dsh-code 1.0.5 → 1.0.7

package/src/render/projection.ts

@@ -1,1621 +1,1974 @@
-/**
- * Pure session-event-to-view projection for the TUI transcript: one reducer
- * over {@link SessionEvent}s producing the ordered entries the renderer draws.
- * Rendering never reads the session directly — this module owns the view
- * model, so tests drive it with plain event arrays.
- *
- * @module @deepseek-ai/dsh-tui/render/projection
- */
-
-import { boundContextSummary, type ContentBlock, type ImageBlock, type MessageId } from '@deepseek-ai/dsh-llm'
-import type { SessionEvent } from '@deepseek-ai/dsh-session'
-import type { TodoItem } from '@deepseek-ai/dsh-tool-todo'
-import { graphemeWidth, splitGraphemes } from './width.ts'
-// Type-only imports merge the plugin-owned SessionEventMap variants
-// (agent/inbox/spliced, command/*, compaction/*, goal/change, llm/retry*,
-// plan/mode, permission/preset, sandbox/mode, session/title) into the union
-// this reducer switches on.
-import type {} from '@deepseek-ai/dsh-agent'
-import type {} from '@deepseek-ai/dsh-commands'
-import type {} from '@deepseek-ai/dsh-compaction'
-import type {} from '@deepseek-ai/dsh-goal'
-import type {} from '@deepseek-ai/dsh-llm-retry'
-import type {} from '@deepseek-ai/dsh-plan-mode'
-import type {} from '@deepseek-ai/dsh-permission-presets'
-import type {} from '@deepseek-ai/dsh-sandbox-policy'
-import type {} from '@deepseek-ai/dsh-session-title'
-import { toolArgumentsPreview, toolPromptPreview } from './tool-preview.ts'
-import { toolResultDetail, type ToolDetail } from './tool-detail.ts'
-
-/** In-flight UI buffers are tails; the assembled assistant message is authoritative. */
-const MAX_STREAMING_CHARS = 65_536
-
-/**
- * Upper bound on remembered `compaction/summary` shadow prices waiting for a
- * matching `compaction/end`. Compactions are sequential and rare, so a few
- * slots suffice; an aborted compaction (summary without end) otherwise leaves
- * an unbounded residue in `anchors.compactionTokens`. An evicted price
- * degrades to the documented `lastPruneTokens` fallback, exactly like a
- * missing summary.
- */
-const MAX_COMPACTION_SUMMARY_RESIDUE = 16
-
-/** Append one delta without retaining an unbounded duplicate of the live reply. */
-function appendStreamingTail(current: string, delta: string): string {
-  const next = current + delta
-  return next.length <= MAX_STREAMING_CHARS ? next : next.slice(-MAX_STREAMING_CHARS)
-}
-
-/** One user prompt line. */
-export interface UserEntry {
-  kind: 'user'
-  /** Joined text blocks of the user message. */
-  text: string
-  /** True for collapsed injected context (plugin/continuation notices), which
-   * the renderer marks with a dim ↳ instead of the user ❯ prompt. */
-  notice: boolean
-  /** Durable image references carried by this prompt. */
-  images?: readonly ImageBlock['attachment'][]
-}
-
-/** One user message waiting in the agent inbox (the web's queued-message row). */
-export interface PendingEntry {
-  kind: 'pending'
-  /** Stable message identity shared with the durable `user/message` that retires it. */
-  messageId: MessageId
-  /** Which inbox list holds the message: steering is consumed at the next step boundary. */
-  target: 'next-turn' | 'next-step'
-  /** Full message text — Codex PendingSteer renders queued prompts exactly like user rows. */
-  text: string
-  /** Durable image references queued with this prompt. */
-  images?: readonly ImageBlock['attachment'][]
-}
-
-/** One authoritative assembled assistant reply. */
-export interface AssistantEntry {
-  kind: 'assistant'
-  /** Joined text blocks of the assistant message. */
-  text: string
-  /** Joined reasoning blocks from the same assembled message. */
-  reasoning: string
-  /** True when a cancelled stream's delivered prefix was finalized as this
-   * entry (rc.8 `assistant/message.interrupted`) — rendered with a marker. */
-  interrupted?: true
-}
-
-/** One model-requested tool invocation and its settled state. */
-export interface ToolEntry {
-  kind: 'tool'
-  /** Correlation id shared with the matching `tool/result`. */
-  callId: string
-  /**
-   * Global tool-call ordinal across the whole transcript (1, 2, 3…, never
-   * reset between turns). The tool-card badge and every error line that
-   * references the failed call share this number, so "call N" in an error
-   * always names the exact card the badge shows.
-   */
-  ordinal: number
-  /** Tool name as the model addressed it. */
-  name: string
-  /** Raw arguments JSON string exactly as the model produced it. */
-  arguments: string
-  /** Bounded human-meaningful arguments preview for the tool card. */
-  preview: string
-  /** Bounded delegation prompt (subagent cards' second row), '' when none. */
-  prompt: string
-  /** Execution state; `running` until the paired result lands. */
-  state: 'running' | 'done' | 'error'
-  /** Bounded first text block of the result, empty until it lands. */
-  summary: string
-  /**
-   * Bounded expansion payload for the verbose transcript (Ctrl+O), derived
-   * from the tool's persisted presentation metadata; undefined until the
-   * result lands and only when something renderable exists.
-   */
-  detail: ToolDetail | undefined
-}
-
-/** One slash-command execution dispatched through `ctx.commands`. */
-export interface CommandEntry {
-  kind: 'command'
-  /** Pairing id shared with the matching `command/done`. */
-  commandId: string
-  /** Lowercase command name without the leading slash. */
-  name: string
-  /** Verbatim text following the command name. */
-  args: string
-  /** Execution state; `running` until the paired lifecycle event lands. */
-  state: 'running' | 'done' | 'error'
-  /** Handler outcome text, empty until it lands. */
-  summary: string
-}
-
-/** One turn-level failure surfaced from `turn/end`. */
-export interface ErrorEntry {
-  kind: 'error'
-  /** `code: message` of the failure. */
-  text: string
-}
-
-/** One non-error turn outcome surfaced from `turn/end`. */
-export interface TurnMarkerEntry {
-  kind: 'turn-marker'
-  /** Human-readable outcome line, dim-rendered. */
-  text: string
-}
-
-/** One completed compaction lifecycle surfaced from `compaction/end`. */
-export interface CompactionEntry {
-  kind: 'compaction'
-  /** True when the compaction completed, false when it failed. */
-  ok: boolean
-  /** Heuristic tokens shadowed by the compaction (summary or prune price). */
-  tokens: number
-  /** Failure text when `ok` is false, empty otherwise. */
-  error: string
-}
-
-/** One provider-routed model-request retry (the `llm/retry` pair). */
-export interface RetryEntry {
-  kind: 'retry'
-  /** Correlation id shared with the matching `llm/retry-started`. */
-  retryId: string
-  /** Retry policy mode from the event: `always` has no attempt cap. */
-  mode: 'normal' | 'always'
-  /** Attempt ordinal and its cap. */
-  attempt: number
-  max: number
-  /** Failure code that triggered the retry. */
-  code: string
-  /** Backoff wait before the next attempt, in ms. */
-  delayMs: number
-  /**
-   * `running` while the backoff waits, `done` once the attempt started — or
-   * when the turn ended first (the turn-end sweep finalizes orphans so they
-   * never pin the settled boundary).
-   */
-  state: 'running' | 'done'
-}
-
-/** Turn-tail deliverables: files mutated by the turn's diff-bearing tools. */
-export interface FilesEntry {
-  kind: 'files'
-  /** Unique mutated paths in call order, bounded. */
-  paths: readonly string[]
-}
-
-/** Ordered transcript items the renderer draws. */
-export type TranscriptEntry = UserEntry | PendingEntry | AssistantEntry | ToolEntry | CommandEntry | ErrorEntry | TurnMarkerEntry | CompactionEntry | RetryEntry | FilesEntry
-
-/** The live goal the status line badges, folded from `goal/change`. */
-export interface GoalFold {
-  /** Human-requested completion objective. */
-  objective: string
-  /** Durable lifecycle phase. */
-  phase: 'active' | 'paused' | 'blocked' | 'complete'
-  /** Highest admitted continuation round and its cap. */
-  rounds: number
-  max: number
-  /** Blocked explanation, empty outside the blocked phase. */
-  blocked: string
-}
-
-/** Cumulative token accounting folded from `assistant/message` usage reports. */
-export interface UsageTotals {
-  /** Prompt-side billed tokens: `inputTokens` plus both cache buckets. */
-  inputTokens: number
-  /** Completion-side tokens over the whole log. */
-  outputTokens: number
-  /** Cache-read tokens over the whole log (0 when the adapter reports none). */
-  cacheReadTokens: number
-}
-
-/**
- * Estimated used tokens per context content type, folded from transcript
- * events via {@link estimateTokens}. The segmented context bar's composition
- * source: proportions across types are meaningful, absolute values are not
- * (they never touch billing or the reported `lastPromptTokens`).
- */
-export interface ContextSegments {
-  /** Rendered system-prompt text (latest `request/header`) plus injected-context notices. */
-  system: number
-  /** Direct human prompts (durable `user/message` rows). */
-  prompt: number
-  /** Assistant text blocks (visible replies). */
-  assistant: number
-  /** Assistant reasoning blocks (hidden thinking). */
-  thinking: number
-  /** Tool call arguments plus result text. */
-  tools: number
-}
-
-/** Window-scoped figures the status line shows; timing uses event timestamps. */
-export interface TranscriptStats {
-  /** Durable turns opened (`turn/start` events). */
-  turns: number
-  /** Model requests made (`step/start` events). */
-  steps: number
-  /** Summed model wall time: `step/start` → `assistant/message`, in ms. */
-  llmMs: number
-  /** Summed tool wall time: `tool/call` → `tool/result`, in ms. */
-  toolMs: number
-  /** Cumulative token accounting; input stays 0 until a report lands. */
-  usage: UsageTotals
-  /** Prompt-side size of the most recent reported request (context pressure). */
-  lastPromptTokens: number
-  /** Newest advertised route capacity, 0 when no adapter ever advertised one. */
-  contextWindow: number
-  /** Estimated used tokens per content type (the segmented bar's composition). */
-  contextSegments: ContextSegments
-  /** Summed first-token waits: `step/start` → first non-empty chunk, in ms. */
-  ttftMs: number
-  /** Steps that produced a first chunk (the TTFT average's denominator). */
-  ttftSteps: number
-  /** Summed decode spans: first chunk → `assistant/message`, in ms. */
-  decodeMs: number
-  /** Completion tokens over timed decode spans (the tok/s numerator). */
-  decodeTokens: number
-  /**
-   * Adapter-owned reasoning effort of the latest `request/header` config —
-   * the EFFECTIVE effort the session actually uses (a materialized model
-   * default is included, exactly as the adapter resolved it). Empty when the
-   * header carried none (provider-default behavior). The status line appends
-   * it to the model segment as `provider/model@effort`.
-   */
-  reasoningEffort: string
-}
-
-/** The complete TUI transcript view for one session. */
-export interface TranscriptView {
-  /** Settled entries in log order. */
-  entries: readonly TranscriptEntry[]
-  /** Bounded text tail accumulated from `assistant/chunk` deltas since the last flush. */
-  streaming: string
-  /** Bounded thinking tail accumulated from reasoning deltas since the last flush. */
-  streamingReasoning: string
-  /** Latest whole-list todo snapshot from `todo/write`, empty when none. */
-  todos: readonly TodoItem[]
-  /**
-   * Global tool-call ordinal counter: the number the NEXT `tool/call` lands
-   * with (1-based). Never reset, so the counter and the badges/error lines
-   * stay consistent across turns and resumed sessions.
-   */
-  toolCallOrdinal: number
-  /** True while a durable turn is open (`turn/start` … `turn/end`). */
-  busy: boolean
-  /** `turn/start` time of the open turn (0 while idle) — the web TurnStatus clock anchor. */
-  busySince: number
-  /** Figures the status line renders. */
-  stats: TranscriptStats
-  /**
-   * The `provider/model` pair of the last `request/header` snapshot — the
-   * session's own model record, which a resumed TUI prefers over the
-   * deployment default (mirrors the web host's resume selection order).
-   * Empty before the session's first request.
-   */
-  model: string
-  /** Plan mode state folded from the last `plan/mode` event. */
-  plan: boolean
-  /** Active permission preset folded from the last `permission/preset` event, empty before one. */
-  permission: string
-  /** Latest session title folded from the last `session/title` event, empty before one. */
-  title: string
-  /** Sandbox-mode override folded from the last `sandbox/mode` event, empty when never switched. */
-  sandbox: string
-  /** Current long-running goal folded from the last `goal/change`, undefined when cleared. */
-  goal: GoalFold | undefined
-  /**
-   * Ordered live message ids per inbox target, mirrored from
-   * `agent/inbox/spliced` exactly like the upstream Inbox projection — the
-   * coordinates later removals resolve against.
-   */
-  pending: { 'next-turn': readonly string[]; 'next-step': readonly string[] }
-  /**
-   * Fold-internal timing anchors, never rendered: open step and tool-call
-   * start timestamps the next `assistant/message` / `tool/result` resolves
-   * against. Keyed `turn:step` and by call id. `turnSteps`/`turnTools`
-   * track which step/tool anchors still belong to the open turn so
-   * `turn/end` (and a superseding `step/start`) can sweep anchors an
-   * interruption left behind; `turnFiles` keys mutated paths by turn.
-   */
-  readonly anchors: {
-    stepStart: Map<string, number>
-    toolStart: Map<string, number>
-    firstChunkAt: Map<string, number>
-    compactionTokens: Map<string, number>
-    lastPruneTokens: number
-    turnFiles: Map<number, Set<string>>
-    turnSteps: Map<number, string>
-    turnTools: Map<number, Set<string>>
-  }
-}
-
-/** Join the text blocks of a content list; non-text blocks contribute nothing. */
-function textOf(content: readonly ContentBlock[]): string {
-  return content.filter(block => block.type === 'text').map(block => block.text).join('')
-}
-
-/**
- * Snapshot-isolate one anchors block (Maps and their nested Sets): a view
- * already handed to the renderer must never observe a later fold through a
- * shared container. The collections are small and turn-bounded, so cloning
- * per event is cheap next to the entries copy the reducer already makes.
- */
-function cloneViewAnchors(anchors: TranscriptView['anchors']): TranscriptView['anchors'] {
-  return {
-    stepStart: new Map(anchors.stepStart),
-    toolStart: new Map(anchors.toolStart),
-    firstChunkAt: new Map(anchors.firstChunkAt),
-    compactionTokens: new Map(anchors.compactionTokens),
-    lastPruneTokens: anchors.lastPruneTokens,
-    turnFiles: new Map([...anchors.turnFiles].map(([turn, files]) => [turn, new Set(files)])),
-    turnSteps: new Map(anchors.turnSteps),
-    turnTools: new Map([...anchors.turnTools].map(([turn, tools]) => [turn, new Set(tools)])),
-  }
-}
-
-/** Durable image references in their model-visible order. */
-function imagesOf(content: readonly ContentBlock[]): readonly ImageBlock['attachment'][] {
-  return content.filter((block): block is ImageBlock => block.type === 'image').map(block => block.attachment)
-}
-
-/** Human-readable bounded image labels for transcript, inspector, and export surfaces. */
-export function imageLabels(images: readonly ImageBlock['attachment'][] | undefined): string {
-  if (images === undefined || images.length === 0) return ''
-  return images.map((image, index) => {
-    const rawName = image.name?.trim() || `image ${index + 1}`
-    const name = rawName.length <= 80 ? rawName : `${rawName.slice(0, 79)}…`
-    const original = image.originalDimensions
-    const dimensions = original === undefined
-      ? `${image.width}×${image.height}`
-      : `${image.width}×${image.height} · original ${original.width}×${original.height}`
-    return `[image: ${name} · ${dimensions} · ${image.bytes} B]`
-  }).join('\n')
-}
-
-/** Prompt text with its durable image labels, without exposing local paths or bytes. */
-export function promptDisplayText(entry: Pick<UserEntry | PendingEntry, 'text' | 'images'>): string {
-  const labels = imageLabels(entry.images)
-  return entry.text === '' ? labels : labels === '' ? entry.text : `${entry.text}\n${labels}`
-}
-
-/** Join the reasoning blocks of a content list; non-reasoning blocks contribute nothing. */
-function reasoningOf(content: readonly ContentBlock[]): string {
-  return content.filter(block => block.type === 'reasoning').map(block => block.text).join('')
-}
-
-/**
- * Rough token estimate for the segmented context bar (pi-nano-context's ~4
- * chars/token heuristic, CJK-aware so a Chinese prompt is not quartered):
- * CJK/wide chars cost ~1 token each, ASCII ~4 chars per token. Estimates
- * drive bar PROPORTIONS, never billing, so precision is not required.
- * @param text - the text to estimate.
- * @returns an integer token estimate, 0 for empty text.
- */
-function estimateTokens(text: string): number {
-  let wide = 0
-  let narrow = 0
-  for (const cluster of splitGraphemes(text)) {
-    if (graphemeWidth(cluster) > 1) wide += 1
-    else narrow += 1
-  }
-  return wide + Math.ceil(narrow / 4)
-}
-
-/** A fresh, empty transcript view. */
-export function createTranscriptView(): TranscriptView {
-  return {
-    entries: [],
-    streaming: '',
-    streamingReasoning: '',
-    todos: [],
-    toolCallOrdinal: 0,
-    busy: false,
-    busySince: 0,
-    model: '',
-    plan: false,
-    permission: '',
-    title: '',
-    sandbox: '',
-    goal: undefined,
-    pending: { 'next-turn': [], 'next-step': [] },
-    stats: { turns: 0, steps: 0, llmMs: 0, toolMs: 0, usage: { inputTokens: 0, outputTokens: 0, cacheReadTokens: 0 }, lastPromptTokens: 0, contextWindow: 0, contextSegments: { system: 0, prompt: 0, assistant: 0, thinking: 0, tools: 0 }, ttftMs: 0, ttftSteps: 0, decodeMs: 0, decodeTokens: 0, reasoningEffort: '' },
-    anchors: { stepStart: new Map(), toolStart: new Map(), firstChunkAt: new Map(), compactionTokens: new Map(), lastPruneTokens: 0, turnFiles: new Map(), turnSteps: new Map(), turnTools: new Map() },
-  }
-}
-
-/** Full prompt text of a queued message (identical to the durable user row it retires into). */
-function pendingText(content: readonly ContentBlock[]): string {
-  return textOf(content)
-}
-
-/**
- * Fold one session event into an updated view (copy-on-write).
- * @param view - the view before the event.
- * @param event - one durable session event from `session/event` or the log.
- * @returns the view after the event; the input view is never mutated.
