Two related fixes for the "switching agent mid-thread loses context"
issue:
1. LLM-context fix
The system prompt now includes a PRIOR HAND-OFF block whenever the
current conversation has been touched by more than one agent. It
lists the prior personas (name + role) and tells the new agent:
"Earlier turns were produced by other personas. Read them as
context, but answer in your own voice as the current persona."
Without this, switching from Atlas (Operator) to Pythia (Researcher)
left Pythia answering as if she'd produced Atlas's prior turns.
Tracked via two-trigger useEffect:
- On agent change with messages already in the thread, the prior
agent gets locked into history.
- On stream finish, the current active agent gets added (it just
produced a turn).
Cleared with the conversation.
2. UI-attribution fix
Each assistant turn now records which agent produced it
(messageAgents map: index -> Agent). The row signature in
MessageRow now reads that stamped agent rather than always echoing
the currently-active one. Switching agents mid-thread no longer
retroactively re-attributes prior responses.
Both maps are wiped by Clear conversation alongside the live snapshot
and initialLive ref, so a fresh thread starts truly fresh.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
useChat exposes reset() which calls setMessages(opts.initialMessages),
and the AI page passes initialLive.current as initialMessages — a ref
captured once on mount from localStorage.
resetAndClear was calling reset() then clearLive(). Sequence problem:
reset() → setMessages(initialLive.current) // populated old array
clearLive() → localStorage.removeItem(LIVE_KEY) // does nothing to memory
The ref still held the original messages, so reset re-seeded them and
the conversation appeared to "come back" the moment you typed anything
(or sometimes immediately, depending on render timing).
Fix: blank the ref, clear localStorage, and call setMessages([])
directly. reset() is no longer needed at this call site so it's been
dropped from the useChat destructure.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The composer's model picker had <DropdownMenuTrigger asChild><button>...
which is the Radix Slot pattern. Project uses @base-ui/react where
Menu.Trigger has no asChild prop and renders its own <button>, so the
result was nested-button-inside-button (DOM-nesting violation) plus
asChild leaking as an unknown DOM attribute (React warning).
Dropped the inner <button> and put className/data-action straight on
the Trigger. Visual output identical, no more console errors.
This pattern is used by ~14 other routes (Button asChild + Link),
mostly behind sign-in-required states. They're broken too but rarely
fire — separate followup.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two bugs in the previous /ai redesign:
1. theme="console" on AppShell put the entire shell (sidebar, appbar,
appbar dropdowns, the lot) inside [data-theme="console"], so the
console palette + JetBrains Mono override leaked into the sidebar
and made light mode look broken on /ai. Scoped now: the AppShell
stays in skyrise (so light/dark toggle keeps working everywhere),
and only the route content area gets data-theme="console" via an
inner wrapper.
2. The Google Fonts @import inside console.css was being silently
dropped because @import rules must precede all other rules in the
final bundle, and skyrise's content lands first. Moved JetBrains
Mono + Newsreader into app.css's top-level @import url() alongside
the existing Inter/Instrument Sans/Geist Mono families.
Atmosphere ::before was also position: fixed, which painted the grain
overlay across the whole viewport (including the sidebar) regardless
of where data-theme lived. Now position: absolute on the wrapper, with
isolation: isolate to keep z-index local.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces the conventional chat aesthetic on /ai with a brutalist-mono
operator deck. The page now reads as a flight recorder — turn numbers
in the gutter, hairline rules, sodium-amber phosphor primary on
deep-ink ground, vim-style modeline at the foot.
Type system is the design's load-bearing element:
- JetBrains Mono for everything system-y (operator lines, signatures,
modeline, session ids, tool calls)
- Newsreader serif for the agent's prose only — the synthesis voice
literally lifts off the page in a different family from the machine
voice. Operator and agent are typographically inseparable from their
speaker.
Layout changes:
- Sticky session header with a giant base36 session id ("3K9P · A4C2")
and a metadata strip showing agent, model, turn count, status. The
status pill flips colour: AMBER on stream, ROSE on awaiting confirm,
MINT on ready, MUTED on mock.
- Empty state is no longer the apologetic "How can I help you today?".
It's "ATLAS. standing by." in oversize mono with the agent name in
italic serif amber, a hairline divider, and a single one-liner
instruction prefixed with ›. Lines stagger in via animation-delay.
