* feat(agent): end-to-end in-product AI agent chat

Phase 2 of #94. Sidebar entry under Brands, /dashboard/agent route,
streaming chat backed by Anthropic Claude Sonnet 4.6, conversation
+ message persistence, monthly token quota per the plan flag landed
in PR #120.

Backend:
- web/src/lib/agent/model.ts — direct @ai-sdk/anthropic provider.
  Issue body cross-cutting requirement says "self-hosted instances
  use their own keys (BYO)"; Vercel AI Gateway would lock self-host
  out, so we mirror the existing server/src/lib/ai-provider.js
  pattern.
- web/src/lib/agent/system-prompt.ts — analyst persona distilled
  from skills/ansvisor-aeo-coach/SKILL.md (never invent numbers,
  always call a tool first, lead with the headline, end with one
  actionable next step).
- web/src/lib/agent/tools.ts — Vercel AI SDK tool defs wrapping the
  MCP data layer fns directly (list_brands, get_visibility_summary,
  get_visibility_trend, get_competitor_comparison, list_citations,
  list_topics, list_prompts, list_content_opportunities). Same
  scoping the MCP server uses; no data path duplication.
- web/src/lib/agent/auth.ts — resolves the dashboard session into
  the McpAuthContext shape so the tools can be reused as-is.
- web/src/lib/agent/token-quota.ts — monthly bucket lookup + upsert.
  Self-host / enterprise get unlimited via aiAgentTokenQuota
  undefined.

API routes:
- POST /api/agent/chat — streaming. Verifies session, conversation
  ownership, plan flag, quota; then streamText with the tool surface
  and toUIMessageStreamResponse({ originalMessages, onFinish }).
  onFinish persists the full UIMessage[] (replace-in-place per
  turn, parts kept as jsonb in tool_calls for replay), auto-titles
  the conversation from the first user message, and bumps the
  updated_at trigger so the sidebar re-sorts.
- GET /api/agent/conversations — sidebar list.
- POST /api/agent/conversations — "+ new chat".
- GET/DELETE /api/agent/conversations/[id] — load thread, drop
  thread.

UI:
- web/src/config/dashboard.ts — "Agent" sidebar entry under
  Brands, requiredFeature 'ai_agent', Sparkles icon.
- web/src/components/layout/sidebar.tsx, web/messages/en.json —
  nav.agent translation.
- /dashboard/agent — single client component with conversation
  sidebar + chat area. useChat from @ai-sdk/react with
  DefaultChatTransport pinned to /api/agent/chat. Plan-gated render
  shows an upgrade card when the user's plan doesn't include
  ai_agent. Messages render text parts inline and tool calls as a
  compact pill so the user sees which tool ran.

Refs #94.

* fix(agent): convert conversation row from button to role=button div

HTML doesn't allow <button> nested inside <button>, which Next.js
surfaces as a hydration error. The sidebar's conversation row was a
real <button> (so the entire row was clickable) and the trash icon
was a nested <button> for delete — that's the disallowed pattern.

Swap the outer <button> for a <div role="button" tabIndex={0}> with
the same onClick plus a keyboard handler for Enter / Space so screen
readers + keyboard users keep the same affordance. Inner trash stays a
real <button>.

* fix(agent): pass conversation id synchronously to sendMessage

POST /api/agent/chat was returning 400 "conversationId required" on
the first message of a brand-new chat. Root cause: setActiveId is an
async React state update, so when sendMessage runs immediately after
in the same tick, prepareSendMessagesRequest reads the still-null
state from its closure and posts conversationId: null.

Two-part fix so neither path can drift again:

- activeIdRef mirrors activeId; useEffect keeps it in sync, and
  prepareSendMessagesRequest reads from the ref instead of the
  stale state closure. Covers the case where a user clicks an
  existing conversation in the sidebar and immediately types.
- onSubmit captures the resolved id into a local, updates the ref
  synchronously when it just created a new conversation, and
  passes the id via sendMessage's body option as well. Belt-and-
  suspenders: even if the ref hadn't been touched yet, the body
  argument arrives in prepareSendMessagesRequest's body param and
  the spread order lets it take priority.

* fix(agent): set stopWhen on streamText so the tool loop continues

streamText is single-step by default in v6 — the model fires a tool
call and ends the turn before ever seeing the result. The chat in the
test session showed exactly this: "Hangi marka için bakmamı
istersiniz?" + tool-list_brands part, then nothing. The user had to
manually re-ask before Claude (now seeing the prior tool result in
the second turn's context) answered with the brand list.

stepCountIs(20) re-runs the generation after each tool call so the
agent interprets the result and either answers or calls another tool.
20 matches the ToolLoopAgent default — plenty of headroom for the
typical 'list_brands → get_visibility_summary → answer' chain
without burning tokens forever on a pathological loop.