- */
-export function projectEvent(view: TranscriptView, event: SessionEvent): TranscriptView {
-  // Fold against a private anchors block so the documented contract holds —
-  // "the input view is never mutated" — even for the in-place anchor sweeps
-  // below; without this, every handed-out view shared live Maps.
-  view = { ...view, anchors: cloneViewAnchors(view.anchors) }
-  switch (event.type) {
-    case 'user/message': {
-      // A queued row retires when its durable user message lands (the agent
-      // claims the inbox and logs the same message identity) — the transient
-      // steering/queued preview yields to the real transcript entry.
-      const message = event.data
-      let entries = view.entries
-      let pending = view.pending
-      for (const target of ['next-turn', 'next-step'] as const) {
-        const index = pending[target].indexOf(message.id)
-        if (index < 0) continue
-        pending = { ...pending, [target]: pending[target].filter((_, i) => i !== index) }
-        entries = entries.filter(entry => !(entry.kind === 'pending' && entry.messageId === message.id))
-      }
-      // Injected context (plugin/model-continuation sources) stays collapsed
-      // to a bounded notice row, exactly like collapsed transcript context
-      // elsewhere in the product; only direct human prompts render in full.
-      const text = textOf(message.content)
-      const images = imagesOf(message.content)
-      if (message.source.kind === 'user') {
-        return {
-          ...view,
-          pending,
-          entries: [...entries, { kind: 'user', text, notice: false, ...(images.length === 0 ? {} : { images }) }],
-          stats: {
-            ...view.stats,
-            contextSegments: {
-              ...view.stats.contextSegments,
-              prompt: view.stats.contextSegments.prompt + estimateTokens(text),
-            },
-          },
-        }
-      }
-      const notice = message.source.kind === 'plugin' && message.source.form === 'notice'
-        ? message.source.summary
-        : message.source.kind
-      const summary = boundContextSummary(notice)
-      return {
-        ...view,
-        pending,
-        entries: [...entries, { kind: 'user', text: summary, notice: true }],
-        stats: {
-          ...view.stats,
-          contextSegments: {
-            ...view.stats.contextSegments,
-            system: view.stats.contextSegments.system + estimateTokens(summary),
-          },
-        },
-      }
-    }
-    case 'agent/inbox/spliced': {
-      // The durable inbox mutation (web queue-mirror contract, event-sourced):
-      // removals drop the projected rows at their inbox coordinates, inserted
-      // messages gain a pending row at their log position.
-      const { target, start, removedCount = 0, inserted } = event.data
-      const ids = view.pending[target]
-      const removed = ids.slice(start, start + removedCount)
-      // In-place upstream semantics: the kernel's authoritative fold is
-      // `inbox.splice(start, removedCount, ...inserted)` — inserted ids land
-      // AT the splice position (prepend/replace shapes), never at the tail.
-      // A tail append diverged the id order, so later coordinate-based events
-      // (next-turn head claims, positioned remove/replace) tombstoned the
-      // wrong pending row.
-      const nextIds = [
-        ...ids.slice(0, start),
-        ...inserted.map(message => message.id),
-        ...ids.slice(start + removedCount),
-      ]
-      let entries = view.entries
-      if (removed.length > 0) {
-        const removedSet = new Set(removed)
-        entries = entries.filter(entry =>
-          !(entry.kind === 'pending' && entry.target === target && removedSet.has(entry.messageId)))
-      }
-      for (const message of inserted) {
-        entries = [...entries, {
-          kind: 'pending',
-          messageId: message.id,
-          target,
-          text: pendingText(message.content),
-          ...imagesOf(message.content).length === 0 ? {} : { images: imagesOf(message.content) },
-        }]
-      }
-      return { ...view, entries, pending: { ...view.pending, [target]: nextIds } }
-    }
-    case 'assistant/chunk': {
-      const chunk = event.data.chunk
-      // First-token latency: the first non-empty delta of a step anchors the
-      // TTFT (empty keep-alive deltas do not count as tokens).
-      const key = `${event.data.turn}:${event.data.step}`
-      const delta = chunk.type === 'text-delta' || chunk.type === 'reasoning-delta' ? chunk.text : ''
-      let stats = view.stats
-      if (delta !== '' && !view.anchors.firstChunkAt.has(key)) {
-        view.anchors.firstChunkAt.set(key, event.time)
-        const started = view.anchors.stepStart.get(key)
-        if (started !== undefined) {
-          stats = {
-            ...stats,
-            ttftMs: stats.ttftMs + Math.max(0, event.time - started),
-            ttftSteps: stats.ttftSteps + 1,
-          }
-        }
-      }
-      if (chunk.type === 'text-delta') {
-        return {
-          ...view,
-          streaming: appendStreamingTail(view.streaming, chunk.text),
-          stats,
-        }
-      }
-      if (chunk.type === 'reasoning-delta') {
-        return {
-          ...view,
-          streamingReasoning: appendStreamingTail(view.streamingReasoning, chunk.text),
-          stats,
-        }
-      }
-      return view
-    }
-    case 'assistant/message': {
-      // The assembled message is authoritative; drop the streamed buffers.
-      const key = `${event.data.turn}:${event.data.step}`
-      const started = view.anchors.stepStart.get(key)
-      view.anchors.stepStart.delete(key)
-      const firstChunk = view.anchors.firstChunkAt.get(key)
-      view.anchors.firstChunkAt.delete(key)
-      // The assembled message consumes the turn's current step anchor; a
-      // later `turn/end` sweep then has nothing left to clean for this step.
-      if (view.anchors.turnSteps.get(event.data.turn) === key) view.anchors.turnSteps.delete(event.data.turn)
-      const usage = event.data.usage
-      const totals = view.stats.usage
-      const text = textOf(event.data.message.content)
-      const reasoning = reasoningOf(event.data.message.content)
-      return {
-        ...view,
-        streaming: '',
-        streamingReasoning: '',
-        entries: [...view.entries, { kind: 'assistant', text, reasoning, interrupted: event.data.interrupted === true ? true : undefined }],
-        stats: {
-          ...view.stats,
-          llmMs: view.stats.llmMs + (started === undefined ? 0 : Math.max(0, event.time - started)),
-          usage: usage === undefined ? totals : {
-            inputTokens: totals.inputTokens + usage.inputTokens + (usage.cacheReadTokens ?? 0) + (usage.cacheWriteTokens ?? 0),
-            outputTokens: totals.outputTokens + usage.outputTokens,
-            cacheReadTokens: totals.cacheReadTokens + (usage.cacheReadTokens ?? 0),
-          },
-          lastPromptTokens: usage === undefined ? view.stats.lastPromptTokens
-            : usage.inputTokens + (usage.cacheReadTokens ?? 0) + (usage.cacheWriteTokens ?? 0),
-          // Decode span and its tokens pair up: an un-timed step (no first
-          // chunk landed) contributes neither, so the rate stays honest.
-          decodeMs: view.stats.decodeMs + (firstChunk === undefined ? 0 : Math.max(0, event.time - firstChunk)),
-          decodeTokens: view.stats.decodeTokens + (firstChunk === undefined || usage === undefined ? 0 : usage.outputTokens),
-          contextSegments: {
-            ...view.stats.contextSegments,
-            thinking: view.stats.contextSegments.thinking + estimateTokens(reasoning),
-            assistant: view.stats.contextSegments.assistant + estimateTokens(text),
-          },
-        },
-      }
-    }
-    case 'tool/call': {
-      const data = event.data
-      view.anchors.toolStart.set(data.callId, event.time)
-      // Remember the call's turn so `turn/end` can sweep a start that never
-      // pairs with a result (an interrupted tool otherwise leaks its anchor).
-      const turnTools = view.anchors.turnTools.get(data.turn) ?? new Set<string>()
-      turnTools.add(data.callId)
-      view.anchors.turnTools.set(data.turn, turnTools)
-      const ordinal = view.toolCallOrdinal + 1
-      return {
-        ...view,
-        toolCallOrdinal: ordinal,
-        entries: [
-          ...view.entries,
-          {
-            kind: 'tool',
-          callId: data.callId,
-          ordinal,
-          name: data.name,
-          arguments: data.arguments,
-          preview: toolArgumentsPreview(data.arguments, data.name),
-          prompt: toolPromptPreview(data.name, data.arguments),
-          state: 'running',
-          summary: '',
-          detail: undefined,
-        }],
-        stats: {
-          ...view.stats,
-          contextSegments: {
-            ...view.stats.contextSegments,
-            tools: view.stats.contextSegments.tools
-              + (typeof data.arguments === 'string' ? estimateTokens(data.arguments) : 0),
-          },
-        },
-      }
-    }
-    case 'tool/result': {
-      const block = event.data.message.content[0]
-      const started = view.anchors.toolStart.get(block.toolCallId)
-      view.anchors.toolStart.delete(block.toolCallId)
-      // Deregister the call from its turn's registry so `turn/end` does not
-      // sweep a start that already paired with a result.
-      const turnTools = view.anchors.turnTools.get(event.data.turn)
-      if (turnTools !== undefined) {
-        turnTools.delete(block.toolCallId)
-        if (turnTools.size === 0) view.anchors.turnTools.delete(event.data.turn)
-      }
-      const rawText = textOf(block.content)
-      const summary = boundContextSummary(rawText)
-      // The verbose expansion self-serves from the persisted presentation
-      // metadata (diffs, read windows, web sources) with the bounded raw text
-      // as the universal fallback — the capable-UI degradation ladder.
-      const detail = toolResultDetail(event.data.meta, rawText)
-      // Turn-tail deliverables: a diff-bearing mutation records its paths.
-      if (detail?.kind === 'diff') {
-        const set = view.anchors.turnFiles.get(event.data.turn) ?? new Set<string>()
-        for (const diff of detail.diffs) set.add(diff.path)
-        view.anchors.turnFiles.set(event.data.turn, set)
-      }
-      const entries = view.entries.map((entry) => {
-        if (entry.kind !== 'tool' || entry.callId !== block.toolCallId) return entry
-        return { ...entry, state: block.isError === true ? 'error' as const : 'done' as const, summary, detail }
-      })
-      return {
-        ...view,
-        entries,
-        stats: {
-          ...view.stats,
-          toolMs: view.stats.toolMs + (started === undefined ? 0 : Math.max(0, event.time - started)),
-          contextSegments: {
-            ...view.stats.contextSegments,
-            tools: view.stats.contextSegments.tools + estimateTokens(rawText),
-          },
-        },
-      }
-    }
-    case 'todo/write':
-      return { ...view, todos: event.data.todos }
-    case 'turn/start':
-      // The web todo projection clears on turn/start: a fresh turn's first
-      // write is the authoritative list, and a stale snapshot must not linger
-      // through a turn that has not written one yet.
-      return {
-        ...view,
-        busy: true,
-        busySince: view.busy ? view.busySince : event.time,
-        todos: [],
-        stats: { ...view.stats, turns: view.stats.turns + 1 },
-      }
-    case 'step/start': {
-      // A step supersedes the turn's previous step: if that step never
-      // assembled a message (interrupted), its timing anchors are stale the
-      // moment the next step opens and are swept here instead of leaking.
-      const key = `${event.data.turn}:${event.data.step}`
-      const previous = view.anchors.turnSteps.get(event.data.turn)
-      if (previous !== undefined && previous !== key) {
-        view.anchors.stepStart.delete(previous)
-        view.anchors.firstChunkAt.delete(previous)
-      }
-      view.anchors.turnSteps.set(event.data.turn, key)
-      view.anchors.stepStart.set(key, event.time)
-      return {
-        ...view,
-        streaming: '',
-        streamingReasoning: '',
-        stats: { ...view.stats, steps: view.stats.steps + 1 },
-      }
-    }
-    case 'turn/end': {
-      const reason = event.data.reason
-      const appended: TranscriptEntry[] = []
-      if (reason.kind === 'error') {
-        const recovery = reason.error.code === 'MISSING_CREDENTIAL'
-          ? ' · open /model to add an API key'
-          : ''
-        appended.push({ kind: 'error', text: `${reason.error.code}: ${reason.error.message}${recovery}` })
-      } else {
-        // Non-error outcomes deserve their own durable row (the web renders
-        // distinct max-tokens / abort / interruption nodes); `completed` stays
-        // silent so an ordinary turn never grows a marker.
-        const marker = reason.kind === 'aborted'
-          ? reason.reason.kind === 'user' ? 'turn cancelled by the user' : `turn cancelled (${reason.reason.kind})`
-          : reason.kind === 'max-tokens'
-            ? 'turn hit the output-token ceiling (max-tokens)'
-            : reason.kind === 'blocked'
-              ? 'turn ended blocked'
-              : reason.kind === 'interrupted'
-                ? 'turn was interrupted by a restart'
-                : undefined
-        if (marker !== undefined) appended.push({ kind: 'turn-marker', text: marker })
-      }
-      // Deliverables ride the turn tail (the web's turnTail chips): the
-      // turn's mutated files flush as one bounded row, then the set resets.
-      const files = view.anchors.turnFiles.get(event.data.turn)
-      view.anchors.turnFiles.delete(event.data.turn)
-      if (files !== undefined && files.size > 0) appended.push({ kind: 'files', paths: [...files].slice(0, 12) })
-      // Derivable boundary sweep: the turn is over, so any step/tool anchors
-      // it left behind (interruptions that never produced their message or
-      // result) can never be resolved and are reclaimed now.
-      const stepKey = view.anchors.turnSteps.get(event.data.turn)
-      if (stepKey !== undefined) {
-        view.anchors.stepStart.delete(stepKey)
-        view.anchors.firstChunkAt.delete(stepKey)
-        view.anchors.turnSteps.delete(event.data.turn)
-      }
-      const turnToolSet = view.anchors.turnTools.get(event.data.turn)
-      if (turnToolSet !== undefined) {
-        for (const callId of turnToolSet) view.anchors.toolStart.delete(callId)
-        view.anchors.turnTools.delete(event.data.turn)
-      }
-      // Orphaned retry/command rows can never be resolved after the turn
-      // ends: an aborted retry backoff returns upstream without its
-      // `llm/retry-started`, and crash repair synthesizes only tool/step/
-      // turn closers. Left `running` they pin the settled boundary forever,
-      // so the turn end finalizes them exactly like the anchor sweep above.
-      let orphans = false
-      const swept = view.entries.map((entry) => {
-        if (entry.kind === 'retry' && entry.state === 'running') {
-          orphans = true
-          return { ...entry, state: 'done' as const }
-        }
-        if (entry.kind === 'command' && entry.state === 'running') {
-          orphans = true
-          return { ...entry, state: 'error' as const, summary: 'interrupted before the turn ended' }
-        }
-        return entry
-      })
-      const entries = orphans ? swept : view.entries
-      if (appended.length === 0) {
-        return { ...view, busy: false, busySince: 0, streaming: '', streamingReasoning: '', entries }
-      }
-      return {
-        ...view,
-        busy: false,
-        busySince: 0,
-        streaming: '',
-        streamingReasoning: '',
-        entries: [...entries, ...appended],
-      }
-    }
-    case 'llm/retry': {
-      const data = event.data
-      return {
-        ...view,
-        streaming: '',
-        streamingReasoning: '',
-        entries: [...view.entries, {
-          kind: 'retry',
-          retryId: data.retryId,
-          mode: data.mode,
-          attempt: data.retry,
-          max: 'maxRetries' in data ? data.maxRetries : data.retry,
-          code: data.failure.code,
-          delayMs: data.delayMs,
-          state: 'running',
-        }],
-      }
-    }
-    case 'llm/retry-started': {
-      const data = event.data
-      const entries = view.entries.map((entry) => {
-        if (entry.kind !== 'retry' || entry.retryId !== data.retryId) return entry
-        return { ...entry, state: 'done' as const }
-      })
-      return { ...view, entries }
-    }
-    case 'sandbox/mode':
-      // Log-only override switch; last write wins for the status badge.
-      return { ...view, sandbox: event.data.mode }
-    case 'goal/change': {
-      const data = event.data
-      const clip = (text: string): string => (text.length > 60 ? `${text.slice(0, 59)}…` : text)
-      if (data.operation === 'clear') {
-        return {
-          ...view,
-          goal: undefined,
-          entries: [...view.entries, { kind: 'turn-marker', text: '◎ goal cleared' }],
-        }
-      }
-      const goal: GoalFold = {
-        objective: data.goal.objective,
-        phase: data.goal.phase,
-        rounds: data.roundsStarted,
-        max: data.goal.maxGoalRounds,
-        blocked: data.goal.blockedReason?.message ?? '',
-      }
-      const line = data.operation === 'create'
-        ? `◎ goal: ${clip(data.goal.objective)}`
-        : data.operation === 'complete'
-          ? '◎ goal complete'
-          : data.operation === 'pause'
-            ? '◎ goal paused'
-            : data.operation === 'resume'
-              ? '◎ goal resumed'
-              : data.operation === 'block'
-                ? `◎ goal blocked: ${clip(goal.blocked)}`
-                : undefined
-      return {
-        ...view,
-        goal,
-        entries: line === undefined ? view.entries : [...view.entries, { kind: 'turn-marker', text: line }],
-      }
-    }
-    case 'session/title':
-      // Latest-wins title snapshot, log-only; the status line prefers it.
-      return { ...view, title: event.data.title }
-    case 'compaction/summary':
-      // Remember the shadow price so the matching `compaction/end` row can
-      // state what the compaction reclaimed. The map is capped so an aborted
-      // compaction (summary without end) cannot leave an unbounded residue.
-      if (view.anchors.compactionTokens.size >= MAX_COMPACTION_SUMMARY_RESIDUE) {
-        const oldest = view.anchors.compactionTokens.keys().next().value
-        if (oldest !== undefined) view.anchors.compactionTokens.delete(oldest)
-      }
-      view.anchors.compactionTokens.set(event.data.compactionId, event.data.shadowedTokenCount)
-      return view
-    case 'compaction/prune':
-      // A model-free prune carries no compaction id; its price serves the next
-      // `compaction/end` that cannot find a summary price.
-      return { ...view, anchors: { ...view.anchors, lastPruneTokens: event.data.shadowedTokenCount } }
-    case 'compaction/end': {
-      const ok = event.data.error === undefined
-      const tokens = view.anchors.compactionTokens.get(event.data.compactionId) ?? view.anchors.lastPruneTokens
-      view.anchors.compactionTokens.delete(event.data.compactionId)
-      return {
-        ...view,
-        entries: [...view.entries, { kind: 'compaction', ok, tokens, error: event.data.error ?? '' }],
-      }
-    }
-    case 'request/context':
-      // Route capacity, logged only when it changes; last one wins.
-      return {
-        ...view,
-        stats: { ...view.stats, contextWindow: event.data.contextWindow ?? view.stats.contextWindow },
-      }
-    case 'request/header': {
-      // The session's own model record: the latest snapshot's provider/model
-      // pair, exactly what a resumed TUI restores as the selection, plus the
-      // effective reasoning effort that snapshot carried (the adapter may
-      // materialize the model default, which is what the status line shows).
