Postgres rejects CHECK constraints that contain subqueries with
SQLSTATE 0A000 — `NOT EXISTS (SELECT 1 FROM jsonb_array_elements(...))`
inside CHECK is illegal even though it parses.
Move per-element validation entirely to the API layer (the upcoming
template-validators in PR 3). The DB CHECK keeps the cheap shape
guard: NULL or an array of length <= 10. Per-element validation
needs specific error messages anyway ("button #3 has no `text` field")
which API-layer validators do much better than a CHECK violation.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Switches message_templates.status from the local TitleCase set
(Draft|Pending|Approved|Rejected) to Meta's raw enum (DRAFT|PENDING|
APPROVED|REJECTED|PAUSED|DISABLED|IN_APPEAL|PENDING_DELETION) and
adds the columns the upcoming submit / edit / resubmit / webhook
flows need. No new user-facing behavior — this is groundwork.
Migration 014 also adds:
* sample_values JSONB Meta requires these for variable templates
* meta_template_id TEXT hsm_id for scoped edit/delete (avoids
nuking sibling-language variants)
* rejection_reason TEXT surfaced from webhook on REJECTED
* quality_score TEXT GREEN | YELLOW | RED, from webhook
* header_handle TEXT from Resumable Upload (media headers)
* header_media_url TEXT URL fallback for media headers (v1 path)
* submission_error TEXT last 4xx from Meta, for retry surface
* last_submitted_at rate-limit awareness (100 creates/hour)
* buttons JSONB shape CHECK permissive array-of-typed-objects guard
(strict per-type validation lives in API)
* UNIQUE (user_id, name, language)
so sync upsert is atomic and prevents
duplicate language variants
* idx on meta_template_id fast webhook lookup
Sync route now:
* persists Meta's status verbatim (PAUSED no longer collapses to
REJECTED — these are different states with different recoverability)
* parses BUTTONS components into the new discriminated-union shape
* extracts sample_values from body.example.body_text + header.example
* persists meta_template_id, quality_score, header_handle
Shared template-status.ts gives the manager / inbox picker / broadcast
picker one source of truth for labels + dark-theme badge classes.
Part of the Issue #147 overhaul plan (PR 2 of 6 in the sequence).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The webhook routes inbound messages by phone_number_id and uses
`.single()` to find the owning config. If two users save the same
phone_number_id, `.single()` errors PGRST116 ("multiple rows") and every
inbound message is silently dropped while the log misleadingly reports
"No config found".
Three layers of fix:
- POST /api/whatsapp/config now rejects with 409 when another user has
already claimed the phone_number_id. The check uses a service-role
client because RLS hides the conflicting row from the saving user.
- Webhook lookup no longer uses `.single()`. It distinguishes 0 rows
(true "no config") from ≥2 rows (duplicate claim) so logs name the
real cause and list conflicting user_ids.
- Migration 013 adds UNIQUE(phone_number_id) to whatsapp_config so the
same can't happen via direct DB writes or a race. Fails loudly with a
copy-pasteable resolution hint if duplicates already exist, rather
than auto-deduping (which would destroy tokens).
Closes#136.
startNewRun read flows.execution_count from a cached row and wrote
`cached + 1`. Two concurrent webhooks starting runs on the same flow
for different contacts both read N and both write N+1 — one count
permanently lost.
Mirrors the automations engine's solution (migration 007): a tiny
SECURITY DEFINER function that does the increment server-side, locked
to the service role. The engine's call becomes a single RPC.
