diff --git a/CLAUDE.md b/CLAUDE.md
index da10214c..56e5938e 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -1,13 +1,13 @@
 # Claude Code Working Guide
 
-_Last synced: v2.8.13 (2026-10-20) — Windows paths in chat open as files; AI Tools npm install falls back to `--prefix ~/.local`; editing a skill artifact writes a new snapshot, a shared row its owner deletes stays in the Hub; Files uploads small files in batches (a member's: one write-as `extract`)._ A change that makes any statement in this file, `AGENTS.md`, or a README wrong — a moved path, a new or changed extension point, a retired rule — updates that file in the same change-set and bumps this marker. **Bump = replace the version/date and a ≤ 1-line summary; never append a changelog here** — details belong in the commit message and `docs/`, because this whole file is loaded into every session.
+_Last synced: v2.8.13 (2026-10-21) — realtime chat (bridge, `/ws-chat`) removed; the terminal chat covers omp/Cursor/agy (agy's step log as a transcript), permission prompts as cards, Stop from the pane's busy line, Codex `/model`·`/permissions` pickers, `transcript_changed` push; artifacts kept._ A change that makes any statement in this file, `AGENTS.md`, or a README wrong — a moved path, a new or changed extension point, a retired rule — updates that file in the same change-set and bumps this marker. **Bump = replace the version/date and a ≤ 1-line summary; never append a changelog here** — details belong in the commit message and `docs/`, because this whole file is loaded into every session.
 
 ## Project mission
 - Maintain a reliable local-first dashboard for Claude Code session monitoring.
 - Preserve real-time behavior (hooks -> API -> SQLite -> WebSocket -> UI).
 
 ## Repo map
-- `server/`: Express API, hook ingestion, SQLite access, websocket broadcast (includes release-based update detection — `lib/update-check.js` reads the repo's own releases API and `routes/updates.js` exposes status/check/apply, refusing to self-update a `~/.claude/plugins/cache/` install, plus `lib/workflow-ingest.js` which ingests on-disk Workflow-tool run journals — fleets that emit no hooks — and `lib/codex-ingest.js`, which mirrors a `codex` lane's running/idle from the mtime of Codex's own rollout journals (skipping a lane whose hooks are current), `lib/codex-usage.js`, which reads the same journals into `sessions` (`tool = 'codex'`) and `token_usage`, and `lib/codex-hooks.js`, which installs AgentHub's opt-in hooks into Codex's `hooks.json`, and their Antigravity counterparts `lib/antigravity-usage.js` (agy's conversation files → `tool = 'antigravity'` sessions and tokens) and `lib/antigravity-hooks.js` (the named hook `agenthub` in `~/.gemini/config/hooks.json`), and `lib/cursor-usage.js` (Cursor's agent transcripts → `tool = 'cursor'` sessions, no tokens), and `lib/omp-usage.js`, `lib/opencode-usage.js` (omp's journals, OpenCode's `opencode.db` → `tool = 'omp'|'opencode'` sessions, tokens and subagent agents)).
+- `server/`: Express API, hook ingestion, SQLite access, websocket broadcast (includes release-based update detection — `lib/update-check.js` reads the repo's own releases API and `routes/updates.js` exposes status/check/apply, refusing to self-update a `~/.claude/plugins/cache/` install, plus `lib/workflow-ingest.js` which ingests on-disk Workflow-tool run journals — fleets that emit no hooks — and `lib/codex-ingest.js`, which mirrors a `codex` lane's running/idle from the mtime of Codex's own rollout journals (skipping a lane whose hooks are current), `lib/codex-usage.js`, which reads the same journals into `sessions` (`tool = 'codex'`) and `token_usage`, and `lib/codex-hooks.js`, which installs AgentHub's opt-in hooks into Codex's `hooks.json`, and their Antigravity counterparts `lib/antigravity-usage.js` (agy's conversation files → `tool = 'antigravity'` sessions and tokens, and its step log as their conversation view) and `lib/antigravity-hooks.js` (the named hook `agenthub` in `~/.gemini/config/hooks.json`), and `lib/cursor-usage.js` (Cursor's agent transcripts → `tool = 'cursor'` sessions and their conversation view, no tokens), and `lib/omp-usage.js`, `lib/opencode-usage.js` (omp's journals, OpenCode's `opencode.db` → `tool = 'omp'|'opencode'` sessions, tokens and subagent agents)).
 - `client/`: React + Vite UI.
 - `scripts/`: hook installer/handler, import, seed, cleanup utilities. (Update detection lives server-side in `server/lib/update-check.js`, keyed on the forge's **newest release** rather than a commit count. `POST /api/updates/apply` has two modes and picks by install kind, never from the request: a fast-forwardable **checkout** merges the release tag `--ff-only`, rebuilds, and self-restarts via `server/lib/self-restart.js` — through our own unit when this process runs in one (`systemctl restart --no-block`, launchd's KeepAlive; a detached replacement beside a `Restart=always` unit crash-loops it), else the detached `scripts/restart-helper.js`; a **plugin-cache** install is handed to `server/lib/plugin-update.js` + the detached `scripts/plugin-update-helper.js`, which runs `claude plugin marketplace update` → `claude plugin update -y` → `plugin-bootstrap.js --force` from the NEW cache dir. Both write the HTTP response before anything stops the process, and every plugin-chain outcome is recorded to `plugin-update.json` in the runtime dir — outside the cache, so it survives the swap and the next server reports it as `last_plugin_update`. Neither mode can run `/reload-plugins`: it is REPL-scoped to the session that invokes it, so an open Claude Code session still reloads by hand for new skills/commands. Elsewhere users run the printed manual command, surfaced in the UI and by `agenthub update-check`. `scripts/cc-update-all-helper.js` is the same detached pattern for the config panel's **Update all** button — marketplaces → every plugin → AgentHub LAST → one reload — journalling each step to `cc-update-all.json` in the runtime dir.)
 - `skills/`: this repo's skills. Because the repo root **is** the `agenthub` plugin root, this is also the plugin's skills directory — every skill here ships to anyone who installs the plugin, addressed as `/agenthub:<name>` (bare `/<name>` works as an alias). Put a skill here only if it should reach plugin users. The plugin ships no subagents and no pipeline skill, and depends on no other plugin (Superpowers included) — do not add an auto-install of one back into the bootstrap.
@@ -64,6 +64,7 @@ A **workspace** — called a **lane** in the API, CLI, database and code, which
 - **A workspace's sessions are the ones it OWNS, never the ones that ran in its folder.** `sessions.lane_id` is stamped by `bindRunSession` (`server/routes/hooks.js`) for the session a dashboard pane holds — first writer wins, a nested `claude -p` excluded — and no route sets it. A Codex/Antigravity session (their hooks are opt-in, so a pane may never learn its session) is also stamped by `claimToolSessions` (`server/lib/dashboard-runs.js`, called by both pollers and once at start): the lane of the newest same-tool dashboard run in the session's folder that was live when the session started, never a Claude session. A new Claude pane is started with `--session-id <uuid>` (`spawnRun`), so its `dashboard_runs.session_id` is known before any hook and a hook that carries no `run_id` (a hook handler older than `AGENTHUB_RUN_ID` forwarding) is still stamped from that run (`laneForSession`). Every workspace session list reads `GET /api/sessions?lane=<id>` — which also counts a still-unstamped session one of the lane's `dashboard_runs` holds, because hooks reach only the lowest-port dashboard on a shared DB and that one may be an older build that never stamps — and shows nothing else (no "other sessions in this folder"). Do not go back to listing a workspace by `cwd`, and do not guess ownership for old history.
 - **A session's worktree is `claude --worktree`'s, resolved from the session row — and removing it never forces.** A Worktree start passes `--worktree=<server-made name>` (the `=` form: `-w [name]` eats the prompt otherwise) and Claude Code makes, populates and cleans up `<repo>/.claude/worktrees/<name>`; AgentHub only excludes `/.claude/worktrees/` at the repo root. A git route, a resume or a removal may work in a session's worktree only through `sessionWorktree` (`server/lib/session-worktree.js`: the session owned by the lane, its cwd not the lane's folder, registered by `git worktree list`) — `?session=` on the git routes answers `ENOWORKTREE` rather than falling back to the main folder. `POST /api/lanes/:id/session-worktree/remove` (same-origin, `confirm: true`, refused while a live pane holds the session) is the ONE dashboard path that deletes a worktree: `git worktree remove` with no `--force`, `git branch -d` never `-D`, a live-pid lock refused. Do not add a force, and do not let it touch a lane's own folder.
 - **Opening a session never resumes it — only SENDING a message to it does.** The sidebar and a workspace home ATTACH pane 0 to a session's live tmux pane (`/workspace?lane=<id>&run=<runId>`, `attachRequest` on `LaneConsolePane`) or, for an ended one, open it in its workspace like Claude Desktop (`/workspace?lane=<id>&session=<id>`, `client/src/components/lanes/SessionReplay.tsx`: its conversation read-only plus a prompt box; the console is only hidden, CSS). A resume re-reads the whole transcript into context, so it happens only on an explicit action: a message SENT from that replay (`resumeSessionId` + `prompt`, first free slot), the transcript page's **Resume**/**Fork** (`SessionResumeButton.tsx`, behind a warning), or the ticked sessions of the **Resume interrupted sessions?** dialog (`InterruptedSessionsPrompt.tsx`, `GET /api/run/interrupted` — runs whose pane died under them; opening it resumes nothing). The analysis page `/sessions/<id>` is reached from the session view's **Analyst** button (beside Files/Changes in the right panel). Do not add a path that resumes on a click. A resume is refused while a live pane holds the session (`409 ESESSIONLIVE` with that pane's `run_id`) — open that pane instead; a fork makes a new session and is allowed.
+- **A chat artifact runs in its own sandbox, never on the dashboard's origin.** The artifact panel (`client/src/components/run/ChatFilePreview.tsx`, `client/src/lib/artifactView.ts`) runs model-written HTML/Mermaid/React in an `<iframe srcdoc sandbox="allow-scripts …">` and an SVG in `sandbox=""` — never add `allow-same-origin` (the script would read the dashboard's cookies and call its API as the user), never render one with `dangerouslySetInnerHTML`, and keep its CDN libraries pinned to exact versions. Versions are rebuilt from Write/Edit inputs (kept up to 256 KB by `server/lib/transcript-message.js`); no copy is stored, and Save to Artifacts uses the Hub's own `pinFile`/upload routes.
 - **A `cwd` can back at most one lane per tool, not one lane overall.** Uniqueness is the composite `idx_lanes_cwd_tool` (`cwd`, `tool`), not `cwd` alone — Repo/Adopt mode (`POST /api/lanes/ensure`) lets the same directory carry a `claude` lane and a `codex` lane side by side, shown as two separate cards. `resolveLaneByCwd(cwd, tool)` takes an optional `tool` filter for exactly this; hook ingestion (`touchLaneFromHook`) pins it to the hook's own tool — `claude`, `codex` for a hook posted with `tool: "codex"`, or `antigravity` for one posted with `tool: "antigravity"` (the Codex and Antigravity hook rules above) — so a hook event never touches the other tool's lane sharing the path, and stage detection (`resolveLaneForWork` + `detect`) runs for `claude` only. Worktree mode is unaffected — it already allocates a unique directory per lane.
 
 See `docs/LANES.md` for full guide: stage reporting, custom templates, lane actions, and the five node states.
diff --git a/client/src/components/Sagi/PetChat.tsx b/client/src/components/Sagi/PetChat.tsx
index 751afb0e..18b28b02 100644
--- a/client/src/components/Sagi/PetChat.tsx
+++ b/client/src/components/Sagi/PetChat.tsx
@@ -34,6 +34,7 @@ import {
   chooseRunOption,
   interruptRun,
   readRunMenu,
+  stepRunMenu,
   submitToRun,
   type PaneMenu,
 } from "../../lib/panePaste";
@@ -505,6 +506,9 @@ export function PetChat({
                 if (answerRunMenu(chat.runId, menu, text)) setMenu(null);
                 else setError(t("sagi.chat.notSent"));
               }}
+              onStep={(dir) => {
+                if (!stepRunMenu(chat.runId, dir)) setError(t("sagi.chat.notSent"));
+              }}
               onCancel={() => {
                 if (interruptRun(chat.runId)) setMenu(null);
                 else setError(t("sagi.chat.notSent"));
diff --git a/client/src/components/conversation/ConversationView.tsx b/client/src/components/conversation/ConversationView.tsx
index 25034145..896327d2 100644
--- a/client/src/components/conversation/ConversationView.tsx
+++ b/client/src/components/conversation/ConversationView.tsx
@@ -96,6 +96,9 @@ const TRANSCRIPTS_REFRESH_MS = 15000;
 // at most this often: the first event still fetches at once, the last is never
 // lost, and a long transcript is not re-read for every single frame.
 const EVENT_FETCH_MS = 1500;
+// The server saw this session's transcript grow (`transcript_changed`, itself
+// at most every 250 ms): the new lines are on disk, so read them nearly at once.
+const PUSH_FETCH_MS = 300;
 
 interface ConversationViewProps {
   sessionId: string;
@@ -108,6 +111,10 @@ interface ConversationViewProps {
   compact?: boolean;
   /** Opens a file named in a turn's "Edited N files" card. */
   onOpenFile?: (path: string) => void;
+  /** Opens an artifact card's file in the chat's artifact panel. */
+  onOpenArtifact?: (path: string) => void;
+  /** The messages on screen (for an artifact's versions), as they change. */
+  onMessages?: (messages: TranscriptMessage[]) => void;
   /** The session runs in a pane: show the step in progress. */
   live?: boolean;
   /** Changed (a chat's send): back to the end, and stick to it. */
@@ -133,6 +140,8 @@ function ConversationViewPane({
   fill = false,
   compact = false,
   onOpenFile,
+  onOpenArtifact,
+  onMessages,
   live = false,
   follow = 0,
   waiting = false,
@@ -144,19 +153,21 @@ function ConversationViewPane({
 }: ConversationViewProps) {
   const { t } = useTranslation("sessions");
   const [messages, setMessages] = useState<TranscriptMessage[]>([]);
+  const shown = messages;
   // A live pane's session has no row until its first hook — and Claude Code
   // runs no hook before the folder is trusted, so a trust prompt (or any
   // other question before the session starts) waits unseen in the terminal.
   const [notStarted, setNotStarted] = useState(false);
   useEffect(() => onNotStarted?.(notStarted), [notStarted, onNotStarted]);
+  useEffect(() => onMessages?.(shown), [shown, onMessages]);
   useEffect(() => {
     if (!onContext) return;
-    for (let i = messages.length - 1; i >= 0; i--) {
-      const m = messages[i]!;
+    for (let i = shown.length - 1; i >= 0; i--) {
+      const m = shown[i]!;
       if (m.type === "assistant" && m.usage)
         return onContext(promptTokens(m.usage), m.model ?? null);
     }
-  }, [messages, onContext]);
+  }, [shown, onContext]);
   const [total, setTotal] = useState(0);
   const [loading, setLoading] = useState(true);
   const [loadingHistory, setLoadingHistory] = useState(false);
@@ -209,12 +220,13 @@ function ConversationViewPane({
   const agentLinks = useMemo<AgentLinks>(() => {
     const ids = new Map<string, string>();
     for (const tr of transcripts) if (tr.tool_use_id) ids.set(tr.tool_use_id, tr.id);
-    return { ids, notices: agentNotices(messages), live, render: renderAgent };
-  }, [transcripts, messages, live, renderAgent]);
+    return { ids, notices: agentNotices(shown), live, render: renderAgent };
+  }, [transcripts, shown, live, renderAgent]);
 
   // Load available transcript list (also rescanned on a short interval so
   // newly-spawned subagents appear in the dropdown without a page reload).
   useEffect(() => {
+    if (!sessionId) return;
     let cancelled = false;
     async function loadTranscripts() {
       try {
@@ -254,6 +266,12 @@ function ConversationViewPane({
     firstLineRef.current = 0;
 
     async function load() {
+      // No session yet: no history.
+      if (!sessionId) {
+        setMessages([]);
+        setLoading(false);
+        return;
+      }
       try {
         setError(null);
         setLoading(true);
@@ -297,6 +315,7 @@ function ConversationViewPane({
   //     parsed message we've seen. The server already de-overlaps via
   //     afterLine, so we can safely append.
   const fetchNewMessages = useCallback(async () => {
+    if (!sessionId) return;
     if (fetchingRef.current) {
       // Coalesce: remember a trigger arrived during this fetch and re-run
       // exactly once when the in-flight request settles.
@@ -358,7 +377,12 @@ function ConversationViewPane({
     () => throttle(() => void fetchNewMessages(), EVENT_FETCH_MS),
     [fetchNewMessages]
   );
+  const fetchPushed = useMemo(
+    () => throttle(() => void fetchNewMessages(), PUSH_FETCH_MS),
+    [fetchNewMessages]
+  );
   useEffect(() => () => fetchSoon.cancel(), [fetchSoon]);
+  useEffect(() => () => fetchPushed.cancel(), [fetchPushed]);
   useEffect(() => {
     const unsubscribe = eventBus.subscribe((msg: WSMessage) => {
       // Hidden tab or hidden pane: catching up once on show is enough.
@@ -367,13 +391,14 @@ function ConversationViewPane({
         fetchSoon();
         return;
       }
-      if (msg.type !== "new_event") return;
+      if (msg.type !== "new_event" && msg.type !== "transcript_changed") return;
       const data = msg.data as { session_id?: string };
       if (data.session_id !== sessionId) return;
-      fetchSoon();
+      if (msg.type === "transcript_changed") fetchPushed();
+      else fetchSoon();
     });
     return unsubscribe;
-  }, [sessionId, fetchSoon, onScreen]);
+  }, [sessionId, fetchSoon, fetchPushed, onScreen]);
 
   onShowRef.current = () => {
     if (document.visibilityState === "visible") void fetchNewMessages();
@@ -535,7 +560,7 @@ function ConversationViewPane({
 
   // Auto-scroll to bottom after initial load
   useEffect(() => {
-    if (!loading && messages.length > 0) {
+    if (!loading && shown.length > 0) {
       scrollToBottom();
     }
   }, [loading, scrollToBottom]); // eslint-disable-line react-hooks/exhaustive-deps
@@ -626,7 +651,7 @@ function ConversationViewPane({
             >
               {t("conversation.loading")}
             </div>
-          ) : messages.length === 0 && notStarted ? (
+          ) : shown.length === 0 && notStarted ? (
             <div className="mx-auto max-w-md py-12 text-center">
               <p className="text-sm text-fg-secondary">{t("conversation.notStarted")}</p>
               <p className="mt-2 text-xs leading-relaxed text-fg-muted">
@@ -642,7 +667,7 @@ function ConversationViewPane({
                 </button>
               )}
             </div>
-          ) : messages.length === 0 ? (
+          ) : shown.length === 0 ? (
             <div className="mx-auto max-w-md py-12 text-center">
               <p className="text-sm text-fg-secondary">{t("conversation.empty")}</p>
               <p className="mt-2 text-xs leading-relaxed text-fg-muted">
@@ -651,9 +676,10 @@ function ConversationViewPane({
             </div>
           ) : (
             <MessageList
-              messages={messages}
+              messages={shown}
               loading={false}
               onOpenFile={onOpenFile}
+              onOpenArtifact={onOpenArtifact}
               live={live}
               waiting={waiting}
               onOpenTerminal={onOpenTerminal}
diff --git a/client/src/components/conversation/MessageList.tsx b/client/src/components/conversation/MessageList.tsx
index 9b38c602..051e411c 100644
--- a/client/src/components/conversation/MessageList.tsx
+++ b/client/src/components/conversation/MessageList.tsx
@@ -112,7 +112,11 @@ import {
   Search,
   Globe,
   Wrench,
+  Image as ImageIcon,
+  Code2,
+  Table2,
 } from "lucide-react";
+import { artifactKind, isArtifact, type ArtifactKind } from "../../lib/artifactView";
 import type { TranscriptMessage, TranscriptContent, TranscriptSender } from "../../lib/types";
 import { api } from "../../lib/api";
 import { copyText } from "../../lib/clipboard";
@@ -148,6 +152,8 @@ interface MessageListProps {
   loading: boolean;
   /** Opens a file named in the "Edited N files" card. */
   onOpenFile?: (path: string) => void;
+  /** Opens an artifact card's file in the chat's artifact panel. */
+  onOpenArtifact?: (path: string) => void;
   /** The session is running in a pane: show the step in progress. */
   live?: boolean;
   /** The CLI is asking the user something (a permission prompt): the live
@@ -604,30 +610,91 @@ export function chatItems(messages: TranscriptMessage[], keys: string[]): Item[]
 
 /** Claude Desktop's end-of-turn card: the files the turn changed, each with
  *  the lines added and removed. A call still running or failed is left out. */
+const ARTIFACT_ICON: Partial<Record<ArtifactKind, typeof FileText>> = {
+  html: Globe,
+  svg: ImageIcon,
+  markdown: FileText,
+  mermaid: Workflow,
+  react: Code2,
+  csv: Table2,
+};
+
+/** Claude Desktop's artifact card: what the model made, opened beside the chat. */
+function ArtifactCard({ path, onOpen }: { path: string; onOpen?: (path: string) => void }) {
+  const { t } = useTranslation("sessions");
+  const kind = artifactKind(path);
+  const Icon = ARTIFACT_ICON[kind] ?? FileText;
+  const name = path.split(/[\\/]/).pop() || path;
+  return (
+    <button
+      type="button"
+      data-testid="artifact-card"
+      title={path}
+      disabled={!onOpen}
+      onClick={() => onOpen?.(path)}
+      draggable
+      onDragStart={(e) => dragAttachments(e, [projectAttachment(path)])}
+      className="flex w-full max-w-md items-center gap-3 rounded-xl border border-border bg-surface-1 px-3 py-2.5 text-left transition-colors enabled:hover:border-border-light enabled:hover:bg-surface-2"
+    >
+      <span className="flex h-9 w-9 flex-shrink-0 items-center justify-center rounded-lg border border-border bg-surface-2">
+        <Icon className="h-4 w-4 text-fg-secondary" />
+      </span>
+      <span className="min-w-0 flex-1">
+        <span className="block truncate text-sm font-medium text-fg-primary">
+          {name.replace(/\.[A-Za-z0-9]+$/, "")}
+        </span>
+        <span className="block truncate text-xs text-fg-muted">
+          {t(`conversation.artifact.${kind}`, { defaultValue: kind })}
+          {onOpen ? ` · ${t("conversation.artifact.open")}` : ""}
+        </span>
+      </span>
+    </button>
+  );
+}
+
 function EditedFilesCard({
   blocks,
   toolResultMap,
   onOpenFile,
+  onOpenArtifact,
+  pending = false,
 }: {
   blocks: TranscriptContent[];
   toolResultMap: Map<string, TranscriptContent>;
   onOpenFile?: (path: string) => void;
+  /** Opens an artifact (a page, document, diagram…) in the chat's own panel. */
+  onOpenArtifact?: (path: string) => void;
+  /** The turn is still working: only its artifact cards so far. */
+  pending?: boolean;
 }) {
   const { t } = useTranslation("sessions");
   // `counted`: a result carried its +/- lines (an older server sends none).
   const files = new Map<string, { added: number; removed: number; counted: boolean }>();
+  // Written whole in this turn (a Write): a component made here is an artifact.
+  const created = new Set<string>();
   for (const b of blocks) {
     const input = (b.input ?? {}) as Record<string, unknown>;
     const p = input.file_path ?? input.notebook_path;
     const r = b.id ? toolResultMap.get(b.id) : undefined;
     if (typeof p !== "string" || !r || r.is_error) continue;
+    if (b.name === "Write") created.add(p);
     const f = files.get(p) ?? { added: 0, removed: 0, counted: false };
     f.added += r.diff?.added ?? 0;
     f.removed += r.diff?.removed ?? 0;
     f.counted ||= !!r.diff;
     files.set(p, f);
   }
-  if (files.size === 0) return null;
+  const artifacts = [...files.keys()].filter((p) => isArtifact(p, created.has(p)));
+  const cards = artifacts.length > 0 && (
+    <div data-testid="artifact-cards" className="flex flex-col gap-2">
+      {artifacts.map((p) => (
+        <ArtifactCard key={p} path={p} onOpen={onOpenArtifact ?? onOpenFile} />
+      ))}
+    </div>
+  );
+  if (pending) return cards || null;
+  for (const p of artifacts) files.delete(p);
+  if (files.size === 0) return cards || null;
   let added = 0;
   let removed = 0;
   let counted = false;
@@ -643,7 +710,7 @@ function EditedFilesCard({
         <span className="text-status-danger">−{r}</span>
       </span>
     );
-  return (
+  const list = (
     <div data-testid="edited-files" className="overflow-hidden rounded-xl border border-border">
       <div
         draggable
@@ -673,11 +740,21 @@ function EditedFilesCard({
       ))}
     </div>
   );
+  return cards ? (
+    <div className="flex flex-col gap-2">
+      {cards}
+      {list}
+    </div>
+  ) : (
+    list
+  );
 }
 const EditedFiles = memo(
   EditedFilesCard,
   (p, n) =>
     p.onOpenFile === n.onOpenFile &&
+    p.onOpenArtifact === n.onOpenArtifact &&
+    p.pending === n.pending &&
     sameBlocks(p.blocks, n.blocks) &&
     sameResults(n.blocks, p.toolResultMap, n.toolResultMap)
 );
@@ -688,6 +765,7 @@ function MessageListView({
   messages,
   loading,
   onOpenFile,
+  onOpenArtifact,
   live = false,
   waiting = false,
   onOpenTerminal,
@@ -780,14 +858,16 @@ function MessageListView({
                   />
                 );
               case "edits":
-                // The card closes a finished turn; one still working has none yet.
-                if (pending && item === items[items.length - 1]) return null;
+                // The card closes a finished turn; one still working shows
+                // only the artifacts it made so far, as Claude Desktop does.
                 return (
                   <EditedFiles
                     key={item.key}
                     blocks={item.blocks}
                     toolResultMap={toolResultMap}
                     onOpenFile={onOpenFile}
+                    onOpenArtifact={onOpenArtifact}
+                    pending={!!pending && item === items[items.length - 1]}
                   />
                 );
               case "done":
diff --git a/client/src/components/conversation/__tests__/ConversationView.pushed.test.tsx b/client/src/components/conversation/__tests__/ConversationView.pushed.test.tsx
new file mode 100644
index 00000000..31a87abd
--- /dev/null
+++ b/client/src/components/conversation/__tests__/ConversationView.pushed.test.tsx
@@ -0,0 +1,74 @@
+/**
+ * @file ConversationView.pushed.test.tsx
+ * @description The server says a session's transcript grew
+ * (`transcript_changed`): the chat showing that session reads the new lines
+ * at once, and again within a short window after a burst — not at the hook
+ * throttle's 1.5 s pace — while another session's notice fetches nothing.
+ * @author Nguyễn Ngọc Trí Vĩ <vinnt@smartgift.vn>
+ */
+
+import { act, render, screen } from "@testing-library/react";
+import { afterEach, describe, expect, it, vi } from "vitest";
+import { ConversationView } from "../ConversationView";
+import { api } from "../../../lib/api";
+import { eventBus } from "../../../lib/eventBus";
+
+vi.mock("../../../lib/api", async (importOriginal) => {
+  const actual = await importOriginal<Record<string, unknown>>();
+  return {
+    ...actual,
+    api: {
+      sessions: {
+        transcripts: vi.fn(async () => ({ transcripts: [] })),
+        transcript: vi.fn(async () => ({
+          messages: [
+            {
+              type: "user",
+              content: [{ type: "text", text: "hello" }],
+              timestamp: "2026-10-21T00:00:00Z",
+            },
+          ],
+          total: 1,
+          has_more: false,
+          first_line: 1,
+          last_line: 1,
+        })),
+      },
+    },
+  };
+});
+
+afterEach(() => {
+  vi.useRealTimers();
+  vi.clearAllMocks();
+});
+
+const pushed = (session_id: string) =>
+  act(() => eventBus.publish({ type: "transcript_changed", data: { session_id } } as never));
+const calls = () => vi.mocked(api.sessions.transcript).mock.calls.length;
+
+describe("ConversationView on transcript_changed", () => {
+  it("reads its own session's new lines at once, and a burst within 300 ms", async () => {
+    render(<ConversationView sessionId="s1" fill live />);
+    await screen.findByText("hello");
+    vi.useFakeTimers();
+    const before = calls();
+    pushed("s1");
+    expect(calls()).toBe(before + 1);
+    pushed("s1");
+    pushed("s1");
+    await act(async () => {
+      await vi.advanceTimersByTimeAsync(350);
+    });
+    expect(calls()).toBe(before + 2);
+  });
+
+  it("ignores another session's transcript", async () => {
+    render(<ConversationView sessionId="s1" fill live />);
+    await screen.findByText("hello");
+    const before = calls();
+    pushed("s2");
+    await act(async () => {});
+    expect(calls()).toBe(before);
+  });
+});
diff --git a/client/src/components/conversation/__tests__/MessageList.edits.test.tsx b/client/src/components/conversation/__tests__/MessageList.edits.test.tsx
index a18fdf52..6dced3fa 100644
--- a/client/src/components/conversation/__tests__/MessageList.edits.test.tsx
+++ b/client/src/components/conversation/__tests__/MessageList.edits.test.tsx
@@ -46,7 +46,7 @@ describe("the Edited N files card", () => {
           user("fix it"),
           call("1", "Edit", "/w/a.ts"),
           result("1", { added: 3, removed: 1 }),
-          call("2", "Write", "/w/b.md"),
+          call("2", "Write", "/w/b.json"),
           result("2", { added: 10, removed: 0 }),
           call("3", "Edit", "/w/a.ts"),
           result("3", { added: 2, removed: 2 }),
@@ -65,8 +65,35 @@ describe("the Edited N files card", () => {
     expect(within(card).getByText("+15")).toBeTruthy();
     expect(within(card).getByText("+5")).toBeTruthy(); // a.ts: 3+2 added
     expect(within(card).queryByText("c.ts")).toBeNull(); // failed
-    fireEvent.click(within(card).getByTitle("/w/b.md"));
-    expect(onOpenFile).toHaveBeenCalledWith("/w/b.md");
+    fireEvent.click(within(card).getByTitle("/w/b.json"));
+    expect(onOpenFile).toHaveBeenCalledWith("/w/b.json");
+  });
+
+  it("shows a page or document the turn made as an artifact card, opened in the chat's panel", () => {
+    const onOpenFile = vi.fn();
+    const onOpenArtifact = vi.fn();
+    render(
+      <MessageList
+        loading={false}
+        onOpenFile={onOpenFile}
+        onOpenArtifact={onOpenArtifact}
+        messages={[
+          user("make a page"),
+          call("1", "Write", "/w/page.html"),
+          result("1", { added: 20, removed: 0 }),
+          call("2", "Edit", "/w/App.tsx"),
+          result("2", { added: 1, removed: 1 }),
+          user("thanks"),
+        ]}
+      />
+    );
+    const cards = screen.getAllByTestId("artifact-card");
+    // A .tsx it only edited is project code, not an artifact.
+    expect(cards).toHaveLength(1);
+    expect(cards[0]!.textContent).toContain("page");
+    fireEvent.click(cards[0]!);
+    expect(onOpenArtifact).toHaveBeenCalledWith("/w/page.html");
+    expect(within(screen.getByTestId("edited-files")).getByTitle("/w/App.tsx")).toBeTruthy();
   });
 
   it("drags a row as its @path attachment, and the header as every file", () => {
@@ -77,7 +104,7 @@ describe("the Edited N files card", () => {
           user("fix it"),
           call("1", "Edit", "/w/a.ts"),
           result("1", { added: 1, removed: 0 }),
-          call("2", "Write", "/w/b.md"),
+          call("2", "Write", "/w/b.json"),
           result("2", { added: 2, removed: 0 }),
         ]}
       />
@@ -90,14 +117,14 @@ describe("the Edited N files card", () => {
       });
       return data;
     };
-    const one = drag(within(card).getByTitle("/w/b.md"));
+    const one = drag(within(card).getByTitle("/w/b.json"));
     expect(JSON.parse(one["application/x-agenthub-attach"]!)).toEqual([
-      { path: "@/w/b.md", name: "b.md" },
+      { path: "@/w/b.json", name: "b.json" },
     ]);
-    expect(one["text/plain"]).toBe("/w/b.md");
+    expect(one["text/plain"]).toBe("/w/b.json");
     const all = drag(within(card).getByText("Edited 2 files"));
     expect(
       JSON.parse(all["application/x-agenthub-attach"]!).map((a: { path: string }) => a.path)
-    ).toEqual(["@/w/a.ts", "@/w/b.md"]);
+    ).toEqual(["@/w/a.ts", "@/w/b.json"]);
   });
 });
diff --git a/client/src/components/run/ChatFilePreview.tsx b/client/src/components/run/ChatFilePreview.tsx
index 07798a6c..e254df9d 100644
--- a/client/src/components/run/ChatFilePreview.tsx
+++ b/client/src/components/run/ChatFilePreview.tsx
@@ -18,15 +18,40 @@
  * @author Nguyễn Ngọc Trí Vĩ <vinnt@smartgift.vn>
  */
 