-      // The snapshot's rendered system prompt is the current system slot, so
-      // it REPLACES the estimate (an older system prompt is not re-sent).
-      const config = event.data.header.config
-      return {
-        ...view,
-        model: `${config.provider}/${config.model}`,
-        stats: {
-          ...view.stats,
-          reasoningEffort: config.reasoningEffort === undefined ? '' : String(config.reasoningEffort),
-          contextSegments: {
-            ...view.stats.contextSegments,
-            system: estimateTokens(event.data.header.system ?? ''),
-          },
-        },
-      }
-    }
-    case 'plan/mode':
-      // Whole-value replace; the last one wins (upstream fold semantics).
-      return { ...view, plan: event.data.active }
-    case 'permission/preset':
-      return { ...view, permission: event.data.preset }
-    case 'command/run': {
-      const data = event.data
-      return {
-        ...view,
-        entries: [...view.entries, {
-          kind: 'command',
-          commandId: data.commandId,
-          name: data.name,
-          args: data.args ?? '',
-          state: 'running',
-          summary: '',
-        }],
-      }
-    }
-    case 'command/done': {
-      const data = event.data
-      const entries = view.entries.map((entry) => {
-        if (entry.kind !== 'command' || entry.commandId !== data.commandId) return entry
-        return {
-          ...entry,
-          state: data.kind === 'success' ? 'done' as const : 'error' as const,
-          summary: boundContextSummary(data.text ?? ''),
-        }
-      })
-      return { ...view, entries }
-    }
-    default:
-      return view
-  }
-}
-
-/**
- * Mutable replay accumulator: folds a persisted log into the identical view
- * `projectEvent` would produce, but in near-linear time. Where `projectEvent`
- * is copy-on-write — every append/scan rebuilds the whole `entries` array, so
- * folding a full log costs O(N²) — the accumulator appends by push, resolves
- * id-keyed updates (tool/result, command/done, retry-started) through index
- * maps, and tombstones retired pending rows, so the whole log folds in O(N)
- * plus one compaction pass when tombstones exist.
- *
- * Index maps never delete: every appended row registers its index, so an id
- * lookup miss provably means no matching row exists and the update is an O(1)
- * no-op (a malicious/orphan-heavy log cannot force per-orphan full-array
- * scans). Each id maps to ALL of its indices, so a duplicate id updates every
- * matching row exactly like the copy-on-write reducer.
- *
- * @internal Exported only so tests can (a) prove replay ≡ sequential
- * `projectEvent` folds and (b) assert the linear complexity deterministically
- * via {@link ReplayAccumulator.ops}, which counts entry-level container work
- * instead of relying on wall-clock thresholds. No public consumer.
- */
-export interface ReplayAccumulator {
-  /** Working entry list; `undefined` marks a retired pending row (tombstone). */
-  entries: (TranscriptEntry | undefined)[]
-  /** callId → every index into `entries` holding a `tool` row with that id. */
-  toolIndex: Map<string, number[]>
-  /** commandId → every index into `entries` holding a `command` row with that id. */
-  commandIndex: Map<string, number[]>
-  /** retryId → every index into `entries` holding a `retry` row with that id. */
-  retryIndex: Map<string, number[]>
-  /** messageId → every index into `entries` holding a `pending` row with that id. */
-  pendingIndex: Map<string, number[]>
-  /** Tombstone count; zero means `entries` is already the final array. */
-  removedCount: number
-  /** Mutable inbox id lists, mirroring `view.pending` order per target. */
-  pendingTurn: string[]
-  pendingStep: string[]
-  streaming: string
-  streamingReasoning: string
-  todos: readonly TodoItem[]
-  /** Global tool-call ordinal counter (see `TranscriptView.toolCallOrdinal`). */
-  toolCallOrdinal: number
-  busy: boolean
-  busySince: number
-  model: string
-  plan: boolean
-  permission: string
-  title: string
-  sandbox: string
-  goal: GoalFold | undefined
-  stats: TranscriptStats
-  stepStart: Map<string, number>
-  toolStart: Map<string, number>
-  firstChunkAt: Map<string, number>
-  compactionTokens: Map<string, number>
-  lastPruneTokens: number
-  turnFiles: Map<number, Set<string>>
-  turnSteps: Map<number, string>
-  turnTools: Map<number, Set<string>>
-  /** Entry-level container operations performed so far (test instrumentation). */
-  ops: number
-}
-
-/** @internal A fresh replay accumulator whose state mirrors `createTranscriptView()`. */
-export function createReplayAccumulator(): ReplayAccumulator {
-  return {
-    entries: [],
-    toolIndex: new Map(),
-    commandIndex: new Map(),
-    retryIndex: new Map(),
-    pendingIndex: new Map(),
-    removedCount: 0,
-    pendingTurn: [],
-    pendingStep: [],
-    streaming: '',
-    streamingReasoning: '',
-    todos: [],
-    toolCallOrdinal: 0,
-    busy: false,
-    busySince: 0,
-    model: '',
-    plan: false,
-    permission: '',
-    title: '',
-    sandbox: '',
-    goal: undefined,
-    stats: { turns: 0, steps: 0, llmMs: 0, toolMs: 0, usage: { inputTokens: 0, outputTokens: 0, cacheReadTokens: 0 }, lastPromptTokens: 0, contextWindow: 0, contextSegments: { system: 0, prompt: 0, assistant: 0, thinking: 0, tools: 0 }, ttftMs: 0, ttftSteps: 0, decodeMs: 0, decodeTokens: 0, reasoningEffort: '' },
-    stepStart: new Map(),
-    toolStart: new Map(),
-    firstChunkAt: new Map(),
-    compactionTokens: new Map(),
-    lastPruneTokens: 0,
-    turnFiles: new Map(),
-    turnSteps: new Map(),
-    turnTools: new Map(),
-    ops: 0,
-  }
-}
-
-/** Append one entry (O(1)) and account the push. */
-function appendReplayEntry(acc: ReplayAccumulator, entry: TranscriptEntry): void {
-  acc.entries.push(entry)
-  acc.ops += 1
-}
-
-/**
- * Get (or create) the index list an id owns. Lists are never removed: every
- * appended row registers its index, so a lookup miss later proves no matching
- * row exists and the caller can no-op in O(1).
- */
-function indexList(map: Map<string, number[]>, id: string): number[] {
-  let list = map.get(id)
-  if (list === undefined) {
-    list = []
-    map.set(id, list)
-  }
-  return list
-}
-
-/**
- * Finalize replay rows the ended turn left `running`, mirroring the reducer's
- * turn-end orphan sweep: an orphaned retry settles `done`, an orphaned command
- * settles `error` with an interruption note. Only the id-indexed rows are
- * visited, so the sweep stays O(retries+commands of the log), never a scan.
- */
-function finalizeReplayOrphans(acc: ReplayAccumulator): void {
-  for (const list of acc.retryIndex.values()) {
-    for (const index of list) {
-      const entry = acc.entries[index]
-      if (entry !== undefined && entry.kind === 'retry' && entry.state === 'running') {
-        acc.entries[index] = { ...entry, state: 'done' }
-      }
-    }
-  }
-  for (const list of acc.commandIndex.values()) {
-    for (const index of list) {
-      const entry = acc.entries[index]
-      if (entry !== undefined && entry.kind === 'command' && entry.state === 'running') {
-        acc.entries[index] = { ...entry, state: 'error', summary: 'interrupted before the turn ended' }
-      }
-    }
-  }
-}
-
-/**
- * Apply an id-keyed update to every row that registered the id, mirroring the
- * copy-on-write reducer's full-array map semantics (all matching rows update,
- * in order). Each registered index is O(1), so a duplicate id costs
- * O(#duplicates) — never a full-array scan. The kind+id re-check is defensive:
- * registered indices are valid by construction, because tool/command/retry
- * rows are never removed and tombstones never shift indices.
- */
-function updateReplayById<T extends TranscriptEntry>(
-  acc: ReplayAccumulator,
-  map: Map<string, number[]>,
-  id: string,
-  isMatch: (entry: T) => boolean,
-  update: (entry: T) => T,
-): void {
-  const list = map.get(id)
-  if (list === undefined) return // miss provably means no matching row
-  for (const index of list) {
-    const entry = acc.entries[index]
-    if (entry === undefined || !isMatch(entry as T)) continue
-    acc.entries[index] = update(entry as T)
-    acc.ops += 1
-  }
-}
-
-/** Tombstone a retired pending row, keeping every other index stable. */
-function retireReplayEntry(acc: ReplayAccumulator, index: number): void {
-  if (acc.entries[index] !== undefined) {
-    acc.entries[index] = undefined
-    acc.removedCount += 1
-    acc.ops += 1
-  }
-}
-
-/**
- * Fold one session event into a replay accumulator. This mirrors
- * {@link projectEvent} case for case — same stats arithmetic, same anchor
- * set/delete behavior, same entry shapes — so the finished view is identical
- * to a sequential fold; only the `entries` container operations are mutable.
- *
- * @internal Test-instrumentation path; `projectEvents` is the public entry.
- * @returns whether the event changed the accumulated state — the live store
- * stays silent and keeps its snapshot identity for ignored events, exactly
- * like the copy-on-write reducer returning its input view unchanged.
- */
-export function replayProjectEvent(acc: ReplayAccumulator, event: SessionEvent): boolean {
-  switch (event.type) {
-    case 'user/message': {
-      const message = event.data
-      for (const target of ['next-turn', 'next-step'] as const) {
-        const ids = target === 'next-turn' ? acc.pendingTurn : acc.pendingStep
-        const index = ids.indexOf(message.id)
-        acc.ops += index < 0 ? ids.length : index + 1
-        if (index < 0) continue
-        ids.splice(index, 1)
-        acc.ops += 1
-        // Retire every pending row carrying this message id (duplicate ids
-        // included), exactly like the reducer's full-array filter.
-        const list = acc.pendingIndex.get(message.id)
-        if (list !== undefined) {
-          for (const entryIndex of list) retireReplayEntry(acc, entryIndex)
-          acc.ops += 1
-        }
-      }
-      const text = textOf(message.content)
-      const images = imagesOf(message.content)
-      if (message.source.kind === 'user') {
-        appendReplayEntry(acc, { kind: 'user', text, notice: false, ...(images.length === 0 ? {} : { images }) })
-        acc.stats = {
-          ...acc.stats,
-          contextSegments: {
-            ...acc.stats.contextSegments,
-            prompt: acc.stats.contextSegments.prompt + estimateTokens(text),
-          },
-        }
-        return true
-      }
-      const notice = message.source.kind === 'plugin' && message.source.form === 'notice'
-        ? message.source.summary
-        : message.source.kind
-      const summary = boundContextSummary(notice)
-      appendReplayEntry(acc, { kind: 'user', text: summary, notice: true })
-      acc.stats = {
-        ...acc.stats,
-        contextSegments: {
-          ...acc.stats.contextSegments,
-          system: acc.stats.contextSegments.system + estimateTokens(summary),
-        },
-      }
-      return true
-    }
-    case 'agent/inbox/spliced': {
-      const { target, start, removedCount = 0, inserted } = event.data
-      const ids = target === 'next-turn' ? acc.pendingTurn : acc.pendingStep
-      const removed = ids.slice(start, start + removedCount)
-      acc.ops += removed.length
-      ids.splice(start, removedCount)
-      acc.ops += removed.length
-      for (const id of removed) {
-        const list = acc.pendingIndex.get(id)
-        if (list === undefined) continue
-        for (const entryIndex of list) {
-          const entry = acc.entries[entryIndex]
-          if (entry !== undefined && entry.kind === 'pending' && entry.target === target) {
-            retireReplayEntry(acc, entryIndex)
-          }
-        }
-      }
-      // Mirror the reducer's in-place order: inserted ids join at the splice
-      // position (upstream `splice(start, removedCount, ...inserted)`), never
-      // at the tail — the id list must stay coordinate-compatible with every
-      // later inbox event.
-      ids.splice(start, 0, ...inserted.map(message => message.id))
-      for (const message of inserted) {
-        const images = imagesOf(message.content)
-        appendReplayEntry(acc, { kind: 'pending', messageId: message.id, target, text: pendingText(message.content), ...(images.length === 0 ? {} : { images }) })
-        indexList(acc.pendingIndex, message.id).push(acc.entries.length - 1)
-        acc.ops += 1
-      }
-      return true
-    }
-    case 'assistant/chunk': {
-      const chunk = event.data.chunk
-      const key = `${event.data.turn}:${event.data.step}`
-      const delta = chunk.type === 'text-delta' || chunk.type === 'reasoning-delta' ? chunk.text : ''
-      if (delta !== '' && !acc.firstChunkAt.has(key)) {
-        acc.firstChunkAt.set(key, event.time)
-        const started = acc.stepStart.get(key)
-        if (started !== undefined) {
-          acc.stats = {
-            ...acc.stats,
-            ttftMs: acc.stats.ttftMs + Math.max(0, event.time - started),
-            ttftSteps: acc.stats.ttftSteps + 1,
-          }
-        }
-      }
-      if (chunk.type === 'text-delta') {
-        acc.streaming = appendStreamingTail(acc.streaming, chunk.text)
-        return true
-      }
-      if (chunk.type === 'reasoning-delta') {
-        acc.streamingReasoning = appendStreamingTail(acc.streamingReasoning, chunk.text)
-        return true
-      }
-      return false
-    }
-    case 'assistant/message': {
-      const key = `${event.data.turn}:${event.data.step}`
-      const started = acc.stepStart.get(key)
-      acc.stepStart.delete(key)
-      const firstChunk = acc.firstChunkAt.get(key)
-      acc.firstChunkAt.delete(key)
-      if (acc.turnSteps.get(event.data.turn) === key) acc.turnSteps.delete(event.data.turn)
-      const usage = event.data.usage
-      const totals = acc.stats.usage
-      const text = textOf(event.data.message.content)
-      const reasoning = reasoningOf(event.data.message.content)
-      acc.streaming = ''
-      acc.streamingReasoning = ''
-      appendReplayEntry(acc, { kind: 'assistant', text, reasoning, interrupted: event.data.interrupted === true ? true : undefined })
-      acc.stats = {
-        ...acc.stats,
-        llmMs: acc.stats.llmMs + (started === undefined ? 0 : Math.max(0, event.time - started)),
-        usage: usage === undefined ? totals : {
-          inputTokens: totals.inputTokens + usage.inputTokens + (usage.cacheReadTokens ?? 0) + (usage.cacheWriteTokens ?? 0),
-          outputTokens: totals.outputTokens + usage.outputTokens,
-          cacheReadTokens: totals.cacheReadTokens + (usage.cacheReadTokens ?? 0),
-        },
-        lastPromptTokens: usage === undefined ? acc.stats.lastPromptTokens
-          : usage.inputTokens + (usage.cacheReadTokens ?? 0) + (usage.cacheWriteTokens ?? 0),
-        decodeMs: acc.stats.decodeMs + (firstChunk === undefined ? 0 : Math.max(0, event.time - firstChunk)),
-        decodeTokens: acc.stats.decodeTokens + (firstChunk === undefined || usage === undefined ? 0 : usage.outputTokens),
-        contextSegments: {
-          ...acc.stats.contextSegments,
-          thinking: acc.stats.contextSegments.thinking + estimateTokens(reasoning),
-          assistant: acc.stats.contextSegments.assistant + estimateTokens(text),
-        },
-      }
-      return true
-    }
-    case 'tool/call': {
-      const data = event.data
-      acc.toolStart.set(data.callId, event.time)
-      const turnTools = acc.turnTools.get(data.turn) ?? new Set<string>()
-      turnTools.add(data.callId)
-      acc.turnTools.set(data.turn, turnTools)
-      acc.toolCallOrdinal += 1
-      appendReplayEntry(acc, {
-        kind: 'tool',
-        callId: data.callId,
-        ordinal: acc.toolCallOrdinal,
-        name: data.name,
-        arguments: data.arguments,
-        preview: toolArgumentsPreview(data.arguments, data.name),
-        prompt: toolPromptPreview(data.name, data.arguments),
-        state: 'running',
-        summary: '',
-        detail: undefined,
-      })
-      indexList(acc.toolIndex, data.callId).push(acc.entries.length - 1)
-      acc.stats = {
-        ...acc.stats,
-        contextSegments: {
-          ...acc.stats.contextSegments,
-          tools: acc.stats.contextSegments.tools
-            + (typeof data.arguments === 'string' ? estimateTokens(data.arguments) : 0),
-        },
-      }
-      return true
-    }
-    case 'tool/result': {
-      const block = event.data.message.content[0]
-      const started = acc.toolStart.get(block.toolCallId)
-      acc.toolStart.delete(block.toolCallId)
-      const turnTools = acc.turnTools.get(event.data.turn)
-      if (turnTools !== undefined) {
-        turnTools.delete(block.toolCallId)
-        if (turnTools.size === 0) acc.turnTools.delete(event.data.turn)
-      }
-      const rawText = textOf(block.content)
-      const summary = boundContextSummary(rawText)
-      const detail = toolResultDetail(event.data.meta, rawText)
-      if (detail?.kind === 'diff') {
-        const set = acc.turnFiles.get(event.data.turn) ?? new Set<string>()
-        for (const diff of detail.diffs) set.add(diff.path)
-        acc.turnFiles.set(event.data.turn, set)
-      }
-      const update = (entry: ToolEntry): ToolEntry => ({
-        ...entry,
-        state: block.isError === true ? 'error' as const : 'done' as const,
-        summary,
-        detail,
-      })
-      // Every matching row updates (duplicate callIds included); an id with no
-      // registered index is a provable no-op — no full-array fallback scan.