- Operator turns: monospace, 14px, sodium-amber › prompt, no bubble.
Hangs from a left gutter with T01/T02… turn number + UTC timestamp.
- Agent turns: serif, 17px/1.55, with a tiny mono signature underneath
("atlas» 03:14:08Z · recv"). Cyan accent column instead of amber.
- Composer: terminal frame (square, 1px border, focus ring is amber
glow). Internal ›_ prompt mark in front of the textarea, mono input.
- Bottom modeline: utc clock + turn count + estimated tokens on the
left, keyboard hints on the right. Streaming flips the right side
to a pulsing phosphor bar + STREAM label.
Atmosphere details:
- 2px scanline overlay (very faint, 1.2% opacity)
- Corner phosphor blooms (amber top-right, cyan bottom-left)
- Inline SVG turbulence grain (3.5% opacity) over the whole theme
- Cursor blink animation on the prompt mark
- Consolas-tier ligatures on the mono via JetBrains Mono ss01/calt
All theming scoped via [data-theme="console"] — picks up automatically
because /ai's AppShell now passes theme="console". Other routes are
untouched. Tool-call cards from @crema/agent-ui inherit the palette
via overridden CSS variables (--card, --border, --primary, etc) plus
a [data-slot="tool-call-card"] override for the frame.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces the single-base-URL LLM settings with the new providers lib
(OpenAI, Anthropic, DeepSeek, Qwen, LM Studio). Settings/LLM hosts the
catalog-aware card; the /ai route builds adapters via buildAdapter()
and resolves API keys from the arcadia vault per-call (direct mode).
Anthropic skips the /v1/models probe (no such endpoint) and uses
catalog defaults; failed probes for keyed providers fall back to the
catalog instead of dropping to mock.
AI conversation now persists across navigation and refresh via a new
crema.ai.live localStorage key (separate from the compact-snapshot
key). useChat hydrates from initialMessages on mount, saves on every
change, and "Clear conversation" wipes both state and storage.
Vite needs explicit resolve.alias for @crema/llm-ui and
@crema/llm-providers-ui — when a sibling lib imports another @crema/*,
tsconfigPaths can't resolve it (the importing file isn't in this
project's tsconfig scope).
Adds docs/LLM_PROXY_CONTRACT.md describing the
POST /api/v1/ai/llm/chat endpoint the backend needs for proxy mode
(keys never leave the server). Direct mode works against today's
arcadia; proxy mode unblocks once that endpoint ships.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three layers:
1. GFM markdown — add remark-gfm so tables, task lists, strikethrough,
autolinks render properly. Style table elements (overflow-aware
container, muted header, divider rows). Render `[ ]` task list items
as visible checkboxes.
2. Structured tool-result rendering — new `tool-result-renderers.tsx`
dispatches by tool name to render a small UI block beneath each
ToolCallCard:
- list_tenants → table with status pills + plan column
- get_tenant → tenant detail card
- get_platform_stats → KPI tiles (total + per-status)
- list_audit_log → timeline rows with actor_type + action
- list_users → user list with role chips
- suspend_tenant / activate_tenant → tenant card with action confirm
ToolCallCard collapses by default — operators expand for raw JSON.
3. Custom ```card``` blocks the LLM can emit inline:
- {"kind":"pill","status":"…"} — status pill
- {"kind":"stat","label":"…","value":…} — stat tile
- {"kind":"callout","tone":"info|warning|danger|success",…} — callout
Malformed blocks fall through to the prose unchanged. Client strips
well-formed blocks from prose and renders them as components.
Domain primer updated to teach the model the card schemas and remind it
NOT to re-render tool-result data as markdown tables (that's done
automatically — it should add commentary only).
Layers are independent: 1 + 2 always work; 3 is purely additive.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Persistent presence is less jumpy than pop-in/out. When the model isn't
actively doing anything, the avatar collapses to the `sm` size variant,
loses its activity label, and dims to 50% opacity. Springs back to
full-size + label when activity resumes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The /ai surface now renders agent-ui primitives instead of homegrown
tool/typing widgets:
- AgentAvatar with activity (thinking / working / waiting / speaking /
idle) replaces TypingIndicator. Pulses while the model is generating,
shows "waiting" while a write is held for confirmation, "working" while
a confirmed write is executing, "speaking" once tokens are streaming.
