The security PR (#352) shipped 031_fix_profiles_update_rls.sql while
031_ai_reply_slot_grant.sql (#345) already held version 031. Supabase
keys schema_migrations on the numeric prefix, so applying both to a fresh
DB fails with `duplicate key ... (version)=(031)` (SQLSTATE 23505) — which
is what broke the Supabase Preview check on this PR.
Rename the newer, colliding file to 034 to restore unique versions. It's
idempotent and independent of the AI tables, so re-sequencing after 033 is
safe; a prod DB that already applied 031 will apply this fresh as 034
without side effects.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Correctness:
- Resume AI now releases ANY assignment (not just the caller's own).
A handoff-assigned thread previously stayed assigned after resume, so
the eligibility gate silently kept the bot muted — Resume was a no-op.
- GET /api/ai/usage: guard days >= 1 (Number(null)===0 is finite, so the
30-day default was unreachable and the window collapsed to 1 day).
- GET /api/ai/usage: align the query cutoff to the start of the oldest
charted local day so rows in the oldest partial day no longer count in
totals while missing from the daily series (totals now match the bars).
- ai-thread-banner: cache account AI-status per accountId and never cache
a transient fetch failure — fixes cross-account leak and the banner
staying hidden all session after one failed /api/ai/config call.
- /api/ai/draft: construct the service-role client inside a guard and
fire-and-forget the usage log, so a missing service-role key (or any
log failure) can't 500 an already-generated, already-billed draft.
Efficiency:
- Don't await usage logging on the reply hot path (auto-reply + draft);
logAiUsage swallows its own errors, so fire-and-forget is safe and
removes a DB round-trip from the customer-facing send.
Reuse / consistency:
- usage route reuses date-utils (daysAgoStart/lastNDayKeys/localDayKey)
and now buckets by LOCAL day like every other dashboard chart (was UTC).
- extract formatCompactNumber (shared with formatCurrencyShort); drop the
hand-rolled month table in favour of date-fns; extract
fetchAccountMembers/memberLabel and use them in the AI settings picker.
Resume still resets the per-conversation reply count (a deliberate,
manual, rate-limited re-engagement — documented in the handler).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Surfaces the ai_usage_log data added in the previous commit:
- GET /api/ai/usage — admin-only aggregation (totals, per-mode,
per-model, zero-filled daily series) over a 7/30/90-day window,
mirroring the ai_usage_log SELECT policy
- AiUsageCard — stat tiles + daily token bar chart + by-model list,
rendered under a new admin-only "Usage" tab on the AI Agents page
Counts only; no message content is stored or shown.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
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>
The send_webhook automation step fetched an account-controlled URL with
attacker-controlled headers and no SSRF protection, following redirects
by default — letting an agent+ reach cloud metadata (169.254.169.254),
loopback, and RFC1918 hosts from the server's network (status code
surfaces in automation logs). The separate webhook_endpoints delivery
path already guards this; the automation path did not.
Route send_webhook through the existing isDeliverableUrl guard before the
fetch, and mirror the deliver.ts hardening: redirect: 'manual' (a public
URL can't 3xx-bounce to an internal one) and a 10s timeout. Add an engine
regression test asserting a link-local target never reaches fetch.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Flows and automations write routes verify only authentication/membership
(via the RLS client, whose *_select policies have no minimum role) and
then mutate with the service-role client, which bypasses the agent-gated
*_insert/update/delete RLS policies. A read-only `viewer` therefore had
full write power.
H3 (High): add requireRole('agent') to the flows write handlers —
PUT/DELETE (flows/[id]), activate (flows/[id]/activate), and create
(flows POST). GET/list stay open to viewers.
M2 (Medium): same gate on automations POST, the engine trigger
(automations/engine, which sends outbound WhatsApp), and the [id]
mutation handlers (PATCH/DELETE/duplicate). The cron route is unchanged —
it is gated by the x-cron-secret shared secret, not user auth.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
C1 (Critical): the profiles_update RLS policy gates only on user_id,
leaving account_role/account_id — the source of truth for
is_account_member() — freely writable by the row's own user. A
browser-client UPDATE could self-promote a viewer to owner or move
into another tenant. Add a BEFORE UPDATE trigger that rejects changes
to those columns when current_user = 'authenticated'; the SECURITY
DEFINER member/invitation RPCs (postgres) and the service_role backend
are unaffected.
H2 (High): match_ai_knowledge_fts/_semantic were SECURITY DEFINER,
filtered only on the caller-supplied p_account_id with no membership
check, and were granted to authenticated — letting any logged-in user
read another account's knowledge base via a foreign p_account_id.
Recreate both as SECURITY INVOKER so the existing
ai_knowledge_chunks_select RLS policy applies; the service_role
auto-reply bot still bypasses RLS.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Clicking "Add node" in the flow builder crashed the whole page to a
"This page couldn't load" error boundary, in both the list and canvas
views.
