fix(tui): reduce cached transcript remount work

packages/tui/perf/tabs.md

@@ -0,0 +1,106 @@
+# Session Tab Switching
+
+## Run
+
+Install OpenCode Drive, then run from the repository root:
+
+```sh
+PERF_RUN=before \
+OPENCODE_DRIVE_MEDIA_DIR="$PWD/.cache/tui-switch/media" \
+opencode-drive run script/bench-tui-tabs.ts
+```
+
+Use a new `PERF_RUN` for each run; existing result directories are not overwritten.
+`PERF_TARGET` selects another source worktree and defaults to the working directory.
+`PERF_OUTPUT` overrides the result root, which defaults to `<target>/.cache/tui-switch`.
+`PERF_CONTENT=prose` replaces the Markdown with equal-byte-size plain text.
+
+Drive checks the script, creates an isolated server/home/project, imports synthetic
+sessions through the real API, and launches the real TUI components. It never
+connects to the elected background service. Only the final streaming correctness
+check prompts a model, and that model is simulated.
+
+Run benchmarks serially, without simultaneous tests or builds. The script records
+the target revision and its production TUI/Client diff, every completed action in
+`samples.jsonl`, summary statistics, terminal frames, and Drive artifact metadata.
+It retains failed runs' completed samples. Result and media directories are local
+artifacts, not files to commit.
+
+## Workload
+
+- SHORT: 20 messages, 256 text bytes per assistant.
+- LONG: 2,000 messages with the same text sizes and a comparable latest page.
+- LARGE: 20 messages, 32 KiB per assistant.
+- Every fifth assistant includes a completed synthetic read-tool result.
+- Historical fixtures carry creation/completion times but omit stream-end/token
+  accounting. Correctness tests exercise complete timing and token metadata.
+- Markdown deliberately repeats small fenced blocks, about 170 per LARGE
+  assistant. It is a stress fixture, not a typical-response latency claim.
+
+The TUI starts after import. It opens sessions through the picker, switches with
+the real keybindings, loads all LONG history, and returns to both its tail and a
+saved head anchor. Each warm category has one explicitly retained warm-up sample
+(`sample: 0`) and eight measured observations. Initial opens and first/last-message
+navigation are single observations and are not included in warm medians.
+
+Location caveat: this runner supplies `info.location` but omits the import API's
+top-level `location`. Imported sessions therefore use the isolated server's
+working directory, not the fixture `files` directory. The results describe warm
+cross-Location sessions, not same-Location restoration or cold project loading.
+Both directories are private synthetic fixtures; no live user Location is used.
+
+`actionMs` is Drive's action RPC duration, including its forced render and UI-tree
+inspection. `visibleMs` additionally waits for a destination marker, with 20 ms
+polling. Neither is physical-terminal input-to-paint latency or a styled-content
+completion guarantee. Fixed inter-action pacing occurs outside the measured
+interval and is not used instead of readiness checks.
+
+The final check streams an incomplete ordinary fence, completes it, verifies its
+final displayed text, and checks the server's completed assistant projection.
+Normal tests separately cover custom Markdown and footer correctness.
+
+## Initial Experiments
+
+Base: `849824efd2`, Bun 1.3.14, OpenTUI 0.5.9, Apple M2 Max, 120x40, source builds,
+DevTools disabled. All results below are local action-RPC medians in milliseconds,
+eight measured observations per cell. They are not release-binary guarantees.
+
+| Scenario                       | Base repeat | Markdown ordering | Ordering + footer index | Footer confirmation |
+| ------------------------------ | ----------: | ----------------: | ----------------------: | ------------------: |
+| Latest 20 of LONG retained     |        28.6 |              28.8 |                    26.6 |                26.2 |
+| All 2,000 retained, tail       |       107.7 |             107.9 |                    66.9 |                60.5 |
+| All 2,000 retained, saved head |       115.3 |             118.7 |                    71.9 |                69.3 |
+| Dense Markdown                 |     1,615.5 |             832.1 |                   830.4 |               820.6 |
+
+These are the `before-02`, `markdown-02`, `footer-01`, and `footer-02` runs. The earlier
+base/Markdown pair independently measured 1,717.3 -> 842.6 ms for dense Markdown;
+that earlier runner did not yet include the saved-head category, so its results
+are not pooled into the table. Ordinary short/latest-page differences are near
+the noise floor and are not claimed as improvements.
+
+### Kept
+
+- Configure Markdown's custom renderer before content. OpenTUI's `renderNode`
+  setter otherwise clears populated parse/block state and repeats preparation.
+  Keep content before `streaming` so the completion update retains final tokens.
+- Share a reactive message-position index across a Session view's footers. Scan
+  only from each position to its preceding user/synthetic input instead of
+  searching and slicing entire history prefixes. This improves both tail and
+  historical positions rather than shifting work to the newer suffix.
+
+Captured content, geometry, and styling matched across the base, Markdown-only,
+and both footer trials after excluding the isolated-project path footer.
+
+The helper dependency test deliberately supplies a known position. It verifies
+that the bounded calculation does not read an unrelated prefix, not that the
+whole Session ignores structural history changes. A real-App regression checks
+the reactive index through prepend/reconcile and a same-length truncate/append.
+
+### Deferred
+
+- The second row reduction after a cache-hit sync still exists. Removing it
+  cleanly needs an explicit cache-hit/synchronization contract; this pass does
+  not change the public Client data API or infer freshness from array identity.
+- Parsed Markdown caches, mounted-view retention, history eviction, and initial
+  window changes were not mixed into these experiments.
+- No memory-leak or retained-heap improvement is claimed by these latency runs.