-import { useEffect, useRef, useState, type DragEvent } from "react";
+import { useEffect, useMemo, useRef, useState, type DragEvent } from "react";
 import { createPortal } from "react-dom";
 import { useTranslation } from "react-i18next";
-import { Copy, FileText, Loader2, X } from "lucide-react";
+import {
+  ChevronLeft,
+  ChevronRight,
+  Copy,
+  Download,
+  FileText,
+  Loader2,
+  Maximize2,
+  Minimize2,
+  Share2,
+  X,
+} from "lucide-react";
 import { api, gitTarget } from "../../lib/api";
 import { copyText } from "../../lib/clipboard";
 import { useDialog } from "../../hooks/useDialog";
 import { dragSkill } from "../../lib/promptAttach";
-import { MarkdownPreview, isMarkdown } from "../MarkdownPreview";
+import { MarkdownPreview } from "../MarkdownPreview";
+import { CodeBlock } from "../conversation/CodeBlock";
+import { useTheme } from "../../hooks/useTheme";
+import { toast } from "../../lib/toast";
+import type { TranscriptMessage } from "../../lib/types";
+import {
+  RENDERED,
+  artifactDocument,
+  artifactKind,
+  artifactLang,
+  artifactVersions,
+  frameSandbox,
+  parseCsv,
+  type ArtifactKind,
+} from "../../lib/artifactView";
 
 /** Where a Chat pane's files can be read from. */
 export interface PreviewScope {
@@ -146,33 +171,89 @@ function useContent(path: string, scope: PreviewScope): Content {
   return content;
 }
 
-function Body({ path, scope }: { path: string; scope: PreviewScope }) {
+/** A CSV/TSV as a table (the first rows; a file is not a spreadsheet). */
+function CsvTable({ text, tab }: { text: string; tab: boolean }) {
+  const rows = useMemo(() => parseCsv(text, tab ? "\t" : ",").slice(0, 1000), [text, tab]);
+  const [head, ...body] = rows;
+  return (
+    <div className="min-h-0 flex-1 overflow-auto p-3">
+      <table className="w-full border-collapse text-xs">
+        {head && (
+          <thead className="sticky top-0 bg-surface-2">
+            <tr>
+              {head.map((h, i) => (
+                <th
+                  key={i}
+                  className="border border-border px-2 py-1 text-left font-medium text-fg-primary"
+                >
+                  {h}
+                </th>
+              ))}
+            </tr>
+          </thead>
+        )}
+        <tbody>
+          {body.map((r, i) => (
+            <tr key={i} className="odd:bg-surface-1 even:bg-surface-2/40">
+              {r.map((c, j) => (
+                <td key={j} className="border border-border px-2 py-1 text-fg-secondary">
+                  {c}
+                </td>
+              ))}
+            </tr>
+          ))}
+        </tbody>
+      </table>
+    </div>
+  );
+}
+
+/** An HTML, SVG, Mermaid or React artifact, running in its own sandbox:
+ *  scripts (none for an SVG) but never the dashboard's origin. */
+function ArtifactFrame({ kind, text, title }: { kind: ArtifactKind; text: string; title: string }) {
+  const { theme } = useTheme();
+  const doc = useMemo(() => artifactDocument(kind, text, theme === "dark"), [kind, text, theme]);
+  return (
+    <iframe
+      data-testid="artifact-frame"
+      title={title}
+      srcDoc={doc ?? ""}
+      sandbox={frameSandbox(kind)}
+      referrerPolicy="no-referrer"
+      className={`min-h-0 w-full flex-1 border-0 ${kind === "html" || kind === "react" ? "bg-white" : ""}`}
+    />
+  );
+}
+
+type View = "preview" | "code";
+
+function Body({ path, content, view }: { path: string; content: Content; view: View }) {
   const { t } = useTranslation("run");
-  const loaded = useContent(path, scope);
+  const kind = artifactKind(path);
   // An image the route refused (gone, server mode) shows like no preview.
   const [broken, setBroken] = useState<string | null>(null);
-  const content: Content =
-    loaded.kind === "image" && loaded.src === broken ? { kind: "none" } : loaded;
-  if (content.kind === "loading") {
+  const shown: Content =
+    content.kind === "image" && content.src === broken ? { kind: "none" } : content;
+  if (shown.kind === "loading") {
     return (
       <div className="flex flex-1 items-center justify-center">
         <Loader2 className="h-5 w-5 animate-spin text-fg-muted" />
       </div>
     );
   }
-  if (content.kind === "image") {
+  if (shown.kind === "image") {
     return (
       <div className="flex flex-1 items-center justify-center overflow-auto p-4">
         <img
-          src={content.src}
+          src={shown.src}
           alt={baseName(path)}
-          onError={() => setBroken(content.src)}
+          onError={() => setBroken(shown.src)}
           className="max-h-full max-w-full object-contain"
         />
       </div>
     );
   }
-  if (content.kind === "none") {
+  if (shown.kind === "none") {
     return (
       <div className="flex flex-1 flex-col items-center justify-center gap-2 p-6 text-center text-sm text-fg-muted">
         <FileText className="h-8 w-8" />
@@ -181,69 +262,244 @@ function Body({ path, scope }: { path: string; scope: PreviewScope }) {
       </div>
     );
   }
-  return (
-    <div className="min-h-0 flex-1 overflow-auto">
-      {isMarkdown(path) ? (
-        <MarkdownPreview text={content.text} className="p-4" />
-      ) : (
-        <pre className="p-4 font-mono text-xs leading-relaxed text-fg-secondary">
-          {content.text}
-        </pre>
-      )}
-      {content.truncated && (
-        <p className="px-4 pb-4 text-xs text-fg-muted">{t("chat.outline.truncated")}</p>
-      )}
-    </div>
+  const note = shown.truncated && (
+    <p className="px-4 pb-4 text-xs text-fg-muted">{t("chat.outline.truncated")}</p>
   );
+  if (view === "code" || !RENDERED.has(kind)) {
+    return (
+      <div className="min-h-0 flex-1 overflow-auto p-3">
+        <CodeBlock code={shown.text} lang={artifactLang(path)} maxHeight="none" showLineNumbers />
+        {note}
+      </div>
+    );
+  }
+  if (kind === "markdown") {
+    return (
+      <div className="min-h-0 flex-1 overflow-auto">
+        <MarkdownPreview text={shown.text} className="p-4" />
+        {note}
+      </div>
+    );
+  }
+  if (kind === "csv") return <CsvTable text={shown.text} tab={/\.tsv$/i.test(path)} />;
+  return <ArtifactFrame kind={kind} text={shown.text} title={baseName(path)} />;
 }
 
-/** One file: header (name · EXT, Copy path, Close) over its content. */
+const iconBtn =
+  "flex h-7 w-7 items-center justify-center rounded text-fg-muted hover:bg-surface-3 hover:text-fg-primary disabled:opacity-40";
+
+/** Saves text (or an image's URL) as a file through the browser. */
+function download(name: string, content: Content) {
+  const a = document.createElement("a");
+  if (content.kind === "image") a.href = content.src;
+  else if (content.kind === "text")
+    a.href = URL.createObjectURL(new Blob([content.text], { type: "text/plain;charset=utf-8" }));
+  else return;
+  a.download = name;
+  a.click();
+  if (content.kind === "text") setTimeout(() => URL.revokeObjectURL(a.href), 1000);
+}
+
+/** One file, Claude Desktop's artifact view: name · EXT, Preview | Code, the
+ *  versions the conversation made of it, Copy, Download, Share to the
+ *  Artifacts Hub, full screen, Close. */
 export function FilePreview({
   path,
   scope,
   onClose,
   className = "",
+  messages,
 }: {
   path: string;
   scope: PreviewScope;
   onClose: () => void;
   className?: string;
+  /** The conversation, for the versions its Write/Edit calls left. */
+  messages?: TranscriptMessage[];
 }) {
   const { t } = useTranslation("run");
-  return (
+  const kind = artifactKind(path);
+  const loaded = useContent(path, scope);
+  const [view, setView] = useState<View>(RENDERED.has(kind) ? "preview" : "code");
+  const [full, setFull] = useState(false);
+  const [sharing, setSharing] = useState(false);
+  const versions = useMemo(
+    () => (messages ? artifactVersions(messages, path) : []),
+    [messages, path]
+  );
+  // null: the newest — the file as it is now.
+  const [picked, setPicked] = useState<number | null>(null);
+  // Another file opens on its own default tab, at its newest version.
+  useEffect(() => {
+    setPicked(null);
+    setView(RENDERED.has(artifactKind(path)) ? "preview" : "code");
+  }, [path]);
+  const last = Math.max(versions.length - 1, 0);
+  const at = picked ?? last;
+  const older = picked !== null && picked < last ? versions[picked] : null;
+  const content: Content =
+    older && older.content !== null
+      ? { kind: "text", text: older.content, truncated: false }
+      : loaded;
+  const known = (i: number) => i === last || versions[i]?.content != null;
+  const step = (dir: -1 | 1) => {
+    for (let i = at + dir; i >= 0 && i <= last; i += dir) {
+      if (known(i)) return setPicked(i === last ? null : i);
+    }
+  };
+  const share = async () => {
+    if (sharing) return;
+    setSharing(true);
+    try {
+      const file = slashed(path);
+      const base = slashed(scope.laneCwd).replace(/\/+$/, "");
+      if (scope.laneId !== null && base && file.startsWith(`${base}/`) && picked === null) {
+        await api.artifacts.pinFile({
+          laneId: scope.laneId,
+          path: file.slice(base.length + 1),
+          title: displayName(path),
+          ...(scope.sessionId ? { sessionId: scope.sessionId } : {}),
+        });
+      } else if (content.kind === "text") {
+        await api.artifacts.uploadFile(new File([content.text], displayName(path)), {
+          title: displayName(path),
+        });
+      } else throw new Error(t("chat.artifact.cannotShare"));
+      toast.success(t("chat.artifact.shared"));
+    } catch (err) {
+      toast.error(err);
+    } finally {
+      setSharing(false);
+    }
+  };
+  const panel = (
     <section
       data-testid="chat-file-preview"
-      className={`flex min-h-0 min-w-0 flex-col overflow-hidden bg-surface-1 ${className}`}
+      className={
+        full
+          ? "fixed inset-4 z-[200] flex min-h-0 min-w-0 flex-col overflow-hidden rounded-xl border border-border bg-surface-1 shadow-2xl"
+          : `flex min-h-0 min-w-0 flex-col overflow-hidden bg-surface-1 ${className}`
+      }
     >
-      <header className="flex items-center gap-2 border-b border-border px-4 py-2.5">
-        <span className="min-w-0 truncate text-sm text-fg-primary" title={path}>
+      <header className="flex items-center gap-2 border-b border-border px-3 py-2">
+        <button
+          type="button"
+          onClick={() => void copyText(path)}
+          title={t("chat.outline.copyPath", { path })}
+          className="min-w-0 truncate text-left text-sm text-fg-primary hover:underline"
+        >
           {displayName(path).replace(/\.[A-Za-z0-9]+$/, "")}
-        </span>
-        <span className="text-xs text-fg-muted">{extOf(path)}</span>
-        <div className="ml-auto flex items-center gap-1">
+        </button>
+        <span className="flex-shrink-0 text-xs text-fg-muted">{extOf(path)}</span>
+        {versions.length > 1 && (
+          <span
+            data-testid="artifact-versions"
+            className="flex flex-shrink-0 items-center text-xs text-fg-muted"
+          >
+            <button
+              type="button"
+              onClick={() => step(-1)}
+              disabled={![...Array(at).keys()].some(known)}
+              aria-label={t("chat.artifact.older")}
+              title={t("chat.artifact.older")}
+              className={iconBtn}
+            >
+              <ChevronLeft className="h-3.5 w-3.5" />
+            </button>
+            <span className="tabular-nums">
+              {t("chat.artifact.version", { n: at + 1, total: versions.length })}
+            </span>
+            <button
+              type="button"
+              onClick={() => step(1)}
+              disabled={at >= last}
+              aria-label={t("chat.artifact.newer")}
+              title={t("chat.artifact.newer")}
+              className={iconBtn}
+            >
+              <ChevronRight className="h-3.5 w-3.5" />
+            </button>
+          </span>
+        )}
+        <div className="ml-auto flex flex-shrink-0 items-center gap-1">
+          {RENDERED.has(kind) && kind !== "image" && (
+            <div role="tablist" className="mr-1 flex rounded-md border border-border p-0.5 text-xs">
+              {(["preview", "code"] as const).map((v) => (
+                <button
+                  key={v}
+                  type="button"
+                  role="tab"
+                  aria-selected={view === v}
+                  data-testid={`artifact-tab-${v}`}
+                  onClick={() => setView(v)}
+                  className={`rounded px-2 py-0.5 ${view === v ? "bg-surface-3 text-fg-primary" : "text-fg-muted hover:text-fg-primary"}`}
+                >
+                  {t(`chat.artifact.${v}`)}
+                </button>
+              ))}
+            </div>
+          )}
           <button
             type="button"
-            onClick={() => void copyText(path)}
-            title={t("chat.outline.copyPath", { path })}
-            aria-label={t("chat.outline.copyPath", { path })}
-            className="flex h-7 w-7 items-center justify-center rounded text-fg-muted hover:bg-surface-3 hover:text-fg-primary"
+            data-testid="artifact-copy"
+            onClick={() => void copyText(content.kind === "text" ? content.text : path)}
+            title={t("chat.artifact.copy")}
+            aria-label={t("chat.artifact.copy")}
+            className={iconBtn}
           >
             <Copy className="h-3.5 w-3.5" />
           </button>
+          <button
+            type="button"
+            onClick={() => download(displayName(path), content)}
+            disabled={content.kind !== "text" && content.kind !== "image"}
+            title={t("chat.artifact.download")}
+            aria-label={t("chat.artifact.download")}
+            className={iconBtn}
+          >
+            <Download className="h-3.5 w-3.5" />
+          </button>
+          <button
+            type="button"
+            data-testid="artifact-share"
+            onClick={() => void share()}
+            disabled={sharing || content.kind === "loading" || content.kind === "none"}
+            title={t("chat.artifact.share")}
+            aria-label={t("chat.artifact.share")}
+            className={iconBtn}
+          >
+            {sharing ? (
+              <Loader2 className="h-3.5 w-3.5 animate-spin" />
+            ) : (
+              <Share2 className="h-3.5 w-3.5" />
+            )}
+          </button>
+          <button
+            type="button"
+            data-testid="artifact-expand"
+            onClick={() => setFull((f) => !f)}
+            title={t(full ? "chat.artifact.collapse" : "chat.artifact.expand")}
+            aria-label={t(full ? "chat.artifact.collapse" : "chat.artifact.expand")}
+            className={iconBtn}
+          >
+            {full ? <Minimize2 className="h-3.5 w-3.5" /> : <Maximize2 className="h-3.5 w-3.5" />}
+          </button>
           <button
             type="button"
             onClick={onClose}
             title={t("chat.outline.close")}
             aria-label={t("chat.outline.close")}
-            className="flex h-7 w-7 items-center justify-center rounded text-fg-muted hover:bg-surface-3 hover:text-fg-primary"
+            className={iconBtn}
           >
             <X className="h-4 w-4" />
           </button>
         </div>
       </header>
-      <Body path={path} scope={scope} />
+      <Body path={path} content={content} view={view} />
     </section>
   );
+  // Full screen: over the whole page, where a pane's edges cannot clip it.
+  return full ? createPortal(panel, document.body) : panel;
 }
 
 /** Claude's Uploads dialog: the files on the left, the picked one on the right. */
diff --git a/client/src/components/run/LaneConsolePane.tsx b/client/src/components/run/LaneConsolePane.tsx
index 04f3e536..4879349f 100644
--- a/client/src/components/run/LaneConsolePane.tsx
+++ b/client/src/components/run/LaneConsolePane.tsx
@@ -12,7 +12,8 @@
  * new lane's own live session for its slot, or an empty pane — never the
  * previous lane's terminal. `attachRequest` (a run picked in the sidebar or a
  * tab) attaches to that tmux session only; nothing is started or resumed.
- * A Claude or Codex pane also has a Chat view (`PaneChat`): the session's
+ * A pane of a `CHAT_TOOLS` tool (Claude, Codex, omp, Cursor, Antigravity) also
+ * has a Chat view (`PaneChat`): the session's
  * conversation laid OVER the terminal, which stays mounted and attached under
  * it. Chat or terminal is each SESSION's own (`useSessionChat`, the toggle
  * left of the pane's split buttons); the Chat UI setting (Settings →
@@ -29,9 +30,7 @@ import type { Lane } from "../../lib/types";
 import { useAiTools } from "../../hooks/useAiTools";
 import { TerminalView } from "./TerminalView";
 import { PaneChat } from "./PaneChat";
-import { useSessionChat } from "../../lib/chatUi";
-
-const CHAT_TOOLS = new Set(["claude", "codex"]);
+import { CHAT_TOOLS, useSessionChat } from "../../lib/chatUi";
 
 export interface LaneConsolePaneProps {
   lanes: Lane[];
diff --git a/client/src/components/run/PaneChat.tsx b/client/src/components/run/PaneChat.tsx
index b6a2d099..bb2bb9d2 100644
--- a/client/src/components/run/PaneChat.tsx
+++ b/client/src/components/run/PaneChat.tsx
@@ -15,7 +15,8 @@
  * in the box beside `+`. A menu the CLI asks with — AskUserQuestion, a
  * permission or trust prompt, Codex's approval — is read off the terminal
  * underneath (`readRunMenu`) and shown in the prompt box's place as buttons
- * that press its keys there (`PaneMenuCard`); its "Type something" row opens
+ * that press its keys there (`PaneMenuCard`) — a permission prompt with its tool, command and a deny-with-note
+ * (`PermissionMenuCard`); its "Type something" row opens
  * a field in the card, whose text is typed there as the answer
  * (`answerRunMenu`).
  * It also has Claude Desktop's activity panel (`ChatOutline`: progress,
@@ -56,7 +57,17 @@
 
 import { memo, useCallback, useEffect, useRef, useState, type CSSProperties } from "react";
 import { useTranslation } from "react-i18next";
-import { ArrowUp, FileText, Square, SquareCheck } from "lucide-react";
+import {
+  ArrowUp,
+  Check,
+  ChevronLeft,
+  ChevronRight,
+  FileText,
+  MessageCircleQuestion,
+  ShieldQuestion,
+  Square,
+  SquareCheck,
+} from "lucide-react";
 import { isSubmitEnter } from "../../lib/keys";
 import { ConversationView } from "../conversation/ConversationView";
 import {
@@ -77,7 +88,11 @@ import { Divider, showTerminal } from "./PaneGrid";
 import {
   answerRunMenu,
   chooseRunOption,
+  notePaneFocus,
+  onPaneOutput,
+  stepRunMenu,
   interruptRun,
+  readRunBusy,
   readRunMenu,
   submitToRun,
   type PaneMenu,
@@ -89,7 +104,9 @@ import { toast } from "../../lib/toast";
 import { confirmDialog } from "../../lib/confirm";
 import { setTabBadge } from "../../lib/tabBadge";
 import { fmt } from "../../lib/format";
-import { agentAwaitingReason } from "../../lib/types";
+import { agentAwaitingReason, type TranscriptMessage, type WSMessage } from "../../lib/types";
+import { eventBus } from "../../lib/eventBus";
+import { throttle } from "../../lib/throttle";
 import { api, type EffortLevel, type PermissionMode, type RunHandle } from "../../lib/api";
 
 const MODES: [PermissionMode, string][] = [
@@ -100,6 +117,11 @@ const MODES: [PermissionMode, string][] = [
   ["bypassPermissions", "Bypass"],
 ];
 const isSettable = (m: unknown): m is PermissionMode => MODES.some(([v]) => v === m);
+// Pane output re-reads the transcript at most this often: a TUI spinner
+// redraws several times a second, and each read walks the transcript file.
+const OUTPUT_FETCH_MS = 1000;
+/** A tool with no known busy line counts as at work this long after its pane last drew. */
+const OUTPUT_BUSY_MS = 2000;
 const SIDE_KEY = "agenthub.chat.panelWidth";
 const clampSide = (f: number) => Math.round(Math.min(0.7, Math.max(0.2, f)) * 100) / 100;
 const outlineKey = (sessionId: string) => `agenthub.chat.panel.${sessionId}`;
@@ -117,6 +139,7 @@ const storedSide = () => {
  *  (LiteLLM's list, 2026-10). Null for another vendor's model. */
 // ponytail: a fixed rule, not a table; read LiteLLM's max_input_tokens if a new 200K model ships.
 export function contextWindow(model: string | null): number | null {
+  // A Codex/omp chat reports its own window (ContextMeter's `window`).
   if (!model?.startsWith("claude-")) return null;
   return /^claude-(?:[23]|instant|(?:opus|sonnet|haiku)-4(?:-[0-5])?(?:-\d{8})?$)/.test(model)
     ? 200_000
@@ -125,7 +148,13 @@ export function contextWindow(model: string | null): number | null {
 
 /** How full the context is: a ring and a percentage, amber then red as
  *  Claude Code's auto-compact comes near. */
-function ContextMeter({ tokens, model }: { tokens: number; model: string | null }) {
+function ContextMeter({
+  tokens,
+  model,
+}: {
+  tokens: number;
+  model: string | null;
+}) {
   const { t } = useTranslation("run");
   const size = contextWindow(model);
   if (!size || !tokens) return null;
@@ -212,33 +241,222 @@ function PermissionChip({ runId, initial }: { runId: string; initial: Permission
   );
 }
 
+/** A permission prompt (Claude's tool approval, Codex's command approval,
+ *  the folder-trust question): it offers a Yes and a No. */
+export function isPermissionMenu(menu: PaneMenu): boolean {
+  if (menu.multi || menu.tabs) return false;
+  const labels = menu.options.map((o) => o.label);
+  return labels.some((l) => /^Yes\b/i.test(l)) && labels.some((l) => /^No\b/i.test(l));
+}
+
+/** A menu row's label without its key hint (`Yes, proceed (y)`). */
+const bareLabel = (label: string) => label.replace(/\s+\((?:[a-z]|esc|shift\+tab)\)$/i, "");
+
+/** A permission prompt read off the terminal: the tool, what it will run, and the CLI's own
+ *  answers as buttons — each still presses its key in the pane. "No, and
+ *  tell … differently" also takes a note, sent as the next message once the
+ *  CLI is back at its prompt. */
+function PermissionMenuCard({
+  menu,
+  onChoose,
+  onDenyNote,
+  onCancel,
+}: {
+  menu: PaneMenu;
+  onChoose: (index: number) => void;
+  onDenyNote: (index: number, note: string) => void;
+  onCancel: () => void;
+}) {
+  const { t } = useTranslation("run");
+  const [noting, setNoting] = useState(false);
+  const [note, setNote] = useState("");
+  const q = menu.question.split("\n");
+  const title = menu.detail?.[0] ?? q[0] ?? "";
+  const body = (menu.detail ? menu.detail.slice(1) : q.slice(1)).join("\n");
+  const ask = menu.detail ? menu.question : "";
+  const yes = menu.options.findIndex((o) => /^Yes\b/i.test(o.label));
+  const tell = menu.options.findIndex((o) => /^No, and tell\b/i.test(o.label));
+  const sendNote = () => {
+    if (note.trim() && tell >= 0) onDenyNote(tell, note.trim());
+  };
+  return (
+    <div data-testid="pane-chat-permission" className="flex flex-col gap-2">
+      <div className="flex items-start gap-2">
+        <ShieldQuestion className="mt-0.5 h-4 w-4 flex-shrink-0 text-status-warning" />
+        <div className="min-w-0 flex-1">
+          <p className="whitespace-pre-wrap break-words text-sm font-medium text-fg-primary">
+            {title}
+          </p>
+          {ask && <p className="text-xs text-fg-muted">{ask}</p>}
+        </div>
+      </div>
+      {body && (
+        <pre className="max-h-40 overflow-auto whitespace-pre-wrap break-all rounded-lg border border-border bg-surface-2 px-3 py-2 font-mono text-xs text-fg-secondary">
+          {body}
+        </pre>
+      )}
+      {noting ? (
+        <div className="flex items-center gap-2 rounded-lg border border-accent/50 bg-surface-2 px-3 py-1">
+          <input
+            autoFocus
+            data-testid="pane-chat-permission-note"
+            value={note}
+            onChange={(e) => setNote(e.target.value)}
+            onKeyDown={(e) => {
+              if (isSubmitEnter(e)) {
+                e.preventDefault();
+                sendNote();
+              } else if (e.key === "Escape") setNoting(false);
+            }}
+            placeholder={t("chat.request.notePlaceholder")}
+            aria-label={t("chat.request.notePlaceholder")}
+            className="min-w-0 flex-1 bg-transparent py-0.5 text-sm text-fg-primary outline-none placeholder:text-fg-muted"
+          />
+          <button
+            type="button"
+            onClick={sendNote}
+            disabled={!note.trim()}
+            className="btn-secondary text-xs disabled:opacity-40"
+          >
+            {t("chat.request.send")}
+          </button>
+        </div>
+      ) : (
+        <div className="flex flex-wrap items-center gap-2">
+          {menu.options.map((o, i) => (
+            <button
+              key={`${i}-${o.label}`}
+              type="button"
+              data-testid="pane-chat-permission-option"
+              autoFocus={i === yes}
+              onClick={() => onChoose(i)}
+              title={o.hint || undefined}
+              className={`${i === yes ? "btn-primary" : "btn-secondary"} max-w-full break-words text-left text-xs`}
+            >
+              {bareLabel(o.label)}
+            </button>
+          ))}
+          {tell >= 0 && (
+            <button
+              type="button"
+              data-testid="pane-chat-permission-tell"
+              onClick={() => setNoting(true)}
+              className="text-xs text-fg-muted underline-offset-2 hover:text-fg-primary hover:underline"
+            >
+              {t("chat.request.denyWithNote")}
+            </button>
+          )}
+          <button
+            type="button"
+            onClick={onCancel}
+            className="ml-auto rounded-md border border-border px-2 py-0.5 text-xs text-fg-muted hover:bg-surface-3"
+          >
+            {t("chat.menu.cancel")}
+          </button>
+        </div>
+      )}
+    </div>
+  );
+}
+
 /** The menu the CLI asks with, as buttons: a pick presses its key in the
  *  pane; its "Type something" row opens a field right there instead —
- *  pressing that row's digit again would type the digit into the CLI's own. */
+ *  pressing that row's digit again would type the digit into the CLI's own.
+ *  A permission prompt takes its own layout (`PermissionMenuCard`), any
+ *  other menu the question layout. */
 export function PaneMenuCard({
   menu,
   onChoose,
   onAnswer,
   onCancel,
+  onStep,
+  onDenyNote,
 }: {
   menu: PaneMenu;
   onChoose: (index: number) => void;
   onAnswer: (text: string) => void;
   onCancel: () => void;
+  /** Previous (-1) / next (1) question of a multi-question AskUserQuestion. */
+  onStep?: (dir: -1 | 1) => void;
+  /** A permission prompt's "No, and tell …", with the note to send after it. */
+  onDenyNote?: (index: number, note: string) => void;
 }) {
   const { t } = useTranslation("run");
   const [typing, setTyping] = useState(false);
   const [draft, setDraft] = useState("");
   const submit = () => draft.trim() && onAnswer(draft);
+  if (isPermissionMenu(menu))
+    return (
+      <div
+        data-testid="pane-chat-menu"
+        className="rounded-2xl border border-border bg-surface-1 p-3 shadow-sm"
+      >
+        <PermissionMenuCard
+          menu={menu}
+          onChoose={onChoose}
+          onDenyNote={onDenyNote ?? ((i) => onChoose(i))}
+          onCancel={onCancel}
+        />
+      </div>
+    );
   return (
     <div
       data-testid="pane-chat-menu"
       className="rounded-2xl border border-border bg-surface-1 p-3 shadow-sm"
     >
+      {menu.tabs && (
+        <div
+          data-testid="pane-chat-menu-tabs"
+          className="mb-2 flex flex-wrap items-center gap-1 text-xs"
+        >
+          {onStep && (
+            <button
+              type="button"
+              onClick={() => onStep(-1)}
+              aria-label={t("chat.menu.prev")}
+              title={t("chat.menu.prev")}
+              className="rounded-md border border-border p-0.5 text-fg-muted hover:bg-surface-3"
+            >
+              <ChevronLeft className="h-3.5 w-3.5" />
+            </button>
+          )}
+          {menu.tabs.map((tab, i) => (
+            <span
+              key={`${i}-${tab.label}`}
+              className={`flex items-center gap-1 rounded-md border px-1.5 py-0.5 ${
+                i === menu.tab
+                  ? "border-accent/50 bg-surface-2 text-fg-primary"
+                  : "border-border text-fg-muted"
+              }`}
+            >
+              {tab.done ? (
+                <Check className="h-3 w-3 text-accent" />
+              ) : (
+                <Square className="h-3 w-3" />
+              )}
+              {tab.label}
+            </span>
+          ))}
+          {onStep && (
+            <button
+              type="button"
+              onClick={() => onStep(1)}
+              aria-label={t("chat.menu.next")}
+              title={t("chat.menu.next")}
+              className="rounded-md border border-border p-0.5 text-fg-muted hover:bg-surface-3"
+            >
+              <ChevronRight className="h-3.5 w-3.5" />
+            </button>
+          )}
+        </div>
+      )}
       {menu.question && (
-        <p className="mb-2 whitespace-pre-wrap break-words text-sm font-medium text-fg-primary">
-          {menu.question}
-        </p>
+        <div className="mb-2 flex items-start gap-2">
+          <MessageCircleQuestion className="mt-0.5 h-4 w-4 flex-shrink-0 text-accent" />
+          <p className="whitespace-pre-wrap break-words text-sm font-medium text-fg-primary">
+            {menu.question}
+          </p>
+        </div>
       )}
       <div className="flex flex-col gap-1">
         {menu.options.map((o, i) =>
@@ -338,9 +556,10 @@ export const PaneChat = memo(function PaneChat({
   const runId = run.id;
   const claude = run.tool === "claude";
   // The unsent prompt outlives the terminal view and a reload (`lib/chatDraft`).
-  const [prompt, setPrompt] = useState(() => readChatDraft(sessionId).prompt);
-  const [images, setImages] = useState<PromptImage[]>(() => readChatDraft(sessionId).images);
-  useEffect(() => writeChatDraft(sessionId, prompt, images), [sessionId, prompt, images]);
+  const draftKey = sessionId;
+  const [prompt, setPrompt] = useState(() => readChatDraft(draftKey).prompt);
+  const [images, setImages] = useState<PromptImage[]>(() => readChatDraft(draftKey).images);
+  useEffect(() => writeChatDraft(draftKey, prompt, images), [draftKey, prompt, images]);
   const [uploading, setUploading] = useState(false);
   const [follow, setFollow] = useState(0);
   const [model, setModel] = useState(run.model ?? "");
@@ -416,6 +635,46 @@ export const PaneChat = memo(function PaneChat({
   useEffect(() => {
     void readWorking();
   }, [readWorking]);
+  // The terminal underneath drew something: the CLI is at work, so its
+  // transcript may have grown — read it now, for any tool, hooks or not. A
+  // spinner redraws all the time, hence the throttle.
+  useEffect(() => {
+    if (!sessionId) return;
+    const poke = throttle(
+      () =>
+        eventBus.publish({
+          type: "transcript_changed",
+          data: { session_id: sessionId },
+        } as unknown as WSMessage),
+      OUTPUT_FETCH_MS
+    );
+    const off = onPaneOutput(runId, poke);
+    return () => {
+      off();
+      poke.cancel();
+    };
+  }, [runId, sessionId]);
+  // The terminal's own word on a turn: its busy line (BUSY_LINE) while it
+  // shows, else — a tool whose line we do not know — output in the last
+  // OUTPUT_BUSY_MS. Hooks miss a tool that sends none (omp, Cursor, Codex
+  // without its opt-in hooks), and arrive late on a dashboard they skip.
+  const [screenBusy, setScreenBusy] = useState(false);
+  useEffect(() => {
+    let last = 0;
+    const off = onPaneOutput(runId, () => (last = Date.now()));
+    const read = () => {
+      const known = readRunBusy(runId, run.tool);
+      setScreenBusy(known ?? (last > 0 && Date.now() - last < OUTPUT_BUSY_MS));
+    };
+    read();
+    const timer = window.setInterval(() => {
+      if (isShown(onScreen)) read();
+    }, 700);
+    return () => {
+      off();
+      window.clearInterval(timer);
+    };
+  }, [runId, run.tool, onScreen]);
   useLiveRefresh(readWorking, {
     throttleMs: 500,
     // While Stop shows, also poll: a missed agent_updated (the hook reached
@@ -444,6 +703,7 @@ export const PaneChat = memo(function PaneChat({
     }, 700);
     return () => window.clearInterval(timer);
   }, [runId, onScreen]);
+  const busy = working || screenBusy;
   const asking = waiting || menu !== null;
   // Gone until the pane redraws: no second press on a stale menu.
   const choose = (index: number) => {
@@ -456,15 +716,22 @@ export const PaneChat = memo(function PaneChat({
       setFollow((n) => n + 1);
     } else toast.error(t("chat.notSent"));
   };
+  // "No, and tell … differently": the CLI drops back to its prompt, where
+  // the note is what the user would type next.
+  const denyWithNote = (index: number, note: string) => {
+    if (!menu || !chooseRunOption(runId, menu, index)) return void toast.error(t("chat.notSent"));
+    setMenu(null);
+    window.setTimeout(() => void command(note), 800);
+  };
 