- ToolCallCard renders each native tool_call with typed status (pending
/ running / success / error). Built from the assistant message's
toolCalls plus the matching tool result message. Tool messages no
longer render standalone — absorbed into their parent assistant turn.
- Empty assistant bubbles (no prose, only tool_calls) collapse so the
ToolCallCards carry the visual weight.
Wiring: add @crema/agent-ui path entry to tsconfig and @source line to
app.css. Sibling lib-agent-ui must be cloned next to arcadia-admin.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
New tools in admin-tools.ts:
- list_audit_log({limit?}) — recent audit entries (terse: actor, action,
target, timestamp). Hits /api/v1/admin/audit-log.
- get_platform_stats() — aggregate counts (tenants by status + by plan),
composed locally from list_tenants until arcadia exposes a real stats
endpoint.
- list_users({limit?}) — users in the currently-selected tenant via
/api/v1/users.
- suspend_tenant({slug}) — write tool, suspends a tenant by slug.
- activate_tenant({slug}) — write tool, restores a suspended/deactivated
tenant.
Inline write confirmation:
- New ConfirmCard component renders below the assistant message that
proposed a write. Shows tool(args) and Confirm/Deny buttons.
- classifyCalls() splits LLM tool calls into reads/writes. Auto-loop
runs reads immediately; for any writes, holds them in pendingConfirm
state instead of dispatching.
- On Confirm: runs writes with allowWrites:true, prepends prior read
results, continueChat to produce the final answer.
- On Deny: synthesises tool-result messages telling the model the user
declined; continueChat so it can acknowledge.
- Arcadia-knowledge primer updated to tell the model the user sees an
inline confirm card automatically — it shouldn't ask in prose first.
Wired into both /ai and /assistant.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Make /ai and /assistant operate as the platform admin's assistant
against arcadia-app's API:
- Add `arcadia-knowledge.ts` — domain primer (multi-tenant Phoenix
backend, tenant lifecycle, platform_admins identity, etc.) baked into
every system prompt.
- Add `admin-tools.ts` — curated tool registry exposing `list_tenants`
and `get_tenant`, callable via OpenAI-native function calling. Tools
hit arcadia through `useArcadiaClient()` and inherit the operator's
JWT + tenant header. `runLLMToolCalls()` returns `tool` role messages
ready to push back into history.
- Add `admin-context.ts` — runtime registry pages publish to so the
assistant can answer factual questions about live UI state without
scraping the DOM. Tenants page registers its summary on mount.
- Replace generic Vibespace personas (Atlas/Forge/Inkwell/Pilot/Cursor)
with arcadia-flavoured ones: Operator, Auditor, Triage, Analyst,
UI Operator. Auto-migrate stored agents from the legacy set.
- /assistant: build admin preface (role + primer + persona + ctx) and
pass it as the `useChat` system at construction. Pass `tools` on every
`send()`. Auto-loop reads `toolCalls` off the streaming assistant
message and uses `continueChat()` to push tool results.
- /ai: same wiring (this is the canonical admin chat surface; the user
prefers its look).
- MessageBody renders tool-result cards (role: "tool") and a "Called X"
pill on assistant messages with toolCalls. Strips Qwen-style
`<tool_call>` XML from prose when the tags were converted to
structured calls.
- Extend ThreadMessage with the `tool` role + tool-call metadata so
conversations round-trip through localStorage.
- Tenants page: row actions get `data-action="tenant-<slug>-{suspend,
activate,deactivate}"` (via lib-table-ui's new dataAction prop);
registers tenant summary into admin-context.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Initial commit. Spun up via the docs/STARTER.md recipe: cp from vibespace,
reset git, rename package, set brand to "Arcadia Admin" with Shield icon
in app/lib/identity.ts.
Inherits the full Crema sibling-lib wiring including @crema/arcadia-client
(typed HTTP + Phoenix Channels realtime against arcadia-core) and
@crema/arcadia-auth-ui (login/signup/password-reset/2FA forms). The /login
route already renders <LoginForm>; <ArcadiaProvider> in app/root.tsx reads
VITE_ARCADIA_URL (default localhost:4000) and VITE_ARCADIA_TENANT (default
"default").
CLAUDE.md and README rewritten to frame this as the admin app for
arcadia-core. docs/STARTER.md removed — arcadia-admin is a leaf consumer,
not a downstream starter.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