The counter is best-effort: if the RPC fails we log + continue rather
than failing the run for a stat error.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Foundation for the upcoming WhatsApp chatbot flow builder. No runner,
no UI yet — just the plumbing that lets the codebase SEND interactive
messages (button replies + lists) and RECEIVE the taps as structured
message rows. Plan: ~/.claude/plans/understood-research-such-functionality-binary-lake.md
## What's in this PR
### Migration 010_flows.sql
- New tables: flows, flow_nodes, flow_runs, flow_run_events
- Linchpin partial unique index on flow_runs(user_id, contact_id) WHERE status='active' — at most one active run per contact, makes concurrent webhook deliveries idempotent at the DB layer
- Widens messages.content_type CHECK to allow 'interactive'
- New messages.interactive_reply_id column for the tapped button/row id
- RLS on every table; flow_runs added to supabase_realtime publication
- Idempotent throughout — matches the 009_message_actions.sql template
### Interactive Meta senders (src/lib/whatsapp/meta-api.ts)
- sendInteractiveButtons: type='interactive', interactive.type='button', up to 3 reply buttons
- sendInteractiveList: type='interactive', interactive.type='list', up to 10 rows across sections
- Limits exported as INTERACTIVE_LIMITS so flow validation can share the constants
- All Meta caps validated BEFORE the network call (button title ≤20 chars, list row title ≤24 chars, ≤10 total rows, no duplicate row ids, body ≤1024 chars, etc.) — misconfigured flows fail at save time instead of as a customer-facing 400 mid-conversation
### Webhook (src/app/api/whatsapp/webhook/route.ts)
- Extends WhatsAppMessage interface with `interactive?: { type, button_reply?, list_reply? }`
- parseMessageContent returns a new `interactiveReplyId` field; refactored to use a shared `empty` base so every return site picks up the new field without 9 explicit copies
- New `case 'interactive'`: uses the tapped option's TITLE as content_text (so the inbox bubble reads "Existing customer", not a UUID) and stashes the id separately
- 'interactive' added to ALLOWED_CONTENT_TYPES so the messages insert no longer coerces it to 'text'
### Inbox bubble (src/components/inbox/message-bubble.tsx)
- New `case "interactive"` in MessageContent renders the tap with a small "↩ Button reply" affordance so agents reading the inbox can tell at a glance that this was a tap rather than the customer typing the same words
### Types (src/types/index.ts)
- ContentType union extended with 'interactive'
- Message.interactive_reply_id field
### Engine helpers (src/lib/flows/)
- admin-client.ts: thin service-role client wrapper, mirrors automations/admin-client.ts
- meta-send.ts: engineSendInteractiveButtons / engineSendInteractiveList — adapts the Meta senders for service-role use, persists messages with sender_type='bot' content_type='interactive', uses the same phone-variant retry as automations/meta-send.ts. Unused in PR #1; ready for the runner in PR #2
### Tests
- 13 new vitest cases in meta-api.test.ts covering Meta's limit validation + payload-shape assertions for both interactive senders
## Validation
- npm test → 102/102 (was 89; 13 new)
- npm run lint → 0 errors
- npm run typecheck → clean
- npm run build → clean
## What's NOT in this PR
- The runner (dispatchInboundToFlows) — PR #2
- Builder UI / API surface — PR #3
- collect_input / condition / set_tag nodes + templates — PR #4
- react-flow canvas + http_fetch — v2 (PR #5)
## Risks tracked from the plan
- messages.content_type CHECK widening: grep'd every consumer; message-bubble.tsx, reply-quote.tsx, send/route.ts, automations/meta-send.ts all handle the new variant correctly (or accept it as a defaulted fallthrough)
- engineSendText / engineSendInteractive share ~80% of code; deliberately not pre-factored yet — let both stabilize, extract a shared base in PR #2 once the shape settles
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Scaffolding for the upcoming WhatsApp Flows feature. The whole Flows
stack (interactive button/list send + parse, server-side runner, no-
code builder, templates) will land across the next ~4 PRs *behind*
this gate so the owner can dogfood the feature on their own account
before exposing it to teammates or downstream pullers of the OSS.
What this lands:
1. `profiles.beta_features TEXT[] NOT NULL DEFAULT ARRAY[]::TEXT[]`
— generic opt-in column, lives on the existing profiles table.