   // A hidden tab says when this session waits on the user or its turn ended.
   const was = useRef({ working: false, asking: false });
   useEffect(() => {
     if (asking) setTabBadge(runId, "waiting");
-    else if (working) setTabBadge(runId, null);
+    else if (busy) setTabBadge(runId, null);
     else if (was.current.working || was.current.asking) setTabBadge(runId, "done");
-    was.current = { working, asking };
-  }, [runId, working, asking]);
+    was.current = { working: busy, asking };
+  }, [runId, busy, asking]);
   useEffect(() => () => setTabBadge(runId, null), [runId]);
   const [context, setContext] = useState<{ tokens: number; model: string | null }>({
     tokens: 0,
@@ -484,6 +751,16 @@ export const PaneChat = memo(function PaneChat({
     () => showTerminal(runId, sessionId, run.tool),
     [runId, sessionId, run.tool]
   );
+  // The conversation on screen: an artifact's versions are rebuilt from it.
+  const [convMessages, setConvMessages] = useState<TranscriptMessage[]>([]);
+  // An artifact card opens in this chat's own panel, like Claude Desktop's —
+  // not the workspace's Changes panel a plain file link goes to.
+  const openArtifact = useCallback(
+    (named: string) => {
+      setPreview(resolveIn(cwd, named));
+    },
+    [cwd]
+  );
   const openFile = useCallback(
     (named: string) => {
       // A relative path in a reply (`server/lib/x.js`) is relative to this pane's folder.
@@ -533,6 +810,8 @@ export const PaneChat = memo(function PaneChat({
     <div
       ref={rootRef}
       data-prompt-drop
+      // The last chat focused is where the Changes panel's comments go.
+      onFocusCapture={() => notePaneFocus(runId)}
       className="relative flex min-h-0 flex-1 [container-type:inline-size]"
       style={{ "--chat-side": `${side * 100}%` } as CSSProperties}
     >
@@ -549,6 +828,8 @@ export const PaneChat = memo(function PaneChat({
             onContext={onContext}
             onNotStarted={setNotStarted}
             onOpenFile={openFile}
+            onOpenArtifact={openArtifact}
+            onMessages={setConvMessages}
           />
           {!outlineOpen && !preview && (
             <button
@@ -573,6 +854,10 @@ export const PaneChat = memo(function PaneChat({
                 menu={menu}
                 onChoose={choose}
                 onAnswer={answer}
+                onDenyNote={denyWithNote}
+                onStep={(dir) => {
+                  if (!stepRunMenu(runId, dir)) toast.error(t("chat.notSent"));
+                }}
                 onCancel={() => {
                   if (interruptRun(runId)) setMenu(null);
                   else toast.error(t("chat.notSent"));
@@ -609,6 +894,26 @@ export const PaneChat = memo(function PaneChat({
                   onAttach={(items) => setImages((prev) => addAttachments(prev, items))}
                 />
                 {claude && <PermissionChip runId={runId} initial={run.permissionMode} />}
+                {run.tool === "codex" &&
+                  // Codex's `/model` (then its effort) and `/permissions` are
+                  // numbered pickers: typed here, answered on the menu card.
+                  (
+                    [
+                      ["/permissions", t("fields.permissionMode")],
+                      ["/model", t("fields.model")],
+                    ] as const
+                  ).map(([line, label]) => (
+                    <button
+                      key={line}
+                      type="button"
+                      data-testid={`pane-chat-codex-${line.slice(1)}`}
+                      onClick={() => void command(line)}
+                      title={t("chat.codexPicker")}
+                      className="rounded-md px-2 py-1 text-xs text-fg-secondary transition-colors hover:bg-surface-3 hover:text-fg-primary"
+                    >
+                      {label}
+                    </button>
+                  ))}
                 <div className="ml-auto flex items-center gap-1">
                   <ContextMeter tokens={context.tokens} model={context.model} />
                   {claude && (
@@ -624,7 +929,7 @@ export const PaneChat = memo(function PaneChat({
                       }}
                     />
                   )}
-                  {working && !text ? (
+                  {busy && !text ? (
                     <button
                       type="button"
                       data-testid="pane-chat-stop"
@@ -669,6 +974,7 @@ export const PaneChat = memo(function PaneChat({
         <FilePreview
           path={preview}
           scope={scope}
+          messages={convMessages}
           onClose={() => setPreview(null)}
           // Beside the chat when the pane has room, over it when not.
           className="absolute inset-x-0 top-0 bottom-[calc(100%-var(--chat-area,100%))] z-30 rounded-xl border border-border [@container(min-width:40rem)]:static [@container(min-width:40rem)]:m-3 [@container(min-width:40rem)]:ml-0 [@container(min-width:40rem)]:w-[calc(var(--chat-side)-0.75rem)] [@container(min-width:40rem)]:flex-shrink-0"
diff --git a/client/src/components/run/PaneGrid.tsx b/client/src/components/run/PaneGrid.tsx
index ed743cfa..0506e72c 100644
--- a/client/src/components/run/PaneGrid.tsx
+++ b/client/src/components/run/PaneGrid.tsx
@@ -70,7 +70,7 @@ import {
 } from "../../lib/paneLayout";
 import { AiToolIcon } from "../AiToolIcon";
 import { RunDot, type RunState } from "./ConsoleTabs";
-import { setSessionChat, useSessionChat } from "../../lib/chatUi";
+import { CHAT_TOOLS, setSessionChat, useSessionChat } from "../../lib/chatUi";
 import { readChatDraft, writeChatDraft } from "../../lib/chatDraft";
 import { pasteToRun, takeRunInput } from "../../lib/panePaste";
 import { promptForPane } from "./RunSetup";
@@ -238,7 +238,7 @@ function PaneMenu({
 const ITEM =
   "flex w-full items-center gap-2.5 rounded-md px-2.5 py-1.5 text-left text-sm text-fg-secondary hover:bg-surface-4 hover:text-fg-primary";
 
-/** Chat UI ⇄ terminal for a Claude/Codex pane — this session's own view
+/** Chat UI ⇄ terminal for a `CHAT_TOOLS` pane — this session's own view
  *  (`lib/chatUi.ts`); a session with no id yet has no chat to show. The
  *  unsent prompt goes along: the terminal's input box into the chat's draft,
  *  the draft (its files as paths) back into the input box, never sent. */
@@ -254,7 +254,7 @@ export function showTerminal(runId: string, sessionId: string, tool?: string) {
 function ChatToggle({ run }: { run?: RunHandle }) {
   const { t } = useTranslation("lanes");
   const chat = useSessionChat(run?.sessionId);
-  if (!run?.sessionId || (run.tool !== "claude" && run.tool !== "codex")) return null;
+  if (!run?.sessionId || !CHAT_TOOLS.has(run.tool)) return null;
   const sessionId = run.sessionId;
   const label = chat ? t("panes.showTerminal") : t("panes.showChat");
   return (
diff --git a/client/src/components/run/TerminalView.tsx b/client/src/components/run/TerminalView.tsx
index 245e4b99..e1728efa 100644
--- a/client/src/components/run/TerminalView.tsx
+++ b/client/src/components/run/TerminalView.tsx
@@ -65,7 +65,12 @@ import { useTheme } from "../../hooks/useTheme";
 import { terminalThemeFor } from "../../lib/terminalTheme";
 import { api } from "../../lib/api";
 import { writeClipboard } from "../../lib/clipboard";
-import { MENU_CURSOR_ROW, notePaneFocus, registerPane } from "../../lib/panePaste";
+import {
+  MENU_CURSOR_ROW,
+  notePaneFocus,
+  notePaneOutput,
+  registerPane,
+} from "../../lib/panePaste";
 import {
   claudePastes,
   droppedFilePaths,
@@ -617,6 +622,7 @@ export const TerminalView = memo(function TerminalView({
               }
             }
             term.write(data);
+            notePaneOutput(runId);
           }
         };
         sock.onclose = (e?: CloseEvent) => {
@@ -779,6 +785,17 @@ export const TerminalView = memo(function TerminalView({
             rows.push(row);
           }
           return { rows, cols: term.cols };
+        },
+        // Plain text of the rows on screen, dim cells kept (a status line's
+        // `esc to interrupt` is dim): the chat's busy read.
+        () => {
+          const term = termRef.current;
+          if (!term) return [];
+          const buf = term.buffer.active;
+          const rows: string[] = [];
+          for (let y = 0; y < term.rows; y++)
+            rows.push(buf.getLine(buf.baseY + y)?.translateToString(true) ?? "");
+          return rows;
         }
       ),
     [runId]
diff --git a/client/src/components/run/__tests__/ChatFilePreview.test.tsx b/client/src/components/run/__tests__/ChatFilePreview.test.tsx
index 227519ea..1af74d51 100644
--- a/client/src/components/run/__tests__/ChatFilePreview.test.tsx
+++ b/client/src/components/run/__tests__/ChatFilePreview.test.tsx
@@ -31,6 +31,10 @@ vi.mock("../../../lib/api", async (importOriginal) => {
         }),
         gitRawUrl: () => "",
       },
+      artifacts: {
+        pinFile: vi.fn().mockResolvedValue({ item: { id: 1 } }),
+        uploadFile: vi.fn().mockResolvedValue({ item: { id: 2 } }),
+      },
       sessions: {
         readImageUrl: (id: string, p: string) => `/img/${id}?path=${p}`,
         readFile: vi.fn((_id: string, p: string) =>
@@ -75,7 +79,11 @@ const scope = { laneId: 4, laneCwd: "/w", cwd: "/w", sessionId: "s1" };
 describe("FilePreview", () => {
   it("reads a workspace file through the lane", async () => {
     render(<FilePreview path="/w/src/a.ts" scope={scope} onClose={() => {}} />);
-    expect(await screen.findByText("export const a = 1;")).toBeTruthy();
+    expect(
+      await screen.findByText(
+        (_t, el) => el?.tagName === "CODE" && !!el.textContent?.includes("export const a = 1;")
+      )
+    ).toBeTruthy();
     expect(vi.mocked(api.lanes.gitFile)).toHaveBeenCalledWith(4, "src/a.ts");
   });
 
@@ -89,7 +97,11 @@ describe("FilePreview", () => {
   it("reads Windows paths the same way", async () => {
     const win = { laneId: 4, laneCwd: "D:\\w", cwd: "D:\\w", sessionId: "s1" };
     render(<FilePreview path={"D:\\w\\src\\a.ts"} scope={win} onClose={() => {}} />);
-    expect(await screen.findByText("export const a = 1;")).toBeTruthy();
+    expect(
+      await screen.findByText(
+        (_t, el) => el?.tagName === "CODE" && !!el.textContent?.includes("export const a = 1;")
+      )
+    ).toBeTruthy();
     expect(vi.mocked(api.lanes.gitFile)).toHaveBeenCalledWith(4, "src/a.ts");
     cleanup();
     const tool = "C:\\Users\\k\\.claude\\projects\\D--w\\s1\\tool-results\\b.txt";
@@ -144,7 +156,11 @@ describe("FileListModal", () => {
     );
     expect(screen.getByText("No preview for this file here")).toBeTruthy();
     fireEvent.click(screen.getByRole("button", { name: "a.ts" }));
-    expect(await screen.findByText("export const a = 1;")).toBeTruthy();
+    expect(
+      await screen.findByText(
+        (_t, el) => el?.tagName === "CODE" && !!el.textContent?.includes("export const a = 1;")
+      )
+    ).toBeTruthy();
   });
 });
 
@@ -172,3 +188,77 @@ describe("SkillListModal", () => {
     expect(screen.getByText("No preview for this file here")).toBeTruthy();
   });
 });
+
+describe("the artifact view", () => {
+  const html = (lines: string[]) =>
+    vi.mocked(api.lanes.gitFile).mockResolvedValueOnce({
+      available: true,
+      path: "page.html",
+      rev: "work",
+      binary: false,
+      truncated: false,
+      lines,
+    } as never);
+
+  it("runs a page in a sandbox with scripts but no same origin, and shows its code", async () => {
+    html(["<button>go</button>"]);
+    render(<FilePreview path="/w/page.html" scope={scope} onClose={() => {}} />);
+    const frame = (await screen.findByTestId("artifact-frame")) as HTMLIFrameElement;
+    expect(frame.getAttribute("sandbox")).toContain("allow-scripts");
+    expect(frame.getAttribute("sandbox")).not.toContain("allow-same-origin");
+    expect(frame.getAttribute("srcdoc")).toContain("<button>go</button>");
+    fireEvent.click(screen.getByTestId("artifact-tab-code"));
+    expect(screen.queryByTestId("artifact-frame")).toBeNull();
+    expect(
+      await screen.findByText(
+        (_t, el) => el?.tagName === "CODE" && !!el.textContent?.includes("<button>go</button>")
+      )
+    ).toBeTruthy();
+    cleanup();
+  });
+
+  it("steps back to an older version the conversation wrote", async () => {
+    html(["<p>two</p>"]);
+    const tool = (id: string, input: Record<string, unknown>, name = "Write") => ({
+      type: "assistant" as const,
+      timestamp: null,
+      content: [{ type: "tool_use" as const, id, name, input }],
+    });
+    const messages = [
+      tool("1", { file_path: "/w/page.html", content: "<p>one</p>" }),
+      tool("2", { file_path: "/w/page.html", old_string: "one", new_string: "two" }, "Edit"),
+    ];
+    render(
+      <FilePreview
+        path="/w/page.html"
+        scope={scope}
+        onClose={() => {}}
+        messages={messages as never}
+      />
+    );
+    const versions = await screen.findByTestId("artifact-versions");
+    expect(versions.textContent).toContain("2");
+    expect(
+      ((await screen.findByTestId("artifact-frame")) as HTMLIFrameElement).getAttribute("srcdoc")
+    ).toContain("two");
+    fireEvent.click(versions.querySelector("button")!);
+    expect(screen.getByTestId("artifact-frame").getAttribute("srcdoc")).toContain("<p>one</p>");
+    cleanup();
+  });
+
+  it("saves a workspace file to the Artifacts library by its path", async () => {
+    html(["<p>x</p>"]);
+    render(<FilePreview path="/w/page.html" scope={scope} onClose={() => {}} />);
+    await screen.findByTestId("artifact-frame");
+    fireEvent.click(screen.getByTestId("artifact-share"));
+    await vi.waitFor(() =>
+      expect(api.artifacts.pinFile).toHaveBeenCalledWith({
+        laneId: 4,
+        path: "page.html",
+        title: "page.html",
+        sessionId: "s1",
+      })
+    );
+    cleanup();
+  });
+});
diff --git a/client/src/components/run/__tests__/PaneChat.test.tsx b/client/src/components/run/__tests__/PaneChat.test.tsx
index 63a377c7..9237939c 100644
--- a/client/src/components/run/__tests__/PaneChat.test.tsx
+++ b/client/src/components/run/__tests__/PaneChat.test.tsx
@@ -7,9 +7,11 @@
  * activity panel, collapsed into a file button with the output count, whose
  * files open a preview in its place; a file in a turn's "Edited N files" card
  * opens in the Workspace's Changes panel when it is in the pane's folder; and
- * Stop in Send's place while the main agent works, pressing Esc in the pane;
+ * Stop in Send's place while the main agent works (its hooks, or the pane's
+ * own busy line / fresh output), pressing Esc in the pane;
  * and the menu the CLI asks with, shown as buttons that press its keys, a
- * message sent meanwhile typed as its "Type something" answer.
+ * message sent meanwhile typed as its "Type something" answer; a Codex
+ * terminal's Model / Permission mode buttons typing its own pickers.
  * @author Nguyễn Ngọc Trí Vĩ <vinnt@smartgift.vn>
  */
 
@@ -22,11 +24,14 @@ import {
   answerRunMenu,
   chooseRunOption,
   interruptRun,
+  onPaneOutput,
+  readRunBusy,
   readRunMenu,
   submitToRun,
   type PaneMenu,
 } from "../../../lib/panePaste";
 import { OPEN_IN_CHANGES_EVENT } from "../../../lib/openInChanges";
+import { eventBus } from "../../../lib/eventBus";
 
 vi.mock("../../conversation/ConversationView", () => ({
   // Stands in for a turn's "Edited N files" card: a click names a file.
@@ -51,8 +56,12 @@ vi.mock("../../../lib/panePaste", () => ({
   submitToRun: vi.fn(() => true),
   interruptRun: vi.fn(() => true),
   readRunMenu: vi.fn(() => null),
+  readRunBusy: vi.fn(() => false),
   chooseRunOption: vi.fn(() => true),
   answerRunMenu: vi.fn(() => true),
+  stepRunMenu: vi.fn(() => true),
+  notePaneFocus: vi.fn(),
+  onPaneOutput: vi.fn(() => () => {}),
 }));
 vi.mock("../../../lib/api", async (importOriginal) => {
   const actual = await importOriginal<Record<string, unknown>>();
@@ -212,6 +221,54 @@ describe("PaneChat — Stop", () => {
     }
   });
 
+  it("shows Stop while the pane's own busy line shows, with no hook saying so", async () => {
+    // omp, Cursor or Codex without its hooks: only the terminal knows.
+    vi.useFakeTimers({ shouldAdvanceTime: true });
+    vi.mocked(readRunBusy).mockReturnValue(true);
+    try {
+      renderChat();
+      expect(await screen.findByTestId("pane-chat-stop")).toBeInTheDocument();
+      expect(readRunBusy).toHaveBeenCalledWith("agenthub-lane-1", "claude");
+      vi.mocked(readRunBusy).mockReturnValue(false);
+      await act(async () => {
+        await vi.advanceTimersByTimeAsync(800);
+      });
+      await waitFor(() => expect(screen.queryByTestId("pane-chat-stop")).toBeNull());
+    } finally {
+      vi.useRealTimers();
+      vi.mocked(readRunBusy).mockReturnValue(false);
+    }
+  });
+
+  it("counts a tool with no known busy line as working while its pane draws", async () => {
+    vi.useFakeTimers({ shouldAdvanceTime: true });
+    vi.mocked(readRunBusy).mockReturnValue(null);
+    const fns: (() => void)[] = [];
+    vi.mocked(onPaneOutput).mockImplementation((_runId, fn) => {
+      fns.push(fn);
+      return () => {};
+    });
+    try {
+      renderChat();
+      await waitFor(() => expect(fns.length).toBeGreaterThan(0));
+      expect(screen.queryByTestId("pane-chat-stop")).toBeNull();
+      fns.forEach((fn) => fn());
+      await act(async () => {
+        await vi.advanceTimersByTimeAsync(800);
+      });
+      expect(await screen.findByTestId("pane-chat-stop")).toBeInTheDocument();
+      // Quiet for longer than OUTPUT_BUSY_MS: at rest again.
+      await act(async () => {
+        await vi.advanceTimersByTimeAsync(2800);
+      });
+      await waitFor(() => expect(screen.queryByTestId("pane-chat-stop")).toBeNull());
+    } finally {
+      vi.useRealTimers();
+      vi.mocked(readRunBusy).mockReturnValue(false);
+      vi.mocked(onPaneOutput).mockImplementation(() => () => {});
+    }
+  });
+
   it("keeps Send for a session at rest", async () => {
     // With session_id the server answers EVERY agent of the session (its
     // status filter does not apply): the main agent is there, waiting.
@@ -553,4 +610,103 @@ describe("PaneChat — a menu the CLI asks with", () => {
       vi.mocked(readRunMenu).mockReturnValue(null);
     }
   });
+
+  it("re-reads its transcript when the pane draws, at most once a second", async () => {
+    // Every listener the chat put on the pane (the re-read, the busy clock).
+    const fns: (() => void)[] = [];
+    const drew = () => fns.forEach((fn) => fn());
+    vi.mocked(onPaneOutput).mockImplementation((_runId, fn) => {
+      fns.push(fn);
+      return () => {};
+    });
+    const seen: unknown[] = [];
+    const off = eventBus.subscribe((m) => {
+      if (m.type === "transcript_changed") seen.push(m.data);
+    });
+    try {
+      renderChat();
+      await waitFor(() => expect(fns.length).toBeGreaterThan(0));
+      vi.useFakeTimers();
+      drew();
+      drew();
+      drew();
+      expect(seen).toEqual([{ session_id: "s1" }]);
+      await act(async () => {
+        await vi.advanceTimersByTimeAsync(1100);
+      });
+      expect(seen).toHaveLength(2);
+    } finally {
+      vi.useRealTimers();
+      off();
+      vi.mocked(onPaneOutput).mockImplementation(() => () => {});
+    }
+  });
+
+  const PERMISSION: PaneMenu = {
+    question: "Do you want to proceed?",
+    detail: ["Bash command", "npm test", "Run the tests"],
+    options: [
+      { n: 1, label: "Yes", hint: "", checked: null },
+      { n: 2, label: "Yes, and don't ask again for npm test", hint: "", checked: null },
+      { n: 3, label: "No, and tell Claude what to do differently (esc)", hint: "", checked: null },
+    ],
+    cursor: 0,
+    multi: false,
+    typeAt: -1,
+  };
+
+  it("lays a permission prompt out as a card, each answer its own key", async () => {
+    vi.mocked(readRunMenu).mockReturnValue(PERMISSION);
+    vi.mocked(chooseRunOption).mockClear();
+    try {
+      renderChat();
+      const card = await screen.findByTestId("pane-chat-permission");
+      expect(card).toHaveTextContent("Bash command");
+      expect(card).toHaveTextContent("npm test");
+      expect(card).toHaveTextContent("Do you want to proceed?");
+      const options = screen.getAllByTestId("pane-chat-permission-option");
+      // The key hint is dropped from the label.
+      expect(options.map((o) => o.textContent)).toEqual([
+        "Yes",
+        "Yes, and don't ask again for npm test",
+        "No, and tell Claude what to do differently",
+      ]);
+      fireEvent.click(options[1]!);
+      expect(chooseRunOption).toHaveBeenCalledWith("agenthub-lane-1", PERMISSION, 1);
+    } finally {
+      vi.mocked(readRunMenu).mockReturnValue(null);
+    }
+  });
+
+  it("denies with a note: picks the No row, then sends the note as the next message", async () => {
+    vi.mocked(readRunMenu).mockReturnValue(PERMISSION);
+    vi.mocked(chooseRunOption).mockClear();
+    vi.mocked(submitToRun).mockClear();
+    try {
+      renderChat();
+      fireEvent.click(await screen.findByTestId("pane-chat-permission-tell"));
+      const field = screen.getByTestId("pane-chat-permission-note");
+      fireEvent.change(field, { target: { value: "use pnpm" } });
+      fireEvent.keyDown(field, { key: "Enter" });
+      expect(chooseRunOption).toHaveBeenCalledWith("agenthub-lane-1", PERMISSION, 2);
+      await waitFor(() => expect(submitToRun).toHaveBeenCalled(), { timeout: 2000 });
+      expect(vi.mocked(submitToRun).mock.calls[0]![1]).toContain("use pnpm");
+    } finally {
+      vi.mocked(readRunMenu).mockReturnValue(null);
+    }
+  });
+});
+
+describe("PaneChat — a Codex terminal", () => {
+  it("types /model and /permissions, whose pickers come back as menu cards", () => {
+    renderChat({ tool: "codex" });
+    fireEvent.click(screen.getByTestId("pane-chat-codex-model"));
+    fireEvent.click(screen.getByTestId("pane-chat-codex-permissions"));
+    expect(vi.mocked(submitToRun).mock.calls.map((c) => c[1])).toEqual(["/model", "/permissions"]);
+  });
+
+  it("offers neither to a Claude pane", () => {
+    renderChat();
+    expect(screen.queryByTestId("pane-chat-codex-model")).toBeNull();
+  });
 });
diff --git a/client/src/i18n/locales/en/run.json b/client/src/i18n/locales/en/run.json
index e3484064..f0de4388 100644
--- a/client/src/i18n/locales/en/run.json
+++ b/client/src/i18n/locales/en/run.json
@@ -89,6 +89,7 @@
     "findClose": "Close find"
   },
   "chat": {
+    "codexPicker": "Opens Codex's own picker — answer it in the chat",
     "placeholder": "Write a message…",
     "send": "Send",
     "notSent": "The terminal is not connected — the message was not sent.",
@@ -139,7 +140,28 @@
     },
     "menu": {
       "placeholder": "Type your answer…",
-      "cancel": "Cancel (Esc)"
+      "cancel": "Cancel (Esc)",
+      "prev": "Previous question",
+      "next": "Next question"
+    },
+    "request": {
+      "notePlaceholder": "Tell the model what to do instead…",
+      "send": "Send",
+      "denyWithNote": "Deny with a note"
+    },
+    "artifact": {
+      "preview": "Preview",
+      "code": "Code",
+      "copy": "Copy",
+      "download": "Download",
+      "share": "Save to Artifacts",
+      "shared": "Saved to your Artifacts library — share it from the Artifacts page.",
+      "cannotShare": "This file cannot be saved to Artifacts.",
+      "expand": "Full screen",
+      "collapse": "Exit full screen",
+      "older": "Older version",
+      "newer": "Newer version",
+      "version": "Version {{n}} of {{total}}"
     }
   },
   "interrupted": {
diff --git a/client/src/i18n/locales/en/sessions.json b/client/src/i18n/locales/en/sessions.json
index 14d12005..f954a23c 100644
--- a/client/src/i18n/locales/en/sessions.json
+++ b/client/src/i18n/locales/en/sessions.json
@@ -230,6 +230,15 @@
         "stopped": "stopped"
       },
       "hide": "Hide conversation"
+    },
+    "artifact": {
+      "html": "Web page",
+      "svg": "Image · SVG",
+      "markdown": "Document",
+      "mermaid": "Diagram",
+      "react": "React component",
+      "csv": "Table",
+      "open": "Click to open"
     }
   },
   "overview": {
diff --git a/client/src/i18n/locales/en/settings.json b/client/src/i18n/locales/en/settings.json
index 708e2a30..2c245a71 100644
--- a/client/src/i18n/locales/en/settings.json
+++ b/client/src/i18n/locales/en/settings.json
@@ -1302,9 +1302,9 @@
   },
   "chatUi": {
     "title": "Chat UI",
-    "description": "How a new Claude or Codex session opens in its workspace. Saved for your account only.",
+    "description": "How a new Claude, Codex, omp, Cursor or Antigravity session opens in its workspace. Saved for your account only.",
     "label": "Open new sessions as a chat",
-    "labelDesc": "A Claude or Codex session first shows its conversation like Claude Desktop, with a message box, instead of its terminal. Each session then keeps its own view: the chat/terminal button beside a pane's split buttons switches it. Permission prompts and menus still need the terminal.",
+    "labelDesc": "A new session first shows its conversation like Claude Desktop, with a message box, instead of its terminal (the terminal stays underneath). Each session then keeps its own view: the chat/terminal button beside a pane's split buttons switches it. Permission prompts and questions are answered from cards in the chat.",
     "saveFailed": "Could not save: {{message}}"
   },
   "sessionDefaults": {
diff --git a/client/src/i18n/locales/vi/run.json b/client/src/i18n/locales/vi/run.json
index 9256674d..3fae175b 100644
--- a/client/src/i18n/locales/vi/run.json
+++ b/client/src/i18n/locales/vi/run.json
@@ -89,6 +89,7 @@
     "findClose": "Đóng tìm kiếm"
   },
   "chat": {
+    "codexPicker": "Mở bộ chọn của chính Codex — trả lời ngay trong chat",
     "placeholder": "Viết tin nhắn…",
     "send": "Gửi",
     "notSent": "Terminal chưa kết nối — tin nhắn chưa được gửi.",
@@ -139,7 +140,28 @@
     },
     "menu": {
       "placeholder": "Nhập câu trả lời của bạn…",
-      "cancel": "Huỷ (Esc)"
+      "cancel": "Huỷ (Esc)",
+      "prev": "Câu trước",
+      "next": "Câu tiếp"
+    },
+    "request": {
+      "notePlaceholder": "Nói cho model biết nên làm gì thay vào đó…",
+      "send": "Gửi",
+      "denyWithNote": "Từ chối kèm lời nhắn"
+    },
+    "artifact": {
+      "preview": "Xem trước",
+      "code": "Mã",
+      "copy": "Sao chép",
+      "download": "Tải xuống",
+      "share": "Lưu vào Artifacts",
+      "shared": "Đã lưu vào thư viện Artifacts của bạn — chia sẻ từ trang Artifacts.",
+      "cannotShare": "Không lưu được file này vào Artifacts.",
+      "expand": "Toàn màn hình",
+      "collapse": "Thoát toàn màn hình",
+      "older": "Phiên bản cũ hơn",
+      "newer": "Phiên bản mới hơn",
+      "version": "Phiên bản {{n}}/{{total}}"
     }
   },
   "interrupted": {
diff --git a/client/src/i18n/locales/vi/sessions.json b/client/src/i18n/locales/vi/sessions.json
index cc72bf71..7c957b49 100644
--- a/client/src/i18n/locales/vi/sessions.json
+++ b/client/src/i18n/locales/vi/sessions.json
@@ -230,6 +230,15 @@
         "stopped": "đã dừng"
       },
       "hide": "Ẩn hội thoại"
+    },
+    "artifact": {
+      "html": "Trang web",
+      "svg": "Ảnh · SVG",
+      "markdown": "Tài liệu",
+      "mermaid": "Sơ đồ",
+      "react": "Component React",
+      "csv": "Bảng",
+      "open": "Bấm để mở"
     }
   },
   "overview": {
diff --git a/client/src/i18n/locales/vi/settings.json b/client/src/i18n/locales/vi/settings.json
index d26b490d..ddaf7544 100644
--- a/client/src/i18n/locales/vi/settings.json
+++ b/client/src/i18n/locales/vi/settings.json
@@ -1302,9 +1302,9 @@
   },
   "chatUi": {
     "title": "Giao diện chat",
-    "description": "Cách một session Claude hoặc Codex mới mở trong workspace. Chỉ lưu cho tài khoản của bạn.",
+    "description": "Cách một session Claude, Codex, omp, Cursor hoặc Antigravity mới mở trong workspace. Chỉ lưu cho tài khoản của bạn.",
     "label": "Mở session mới dạng chat",
-    "labelDesc": "Session Claude hoặc Codex sẽ mở hội thoại như Claude Desktop, kèm ô nhập tin, thay cho terminal. Sau đó mỗi session giữ dạng xem riêng: nút chat/terminal cạnh nút split của pane để chuyển. Hỏi quyền và menu vẫn cần terminal.",
+    "labelDesc": "Session mới sẽ mở hội thoại như Claude Desktop, kèm ô nhập tin, thay cho terminal (terminal vẫn chạy bên dưới). Sau đó mỗi session giữ dạng xem riêng: nút chat/terminal cạnh nút split của pane để chuyển. Hỏi quyền và câu hỏi được trả lời bằng thẻ ngay trong chat.",
     "saveFailed": "Không lưu được: {{message}}"
   },
   "sessionDefaults": {
diff --git a/client/src/lib/__tests__/artifactView.test.ts b/client/src/lib/__tests__/artifactView.test.ts
new file mode 100644
index 00000000..79a446d9
--- /dev/null
+++ b/client/src/lib/__tests__/artifactView.test.ts
@@ -0,0 +1,102 @@
+/**
+ * @file artifactView.test.ts
+ * @description What counts as an artifact and how it is shown: kinds by
+ * name, a React file only when written whole, a file's versions rebuilt from
+ * the conversation's Write/Edit/MultiEdit inputs (a failed call skipped, an
+ * edit to unseen text unknown), CSV rows, and the sandboxed documents.
+ * @author Nguyễn Ngọc Trí Vĩ <vinnt@smartgift.vn>
+ */
+
+import { describe, expect, it } from "vitest";
+import {
+  artifactDocument,
+  artifactKind,
+  artifactVersions,
+  frameSandbox,
+  isArtifact,
+  parseCsv,
+} from "../artifactView";
+import type { TranscriptMessage } from "../types";
+
+const use = (id: string, name: string, input: Record<string, unknown>): TranscriptMessage => ({
+  type: "assistant",
+  timestamp: `2026-10-21T00:00:0${id}.000Z`,
+  content: [{ type: "tool_use", id, name, input } as TranscriptMessage["content"][number]],
+});
+const result = (id: string, isError = false): TranscriptMessage => ({
+  type: "user",
+  sender: "tool",
+  timestamp: null,
+  content: [
+    {
+      type: "tool_result",
+      id,
+      output: "",
+      is_error: isError,
+    } as TranscriptMessage["content"][number],
+  ],
+});
+
+describe("artifactView", () => {
+  it("knows an artifact by its name", () => {
+    expect(artifactKind("/a/page.HTML")).toBe("html");
+    expect(artifactKind("d.mmd")).toBe("mermaid");
+    expect(artifactKind("App.tsx")).toBe("react");
+    expect(artifactKind("x.go")).toBe("code");
+    expect(isArtifact("report.md", false)).toBe(true);
+    expect(isArtifact("App.tsx", false)).toBe(false);
+    expect(isArtifact("App.tsx", true)).toBe(true);
+    expect(isArtifact("server.js", true)).toBe(false);
+  });
+
+  it("rebuilds a file's versions from Write, Edit and MultiEdit", () => {
+    const msgs = [
+      use("1", "Write", { file_path: "/w/a.md", content: "# A\nold line\n" }),
+      result("1"),
+      use("2", "Edit", { file_path: "/w/a.md", old_string: "old", new_string: "new" }),
+      result("2"),
+      use("3", "Edit", { file_path: "/w/a.md", old_string: "missing", new_string: "x" }),
+      result("3", true),
+      use("4", "MultiEdit", {
+        file_path: "/w/a.md",
+        edits: [{ old_string: "# A", new_string: "# B" }],
+      }),
+      result("4"),
+      use("5", "Write", { file_path: "/w/other.md", content: "no" }),
+    ];
+    const v = artifactVersions(msgs, "/w/a.md");
+    expect(v.map((x) => x.content)).toEqual([
+      "# A\nold line\n",
+      "# A\nnew line\n",
+      "# B\nnew line\n",
+    ]);
+  });
+
+  it("says an edit to text it never saw is of unknown content", () => {
+    const v = artifactVersions(
+      [use("1", "Edit", { file_path: "a.md", old_string: "a", new_string: "b" })],
+      "a.md"
+    );
+    expect(v).toHaveLength(1);
+    expect(v[0]!.content).toBeNull();
+  });
+
+  it("reads CSV with quotes and newlines inside them", () => {
+    expect(parseCsv('a,b\n"x, y","say ""hi""\nthere"\n')).toEqual([
+      ["a", "b"],
+      ["x, y", 'say "hi"\nthere'],
+    ]);
+  });
+
+  it("builds sandboxed documents that keep the source inert inside a script", () => {
+    const doc = artifactDocument("react", "export default () => <p>a</script>b</p>", false)!;
+    expect(doc).toContain("babel");
+    expect(doc).not.toContain("a</script>b");
+    expect(artifactDocument("mermaid", "graph TD; A-->B<img>", true)).toContain(
+      "A--&gt;B&lt;img&gt;"
+    );
+    expect(artifactDocument("code", "x", false)).toBeNull();
+    expect(frameSandbox("svg")).toBe("");
+    expect(frameSandbox("html")).not.toContain("allow-same-origin");
+  });
+});
diff --git a/client/src/lib/__tests__/panePaste.test.ts b/client/src/lib/__tests__/panePaste.test.ts
index 61b78cc1..0858ed7c 100644
--- a/client/src/lib/__tests__/panePaste.test.ts
+++ b/client/src/lib/__tests__/panePaste.test.ts
@@ -21,6 +21,7 @@ import {
   chooseRunOption,
   parsePaneMenu,
   pasteToLane,
+  readRunBusy,
   readRunMenu,
   registerPane,
   takeRunInput,
@@ -160,6 +161,33 @@ describe("panePaste", () => {
     expect(m.typeAt).toBe(2);
   });
 