-      updateReplayById<ToolEntry>(acc, acc.toolIndex, block.toolCallId, entry => entry.callId === block.toolCallId, update)
-      acc.stats = {
-        ...acc.stats,
-        toolMs: acc.stats.toolMs + (started === undefined ? 0 : Math.max(0, event.time - started)),
-        contextSegments: {
-          ...acc.stats.contextSegments,
-          tools: acc.stats.contextSegments.tools + estimateTokens(rawText),
-        },
-      }
-      return true
-    }
-    case 'todo/write':
-      acc.todos = event.data.todos
-      return true
-    case 'turn/start': {
-      const wasBusy = acc.busy
-      acc.busy = true
-      acc.busySince = wasBusy ? acc.busySince : event.time
-      acc.todos = []
-      acc.stats = { ...acc.stats, turns: acc.stats.turns + 1 }
-      return true
-    }
-    case 'step/start': {
-      const key = `${event.data.turn}:${event.data.step}`
-      const previous = acc.turnSteps.get(event.data.turn)
-      if (previous !== undefined && previous !== key) {
-        acc.stepStart.delete(previous)
-        acc.firstChunkAt.delete(previous)
-      }
-      acc.turnSteps.set(event.data.turn, key)
-      acc.stepStart.set(key, event.time)
-      acc.streaming = ''
-      acc.streamingReasoning = ''
-      acc.stats = { ...acc.stats, steps: acc.stats.steps + 1 }
-      return true
-    }
-    case 'turn/end': {
-      const reason = event.data.reason
-      const appended: TranscriptEntry[] = []
-      acc.streamingReasoning = ''
-      acc.streaming = ''
-      if (reason.kind === 'error') {
-        const recovery = reason.error.code === 'MISSING_CREDENTIAL'
-          ? ' · open /model to add an API key'
-          : ''
-        appended.push({ kind: 'error', text: `${reason.error.code}: ${reason.error.message}${recovery}` })
-      } else {
-        const marker = reason.kind === 'aborted'
-          ? reason.reason.kind === 'user' ? 'turn cancelled by the user' : `turn cancelled (${reason.reason.kind})`
-          : reason.kind === 'max-tokens'
-            ? 'turn hit the output-token ceiling (max-tokens)'
-            : reason.kind === 'blocked'
-              ? 'turn ended blocked'
-              : reason.kind === 'interrupted'
-                ? 'turn was interrupted by a restart'
-                : undefined
-        if (marker !== undefined) appended.push({ kind: 'turn-marker', text: marker })
-      }
-      const files = acc.turnFiles.get(event.data.turn)
-      acc.turnFiles.delete(event.data.turn)
-      if (files !== undefined && files.size > 0) appended.push({ kind: 'files', paths: [...files].slice(0, 12) })
-      const stepKey = acc.turnSteps.get(event.data.turn)
-      if (stepKey !== undefined) {
-        acc.stepStart.delete(stepKey)
-        acc.firstChunkAt.delete(stepKey)
-        acc.turnSteps.delete(event.data.turn)
-      }
-      const turnToolSet = acc.turnTools.get(event.data.turn)
-      if (turnToolSet !== undefined) {
-        for (const callId of turnToolSet) acc.toolStart.delete(callId)
-        acc.turnTools.delete(event.data.turn)
-      }
-      finalizeReplayOrphans(acc)
-      acc.busy = false
-      acc.busySince = 0
-      for (const entry of appended) appendReplayEntry(acc, entry)
-      return true
-    }
-    case 'llm/retry': {
-      const data = event.data
-      acc.streaming = ''
-      acc.streamingReasoning = ''
-      appendReplayEntry(acc, {
-        kind: 'retry',
-        retryId: data.retryId,
-        mode: data.mode,
-        attempt: data.retry,
-        max: 'maxRetries' in data ? data.maxRetries : data.retry,
-        code: data.failure.code,
-        delayMs: data.delayMs,
-        state: 'running',
-      })
-      indexList(acc.retryIndex, data.retryId).push(acc.entries.length - 1)
-      return true
-    }
-    case 'llm/retry-started': {
-      const data = event.data
-      updateReplayById<RetryEntry>(acc, acc.retryIndex, data.retryId, entry => entry.retryId === data.retryId, entry => ({ ...entry, state: 'done' as const }))
-      return true
-    }
-    case 'sandbox/mode':
-      acc.sandbox = event.data.mode
-      return true
-    case 'goal/change': {
-      const data = event.data
-      const clip = (text: string): string => (text.length > 60 ? `${text.slice(0, 59)}…` : text)
-      if (data.operation === 'clear') {
-        acc.goal = undefined
-        appendReplayEntry(acc, { kind: 'turn-marker', text: '◎ goal cleared' })
-        return true
-      }
-      const goal: GoalFold = {
-        objective: data.goal.objective,
-        phase: data.goal.phase,
-        rounds: data.roundsStarted,
-        max: data.goal.maxGoalRounds,
-        blocked: data.goal.blockedReason?.message ?? '',
-      }
-      const line = data.operation === 'create'
-        ? `◎ goal: ${clip(data.goal.objective)}`
-        : data.operation === 'complete'
-          ? '◎ goal complete'
-          : data.operation === 'pause'
-            ? '◎ goal paused'
-            : data.operation === 'resume'
-              ? '◎ goal resumed'
-              : data.operation === 'block'
-                ? `◎ goal blocked: ${clip(goal.blocked)}`
-                : undefined
-      acc.goal = goal
-      if (line !== undefined) appendReplayEntry(acc, { kind: 'turn-marker', text: line })
-      return true
-    }
-    case 'session/title':
-      acc.title = event.data.title
-      return true
-    case 'compaction/summary':
-      if (acc.compactionTokens.size >= MAX_COMPACTION_SUMMARY_RESIDUE) {
-        const oldest = acc.compactionTokens.keys().next().value
-        if (oldest !== undefined) acc.compactionTokens.delete(oldest)
-      }
-      acc.compactionTokens.set(event.data.compactionId, event.data.shadowedTokenCount)
-      return true
-    case 'compaction/prune':
-      acc.lastPruneTokens = event.data.shadowedTokenCount
-      return true
-    case 'compaction/end': {
-      const ok = event.data.error === undefined
-      const tokens = acc.compactionTokens.get(event.data.compactionId) ?? acc.lastPruneTokens
-      acc.compactionTokens.delete(event.data.compactionId)
-      appendReplayEntry(acc, { kind: 'compaction', ok, tokens, error: event.data.error ?? '' })
-      return true
-    }
-    case 'request/context':
-      acc.stats = { ...acc.stats, contextWindow: event.data.contextWindow ?? acc.stats.contextWindow }
-      return true
-    case 'request/header': {
-      const config = event.data.header.config
-      acc.model = `${config.provider}/${config.model}`
-      acc.stats = {
-        ...acc.stats,
-        reasoningEffort: config.reasoningEffort === undefined ? '' : String(config.reasoningEffort),
-        contextSegments: {
-          ...acc.stats.contextSegments,
-          system: estimateTokens(event.data.header.system ?? ''),
-        },
-      }
-      return true
-    }
-    case 'plan/mode':
-      acc.plan = event.data.active
-      return true
-    case 'permission/preset':
-      acc.permission = event.data.preset
-      return true
-    case 'command/run': {
-      const data = event.data
-      appendReplayEntry(acc, {
-        kind: 'command',
-        commandId: data.commandId,
-        name: data.name,
-        args: data.args ?? '',
-        state: 'running',
-        summary: '',
-      })
-      indexList(acc.commandIndex, data.commandId).push(acc.entries.length - 1)
-      return true
-    }
-    case 'command/done': {
-      const data = event.data
-      const update = (candidate: CommandEntry): CommandEntry => ({
-        ...candidate,
-        state: data.kind === 'success' ? 'done' as const : 'error' as const,
-        summary: boundContextSummary(data.text ?? ''),
-      })
-      updateReplayById<CommandEntry>(acc, acc.commandIndex, data.commandId, entry => entry.commandId === data.commandId, update)
-      return true
-    }
-    default:
-      return false
-  }
-}
-
-/**
- * Materialize the accumulated fold as a `TranscriptView`, compacting any
- * retired tombstones. The anchors maps are handed through as-is (their
- * content is identical to a sequential fold's).
- *
- * @internal Test-instrumentation path; `projectEvents` is the public entry.
- */
-export function finishReplay(acc: ReplayAccumulator): TranscriptView {
-  return materializeReplayView(acc, false)
-}
-
-/**
- * Materialize the accumulated fold as a fresh immutable snapshot for the
- * live store. Unlike {@link finishReplay} — the one-shot replay entry, which
- * hands the accumulator's own arrays through because the accumulator is
- * discarded — every array a renderer can hold is copied here, so later
- * folds never mutate a snapshot already handed out. Same fields, same
- * tombstone compaction.
- *
- * @internal Live-store path; `projectEvents` is the public entry.
- */
-export function snapshotReplayView(acc: ReplayAccumulator): TranscriptView {
-  return materializeReplayView(acc, true)
-}
-
-/** Field-for-field materialization; `copy` selects snapshot array isolation. */
-function materializeReplayView(acc: ReplayAccumulator, copy: boolean): TranscriptView {
-  const entries: readonly TranscriptEntry[] = acc.removedCount === 0
-    ? (copy ? [...acc.entries] : acc.entries) as TranscriptEntry[]
-    : acc.entries.filter((entry): entry is TranscriptEntry => entry !== undefined)
-  if (acc.removedCount > 0) acc.ops += acc.entries.length
-  return {
-    entries,
-    streaming: acc.streaming,
-    streamingReasoning: acc.streamingReasoning,
-    todos: acc.todos,
-    toolCallOrdinal: acc.toolCallOrdinal,
-    busy: acc.busy,
-    busySince: acc.busySince,
-    model: acc.model,
-    plan: acc.plan,
-    permission: acc.permission,
-    title: acc.title,
-    sandbox: acc.sandbox,
-    goal: acc.goal,
-    pending: { 'next-turn': [...acc.pendingTurn], 'next-step': [...acc.pendingStep] },
-    stats: acc.stats,
-    // Handed-out views get their own anchors snapshot: the accumulator keeps
-    // folding its live containers, and no consumer may observe that.
-    anchors: {
-      stepStart: new Map(acc.stepStart),
-      toolStart: new Map(acc.toolStart),
-      firstChunkAt: new Map(acc.firstChunkAt),
-      compactionTokens: new Map(acc.compactionTokens),
-      lastPruneTokens: acc.lastPruneTokens,
-      turnFiles: new Map([...acc.turnFiles].map(([turn, files]) => [turn, new Set(files)])),
-      turnSteps: new Map(acc.turnSteps),
-      turnTools: new Map([...acc.turnTools].map(([turn, tools]) => [turn, new Set(tools)])),
-    },
-  }
-}
-
-/**
- * Fold a replayed event history into one view.
- *
- * Folding is near-linear in the log size: the mutable replay accumulator
- * appends in place and resolves id-keyed updates through index maps, so a
- * long persisted session replays without the O(N²) copy-on-write rebuilds a
- * naive sequential fold would incur. The result is identical to folding
- * {@link projectEvent} per event in order.
- * @param events - events in `seq` order.
- * @returns the folded view.
- */
-export function projectEvents(events: readonly SessionEvent[]): TranscriptView {
-  const acc = createReplayAccumulator()
-  for (const event of events) replayProjectEvent(acc, event)
-  return finishReplay(acc)
-}
-
-/**
- * The append-only flush boundary for a transcript view: the count of entries
- * no later event can remove. Entries at or beyond this index are mutable and
- * must stay in the live tree.
- *
- * `pending` rows are excluded even though they are not a running tool/retry:
- * the inbox claims or cancels them durably (`agent/inbox/spliced` removals,
- * `user/message` retirement), and an append-only `<Static>` flush cannot
- * erase a row that vanishes from the view — the retired row would ghost on
- * screen until the next source-backed replay. Running commands join the
- * mutable boundary for the same reason in reverse: `command/done` mutates the
- * row's state/summary, so a flushed row would keep its stale running mark
- * until a resize-triggered replay. Everything else (including a completed
- * tail) is final: later events only APPEND new rows.
- * @param entries - the view's transcript entries in order.
- * @returns the count of entries safe to flush (0 for an empty transcript).
- */
-export function settledEntryCount(entries: readonly TranscriptEntry[]): number {
-  for (let index = 0; index < entries.length; index++) {
-    const entry = entries[index]
-    if (entry.kind === 'pending') return index
-    if (entry.kind === 'tool' && entry.state === 'running') return index
-    if (entry.kind === 'retry' && entry.state === 'running') return index
-    if (entry.kind === 'command' && entry.state === 'running') return index
-  }
-  return entries.length
-}
+/**
+ * Pure session-event-to-view projection for the TUI transcript: one reducer
+ * over {@link SessionEvent}s producing the ordered entries the renderer draws.
+ * Rendering never reads the session directly — this module owns the view
+ * model, so tests drive it with plain event arrays.
+ *
+ * @module @deepseek-ai/dsh-tui/render/projection
+ */
+
+import { assistantStreamFirstTokenTime, boundContextSummary, isTokenDelta, type ContentBlock, type FileBlock, type ImageBlock, type MessageId, type StreamChunk } from '@deepseek-ai/dsh-llm'
+import type { FileAttachmentRef } from '@deepseek-ai/dsh-attachment'
+import type { SessionEvent } from '@deepseek-ai/dsh-session'
+import type { TodoItem } from '@deepseek-ai/dsh-tool-todo'
+import { graphemeWidth, splitGraphemes } from './width.ts'
+// Type-only imports merge the plugin-owned SessionEventMap variants
+// (agent/inbox/spliced, command/*, compaction/*, goal/change, llm/retry*,
+// plan/mode, permission/preset, sandbox/mode, session/title) into the union
+// this reducer switches on.
+import type {} from '@deepseek-ai/dsh-agent'
+import type {} from '@deepseek-ai/dsh-commands'
+import type {} from '@deepseek-ai/dsh-compaction'
+import type {} from '@deepseek-ai/dsh-goal'
+import type {} from '@deepseek-ai/dsh-llm-retry'
+import type {} from '@deepseek-ai/dsh-plan-mode'
+import type {} from '@deepseek-ai/dsh-permission-presets'
+import type {} from '@deepseek-ai/dsh-schedule'
+import type {} from '@deepseek-ai/dsh-sandbox-policy'
+import type {} from '@deepseek-ai/dsh-session-title'
+// The subagent package's durable catalog event joins the union the same way
+// (parent-owned facts; the fold itself lives with the live feed).
+import type {} from '@deepseek-ai/dsh-subagent'
+import { toolArgumentsPreview, toolPromptPreview } from './tool-preview.ts'
+import { toolResultDetail, type ToolDetail } from './tool-detail.ts'
+
+/** In-flight UI buffers are tails; the assembled assistant message is authoritative. */
+const MAX_STREAMING_CHARS = 65_536
+
+/**
+ * Upper bound on remembered `compaction/summary` shadow prices waiting for a
+ * matching `compaction/end`. Compactions are sequential and rare, so a few
+ * slots suffice; an aborted compaction (summary without end) otherwise leaves
+ * an unbounded residue in `anchors.compactionTokens`. An evicted price
+ * degrades to the documented `lastPruneTokens` fallback, exactly like a
+ * missing summary.
+ */
+const MAX_COMPACTION_SUMMARY_RESIDUE = 16
+
+/** Append one delta without retaining an unbounded duplicate of the live reply. */
+function appendStreamingTail(current: string, delta: string): string {
+  const next = current + delta
+  return next.length <= MAX_STREAMING_CHARS ? next : next.slice(-MAX_STREAMING_CHARS)
+}
+
+/** One user prompt line. */
+export interface UserEntry {
+  kind: 'user'
+  /** Joined text blocks of the user message. */
+  text: string
+  /** True for collapsed injected context (plugin/continuation notices), which
+   * the renderer marks with a dim ↳ instead of the user ❯ prompt. */
+  notice: boolean
+  /** Durable image references carried by this prompt. */
+  images?: readonly ImageBlock['attachment'][]
+  /** Durable file references carried by this prompt (0.1.5 file blocks). */
+  files?: readonly FileAttachmentRef[]
+}
+
+/** One user message waiting in the agent inbox (the web's queued-message row). */
+export interface PendingEntry {
+  kind: 'pending'
+  /** Stable message identity shared with the durable `user/message` that retires it. */
+  messageId: MessageId
+  /** Which inbox list holds the message: steering is consumed at the next step boundary. */
+  target: 'next-turn' | 'next-step'
+  /** Full message text — Codex PendingSteer renders queued prompts exactly like user rows. */
+  text: string
+  /** Durable image references queued with this prompt. */
+  images?: readonly ImageBlock['attachment'][]
+  /** Durable file references queued with this prompt (0.1.5 file blocks). */
+  files?: readonly FileAttachmentRef[]
+}
+
+/** One authoritative assembled assistant reply. */
+export interface AssistantEntry {
+  kind: 'assistant'
+  /** Joined text blocks of the assistant message. */
+  text: string
+  /** Joined reasoning blocks from the same assembled message. */
+  reasoning: string
+  /** True when a cancelled stream's delivered prefix was finalized as this
+   * entry (rc.8 `assistant/message.interrupted`) — rendered with a marker. */
+  interrupted?: true
+}
+
+/** One model-requested tool invocation and its settled state. */
+export interface ToolEntry {
+  kind: 'tool'
+  /** Correlation id shared with the matching `tool/result`. */
+  callId: string
+  /**
+   * Global tool-call ordinal across the whole transcript (1, 2, 3…, never
+   * reset between turns). The tool-card badge and every error line that
+   * references the failed call share this number, so "call N" in an error
+   * always names the exact card the badge shows.
+   */
+  ordinal: number
+  /** Tool name as the model addressed it. */
+  name: string
+  /** Raw arguments JSON string exactly as the model produced it. */
+  arguments: string
+  /** Bounded human-meaningful arguments preview for the tool card. */
+  preview: string
+  /** Bounded delegation prompt (subagent cards' second row), '' when none. */
+  prompt: string
+  /** Execution state; `running` until the paired result lands. */
+  state: 'running' | 'done' | 'error'
+  /** Bounded first text block of the result, empty until it lands. */
+  summary: string
+  /**
+   * Bounded expansion payload for the verbose transcript (Ctrl+O), derived
+   * from the tool's persisted presentation metadata; undefined until the
+   * result lands and only when something renderable exists.
+   */
+  detail: ToolDetail | undefined
+}
+
+/** One slash-command execution dispatched through `ctx.commands`. */
+export interface CommandEntry {
+  kind: 'command'
+  /** Pairing id shared with the matching `command/done`. */
+  commandId: string
+  /** Lowercase command name without the leading slash. */
+  name: string
+  /** Verbatim text following the command name. */
+  args: string
+  /** Execution state; `running` until the paired lifecycle event lands. */
+  state: 'running' | 'done' | 'error'
+  /** Handler outcome text, empty until it lands. */
+  summary: string
+}
+
+/** One turn-level failure surfaced from `turn/end`. */
+export interface ErrorEntry {
+  kind: 'error'
+  /** `code: message` of the failure. */
+  text: string
+}
+
+/** One non-error turn outcome surfaced from `turn/end`. */
+export interface TurnMarkerEntry {
+  kind: 'turn-marker'
+  /** Human-readable outcome line, dim-rendered. */
+  text: string
+}
+
+/** One completed compaction lifecycle surfaced from `compaction/end`. */
+export interface CompactionEntry {
+  kind: 'compaction'
+  /** True when the compaction completed, false when it failed. */
+  ok: boolean
+  /** Heuristic tokens shadowed by the compaction (summary or prune price). */
+  tokens: number
+  /** Failure text when `ok` is false, empty otherwise. */
+  error: string
+}
+
+/** One provider-routed model-request retry (the `llm/retry` pair). */
+export interface RetryEntry {
+  kind: 'retry'
+  /** Correlation id shared with the matching `llm/retry-started`. */
+  retryId: string
+  /** Retry policy mode from the event: `always` has no attempt cap. */
+  mode: 'normal' | 'always'
+  /** Attempt ordinal and its cap. */
+  attempt: number
+  max: number
+  /** Failure code that triggered the retry. */
+  code: string
+  /** Backoff wait before the next attempt, in ms. */
+  delayMs: number
+  /**
+   * `running` while the backoff waits, `done` once the attempt started — or
+   * when the turn ended first (the turn-end sweep finalizes orphans so they
+   * never pin the settled boundary).