The add-node menus render a DropdownMenuLabel per category. That is
base-ui's Menu.GroupLabel, which calls useMenuGroupRootContext() during
render and THROWS ("MenuGroupContext is missing") when it has no
Menu.Group ancestor. Both menus wrapped each category in a plain <div>
instead of a DropdownMenuGroup, so the label had no group context. The
menu content only mounts when the menu opens, which is why initial
render was fine and the crash fired exactly on click.
Wrap each category's label + items in DropdownMenuGroup (base-ui
Menu.Group), which provides the context Menu.GroupLabel needs. The
between-group separator moves into a keyed Fragment so it stays a
sibling of the groups.
Adds a regression test pinning the contract: a bare DropdownMenuLabel
throws, a group-wrapped one does not.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Migration 029 created claim_ai_reply_slot(uuid, integer) as SECURITY
DEFINER but never GRANTed EXECUTE — the only function in the schema
missing its grant (007/012/017/018/019/025/030 all grant theirs).
SECURITY DEFINER sets the privileges a function runs *with*, not who may
*call* it: the caller still needs EXECUTE. On instances where the default
PUBLIC execute privilege on public-schema functions has been revoked
(hardened / self-hosted Supabase), the service_role that PostgREST uses
cannot invoke it. The AI auto-reply path runs entirely under the
service-role client (the inbound webhook has no auth.uid()), so
db.rpc('claim_ai_reply_slot') fails permission-denied, the dispatch bails
before sending, and the bot silently never answers ANY inbound message.
The Playground works because it never claims a slot — matching the report.
- 031_ai_reply_slot_grant.sql: new migration granting EXECUTE to
service_role (029 is already applied on existing installs, so it needs
its own migration to reach them). service_role only — the sole caller —
matching the 007/012 increment-counter precedent.
- 029_ai_reply.sql: add the same grant next to the definition so fresh
installs are correct.
- auto-reply.ts: log a real claim RPC error instead of returning
silently, so this class of deploy failure is diagnosable.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Promotes the AI assistant out of Settings into a first-class "AI Agents"
sidebar area.
- Sidebar: new "AI Agents" item → /agents.
- /agents page with two tabs:
- Playground — a test chat (POST /api/ai/playground) that runs the
real auto-reply path (KB retrieval + provider, auto_reply prompt,
multi-turn, shows handoff). loadAiConfig gains requireActive so you
can test before enabling.
- Setup — the existing provider/key + knowledge-base config, relocated.
- Removed the AI section from Settings (settings-sections + page); same
account config, new home. No migration.
Tests +6 (query, loadAiConfig requireActive). 593 pass; typecheck, lint,
build green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Security: REVOKE ALL FROM PUBLIC + GRANT EXECUTE (authenticated,
service_role) on the match RPCs so anon can't read a KB by supplying
an account_id (mirrors migrations 018/025).
- Ingest: store chunks even when embedding fails (insert first, rethrow
after) so a doc stays lexically searchable; the "lexical still works"
warning is now truthful.
- Retrieval: short-circuit on an empty KB (cheap COUNT) to avoid a paid
query embedding + 2 RPCs on every draft/auto-reply.
- Migration: HNSW instead of IVFFlat (robust on small/growing tables);
language-neutral 'simple' FTS config (was 'english'); semantic RPC
takes text and casts ::vector(1536) to remove PostgREST bind ambiguity.
- Embeddings key: distinguish corrupt-from-absent (loadEmbeddingsKey
returns {key,corrupt}); log on decrypt failure; reindex now reports
failure instead of false success; POST/PATCH warn on keyword-only save.
- embedTexts: fail loud when the provider omits result indices (was
silently misaligning chunks with vectors).
Tests: +3 (empty-KB guard, embed-failure-still-indexes, missing-index).
587 pass; typecheck/lint/build green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Grounds the AI assistant (0.5.0) in the account's own content so it can
answer instead of handing off.
- Hybrid retrieval: Postgres full-text search for every account; optional
semantic search (pgvector + OpenAI text-embedding-3-small) when an
embeddings key is set. Semantic-primary, topped up with FTS.
- src/lib/ai: chunk, embeddings, knowledge (ingest + retrieve), query;
buildSystemPrompt gains a knowledge block; config carries an
(optional) embeddings key.
- Wired into /api/ai/draft and the auto-reply bot (best-effort — never
blocks a reply).
- KB API (/api/ai/knowledge CRUD + /reindex) and a Settings → AI
Assistant knowledge-base manager + embeddings-key field.
- Migration 030: pgvector, ai_knowledge_documents, ai_knowledge_chunks
(fts + embedding), match RPCs, ai_configs.embeddings_api_key.