+  // Several questions: the unnumbered row is Next until the last one, and the
+  // tab bar says which questions have an answer.
+  it("reads a multi-question multiSelect with its Next row and question tabs", () => {
+    const m = parsePaneMenu(MULTI_Q)!;
+    expect(m.options.map((o) => [o.n, o.label, o.checked])).toEqual([
+      [1, "JWT HS256 + default secret", true],
+      [2, "Product code issue 2 lần", true],
+      [3, "Type something", false],
+      [null, "Next", null],
+      [4, "Chat about this", null],
+    ]);
+    expect(m.options[2]!.hint).toBe("");
+    expect(m.cursor).toBe(0);
+    expect(m.tabs).toEqual([
+      { label: "Critical", done: true },
+      { label: "High 1/3", done: false },
+      { label: "High 2/3", done: false },
+      { label: "High 3/3", done: false },
+    ]);
+    expect(m.tab).toBe(0);
+    expect(MENU_CURSOR_ROW.test("❯ Next")).toBe(true);
+  });
+
+  it("reads no tabs for a single question", () => {
+    expect(parsePaneMenu(ASK)!.tabs).toBeUndefined();
+  });
+
   // The "Type something" row once something is typed in it: its label is the
   // text. It is still the row right above "Chat about this".
   it("finds Type something by its place once text is typed in it, and replaces that text", () => {
@@ -188,8 +216,11 @@ describe("panePaste", () => {
     const c = parsePaneMenu(PERMISSION)!;
     expect(c.question).toBe("Do you want to proceed?");
     expect(c.options.map((o) => o.label)).toEqual(["Yes", "Yes, and always allow", "No"]);
+    // The tool and its command, drawn over the dashed rule.
+    expect(c.detail).toEqual(["Bash command", "touch probe-file.txt"]);
     const x = parsePaneMenu(CODEX_APPROVAL)!;
     expect(x.question).toBe("Would you like to run the following command?\n$ touch probe.txt");
+    expect(x.detail).toBeUndefined();
     expect(x.options).toHaveLength(3);
     expect(x.cursor).toBe(0);
   });
@@ -317,6 +348,23 @@ const MULTI = [
   "",
   "Enter to select · Tab/Arrow keys to navigate · Esc to cancel",
 ];
+const MULTI_Q = [
+  RULE,
+  "←  ☒ Critical  □ High 1/3  □ High 2/3  □ High 3/3  ✔ Submit  →",
+  "",
+  "Critical (đã verify trong code). Chọn issue để log BUG:",
+  "",
+  "❯ 1. [✔] JWT HS256 + default secret",
+  "        pkg/middleware/auth.go:156,184 — HMAC only",
+  "  2. [✔] Product code issue 2 lần",
+  "        gift product_code_repository.go:57 — không FOR UPDATE",
+  "  3. [ ] Type something",
+  "     Next",
+  RULE,
+  "  4. Chat about this",
+  "",
+  "Enter to select · Tab/Arrow keys to navigate · Esc to cancel",
+];
 const PERMISSION = [
   " Bash command",
   " touch probe-file.txt",
@@ -378,3 +426,51 @@ describe("MENU_CURSOR_ROW", () => {
     expect(has(CODEX)).toBe(false);
   });
 });
+
+describe("readRunBusy", () => {
+  it("reads each tool's own busy line, and nothing for a tool it does not know", () => {
+    let rows = ["✻ Thinking… (4s · ↓ 120 tokens · esc to interrupt)"];
+    const off = registerPane("agenthub-lane-8", () => {}, () => true, undefined, undefined, () => rows);
+    expect(readRunBusy("agenthub-lane-8", "claude")).toBe(true);
+    // Claude's permission footer is not a turn at work.
+    rows = ["Esc to cancel · Tab to amend"];
+    expect(readRunBusy("agenthub-lane-8", "claude")).toBe(false);
+    expect(readRunBusy("agenthub-lane-8", "antigravity")).toBe(false);
+    rows = ["Thinking… (3s · esc to cancel)"];
+    expect(readRunBusy("agenthub-lane-8", "antigravity")).toBe(true);
+    expect(readRunBusy("agenthub-lane-8", "omp")).toBeNull();
+    off();
+    expect(readRunBusy("agenthub-lane-8", "claude")).toBeNull();
+  });
+});
+
+describe("Codex's pickers", () => {
+  it("reads /permissions with each row's description column as its hint", () => {
+    const menu = parsePaneMenu([
+      "  >_ OpenAI Codex (v0.163.0)",
+      "",
+      "  Update Model Permissions",
+      "",
+      "› 1. Ask for approval (current)  Read and edit workspace files and run commands, with approval req",
+      "                                 access or edits outside the workspace",
+      "  2. Approve for me              Only ask for actions detected as potentially unsafe",
+      "  3. Full Access                 Use with caution: Codex can edit files",
+      "                                 without approval",
+      "  4. Read Only                   Read workspace files",
+      "",
+      "  enter select · esc back",
+    ]);
+    expect(menu?.question).toBe("Update Model Permissions");
+    expect(menu?.cursor).toBe(0);
+    expect(menu?.options.map((o) => [o.n, o.label, o.hint])).toEqual([
+      [
+        1,
+        "Ask for approval (current)",
+        "Read and edit workspace files and run commands, with approval req access or edits outside the workspace",
+      ],
+      [2, "Approve for me", "Only ask for actions detected as potentially unsafe"],
+      [3, "Full Access", "Use with caution: Codex can edit files without approval"],
+      [4, "Read Only", "Read workspace files"],
+    ]);
+  });
+});
diff --git a/client/src/lib/artifactView.ts b/client/src/lib/artifactView.ts
new file mode 100644
index 00000000..3701ca69
--- /dev/null
+++ b/client/src/lib/artifactView.ts
@@ -0,0 +1,227 @@
+/**
+ * @file artifactView.ts
+ * @description What makes a file an artifact, Claude Desktop's way, and how
+ * one is shown: its kind from its name (a page, an image, a diagram, a React
+ * component, a document, a table, code), the sandboxed document an HTML,
+ * SVG, Mermaid or React artifact runs in (an iframe `srcdoc` with no
+ * same-origin access, its libraries from a pinned CDN build, as Claude's own
+ * artifacts load theirs), a CSV's rows, and the versions a conversation's
+ * Write/Edit calls left of one file — rebuilt from their own inputs, so no
+ * server keeps copies.
+ * @author Nguyễn Ngọc Trí Vĩ <vinnt@smartgift.vn>
+ */
+
+import type { TranscriptMessage } from "./types";
+
+export type ArtifactKind =
+  "html" | "svg" | "markdown" | "mermaid" | "react" | "csv" | "image" | "code";
+
+const EXT = (p: string) =>
+  (/\.([A-Za-z0-9]+)$/.exec(p.split(/[\\/]/).pop() || "")?.[1] || "").toLowerCase();
+
+/** The kind a file shows as. */
+export function artifactKind(path: string): ArtifactKind {
+  const e = EXT(path);
+  if (e === "html" || e === "htm") return "html";
+  if (e === "svg") return "svg";
+  if (e === "md" || e === "markdown" || e === "mdx") return "markdown";
+  if (e === "mmd" || e === "mermaid") return "mermaid";
+  if (e === "jsx" || e === "tsx") return "react";
+  if (e === "csv" || e === "tsv") return "csv";
+  if (["png", "jpg", "jpeg", "gif", "webp", "avif", "bmp", "ico"].includes(e)) return "image";
+  return "code";
+}
+
+/** Kinds with a rendered Preview (the rest show their code only). */
+export const RENDERED: ReadonlySet<ArtifactKind> = new Set([
+  "html",
+  "svg",
+  "markdown",
+  "mermaid",
+  "react",
+  "csv",
+  "image",
+]);
+
+/** An edit worth a card in the conversation: a page, a document, a diagram, a
+ *  table — and a React component the model wrote whole (a `.tsx` it merely
+ *  edited is project code, not an artifact). */
+export function isArtifact(path: string, created: boolean): boolean {
+  const k = artifactKind(path);
+  if (k === "react") return created;
+  return k === "html" || k === "svg" || k === "markdown" || k === "mermaid" || k === "csv";
+}
+
+/** A code fence language for the Code tab. */
+export function artifactLang(path: string): string {
+  const e = EXT(path);
+  const map: Record<string, string> = {
+    htm: "html",
+    mmd: "text",
+    mermaid: "text",
+    md: "markdown",
+    mdx: "markdown",
+    jsx: "jsx",
+    tsx: "tsx",
+    js: "javascript",
+    mjs: "javascript",
+    cjs: "javascript",
+    ts: "typescript",
+    py: "python",
+    sh: "bash",
+    yml: "yaml",
+  };
+  return map[e] || e;
+}
+
+// Pinned builds (all on npm for well over two weeks), as Claude's own
+// artifacts load React, Tailwind and Mermaid from a CDN.
+const CDN = {
+  react: "https://cdn.jsdelivr.net/npm/react@18.3.1/umd/react.production.min.js",
+  reactDom: "https://cdn.jsdelivr.net/npm/react-dom@18.3.1/umd/react-dom.production.min.js",
+  babel: "https://cdn.jsdelivr.net/npm/@babel/standalone@7.26.4/babel.min.js",
+  tailwind: "https://cdn.tailwindcss.com/3.4.16",
+  mermaid: "https://cdn.jsdelivr.net/npm/mermaid@11.4.1/dist/mermaid.esm.min.mjs",
+};
+
+/** Text safe inside a `<script>` element. */
+const scriptSafe = (s: string) => s.replace(/<\/(script)/gi, "<\\/$1");
+const htmlEscape = (s: string) =>
+  s.replace(/[&<>"]/g, (c) => ({ "&": "&amp;", "<": "&lt;", ">": "&gt;", '"': "&quot;" })[c]!);
+
+/** Shown inside the frame when the artifact itself fails. */
+const ERROR_BOX = `<pre id="__err" style="display:none;margin:16px;padding:12px;border-radius:8px;background:#fee2e2;color:#991b1b;font:12px/1.5 ui-monospace,monospace;white-space:pre-wrap"></pre>
+<script>window.addEventListener("error",function(e){var b=document.getElementById("__err");if(b){b.style.display="block";b.textContent=String(e.message||e.error||e);}});</script>`;
+
+/** The document a rendered artifact runs in (null: it is not a frame kind). */
+export function artifactDocument(kind: ArtifactKind, source: string, dark: boolean): string | null {
+  switch (kind) {
+    case "html":
+      return /<html[\s>]|<!doctype/i.test(source)
+        ? source
+        : `<!doctype html><html><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"></head><body>${source}</body></html>`;
+    case "svg":
+      return `<!doctype html><html><head><meta charset="utf-8"><style>html,body{margin:0;height:100%}body{display:flex;align-items:center;justify-content:center;background:${dark ? "#1f1f1f" : "#fff"}}svg{max-width:100%;max-height:100vh;height:auto}</style></head><body>${source}</body></html>`;
+    case "mermaid":
+      return `<!doctype html><html><head><meta charset="utf-8"><style>body{margin:0;padding:16px;background:${dark ? "#1f1f1f" : "#fff"};display:flex;justify-content:center}</style></head><body>${ERROR_BOX}<pre class="mermaid">${htmlEscape(source)}</pre>
+<script type="module">import mermaid from "${CDN.mermaid}";mermaid.initialize({startOnLoad:false,theme:${dark ? '"dark"' : '"default"'},securityLevel:"strict"});mermaid.run().catch(function(e){var b=document.getElementById("__err");b.style.display="block";b.textContent=String(e&&e.message||e);});</script></body></html>`;
+    case "react":
+      return `<!doctype html><html><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<script src="${CDN.react}"></script><script src="${CDN.reactDom}"></script><script src="${CDN.babel}"></script><script src="${CDN.tailwind}"></script></head>
+<body><div id="root"></div>${ERROR_BOX}
+<script>
+(function(){
+  var src = ${JSON.stringify(scriptSafe(source))};
+  var mods = { react: window.React, "react-dom": window.ReactDOM, "react-dom/client": window.ReactDOM };
+  function req(name){ if (mods[name]) return mods[name]; throw new Error("This preview does not provide \\"" + name + "\\" — only react and react-dom are available."); }
+  try {
+    var out = Babel.transform(src, { filename: "Artifact.tsx", presets: [["typescript", { isTSX: true, allExtensions: true }], "react"], plugins: ["transform-modules-commonjs"] }).code;
+    var module = { exports: {} };
+    new Function("require", "module", "exports", "React", out)(req, module, module.exports, window.React);
+    var C = module.exports.default || module.exports.App || Object.values(module.exports).find(function(v){ return typeof v === "function"; });
+    if (!C) throw new Error("No component exported: export a default React component.");
+    ReactDOM.createRoot(document.getElementById("root")).render(React.createElement(C));
+  } catch (e) { var b = document.getElementById("__err"); b.style.display = "block"; b.textContent = String(e && e.message || e); }
+})();
+</script></body></html>`;
+    default:
+      return null;
+  }
+}
+
+/** The sandbox a frame kind gets: scripts but never the dashboard's origin
+ *  (no `allow-same-origin`), and none at all for an SVG. */
+export const frameSandbox = (kind: ArtifactKind) =>
+  kind === "svg" ? "" : "allow-scripts allow-forms allow-modals allow-popups allow-downloads";
+
+/** A CSV/TSV's rows (quoted fields, doubled quotes, newlines inside quotes). */
+export function parseCsv(text: string, sep = ","): string[][] {
+  const rows: string[][] = [];
+  let row: string[] = [];
+  let field = "";
+  let quoted = false;
+  for (let i = 0; i < text.length; i++) {
+    const c = text[i]!;
+    if (quoted) {
+      if (c === '"' && text[i + 1] === '"') {
+        field += '"';
+        i++;
+      } else if (c === '"') quoted = false;
+      else field += c;
+    } else if (c === '"' && field === "") quoted = true;
+    else if (c === sep) {
+      row.push(field);
+      field = "";
+    } else if (c === "\n" || c === "\r") {
+      if (c === "\r" && text[i + 1] === "\n") i++;
+      row.push(field);
+      rows.push(row);
+      row = [];
+      field = "";
+    } else field += c;
+  }
+  if (field !== "" || row.length) {
+    row.push(field);
+    rows.push(row);
+  }
+  return rows;
+}
+
+/** One version of a file, as a Write/Edit call left it. */
+export interface ArtifactVersion {
+  /** The call that made it. */
+  toolUseId: string | null;
+  timestamp: string | null;
+  /** null: the call is known, its result text is not (an edit to a file
+   *  whose earlier text the conversation never showed). */
+  content: string | null;
+}
+
+const samePath = (a: unknown, b: string) =>
+  typeof a === "string" && a.replace(/\\/g, "/") === b.replace(/\\/g, "/");
+
+/** Applies one Edit (old → new, once or everywhere) to known text. */
+function applyEdit(text: string, oldS: unknown, newS: unknown, all: unknown): string | null {
+  if (typeof oldS !== "string" || typeof newS !== "string") return null;
+  if (!text.includes(oldS)) return null;
+  return all ? text.split(oldS).join(newS) : text.replace(oldS, () => newS);
+}
+
+/**
+ * The versions of `path` in a conversation, oldest first: each successful
+ * Write gives its whole text, and each Edit/MultiEdit after a known text gives
+ * that text with its edits applied. A call whose input the transcript cut, or
+ * an edit to text never shown, is a version of unknown content (the current
+ * file on disk stands in for the newest one).
+ */
+export function artifactVersions(messages: TranscriptMessage[], path: string): ArtifactVersion[] {
+  const failed = new Set<string>();
+  for (const m of messages)
+    for (const c of m.content) if (c.type === "tool_result" && c.is_error && c.id) failed.add(c.id);
+  const out: ArtifactVersion[] = [];
+  let text: string | null = null;
+  for (const m of messages) {
+    if (m.type !== "assistant") continue;
+    for (const c of m.content) {
+      if (c.type !== "tool_use" || (c.id && failed.has(c.id))) continue;
+      const input = (c.input || {}) as Record<string, unknown>;
+      if (!samePath(input.file_path, path)) continue;
+      if (c.name === "Write") {
+        text = typeof input.content === "string" ? input.content : null;
+      } else if (c.name === "Edit") {
+        text =
+          text === null
+            ? null
+            : applyEdit(text, input.old_string, input.new_string, input.replace_all);
+      } else if (c.name === "MultiEdit") {
+        const edits = Array.isArray(input.edits) ? (input.edits as Record<string, unknown>[]) : [];
+        for (const e of edits) {
+          if (text === null) break;
+          text = applyEdit(text, e.old_string, e.new_string, e.replace_all);
+        }
+      } else continue;
+      out.push({ toolUseId: c.id ?? null, timestamp: m.timestamp, content: text });
+    }
+  }
+  return out;
+}
diff --git a/client/src/lib/chatUi.ts b/client/src/lib/chatUi.ts
index 5e126967..145985e3 100644
--- a/client/src/lib/chatUi.ts
+++ b/client/src/lib/chatUi.ts
@@ -1,6 +1,7 @@
 /**
- * @file chatUi.ts — whether a Claude/Codex session shows as a chat
- * (`PaneChat`, like Claude Desktop) or as its terminal. Each SESSION has its
+ * @file chatUi.ts — which tools' terminal panes have a Chat view
+ * (`CHAT_TOOLS`), and whether such a session shows as a chat (`PaneChat`,
+ * like Claude Desktop) or as its terminal. Each SESSION has its
  * own view (`useSessionChat`, toggled by `setSessionChat` from the pane's
  * chat/terminal button), kept in this browser's localStorage
  * (`agenthub.chatUi.sessions`, the newest `MAX_VIEWS`). The account's Chat UI
@@ -18,6 +19,19 @@ import { api } from "./api";
 
 export const CHAT_UI_KEY = "agenthub.chatUi";
 
+/** The tools whose terminal pane can show a chat: each one's pane learns its
+ *  session id (Claude's `--session-id`, the others' pollers through
+ *  `claimToolSessions`) and the server reads that session's conversation
+ *  (`GET /api/sessions/:id/transcript`). omp and Cursor send no hooks, so
+ *  their chat has no Stop button and follows at the pollers' pace. */
+export const CHAT_TOOLS: ReadonlySet<string> = new Set([
+  "claude",
+  "codex",
+  "omp",
+  "cursor",
+  "antigravity",
+]);
+
 export const VIEWS_KEY = "agenthub.chatUi.sessions";
 const MAX_VIEWS = 200;
 
diff --git a/client/src/lib/panePaste.ts b/client/src/lib/panePaste.ts
index bc3b536d..8fc0aeda 100644
--- a/client/src/lib/panePaste.ts
+++ b/client/src/lib/panePaste.ts
@@ -34,20 +34,24 @@ interface Pane {
   key?: (data: string) => void;
   /** `menuOnly`: null at once unless a row could be a menu's cursor row. */
   screen?: (menuOnly?: boolean) => PaneScreen | null;
+  /** The rows on screen as plain text, dim cells kept. */
+  lines?: () => string[];
 }
 const panes = new Map<string, Pane>();
 let lastFocused: string | null = null;
 
 /** Registers a mounted terminal (`visible`: is it on screen now; `key`: send
- *  raw input, as a key press; `screen`: its rows now); returns its unregister. */
+ *  raw input, as a key press; `screen`: its rows now; `lines`: the same as
+ *  plain text); returns its unregister. */
 export function registerPane(
   runId: string,
   paste: (text: string, submit?: boolean) => void,
   visible: () => boolean = () => true,
   key?: (data: string) => void,
-  screen?: (menuOnly?: boolean) => PaneScreen | null
+  screen?: (menuOnly?: boolean) => PaneScreen | null,
+  lines?: () => string[]
 ): () => void {
-  const entry = { paste, visible, key, screen };
+  const entry = { paste, visible, key, screen, lines };
   panes.set(runId, entry);
   return () => {
     if (panes.get(runId) === entry) panes.delete(runId);
@@ -59,6 +63,26 @@ export function notePaneFocus(runId: string): void {
   lastFocused = runId;
 }
 
+/** Who wants to hear that a pane drew something (its chat: the CLI is doing
+ *  something, so its transcript may have grown). */
+const outputListeners = new Map<string, Set<() => void>>();
+
+/** A terminal received output for `runId` (TerminalView, every PTY frame). */
+export function notePaneOutput(runId: string): void {
+  outputListeners.get(runId)?.forEach((fn) => fn());
+}
+
+/** Calls `fn` on every output `runId`'s terminal receives; returns its unregister. */
+export function onPaneOutput(runId: string, fn: () => void): () => void {
+  let set = outputListeners.get(runId);
+  if (!set) outputListeners.set(runId, (set = new Set()));
+  set.add(fn);
+  return () => {
+    set!.delete(fn);
+    if (!set!.size) outputListeners.delete(runId);
+  };
+}
+
 /** The tmux names of workspace `laneId`'s panes: `agenthub-lane-<id>[-slot]`. */
 function ofLane(runId: string, laneId: number): boolean {
   const base = `agenthub-lane-${laneId}`;
@@ -162,32 +186,47 @@ export function takeRunInput(runId: string): string {
 }
 
 /** One choice of a menu the CLI shows: `n` its number (null for Claude's
- *  unnumbered Submit row), `checked` a multiSelect box's state. */
+ *  unnumbered Submit / Next row), `checked` a multiSelect box's state. */
 export type PaneMenuOption = {
   n: number | null;
   label: string;
   hint: string;
   checked: boolean | null;
 };
+/** One question tab of an AskUserQuestion that asks several: `done` once it
+ *  has an answer (`☒`). */
+export type PaneMenuTab = { label: string; done: boolean };
 /** A menu the CLI asks with (AskUserQuestion, a permission or trust prompt,
  *  Codex's approval): `cursor` the row its `❯`/`›` is on, `typeAt` the
- *  "Type something" row (-1: none). */
+ *  "Type something" row (-1: none). `tabs` only when it asks several
+ *  questions (`← ☒ A  □ B  ✔ Submit →`), `tab` the one shown (-1: unknown).
+ *  `detail` what a permission prompt asks about (its tool, then the command),
+ *  when the screen shows it apart from the question. */
 export type PaneMenu = {
   question: string;
   options: PaneMenuOption[];
   cursor: number;
   multi: boolean;
   typeAt: number;
+  tabs?: PaneMenuTab[];
+  tab?: number;
+  detail?: string[];
 };
 
 const OPTION = /^ {0,4}([❯›][ \u00a0]?)? *(\d{1,2})\. +(.*)$/;
 /** A row only a menu shows: its cursor on a numbered option or Submit, or the
  *  "Enter to confirm" footer of Claude's unnumbered menus (the trust prompt). */
-export const MENU_CURSOR_ROW = /^ *(?:[❯›][ \u00a0]* *(?:\d{1,2}\. |Submit\s*$)|Enter to confirm)/;
-const SUBMIT = /^ *([❯›])?[ \u00a0]*Submit$/;
+export const MENU_CURSOR_ROW =
+  /^ *(?:[❯›][ \u00a0]* *(?:\d{1,2}\. |(?:Submit|Next)\s*$)|Enter to confirm)/;
+// The unnumbered button under a multiSelect's rows: Next while more questions
+// follow, Submit on the last one.
+const SUBMIT = /^ *([❯›])?[ \u00a0]*(Submit|Next)$/;
+const TAB_BAR = /^ *← +(.+?) +→ *$/;
 const RULE_ROW = /^ *[─╌]{3,} *$/;
+const DASHED_ROW = /^ *╌{3,} *$/;
 const CHECK = /^\[([ ✔])\] *(.*)$/;
 const PLACEHOLDER = /^Type something\.?$/;
+const COLUMN = /^(\S.*?\S) {2,}(\S.*)$/;
 