+   */
+  state: 'running' | 'done'
+}
+
+/** Turn-tail deliverables: files mutated by the turn's diff-bearing tools. */
+export interface FilesEntry {
+  kind: 'files'
+  /** Unique mutated paths in call order, bounded. */
+  paths: readonly string[]
+}
+
+/** Ordered transcript items the renderer draws. */
+export type TranscriptEntry = UserEntry | PendingEntry | AssistantEntry | ToolEntry | CommandEntry | ErrorEntry | TurnMarkerEntry | CompactionEntry | RetryEntry | FilesEntry
+
+/** The live goal the status line badges, folded from `goal/change`. */
+export interface GoalFold {
+  /** Human-requested completion objective. */
+  objective: string
+  /** Durable lifecycle phase. */
+  phase: 'active' | 'paused' | 'blocked' | 'complete'
+  /** Highest admitted continuation round and its cap. */
+  rounds: number
+  max: number
+  /** Blocked explanation, empty outside the blocked phase. */
+  blocked: string
+}
+
+/** Cumulative token accounting folded from `assistant/message` usage reports. */
+export interface UsageTotals {
+  /** Prompt-side billed tokens: `inputTokens` plus both cache buckets. */
+  inputTokens: number
+  /** Completion-side tokens over the whole log. */
+  outputTokens: number
+  /** Cache-read tokens over the whole log (0 when the adapter reports none). */
+  cacheReadTokens: number
+}
+
+/**
+ * Estimated used tokens per context content type, folded from transcript
+ * events via {@link estimateTokens}. The segmented context bar's composition
+ * source: proportions across types are meaningful, absolute values are not
+ * (they never touch billing or the reported `lastPromptTokens`).
+ */
+export interface ContextSegments {
+  /** Rendered system-prompt text (latest `request/header`) plus injected-context notices. */
+  system: number
+  /** Direct human prompts (durable `user/message` rows). */
+  prompt: number
+  /** Assistant text blocks (visible replies). */
+  assistant: number
+  /** Assistant reasoning blocks (hidden thinking). */
+  thinking: number
+  /** Tool call arguments plus result text. */
+  tools: number
+}
+
+/** Window-scoped figures the status line shows; timing uses event timestamps. */
+export interface TranscriptStats {
+  /** Durable turns opened (`turn/start` events). */
+  turns: number
+  /** Model requests made (`step/start` events). */
+  steps: number
+  /** Summed model wall time: `step/start` → `assistant/message`, in ms. */
+  llmMs: number
+  /** Summed tool wall time: `tool/call` → `tool/result`, in ms. */
+  toolMs: number
+  /** Cumulative token accounting; input stays 0 until a report lands. */
+  usage: UsageTotals
+  /** Prompt-side size of the most recent reported request (context pressure). */
+  lastPromptTokens: number
+  /** Newest advertised route capacity, 0 when no adapter ever advertised one. */
+  contextWindow: number
+  /** Estimated used tokens per content type (the segmented bar's composition). */
+  contextSegments: ContextSegments
+  /** Summed first-token waits: `step/start` → first non-empty chunk, in ms. */
+  ttftMs: number
+  /** Steps that produced a first chunk (the TTFT average's denominator). */
+  ttftSteps: number
+  /** Summed decode spans: first chunk → `assistant/message`, in ms. */
+  decodeMs: number
+  /** Completion tokens over timed decode spans (the tok/s numerator). */
+  decodeTokens: number
+  /**
+   * Adapter-owned reasoning effort of the latest `request/header` config —
+   * the EFFECTIVE effort the session actually uses (a materialized model
+   * default is included, exactly as the adapter resolved it). Empty when the
+   * header carried none (provider-default behavior). The status line appends
+   * it to the model segment as `provider/model@effort`.
+   */
+  reasoningEffort: string
+}
+
+/** One active reminder folded from durable `schedule/change` events. */
+export interface ScheduleRow {
+  readonly id: string
+  readonly kind: 'after' | 'at' | 'every'
+  readonly prompt: string
+  /** Next due time (epoch ms); the /schedule panel derives overdue/relative labels. */
+  readonly targetAt: number
+  /** Recurrence seconds for 'every' rows, undefined otherwise. */
+  readonly everySeconds?: number
+}
+
+/**
+ * The durable `schedule/change` payload shape this fold consumes. Upstream
+ * strict-decodes the whole transition stream before appending, so unknown
+ * ids here are corrupt-input edges that degrade to a no-op.
+ */
+export interface ScheduleChangeLike {
+  readonly operation: 'create' | 'delete' | 'dispatch'
+  readonly schedule?: {
+    readonly id: string
+    readonly kind: 'after' | 'at' | 'every'
+    readonly prompt: string
+    readonly afterSeconds?: number
+    readonly everySeconds?: number
+    readonly scheduledAt: string
+  }
+  readonly id?: string
+  readonly acceptedAt?: string
+}
+
+/*
+ * Upstream record semantics (dsh-schedule types): `scheduledAt` is ALREADY
+ * the due instant — AfterScheduleRecord carries the RFC 3339 UTC target
+ * (delay included), EveryScheduleRecord carries the earliest anchor-aligned
+ * occurrence not yet dispatched. No kind ever adds its own interval on top.
+ */
+
+/** Fold one `schedule/change` into the active-reminder list (create/delete/dispatch). */
+export function applyScheduleChange(rows: readonly ScheduleRow[], data: ScheduleChangeLike): readonly ScheduleRow[] {
+  if (data.operation === 'create' && data.schedule !== undefined) {
+    const schedule = data.schedule
+    const row: ScheduleRow = {
+      id: schedule.id,
+      kind: schedule.kind,
+      prompt: schedule.prompt,
+      targetAt: Date.parse(schedule.scheduledAt),
+      ...(schedule.kind === 'every' ? { everySeconds: schedule.everySeconds ?? 0 } : {}),
+    }
+    return [...rows.filter(existing => existing.id !== schedule.id), row]
+  }
+  if (data.operation === 'delete' && data.id !== undefined) {
+    return rows.filter(existing => existing.id !== data.id)
+  }
+  if (data.operation === 'dispatch' && data.id !== undefined) {
+    // A dispatched one-shot reminder is finished. An 'every' reminder
+    // advances PAST every missed occurrence in one step: the next target is
+    // the first anchor-aligned instant strictly after acceptedAt, stepping
+    // from the previous aligned target (upstream advances the same way).
+    if (data.acceptedAt === undefined) return rows.filter(existing => existing.id !== data.id)
+    const accepted = Date.parse(data.acceptedAt)
+    return rows.map(existing => existing.id === data.id
+      ? { ...existing, targetAt: nextEveryTarget(existing.targetAt, accepted, existing.everySeconds ?? 0) }
+      : existing)
+  }
+  return rows
+}
+
+/** First anchor-aligned target after `acceptedAt`, stepping from the previous aligned target. */
+export function nextEveryTarget(previousTarget: number, acceptedAt: number, everySeconds: number): number {
+  const interval = Math.max(1, everySeconds) * 1000
+  if (acceptedAt <= previousTarget) return previousTarget + interval
+  const missed = Math.ceil((acceptedAt - previousTarget + 1) / interval)
+  return previousTarget + missed * interval
+}
+
+/** Plugin snapshot sources folded into token stats but never rendered as rows. */
+const HIDDEN_SNAPSHOT_PLUGINS = new Set(['time-context', 'tmux-context'])
+/** Plugin prompt sources rendered as full user rows (they ARE the conversation). */
+const REMINDER_PLUGINS = new Set(['schedule'])
+
+/** The complete TUI transcript view for one session. */
+export interface TranscriptView {
+  /** Settled entries in log order. */
+  entries: readonly TranscriptEntry[]
+  /** Bounded text tail accumulated from live stream frames since the last settlement. */
+  streaming: string
+  /** Bounded thinking tail accumulated from reasoning deltas since the last flush. */
+  streamingReasoning: string
+  /** Latest whole-list todo snapshot from `todo/write`, empty when none. */
+  todos: readonly TodoItem[]
+  /**
+   * Global tool-call ordinal counter: the number the NEXT `tool/call` lands
+   * with (1-based). Never reset, so the counter and the badges/error lines
+   * stay consistent across turns and resumed sessions.
+   */
+  toolCallOrdinal: number
+  /** True while a durable turn is open (`turn/start` … `turn/end`). */
+  busy: boolean
+  /** `turn/start` time of the open turn (0 while idle) — the web TurnStatus clock anchor. */
+  busySince: number
+  /** Figures the status line renders. */
+  stats: TranscriptStats
+  /**
+   * The `provider/model` pair of the last `request/header` snapshot — the
+   * session's own model record, which a resumed TUI prefers over the
+   * deployment default (mirrors the web host's resume selection order).
+   * Empty before the session's first request.
+   */
+  model: string
+  /** Plan mode state folded from the last `plan/mode` event. */
+  plan: boolean
+  /** Active permission preset folded from the last `permission/preset` event, empty before one. */
+  permission: string
+  /** Latest session title folded from the last `session/title` event, empty before one. */
+  title: string
+  /**
+   * Effective system prompt assembled from `system/message` surface nodes
+   * (v3): the head node's text joined with every later non-empty node, blank
+   * lines between. Empty before the first system node or when every node is
+   * empty ("no system prompt").
+   */
+  systemPrompt: string
+  /** Sandbox-mode override folded from the last `sandbox/mode` event, empty when never switched. */
+  sandbox: string
+  /** Current long-running goal folded from the last `goal/change`, undefined when cleared. */
+  goal: GoalFold | undefined
+  /** Active reminders folded from `schedule/change` events, oldest target first at render. */
+  schedules: readonly ScheduleRow[]
+  /**
+   * Ordered live message ids per inbox target, mirrored from
+   * `agent/inbox/spliced` exactly like the upstream Inbox projection — the
+   * coordinates later removals resolve against.
+   */
+  pending: { 'next-turn': readonly string[]; 'next-step': readonly string[] }
+  /**
+   * Fold-internal timing anchors, never rendered: open step and tool-call
+   * start timestamps the next `assistant/message` / `tool/result` resolves
+   * against. Keyed `turn:step` and by call id. `turnSteps`/`turnTools`
+   * track which step/tool anchors still belong to the open turn so
+   * `turn/end` (and a superseding `step/start`) can sweep anchors an
+   * interruption left behind; `turnFiles` keys mutated paths by turn.
+   */
+  readonly anchors: {
+    stepStart: Map<string, number>
+    toolStart: Map<string, number>
+    firstChunkAt: Map<string, number>
+    compactionTokens: Map<string, number>
+    lastPruneTokens: number
+    turnFiles: Map<number, Set<string>>
+    turnSteps: Map<number, string>
+    turnTools: Map<number, Set<string>>
+    /** Live `system/message` surface nodes by event seq (empty string = an empty node). */
+    systemNodes: Map<number, string>
+  }
+}
+
+/** Assemble the effective system prompt from surface nodes: head text plus every later non-empty node. */
+function assembleSystemPrompt(nodes: ReadonlyMap<number, string>): string {
+  if (nodes.size === 0) return ''
+  const ordered = [...nodes.entries()].sort((left, right) => left[0] - right[0])
+  const texts = ordered
+    .map(([, text]) => text)
+    .filter(text => text !== '')
+  return texts.join('\n\n')
+}
+
+/**
+ * Apply one surface event's replace to the live system nodes. Any surface
+ * event may shadow system nodes — the kernel's compaction summary lands as a
+ * `user/message` replace whose range can cover later system nodes (only node
+ * 0 is compaction-protected upstream) — so every surface fold retires covered
+ * nodes, not just `system/message` itself.
+ * @param nodes - the live system-node map (mutated when the event replaces).
+ * @param surfaceOp - the surface operation the event carries, when it is a
+ * surface event (log-only events have none and change nothing).
+ * @returns the reassembled prompt when nodes were retired, `changed: false`
+ * when the event shadows nothing.
+ */
+function retireShadowedSystemNodes(
+  nodes: Map<number, string>,
+  surfaceOp: 'append' | { readonly op: 'replace'; readonly startSeq: number; readonly endSeq: number } | undefined,
+): { readonly prompt: string; readonly changed: boolean } {
+  if (surfaceOp === undefined || surfaceOp === 'append') return { prompt: '', changed: false }
+  let changed = false
+  for (const seq of nodes.keys()) {
+    if (seq >= surfaceOp.startSeq && seq <= surfaceOp.endSeq) {
+      nodes.delete(seq)
+      changed = true
+    }
+  }
+  return changed ? { prompt: assembleSystemPrompt(nodes), changed } : { prompt: '', changed: false }
+}
+
+/** Join the text blocks of a content list; non-text blocks contribute nothing. */
+function textOf(content: readonly ContentBlock[]): string {
+  return content.filter(block => block.type === 'text').map(block => block.text).join('')
+}
+
+/**
+ * Snapshot-isolate one anchors block (Maps and their nested Sets): a view
+ * already handed to the renderer must never observe a later fold through a
+ * shared container. The collections are small and turn-bounded, so cloning
+ * per event is cheap next to the entries copy the reducer already makes.
+ */
+function cloneViewAnchors(anchors: TranscriptView['anchors']): TranscriptView['anchors'] {
+  return {
+    stepStart: new Map(anchors.stepStart),
+    toolStart: new Map(anchors.toolStart),
+    firstChunkAt: new Map(anchors.firstChunkAt),
+    compactionTokens: new Map(anchors.compactionTokens),
+    lastPruneTokens: anchors.lastPruneTokens,
+    turnFiles: new Map([...anchors.turnFiles].map(([turn, files]) => [turn, new Set(files)])),
+    turnSteps: new Map(anchors.turnSteps),
+    turnTools: new Map([...anchors.turnTools].map(([turn, tools]) => [turn, new Set(tools)])),
+    systemNodes: new Map(anchors.systemNodes),
+  }
+}
+
+/** Durable image references in their model-visible order. */
+function imagesOf(content: readonly ContentBlock[]): readonly ImageBlock['attachment'][] {
+  return content.filter((block): block is ImageBlock => block.type === 'image').map(block => block.attachment)
+}
+
+/** Durable file references in their model-visible order. */
+function filesOf(content: readonly ContentBlock[]): readonly FileAttachmentRef[] {
+  return content.filter((block): block is FileBlock => block.type === 'file').map(block => block.attachment)
+}
+
+/** Human-readable bounded image labels for transcript, inspector, and export surfaces. */
+export function imageLabels(images: readonly ImageBlock['attachment'][] | undefined): string {
+  if (images === undefined || images.length === 0) return ''
+  return images.map((image, index) => {
+    const rawName = image.name?.trim() || `image ${index + 1}`
+    const name = rawName.length <= 80 ? rawName : `${rawName.slice(0, 79)}…`
+    const original = image.originalDimensions
+    const dimensions = original === undefined
+      ? `${image.width}×${image.height}`
+      : `${image.width}×${image.height} · original ${original.width}×${original.height}`
+    return `[image: ${name} · ${dimensions} · ${image.bytes} B]`
+  }).join('\n')
+}
+
+/** Human-readable bounded file labels for the same surfaces (0.1.5 file blocks). */
+export function fileLabels(files: readonly FileAttachmentRef[] | undefined): string {
+  if (files === undefined || files.length === 0) return ''
+  return files.map((file, index) => {
+    const rawName = file.name?.trim() || `file ${index + 1}`
+    const name = rawName.length <= 80 ? rawName : `${rawName.slice(0, 79)}…`
+    return `[file: ${name} · ${file.bytes} B]`
+  }).join('\n')
+}
+
+/** Prompt text with its durable image and file labels, without exposing local paths or bytes. */
+export function promptDisplayText(entry: Pick<UserEntry | PendingEntry, 'text' | 'images' | 'files'>): string {
+  const imageText = imageLabels(entry.images)
+  const fileText = fileLabels(entry.files)
+  const labels = imageText === '' ? fileText : fileText === '' ? imageText : `${imageText}\n${fileText}`
+  return entry.text === '' ? labels : labels === '' ? entry.text : `${entry.text}\n${labels}`
+}
+
+/** Join the reasoning blocks of a content list; non-reasoning blocks contribute nothing. */
+function reasoningOf(content: readonly ContentBlock[]): string {
+  return content.filter(block => block.type === 'reasoning').map(block => block.text).join('')
+}
+
+/**
+ * Rough token estimate for the segmented context bar (pi-nano-context's ~4
+ * chars/token heuristic, CJK-aware so a Chinese prompt is not quartered):
+ * CJK/wide chars cost ~1 token each, ASCII ~4 chars per token. Estimates
+ * drive bar PROPORTIONS, never billing, so precision is not required.
+ * @param text - the text to estimate.
+ * @returns an integer token estimate, 0 for empty text.
+ */
+function estimateTokens(text: string): number {
+  let wide = 0
+  let narrow = 0
+  for (const cluster of splitGraphemes(text)) {
+    if (graphemeWidth(cluster) > 1) wide += 1
+    else narrow += 1
+  }
+  return wide + Math.ceil(narrow / 4)
+}
+
+/** A fresh, empty transcript view. */
+export function createTranscriptView(): TranscriptView {
+  return {
+    entries: [],
+    streaming: '',
+    streamingReasoning: '',
+    todos: [],
+    toolCallOrdinal: 0,
+    busy: false,
+    busySince: 0,
+    model: '',
+    plan: false,
+    permission: '',
+    title: '',
+    systemPrompt: '',
+    sandbox: '',
+    goal: undefined,
+    schedules: [],
+    pending: { 'next-turn': [], 'next-step': [] },
+    stats: { turns: 0, steps: 0, llmMs: 0, toolMs: 0, usage: { inputTokens: 0, outputTokens: 0, cacheReadTokens: 0 }, lastPromptTokens: 0, contextWindow: 0, contextSegments: { system: 0, prompt: 0, assistant: 0, thinking: 0, tools: 0 }, ttftMs: 0, ttftSteps: 0, decodeMs: 0, decodeTokens: 0, reasoningEffort: '' },
+    anchors: { stepStart: new Map(), toolStart: new Map(), firstChunkAt: new Map(), compactionTokens: new Map(), lastPruneTokens: 0, turnFiles: new Map(), turnSteps: new Map(), turnTools: new Map(), systemNodes: new Map() },
+  }
+}
+
+/** Full prompt text of a queued message (identical to the durable user row it retires into). */
+function pendingText(content: readonly ContentBlock[]): string {
+  return textOf(content)
+}
+
+/**
+ * Fold one session event into an updated view (copy-on-write).
+ * @param view - the view before the event.