Migration required: supabase/migrations/030_ai_knowledge.sql.
Stacked on feat/ai-reply-assistant (#329).
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>
This introduces the base i18n setup using next-intl without routing. It adds the default messages/en.json and sets up the app layout and next.config.ts to consume the locale from the environment. The login page has been migrated as a first example.
Completes the public API roadmap (#245): register HTTPS endpoints to be
POSTed to when things happen in an account, so automations can react to
inbound activity instead of polling.
- Migration 028: account-scoped webhook_endpoints (secret stored
AES-GCM-encrypted) + an atomic record_webhook_failure() function.
- Scope webhooks:manage; management API GET/POST /api/v1/webhooks and
GET/PATCH/DELETE /api/v1/webhooks/{id} (secret returned once).
- Events message.received / conversation.created / message.status_updated
dispatched from the inbound webhook handler's after() block.
- Signed delivery (X-Wacrm-Signature = HMAC-SHA256 over t.body), single
best-effort attempt, auto-disable after consecutive failures.
- Versioned 0.4.0; docs/public-api.md Webhooks section; CHANGELOG.
Fixes from the adversarial code review of this diff:
- SSRF: block private/loopback/link-local (incl. cloud metadata) targets
at delivery time + redirect:'manual' so a public URL can't bounce
internal (src/lib/webhooks/ssrf.ts).
- Status mirror UPDATE no longer uses .select().maybeSingle() (message_id
is non-unique); account resolved via a bounded limit(1) read.
- Atomic failure counter via SQL RPC (was a racy read-modify-write).
- conversation.created emitted before the reaction short-circuit.
- Per-delivery id added for receiver dedupe; docs state at-least-once /
unordered delivery.
- Signature verify tolerates hex case / header whitespace.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Complete the public API roadmap on top of the existing key-auth layer
(#290). Re-lands POST /api/v1/messages (lost when the stacked #291 never
reached main) by extracting the dashboard send route's core into
`sendMessageToConversation`, then adds:
- Contacts: GET/POST /api/v1/contacts, GET/PATCH /api/v1/contacts/{id}
- Conversations: GET list + /{id} + /{id}/messages
- Broadcasts: POST /api/v1/broadcasts + GET /{id} (persist rows, fan out
in after(), poll for status)
All list endpoints share one keyset-pagination contract
({ data, meta: { next_cursor } }). No migration required — scopes and
tables already exist. Versioned as 0.3.0 (folded into [Unreleased]);
docs/public-api.md updated; outbound webhooks left as the sole roadmap
item. API stays v1 (additive).
Includes fixes from an adversarial code review: cursor validation
against filter injection, broadcast counts deferred to the DB aggregate
trigger, terminal broadcast status reflects failures, duplicate-phone
de-dup, up-front send-param validation (no orphan conversations),
PATCH null-clear semantics, error-checked tag sync, bounded tag filter,
unified audit-user attribution, and a maxDuration on the broadcast
fan-out.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds contact-based filters to the Inbox conversation list (issue #272):
- Multi-select tag filter with OR logic (a conversation matches if its
contact has any selected tag), consistent with Broadcast audiences.
- Company filter (exact match) over companies present in the inbox.
- Filters stack with the existing status filter and search; active
selections show as removable chips with a clear-all control.
Tags are hydrated via a shared CONVERSATION_SELECT embed and normalized
onto contact.tags, so the realtime hydrate path keeps new conversations
matching an active filter. Filtering logic lives in a pure, unit-tested
matchesContactFilters helper.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The WhatsApp settings load effect re-ran fetchConfig() whenever `user`
or `profileLoading` changed. Supabase's onAuthStateChange fires a token
refresh (new `user` object, profileLoading toggles) when the browser
tab regains focus, so returning to the tab re-hydrated the form and
overwrote any unsaved credential input.
Guard the effect with a ref tracking the account id already loaded, so
it only fetches once per account and ignores auth-state churn that
doesn't actually change the account. Explicit reloads after save,
registration verification, and reset still call fetchConfig directly.
Closes#307
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replace `user` object with `user?.id` in dependency arrays to prevent unnecessary re-renders. Add `lastFetchedUserIdRef` to track the user ID for successfully initiated profile fetches, preventing redundant API calls when the window refocuses or the auth token is refreshed.
The inbox/contact TemplatePicker fetched approved templates filtered by
`user_id = current user`, so in a shared account a teammate only saw
templates they personally created — others' approved templates were
hidden, leaving them unable to send those templates.
message_templates is account-owned and its RLS SELECT policy already
scopes rows to account members (is_account_member, migration 017), so
the user_id filter was both redundant and over-restrictive. Drop it and
rely on RLS, matching the broadcast template picker.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