 /** The numbered menu at the bottom of `rows`, or null. A menu has two
  *  numbered rows or more from 1, the cursor on one of them, and no rule right
@@ -207,10 +246,13 @@ export function parsePaneMenu(rows: string[]): PaneMenu | null {
     const m = OPTION.exec(row);
     const s = m ? null : SUBMIT.exec(row);
     if (m || s) {
-      const text = m ? (m[3] ?? "") : "Submit";
+      const text = m ? (m[3] ?? "") : (s?.[2] ?? "Submit");
       const c = m ? CHECK.exec(text) : null;
-      const label = c ? (c[2] ?? "") : text;
-      const h = hint.join(" ");
+      // Codex's pickers (`/model`, `/permissions`) draw each row's
+      // description in a column after the name: split it off as the hint.
+      const col = COLUMN.exec(c ? (c[2] ?? "") : text);
+      const label = col ? col[1]! : c ? (c[2] ?? "") : text;
+      const h = [col?.[2], ...hint].filter(Boolean).join(" ");
       options.unshift({
         n: m ? Number(m[2]) : null,
         label,
@@ -239,15 +281,47 @@ export function parsePaneMenu(rows: string[]): PaneMenu | null {
   const lines: string[] = [];
   for (; q >= 0 && r[q] && !RULE_ROW.test(r[q]!) && lines.length < 4; q--)
     lines.unshift(r[q]!.trim());
+  const question = lines.join("\n");
+  const tabs = tabBar(r.slice(Math.max(0, q - 4), Math.max(0, q + 1)));
+  const low = question.toLowerCase();
+  // What a permission prompt asks about: Claude draws the tool and its
+  // command over a dashed rule above the question, up to the dialog's top rule.
+  const detail: string[] = [];
+  let d = q;
+  while (d >= 0 && !r[d]) d--;
+  if (d >= 0 && DASHED_ROW.test(r[d]!))
+    for (d--; d >= 0 && !RULE_ROW.test(r[d]!) && detail.length < 12; d--)
+      if (r[d]) detail.unshift(r[d]!.trim());
   return {
-    question: lines.join("\n"),
+    question,
     options,
     cursor: options.length - cursor,
     multi: options.some((o) => o.checked !== null),
     typeAt: typeRow(options),
+    ...(tabs.length > 1
+      ? { tabs, tab: tabs.findIndex((t) => low.startsWith(t.label.toLowerCase())) }
+      : {}),
+    ...(detail.length ? { detail } : {}),
   };
 }
 
+/** The question tabs of a multi-question AskUserQuestion, read off its bar
+ *  (`←  ☒ Critical  □ High 1/3  ✔ Submit  →`) among `rows`; the Submit tab is
+ *  no question. [] when there is no bar. */
+function tabBar(rows: string[]): PaneMenuTab[] {
+  for (let i = rows.length - 1; i >= 0; i--) {
+    const m = TAB_BAR.exec(rows[i] ?? "");
+    if (!m) continue;
+    return (m[1] ?? "")
+      .split(/(?=[☒☐□✔☑■] )/)
+      .map((x) => x.trim())
+      .filter(Boolean)
+      .map((x) => ({ label: x.slice(1).trim(), done: /^[☒✔☑■]/.test(x) }))
+      .filter((t) => t.label && t.label !== "Submit");
+  }
+  return [];
+}
+
 const BARE_CURSOR = /^( {0,4})❯[ \u00a0]+(\S.*)$/;
 
 /** Claude's UNNUMBERED menu at the bottom of `r` (the folder-trust prompt:
@@ -291,6 +365,26 @@ function typeRow(options: PaneMenuOption[]): number {
   return -1;
 }
 
+/** The status line a tool's TUI shows only while a turn runs, read off its
+ *  binary's own strings: Claude's and Codex's `esc to interrupt`, agy's
+ *  `Thinking… (3s · esc to cancel)` — in parentheses after an ellipsis, so
+ *  Claude's permission footer (`Esc to cancel · Tab to amend`) is not it. A
+ *  tool missing here (omp, Cursor) has no line we know. */
+export const BUSY_LINE: Readonly<Record<string, RegExp>> = {
+  claude: /\besc to interrupt\b/i,
+  codex: /\besc to interrupt\b/i,
+  antigravity: /…\s*\([^)]*\besc to cancel\)/i,
+};
+
+/** Whether `runId`'s pane shows `tool` at work now (its BUSY_LINE), or null
+ *  when the tool's line is unknown or the pane is not mounted. */
+export function readRunBusy(runId: string, tool: string): boolean | null {
+  const re = BUSY_LINE[tool];
+  const p = panes.get(runId);
+  if (!re || !p?.lines) return null;
+  return p.lines().some((l) => re.test(l));
+}
+
 /** The menu `runId`'s visible pane shows now, or null. */
 export function readRunMenu(runId: string): PaneMenu | null {
   const p = panes.get(runId);
@@ -322,6 +416,16 @@ export function chooseRunOption(runId: string, menu: PaneMenu, index: number): b
   return true;
 }
 