Empty by default for every existing row.
2. `src/lib/flows/feature-flag.ts` — `isFlowsEnabled(profile)`
predicate + `BETA_FEATURE_FLOWS` constant. Single source of
truth for the gate; later PRs import this in every guarded
surface (sidebar nav, /flows page, /api/flows routes, webhook
dispatcher).
3. `Profile.beta_features` added to the shared TS type and the
auth hook's local copy. Auth hook loads the column and
defaults it to `[]` defensively if the DB read returns null.
4. 11 unit tests covering null/undefined/empty/mismatched cases,
including case sensitivity (only lowercase `'flows'` opts in)
and defensive guards against malformed input.
How the owner enables it for their own account, once merged:
UPDATE profiles
SET beta_features = array_append(beta_features, 'flows')
WHERE user_id = '<their-auth-uid>';
How we ship it to everyone, when ready: grep-and-delete the
`isFlowsEnabled()` callsites in one cleanup PR.
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>
Adds a profile tab to /settings with the five things every B2B SaaS
user expects:
- Display name — edits profiles.full_name.
- Avatar — upload / replace / remove, stored in a new `avatars`
Supabase Storage bucket with RLS scoped to the owner's folder.
Migration 008_profile_avatars_storage.sql creates the bucket (public
read, 2 MB cap, image mimes only) and the four storage.objects
policies.
- Email — goes through supabase.auth.updateUser({ email }), which
confirms to both the old and new addresses. UI surfaces a "check
both inboxes" banner after save.
- Password — three-field form with a current-password gate implemented
via a throwaway signInWithPassword() before calling updateUser.
- Active sessions — one-click "sign out of all devices" using
supabase.auth.signOut({ scope: 'global' }), with a confirm dialog.
Plus read-only role, joined date, and user id for operator debugging.
The header avatar and sidebar user row are now both dropdown menus
linking to /settings?tab=profile and /settings?tab=whatsapp; the
settings page reads ?tab= to land on the right tab. useAuth gets a
refreshProfile() callback so the header/sidebar update immediately
after a save without a full reload.
Deferred (separate work): account deletion, 2FA, phone/timezone.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The engine did a client-side read-modify-write on
automations.execution_count, so two concurrent dispatches of the
same automation (e.g. fired for two different contacts within the
same second) could both read N and both write N+1, permanently
losing one count.
- Migration 007 adds increment_automation_execution_count(p_automation_id)
as a SECURITY DEFINER sql function. Only service_role can EXECUTE
it — an authenticated user can't juice someone else's counter by
calling the RPC directly.
- Engine now calls this RPC instead of reading automation.execution_count
and adding 1. Atomic in one statement; concurrent dispatches both
land correctly.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces the /automations placeholder with a full feature:
- Migration 006 adds automations, automation_steps, automation_logs,
and automation_pending_executions with RLS matching existing tables.
- List page with active/inactive switch, quick-start templates,
duplicate, delete, and View Logs actions.
- Full-screen vertical flow builder with Trigger + Action + Condition
(Yes/No branches) + Wait cards. Inline edit forms, up/down reorder,
dotted-grid canvas.
- Execution engine (src/lib/automations/engine.ts) dispatched from the
WhatsApp webhook for new_message_received, new_contact_created, and
keyword_match. Wait steps enqueue to automation_pending_executions;
/api/automations/cron drains them behind AUTOMATION_CRON_SECRET.
- Logs viewer with per-step checkmarks and expandable detail.
Existing files touched only in webhook/route.ts (fire-and-forget
trigger dispatch) and types/index.ts (appended automation types).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Fixes 8 real issues found during a thorough review of the broadcast
feature (PRs #21-#29).