+ * @param event - one durable session event from `session/event` or the log.
+ * @returns the view after the event; the input view is never mutated.
+ */
+export function projectEvent(view: TranscriptView, event: SessionEvent): TranscriptView {
+  // Fold against a private anchors block so the documented contract holds —
+  // "the input view is never mutated" — even for the in-place anchor sweeps
+  // below; without this, every handed-out view shared live Maps.
+  view = { ...view, anchors: cloneViewAnchors(view.anchors) }
+  // Any surface event's replace may shadow system nodes (a compaction
+  // summary lands as a user/message replace whose range can cover later
+  // system nodes); the switch below folds against the shadowed view.
+  const shadow = retireShadowedSystemNodes(view.anchors.systemNodes, event.surfaceOp)
+  if (shadow.changed) {
+    view = {
+      ...view,
+      systemPrompt: shadow.prompt,
+      stats: { ...view.stats, contextSegments: { ...view.stats.contextSegments, system: estimateTokens(shadow.prompt) } },
+    }
+  }
+  switch (event.type) {
+    case 'user/message': {
+      // A queued row retires when its durable user message lands (the agent
+      // claims the inbox and logs the same message identity) — the transient
+      // steering/queued preview yields to the real transcript entry.
+      const message = event.data
+      let entries = view.entries
+      let pending = view.pending
+      for (const target of ['next-turn', 'next-step'] as const) {
+        const index = pending[target].indexOf(message.id)
+        if (index < 0) continue
+        pending = { ...pending, [target]: pending[target].filter((_, i) => i !== index) }
+        entries = entries.filter(entry => !(entry.kind === 'pending' && entry.messageId === message.id))
+      }
+      // Injected context (plugin/model-continuation sources) stays collapsed
+      // to a bounded notice row, exactly like collapsed transcript context
+      // elsewhere in the product; only direct human prompts render in full.
+      const text = textOf(message.content)
+      const images = imagesOf(message.content)
+      const files = filesOf(message.content)
+      if (message.source.kind === 'user') {
+        return {
+          ...view,
+          pending,
+          entries: [...entries, { kind: 'user', text, notice: false, ...(images.length === 0 ? {} : { images }), ...(files.length === 0 ? {} : { files }) }],
+          stats: {
+            ...view.stats,
+            contextSegments: {
+              ...view.stats.contextSegments,
+              prompt: view.stats.contextSegments.prompt + estimateTokens(text),
+            },
+          },
+        }
+      }
+      // Snapshot injections (time/tmux context) still spend model context but
+      // render nothing; the schedule reminder is a real prompt and renders in
+      // full — the model acts on it, so the transcript must show it.
+      if (message.source.kind === 'plugin' && HIDDEN_SNAPSHOT_PLUGINS.has(message.source.plugin)) {
+        return {
+          ...view,
+          pending,
+          entries,
+          stats: {
+            ...view.stats,
+            contextSegments: {
+              ...view.stats.contextSegments,
+              system: view.stats.contextSegments.system + estimateTokens(text),
+            },
+          },
+        }
+      }
+      if (message.source.kind === 'plugin' && REMINDER_PLUGINS.has(message.source.plugin)) {
+        return {
+          ...view,
+          pending,
+          entries: [...entries, { kind: 'user', text, notice: false, ...(images.length === 0 ? {} : { images }), ...(files.length === 0 ? {} : { files }) }],
+          stats: {
+            ...view.stats,
+            contextSegments: {
+              ...view.stats.contextSegments,
+              prompt: view.stats.contextSegments.prompt + estimateTokens(text),
+            },
+          },
+        }
+      }
+      const notice = message.source.kind === 'plugin' && message.source.form === 'notice'
+        ? message.source.summary
+        : message.source.kind === 'plugin'
+          ? message.source.plugin
+          : message.source.kind
+      const summary = boundContextSummary(notice)
+      return {
+        ...view,
+        pending,
+        entries: [...entries, { kind: 'user', text: summary, notice: true }],
+        stats: {
+          ...view.stats,
+          contextSegments: {
+            ...view.stats.contextSegments,
+            system: view.stats.contextSegments.system + estimateTokens(summary),
+          },
+        },
+      }
+    }
+    case 'agent/inbox/spliced': {
+      // The durable inbox mutation (web queue-mirror contract, event-sourced):
+      // removals drop the projected rows at their inbox coordinates, inserted
+      // messages gain a pending row at their log position.
+      const { target, start, removedCount = 0, inserted } = event.data
+      const ids = view.pending[target]
+      const removed = ids.slice(start, start + removedCount)
+      // In-place upstream semantics: the kernel's authoritative fold is
+      // `inbox.splice(start, removedCount, ...inserted)` — inserted ids land
+      // AT the splice position (prepend/replace shapes), never at the tail.
+      // A tail append diverged the id order, so later coordinate-based events
+      // (next-turn head claims, positioned remove/replace) tombstoned the
+      // wrong pending row.
+      const nextIds = [
+        ...ids.slice(0, start),
+        ...inserted.map(message => message.id),
+        ...ids.slice(start + removedCount),
+      ]
+      let entries = view.entries
+      if (removed.length > 0) {
+        const removedSet = new Set(removed)
+        entries = entries.filter(entry =>
+          !(entry.kind === 'pending' && entry.target === target && removedSet.has(entry.messageId)))
+      }
+      for (const message of inserted) {
+        entries = [...entries, {
+          kind: 'pending',
+          messageId: message.id,
+          target,
+          text: pendingText(message.content),
+          ...imagesOf(message.content).length === 0 ? {} : { images: imagesOf(message.content) },
+          ...filesOf(message.content).length === 0 ? {} : { files: filesOf(message.content) },
+        }]
+      }
+      return { ...view, entries, pending: { ...view.pending, [target]: nextIds } }
+    }
+    case 'system/message': {
+      // Session-log v3 carries the system prompt as surface nodes (the
+      // `request/header.system` field is gone): append adds one node;
+      // replace(startSeq, endSeq) retires the covered nodes and this event's
+      // node takes their place. The effective prompt is the head text joined
+      // with every later non-empty node (kernel in-history assembly), and
+      // the context estimate prices that assembly — a tail-clearing
+      // replacement never zeroes a surviving head.
+      const text = textOf(event.data.message.content)
+      const nodes = view.anchors.systemNodes
+      retireShadowedSystemNodes(nodes, event.surfaceOp)
+      nodes.set(event.seq, text)
+      const systemPrompt = assembleSystemPrompt(nodes)
+      return {
+        ...view,
+        systemPrompt,
+        stats: {
+          ...view.stats,
+          contextSegments: {
+            ...view.stats.contextSegments,
+            system: estimateTokens(systemPrompt),
+          },
+        },
+      }
+    }
+    case 'assistant/attempt': {
+      // A failed, retried, cancelled, or stream-error attempt that produced no
+      // surface message (session-log v2+ folds its chunk stream in here). The
+      // step is NOT closed — the kernel's session-stats keeps one step start
+      // across in-step retries, so llmMs spans them; keep the anchors so a
+      // retrying attempt and its final settlement time the step from one
+      // start. An attempt whose first token streamed only live (the frames
+      // died before the fold) restores its first-token anchor from the
+      // embedded stream exactly like a replayed one.
+      const key = `${event.data.turn}:${event.data.step}`
+      let stats = view.stats
+      if (!view.anchors.firstChunkAt.has(key)) {
+        const first = assistantStreamFirstTokenTime(event.data.stream ?? [])
+        if (first !== undefined) {
+          view.anchors.firstChunkAt.set(key, first)
+          const started = view.anchors.stepStart.get(key)
+          if (started !== undefined) {
+            stats = {
+              ...stats,
+              ttftMs: stats.ttftMs + Math.max(0, first - started),
+              ttftSteps: stats.ttftSteps + 1,
+            }
+          }
+        }
+      }
+      if (view.streaming === '' && view.streamingReasoning === '' && stats === view.stats) return view
+      return { ...view, streaming: '', streamingReasoning: '', stats }
+    }
+    case 'assistant/message': {
+      // The assembled message is authoritative; drop the streamed buffers.
+      const key = `${event.data.turn}:${event.data.step}`
+      const started = view.anchors.stepStart.get(key)
+      view.anchors.stepStart.delete(key)
+      // Live streaming anchored the first token through the process-local
+      // assistant-stream frames; a replayed settlement has no live frames, so
+      // the embedded stream's own first-token time restores the same anchor
+      // AND the TTFT figures live frames accumulate on the live path.
+      let firstChunk = view.anchors.firstChunkAt.get(key)
+      let stats = view.stats
+      if (firstChunk === undefined) {
+        firstChunk = assistantStreamFirstTokenTime(event.data.stream ?? [])
+        if (firstChunk !== undefined && started !== undefined) {
+          stats = {
+            ...stats,
+            ttftMs: stats.ttftMs + Math.max(0, firstChunk - started),
+            ttftSteps: stats.ttftSteps + 1,
+          }
+        }
+      }
+      view.anchors.firstChunkAt.delete(key)
+      // The assembled message consumes the turn's current step anchor; a
+      // later `turn/end` sweep then has nothing left to clean for this step.
+      if (view.anchors.turnSteps.get(event.data.turn) === key) view.anchors.turnSteps.delete(event.data.turn)
+      const usage = event.data.usage
+      const totals = stats.usage
+      const text = textOf(event.data.message.content)
+      const reasoning = reasoningOf(event.data.message.content)
+      return {
+        ...view,
+        streaming: '',
+        streamingReasoning: '',
+        entries: [...view.entries, { kind: 'assistant', text, reasoning, interrupted: event.data.interrupted === true ? true : undefined }],
+        stats: {
+          ...stats,
+          llmMs: stats.llmMs + (started === undefined ? 0 : Math.max(0, event.time - started)),
+          usage: usage === undefined ? totals : {
+            inputTokens: totals.inputTokens + usage.inputTokens + (usage.cacheReadTokens ?? 0) + (usage.cacheWriteTokens ?? 0),
+            outputTokens: totals.outputTokens + usage.outputTokens,
+            cacheReadTokens: totals.cacheReadTokens + (usage.cacheReadTokens ?? 0),
+          },
+          lastPromptTokens: usage === undefined ? stats.lastPromptTokens
+            : usage.inputTokens + (usage.cacheReadTokens ?? 0) + (usage.cacheWriteTokens ?? 0),
+          // Decode span and its tokens pair up: an un-timed step (no first
+          // chunk landed) contributes neither, so the rate stays honest.
+          decodeMs: stats.decodeMs + (firstChunk === undefined ? 0 : Math.max(0, event.time - firstChunk)),
+          decodeTokens: stats.decodeTokens + (firstChunk === undefined || usage === undefined ? 0 : usage.outputTokens),
+          contextSegments: {
+            ...stats.contextSegments,
+            thinking: stats.contextSegments.thinking + estimateTokens(reasoning),
+            assistant: stats.contextSegments.assistant + estimateTokens(text),
+          },
+        },
+      }
+    }
+    case 'tool/call': {
+      const data = event.data
+      view.anchors.toolStart.set(data.callId, event.time)
+      // Remember the call's turn so `turn/end` can sweep a start that never
+      // pairs with a result (an interrupted tool otherwise leaks its anchor).
+      const turnTools = view.anchors.turnTools.get(data.turn) ?? new Set<string>()
+      turnTools.add(data.callId)
+      view.anchors.turnTools.set(data.turn, turnTools)
+      const ordinal = view.toolCallOrdinal + 1
+      return {
+        ...view,
+        toolCallOrdinal: ordinal,
+        entries: [
+          ...view.entries,
+          {
+            kind: 'tool',
+          callId: data.callId,
+          ordinal,
+          name: data.name,
+          arguments: data.arguments,
+          preview: toolArgumentsPreview(data.arguments, data.name),
+          prompt: toolPromptPreview(data.name, data.arguments),
+          state: 'running',
+          summary: '',
+          detail: undefined,
+        }],
+        stats: {
+          ...view.stats,
+          contextSegments: {
+            ...view.stats.contextSegments,
+            tools: view.stats.contextSegments.tools
+              + (typeof data.arguments === 'string' ? estimateTokens(data.arguments) : 0),
+          },
+        },
+      }
+    }
+    case 'tool/result': {
+      const block = event.data.message.content[0]
+      const started = view.anchors.toolStart.get(block.toolCallId)
+      view.anchors.toolStart.delete(block.toolCallId)
+      // Deregister the call from its turn's registry so `turn/end` does not
+      // sweep a start that already paired with a result.
+      const turnTools = view.anchors.turnTools.get(event.data.turn)
+      if (turnTools !== undefined) {
+        turnTools.delete(block.toolCallId)
+        if (turnTools.size === 0) view.anchors.turnTools.delete(event.data.turn)
+      }
+      const rawText = textOf(block.content)
+      const summary = boundContextSummary(rawText)
+      // The verbose expansion self-serves from the persisted presentation
+      // metadata (diffs, read windows, web sources) with the bounded raw text
+      // as the universal fallback — the capable-UI degradation ladder.
+      const detail = toolResultDetail(event.data.meta, rawText)
+      // Turn-tail deliverables: a diff-bearing mutation records its paths.
+      if (detail?.kind === 'diff') {
+        const set = view.anchors.turnFiles.get(event.data.turn) ?? new Set<string>()
+        for (const diff of detail.diffs) set.add(diff.path)
+        view.anchors.turnFiles.set(event.data.turn, set)
+      }
+      const entries = view.entries.map((entry) => {
+        if (entry.kind !== 'tool' || entry.callId !== block.toolCallId) return entry
+        return { ...entry, state: block.isError === true ? 'error' as const : 'done' as const, summary, detail }
+      })
+      return {
+        ...view,
+        entries,
+        stats: {
+          ...view.stats,
+          toolMs: view.stats.toolMs + (started === undefined ? 0 : Math.max(0, event.time - started)),
+          contextSegments: {
+            ...view.stats.contextSegments,
+            tools: view.stats.contextSegments.tools + estimateTokens(rawText),
+          },
+        },
+      }
+    }
+    case 'todo/write':
+      return { ...view, todos: event.data.todos }
+    case 'turn/start':
+      // The web todo projection clears on turn/start: a fresh turn's first
+      // write is the authoritative list, and a stale snapshot must not linger
+      // through a turn that has not written one yet.
+      return {
+        ...view,
+        busy: true,
+        busySince: view.busy ? view.busySince : event.time,
+        todos: [],
+        stats: { ...view.stats, turns: view.stats.turns + 1 },
+      }
+    case 'step/start': {
+      // A step supersedes the turn's previous step: if that step never
+      // assembled a message (interrupted), its timing anchors are stale the
+      // moment the next step opens and are swept here instead of leaking.
+      const key = `${event.data.turn}:${event.data.step}`
+      const previous = view.anchors.turnSteps.get(event.data.turn)
+      if (previous !== undefined && previous !== key) {
+        view.anchors.stepStart.delete(previous)
+        view.anchors.firstChunkAt.delete(previous)
+      }
+      view.anchors.turnSteps.set(event.data.turn, key)
+      view.anchors.stepStart.set(key, event.time)
+      return {
+        ...view,
+        streaming: '',
+        streamingReasoning: '',
+        stats: { ...view.stats, steps: view.stats.steps + 1 },
+      }
+    }
+    case 'turn/end': {
+      const reason = event.data.reason
+      const appended: TranscriptEntry[] = []
+      if (reason.kind === 'error') {
+        const recovery = reason.error.code === 'MISSING_CREDENTIAL'
+          ? ' · open /model to add an API key'
+          : ''
+        appended.push({ kind: 'error', text: `${reason.error.code}: ${reason.error.message}${recovery}` })
+      } else {
+        // Non-error outcomes deserve their own durable row (the web renders
+        // distinct max-tokens / abort / interruption nodes); `completed` stays
+        // silent so an ordinary turn never grows a marker.
+        const marker = reason.kind === 'aborted'
+          ? reason.reason.kind === 'user' ? 'turn cancelled by the user' : `turn cancelled (${reason.reason.kind})`
+          : reason.kind === 'max-tokens'
+            ? 'turn hit the output-token ceiling (max-tokens)'
+            : reason.kind === 'blocked'
+              ? 'turn ended blocked'
+              : reason.kind === 'interrupted'
+                ? 'turn was interrupted by a restart'
+                : undefined
+        if (marker !== undefined) appended.push({ kind: 'turn-marker', text: marker })
+      }
+      // Deliverables ride the turn tail (the web's turnTail chips): the
+      // turn's mutated files flush as one bounded row, then the set resets.
+      const files = view.anchors.turnFiles.get(event.data.turn)
+      view.anchors.turnFiles.delete(event.data.turn)
+      if (files !== undefined && files.size > 0) appended.push({ kind: 'files', paths: [...files].slice(0, 12) })
+      // Derivable boundary sweep: the turn is over, so any step/tool anchors
+      // it left behind (interruptions that never produced their message or
+      // result) can never be resolved and are reclaimed now.
+      const stepKey = view.anchors.turnSteps.get(event.data.turn)
+      if (stepKey !== undefined) {
+        view.anchors.stepStart.delete(stepKey)
+        view.anchors.firstChunkAt.delete(stepKey)
+        view.anchors.turnSteps.delete(event.data.turn)
+      }
+      const turnToolSet = view.anchors.turnTools.get(event.data.turn)
+      if (turnToolSet !== undefined) {
+        for (const callId of turnToolSet) view.anchors.toolStart.delete(callId)
+        view.anchors.turnTools.delete(event.data.turn)
+      }
+      // Orphaned retry/command rows can never be resolved after the turn
+      // ends: an aborted retry backoff returns upstream without its
+      // `llm/retry-started`, and crash repair synthesizes only tool/step/
+      // turn closers. Left `running` they pin the settled boundary forever,
+      // so the turn end finalizes them exactly like the anchor sweep above.
+      let orphans = false
+      const swept = view.entries.map((entry) => {
+        if (entry.kind === 'retry' && entry.state === 'running') {
+          orphans = true
+          return { ...entry, state: 'done' as const }
+        }
+        if (entry.kind === 'command' && entry.state === 'running') {
+          orphans = true
+          return { ...entry, state: 'error' as const, summary: 'interrupted before the turn ended' }
+        }
+        return entry
+      })
+      const entries = orphans ? swept : view.entries
+      if (appended.length === 0) {
+        return { ...view, busy: false, busySince: 0, streaming: '', streamingReasoning: '', entries }
+      }
+      return {
+        ...view,
+        busy: false,
+        busySince: 0,
+        streaming: '',
+        streamingReasoning: '',
+        entries: [...entries, ...appended],
+      }
+    }
+    case 'llm/retry': {
+      const data = event.data
+      return {
+        ...view,
+        streaming: '',
+        streamingReasoning: '',
+        entries: [...view.entries, {
+          kind: 'retry',
+          retryId: data.retryId,
+          mode: data.mode,
+          attempt: data.retry,
+          max: 'maxRetries' in data ? data.maxRetries : data.retry,
+          code: data.failure.code,
+          delayMs: data.delayMs,
+          state: 'running',
+        }],
+      }
+    }
+    case 'llm/retry-started': {
+      const data = event.data
+      const entries = view.entries.map((entry) => {
+        if (entry.kind !== 'retry' || entry.retryId !== data.retryId) return entry
+        return { ...entry, state: 'done' as const }
+      })
+      return { ...view, entries }
+    }
+    case 'sandbox/mode':
+      // Log-only override switch; last write wins for the status badge.