+/** Moves a multi-question AskUserQuestion to its previous (-1) or next (1)
+ *  question tab, the arrows Claude's footer names. False when the pane is not
+ *  visible or not attached. */
+export function stepRunMenu(runId: string, dir: -1 | 1): boolean {
+  const p = panes.get(runId);
+  if (!p?.visible() || !p.key) return false;
+  p.key(dir < 0 ? "\x1b[D" : "\x1b[C");
+  return true;
+}
+
 /** Answers `menu` with typed text: the arrows to its "Type something" row,
  *  Ctrl+U over what is typed there already (one key clears it, however long
  *  it is), the text as a bracketed
diff --git a/client/src/lib/types.ts b/client/src/lib/types.ts
index b62272fd..8646a3e3 100644
--- a/client/src/lib/types.ts
+++ b/client/src/lib/types.ts
@@ -1930,7 +1930,10 @@ export interface WSMessage {
     /** A pane ended; refetch GET /api/run/interrupted (no payload). */
     | "runs_interrupted"
     /** An artifact was added, changed, shared or deleted; refetch the list. */
-    | "artifact_changed";
+    | "artifact_changed"
+    /** A session's transcript grew on disk (`{session_id}`); a chat showing
+     *  it reads the new lines now. */
+    | "transcript_changed";
   /** The message body, whose concrete shape is selected by `type` above. */
   data:
     | Session
diff --git a/docs/DATABASE.md b/docs/DATABASE.md
index 7c97e7a7..b68e97a6 100644
--- a/docs/DATABASE.md
+++ b/docs/DATABASE.md
@@ -213,7 +213,7 @@ CREATE TABLE sessions (
 | `awaiting_reason` | TEXT | YES | Why the row is waiting: `notification`, `stop`, `session_start`, or `interrupted`. Set/cleared in lock-step with `awaiting_input_since` (SessionStart→`session_start`, Stop→`stop`, permission/input Notification→`notification`, watchdog/Esc recovery→`interrupted`). NULL otherwise. Exception: a `compact`-source SessionStart preserves the existing value (neither stamps `session_start` nor clears it) |
 | `transcript_path` | TEXT | YES | Absolute path to the session's JSONL transcript. Written by `routes/hooks.js` on the first event that carries it (subsequent events no-op via a SQL guard) and read by the periodic compaction sweep — so the sweep touches only active session rows instead of scanning the entire `events` table for `json_extract(data,'$.transcript_path')`. Backfilled once from `events` by the `db.js` migration |
 | `source` | TEXT | NO | Data source this session was captured from. `'local'` for this machine's own Claude Code history (the default); otherwise the `remote_sources.id` of the remote SSH machine it was pulled from, or `push:<name>` for a session another machine pushed to `POST /api/hooks/ingest-batch` (a pushed hook may only extend a session its own source created). Powers the `sources` query filter on `/api/sessions`, `/api/events`, `/api/agents`, `/api/stats`, and `/api/analytics`, and the `sources` facet on `/api/sessions/facets`. Indexed by `idx_sessions_source` |
-| `tool` | TEXT | NO | The AI coding CLI that produced the session. `'claude'` for every hook-created and imported row (the default); `'codex'` for a session ingested from Codex's rollout journals by `server/lib/codex-usage.js` (see `codex_rollouts`) or created by a Codex hook (`tool: "codex"`, `server/lib/codex-hooks.js`) — a Codex session keeps the status its hooks give it while they are current (its newest event at or after its newest journal write minus two minutes); once they stop (removed, untrusted) the journal decides again. A codex row has no `transcript_path`, and, going by the journal, its `status` is `active` only while one of its journals was written in the last two minutes. `'antigravity'` for an agy CLI conversation, ingested by `server/lib/antigravity-usage.js` (see `antigravity_conversations`) or created by an Antigravity hook (`server/lib/antigravity-hooks.js`); no `transcript_path`, and `active` until it has been quiet — no write to its conversation file (or a non-empty `-wal`) or transcript, no hook event — for ten minutes when it has hook events, two minutes when it has none; it then settles to `completed`, its agents too. `'cursor'` for a Cursor agent transcript, ingested by `server/lib/cursor-usage.js` (see `cursor_transcripts`): no `transcript_path`, no `model` and no `token_usage` (the transcripts record none), `active` while its transcript was written in the last two minutes, else `completed`. The startup cleanup of stale `active` sessions and a workspace purge touch only the matching tool's rows (a purge takes the sessions of the workspace's own `tool`). The Workspace's attach/resume list offers `'claude'` rows only |
+| `tool` | TEXT | NO | The AI coding CLI that produced the session. `'claude'` for every hook-created and imported row (the default); `'codex'` for a session ingested from Codex's rollout journals by `server/lib/codex-usage.js` (see `codex_rollouts`) or created by a Codex hook (`tool: "codex"`, `server/lib/codex-hooks.js`) — a Codex session keeps the status its hooks give it while they are current (its newest event at or after its newest journal write minus two minutes); once they stop (removed, untrusted) the journal decides again. A codex row has no `transcript_path`, and, going by the journal, its `status` is `active` only while one of its journals was written in the last two minutes. `'antigravity'` for an agy CLI conversation, ingested by `server/lib/antigravity-usage.js` (see `antigravity_conversations`) or created by an Antigravity hook (`server/lib/antigravity-hooks.js`); no `transcript_path`, and `active` until it has been quiet — no write to its conversation file (or a non-empty `-wal`) or transcript, no hook event — for ten minutes when it has hook events, two minutes when it has none; it then settles to `completed`, its agents too. `'cursor'` for a Cursor agent transcript, ingested by `server/lib/cursor-usage.js` (see `cursor_transcripts`): no `transcript_path`, no `model` and no `token_usage` (the transcripts record none); its conversation is read from the file `cursor_transcripts` recorded for it, `active` while its transcript was written in the last two minutes, else `completed`. The startup cleanup of stale `active` sessions and a workspace purge touch only the matching tool's rows (a purge takes the sessions of the workspace's own `tool`). The Workspace's attach/resume list offers `'claude'` rows only |
 | `lane_id` | INTEGER | YES | The workspace that OWNS the session — the only sessions a workspace lists: the lane whose dashboard pane held it, stamped by `bindRunSession` (`server/routes/hooks.js`) from the hook's `run_id` — first writer wins, a nested `claude -p` excluded — or, for a Codex/Antigravity session, by `claimToolSessions` (`server/lib/dashboard-runs.js`) from the same-tool run that was live in its folder when it started. NULL for a session no workspace started (a terminal, an import, a tool with no hooks). No FK: removing a workspace leaves its sessions. Backfilled once from `dashboard_runs`. Indexed by `idx_sessions_lane` |
 | `archived_at` | TEXT | YES | When the user archived the session out of its workspace's lists (`PATCH /api/sessions/:id {archived}`). Hides only — nothing is deleted |
 | `pinned_at` | TEXT | YES | When the user pinned the session to the top of its workspace's sidebar list (`PATCH /api/sessions/:id {pinned}`; `GET /api/sessions?pinned_first=1` orders by it). Order only; NULL = not pinned |
diff --git a/docs/LANES.md b/docs/LANES.md
index b8c705d0..6d62e2c7 100644
--- a/docs/LANES.md
+++ b/docs/LANES.md
@@ -396,7 +396,8 @@ The dashboard web UI merges workspaces and runs into a single **Workspace** page
 - **Row menu** — right-click a row, or its `⋯` button: Rename, the Model (AI Proxy) select, New workspace here…, Copy path, Reset worktree… (managed only) and Delete workspace…, each the call the rest of the UI already makes. Delete is record-only exactly as before (see "Removing a workspace").
 - **Stop archives, Sleep keeps** — every UI path that ends a live pane goes through `endRuns` (`client/src/lib/stopSession.ts`). **Stop** — a tab's **×**, a pane header's **×**, **Close tab** in a right-click menu — kills the pane AND archives the session it held (`DELETE /api/run/:id?archive=1`, which sets `sessions.archived_at` server-side and broadcasts `session_updated`), so a session the user closed leaves the workspace's list (the home's archived fold brings it back with **Unarchive**). It only hides: a session worktree is NOT removed here, unlike the row's own Archive button. **Sleep** — the moon left of each × (AI panes only), and **Sleep** in the tab, pane and session right-click menus — kills the pane but keeps the session listed, so it can be opened and resumed later by sending it a message. Both ask first for an AI pane (it may be mid-turn); a shell closes at once. A pane that dies on its own (exit, crash, reboot) is not archived, and the CLI's `DELETE` without `archive` keeps its old behaviour.
 - **Right-click menus** — every item that has actions opens them on a right-click too, in one shell (`client/src/components/ContextMenu.tsx`: placed at the pointer, kept inside the viewport, closed by Escape, a click outside or an item). The buttons stay; the menu repeats what they do and nothing more. A session row — in the sidebar, the home's Recents and the archived fold (`lanes/SessionMenu.tsx`) — offers Open, Open in split (live), Rename and Archive (sidebar), Unarchive (archived), Session details, Copy session ID, Sleep and Close tab (live); the sidebar's **New session** row Open, Open in split, Sleep and Close tab. A console tab offers Open, Open in split, Session details, Copy session ID, Sleep and Close tab; a pane header (split view) Show <session> here, Split right/down with a new terminal, Close pane, Sleep and Stop session. Sleep is offered only for an AI pane (a shell holds no session). In the Changes panel a changed file (overview row or review header) offers Open, Stage/Unstage, Discard — two clicks, as on its trash button — or the conflict's Keep mine / Take theirs / Mark resolved, and Copy relative path; a Files tree entry Open and Copy relative path; a History commit View commit, Copy commit SHA and Copy commit message. The home's title opens the row menu, and a Context file or extra folder offers Copy path and Remove. Outside the workspace the same shell serves the Sessions table, the Dashboard's agent cards and recent events, the Activity feed and Workflow runs (`components/SessionRowMenu.tsx`: Open session details, Open in its workspace when it has one, Copy session ID, Copy folder path), an attached prompt image (View / Remove), the managed workspace card (Reset), every Settings list (Git credentials, AI Proxy, Pricing, alert rules, webhooks, remote sources, AI tools) and every Claude Code / Codex config list (skills, agents, commands, output styles and memory files via `FileItemMenu` in `components/config/MdItemList.tsx`; plugins, MCP servers, marketplaces, auth profiles), an Artifacts card (its buttons' actions, plus Copy link for a page) and a Users row (Edit and Open home in Files for an admin, Copy username / email). In the File Explorer (`FileExplorer.tsx`), besides an item and the empty area: a sidebar workspace (Open, Copy path) or home (Open, Open Trash), a path-bar crumb (Open, Download as zip, Paste into this folder, Copy path), a Trash row (Restore, Copy path, Delete forever), the Trash's empty area (Refresh, Empty Trash) and an admin's per-user Trash folder (Open, Empty Trash — asked first, off for an empty one). A delete or remove there takes the button's own confirm — a modal, or two clicks. Text areas (the terminal, a conversation, a diff) keep the browser's own menu. The shell stops the menu's clicks from reaching the card it was opened from.
-- **Chat view of a live pane** — each Claude or Codex SESSION shows either a chat or its terminal, its own choice: the chat/terminal button left of the pane's split buttons (`ChatToggle` in `client/src/components/run/PaneGrid.tsx`) switches it, and the choice is kept in this browser's localStorage by session id (`agenthub.chatUi.sessions`, the newest 200; `useSessionChat` in `client/src/lib/chatUi.ts`). **Settings → Appearance → Chat UI** (`ChatUiSettings.tsx`, saved per ACCOUNT on the server: `GET/PUT /api/me/prefs`, key `agenthub.chatUi`) is only the default — written onto a session the first time it is shown, so changing it later leaves every session already seen as it is. The chat needs the session id, so a pane shows it only once its session is known (`client/src/components/run/LaneConsolePane.tsx`). Chat (`client/src/components/run/PaneChat.tsx`) shows the session's conversation the way Claude Desktop does — the shared `ConversationView` in its `fill` layout: one centred column, your messages as bubbles, the reply as plain prose, every run of tool calls folded into one line named by what it did, in Claude Desktop's fixed order ("Ran 3 commands, edited a file, used a tool ›", `workLabel`) — a thinking block that carries text is the model's narration (Claude Code ≥ 2.1.280 stores real thinking empty) and reads as prose between those lines, as Claude Desktop shows it — that opens onto a framed list of its steps — each titled by its own description, the call in a capped scroll box, its output one click further — an Agent/Task step shows instead one muted line — its status as an icon, then type · model · time · tokens · tool calls — and its report as prose, with **View conversation ›**, which puts that subagent's own conversation in the report's place as a thread hung under the step by a line on the left (a nested `ConversationView` in its `compact` layout: its task as the prompt bubble, every step, its reply; as tall as it is up to 24rem, then it scrolls) (`client/src/components/conversation/AgentDetails.tsx`; the agent is found by the result's `agent` facts, the `agentId:` it names or the `tool_use_id` of `GET /api/sessions/:id/transcripts`, and a background one is closed by its `<task-notification>`), as does a click on a line of the background-agents card, under that line — and a round ↓ back to the end — above a prompt box that grows with its text up to 40% of the screen, then scrolls. The conversation stays at its end while you are there: a new line, the running step, an image loading or the pane shown again after being hidden keeps the newest in sight (a `ResizeObserver` in `ConversationView`), until you scroll up; sending a message takes it back to the end. It lies OVER the terminal, which stays mounted and attached (the console never unmounts a terminal); a message is sent through that terminal as a bracketed paste and Enter (`submitToRun` in `client/src/lib/panePaste.ts`), exactly as if typed, so no new route or server write exists for it. On a Windows dashboard (psmux) the attach (`server/lib/pty-attach.js`) holds any input that follows a paste until 700 ms after it (`PASTE_HOLD_MS`): psmux delays a paste ~350 ms and wrote an Enter sent meanwhile INSIDE it (psmux#766), so the prompt sat in Claude's composer as `[Pasted text #1]` instead of being sent. While the session's main agent is `working` (its `UserPromptSubmit` hook seen, no `Stop` yet — `GET /api/agents?session_id=&status=working`, re-read on `agent_updated`), an empty box shows **Stop** in Send's place, like Claude Desktop: it presses Esc in the pane (`interruptRun` in the same file), the key `claude` and `codex` stop a turn with; typing a message brings Send back (Claude Code queues it). Stop also ends the conversation's live line (`stopped` on `ConversationView`/`MessageList`) until the transcript gains a newer line: an Esc before the model wrote anything puts the prompt back in Claude Code's input box and writes no `[Request interrupted by user]` line, so the transcript still ends on the prompt and would read as **Thinking…** for 15 minutes (issue #74). An Esc pressed in the terminal itself is not seen by the chat. A Codex pane without AgentHub's hooks never shows it. An image attached in the box reaches the transcript as its saved path — in a Claude pane as an `@` mention (`claudeMention` in `client/src/lib/promptAttach.ts`, which Claude reads at submit, an image as an image; a trailing space keeps its `@` suggester from taking the Enter), as is a file dropped or pasted on a Claude terminal and every file a new or resumed Claude session is started with: a bare pasted path became `[Image #n]` only when the TUI caught the paste in time, and stayed a plain `C:\…\drops\…png` on a Windows dashboard (psmux), while a path in the command-line prompt is never an image; the user's turn shows the image itself (`GET /api/run/drop/:name`) — any other attached file as a chip (name · EXT) that opens its preview like an output — and opens it full size over the whole page (`ImageLightbox`, rendered into `document.body`, so nothing stacked over a pane covers it). The box carries Claude Desktop's controls: a `+` menu (`ChatAddMenu.tsx`, also in the workspace home's prompt box) — Upload or add files; what already exists, attached like an upload (a chip, its path sent with the prompt): **Project files** (a search of the box's folder, folders too — `GET /api/run/files?dirs=1` — sent as `@path`), **Context files** (the workspace's own, `GET /api/lanes/:id/sources`, sent as `@<stored path>`) and **Recent uploads** (files uploaded through a prompt box in the last day, kept in this browser's localStorage `agenthub.prompt.recentUploads` — the server prunes an upload after a day); a file or folder dragged onto anywhere in a prompt box (not only its text area) attaches the same way — from the Changes panel's Files tree, its search, a Changes-list row, or a workspace's Context files and Folder list; a row from an extra folder's section (`data-drag-root`) and a context file or folder carry their absolute path, since they lie outside the pane's folder (`client/src/lib/promptAttach.ts`); Skills (yours and the project's), Plugins → a plugin → the skills and commands it ships, each put at the front of the message as `/<name>` / `/<plugin>:<name>`, and Manage links to the Claude config page — and, for a Claude pane, the Model/Effort chip, a change typed into the pane as `/model <id>` / `/effort <level>`, and the permission mode beside `+` in the box. A menu opened in a split pane (a `Select`, the Model chip) flips up or down by the room inside ITS pane (`data-menu-bounds` on the pane, `menuRoom` in `client/src/components/Select.tsx`), never the window's — a menu running past the pane's edge is painted over by the pane below. A running session changes mode only by Shift+Tab, so the server steps it and reads each step back off the pane (`POST /api/run/:id/permission-mode`, `server/lib/pane-mode.js`): it lands on exactly the mode picked, presses nothing while a menu hides the mode line, and returns to where it began when the mode is not in that session's cycle (Bypass only when the session was started with it). The mode belongs to that session only. A menu the CLI asks with — AskUserQuestion (one question or several, multiSelect too), a permission or trust prompt, Codex's approval, a first-start login menu — is answered from the chat: it is read off the terminal underneath every 700 ms while the pane is on screen (`parsePaneMenu`/`readRunMenu` in `client/src/lib/panePaste.ts`: two numbered rows or more from 1, the `❯`/`›` cursor on one of them, no `─` rule right above the first — a reply's numbered list has no cursor, a list typed into Claude's input box sits under its rule) and shown in the prompt box's place as a card of buttons (`PaneMenuCard` in `PaneChat.tsx`, with each option's description and a multiSelect box's state; the box is hidden with CSS, never unmounted, so its draft survives the question); a pick presses its digit in the pane (which picks it, or ticks its box), Claude's unnumbered **Submit** row is reached with the arrows and Enter, **Cancel (Esc)** presses Esc, and the **Type something** row opens a field inside the card whose text, on Enter, is typed there as the answer (`answerRunMenu`: the arrows to that row, Ctrl+U over any text already in it, the text as a bracketed paste, Enter — none in a multiSelect, where Submit is still to pick). That row presses no key: a second press of its digit would type the digit into the CLI's own field. Once text is typed in that row its label is the text, so it is found as the last numbered row above **Chat about this**. Like Send, it only presses keys in the pane, so no new route exists for it. A menu with no numbers (the first-start theme picker) still needs the terminal. A live pane's chat ends, until its turn is over, with what the model is doing now and — like the terminal's status line — how long the turn has run since your prompt and the output tokens it has made so far, e.g. `Thinking… (2m 8s · ↑ 3.5K tokens)` (`workingOn` and `turnStats` in `MessageList.tsx`; Claude Code repeats a call's usage on each of its lines, so consecutive lines with the same prompt size count once), the words lit by a shimmer that sweeps a few times and rests (`.text-shimmer`, stopped under reduced motion) and the token count rolling up rather than jumping. It names every call of the turn still waiting for its result — the oldest first and **+N more** for the rest, a click listing them all, since Claude Code writes parallel calls on separate lines and their results arrive in any order — by its description (Claude Code's Bash `description`) or tool and target, with that call's own time once it has run 30 s; else **Thinking…**. A line with no new transcript line for 3 minutes turns amber (the step may be stuck). While the CLI asks you something — the main agent awaiting input on a `Notification` hook (a permission prompt; `awaiting_reason = 'notification'`, read with the Stop button's agent poll) — it reads **Waiting for your approval: <step>** in amber with an **Open terminal** button (`showTerminal` in `PaneGrid.tsx`, the chat/terminal toggle's own switch: the unsent prompt goes into the CLI's input box); the menu card above answers it from the chat. AskUserQuestion fires no `Notification`, so a menu read off the pane counts as waiting too — for this line and the hidden tab's **● Waiting**. Before the session has started — no session row yet, because Claude Code runs no hook until its folder is trusted — the chat says so instead of the empty-transcript text, with the same **Open terminal** button (`notStarted` in `ConversationView.tsx`), and Send and every control typed into the pane (`/model`, `/effort`) ask first (`confirmDialog`, **Send anyway**): an Enter there would answer the trust prompt, whose preselected choice is **No, exit** (issue #73). It asks rather than refuses because the row appears only with the session's first hook, and a hard lock would strand the chat if none came right after the answer. A turn that ended (`end_turn`) closes with **✓ Done in 2m 8s · ↑ 3.5K tokens** — no cost: pricing one turn would need a second cost formula. A subagent launched in the background returns at once (its result is "Async agent launched…"), so the turn ends while it still works: until a `<task-notification>` names its call, a live chat lists it under the conversation — its type, its description, how long it has run and, once its own transcript was read, the step it is on and its context size (↓, its newest prompt), like the terminal's footer (`backgroundAgents`/`BackgroundAgents`/`agentProgress` in `MessageList.tsx`; the tail is re-read every 4 s through the existing `GET /api/sessions/:id/transcript?agent_id=<agentId from the launch result>`, only the lines after those already read, and only while the card is on screen) — and the last turn reads **Replied in … · N agents still running** with a spinner instead of Done (issue #81). A launch older than 6 hours with no notification is taken as ended (a CLI that quit, a resumed old session). The prompt box shows how full the context is (`ContextMeter` in `PaneChat.tsx`): the newest reply's prompt size (input + cache) against the model's window — 200K for Claude's 4.5 line and older, 1M for every model since (`contextWindow`, from LiteLLM's list; none for another vendor) — amber from 75 %, red from 90 %, the numbers on hover. While the browser tab is hidden its title is marked **● Waiting** while a chat pane's session waits on you and **✓ Done** once a turn ended (`client/src/lib/tabBadge.ts`); showing the tab drops the Done mark. Thinking is shown (after your prompt, after a tool result, or after narration whose `stop_reason` is `tool_use`). The transcript route passes each assistant line's `stop_reason`; `end_turn` ends the turn, so does an interrupt, and a turn with no new line for 15 minutes and no call running is treated as dead. The chat re-reads the transcript every 3 s while the page is visible and the pane is on screen (`useOnScreen` in `client/src/hooks/useOnScreen.ts`: a pane of another workspace or tab stays mounted but `display:none`, and polls nothing until it is shown again, then catches up once), and on a `new_event` of its session at most every 1.5 s (`EVENT_FETCH_MS`). The tail read (`?after=`) skips the lines the chat already has before parsing them. While the chat covers the terminal, the terminal is moved out of the viewport with a transform (`covered` on `TerminalView`), which stops xterm drawing without changing its size — the menu reading below still works. Each message's hover line shows ↓ the whole prompt it was answered from (input + cache read + cache write; the split on hover) and ↑ its output. An AskUserQuestion is not folded into the collapsed work line: it is its own card (`AskCard`) showing each question with the option the user picked or what they typed — the `answers` the transcript route reads off Claude Code's `toolUseResult` —, "declined" when they refused it. Each assistant turn that changed files ends, like Claude Desktop's, with an **Edited N files** card (`EditedFiles` in `client/src/components/conversation/MessageList.tsx`): one row per file with the lines it added and removed over the turn — the `diff` the transcript route reads off Claude Code's `structuredPatch` — and, in a live pane's chat, a click on a row opens that file in the workspace's Changes panel when it lies in the pane's folder, else its preview (beside the chat, or over it when the pane is narrow). It has no Undo: reverting a file is a write the dashboard does not take from a click. The conversation column and the prompt box are `max(48rem, 70%)` of the pane wide — never narrower than Claude's fixed 48rem, wider on a large screen. Every chat pane, split or not, has Claude Desktop's activity panel (`client/src/components/run/ChatOutline.tsx`, `GET /api/sessions/:id/outline`, re-read every 5 s while the tab is visible and the pane is on screen; the server keeps the last answer per session until its events change). It starts open in a pane that spans the grid's full width or height (`spanning` in `ConsoleArea`, `client/src/pages/Workspace.tsx`, from `layoutRects`) once the session has an output, and otherwise — no output yet, or a quarter-sized pane (e.g. four panes at once) — collapsed into a file button with the number of outputs at the chat's top right. The user's own open or close is kept per session for the browser tab (`sessionStorage` `agenthub.chat.panel.<sessionId>`), so switching the pane to its terminal and back keeps it; a split or a close that changes the pane's shape drops it and brings the default back. Whether it sits beside the conversation or floats over it is decided by the PANE's width, not the window's (a CSS container query on `PaneChat`'s root, 40rem): beside it, a splitter on its left edge (`Divider` from `PaneGrid.tsx`, arrow keys too) drags the panel — and a file preview in its place — between 20% and 70% of the pane, remembered per browser (`localStorage` `agenthub.chat.panelWidth`, one fraction for every pane). Open, it shows **Progress** — the session's `TaskCreate`/`TaskUpdate` (or `TodoWrite`) plan, only when there is one: how many steps are done of how many (`3/7`) over a bar, done items struck through and the step in progress highlighted; **Outputs** — files it wrote or edited (`Write`/`Edit`/`MultiEdit`/`NotebookEdit`), newest first, each with its extension tag (a file made by a Bash command is not listed); and **Used in this session** — **Uploads** (files attached through the box), **Files read**, **Memory** (Claude Code memory files), **Skills** and **Connectors** (MCP servers). A file path in a reply or your message is a link — absolute, or relative with a folder and an extension (`server/lib/x.js`, resolved against the pane's folder by `resolveIn` in `client/src/lib/resolvePath.ts`), a `:line` suffix included, bare, in inline code or as a markdown link's target (`[x](src/x.ts#L12)`; `MarkdownContent.tsx`) — and opens like an Edited-files row: in the Changes panel when it lies in the workspace's folder, else in the preview. A click on an output opens its preview in the panel's place (`FilePreview` in `client/src/components/run/ChatFilePreview.tsx`: name · EXT, Copy path, Close back to the panel); Uploads, Files read and Memory open their list with a preview beside it (`FileListModal`), and Skills opens the same dialog over the session's skills, each previewed from its SKILL.md (`SkillListModal`: a plugin's `plugin:name` from `GET /api/cc-config/plugins`, else the user's or the project's skill; a built-in skill has no file and shows by name); Connectors opens the Claude config page. Every row but Connectors also drags onto the prompt box (`dragAttachments`/`dragSkill` in `client/src/lib/promptAttach.ts`): an output, or a whole Uploads/Files read/Memory group, attaches its files like the `+` menu does (an upload by its bare path, anything else as `@<absolute path>`); a skill becomes the message's leading `/<name>` — the Skills row only when the session used one skill (a prompt takes one command), else drag it from its list. A row dragged out of one of those list dialogs closes the dialog, so the box is reachable. A preview reads only through routes that already guard it — a file under the pane's folder through `GET /api/lanes/:id/git/file` (an image through `/git/raw`), a memory file under `~/.claude` through `GET /api/cc-config/file`, an attached image or text file (`.txt`, `.md`, `.csv`, `.log`, `.json`… — a long paste is one) through `GET /api/run/drop/:name` — and shows just the path for anything else (an attached spreadsheet, a file outside the workspace, one since deleted). A Windows path (`C:\…\.claude\…`) is matched like a POSIX one. The ended-session replay and the Analyst page's conversation use the same layout. The chat's unsent prompt and its attachments are kept per session in this browser (`agenthub.chat.draft.<sessionId>`, `client/src/lib/chatDraft.ts`), so switching the pane to its terminal and back, or reloading the page, leaves them as typed — an attached image keeps its path, not its preview; a send clears it. Switching carries the unsent prompt across: to the terminal, the draft (its files as paths) is pasted into the CLI's input box without Enter and the draft is cleared; to the chat, the text in the CLI's input box is read off the screen (`parsePromptInput` in `client/src/lib/panePaste.ts`: Claude's `❯` box between its `─` rules, Codex's last `›` row up to its footer; a soft-wrapped row is joined back) into the draft and deleted there with Delete/Backspace — never Ctrl+C or Esc, which would stop a working turn. A box holding `[Pasted text #N]` or `[Image #N]` is left as it is, since the screen does not show what it stands for. A row of a turn's **Edited N files** card drags into the prompt box as an `@path` attachment (its header drags every file in the card); dropped elsewhere it is the plain path.
+- **Chat view of a live pane** — each SESSION of a Claude, Codex, omp, Cursor or Antigravity pane (`CHAT_TOOLS` in `client/src/lib/chatUi.ts`; the last three learn their session id from their pollers' `claimToolSessions`, and the transcript route reads omp's journal, Cursor's agent transcript and agy's `transcript_full.jsonl` step log) shows either a chat or its terminal, its own choice: the chat/terminal button left of the pane's split buttons (`ChatToggle` in `client/src/components/run/PaneGrid.tsx`) switches it, and the choice is kept in this browser's localStorage by session id (`agenthub.chatUi.sessions`, the newest 200; `useSessionChat` in `client/src/lib/chatUi.ts`). **Settings → Appearance → Chat UI** (`ChatUiSettings.tsx`, saved per ACCOUNT on the server: `GET/PUT /api/me/prefs`, key `agenthub.chatUi`) is only the default — written onto a session the first time it is shown, so changing it later leaves every session already seen as it is. The chat needs the session id, so a pane shows it only once its session is known (`client/src/components/run/LaneConsolePane.tsx`). Chat (`client/src/components/run/PaneChat.tsx`) shows the session's conversation the way Claude Desktop does — the shared `ConversationView` in its `fill` layout: one centred column, your messages as bubbles, the reply as plain prose, every run of tool calls folded into one line named by what it did, in Claude Desktop's fixed order ("Ran 3 commands, edited a file, used a tool ›", `workLabel`) — a thinking block that carries text is the model's narration (Claude Code ≥ 2.1.280 stores real thinking empty) and reads as prose between those lines, as Claude Desktop shows it — that opens onto a framed list of its steps — each titled by its own description, the call in a capped scroll box, its output one click further — an Agent/Task step shows instead one muted line — its status as an icon, then type · model · time · tokens · tool calls — and its report as prose, with **View conversation ›**, which puts that subagent's own conversation in the report's place as a thread hung under the step by a line on the left (a nested `ConversationView` in its `compact` layout: its task as the prompt bubble, every step, its reply; as tall as it is up to 24rem, then it scrolls) (`client/src/components/conversation/AgentDetails.tsx`; the agent is found by the result's `agent` facts, the `agentId:` it names or the `tool_use_id` of `GET /api/sessions/:id/transcripts`, and a background one is closed by its `<task-notification>`), as does a click on a line of the background-agents card, under that line — and a round ↓ back to the end — above a prompt box that grows with its text up to 40% of the screen, then scrolls. The conversation stays at its end while you are there: a new line, the running step, an image loading or the pane shown again after being hidden keeps the newest in sight (a `ResizeObserver` in `ConversationView`), until you scroll up; sending a message takes it back to the end. It lies OVER the terminal, which stays mounted and attached (the console never unmounts a terminal); a message is sent through that terminal as a bracketed paste and Enter (`submitToRun` in `client/src/lib/panePaste.ts`), exactly as if typed, so no new route or server write exists for it. On a Windows dashboard (psmux) the attach (`server/lib/pty-attach.js`) holds any input that follows a paste until 700 ms after it (`PASTE_HOLD_MS`): psmux delays a paste ~350 ms and wrote an Enter sent meanwhile INSIDE it (psmux#766), so the prompt sat in Claude's composer as `[Pasted text #1]` instead of being sent. While the session's main agent is `working` (its `UserPromptSubmit` hook seen, no `Stop` yet — `GET /api/agents?session_id=&status=working`, re-read on `agent_updated`) — or while the terminal itself shows the turn at work (its own busy line, `BUSY_LINE`/`readRunBusy` in `client/src/lib/panePaste.ts`: Claude's and Codex's `esc to interrupt`, agy's `Thinking… (… esc to cancel)`; for omp and Cursor, whose line is unknown, output drawn in the last 2 s), read every 700 ms while the pane is on screen, so a tool that sends no hooks gets it too —, an empty box shows **Stop** in Send's place, like Claude Desktop: it presses Esc in the pane (`interruptRun` in the same file), the key `claude` and `codex` stop a turn with; typing a message brings Send back (Claude Code queues it). Stop also ends the conversation's live line (`stopped` on `ConversationView`/`MessageList`) until the transcript gains a newer line: an Esc before the model wrote anything puts the prompt back in Claude Code's input box and writes no `[Request interrupted by user]` line, so the transcript still ends on the prompt and would read as **Thinking…** for 15 minutes (issue #74). An Esc pressed in the terminal itself is not seen by the chat. An image attached in the box reaches the transcript as its saved path — in a Claude pane as an `@` mention (`claudeMention` in `client/src/lib/promptAttach.ts`, which Claude reads at submit, an image as an image; a trailing space keeps its `@` suggester from taking the Enter), as is a file dropped or pasted on a Claude terminal and every file a new or resumed Claude session is started with: a bare pasted path became `[Image #n]` only when the TUI caught the paste in time, and stayed a plain `C:\…\drops\…png` on a Windows dashboard (psmux), while a path in the command-line prompt is never an image; the user's turn shows the image itself (`GET /api/run/drop/:name`) — any other attached file as a chip (name · EXT) that opens its preview like an output — and opens it full size over the whole page (`ImageLightbox`, rendered into `document.body`, so nothing stacked over a pane covers it). The box carries Claude Desktop's controls: a `+` menu (`ChatAddMenu.tsx`, also in the workspace home's prompt box) — Upload or add files; what already exists, attached like an upload (a chip, its path sent with the prompt): **Project files** (a search of the box's folder, folders too — `GET /api/run/files?dirs=1` — sent as `@path`), **Context files** (the workspace's own, `GET /api/lanes/:id/sources`, sent as `@<stored path>`) and **Recent uploads** (files uploaded through a prompt box in the last day, kept in this browser's localStorage `agenthub.prompt.recentUploads` — the server prunes an upload after a day); a file or folder dragged onto anywhere in a prompt box (not only its text area) attaches the same way — from the Changes panel's Files tree, its search, a Changes-list row, or a workspace's Context files and Folder list; a row from an extra folder's section (`data-drag-root`) and a context file or folder carry their absolute path, since they lie outside the pane's folder (`client/src/lib/promptAttach.ts`); Skills (yours and the project's), Plugins → a plugin → the skills and commands it ships, each put at the front of the message as `/<name>` / `/<plugin>:<name>`, and Manage links to the Claude config page — and, for a Claude pane, the Model/Effort chip, a change typed into the pane as `/model <id>` / `/effort <level>`, and the permission mode beside `+` in the box; a Codex pane has **Permission mode** and **Model** buttons that type `/permissions` and `/model` — Codex's own numbered pickers (the model, then its reasoning level), answered on the menu card below, each row's description column read as its hint. A menu opened in a split pane (a `Select`, the Model chip) flips up or down by the room inside ITS pane (`data-menu-bounds` on the pane, `menuRoom` in `client/src/components/Select.tsx`), never the window's — a menu running past the pane's edge is painted over by the pane below. A running session changes mode only by Shift+Tab, so the server steps it and reads each step back off the pane (`POST /api/run/:id/permission-mode`, `server/lib/pane-mode.js`): it lands on exactly the mode picked, presses nothing while a menu hides the mode line, and returns to where it began when the mode is not in that session's cycle (Bypass only when the session was started with it). The mode belongs to that session only. A menu the CLI asks with — AskUserQuestion (one question or several, multiSelect too), a permission or trust prompt, Codex's approval, a first-start login menu — is answered from the chat: it is read off the terminal underneath every 700 ms while the pane is on screen (`parsePaneMenu`/`readRunMenu` in `client/src/lib/panePaste.ts`: two numbered rows or more from 1, the `❯`/`›` cursor on one of them, no `─` rule right above the first — a reply's numbered list has no cursor, a list typed into Claude's input box sits under its rule) and shown in the prompt box's place as a card of buttons (`PaneMenuCard` in `PaneChat.tsx`, with each option's description and a multiSelect box's state; the box is hidden with CSS, never unmounted, so its draft survives the question); a permission prompt — a menu with a **Yes…** and a **No…** row: Claude's tool approval, Codex's command approval, the folder-trust question — is laid out as its own card (`PermissionMenuCard`): the tool and its command (Claude draws them over a dashed `╌` rule above the question, read as the menu's `detail`), the CLI's own answers as buttons, and **Deny with a note**, which picks the "No, and tell … differently" row and then sends the note as the next message once the CLI is back at its prompt; a pick presses its digit in the pane (which picks it, or ticks its box), Claude's unnumbered **Submit** row — **Next** while more questions follow — is reached with the arrows and Enter; an AskUserQuestion that asks several questions shows its question tabs above the card, each ticked once answered, with ‹ › buttons that press ←/→ in the pane (`stepRunMenu`), **Cancel (Esc)** presses Esc, and the **Type something** row opens a field inside the card whose text, on Enter, is typed there as the answer (`answerRunMenu`: the arrows to that row, Ctrl+U over any text already in it, the text as a bracketed paste, Enter — none in a multiSelect, where Submit is still to pick). That row presses no key: a second press of its digit would type the digit into the CLI's own field. Once text is typed in that row its label is the text, so it is found as the last numbered row above **Chat about this**. Like Send, it only presses keys in the pane, so no new route exists for it. A menu with no numbers (the first-start theme picker) still needs the terminal. A live pane's chat ends, until its turn is over, with what the model is doing now and — like the terminal's status line — how long the turn has run since your prompt and the output tokens it has made so far, e.g. `Thinking… (2m 8s · ↑ 3.5K tokens)` (`workingOn` and `turnStats` in `MessageList.tsx`; Claude Code repeats a call's usage on each of its lines, so consecutive lines with the same prompt size count once), the words lit by a shimmer that sweeps a few times and rests (`.text-shimmer`, stopped under reduced motion) and the token count rolling up rather than jumping. It names every call of the turn still waiting for its result — the oldest first and **+N more** for the rest, a click listing them all, since Claude Code writes parallel calls on separate lines and their results arrive in any order — by its description (Claude Code's Bash `description`) or tool and target, with that call's own time once it has run 30 s; else **Thinking…**. A line with no new transcript line for 3 minutes turns amber (the step may be stuck). While the CLI asks you something — the main agent awaiting input on a `Notification` hook (a permission prompt; `awaiting_reason = 'notification'`, read with the Stop button's agent poll) — it reads **Waiting for your approval: <step>** in amber with an **Open terminal** button (`showTerminal` in `PaneGrid.tsx`, the chat/terminal toggle's own switch: the unsent prompt goes into the CLI's input box); the menu card above answers it from the chat. AskUserQuestion fires no `Notification`, so a menu read off the pane counts as waiting too — for this line and the hidden tab's **● Waiting**. Before the session has started — no session row yet, because Claude Code runs no hook until its folder is trusted — the chat says so instead of the empty-transcript text, with the same **Open terminal** button (`notStarted` in `ConversationView.tsx`), and Send and every control typed into the pane (`/model`, `/effort`) ask first (`confirmDialog`, **Send anyway**): an Enter there would answer the trust prompt, whose preselected choice is **No, exit** (issue #73). It asks rather than refuses because the row appears only with the session's first hook, and a hard lock would strand the chat if none came right after the answer. A turn that ended (`end_turn`) closes with **✓ Done in 2m 8s · ↑ 3.5K tokens** — no cost: pricing one turn would need a second cost formula. A subagent launched in the background returns at once (its result is "Async agent launched…"), so the turn ends while it still works: until a `<task-notification>` names its call, a live chat lists it under the conversation — its type, its description, how long it has run and, once its own transcript was read, the step it is on and its context size (↓, its newest prompt), like the terminal's footer (`backgroundAgents`/`BackgroundAgents`/`agentProgress` in `MessageList.tsx`; the tail is re-read every 4 s through the existing `GET /api/sessions/:id/transcript?agent_id=<agentId from the launch result>`, only the lines after those already read, and only while the card is on screen) — and the last turn reads **Replied in … · N agents still running** with a spinner instead of Done (issue #81). A launch older than 6 hours with no notification is taken as ended (a CLI that quit, a resumed old session). The prompt box shows how full the context is (`ContextMeter` in `PaneChat.tsx`): the newest reply's prompt size (input + cache) against the model's window — 200K for Claude's 4.5 line and older, 1M for every model since (`contextWindow`, from LiteLLM's list; none for another vendor) — amber from 75 %, red from 90 %, the numbers on hover. While the browser tab is hidden its title is marked **● Waiting** while a chat pane's session waits on you and **✓ Done** once a turn ended (`client/src/lib/tabBadge.ts`); showing the tab drops the Done mark. Thinking is shown (after your prompt, after a tool result, or after narration whose `stop_reason` is `tool_use`). The transcript route passes each assistant line's `stop_reason`; `end_turn` ends the turn, so does an interrupt, and a turn with no new line for 15 minutes and no call running is treated as dead. The chat re-reads the transcript every 3 s while the page is visible and the pane is on screen (`useOnScreen` in `client/src/hooks/useOnScreen.ts`: a pane of another workspace or tab stays mounted but `display:none`, and polls nothing until it is shown again, then catches up once), and on a `new_event` of its session at most every 1.5 s (`EVENT_FETCH_MS`), and within 300 ms of a `transcript_changed` of its session (`PUSH_FETCH_MS`: the server's projects-tree watcher saw that Claude transcript grow, `pushTranscript` in `server/index.js`, at most every 250 ms per session), and at most once a second while the terminal underneath draws (`notePaneOutput`/`onPaneOutput` in `client/src/lib/panePaste.ts`, fed by `TerminalView` — any tool, hooks or not) — so a reply appears as soon as Claude writes it, not at the next poll. The TUI draws no reply as it streams (it shows the spinner and a token count until the block is done), so the chat has no streaming text. The tail read (`?after=`) skips the lines the chat already has before parsing them. While the chat covers the terminal, the terminal is moved out of the viewport with a transform (`covered` on `TerminalView`), which stops xterm drawing without changing its size — the menu reading below still works. Each message's hover line shows ↓ the whole prompt it was answered from (input + cache read + cache write; the split on hover) and ↑ its output. An AskUserQuestion is not folded into the collapsed work line: it is its own card (`AskCard`) showing each question with the option the user picked or what they typed — the `answers` the transcript route reads off Claude Code's `toolUseResult` —, "declined" when they refused it. Each assistant turn that changed files ends, like Claude Desktop's, with an **Edited N files** card (`EditedFiles` in `client/src/components/conversation/MessageList.tsx`): one row per file with the lines it added and removed over the turn — the `diff` the transcript route reads off Claude Code's `structuredPatch` — and, in a live pane's chat, a click on a row opens that file in the workspace's Changes panel when it lies in the pane's folder, else its preview (beside the chat, or over it when the pane is narrow). It has no Undo: reverting a file is a write the dashboard does not take from a click. The conversation column and the prompt box are `max(48rem, 70%)` of the pane wide — never narrower than Claude's fixed 48rem, wider on a large screen. Every chat pane, split or not, has Claude Desktop's activity panel (`client/src/components/run/ChatOutline.tsx`, `GET /api/sessions/:id/outline`, re-read every 5 s while the tab is visible and the pane is on screen; the server keeps the last answer per session until its events change). It starts open in a pane that spans the grid's full width or height (`spanning` in `ConsoleArea`, `client/src/pages/Workspace.tsx`, from `layoutRects`) once the session has an output, and otherwise — no output yet, or a quarter-sized pane (e.g. four panes at once) — collapsed into a file button with the number of outputs at the chat's top right. The user's own open or close is kept per session for the browser tab (`sessionStorage` `agenthub.chat.panel.<sessionId>`), so switching the pane to its terminal and back keeps it; a split or a close that changes the pane's shape drops it and brings the default back. Whether it sits beside the conversation or floats over it is decided by the PANE's width, not the window's (a CSS container query on `PaneChat`'s root, 40rem): beside it, a splitter on its left edge (`Divider` from `PaneGrid.tsx`, arrow keys too) drags the panel — and a file preview in its place — between 20% and 70% of the pane, remembered per browser (`localStorage` `agenthub.chat.panelWidth`, one fraction for every pane). Open, it shows **Progress** — the session's `TaskCreate`/`TaskUpdate` (or `TodoWrite`) plan, only when there is one (Claude Code offers those tools to its newer models only with `CLAUDE_CODE_ENABLE_TODO_TOOLS=1`, so `spawnRun` puts it in every Claude pane's env; a user's `settings.json` `env` can still set `"0"`): how many steps are done of how many (`3/7`) over a bar, done items struck through and the step in progress highlighted; **Outputs** — files it wrote or edited (`Write`/`Edit`/`MultiEdit`/`NotebookEdit`), newest first, each with its extension tag (a file made by a Bash command is not listed); and **Used in this session** — **Uploads** (files attached through the box), **Files read**, **Memory** (Claude Code memory files), **Skills** and **Connectors** (MCP servers). A file path in a reply or your message is a link — absolute, or relative with a folder and an extension (`server/lib/x.js`, resolved against the pane's folder by `resolveIn` in `client/src/lib/resolvePath.ts`), a `:line` suffix included, bare, in inline code or as a markdown link's target (`[x](src/x.ts#L12)`; `MarkdownContent.tsx`) — and opens like an Edited-files row: in the Changes panel when it lies in the workspace's folder, else in the preview. A click on an output opens its preview in the panel's place (`FilePreview` in `client/src/components/run/ChatFilePreview.tsx`: name · EXT, Copy path, Close back to the panel); Uploads, Files read and Memory open their list with a preview beside it (`FileListModal`), and Skills opens the same dialog over the session's skills, each previewed from its SKILL.md (`SkillListModal`: a plugin's `plugin:name` from `GET /api/cc-config/plugins`, else the user's or the project's skill; a built-in skill has no file and shows by name); Connectors opens the Claude config page. Every row but Connectors also drags onto the prompt box (`dragAttachments`/`dragSkill` in `client/src/lib/promptAttach.ts`): an output, or a whole Uploads/Files read/Memory group, attaches its files like the `+` menu does (an upload by its bare path, anything else as `@<absolute path>`); a skill becomes the message's leading `/<name>` — the Skills row only when the session used one skill (a prompt takes one command), else drag it from its list. A row dragged out of one of those list dialogs closes the dialog, so the box is reachable. A preview reads only through routes that already guard it — a file under the pane's folder through `GET /api/lanes/:id/git/file` (an image through `/git/raw`), a memory file under `~/.claude` through `GET /api/cc-config/file`, an attached image or text file (`.txt`, `.md`, `.csv`, `.log`, `.json`… — a long paste is one) through `GET /api/run/drop/:name` — and shows just the path for anything else (an attached spreadsheet, a file outside the workspace, one since deleted). A Windows path (`C:\…\.claude\…`) is matched like a POSIX one. The ended-session replay and the Analyst page's conversation use the same layout. The chat's unsent prompt and its attachments are kept per session in this browser (`agenthub.chat.draft.<sessionId>`, `client/src/lib/chatDraft.ts`), so switching the pane to its terminal and back, or reloading the page, leaves them as typed — an attached image keeps its path, not its preview; a send clears it. Switching carries the unsent prompt across: to the terminal, the draft (its files as paths) is pasted into the CLI's input box without Enter and the draft is cleared; to the chat, the text in the CLI's input box is read off the screen (`parsePromptInput` in `client/src/lib/panePaste.ts`: Claude's `❯` box between its `─` rules, Codex's last `›` row up to its footer; a soft-wrapped row is joined back) into the draft and deleted there with Delete/Backspace — never Ctrl+C or Esc, which would stop a working turn. A box holding `[Pasted text #N]` or `[Image #N]` is left as it is, since the screen does not show what it stands for. A row of a turn's **Edited N files** card drags into the prompt box as an `@path` attachment (its header drags every file in the card); dropped elsewhere it is the plain path.
+- **Artifacts in the chat** — like Claude Desktop, a page, document, diagram, table or React component the model wrote (`isArtifact` in `client/src/lib/artifactView.ts`: `.html/.htm`, `.svg`, `.md/.markdown/.mdx`, `.mmd/.mermaid`, `.csv/.tsv`, and a `.jsx/.tsx` written whole by a Write — one it merely edited is project code) gets an **artifact card** in the conversation (`ArtifactCard` in `MessageList.tsx`: an icon by kind, its name, "Web page / Document / Diagram / React component / Table · Click to open"), shown as soon as the file is made, even mid-turn; the turn's **Edited N files** card lists the other files. A click opens it in the chat's own side panel (`onOpenArtifact` → `PaneChat`, not the workspace's Changes panel), which the Outputs list opens too: `FilePreview` in `client/src/components/run/ChatFilePreview.tsx` with **Preview | Code** (Code is the file highlighted with line numbers, `CodeBlock`), its **versions** (‹ Version n of m ›, rebuilt from the conversation's Write/Edit/MultiEdit inputs by `artifactVersions` — the transcript keeps those inputs up to 256 KB, `server/lib/transcript-message.js`; an edit to text the conversation never showed is a version of unknown content, the file on disk standing in for the newest), **Copy** (the text), **Download**, **Save to Artifacts** (a workspace file through `POST /api/artifacts/file` by its path, anything else — an older version, a file outside the workspace — through `POST /api/artifacts/file/upload`; shared to others from the Artifacts page), **Full screen** (over the whole page) and Close; clicking the name copies its path. Preview renders Markdown, a CSV as a table, an image, and runs HTML, Mermaid (pinned `mermaid@11.4.1`) and React (`react@18.3.1` + `@babel/standalone@7.26.4` + Tailwind's play CDN, a default-exported component, `react`/`react-dom` the only imports) in an `<iframe srcdoc>` sandboxed with scripts but WITHOUT `allow-same-origin` — an opaque origin that can reach neither the dashboard's cookies nor its API — and an SVG with no scripts at all (`artifactDocument`/`frameSandbox`). The libraries come from jsDelivr, as Claude's own artifacts load theirs; offline, those kinds show their error inside the frame and Code still works.
 - **The workspace home** — `/workspace?lane=<id>` opens on the workspace's home (`client/src/components/lanes/WorkspaceHome.tsx`), shaped like a Claude Desktop project page: the name (rename from its `⋯` menu, which is the sidebar row menu) with the repository it belongs to under it (root, the main worktree's branch, the worktree count, from `GET /api/projects` — the facts Settings → Projects used to show), a prompt box — the one place a session starts WITH a prompt (a pane has no run form; the session view's tab bar `+` only starts a bare CLI or a shell) — whose fields run in order **AI** (the session's own tool: every installed, enabled one, the workspace's by default), **AI Proxy** (Native or a saved profile; Native alone, disabled, for a tool that cannot use one), **Model** (the tool's own list, or on a router that router's models), **Permission** (Claude's permission mode; Codex's sandbox and approval; a catalog tool's per-session bypass switch) and **Effort** (Claude and Codex), whose half-typed text is kept per workspace in `agenthub.workspace.draft.<lane id>` (localStorage, so it survives leaving the page or switching workspace; cleared once sent — attached images are not kept), and whose **Start** opens a NEW session in the first slot with no live run (0–9, `POST /api/lanes/:id/start {slot}`; all ten busy is refused) and then shows it in a tab of its own, and **Recents** — the workspace's own sessions (`?lane=<id>`) with their cost and time, a live one opening its pane and any other its conversation (`&session=`), exactly like the sidebar. The transcript page (`/sessions/<id>`) of a Claude session started in a workspace's own folder carries an explicit **Resume** (`client/src/components/lanes/SessionResumeButton.tsx`, the one place that resumes) and beside it **Fork**, which resumes the history into a NEW session id (`{forkSession: true}` → `--fork-session`, the original untouched, offered even while the original has a live pane): it warns that a resume re-reads the whole transcript into context, re-reads the live panes, then starts it in the workspace's first free pane (`POST /api/lanes/:id/start {slot, resumeSessionId}`) and opens it. A session from a SUBFOLDER gets no Resume — the pane starts in the workspace's folder and Claude finds a session by the folder it ran in — and a session that still has a live pane shows **Open pane** instead. On the right sit the workspace's own context: **Instructions** (appended to every new session's system prompt, after the workspace owner's own **My agent instructions** from Settings → AI tools — one markdown text per account (typed, or read from a local `.md`/`.txt` with **Choose file**, saved only on Save), shared by every workspace they own), **Memory** (read-only: the folder's CLAUDE.md / AGENTS.md family and Claude's auto memory), **Context** (uploaded files, ≤ 10 MB each, stored by AgentHub) and **Folder** (the workspace's own folder plus extra ones, with the repo's untracked worktrees suggested). Only Claude and Codex panes can take them (`--add-dir` plus the instructions as a `CLAUDE.md` in an `--add-dir` folder of their own / `-c developer_instructions=`); for any other tool the page says so and its panes start without them. See `docs/API.md` → Workspace context. Each extra folder is also its own repo in the **Changes** panel, stacked under the workspace folder's like VS Code's Source Control: once the workspace has one, every repo is a section with a foldable header, each a whole panel — Changes, Files, Review, History, commit and push — on its own folder (every git route with `?source=<id>`, resolved from the `lane_sources` row; `ENOSOURCE` otherwise, never the workspace folder); the open sections share the height, and a folded one stays mounted. A chat's request to open a file unfolds the workspace folder's section. Right-clicking a section header offers **Close repository** (extra folders only — the workspace folder never closes) and **Close other repositories**; closing REMOVES the folder from the workspace (`DELETE /api/lanes/:id/sources/:sourceId`, the folder stays on disk), asking first only if that section holds an unsaved edit. The panel's bottom fold, **Workspace context** (folded by default, kept in `agenthub.workspace.contextOpen`), is the home's context column made compact (`client/src/components/lanes/WorkspaceContext.tsx`, the same component): Instructions, Memory, Context files and Folders are edited from the session view without going back to the home, and a folder added or removed there joins or leaves the stack above at once. A plain-text paste over 5,000 characters into any prompt box (the home, a chat, a finished session's reply) is attached as `pasted-text.txt` like a pasted file (`PASTE_AS_FILE_CHARS` in `client/src/components/run/RunSetup.tsx`): the prompt carries its path, never the text (#68). The home HIDES the session view with CSS — the console never unmounts behind it.
 - **A session in its own worktree** — Claude Desktop's worktree toggle: the row under the home's prompt box has a **Local / New worktree** select (Claude workspaces in a git repo). Claude Code itself makes a worktree only when started with `--worktree` or when asked mid-session — and a worktree it enters mid-session is invisible to AgentHub, which reads a session's folder once, at its start, and the new session's pane runs `claude --worktree=<name>`, so CLAUDE CODE makes `<repo>/.claude/worktrees/<name>` on branch `worktree-<name>`, copies `.worktreeinclude`, runs your `WorktreeCreate` hook, and on exit removes a worktree with no changes (or asks). AgentHub adds `/.claude/worktrees/` to the repo's `info/exclude` (Claude does not, and the main repo would list it). The session stays this workspace's own — listed with a branch icon and the worktree's name — and the session view's Changes panel works in ITS worktree (every git route with `?session=`, resolved from the session row: owned by the workspace, registered by `git worktree list`; `ENOWORKTREE` otherwise, never the main folder). Resume and Fork start in that worktree. **Archive** of a worktree session with no live pane asks first and removes the worktree like Claude Desktop — `git worktree remove` without force (git refuses unsaved work, and then nothing is archived) and `git branch -d` (unmerged work keeps its branch); a lock held by a live process refuses, a stale one is cleared. The workspace's own folder is never removed.
 - **Resume interrupted sessions** — when a session's pane dies under it (the machine shut down, tmux was killed, the CLI crashed), its run is recorded `interrupted` instead of `abandoned`, and every page of the dashboard opens the **Resume interrupted sessions?** dialog (`client/src/components/lanes/InterruptedSessionsPrompt.tsx`, from `GET /api/run/interrupted`). Each session is listed with its workspace, tool, name or first prompt and when it stopped, all ticked: **Resume** starts each ticked one in its workspace's first free tab (`POST /api/lanes/:id/start {slot: "auto", resumeSessionId}`, with the run's model, effort and permission mode — never `bypassPermissions`) and opens the first; **Don't resume** (also ×, Escape, the backdrop) resumes none. Either answer dismisses the whole list (`POST /api/run/interrupted/dismiss`), so it is asked once; a resume that failed stays listed. What counts as interrupted: a pane the dashboard did not close (Stop, ×, Quit are `killed`), not a Terminal tab, with a known session — and for Claude, not one the user ended themselves (`SessionEnd` reason `prompt_input_exit`/`logout`). Codex and other tools report no such reason, so their every vanished pane counts; a Codex pane without its opt-in hooks has no known session and is never listed. A session resumed by any other path drops off the list. A run younger than 20 s is never judged dead (a tmux list is a cached answer, and another dashboard on the same database may not have listed a brand-new pane yet). In server mode each user is offered only their own workspaces' sessions.
diff --git a/server/README.md b/server/README.md
index 6e2764ae..bff9a614 100644
--- a/server/README.md
+++ b/server/README.md
@@ -632,12 +632,19 @@ An `antigravity` workspace carries no pipeline either.
   per-conversation SQLite files (`~/.gemini/antigravity-cli/conversations/<id>.db`)
   every `DASHBOARD_ANTIGRAVITY_POLL_MS` ms (default 15000, `0` disables). They
   become `sessions` rows (`tool = 'antigravity'`) and `token_usage` rows, and
-  `antigravity_conversations` holds the mtime read for each file.
+  `antigravity_conversations` holds the mtime read for each file. The
+  session's conversation (`GET /api/sessions/:id/transcript`) is read from its
+  step log (`brain/<id>/.system_generated/logs/transcript_full.jsonl`,
+  `transcriptPath`): prompts, the model's text, thinking, tool calls and
+  tokens, and each tool's result tied to its call by step.
 - **Cursor sessions:** `server/lib/cursor-usage.js` reads Cursor's agent
   transcripts (`~/.cursor/projects/*/agent-transcripts/<id>/<id>.jsonl`) every
   `DASHBOARD_CURSOR_POLL_MS` ms (default 15000, `0` disables) into `sessions`
   rows (`tool = 'cursor'`) — no `token_usage`, the transcripts carry none —
-  and `cursor_transcripts` holds the mtime read for each file.
+  and `cursor_transcripts` holds the mtime read for each file. The session's
+  conversation (`GET /api/sessions/:id/transcript`) is read from that recorded
+  file: prompts, replies, thinking, tool calls and results, with no timestamps
+  or model (Cursor writes none).
 - **Oh My Pi sessions:** `server/lib/omp-usage.js` reads omp's session journals
   (`~/.omp/agent/sessions/<encoded cwd>/<timestamp>_<id>.jsonl`, plus every
   subagent journal under `<timestamp>_<id>/`) every `DASHBOARD_OMP_POLL_MS` ms
@@ -803,7 +810,7 @@ HTTP surface for the tmux-backed terminal run: one session per workspace, named
 
 Codex-only spawn options (`sandbox`, `approval`, `resumeLast`) are accepted by the workspace action `POST /api/lanes/:id/start`, not by `POST /api/run` — see [`docs/API.md`](../docs/API.md).
 