packages/tui/src/routes/session/index.tsx

@@ -141,6 +141,7 @@ const context = createContext<{
   groupExploration: () => boolean
   diffWrapMode: () => "word" | "none"
   models: () => ModelInfo[]
+  messageIndex: (messageID: string) => number | undefined
   config: ReturnType<typeof useConfig>["data"]
   mutatePending: (action: PendingAction, inboxID: string) => Promise<boolean>
   pendingDelivery: (inboxID: string) => SessionInbox.Delivery | undefined
@@ -180,6 +181,7 @@ export function Session(props: {
   const promptRef = usePromptRef()
   const session = createMemo(() => data.session.get(route.sessionID))
   const messages = () => data.session.message.list(route.sessionID)
+  const messageIndexes = createMemo(() => new Map(messages().map((message, index) => [message.id, index])))
   const messagesBeforeRevert = () => {
     const messageID = session()?.revert?.messageID
     if (!messageID) return messages()
@@ -1349,6 +1351,7 @@ export function Session(props: {
         groupExploration,
         diffWrapMode,
         models,
+        messageIndex: (messageID) => messageIndexes().get(messageID),
         config,
         mutatePending,
         pendingDelivery: (inboxID) => pendingDeliveries().get(inboxID),
@@ -2031,8 +2034,10 @@ function AssistantFooter(props: { message: SessionMessageAssistant }) {
         ?.name ?? `${props.message.model.providerID}/${props.message.model.id}`,
   )
   const messages = createMemo(() => data.session.message.list(ctx.sessionID))
-  const duration = createMemo(() => turnDuration(props.message, messages()))
-  const tokensPerSecond = createMemo(() => turnTokensPerSecond(props.message, messages()))
+  const duration = createMemo(() => turnDuration(props.message, messages(), ctx.messageIndex(props.message.id)))
+  const tokensPerSecond = createMemo(() =>
+    turnTokensPerSecond(props.message, messages(), ctx.messageIndex(props.message.id)),
+  )
   const interrupted = createMemo(() => props.message.error?.message === "Step interrupted")
   return (
     <>
@@ -2211,14 +2216,14 @@ function CompactionMessage(props: { message: Extract<SessionMessageInfo, { type:
         <box paddingTop={1} paddingLeft={3}>
           <markdown
             syntaxStyle={syntax()}
+            renderNode={plugins.markdown()}
             streaming={true}
             internalBlockMode="top-level"
             content={content()}
             tableOptions={{ style: "grid", cellPaddingX: 1 }}
             conceal={ctx.markdownMode() === "rendered"}
             fg={theme.markdown.text}
             bg={theme.background.default}
-            renderNode={plugins.markdown()}
           />
         </box>
       </Show>
@@ -2658,17 +2663,17 @@ function TextPart(props: { last: boolean; part: SessionMessageAssistantText; mes
   return (
     <Show when={props.part.text.trim()}>
       <box paddingLeft={3} flexShrink={0}>
-        {/* Apply content before streaming so completion does not freeze the previous Markdown tokens. */}
+        {/* Configure custom nodes before parsing; apply content before streaming so completion keeps the final tokens. */}
         <markdown
           syntaxStyle={syntax()}
+          renderNode={plugins.markdown()}
           content={props.part.text.trim()}
           streaming={props.message.time.completed === undefined}
           internalBlockMode="top-level"
           tableOptions={{ style: "grid", cellPaddingX: 1 }}
           conceal={ctx.markdownMode() === "rendered"}
           fg={theme.markdown.text}
           bg={theme.background.default}
-          renderNode={plugins.markdown()}
         />
       </box>
     </Show>

