hardening.
Verification
- `npm run lint` → 0 errors (same 19 pre-existing warnings)
- `npm run typecheck` → clean
- `npm test` → 366/366 pass
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PR 4 of the multi-user accounts series. Lands the link-only
invitation flow — admin generates a shareable URL, copies it,
sends it via any channel. No email service needed.
API surface
- GET /api/account/invitations — list pending (admin+)
- POST /api/account/invitations — create (admin+)
- DELETE /api/account/invitations/[id] — revoke (admin+)
- GET /api/invitations/[token]/peek — public, rate-limited per-IP
- POST /api/invitations/[token]/redeem — authenticated, rate-limited per-IP
Token lifecycle
- POST returns the plaintext token + share URL exactly ONCE in
the response. We store only the SHA-256 hash. There is no
endpoint that resurfaces it; if the admin dismisses the dialog
without copying, the only recourse is to revoke and re-issue.
- Peek and redeem take the plaintext token in the URL path, hash
it in TS via `hashInviteToken`, and look up by `token_hash`.
Two new SECURITY DEFINER RPCs back the cross-RLS work
(migration 019):
- `peek_invitation` — anonymous read. Returns a uniform JSON
envelope: `{ ok, account_name, role, expires_at }` or
`{ ok: false, reason: not_found|used|expired }`. The route
forwards verbatim so the join page can branch cleanly.
- `redeem_invitation` — authenticated. Locks the invite row
`FOR UPDATE`, validates expiry/usage, refuses if the caller is
already in a shared account OR their personal account has any
domain rows (checked across all 11 user-creatable parent
tables), then atomically moves profile.account_id +
account_role, stamps the invitation accepted, and deletes the
orphan personal account. Single transaction.
Error contract
- 22023 (invalid_parameter_value) → 400 — not_found / used / expired
- 42501 (insufficient_privilege) → 401 — caller not authenticated
- 23505 (unique_violation, reused) → 409 — caller's account has data,
or they're already a
member of the inviter's
account / another shared
account
Rate limiting
- Two new buckets in `RATE_LIMITS`: invitationPeek (30/min/IP)
and invitationRedeem (10/min/IP). Belt-and-braces against
token enumeration even though 256-bit entropy makes it
theoretical. Uses the existing fixed-window in-memory limiter.
Why no peek-then-redeem two-step on the server
- Peek is read-only and idempotent; redeem locks `FOR UPDATE` so
two concurrent redemptions can't both win. The peek result is
cosmetic ("show me what I'm joining"); the actual write
happens entirely inside the redeem RPC's transaction, so a
TOCTOU between peek and redeem just means the user sees the
same `{ ok: false, reason }` on the next render.
Verification
- `npm run lint` → 0 errors (same 20 pre-existing warnings)
- `npm run typecheck` → clean
- `npm test` → 366/366 pass
- `npm run build` → succeeds; the 4 new routes register as ƒ
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds three WhatsApp-standard affordances to the inbox: per-message
emoji reactions, reply-with-quote, and copy-text. Bubbles stay pure
presenters — actions layer on by composition (MessageActions wraps
MessageBubble). Hover toolbar on desktop, long-press (onContextMenu)
on touch.
Data model (migration 009):
- messages.reply_to_message_id (self-FK, ON DELETE SET NULL) so replies
survive a deleted parent
- message_reactions table — one row per (message, actor) with a unique
constraint, conversation_id denormalised so Supabase Realtime can
filter without a join
- RLS scoped via conversation ownership; service-role bypass for the
webhook
API:
- /api/whatsapp/send accepts reply_to_message_id, resolves the parent's
Meta message_id and forwards it as context.message_id
- /api/whatsapp/react (new) sends Meta reaction payloads and mirrors
the upsert/delete into message_reactions
- Webhook stops storing inbound reactions as fake text messages and
instead upserts/deletes message_reactions; inbound replies (context.id)
resolve to internal UUID and populate reply_to_message_id
UI:
- MessageActions hover toolbar with inline 6-emoji Popover (static set;
no emoji-picker library)
- ReplyQuote shared between the in-bubble quote and the composer chip
- MessageReactions pill row with click-to-toggle
- Thread subscribes to message_reactions filtered by conversation_id;
optimistic upsert with rollback on failure
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
A fork running WaCRM publicly had no server-side guard on the
WhatsApp send endpoints — a compromised session or an aggressive UI
loop could call /api/whatsapp/send or /broadcast as fast as the
network allowed and blow through Meta's per-number quota (or worse,
get the number suspended).
This PR adds a minimal in-memory fixed-window rate limiter keyed by
user id.
Budgets (tuned for a template default — forkers can edit):
- /api/whatsapp/send 60 per minute per user (1/s sustained)
- /api/whatsapp/broadcast 5 per minute per user (campaign dispatch
rate, not fan-out)
Behaviour:
- Check runs right after the Supabase auth check; unauthenticated
calls 401 as before and never consume a bucket entry.
- When a user exceeds the budget the route returns 429 with
Retry-After, X-RateLimit-Limit, X-RateLimit-Remaining,
X-RateLimit-Reset headers.
- Send and broadcast buckets are independent; two users are
independent.
Design trade-off:
The limiter lives in a process-local Map. On a single Hostinger
Managed Node.js instance — the template's default deployment —
that's the whole limit. On a horizontally-scaled deploy (multiple
regions, serverless fan-out) each instance has its own Map and the
effective rate is N * limit. Swap src/lib/rate-limit.ts for a
Redis/Upstash/Durable-Objects backend if you scale past one
instance; keep the checkRateLimit return shape and the call sites
don't move.
Memory is self-cleaning: a 1-in-1000 opportunistic sweep drops
expired keys without a background timer, so the Map stays bounded
without running anything outside the request path.
Smoke test (tsx):
- 7 broadcasts → 5 ok, 2 denied.
- Different user → independent bucket.
- Same user, different route → independent bucket.
- Retry-after = 60s.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