-The pane itself is not JSON: its bytes stream over `/ws-pty/:runId` as binary WebSocket frames. The JSON WebSocket surface adds `cc_config_changed` (broadcast by `lib/cc-watcher.js` on `fs.watch` events under `~/.claude/` and by `routes/cc-config.js` after every successful PUT/DELETE — debounced at 500 ms, payload `{ source: "dashboard"|"fs", action?, scope?, type?, name?, paths? }`) and its Codex counterpart `codex_config_changed` (broadcast by `routes/codex-config.js` after every successful file mutation and CLI-backed mcp/plugin/marketplace action — always `source: "dashboard"`, there is no `$CODEX_HOME` filesystem watcher).
+The pane itself is not JSON: its bytes stream over `/ws-pty/:runId` as binary WebSocket frames. The JSON WebSocket surface adds `cc_config_changed` (broadcast by `lib/cc-watcher.js` on `fs.watch` events under `~/.claude/` and by `routes/cc-config.js` after every successful PUT/DELETE — debounced at 500 ms, payload `{ source: "dashboard"|"fs", action?, scope?, type?, name?, paths? }`) and its Codex counterpart `codex_config_changed` (broadcast by `routes/codex-config.js` after every successful file mutation and CLI-backed mcp/plugin/marketplace action — always `source: "dashboard"`, there is no `$CODEX_HOME` filesystem watcher), and `transcript_changed` `{ session_id }` — the session sync's `fs.watch` on the projects tree saw a Claude transcript grow (`pushTranscript` in `server/index.js`, at most every 250 ms per session; nothing is parsed for it), so a chat showing that session reads its new lines at once instead of at its next 3 s poll. It reaches whoever the session's own `new_event` reaches (`broadcastAudience`).
 
 ### Import History
 
diff --git a/server/__tests__/antigravity-transcript.test.js b/server/__tests__/antigravity-transcript.test.js
new file mode 100644
index 00000000..4737105e
--- /dev/null
+++ b/server/__tests__/antigravity-transcript.test.js
@@ -0,0 +1,146 @@
+/**
+ * @file antigravity-transcript.test.js
+ * @description An Antigravity (agy) session's conversation view: its
+ * `transcript_full.jsonl` step log mapped to the Claude transcript shape —
+ * the prompt out of `<USER_REQUEST>`, the model's thinking, text, tool calls
+ * and tokens, each tool result tied to its call by step and an invalid call
+ * marked an error — one "main" transcript, no subagent files, and a session
+ * id that is no UUID finding nothing. The agy dir and the DB are temp dirs;
+ * the steps are synthetic, in the shape agy 1.4 writes.
+ * @author Nguyễn Ngọc Trí Vĩ <vinnt@smartgift.vn>
+ */
+const os = require("node:os");
+const fs = require("node:fs");
+const path = require("node:path");
+
+const ROOT = fs.mkdtempSync(path.join(os.tmpdir(), "agenthub-agy-transcript-"));
+process.env.DASHBOARD_DB_PATH = path.join(ROOT, "dashboard.db");
+process.env.AGENTHUB_GEMINI_DIR = path.join(ROOT, "gemini");
+process.env.CLAUDE_HOME = path.join(ROOT, "claude");
+
+const { describe, it, after } = require("node:test");
+const assert = require("node:assert/strict");
+const { db } = require("../db");
+const agy = require("../lib/antigravity-usage");
+
+let server;
+after(() => {
+  server?.close();
+  fs.rmSync(ROOT, { recursive: true, force: true });
+});
+
+const ID = "57154e1a-65e5-4b51-a116-1890f5ca831f";
+const AT = "2026-10-21T05:08:24Z";
+
+function steps(id) {
+  const dir = path.join(
+    process.env.AGENTHUB_GEMINI_DIR,
+    "antigravity-cli",
+    "brain",
+    id,
+    ".system_generated",
+    "logs"
+  );
+  fs.mkdirSync(dir, { recursive: true });
+  const file = path.join(dir, "transcript_full.jsonl");
+  const lines = [
+    {
+      step_index: 0,
+      source: "USER_EXPLICIT",
+      type: "USER_INPUT",
+      status: "DONE",
+      created_at: AT,
+      content:
+        "<USER_REQUEST>\nCreate c.txt\n</USER_REQUEST>\n<ADDITIONAL_METADATA>\nThe current local time is: x.\n</ADDITIONAL_METADATA>",
+    },
+    {
+      step_index: 1,
+      source: "MODEL",
+      type: "PLANNER_RESPONSE",
+      status: "DONE",
+      created_at: AT,
+      input_tokens: 3889,
+      cache_read_tokens: 768,
+      output_tokens: 12,
+      thinking: "Two commands.",
+      content: "Running them now.",
+      tool_calls: [
+        { name: "run_command", args: { CommandLine: "echo ok > c.txt" } },
+        { name: "read", args: { path: "c.txt" } },
+      ],
+    },
+    {
+      step_index: 2,
+      source: "MODEL",
+      type: "GENERIC",
+      status: "DONE",
+      created_at: AT,
+      content: "Created At: x\nCompleted At: y\n\nThe command exited with code 0.\n",
+    },
+    {
+      step_index: 3,
+      source: "MODEL",
+      type: "GENERIC",
+      status: "INVALID",
+      error: 'unknown tool: "read"',
+      created_at: AT,
+      content: 'Created At: x\nCompleted At: y\nError invalid tool call: unknown tool: "read"',
+    },
+    {
+      step_index: 4,
+      source: "MODEL",
+      type: "PLANNER_RESPONSE",
+      status: "DONE",
+      created_at: AT,
+      content: "Done.",
+    },
+  ];
+  fs.writeFileSync(file, lines.map((l) => JSON.stringify(l)).join("\n") + "\n");
+  return file;
+}
+
+describe("an Antigravity session's conversation", () => {
+  it("reads agy's step log in the Claude transcript shape", async () => {
+    const file = steps(ID);
+    assert.equal(agy.transcriptPath(ID), file);
+    assert.equal(agy.transcriptPath("../../etc"), null);
+    db.prepare(
+      "INSERT INTO sessions (id, name, status, started_at, tool) VALUES (?, 'c', 'completed', ?, 'antigravity')"
+    ).run(ID, AT);
+
+    const { createApp, startServer } = require("../index");
+    server ||= await startServer(createApp(), 0);
+    const BASE = `http://127.0.0.1:${server.address().port}`;
+    const list = await (await fetch(`${BASE}/api/sessions/${ID}/transcripts`)).json();
+    assert.deepEqual(
+      list.transcripts.map((t) => [t.id, t.type]),
+      [["main", "main"]]
+    );
+    const body = await (await fetch(`${BASE}/api/sessions/${ID}/transcript`)).json();
+    assert.deepEqual(
+      body.messages.map((m) => [m.sender, m.content.map((c) => c.type)]),
+      [
+        ["user", ["text"]],
+        ["assistant", ["thinking", "text", "tool_use", "tool_use"]],
+        ["tool", ["tool_result"]],
+        ["tool", ["tool_result"]],
+        ["assistant", ["text"]],
+      ]
+    );
+    const [user, plan, ok, bad] = body.messages;
+    assert.equal(user.content[0].text, "Create c.txt");
+    assert.equal(user.timestamp, AT);
+    // Each result names the call it answers: call i of step s lands at step s+1+i.
+    assert.deepEqual(
+      plan.content.filter((c) => c.type === "tool_use").map((c) => c.id),
+      [ok.content[0].id, bad.content[0].id]
+    );
+    assert.equal(ok.content[0].output, "The command exited with code 0.\n");
+    assert.equal(ok.content[0].is_error, false);
+    assert.equal(bad.content[0].is_error, true);
+    assert.equal(plan.usage.input_tokens, 3889);
+    assert.equal(plan.usage.cache_read_input_tokens, 768);
+    const sub = await (await fetch(`${BASE}/api/sessions/${ID}/transcript?agent_id=x`)).json();
+    assert.equal(sub.messages.length, 0);
+  });
+});
diff --git a/server/__tests__/cursor-usage.test.js b/server/__tests__/cursor-usage.test.js
index f394a40e..94acc696 100644
--- a/server/__tests__/cursor-usage.test.js
+++ b/server/__tests__/cursor-usage.test.js
@@ -4,8 +4,10 @@
  * project slug walked back onto the filesystem) and no tokens, an unchanged
  * file never re-read, a quiet one settled to completed, a session the user
  * deleted staying deleted, and a non-cursor row with the same id never
- * touched. The Cursor dir and the DB are temp dirs; the transcripts are
- * synthetic, in the shape Cursor writes.
+ * touched, a session claimed by the Cursor pane it started in, and the
+ * transcript mapped to the Claude transcript shape by the
+ * transcript routes. The Cursor dir and the DB are temp dirs; the transcripts
+ * are synthetic, in the shape Cursor writes.
  * @author Nguyễn Ngọc Trí Vĩ <vinnt@smartgift.vn>
  */
 
@@ -16,13 +18,18 @@ const path = require("node:path");
 const ROOT = fs.mkdtempSync(path.join(os.tmpdir(), "agenthub-cursor-usage-"));
 process.env.DASHBOARD_DB_PATH = path.join(ROOT, "dashboard.db");
 process.env.AGENTHUB_CURSOR_DIR = path.join(ROOT, "cursor");
+process.env.CLAUDE_HOME = path.join(ROOT, "claude");
 
 const { describe, it, beforeEach, after } = require("node:test");
 const assert = require("node:assert/strict");
 const { db } = require("../db");
 const cursor = require("../lib/cursor-usage");
 
-after(() => fs.rmSync(ROOT, { recursive: true, force: true }));
+let server;
+after(() => {
+  server?.close();
+  fs.rmSync(ROOT, { recursive: true, force: true });
+});
 
 const ID = "7481e37e-9157-4145-b985-2ea9a2117f52";
 const PROJECT = path.join(fs.realpathSync(ROOT), "work", "my app");
@@ -65,7 +72,8 @@ const session = (id) => db.prepare("SELECT * FROM sessions WHERE id = ?").get(id
 beforeEach(() => {
   fs.rmSync(process.env.AGENTHUB_CURSOR_DIR, { recursive: true, force: true });
   db.exec(
-    "DELETE FROM token_usage; DELETE FROM agents; DELETE FROM sessions; DELETE FROM cursor_transcripts;"
+    "DELETE FROM token_usage; DELETE FROM agents; DELETE FROM sessions; DELETE FROM cursor_transcripts;" +
+      " DELETE FROM dashboard_runs; DELETE FROM lanes;"
   );
 });
 
@@ -175,4 +183,64 @@ describe("cursor usage", () => {
     assert.equal(session(ID).name, `Cursor ${ID.slice(0, 8)}`);
     assert.equal(session(ID).cwd, null);
   });
+
+  it("maps the transcript to the transcript shape", async () => {
+    const file = transcript(ID);
+    const more = [
+      {
+        role: "assistant",
+        message: {
+          content: [
+            { type: "text", text: "Reading it now." },
+            { type: "tool_use", id: "t1", name: "Grep", input: { pattern: "x" } },
+          ],
+        },
+      },
+      {
+        role: "user",
+        message: { content: [{ type: "tool_result", tool_use_id: "t1", content: "no match" }] },
+      },
+      { role: "assistant", message: { content: [{ type: "text", text: "Done." }] } },
+    ];
+    fs.appendFileSync(file, more.map((l) => JSON.stringify(l)).join("\n") + "\n");
+    await cursor.ingestCursorUsage();
+    assert.equal(cursor.transcriptPath(ID), file);
+
+    const { createApp, startServer } = require("../index");
+    server ||= await startServer(createApp(), 0);
+    const BASE = `http://127.0.0.1:${server.address().port}`;
+    const list = await (await fetch(`${BASE}/api/sessions/${ID}/transcripts`)).json();
+    assert.deepEqual(
+      list.transcripts.map((t) => [t.id, t.type]),
+      [["main", "main"]]
+    );
+    const body = await (await fetch(`${BASE}/api/sessions/${ID}/transcript`)).json();
+    assert.deepEqual(
+      body.messages.map((m) => [m.sender, m.content.map((c) => c.type)]),
+      [
+        ["user", ["text"]],
+        ["assistant", ["tool_use"]],
+        ["assistant", ["text", "tool_use"]],
+        ["tool", ["tool_result"]],
+        ["assistant", ["text"]],
+      ]
+    );
+    assert.equal(body.messages[0].content[0].text, "Review the campaign code\nplease");
+    assert.equal(body.messages[3].content[0].output, "no match");
+    const sub = await (await fetch(`${BASE}/api/sessions/${ID}/transcript?agent_id=x`)).json();
+    assert.equal(sub.messages.length, 0);
+  });
+
+  it("gives the session to the Cursor pane it started in", async () => {
+    db.prepare("INSERT INTO lanes (id, title, cwd) VALUES (5, 'w', ?)").run(PROJECT);
+    db.prepare(
+      `INSERT INTO dashboard_runs (id, lane_id, cwd, status, started_at, tool)
+       VALUES ('agenthub-lane-5', 5, ?, 'running', ?, 'cursor')`
+    ).run(PROJECT, new Date(Date.now() - 60_000).toISOString());
+    transcript(ID);
+    await cursor.ingestCursorUsage();
+    assert.equal(session(ID).lane_id, 5);
+    const run = db.prepare("SELECT session_id FROM dashboard_runs WHERE id = 'agenthub-lane-5'");
+    assert.equal(run.get().session_id, ID);
+  });
 });
diff --git a/server/__tests__/lane-sources.test.js b/server/__tests__/lane-sources.test.js
index 2c2d91bb..f2f829ac 100644
--- a/server/__tests__/lane-sources.test.js
+++ b/server/__tests__/lane-sources.test.js
@@ -205,6 +205,9 @@ describe("lane-sources instructions and launch", () => {
     assert.equal(argv[5], "--permission-mode");
     assert.ok(!argv.some((a) => a.startsWith("--append-system-prompt")));
     assert.ok(newSession.includes("CLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD=1"));
+    // Claude Code hides its task tools from newer models without this, and the
+    // chat Progress panel is drawn from those calls.
+    assert.ok(newSession.includes("CLAUDE_CODE_ENABLE_TODO_TOOLS=1"));
     assert.deepEqual(fs.readdirSync(dir), ["CLAUDE.md"]);
     const text = fs.readFileSync(path.join(dir, "CLAUDE.md"), "utf8");
     assert.match(text, /^- always answer in Vietnamese/);
diff --git a/server/__tests__/scope.test.js b/server/__tests__/scope.test.js
index a79130c6..64462a5f 100644
--- a/server/__tests__/scope.test.js
+++ b/server/__tests__/scope.test.js
@@ -227,6 +227,9 @@ describe("broadcastAudience", () => {
     assert.equal(reaches("agent_updated", { session_id: "s3" }, m1), false);
     assert.equal(reaches("new_event", { session_id: "s2" }, admin.user), true);
     assert.equal(reaches("new_event", { session_id: "s2" }, undefined), true);
+    // A transcript's growth notice goes where the session's events go.
+    assert.equal(reaches("transcript_changed", { session_id: "s2" }, m2), true);
+    assert.equal(reaches("transcript_changed", { session_id: "s2" }, m1), false);
   });
 
   it("routes workspace messages by owner and keeps unknown types from members", () => {
diff --git a/server/db.js b/server/db.js
index 3ce412c7..86505600 100644
--- a/server/db.js
+++ b/server/db.js
@@ -1827,6 +1827,15 @@ try {
   db.prepare("ALTER TABLE dashboard_runs ADD COLUMN pane_token_hash TEXT").run();
 }
 
+// A realtime chat run's bridge secret (`AGENTHUB_CHAT_TOKEN`), SHA-256: only the
+// pane's own bridge may stream that run's events (server/routes/chat.js).
+// Never listed.
+try {
+  db.prepare("SELECT chat_token_hash FROM dashboard_runs LIMIT 1").get();
+} catch {
+  db.prepare("ALTER TABLE dashboard_runs ADD COLUMN chat_token_hash TEXT").run();
+}
+
 // Server mode: whose browser a push subscription is (`users.id`); NULL in
 // local mode. A user's notifications go only to their own subscriptions.
 try {
diff --git a/server/index.js b/server/index.js
index 1577dbb1..435004cf 100644
--- a/server/index.js
+++ b/server/index.js
@@ -1032,7 +1032,9 @@ function startWorkflowPoll(broadcast) {
  *      sessions show up immediately rather than on the next poll. Events for
  *      files already in `mtimeCache` (active transcripts being appended to) are
  *      ignored, so a busy session never thrashes the importer — the poll picks
- *      up its growth. Recursive watching is used only on macOS/Windows (native,
+ *      up its growth; such an append only broadcasts `transcript_changed`
+ *      `{session_id}` (at most every 250 ms per session), so an open chat reads
+ *      the new lines at once. Recursive watching is used only on macOS/Windows (native,
  *      stable); on Linux, where Node's userland recursive watcher trips on the
  *      high-churn projects tree (see lib/cc-watcher.js), we watch the root plus
  *      each immediate child folder non-recursively instead.
@@ -1148,11 +1150,37 @@ function startSessionSync(broadcast) {
     }, DEBOUNCE_MS);
     if (debounce.unref) debounce.unref();
   }
-  // Only a path we don't already track is interesting (a new session file or a
-  // new project folder). Appends to a known active transcript are left to the
-  // poll, so the watcher never re-parses a busy session every write.
+  // An append to a session's own transcript: a chat showing that session reads
+  // it now instead of at its next poll. One `transcript_changed` per session
+  // per PUSH_MS — a notice only, nothing is parsed here.
+  const PUSH_MS = 250;
+  const pushing = new Map(); // session id → pending timer
+  function pushTranscript(fullPath) {
+    const m =
+      /(?:^|[\\/])([0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12})\.jsonl$/i.exec(
+        fullPath
+      );
+    if (!m || pushing.has(m[1])) return;
+    const sessionId = m[1];
+    const timer = setTimeout(() => {
+      pushing.delete(sessionId);
+      try {
+        broadcast("transcript_changed", { session_id: sessionId });
+      } catch {
+        /* best-effort — the chat's poll still catches up */
+      }
+    }, PUSH_MS);
+    if (timer.unref) timer.unref();
+    pushing.set(sessionId, timer);
+  }
+
+  // Only a path we don't already track is interesting to the importer (a new
+  // session file or a new project folder). Appends to a known active
+  // transcript are left to the poll, so the watcher never re-parses a busy
+  // session every write — it only tells the chats (`pushTranscript`).
   function onFsEvent(fullPath) {
-    if (fullPath && mtimeCache.has(fullPath)) return;
+    if (fullPath && mtimeCache.has(fullPath)) return void pushTranscript(fullPath);
+    if (fullPath) pushTranscript(fullPath);
     scheduleSweep();
   }
 
diff --git a/server/lib/antigravity-usage.js b/server/lib/antigravity-usage.js
index ab4453a2..c4ce0e8d 100644
--- a/server/lib/antigravity-usage.js
+++ b/server/lib/antigravity-usage.js
@@ -17,7 +17,8 @@
  *   5 cache read.
  * The session's name and start come from its transcript
  * (`brain/<id>/.system_generated/logs/transcript_full.jsonl`, first
- * `USER_INPUT`). A file is re-read whole when its mtime (or its non-empty
+ * `USER_INPUT`), and the same file is the conversation view's
+ * (`transcriptPath`, read by `GET /api/sessions/:id/transcript`). A file is re-read whole when its mtime (or its non-empty
  * `-wal`'s) changes — they are small — and the mtime per file lives in
  * `antigravity_conversations`, so a session the user deleted stays deleted
  * until agy writes to it again. Opening read-only still lets SQLite leave an
@@ -157,6 +158,21 @@ const transcriptFiles = (id) =>
     return ["transcript_full.jsonl", "transcript.jsonl"].map((name) => path.join(dir, name));
   });
 
+/** The transcript of conversation `id` (this account's or a member's) the
+ *  conversation view reads, or null. Only a UUID names one. */
+function transcriptPath(id) {
+  if (!/^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i.test(String(id || "")))
+    return null;
+  for (const file of transcriptFiles(id)) {
+    try {
+      if (fs.statSync(file).isFile()) return file;
+    } catch {
+      /* try the next name */
+    }
+  }
+  return null;
+}
+
 /** The transcript's last write, 0 when it has none. */
 function transcriptMtime(id) {
   for (const file of transcriptFiles(id)) {
@@ -531,4 +547,5 @@ module.exports = {
   readConversation,
   ingestAntigravityUsage,
   startAntigravityUsagePoll,
+  transcriptPath,
 };
diff --git a/server/lib/cursor-usage.js b/server/lib/cursor-usage.js
index 053d3a2b..136e318b 100644
--- a/server/lib/cursor-usage.js
+++ b/server/lib/cursor-usage.js
@@ -18,7 +18,11 @@
  *   is resolved only when exactly such a path still exists, else left empty.
  * A file is re-read when its mtime changes, and the mtime per file lives in
  * `cursor_transcripts`, so a session the user deleted stays deleted until
- * Cursor writes to it again. No transcript is ever stored as `transcript_path`.
+ * Cursor writes to it again. A session a dashboard pane started is claimed
+ * for that pane's workspace and run (`claimToolSessions`), which gives the
+ * pane its session id and so its Chat view. No transcript is ever stored as
+ * `transcript_path`; the conversation view finds the file through
+ * `transcriptPath` instead.
  * @author Nguyễn Ngọc Trí Vĩ <vinnt@smartgift.vn>
  */
 
@@ -26,6 +30,7 @@ const fs = require("node:fs");
 const os = require("node:os");
 const path = require("node:path");
 const { db, stmts } = require("../db");
+const { claimToolSessions } = require("./dashboard-runs");
 
 const ACTIVE_MS = 120_000; // the window codex-usage.js calls a session active
 const FUTURE_SLACK_MS = 1000;
@@ -148,6 +153,14 @@ function transcriptFilesIn(cursorDir) {
 }
 
 const getSeen = db.prepare("SELECT * FROM cursor_transcripts WHERE path = ?");
+const getPathOf = db.prepare(
+  "SELECT path FROM cursor_transcripts WHERE session_id = ? ORDER BY mtime_ms DESC LIMIT 1"
+);
+
+/** The transcript a Cursor session was read from (recorded by the poller), or null. */
+function transcriptPath(sessionId) {
+  return getPathOf.get(sessionId)?.path || null;
+}
 const upsertSeen = db.prepare(
   `INSERT INTO cursor_transcripts (path, session_id, mtime_ms, updated_at)
    VALUES (?, ?, ?, strftime('%Y-%m-%dT%H:%M:%fZ', 'now'))
@@ -235,6 +248,7 @@ function ingestCursorUsage({ broadcast = () => {}, now = Date.now() } = {}) {
         upsertSeen.run(f.path, f.id, f.mtime);
         if (type) {
           written += 1;
+          claimToolSessions(f.id); // its workspace and live pane, before the broadcast
           broadcast(type, stmts.getSession.get(f.id));
         }
       } catch (err) {
@@ -273,4 +287,5 @@ module.exports = {
   firstPrompt,
   ingestCursorUsage,
   startCursorUsagePoll,
+  transcriptPath,
 };
diff --git a/server/lib/data-transfer.js b/server/lib/data-transfer.js
index 0ef431f6..741e0a2f 100644
--- a/server/lib/data-transfer.js
+++ b/server/lib/data-transfer.js
@@ -90,7 +90,7 @@ function bindable(v) {
  */
 /** A run row as it may leave this machine: never a pane credential's hash. */
 function exportableRun(row) {
-  const { pane_token_hash: _secret, ...rest } = row;
+  const { pane_token_hash: _secret, chat_token_hash: _chat, ...rest } = row;
   return rest;
 }
 