Security
* Webhook POSTs had zero spoof protection. Anyone could forge a
status update and drift broadcast counts arbitrarily. Added
HMAC-SHA256 signature verification via the X-Hub-Signature-256
header (Meta's standard). New helper src/lib/whatsapp/webhook-
signature.ts with timing-safe comparison. Rollout fails open
with a warning when META_APP_SECRET is unset so existing
deploys don't break at rollout; set the env var to enforce.
Correctness
* CSV audience was completely broken — synthesized "csv-N" strings
as contact_id, which failed the UUID cast on insert. Every CSV
broadcast silently created zero recipients. Replaced with a
real find-or-create against the contacts table (batched lookup
by phone, single INSERT for missing rows, de-duplicates CSV
inputs). CSV contacts now end up as real contacts.rows with
real UUIDs flowing through the rest of the pipeline.
* Recipient insert errors were logged with console.error and the
loop marched on — broadcasts then ran with incomplete recipient
sets, webhook updates couldn't find rows, and aggregate counts
drifted. Now: on batch insert failure, flip the broadcast to
failed + failed_count=total + throw so handleSend's toast fires.
* Webhook status guard ordered "failed" above "replied", which
meant a webhook replay delivering "failed" for an already-
replied recipient would upgrade replied → failed. Split the
ladder: forward-only ordering across pending → sent →
delivered → read → replied; failed is a terminal side branch
that can only be reached from pending or sent (never from any
success state). Net effect: a message that was read can no
longer regress to failed on a replay.
* Template sync only pulled page 1 (ignored paging.next). Users
with >100 templates silently missed some, and broadcasts with
those names failed with Meta's "template not found". Now
follows paging.next with a 20-page safety cap, reports
truncated=true in the response if hit.
* Template sync reported success:true even when individual
upserts failed. Now accumulates per-template errors and
surfaces them via a separate toast.error in the UI — green
toast no longer hides silent drift.
Performance
* Migration 005: replace the broadcast_recipients aggregate
trigger installed in migration 003. Old version ran
COUNT(*) FILTER on every row change — O(n²) across a send
loop (10k recipients × 10k index scans). New version
increments / decrements by ±1 per status transition: O(1)
per change. Kept the old recompute_broadcast_counts() as a
callable safety-net in case of drift.
* List page poll kept running while the tab was hidden. Added
visibilitychange handler — polling pauses when hidden,
resumes (and does an immediate fetch) on refocus.
UX
* Step 3 (Personalize) let users skip past unmapped placeholders
→ broadcasts then sent to Meta with empty-string params.
Next is now disabled until every {{N}} has a value; an amber
warning lists the missing placeholders.
Documented accepted risks in the PR body.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Root cause (error 23503):
broadcast_recipients.contact_id and deals.contact_id were declared
NOT NULL REFERENCES contacts(id) with no ON DELETE clause, so
Postgres defaulted to NO ACTION. Any contact who'd ever received a
broadcast or been attached to a deal became un-deletable with
"violates foreign key constraint broadcast_recipients_contact_id_fkey".
Fix (migration 004):
Relax both FKs to ON DELETE SET NULL + drop NOT NULL on both
columns. History rows survive with contact_id = NULL. The UI is
already null-safe:
- broadcast detail: contact?.name ?? 'Unknown'
- deal card: contact?.name || contact?.phone || 'No contact'
- CSV export: contact?.name ?? ''
Why not CASCADE:
CASCADE would wipe broadcast recipient rows and trigger the
aggregate recount on broadcasts (migration 003), retroactively
changing sent_count / delivered_count / read_count. It would also
delete deal history. SET NULL preserves both counts and history.
Type/code:
- Deal.contact_id + BroadcastRecipient.contact_id widen to
string | null
- DealForm: setContactId(deal.contact_id ?? "") — the contact
picker uses "" to mean unselected
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three correctness bugs were making the existing broadcast flow wrong,
not just incomplete.
1. Per-recipient variables were broken
use-broadcast-sending.ts computed templateParams per contact but shipped
only templateParams[0] to the API, so every contact in a batch received
contact 0's personalization (name/email/etc).