packages/tui/src/routes/session/rows.ts

@@ -346,21 +346,21 @@ export function cacheReuseDrop(previous: CacheUsage | undefined, current: CacheU
   return drop > 0 ? drop : undefined
 }
 
-export function turnDuration(message: SessionMessageAssistant, messages: SessionMessageInfo[]) {
+export function turnDuration(message: SessionMessageAssistant, messages: SessionMessageInfo[], position?: number) {
   if (message.time.completed === undefined) return 0
-  const index = messages.findIndex((item) => item.id === message.id)
-  const input = messages
-    .slice(0, index === -1 ? messages.length : index)
-    .findLast((item) => item.type === "user" || item.type === "synthetic")
+  const index = position ?? messages.findIndex((item) => item.id === message.id)
+  const input = messages[inputIndex(messages, index === -1 ? messages.length : index)]
   return Math.max(0, message.time.completed - (input?.time.created ?? message.time.created))
 }
 
-export function turnTokensPerSecond(message: SessionMessageAssistant, messages: SessionMessageInfo[]) {
-  const index = messages.findIndex((item) => item.id === message.id)
+export function turnTokensPerSecond(
+  message: SessionMessageAssistant,
+  messages: SessionMessageInfo[],
+  position?: number,
+) {
+  const index = position ?? messages.findIndex((item) => item.id === message.id)
   const end = index === -1 ? messages.length : index + 1
-  const start = messages
-    .slice(0, end)
-    .findLastIndex((item) => item.type === "user" || item.type === "synthetic")
+  const start = inputIndex(messages, end)
   const steps = messages
     .slice(start + 1, end)
     .filter((item): item is SessionMessageAssistant => item.type === "assistant")
@@ -374,6 +374,15 @@ export function turnTokensPerSecond(message: SessionMessageAssistant, messages:
   return output / (duration / 1_000)
 }
 
+function inputIndex(messages: SessionMessageInfo[], end: number) {
+  // Reading a sliced prefix subscribes every footer to unrelated historical messages.
+  for (let index = end - 1; index >= 0; index--) {
+    const message = messages[index]
+    if (message.type === "user" || message.type === "synthetic") return index
+  }
+  return -1
+}
+
 function hasTokenUsage(
   message: SessionMessageAssistant,
 ): message is SessionMessageAssistant & { tokens: NonNullable<SessionMessageAssistant["tokens"]> } {