* fix(agent): render assistant messages through the existing Markdown component

The agent answers in markdown — lists, **bold**, headings, links —
but the chat bubble was rendering text parts as plain <p> with
whitespace-pre-wrap, so the user saw the raw markup characters in
the response. Reuse @/components/ui/markdown (the same component the
insights detail page uses for AI responses) for assistant text. User
messages stay plain — they're typed input, not authored markdown,
and the chat bubble contrast reads cleaner without prose styling.

Tool-call pills are unchanged.

* fix(agent): fill viewport height so sidebar + composer reach the bottom

Desktop has no top bar in the dashboard layout (the mobile nav is md:hidden),
so subtracting 3.5rem from 100vh left a ~56px gap below both the conversation
list and the message composer — looking like a phantom padding-bottom. On
mobile the same calc was 8px short (mobile nav is h-16 = 4rem, not 3.5rem).

Switch to h-[calc(100vh-4rem)] md:h-screen so the agent shell matches the
parent main element's border box on both breakpoints. -m-6 already cancels
the parent's p-6, so border-box height = main's border-box height and the
scroll content height = main's content area (no overflow on either axis).

* feat(agent): expandable tool call disclosure with input + output

The previous chip showed the raw AI SDK state string ('output-available')
next to the tool name — internal jargon that doesn't help a user understand
what the agent actually did. It also wasted the only interactive surface we
had on a part that carries the most debug-worthy payload of the turn.

The new ToolCallDisclosure renders the chip as a button: spinner while the
call is in-flight, wrench + chevron once it has a result. Clicking expands
an inline console-style panel with the input args and output result as
pretty-printed JSON, plus an error block if the call failed. Same pattern
Claude / ChatGPT / Cursor ship for tool runs, and it works for every tool
we already register without per-tool formatting code.

* feat(agent): bring-your-own-key on cloud, drop platform-side token quota

Cloud rollout for the in-product agent had two open problems: every turn
billed Ansvisor's Anthropic account, and the feature was gated to Growth+
to keep that bill bounded. Both go away with BYOK.

What's new
- Settings → Agent: org admins paste their own Anthropic key. Stored
  encrypted at rest (AES-256-GCM) in organizations; the master key lives
  in the new ANSVISOR_ENCRYPTION_KEY env, never touches Postgres. UI
  shows configured/last4/set-at metadata, never the plaintext.
- /api/settings/anthropic-key (GET / PUT / DELETE): GET surfaces only the
  display metadata (any member can read), PUT and DELETE are admin-only.
  Soft format validation on save; first chat turn is the rigorous check.
- /api/agent/chat + /api/agent/conversations now gate on 'is a key
  configured?' instead of token quota. Missing key returns 403 with
  code 'missing_anthropic_key' so the UI can route to Settings.
- web/src/lib/agent/key.ts resolves the key per request: cloud → org
  (decrypt envelope), self-host → ANTHROPIC_API_KEY env. Self-host gets
  the feature unconditionally; cloud gates on BYOK.
- model.ts is now a buildAgentModel(key) factory so streamText can be
  driven with the caller's key per turn.
- Agent page renders a KeyMissingState (CTA → Settings → Agent) when
  cloud and no key set, instead of letting the user submit and 403.

Plans
- ai_agent feature added to Starter; no longer a Growth+ gated feature
  on cloud. The real gate is BYOK.
- aiAgentTokenQuota removed entirely from PlanLimits — no Growth quota,
  no Omit gymnastics in isWithinLimit / useFeatureGate.withinLimit /
  plan-guard.enforceLimit. agent_token_usage table stays for analytics
  (recordAgentTokenUsage still writes on each turn).

Encryption
- AES-256-GCM, 12-byte IV per encryption, GCM auth tag attached. Stored
  as a JSON envelope { v, iv, tag, ct } so we can version-bump the
  scheme later. Master key validated at load — refuses to write unless
  ANSVISOR_ENCRYPTION_KEY decodes to exactly 32 bytes.

Migration
- 00010_org_anthropic_keys.sql adds anthropic_api_key_encrypted,
  anthropic_api_key_last4, anthropic_api_key_set_at,
  anthropic_api_key_set_by to organizations. set_by FK to profiles is
  ON DELETE SET NULL so dropping a profile preserves the audit trail.

Deploy notes
- Generate ANSVISOR_ENCRYPTION_KEY: 'openssl rand -base64 32'. Add it
  to web/.env locally and to Vercel env (production + preview +
  development) before the cloud rollout. Losing or rotating it makes
  saved keys unrecoverable — users will see the empty state and need
  to re-paste, no data loss otherwise.

Bundles the in-product Agent epic (#120, #121), the BYOK rollout on
cloud, the MCP toolkit expansion (#109, #110, #116, #117, #118), the
insights performance fix (#114), and the invite-accept flow rebuild
(#127, #129, #130) into a tagged release.

See CHANGELOG.md for the full notes.

Bundles the in-product Agent epic (#120, #121), the BYOK rollout on
cloud, the MCP toolkit expansion (#109, #110, #116, #117, #118), the
insights performance fix (#114), and the invite-accept flow rebuild
(#127, #129, #130) into a tagged release.

See CHANGELOG.md for the full notes.