Fix: extend POST /api/whatsapp/broadcast to accept a per-recipient
shape `{ recipients: [{ phone, params }], ... }`. Legacy
`{ phone_numbers, template_params }` still accepted. Hook sends per-
recipient payloads; API returns whatsapp_message_id per phone.
2. Webhook never updated broadcast_recipients
handleStatusUpdate in webhook/route.ts only mirrored Meta status onto
the `messages` table. broadcast_recipients had no correlation column,
so delivered_count / read_count / replied_count never advanced for
real broadcasts.
Fix (migration 003):
- broadcast_recipients.whatsapp_message_id (+ UNIQUE index) for
Meta-message-id correlation
- composite index (broadcast_id, status) for fast aggregate scans
- AFTER INSERT/UPDATE/DELETE trigger that recomputes parent
broadcasts counts via COUNT(*) FILTER — sender + webhook only
touch the recipient row, aggregates stay consistent
Webhook:
- handleStatusUpdate also updates broadcast_recipients with forward-
only status guard (read cannot regress to sent on a replay)
- processMessage flags the replied recipient row when an inbound
message arrives from a recent broadcast contact
3. Rate limiting did not match spec
Batch 50 / no delay → Batch 10 / 1s between batches, matching the spec
and Meta's per-phone-number rate ceiling.
Also: extract statusConfig from both page components into
src/lib/broadcast-status.ts (single source of truth for PR B). Add
whatsapp_message_id to BroadcastRecipient type.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Rebuild the Pipelines page to match spec.
Highlights:
- Drag-and-drop deal cards between stages via @dnd-kit (200ms drop animation,
emerald dashed drop zone, opacity fade while dragging)
- First-visit seed: auto-creates "Sales Pipeline" with spec-defined stages
(New Lead / Qualified / Proposal Sent / Negotiation / Won)
- Top bar: always-visible pipeline dropdown with "Manage Pipelines" option,
plus "Add Pipeline" and emerald "Add Deal" buttons
- Deal form converted from Dialog to right-side Sheet, adds assigned_to
(profile FK), "Link to Conversation" shortcut, "Mark as Won/Lost" status
buttons, and "Delete Deal"
- Pipeline settings modal: drag-to-reorder stages (@dnd-kit sortable),
per-stage color picker, "Create new pipeline" shortcut, archive
confirmation copy
- Analytics bar below board: total deals, pipeline value, avg deal size,
weighted pipeline value (probability interpolated across stages),
won/lost this month
- Contact detail sheet gains a "Deals" tab showing all deals for that
contact (value + stage badge + status)
- Migration 002: adds deals.assigned_to FK→profiles, normalizes status
CHECK to ('open','won','lost')
Note: spec called for light-gray Kanban columns (#f9fafb) + white cards
inside an otherwise-dark app. Followed spec literally; easy to retheme.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The Supabase GitHub integration re-runs migrations on every push.
The original 001_initial_schema.sql used plain CREATE TABLE without
IF NOT EXISTS, so it failed with "relation 'profiles' already exists"
after the first successful run.
Changes:
- CREATE TABLE -> CREATE TABLE IF NOT EXISTS (all 16 tables)
- CREATE INDEX -> CREATE INDEX IF NOT EXISTS
- DROP POLICY IF EXISTS before each CREATE POLICY
(Postgres has no CREATE POLICY IF NOT EXISTS)
- DROP TRIGGER IF EXISTS before each CREATE TRIGGER
- Wrapped ALTER PUBLICATION in DO block that checks pg_publication_tables
- handle_new_user() trigger now uses SECURITY DEFINER, SET search_path,
EXCEPTION handler, and owner=postgres to properly bypass RLS on the
profiles INSERT during signup (fixes "Database error saving new user")
The migration is now safe to re-run any number of times.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