@@ -185,7 +185,9 @@ function importExportBundle(db, bundle) {
     events: makeInserter(db, "events", { omit: ["id"], wrap: { data: "trim_event_data" } }),
     token_usage: makeInserter(db, "token_usage"),
     workflows: makeInserter(db, "workflows"),
-    dashboard_runs: makeInserter(db, "dashboard_runs", { omit: ["pane_token_hash"] }),
+    dashboard_runs: makeInserter(db, "dashboard_runs", {
+      omit: ["pane_token_hash", "chat_token_hash"],
+    }),
     alert_rules: makeInserter(db, "alert_rules"),
     model_pricing: makeInserter(db, "model_pricing"),
   };
diff --git a/server/lib/pty-run.js b/server/lib/pty-run.js
index bec7bc19..a0abf5cc 100644
--- a/server/lib/pty-run.js
+++ b/server/lib/pty-run.js
@@ -640,6 +640,11 @@ function spawnRun(args) {
     ...(entry.launcher === "claude" && systemPromptFile
       ? { CLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD: "1" }
       : {}),
+    // Claude Code hides TaskCreate/TaskUpdate from its newer models unless
+    // this is set, and those calls are all the chat Progress panel draws
+    // (routes/sessions.js sessionOutline). A user's settings.json `env` still
+    // wins, so `"0"` there turns them back off.
+    ...(entry.launcher === "claude" ? { CLAUDE_CODE_ENABLE_TODO_TOOLS: "1" } : {}),
     // No CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN: an inline claude duplicated its
     // output whenever the pane narrowed (tmux re-wraps the full-width rules and
     // pushes lines into history its redraw cannot clear). The user's own
diff --git a/server/lib/scope.js b/server/lib/scope.js
index d5bcbba0..59082a99 100644
--- a/server/lib/scope.js
+++ b/server/lib/scope.js
@@ -288,6 +288,7 @@ function broadcastAudience(type, data) {
     case "agent_updated":
     case "new_event":
     case "workflow_upserted":
+    case "transcript_changed":
       return { owners: sessionOwners(d.session_id) };
     case "alert_triggered":
     case "alert_updated":
diff --git a/server/lib/transcript-message.js b/server/lib/transcript-message.js
new file mode 100644
index 00000000..4bc9011f
--- /dev/null
+++ b/server/lib/transcript-message.js
@@ -0,0 +1,324 @@
+/**
+ * @file Turns one Claude Code transcript line (or a stream-json event shaped
+ * like one) into the conversation message the chat renders: sender, text,
+ * thinking, tool calls and results (with an edit's +/- counts, an
+ * AskUserQuestion's answers, a subagent's facts), model, usage. Shared by the
+ * transcript route and the realtime chat runtime so both draw the same thing.
+ * @author Nguyễn Ngọc Trí Vĩ <vinnt@smartgift.vn>
+ */
+
+/**
+ * Classify the TRUE sender of a transcript entry. A JSONL `type:"user"` line is
+ * not always the human: it also carries tool results, harness-injected
+ * task-notifications, /loop re-injections (`isMeta`), and — in a subagent
+ * transcript — the task prompt handed down by the orchestrator. Attributing all
+ * of those to "User" is wrong; the UI styles each sender distinctly.
+ *
+ * Returns: "user" | "assistant" | "orchestrator" | "system" | "tool".
+ *   user         — a real message typed by the human
+ *   assistant    — the agent's own turn
+ *   orchestrator — a subagent's task, assigned by its parent/main agent
+ *   system       — harness/tooling injection (task-notification, /loop meta, …)
+ *   tool         — a tool_result echoed back on a `user` line
+ */
+function classifyTranscriptSender(entry, isSubagentFile) {
+  if (entry.type === "assistant") return "assistant";
+  // `system`/local_command lines are surfaced as the human's slash-command I/O.
+  if (entry.type !== "user") return "user";
+
+  const content = entry.message ? entry.message.content : undefined;
+  const onlyToolResults =
+    Array.isArray(content) &&
+    content.length > 0 &&
+    content.every((b) => b && b.type === "tool_result");
+  if (entry.toolUseResult !== undefined || onlyToolResults) return "tool";
+
+  const text =
+    typeof content === "string"
+      ? content
+      : Array.isArray(content)
+        ? (content.find((b) => b && b.type === "text") || {}).text || ""
+        : "";
+  const lead = text.replace(/^\s+/, "");
+
+  // Harness injections that masquerade as a user line.
+  if (entry.isMeta === true) return "system";
+  if (lead.startsWith("<task-notification>") || lead.startsWith("<task-notification ")) {
+    return "system";
+  }
+  // Background-task event banner ("[SYSTEM NOTIFICATION - NOT USER INPUT] …")
+  // that newer harness builds prefix ahead of the <task-notification> payload.
+  if (lead.startsWith("[SYSTEM NOTIFICATION")) return "system";
+
+  // In a subagent transcript, a user line with no human prompt provenance is the
+  // task injected by the Task/Agent tool. A real human message to the subagent
+  // (rare, but allowed) carries promptSource/origin and stays "user".
+  if (isSubagentFile && entry.promptSource === undefined && entry.origin === undefined) {
+    return "orchestrator";
+  }
+
+  return "user";
+}
+
+/** The +/- line counts of a Write/Edit/MultiEdit result (`toolUseResult`):
+ *  its `structuredPatch` hunks' `+`/`-` lines, or every line of a file it
+ *  created. Null for any other result. */
+function patchCounts(r) {
+  if (!r || typeof r !== "object") return null;
+  if (Array.isArray(r.structuredPatch) && r.structuredPatch.length) {
+    let added = 0;
+    let removed = 0;
+    for (const h of r.structuredPatch) {
+      for (const l of Array.isArray(h?.lines) ? h.lines : []) {
+        if (typeof l !== "string") continue;
+        if (l.startsWith("+")) added++;
+        else if (l.startsWith("-")) removed++;
+      }
+    }
+    return { added, removed };
+  }
+  if (r.type === "create" && typeof r.content === "string") {
+    return { added: r.content.split("\n").length - (r.content.endsWith("\n") ? 1 : 0), removed: 0 };
+  }
+  return null;
+}
+
+/** An AskUserQuestion result's answers (`toolUseResult.answers`): question
+ *  text → the option picked, or what the user typed. Null for any other result. */
+function askAnswers(r) {
+  if (!r || typeof r !== "object" || !r.answers || typeof r.answers !== "object") return null;
+  if (Array.isArray(r.answers)) return null;
+  const out = {};
+  for (const [q, a] of Object.entries(r.answers)) if (typeof a === "string") out[q] = a;
+  return Object.keys(out).length ? out : null;
+}
+
+/** An Agent/Task result's facts (`toolUseResult`): the subagent's id (its
+ *  transcript is `agent-<id>.jsonl`), the model it ran on and, once it
+ *  finished in the foreground, its time, tokens and tool calls. Null for any
+ *  other result. */
+function agentInfo(r) {
+  if (!r || typeof r !== "object" || typeof r.agentId !== "string") return null;
+  if (!/^[\w-]{1,200}$/.test(r.agentId)) return null;
+  const num = (v) => (Number.isFinite(v) ? v : null);
+  return {
+    id: r.agentId,
+    model: typeof r.resolvedModel === "string" ? r.resolvedModel : null,
+    status: typeof r.status === "string" ? r.status : null,
+    duration_ms: num(r.totalDurationMs),
+    tokens: num(r.totalTokens),
+    tool_uses: num(r.totalToolUseCount),
+  };
+}
+
+// A conversation turn — a prompt, an answer — is shown whole: it is bounded by
+// the model's output and what the user typed. Only tool payloads and thinking
+// (the parts that can run to megabytes) keep the 10 KB cap. This one guards
+// against a pathological line, not a long answer.
+const TEXT_MAX = 1024 * 1024;
+// The tools whose input carries a file's text (artifactVersions on the client).
+const EDIT_TOOLS = new Set(["Write", "Edit", "MultiEdit"]);
+const EDIT_INPUT_MAX = 256 * 1024;
+
+function truncate(str, maxLen) {
+  if (!str || str.length <= maxLen) return str;
+  return str.slice(0, maxLen) + "[truncated]";
+}
+
+function truncateObj(obj, maxLen) {
+  if (!obj) return obj;
+  const json = JSON.stringify(obj);
+  if (json.length <= maxLen) return obj;
+  return { _truncated: truncate(json, maxLen) };
+}
+
+/** A parser for one transcript: it remembers the last rename it surfaced, so
+ *  a repeated `custom-title` line shows once. `isSubagentFile` tells a
+ *  subagent's task (orchestrator) from a human prompt. */
+function createClaudeLineParser(isSubagentFile = false) {
+  // Dedupe state for synthetic rename markers: custom-title lines can repeat
+  // with the same value across a transcript, so only emit when the title
+  // actually changes from the last one we surfaced.
+  let lastRenameTitle = null;
+
+  // Helper: parse a JSONL line into a message object, or null if not a displayable message
+  function parseMessage(entry, num) {
+    // /rename, `claude -n`, picker Ctrl+R → custom-title metadata line. These
+    // commands produce no user/assistant turn, so without this they'd be
+    // invisible. Surface a compact "renamed" marker instead.
+    if (entry.type === "custom-title") {
+      const title = typeof entry.customTitle === "string" ? entry.customTitle.trim() : "";
+      if (!title || title === lastRenameTitle) return null;
+      lastRenameTitle = title;
+      return {
+        type: "session_event",
+        event_kind: "rename",
+        title,
+        timestamp: entry.timestamp || null,
+        content: [],
+        line: num,
+      };
+    }
+
+    // Local slash-command I/O. Newer Claude Code builds write the command
+    // invocation and its captured output as `system`/`local_command` lines
+    // with the TUI markup in a top-level `content` string (older builds used
+    // `user` messages, handled below). Surface those as user-side text so the
+    // client's tuiSegments parser renders the command pill + stdout/stderr
+    // (e.g. /color → "/color" pill + "Session color set to: cyan"). Skip
+    // every other system subtype (turn_duration, stop_hook_summary, …) and
+    // empty local_command lines (e.g. /clear writes a content-less one).
+    if (entry.type === "system") {
+      if (entry.subtype !== "local_command") return null;
+      const sysText = typeof entry.content === "string" ? entry.content : "";
+      if (!sysText.trim()) return null;
+      return {
+        type: "user",
+        sender: "user", // local slash-command I/O is the human's own action
+        timestamp: entry.timestamp || null,
+        content: [{ type: "text", text: truncate(sysText, TEXT_MAX) }],
+        line: num,
+      };
+    }
+
+    // Mid-turn queued message: journaled as a `queued_command` attachment
+    // (never as a `user` line). Surface it at the position the model actually
+    // received it. NOT everything in the queue is the human, though — the
+    // harness delivers its own injections (task-notifications from background
+    // agents, "[SYSTEM NOTIFICATION …]" banners) through the same queue, and
+    // those attachments carry NO `origin` field, while a genuinely typed
+    // message carries `origin.kind = "human"`. So: harness-marker text or a
+    // non-human origin → "system"; everything else → "user". Other attachment
+    // subtypes are harness noise → dropped.
+    if (entry.type === "attachment") {
+      const att = entry.attachment;
+      if (!att || att.type !== "queued_command") return null;
+      const prompt = typeof att.prompt === "string" ? att.prompt : "";
+      if (!prompt.trim()) return null;
+      const lead = prompt.replace(/^\s+/, "");
+      // Same harness markers classifyTranscriptSender strips off user lines.
+      const isHarnessText =
+        lead.startsWith("<task-notification") || lead.startsWith("[SYSTEM NOTIFICATION");
+      const kind = att.origin && typeof att.origin.kind === "string" ? att.origin.kind : null;
+      const isSystem = isHarnessText || (kind !== null && kind !== "human");
+      return {
+        type: "user",
+        sender: isSystem ? "system" : "user",
+        timestamp: entry.timestamp || att.timestamp || null,
+        content: [{ type: "text", text: truncate(prompt, TEXT_MAX) }],
+        line: num,
+      };
+    }
+
+    const msg = entry.type === "assistant" ? entry.message || {} : {};
+    const content = [];
+
+    if (entry.type === "user") {
+      const msgContent = entry.message?.content;
+      if (typeof msgContent === "string") {
+        content.push({ type: "text", text: truncate(msgContent, TEXT_MAX) });
+      } else if (Array.isArray(msgContent)) {
+        for (const block of msgContent) {
+          if (block.type === "text" && block.text) {
+            content.push({ type: "text", text: truncate(block.text, TEXT_MAX) });
+          } else if (block.type === "tool_result") {
+            // A line carries one tool result; an edit's +/- counts ride on it.
+            const diff = patchCounts(entry.toolUseResult);
+            const answers = askAnswers(entry.toolUseResult);
+            const agent = agentInfo(entry.toolUseResult);
+            // A subagent's report is a reply, read as prose: its text blocks
+            // whole, not a 10 KB slice of their JSON.
+            const report =
+              agent && Array.isArray(block.content)
+                ? block.content
+                    .filter((c) => c?.type === "text" && typeof c.text === "string")
+                    .map((c) => c.text)
+                    .join("\n\n")
+                : null;
+            content.push({
+              type: "tool_result",
+              id: block.tool_use_id || null,
+              output:
+                report !== null
+                  ? truncate(report, TEXT_MAX)
+                  : truncate(
+                      typeof block.content === "string"
+                        ? block.content
+                        : JSON.stringify(block.content || ""),
+                      10240
+                    ),
+              is_error: !!block.is_error,
+              ...(diff ? { diff } : {}),
+              ...(answers ? { answers } : {}),
+              ...(agent ? { agent } : {}),
+            });
+          }
+        }
+      } else if (msgContent === undefined || msgContent === null) {
+        return null;
+      }
+    } else {
+      const msgContent = msg.content || [];
+      if (Array.isArray(msgContent)) {
+        for (const block of msgContent) {
+          if (block.type === "text" && block.text) {
+            content.push({ type: "text", text: truncate(block.text, TEXT_MAX) });
+          } else if (block.type === "thinking" && block.thinking) {
+            content.push({ type: "thinking", text: truncate(block.thinking, 10240) });
+          } else if (block.type === "tool_use") {
+            content.push({
+              type: "tool_use",
+              name: block.name || "unknown",
+              id: block.id || null,
+              // A file a Write/Edit made is an artifact the chat can show in
+              // versions, rebuilt from these inputs: they keep up to 256 KB.
+              input: truncateObj(block.input, EDIT_TOOLS.has(block.name) ? EDIT_INPUT_MAX : 10240),
+            });
+          }
+        }
+      }
+    }
+
+    if (content.length === 0) return null;
+
+    const message = {
+      type: entry.type,
+      sender: classifyTranscriptSender(entry, isSubagentFile),
+      timestamp: entry.timestamp
+        ? typeof entry.timestamp === "number"
+          ? new Date(entry.timestamp).toISOString()
+          : entry.timestamp
+        : null,
+      content,
+      line: num,
+    };
+
+    if (entry.type === "assistant") {
+      if (msg.model) message.model = msg.model;
+      // "end_turn" closes the turn; "tool_use" means more work follows.
+      if (typeof msg.stop_reason === "string") message.stop_reason = msg.stop_reason;
+      if (msg.usage) {
+        message.usage = {
+          input_tokens: msg.usage.input_tokens || 0,
+          output_tokens: msg.usage.output_tokens || 0,
+          cache_read_input_tokens: msg.usage.cache_read_input_tokens || 0,
+          cache_creation_input_tokens: msg.usage.cache_creation_input_tokens || 0,
+        };
+      }
+    }
+
+    return message;
+  }
+  return parseMessage;
+}
+
+module.exports = {
+  classifyTranscriptSender,
+  patchCounts,
+  askAnswers,
+  agentInfo,
+  truncate,
+  truncateObj,
+  TEXT_MAX,
+  createClaudeLineParser,
+};
diff --git a/server/routes/sessions.js b/server/routes/sessions.js
index bfd3f8ae..2cf16e0a 100644
--- a/server/routes/sessions.js
+++ b/server/routes/sessions.js
@@ -22,6 +22,16 @@ const {
   findTranscriptPath,
   findSubagentTranscriptPath,
 } = require("../lib/claude-home");
+const {
+  classifyTranscriptSender,
+  patchCounts,
+  askAnswers,
+  agentInfo,
+  truncate,
+  truncateObj,
+  TEXT_MAX,
+  createClaudeLineParser,
+} = require("../lib/transcript-message");
 
 const router = Router();
 
@@ -52,59 +62,6 @@ const TRANSCRIPT_RENDER_TYPES = new Set([
   "attachment",
 ]);
 
-/**
- * Classify the TRUE sender of a transcript entry. A JSONL `type:"user"` line is
- * not always the human: it also carries tool results, harness-injected
- * task-notifications, /loop re-injections (`isMeta`), and — in a subagent
- * transcript — the task prompt handed down by the orchestrator. Attributing all
- * of those to "User" is wrong; the UI styles each sender distinctly.
- *
- * Returns: "user" | "assistant" | "orchestrator" | "system" | "tool".
- *   user         — a real message typed by the human
- *   assistant    — the agent's own turn
- *   orchestrator — a subagent's task, assigned by its parent/main agent
- *   system       — harness/tooling injection (task-notification, /loop meta, …)
- *   tool         — a tool_result echoed back on a `user` line
- */
-function classifyTranscriptSender(entry, isSubagentFile) {
-  if (entry.type === "assistant") return "assistant";
-  // `system`/local_command lines are surfaced as the human's slash-command I/O.
-  if (entry.type !== "user") return "user";
-
-  const content = entry.message ? entry.message.content : undefined;
-  const onlyToolResults =
-    Array.isArray(content) &&
-    content.length > 0 &&
-    content.every((b) => b && b.type === "tool_result");
-  if (entry.toolUseResult !== undefined || onlyToolResults) return "tool";
-
-  const text =
-    typeof content === "string"
-      ? content
-      : Array.isArray(content)
-        ? (content.find((b) => b && b.type === "text") || {}).text || ""
-        : "";
-  const lead = text.replace(/^\s+/, "");
-
-  // Harness injections that masquerade as a user line.
-  if (entry.isMeta === true) return "system";
-  if (lead.startsWith("<task-notification>") || lead.startsWith("<task-notification ")) {
-    return "system";
-  }
-  // Background-task event banner ("[SYSTEM NOTIFICATION - NOT USER INPUT] …")
-  // that newer harness builds prefix ahead of the <task-notification> payload.
-  if (lead.startsWith("[SYSTEM NOTIFICATION")) return "system";
-
-  // In a subagent transcript, a user line with no human prompt provenance is the
-  // task injected by the Task/Agent tool. A real human message to the subagent
-  // (rare, but allowed) carries promptSource/origin and stays "user".
-  if (isSubagentFile && entry.promptSource === undefined && entry.origin === undefined) {
-    return "orchestrator";
-  }
-
-  return "user";
-}
-
 /**
  * Read only the first non-empty line from a JSONL file using streaming.
  * Avoids loading the entire file into memory.
@@ -866,6 +823,26 @@ router.get("/:id/transcripts", async (req, res) => {
     }
     return res.json({ transcripts: list });
   }
+  // Cursor and Antigravity: one transcript per session, no subagent files.
+  if (session.tool === "cursor" || session.tool === "antigravity") {
+    const file =
+      session.tool === "cursor"
+        ? require("../lib/cursor-usage").transcriptPath(req.params.id)
+        : require("../lib/antigravity-usage").transcriptPath(req.params.id);
+    const list =
+      file && fs.existsSync(file)
+        ? [
+            {
+              id: "main",
+              name: "Main Agent",
+              type: "main",
+              has_transcript: true,
+              db_agent_id: null,
+            },
+          ]
+        : [];
+    return res.json({ transcripts: list });
+  }
   // OpenCode: a subagent is a child session of the store, tied to the agent
   // row its poller made.
   if (session.tool === "opencode") {
@@ -1120,6 +1097,10 @@ router.get("/:id/transcript", async (req, res) => {
   const isCodex = session.tool === "codex";
   // An omp session's (or subagent's) journal, recorded by omp-usage.js.
   const isOmp = session.tool === "omp";
+  // A Cursor session's agent transcript, recorded by cursor-usage.js.
+  const isCursor = session.tool === "cursor";
+  // An agy conversation's step log (`transcript_full.jsonl`), found by antigravity-usage.js.
+  const isAgy = session.tool === "antigravity";
   // An OpenCode conversation lives in its SQLite store, not a JSONL file: read
   // whole (one session's rows), then paged like the files below.
   if (session.tool === "opencode") {
@@ -1136,6 +1117,14 @@ router.get("/:id/transcript", async (req, res) => {
   }
   if (isOmp) {
     jsonlPath = require("../lib/omp-usage").journalPath(req.params.id, agentId || "");
+  } else if (isCursor) {
+    jsonlPath = isSubagentFile
+      ? null
+      : require("../lib/cursor-usage").transcriptPath(req.params.id);
+  } else if (isAgy) {
+    jsonlPath = isSubagentFile
+      ? null
+      : require("../lib/antigravity-usage").transcriptPath(req.params.id);
   } else if (isCodex) {
     jsonlPath = db
       .prepare("SELECT path FROM codex_rollouts WHERE session_id = ? AND thread_id = ?")
@@ -1172,184 +1161,22 @@ router.get("/:id/transcript", async (req, res) => {
       ? (e) => e.type === "response_item"
       : isOmp
         ? (e) => e.type === "message"
-        : (e) => TRANSCRIPT_RENDER_TYPES.has(e.type);
+        : isCursor
+          ? (e) => e.role === "user" || e.role === "assistant"
+          : isAgy
+            ? (e) => AGY_RENDER_TYPES.has(e.type)
+            : (e) => TRANSCRIPT_RENDER_TYPES.has(e.type);
     const toMessage = isCodex
       ? codexMessage
       : isOmp
         ? (e, n) => ompMessage(e, n, isSubagentFile)
-        : (e, n) => parseMessage(e, n);
+        : isCursor
+          ? cursorMessage
+          : isAgy
+            ? agyMessage
+            : (e, n) => parseMessage(e, n);
 
-    // Dedupe state for synthetic rename markers: custom-title lines can repeat
-    // with the same value across a transcript, so only emit when the title
-    // actually changes from the last one we surfaced.
-    let lastRenameTitle = null;
-
-    // Helper: parse a JSONL line into a message object, or null if not a displayable message
-    function parseMessage(entry, num) {
-      // /rename, `claude -n`, picker Ctrl+R → custom-title metadata line. These
-      // commands produce no user/assistant turn, so without this they'd be
-      // invisible. Surface a compact "renamed" marker instead.
-      if (entry.type === "custom-title") {
-        const title = typeof entry.customTitle === "string" ? entry.customTitle.trim() : "";
-        if (!title || title === lastRenameTitle) return null;
-        lastRenameTitle = title;
-        return {
-          type: "session_event",
-          event_kind: "rename",
-          title,
-          timestamp: entry.timestamp || null,
-          content: [],
-          line: num,
-        };
-      }
-
-      // Local slash-command I/O. Newer Claude Code builds write the command
-      // invocation and its captured output as `system`/`local_command` lines
-      // with the TUI markup in a top-level `content` string (older builds used
-      // `user` messages, handled below). Surface those as user-side text so the
-      // client's tuiSegments parser renders the command pill + stdout/stderr
-      // (e.g. /color → "/color" pill + "Session color set to: cyan"). Skip
-      // every other system subtype (turn_duration, stop_hook_summary, …) and
-      // empty local_command lines (e.g. /clear writes a content-less one).
-      if (entry.type === "system") {
-        if (entry.subtype !== "local_command") return null;
-        const sysText = typeof entry.content === "string" ? entry.content : "";
-        if (!sysText.trim()) return null;
-        return {
-          type: "user",
-          sender: "user", // local slash-command I/O is the human's own action
-          timestamp: entry.timestamp || null,
-          content: [{ type: "text", text: truncate(sysText, TEXT_MAX) }],
-          line: num,
-        };
-      }
-
-      // Mid-turn queued message: journaled as a `queued_command` attachment
-      // (never as a `user` line). Surface it at the position the model actually
-      // received it. NOT everything in the queue is the human, though — the
-      // harness delivers its own injections (task-notifications from background
-      // agents, "[SYSTEM NOTIFICATION …]" banners) through the same queue, and
-      // those attachments carry NO `origin` field, while a genuinely typed
-      // message carries `origin.kind = "human"`. So: harness-marker text or a
-      // non-human origin → "system"; everything else → "user". Other attachment
-      // subtypes are harness noise → dropped.
-      if (entry.type === "attachment") {
-        const att = entry.attachment;
-        if (!att || att.type !== "queued_command") return null;
-        const prompt = typeof att.prompt === "string" ? att.prompt : "";
-        if (!prompt.trim()) return null;
-        const lead = prompt.replace(/^\s+/, "");
-        // Same harness markers classifyTranscriptSender strips off user lines.
-        const isHarnessText =
-          lead.startsWith("<task-notification") || lead.startsWith("[SYSTEM NOTIFICATION");
-        const kind = att.origin && typeof att.origin.kind === "string" ? att.origin.kind : null;
-        const isSystem = isHarnessText || (kind !== null && kind !== "human");
-        return {
-          type: "user",
-          sender: isSystem ? "system" : "user",
-          timestamp: entry.timestamp || att.timestamp || null,
-          content: [{ type: "text", text: truncate(prompt, TEXT_MAX) }],
-          line: num,
-        };
-      }
-
-      const msg = entry.type === "assistant" ? entry.message || {} : {};
-      const content = [];
-
-      if (entry.type === "user") {
-        const msgContent = entry.message?.content;
-        if (typeof msgContent === "string") {
-          content.push({ type: "text", text: truncate(msgContent, TEXT_MAX) });
-        } else if (Array.isArray(msgContent)) {
-          for (const block of msgContent) {
-            if (block.type === "text" && block.text) {
-              content.push({ type: "text", text: truncate(block.text, TEXT_MAX) });
-            } else if (block.type === "tool_result") {
-              // A line carries one tool result; an edit's +/- counts ride on it.
-              const diff = patchCounts(entry.toolUseResult);
-              const answers = askAnswers(entry.toolUseResult);
-              const agent = agentInfo(entry.toolUseResult);
-              // A subagent's report is a reply, read as prose: its text blocks
-              // whole, not a 10 KB slice of their JSON.
-              const report =
-                agent && Array.isArray(block.content)
-                  ? block.content
-                      .filter((c) => c?.type === "text" && typeof c.text === "string")
-                      .map((c) => c.text)
-                      .join("\n\n")
-                  : null;
-              content.push({
-                type: "tool_result",
-                id: block.tool_use_id || null,
-                output:
-                  report !== null
-                    ? truncate(report, TEXT_MAX)
-                    : truncate(
-                        typeof block.content === "string"
-                          ? block.content
-                          : JSON.stringify(block.content || ""),
-                        10240
-                      ),
-                is_error: !!block.is_error,
-                ...(diff ? { diff } : {}),
-                ...(answers ? { answers } : {}),
-                ...(agent ? { agent } : {}),
-              });
-            }
-          }
-        } else if (msgContent === undefined || msgContent === null) {
-          return null;
-        }
-      } else {
-        const msgContent = msg.content || [];
-        if (Array.isArray(msgContent)) {
-          for (const block of msgContent) {
-            if (block.type === "text" && block.text) {
-              content.push({ type: "text", text: truncate(block.text, TEXT_MAX) });
-            } else if (block.type === "thinking" && block.thinking) {
-              content.push({ type: "thinking", text: truncate(block.thinking, 10240) });
-            } else if (block.type === "tool_use") {
-              content.push({
-                type: "tool_use",
-                name: block.name || "unknown",
-                id: block.id || null,
-                input: truncateObj(block.input, 10240),
-              });
-            }
-          }
-        }
-      }
-
-      if (content.length === 0) return null;
-
-      const message = {
-        type: entry.type,
-        sender: classifyTranscriptSender(entry, isSubagentFile),
-        timestamp: entry.timestamp
-          ? typeof entry.timestamp === "number"
-            ? new Date(entry.timestamp).toISOString()
-            : entry.timestamp
-          : null,
-        content,
-        line: num,
-      };
-
-      if (entry.type === "assistant") {
-        if (msg.model) message.model = msg.model;
-        // "end_turn" closes the turn; "tool_use" means more work follows.
-        if (typeof msg.stop_reason === "string") message.stop_reason = msg.stop_reason;
-        if (msg.usage) {
-          message.usage = {
-            input_tokens: msg.usage.input_tokens || 0,
-            output_tokens: msg.usage.output_tokens || 0,
-            cache_read_input_tokens: msg.usage.cache_read_input_tokens || 0,
-            cache_creation_input_tokens: msg.usage.cache_creation_input_tokens || 0,
-          };
-        }
-      }
-
-      return message;
-    }
+    const parseMessage = createClaudeLineParser(isSubagentFile);
 
     if (afterLine !== null) {
       // Incremental mode: skip lines until after afterLine, collect up to limit, then stop.
@@ -1626,6 +1453,131 @@ function ompMessage(entry, num, isSubagent = false) {
   }
 }
 
+/**
+ * One line of a Cursor agent transcript (`{role, message:{content}}`, see
+ * cursor-usage.js) as a transcript message. Cursor writes no timestamps and no
+ * model, so neither is set. A user line wraps the typed prompt in
+ * `<user_query>` after a `<timestamp>` block; only the prompt is shown.
+ */
+function cursorMessage(entry, num) {
+  const m = entry.message || {};
+  const parts = Array.isArray(m.content)
+    ? m.content
+    : typeof m.content === "string"
+      ? [{ type: "text", text: m.content }]
+      : [];
+  const msg = (type, sender, content) => ({ type, sender, timestamp: null, content, line: num });
+  const content = [];
+  if (entry.role === "user") {
+    for (const c of parts) {
+      if (c && c.type === "text" && typeof c.text === "string") {
+        const q = /<user_query>\s*([\s\S]*?)\s*<\/user_query>/.exec(c.text);
+        const text = (q ? q[1] : c.text.replace(/<timestamp>[\s\S]*?<\/timestamp>/, "")).trim();
+        if (text) content.push({ type: "text", text: truncate(text, TEXT_MAX) });
+      } else if (c && c.type === "tool_result") {
+        const out = Array.isArray(c.content)
+          ? c.content.map((x) => (x && typeof x.text === "string" ? x.text : "")).join("")
+          : typeof c.content === "string"
+            ? c.content
+            : "";
+        content.push({
+          type: "tool_result",
+          id: c.tool_use_id || null,
+          output: truncate(out, 10240),
+          is_error: c.is_error === true,
+        });
+      }
+    }
+    if (!content.length) return null;
+    const onlyResults = content.every((c) => c.type === "tool_result");
+    return msg("user", onlyResults ? "tool" : "user", content);
+  }
+  for (const c of parts) {
+    if (!c) continue;
+    if (c.type === "text" && typeof c.text === "string" && c.text.trim()) {
+      content.push({ type: "text", text: truncate(c.text, TEXT_MAX) });
+    } else if (c.type === "thinking" && typeof c.thinking === "string" && c.thinking.trim()) {
+      content.push({ type: "thinking", text: truncate(c.thinking, 10240) });
+    } else if (c.type === "tool_use") {
+      content.push({
+        type: "tool_use",
+        name: c.name || "unknown",
+        id: c.id || null,
+        input: truncateObj(c.input && typeof c.input === "object" ? c.input : {}, 10240),
+      });
+    }
+  }
+  return content.length ? msg("assistant", "assistant", content) : null;
+}
+
+/** The agy steps a conversation view shows (see `agyMessage`). */
+const AGY_RENDER_TYPES = new Set(["USER_INPUT", "PLANNER_RESPONSE", "GENERIC"]);
+
+/**
+ * One step of an agy conversation (`brain/<id>/.system_generated/logs/
+ * transcript_full.jsonl`, see antigravity-usage.js) as a transcript message:
+ * `USER_INPUT` the prompt (inside `<USER_REQUEST>`), `PLANNER_RESPONSE` the
+ * model's text, thinking, tool calls and tokens, `GENERIC` a tool's result
+ * (`status: "INVALID"` + `error` an error). agy gives a call no id: its result
+ * is the step right after it (call `i` of step `s` lands at step `s + 1 + i`),
+ * so both are named by that step — the same on every page of the file.
+ */
+function agyMessage(entry, num) {
+  const at = typeof entry.created_at === "string" ? entry.created_at : null;
+  const step = Number.isInteger(entry.step_index) ? entry.step_index : num;
+  const msg = (type, sender, content) => ({ type, sender, timestamp: at, content, line: num });
+  const text = typeof entry.content === "string" ? entry.content : "";
+  switch (entry.type) {
+    case "USER_INPUT": {
+      const req = /<USER_REQUEST>\s*([\s\S]*?)\s*<\/USER_REQUEST>/.exec(text);
+      const prompt = (req ? req[1] : text).trim();
+      if (!prompt) return null;
+      return msg("user", entry.source === "USER_EXPLICIT" ? "user" : "system", [
+        { type: "text", text: truncate(prompt, TEXT_MAX) },
+      ]);
+    }
+    case "PLANNER_RESPONSE": {
+      const content = [];
+      if (typeof entry.thinking === "string" && entry.thinking.trim())
+        content.push({ type: "thinking", text: truncate(entry.thinking, 10240) });
+      if (text.trim()) content.push({ type: "text", text: truncate(text, TEXT_MAX) });
+      (Array.isArray(entry.tool_calls) ? entry.tool_calls : []).forEach((c, i) => {
+        if (!c) return;
+        content.push({
+          type: "tool_use",
+          name: c.name || "unknown",
+          id: `agy-${step + 1 + i}`,
+          input: truncateObj(c.args && typeof c.args === "object" ? c.args : {}, 10240),
+        });
+      });
+      if (!content.length) return null;
+      const out = msg("assistant", "assistant", content);
+      if (Number.isFinite(entry.input_tokens) || Number.isFinite(entry.output_tokens))
+        out.usage = {
+          input_tokens: entry.input_tokens || 0,
+          output_tokens: entry.output_tokens || 0,
+          cache_read_input_tokens: entry.cache_read_tokens || 0,
+          cache_creation_input_tokens: 0,
+        };
+      return out;
+    }
+    case "GENERIC": {
+      // Its first lines say when the call ran: not part of what it returned.
+      const output = text.replace(/^Created At: [^\n]*\nCompleted At: [^\n]*\n\n?/, "");
+      return msg("user", "tool", [
+        {
+          type: "tool_result",
+          id: `agy-${step}`,
+          output: truncate(output, 10240),
+          is_error: entry.status === "INVALID" || entry.status === "ERROR" || !!entry.error,
+        },
+      ]);
+    }
+    default:
+      return null;
+  }
+}
+
 /**
  * The transcript route's paging over a conversation already in memory (an
  * OpenCode store): message N is "line" N, so `after`/`before`/`offset` and the
@@ -1662,75 +1614,6 @@ function pageMessages(all, { limit, afterLine, beforeLine, offset }) {
   };
 }
 
-/** The +/- line counts of a Write/Edit/MultiEdit result (`toolUseResult`):
- *  its `structuredPatch` hunks' `+`/`-` lines, or every line of a file it
- *  created. Null for any other result. */
-function patchCounts(r) {
-  if (!r || typeof r !== "object") return null;
-  if (Array.isArray(r.structuredPatch) && r.structuredPatch.length) {
-    let added = 0;
-    let removed = 0;
-    for (const h of r.structuredPatch) {
-      for (const l of Array.isArray(h?.lines) ? h.lines : []) {
-        if (typeof l !== "string") continue;
-        if (l.startsWith("+")) added++;
-        else if (l.startsWith("-")) removed++;
-      }
-    }
-    return { added, removed };
-  }
-  if (r.type === "create" && typeof r.content === "string") {
-    return { added: r.content.split("\n").length - (r.content.endsWith("\n") ? 1 : 0), removed: 0 };
-  }
-  return null;
-}
-
-/** An AskUserQuestion result's answers (`toolUseResult.answers`): question
- *  text → the option picked, or what the user typed. Null for any other result. */
-function askAnswers(r) {
-  if (!r || typeof r !== "object" || !r.answers || typeof r.answers !== "object") return null;
-  if (Array.isArray(r.answers)) return null;
-  const out = {};
-  for (const [q, a] of Object.entries(r.answers)) if (typeof a === "string") out[q] = a;
-  return Object.keys(out).length ? out : null;
-}
-
-/** An Agent/Task result's facts (`toolUseResult`): the subagent's id (its
- *  transcript is `agent-<id>.jsonl`), the model it ran on and, once it
- *  finished in the foreground, its time, tokens and tool calls. Null for any
- *  other result. */
-function agentInfo(r) {
-  if (!r || typeof r !== "object" || typeof r.agentId !== "string") return null;
-  if (!/^[\w-]{1,200}$/.test(r.agentId)) return null;
-  const num = (v) => (Number.isFinite(v) ? v : null);
-  return {
-    id: r.agentId,
-    model: typeof r.resolvedModel === "string" ? r.resolvedModel : null,
-    status: typeof r.status === "string" ? r.status : null,
-    duration_ms: num(r.totalDurationMs),
-    tokens: num(r.totalTokens),
-    tool_uses: num(r.totalToolUseCount),
-  };
-}
-
-// A conversation turn — a prompt, an answer — is shown whole: it is bounded by
-// the model's output and what the user typed. Only tool payloads and thinking
-// (the parts that can run to megabytes) keep the 10 KB cap. This one guards
-// against a pathological line, not a long answer.
-const TEXT_MAX = 1024 * 1024;
-
-function truncate(str, maxLen) {
-  if (!str || str.length <= maxLen) return str;
-  return str.slice(0, maxLen) + "[truncated]";
-}
-
-function truncateObj(obj, maxLen) {
-  if (!obj) return obj;
-  const json = JSON.stringify(obj);
-  if (json.length <= maxLen) return obj;
-  return { _truncated: truncate(json, maxLen) };
-}
-
 module.exports = router;
 module.exports.sessionStats = sessionStats;
 // Exported for unit tests — sender attribution is correctness-critical.