+      return { ...view, sandbox: event.data.mode }
+    case 'goal/change': {
+      const data = event.data
+      const clip = (text: string): string => (text.length > 60 ? `${text.slice(0, 59)}…` : text)
+      if (data.operation === 'clear') {
+        return {
+          ...view,
+          goal: undefined,
+          entries: [...view.entries, { kind: 'turn-marker', text: '◎ goal cleared' }],
+        }
+      }
+      const goal: GoalFold = {
+        objective: data.goal.objective,
+        phase: data.goal.phase,
+        rounds: data.roundsStarted,
+        max: data.goal.maxGoalRounds,
+        blocked: data.goal.blockedReason?.message ?? '',
+      }
+      const line = data.operation === 'create'
+        ? `◎ goal: ${clip(data.goal.objective)}`
+        : data.operation === 'complete'
+          ? '◎ goal complete'
+          : data.operation === 'pause'
+            ? '◎ goal paused'
+            : data.operation === 'resume'
+              ? '◎ goal resumed'
+              : data.operation === 'block'
+                ? `◎ goal blocked: ${clip(goal.blocked)}`
+                : undefined
+      return {
+        ...view,
+        goal,
+        entries: line === undefined ? view.entries : [...view.entries, { kind: 'turn-marker', text: line }],
+      }
+    }
+    case 'schedule/change':
+      // Non-conversational catalog state: the /schedule panel renders the
+      // active list, the transcript shows only the reminder prompts
+      // (handled at user/message above).
+      return { ...view, schedules: applyScheduleChange(view.schedules, event.data) }
+    case 'session/title':
+      // Latest-wins title snapshot, log-only; the status line prefers it.
+      return { ...view, title: event.data.title }
+    case 'compaction/summary':
+      // Remember the shadow price so the matching `compaction/end` row can
+      // state what the compaction reclaimed. The map is capped so an aborted
+      // compaction (summary without end) cannot leave an unbounded residue.
+      if (view.anchors.compactionTokens.size >= MAX_COMPACTION_SUMMARY_RESIDUE) {
+        const oldest = view.anchors.compactionTokens.keys().next().value
+        if (oldest !== undefined) view.anchors.compactionTokens.delete(oldest)
+      }
+      view.anchors.compactionTokens.set(event.data.compactionId, event.data.shadowedTokenCount)
+      return view
+    case 'compaction/prune':
+      // A model-free prune carries no compaction id; its price serves the next
+      // `compaction/end` that cannot find a summary price.
+      return { ...view, anchors: { ...view.anchors, lastPruneTokens: event.data.shadowedTokenCount } }
+    case 'compaction/end': {
+      const ok = event.data.error === undefined
+      const tokens = view.anchors.compactionTokens.get(event.data.compactionId) ?? view.anchors.lastPruneTokens
+      view.anchors.compactionTokens.delete(event.data.compactionId)
+      return {
+        ...view,
+        entries: [...view.entries, { kind: 'compaction', ok, tokens, error: event.data.error ?? '' }],
+      }
+    }
+    case 'request/context':
+      // Route capacity, logged only when it changes; last one wins.
+      return {
+        ...view,
+        stats: { ...view.stats, contextWindow: event.data.contextWindow ?? view.stats.contextWindow },
+      }
+    case 'request/header': {
+      // The session's own model record: the latest snapshot's provider/model
+      // pair, exactly what a resumed TUI restores as the selection, plus the
+      // effective reasoning effort that snapshot carried (the adapter may
+      // materialize the model default, which is what the status line shows).
+      // The system-prompt estimate lives with `system/message` events since
+      // session-log v3 removed the header's `system` field.
+      const config = event.data.header.config
+      return {
+        ...view,
+        model: `${config.provider}/${config.model}`,
+        stats: {
+          ...view.stats,
+          reasoningEffort: config.reasoningEffort === undefined ? '' : String(config.reasoningEffort),
+        },
+      }
+    }
+    case 'plan/mode':
+      // Whole-value replace; the last one wins (upstream fold semantics).
+      return { ...view, plan: event.data.active }
+    case 'permission/preset':
+      return { ...view, permission: event.data.preset }
+    case 'command/run': {
+      const data = event.data
+      return {
+        ...view,
+        entries: [...view.entries, {
+          kind: 'command',
+          commandId: data.commandId,
+          name: data.name,
+          args: data.args ?? '',
+          state: 'running',
+          summary: '',
+        }],
+      }
+    }
+    case 'command/done': {
+      const data = event.data
+      const entries = view.entries.map((entry) => {
+        if (entry.kind !== 'command' || entry.commandId !== data.commandId) return entry
+        return {
+          ...entry,
+          state: data.kind === 'success' ? 'done' as const : 'error' as const,
+          summary: boundContextSummary(data.text ?? ''),
+        }
+      })
+      return { ...view, entries }
+    }
+    default:
+      return view
+  }
+}
+
+/**
+ * Mutable replay accumulator: folds a persisted log into the identical view
+ * `projectEvent` would produce, but in near-linear time. Where `projectEvent`
+ * is copy-on-write — every append/scan rebuilds the whole `entries` array, so
+ * folding a full log costs O(N²) — the accumulator appends by push, resolves
+ * id-keyed updates (tool/result, command/done, retry-started) through index
+ * maps, and tombstones retired pending rows, so the whole log folds in O(N)
+ * plus one compaction pass when tombstones exist.
+ *
+ * Index maps never delete: every appended row registers its index, so an id
+ * lookup miss provably means no matching row exists and the update is an O(1)
+ * no-op (a malicious/orphan-heavy log cannot force per-orphan full-array
+ * scans). Each id maps to ALL of its indices, so a duplicate id updates every
+ * matching row exactly like the copy-on-write reducer.
+ *
+ * @internal Exported only so tests can (a) prove replay ≡ sequential
+ * `projectEvent` folds and (b) assert the linear complexity deterministically
+ * via {@link ReplayAccumulator.ops}, which counts entry-level container work
+ * instead of relying on wall-clock thresholds. No public consumer.
+ */
+export interface ReplayAccumulator {
+  /** Working entry list; `undefined` marks a retired pending row (tombstone). */
+  entries: (TranscriptEntry | undefined)[]
+  /** callId → every index into `entries` holding a `tool` row with that id. */
+  toolIndex: Map<string, number[]>
+  /** commandId → every index into `entries` holding a `command` row with that id. */
+  commandIndex: Map<string, number[]>
+  /** retryId → every index into `entries` holding a `retry` row with that id. */
+  retryIndex: Map<string, number[]>
+  /** messageId → every index into `entries` holding a `pending` row with that id. */
+  pendingIndex: Map<string, number[]>
+  /** Tombstone count; zero means `entries` is already the final array. */
+  removedCount: number
+  /** Mutable inbox id lists, mirroring `view.pending` order per target. */
+  pendingTurn: string[]
+  pendingStep: string[]
+  streaming: string
+  streamingReasoning: string
+  todos: readonly TodoItem[]
+  /** Global tool-call ordinal counter (see `TranscriptView.toolCallOrdinal`). */
+  toolCallOrdinal: number
+  busy: boolean
+  busySince: number
+  model: string
+  plan: boolean
+  permission: string
+  title: string
+  systemPrompt: string
+  sandbox: string
+  goal: GoalFold | undefined
+  schedules: readonly ScheduleRow[]
+  stats: TranscriptStats
+  stepStart: Map<string, number>
+  toolStart: Map<string, number>
+  firstChunkAt: Map<string, number>
+  compactionTokens: Map<string, number>
+  lastPruneTokens: number
+  turnFiles: Map<number, Set<string>>
+  turnSteps: Map<number, string>
+  turnTools: Map<number, Set<string>>
+  /** Live `system/message` surface nodes by event seq (empty string = an empty node). */
+  systemNodes: Map<number, string>
+  /** Entry-level container operations performed so far (test instrumentation). */
+  ops: number
+}
+
+/** @internal A fresh replay accumulator whose state mirrors `createTranscriptView()`. */
+export function createReplayAccumulator(): ReplayAccumulator {
+  return {
+    entries: [],
+    toolIndex: new Map(),
+    commandIndex: new Map(),
+    retryIndex: new Map(),
+    pendingIndex: new Map(),
+    removedCount: 0,
+    pendingTurn: [],
+    pendingStep: [],
+    streaming: '',
+    streamingReasoning: '',
+    todos: [],
+    toolCallOrdinal: 0,
+    busy: false,
+    busySince: 0,
+    model: '',
+    plan: false,
+    permission: '',
+    title: '',
+    systemPrompt: '',
+    sandbox: '',
+    goal: undefined,
+    schedules: [],
+    stats: { turns: 0, steps: 0, llmMs: 0, toolMs: 0, usage: { inputTokens: 0, outputTokens: 0, cacheReadTokens: 0 }, lastPromptTokens: 0, contextWindow: 0, contextSegments: { system: 0, prompt: 0, assistant: 0, thinking: 0, tools: 0 }, ttftMs: 0, ttftSteps: 0, decodeMs: 0, decodeTokens: 0, reasoningEffort: '' },
+    stepStart: new Map(),
+    toolStart: new Map(),
+    firstChunkAt: new Map(),
+    compactionTokens: new Map(),
+    lastPruneTokens: 0,
+    turnFiles: new Map(),
+    turnSteps: new Map(),
+    turnTools: new Map(),
+    systemNodes: new Map(),
+    ops: 0,
+  }
+}
+
+/** Append one entry (O(1)) and account the push. */
+function appendReplayEntry(acc: ReplayAccumulator, entry: TranscriptEntry): void {
+  acc.entries.push(entry)
+  acc.ops += 1
+}
+
+/**
+ * Get (or create) the index list an id owns. Lists are never removed: every
+ * appended row registers its index, so a lookup miss later proves no matching
+ * row exists and the caller can no-op in O(1).
+ */
+function indexList(map: Map<string, number[]>, id: string): number[] {
+  let list = map.get(id)
+  if (list === undefined) {
+    list = []
+    map.set(id, list)
+  }
+  return list
+}
+
+/**
+ * Finalize replay rows the ended turn left `running`, mirroring the reducer's
+ * turn-end orphan sweep: an orphaned retry settles `done`, an orphaned command
+ * settles `error` with an interruption note. Only the id-indexed rows are
+ * visited, so the sweep stays O(retries+commands of the log), never a scan.
+ */
+function finalizeReplayOrphans(acc: ReplayAccumulator): void {
+  for (const list of acc.retryIndex.values()) {
+    for (const index of list) {
+      const entry = acc.entries[index]
+      if (entry !== undefined && entry.kind === 'retry' && entry.state === 'running') {
+        acc.entries[index] = { ...entry, state: 'done' }
+      }
+    }
+  }
+  for (const list of acc.commandIndex.values()) {
+    for (const index of list) {
+      const entry = acc.entries[index]
+      if (entry !== undefined && entry.kind === 'command' && entry.state === 'running') {
+        acc.entries[index] = { ...entry, state: 'error', summary: 'interrupted before the turn ended' }
+      }
+    }
+  }
+}
+
+/**
+ * Apply an id-keyed update to every row that registered the id, mirroring the
+ * copy-on-write reducer's full-array map semantics (all matching rows update,
+ * in order). Each registered index is O(1), so a duplicate id costs
+ * O(#duplicates) — never a full-array scan. The kind+id re-check is defensive:
+ * registered indices are valid by construction, because tool/command/retry
+ * rows are never removed and tombstones never shift indices.
+ */
+function updateReplayById<T extends TranscriptEntry>(
+  acc: ReplayAccumulator,
+  map: Map<string, number[]>,
+  id: string,
+  isMatch: (entry: T) => boolean,
+  update: (entry: T) => T,
+): void {
+  const list = map.get(id)
+  if (list === undefined) return // miss provably means no matching row
+  for (const index of list) {
+    const entry = acc.entries[index]
+    if (entry === undefined || !isMatch(entry as T)) continue
+    acc.entries[index] = update(entry as T)
+    acc.ops += 1
+  }
+}
+
+/** Tombstone a retired pending row, keeping every other index stable. */
+function retireReplayEntry(acc: ReplayAccumulator, index: number): void {
+  if (acc.entries[index] !== undefined) {
+    acc.entries[index] = undefined
+    acc.removedCount += 1
+    acc.ops += 1
+  }
+}
+
+/**
+ * Fold one session event into a replay accumulator. This mirrors
+ * {@link projectEvent} case for case — same stats arithmetic, same anchor
+ * set/delete behavior, same entry shapes — so the finished view is identical
+ * to a sequential fold; only the `entries` container operations are mutable.
+ *
+ * @internal Test-instrumentation path; `projectEvents` is the public entry.
+ * @returns whether the event changed the accumulated state — the live store
+ * stays silent and keeps its snapshot identity for ignored events, exactly
+ * like the copy-on-write reducer returning its input view unchanged.
+ */
+export function replayProjectEvent(acc: ReplayAccumulator, event: SessionEvent): boolean {
+  // Mirror of the reducer's entry shadow: any surface replace retires the
+  // system nodes it covers before the per-event fold runs.
+  const shadow = retireShadowedSystemNodes(acc.systemNodes, event.surfaceOp)
+  if (shadow.changed) {
+    acc.systemPrompt = shadow.prompt
+    acc.stats = { ...acc.stats, contextSegments: { ...acc.stats.contextSegments, system: estimateTokens(shadow.prompt) } }
+  }
+  switch (event.type) {
+    case 'user/message': {
+      const message = event.data
+      for (const target of ['next-turn', 'next-step'] as const) {
+        const ids = target === 'next-turn' ? acc.pendingTurn : acc.pendingStep
+        const index = ids.indexOf(message.id)
+        acc.ops += index < 0 ? ids.length : index + 1
+        if (index < 0) continue
+        ids.splice(index, 1)
+        acc.ops += 1
+        // Retire every pending row carrying this message id (duplicate ids
+        // included), exactly like the reducer's full-array filter.
+        const list = acc.pendingIndex.get(message.id)
+        if (list !== undefined) {
+          for (const entryIndex of list) retireReplayEntry(acc, entryIndex)
+          acc.ops += 1
+        }
+      }
+      const text = textOf(message.content)
+      const images = imagesOf(message.content)
+      const files = filesOf(message.content)
+      if (message.source.kind === 'user' || (message.source.kind === 'plugin' && REMINDER_PLUGINS.has(message.source.plugin))) {
+        appendReplayEntry(acc, { kind: 'user', text, notice: false, ...(images.length === 0 ? {} : { images }), ...(files.length === 0 ? {} : { files }) })
+        acc.stats = {
+          ...acc.stats,
+          contextSegments: {
+            ...acc.stats.contextSegments,
+            prompt: acc.stats.contextSegments.prompt + estimateTokens(text),
+          },
+        }
+        return true
+      }
+      if (message.source.kind === 'plugin' && HIDDEN_SNAPSHOT_PLUGINS.has(message.source.plugin)) {
+        acc.stats = {
+          ...acc.stats,
+          contextSegments: {
+            ...acc.stats.contextSegments,
+            system: acc.stats.contextSegments.system + estimateTokens(text),
+          },
+        }
+        return true
+      }
+      const notice = message.source.kind === 'plugin' && message.source.form === 'notice'
+        ? message.source.summary
+        : message.source.kind === 'plugin'
+          ? message.source.plugin
+          : message.source.kind
+      const summary = boundContextSummary(notice)
+      appendReplayEntry(acc, { kind: 'user', text: summary, notice: true })
+      acc.stats = {
+        ...acc.stats,
+        contextSegments: {
+          ...acc.stats.contextSegments,
+          system: acc.stats.contextSegments.system + estimateTokens(summary),
+        },
+      }
+      return true
+    }
+    case 'agent/inbox/spliced': {
+      const { target, start, removedCount = 0, inserted } = event.data
+      const ids = target === 'next-turn' ? acc.pendingTurn : acc.pendingStep
+      const removed = ids.slice(start, start + removedCount)
+      acc.ops += removed.length
+      ids.splice(start, removedCount)
+      acc.ops += removed.length
+      for (const id of removed) {
+        const list = acc.pendingIndex.get(id)
+        if (list === undefined) continue
+        for (const entryIndex of list) {
+          const entry = acc.entries[entryIndex]
+          if (entry !== undefined && entry.kind === 'pending' && entry.target === target) {
+            retireReplayEntry(acc, entryIndex)
+          }
+        }
+      }
+      // Mirror the reducer's in-place order: inserted ids join at the splice
+      // position (upstream `splice(start, removedCount, ...inserted)`), never
+      // at the tail — the id list must stay coordinate-compatible with every
+      // later inbox event.
+      ids.splice(start, 0, ...inserted.map(message => message.id))
+      for (const message of inserted) {
+        const images = imagesOf(message.content)
+        const files = filesOf(message.content)
+        appendReplayEntry(acc, { kind: 'pending', messageId: message.id, target, text: pendingText(message.content), ...(images.length === 0 ? {} : { images }), ...(files.length === 0 ? {} : { files }) })
+        indexList(acc.pendingIndex, message.id).push(acc.entries.length - 1)
+        acc.ops += 1
+      }
+      return true
+    }
+    case 'system/message': {
+      // Mirrors the reducer's per-node fold: append adds, replace retires the
+      // covered seq range, and the estimate prices the assembled prompt.
+      const text = textOf(event.data.message.content)
+      retireShadowedSystemNodes(acc.systemNodes, event.surfaceOp)
+      acc.systemNodes.set(event.seq, text)
+      acc.systemPrompt = assembleSystemPrompt(acc.systemNodes)
+      acc.stats = {
+        ...acc.stats,
+        contextSegments: {
+          ...acc.stats.contextSegments,
+          system: estimateTokens(acc.systemPrompt),
+        },
+      }
+      return true
+    }
+    case 'assistant/attempt': {
+      // Mirrors the reducer: the step stays open across in-step retries; a
+      // missing first-token anchor is restored (and accrued) from the
+      // attempt's embedded stream, and only the live tails are dropped.
+      const key = `${event.data.turn}:${event.data.step}`
+      let changed = false
+      if (!acc.firstChunkAt.has(key)) {
+        const first = assistantStreamFirstTokenTime(event.data.stream ?? [])
+        if (first !== undefined) {
+          acc.firstChunkAt.set(key, first)
+          const started = acc.stepStart.get(key)
+          if (started !== undefined) {
+            acc.stats = {
+              ...acc.stats,
+              ttftMs: acc.stats.ttftMs + Math.max(0, first - started),
+              ttftSteps: acc.stats.ttftSteps + 1,
+            }
+          }
+          changed = true
+        }
+      }
+      const streamed = acc.streaming !== '' || acc.streamingReasoning !== ''
+      acc.streaming = ''
+      acc.streamingReasoning = ''
+      return changed || streamed
+    }
+    case 'assistant/message': {
+      const key = `${event.data.turn}:${event.data.step}`
+      const started = acc.stepStart.get(key)
+      acc.stepStart.delete(key)
+      let firstChunk = acc.firstChunkAt.get(key)
+      if (firstChunk === undefined) {
+        // A replayed settlement has no live frames; the embedded stream's
+        // first-token time restores both the anchor and the TTFT figures the
+        // live path accumulates in applyAssistantStreamChunk.
+        firstChunk = assistantStreamFirstTokenTime(event.data.stream ?? [])
+        if (firstChunk !== undefined && started !== undefined) {
+          acc.stats = {
+            ...acc.stats,
+            ttftMs: acc.stats.ttftMs + Math.max(0, firstChunk - started),
+            ttftSteps: acc.stats.ttftSteps + 1,
+          }
+        }
+      }
+      acc.firstChunkAt.delete(key)
+      if (acc.turnSteps.get(event.data.turn) === key) acc.turnSteps.delete(event.data.turn)
+      const usage = event.data.usage
+      const totals = acc.stats.usage
+      const text = textOf(event.data.message.content)
+      const reasoning = reasoningOf(event.data.message.content)
+      acc.streaming = ''
+      acc.streamingReasoning = ''
+      appendReplayEntry(acc, { kind: 'assistant', text, reasoning, interrupted: event.data.interrupted === true ? true : undefined })
+      acc.stats = {
+        ...acc.stats,
+        llmMs: acc.stats.llmMs + (started === undefined ? 0 : Math.max(0, event.time - started)),
+        usage: usage === undefined ? totals : {
+          inputTokens: totals.inputTokens + usage.inputTokens + (usage.cacheReadTokens ?? 0) + (usage.cacheWriteTokens ?? 0),
+          outputTokens: totals.outputTokens + usage.outputTokens,
+          cacheReadTokens: totals.cacheReadTokens + (usage.cacheReadTokens ?? 0),
+        },
+        lastPromptTokens: usage === undefined ? acc.stats.lastPromptTokens
+          : usage.inputTokens + (usage.cacheReadTokens ?? 0) + (usage.cacheWriteTokens ?? 0),
+        decodeMs: acc.stats.decodeMs + (firstChunk === undefined ? 0 : Math.max(0, event.time - firstChunk)),
+        decodeTokens: acc.stats.decodeTokens + (firstChunk === undefined || usage === undefined ? 0 : usage.outputTokens),
+        contextSegments: {
+          ...acc.stats.contextSegments,
+          thinking: acc.stats.contextSegments.thinking + estimateTokens(reasoning),
+          assistant: acc.stats.contextSegments.assistant + estimateTokens(text),
+        },
+      }
+      return true
+    }
+    case 'tool/call': {
+      const data = event.data
+      acc.toolStart.set(data.callId, event.time)
+      const turnTools = acc.turnTools.get(data.turn) ?? new Set<string>()
+      turnTools.add(data.callId)
+      acc.turnTools.set(data.turn, turnTools)
+      acc.toolCallOrdinal += 1
+      appendReplayEntry(acc, {
+        kind: 'tool',
+        callId: data.callId,
+        ordinal: acc.toolCallOrdinal,
+        name: data.name,
+        arguments: data.arguments,
+        preview: toolArgumentsPreview(data.arguments, data.name),
+        prompt: toolPromptPreview(data.name, data.arguments),
+        state: 'running',
+        summary: '',
+        detail: undefined,
+      })
+      indexList(acc.toolIndex, data.callId).push(acc.entries.length - 1)
+      acc.stats = {
+        ...acc.stats,
+        contextSegments: {
+          ...acc.stats.contextSegments,
+          tools: acc.stats.contextSegments.tools
+            + (typeof data.arguments === 'string' ? estimateTokens(data.arguments) : 0),
+        },
+      }
+      return true
+    }
+    case 'tool/result': {
+      const block = event.data.message.content[0]
+      const started = acc.toolStart.get(block.toolCallId)
+      acc.toolStart.delete(block.toolCallId)
+      const turnTools = acc.turnTools.get(event.data.turn)
+      if (turnTools !== undefined) {
+        turnTools.delete(block.toolCallId)
+        if (turnTools.size === 0) acc.turnTools.delete(event.data.turn)
+      }
+      const rawText = textOf(block.content)
+      const summary = boundContextSummary(rawText)
+      const detail = toolResultDetail(event.data.meta, rawText)
+      if (detail?.kind === 'diff') {
+        const set = acc.turnFiles.get(event.data.turn) ?? new Set<string>()
+        for (const diff of detail.diffs) set.add(diff.path)
+        acc.turnFiles.set(event.data.turn, set)
+      }
+      const update = (entry: ToolEntry): ToolEntry => ({
+        ...entry,
+        state: block.isError === true ? 'error' as const : 'done' as const,
+        summary,
+        detail,
+      })
+      // Every matching row updates (duplicate callIds included); an id with no
+      // registered index is a provable no-op — no full-array fallback scan.
+      updateReplayById<ToolEntry>(acc, acc.toolIndex, block.toolCallId, entry => entry.callId === block.toolCallId, update)
+      acc.stats = {
+        ...acc.stats,
+        toolMs: acc.stats.toolMs + (started === undefined ? 0 : Math.max(0, event.time - started)),
+        contextSegments: {
+          ...acc.stats.contextSegments,
+          tools: acc.stats.contextSegments.tools + estimateTokens(rawText),
+        },
+      }
+      return true
+    }
+    case 'todo/write':
+      acc.todos = event.data.todos
+      return true
+    case 'turn/start': {
+      const wasBusy = acc.busy
+      acc.busy = true
+      acc.busySince = wasBusy ? acc.busySince : event.time
+      acc.todos = []
+      acc.stats = { ...acc.stats, turns: acc.stats.turns + 1 }
+      return true
+    }
+    case 'step/start': {
+      const key = `${event.data.turn}:${event.data.step}`
+      const previous = acc.turnSteps.get(event.data.turn)
+      if (previous !== undefined && previous !== key) {
+        acc.stepStart.delete(previous)
+        acc.firstChunkAt.delete(previous)
+      }
+      acc.turnSteps.set(event.data.turn, key)
+      acc.stepStart.set(key, event.time)
+      acc.streaming = ''
+      acc.streamingReasoning = ''
+      acc.stats = { ...acc.stats, steps: acc.stats.steps + 1 }
+      return true
+    }
+    case 'turn/end': {
+      const reason = event.data.reason
+      const appended: TranscriptEntry[] = []
+      acc.streamingReasoning = ''
+      acc.streaming = ''
+      if (reason.kind === 'error') {
+        const recovery = reason.error.code === 'MISSING_CREDENTIAL'
+          ? ' · open /model to add an API key'
+          : ''
+        appended.push({ kind: 'error', text: `${reason.error.code}: ${reason.error.message}${recovery}` })
+      } else {
+        const marker = reason.kind === 'aborted'
+          ? reason.reason.kind === 'user' ? 'turn cancelled by the user' : `turn cancelled (${reason.reason.kind})`
+          : reason.kind === 'max-tokens'
+            ? 'turn hit the output-token ceiling (max-tokens)'
+            : reason.kind === 'blocked'
+              ? 'turn ended blocked'
+              : reason.kind === 'interrupted'
+                ? 'turn was interrupted by a restart'
+                : undefined
+        if (marker !== undefined) appended.push({ kind: 'turn-marker', text: marker })
+      }
+      const files = acc.turnFiles.get(event.data.turn)
+      acc.turnFiles.delete(event.data.turn)
+      if (files !== undefined && files.size > 0) appended.push({ kind: 'files', paths: [...files].slice(0, 12) })
+      const stepKey = acc.turnSteps.get(event.data.turn)
+      if (stepKey !== undefined) {
+        acc.stepStart.delete(stepKey)
+        acc.firstChunkAt.delete(stepKey)
+        acc.turnSteps.delete(event.data.turn)
+      }
+      const turnToolSet = acc.turnTools.get(event.data.turn)
+      if (turnToolSet !== undefined) {
+        for (const callId of turnToolSet) acc.toolStart.delete(callId)
+        acc.turnTools.delete(event.data.turn)
+      }
+      finalizeReplayOrphans(acc)
+      acc.busy = false
+      acc.busySince = 0
+      for (const entry of appended) appendReplayEntry(acc, entry)
+      return true
+    }
+    case 'llm/retry': {
+      const data = event.data
+      acc.streaming = ''
+      acc.streamingReasoning = ''
+      appendReplayEntry(acc, {
+        kind: 'retry',
+        retryId: data.retryId,
+        mode: data.mode,
+        attempt: data.retry,
+        max: 'maxRetries' in data ? data.maxRetries : data.retry,
+        code: data.failure.code,
+        delayMs: data.delayMs,
+        state: 'running',
+      })
+      indexList(acc.retryIndex, data.retryId).push(acc.entries.length - 1)
+      return true
+    }
+    case 'llm/retry-started': {
+      const data = event.data
+      updateReplayById<RetryEntry>(acc, acc.retryIndex, data.retryId, entry => entry.retryId === data.retryId, entry => ({ ...entry, state: 'done' as const }))
+      return true
+    }
+    case 'sandbox/mode':
+      acc.sandbox = event.data.mode
+      return true
+    case 'goal/change': {
+      const data = event.data
+      const clip = (text: string): string => (text.length > 60 ? `${text.slice(0, 59)}…` : text)
+      if (data.operation === 'clear') {
+        acc.goal = undefined
+        appendReplayEntry(acc, { kind: 'turn-marker', text: '◎ goal cleared' })
+        return true
+      }
+      const goal: GoalFold = {
+        objective: data.goal.objective,
+        phase: data.goal.phase,
+        rounds: data.roundsStarted,
+        max: data.goal.maxGoalRounds,
+        blocked: data.goal.blockedReason?.message ?? '',
+      }
+      const line = data.operation === 'create'
+        ? `◎ goal: ${clip(data.goal.objective)}`
+        : data.operation === 'complete'
+          ? '◎ goal complete'
+          : data.operation === 'pause'
+            ? '◎ goal paused'
+            : data.operation === 'resume'
+              ? '◎ goal resumed'
+              : data.operation === 'block'
+                ? `◎ goal blocked: ${clip(goal.blocked)}`
+                : undefined
+      acc.goal = goal
+      if (line !== undefined) appendReplayEntry(acc, { kind: 'turn-marker', text: line })
+      return true
+    }
+    case 'schedule/change':
+      acc.schedules = applyScheduleChange(acc.schedules, event.data)
+      acc.ops += 1
+      return true
+    case 'session/title':
+      acc.title = event.data.title
+      return true
+    case 'compaction/summary':
+      if (acc.compactionTokens.size >= MAX_COMPACTION_SUMMARY_RESIDUE) {
+        const oldest = acc.compactionTokens.keys().next().value
+        if (oldest !== undefined) acc.compactionTokens.delete(oldest)
+      }
+      acc.compactionTokens.set(event.data.compactionId, event.data.shadowedTokenCount)
+      return true
+    case 'compaction/prune':
+      acc.lastPruneTokens = event.data.shadowedTokenCount
+      return true
+    case 'compaction/end': {
+      const ok = event.data.error === undefined
+      const tokens = acc.compactionTokens.get(event.data.compactionId) ?? acc.lastPruneTokens
+      acc.compactionTokens.delete(event.data.compactionId)
+      appendReplayEntry(acc, { kind: 'compaction', ok, tokens, error: event.data.error ?? '' })
+      return true
+    }
+    case 'request/context':
+      acc.stats = { ...acc.stats, contextWindow: event.data.contextWindow ?? acc.stats.contextWindow }
+      return true
+    case 'request/header': {
+      const config = event.data.header.config
+      acc.model = `${config.provider}/${config.model}`
+      acc.stats = {
+        ...acc.stats,
+        reasoningEffort: config.reasoningEffort === undefined ? '' : String(config.reasoningEffort),
+      }
+      return true
+    }
+    case 'plan/mode':
+      acc.plan = event.data.active
+      return true
+    case 'permission/preset':
+      acc.permission = event.data.preset
+      return true
+    case 'command/run': {
+      const data = event.data
+      appendReplayEntry(acc, {
+        kind: 'command',
+        commandId: data.commandId,
+        name: data.name,
+        args: data.args ?? '',
+        state: 'running',
+        summary: '',
+      })
+      indexList(acc.commandIndex, data.commandId).push(acc.entries.length - 1)
+      return true
+    }
+    case 'command/done': {
+      const data = event.data
+      const update = (candidate: CommandEntry): CommandEntry => ({
+        ...candidate,
+        state: data.kind === 'success' ? 'done' as const : 'error' as const,
+        summary: boundContextSummary(data.text ?? ''),
+      })
+      updateReplayById<CommandEntry>(acc, acc.commandIndex, data.commandId, entry => entry.commandId === data.commandId, update)
+      return true
+    }
+    default:
+      return false
+  }
+}
+
+/**
+ * Materialize the accumulated fold as a `TranscriptView`, compacting any
+ * retired tombstones. The anchors maps are handed through as-is (their
+ * content is identical to a sequential fold's).
+ *
+ * @internal Test-instrumentation path; `projectEvents` is the public entry.
+ */
+export function finishReplay(acc: ReplayAccumulator): TranscriptView {
+  return materializeReplayView(acc, false)
+}
+
+/**
+ * Materialize the accumulated fold as a fresh immutable snapshot for the
+ * live store. Unlike {@link finishReplay} — the one-shot replay entry, which
+ * hands the accumulator's own arrays through because the accumulator is
+ * discarded — every array a renderer can hold is copied here, so later
+ * folds never mutate a snapshot already handed out. Same fields, same
+ * tombstone compaction.
+ *
+ * @internal Live-store path; `projectEvents` is the public entry.
+ */
+export function snapshotReplayView(acc: ReplayAccumulator): TranscriptView {
+  return materializeReplayView(acc, true)
+}
+
+/** Field-for-field materialization; `copy` selects snapshot array isolation. */
+function materializeReplayView(acc: ReplayAccumulator, copy: boolean): TranscriptView {
+  const entries: readonly TranscriptEntry[] = acc.removedCount === 0
+    ? (copy ? [...acc.entries] : acc.entries) as TranscriptEntry[]
+    : acc.entries.filter((entry): entry is TranscriptEntry => entry !== undefined)
+  if (acc.removedCount > 0) acc.ops += acc.entries.length
+  return {
+    entries,
+    streaming: acc.streaming,
+    streamingReasoning: acc.streamingReasoning,
+    todos: acc.todos,
+    toolCallOrdinal: acc.toolCallOrdinal,
+    busy: acc.busy,
+    busySince: acc.busySince,
+    model: acc.model,
+    plan: acc.plan,
+    permission: acc.permission,
+    title: acc.title,
+    systemPrompt: acc.systemPrompt,
+    sandbox: acc.sandbox,
+    goal: acc.goal,
+    schedules: acc.schedules,
+    pending: { 'next-turn': [...acc.pendingTurn], 'next-step': [...acc.pendingStep] },
+    stats: acc.stats,
+    // Handed-out views get their own anchors snapshot: the accumulator keeps
+    // folding its live containers, and no consumer may observe that.
+    anchors: {
+      stepStart: new Map(acc.stepStart),
+      toolStart: new Map(acc.toolStart),
+      firstChunkAt: new Map(acc.firstChunkAt),
+      compactionTokens: new Map(acc.compactionTokens),
+      lastPruneTokens: acc.lastPruneTokens,
+      turnFiles: new Map([...acc.turnFiles].map(([turn, files]) => [turn, new Set(files)])),
+      turnSteps: new Map(acc.turnSteps),
+      turnTools: new Map([...acc.turnTools].map(([turn, tools]) => [turn, new Set(tools)])),
+      systemNodes: new Map(acc.systemNodes),
+    },
+  }
+}
+
+/**
+ * Fold a replayed event history into one view.
+ *
+ * Folding is near-linear in the log size: the mutable replay accumulator
+ * appends in place and resolves id-keyed updates through index maps, so a
+ * long persisted session replays without the O(N²) copy-on-write rebuilds a
+ * naive sequential fold would incur. The result is identical to folding
+ * {@link projectEvent} per event in order.
+ * @param events - events in `seq` order.
+ * @returns the folded view.
+ */
+export function projectEvents(events: readonly SessionEvent[]): TranscriptView {
+  const acc = createReplayAccumulator()
+  for (const event of events) replayProjectEvent(acc, event)
+  return finishReplay(acc)
+}
+
+/**
+ * Fold one process-local assistant-stream chunk frame (session-log v2+ keeps
+ * durable logs settlement-only; live typing rides the `agent/assistant-stream`
+ * agent event). Same first-token anchoring the durable `assistant/chunk` event
+ * used to carry: the first non-empty delta anchors the TTFT and empty
+ * keep-alive deltas do not count. The caller maps the frame's attempt to the
+ * `turn:step` key (the start frame owns turn/step; chunk frames do not).
+ * @param acc - the live replay accumulator.
+ * @param key - the `turn:step` key the attempt's start frame declared.
+ * @param time - the frame's safe-integer timestamp.
+ * @param chunk - the model chunk the frame carries.
+ * @returns whether the accumulator changed (the store stays silent otherwise).
+ */
+export function applyAssistantStreamChunk(acc: ReplayAccumulator, key: string, time: number, chunk: StreamChunk): boolean {
+  // First-token latency uses the kernel's isTokenDelta rule (a non-empty
+  // text, reasoning, or tool-call fragment counts; block/usage/finish chunks
+  // do not), so live frames and replayed embedded streams time the same
+  // token.
+  if (isTokenDelta(chunk) && !acc.firstChunkAt.has(key)) {
+    acc.firstChunkAt.set(key, time)
+    const started = acc.stepStart.get(key)
+    if (started !== undefined) {
+      acc.stats = {
+        ...acc.stats,
+        ttftMs: acc.stats.ttftMs + Math.max(0, time - started),
+        ttftSteps: acc.stats.ttftSteps + 1,
+      }
+    }
+  }
+  if (chunk.type === 'text-delta') {
+    acc.streaming = appendStreamingTail(acc.streaming, chunk.text)
+    return true
+  }
+  if (chunk.type === 'reasoning-delta') {
+    acc.streamingReasoning = appendStreamingTail(acc.streamingReasoning, chunk.text)
+    return true
+  }
+  return false
+}
+
+/**
+ * Drop the live streaming tails without a settlement (an `agent/assistant-stream`
+ * end frame with an `abandoned` outcome, or a session switch). The next start
+ * frame rebuilds from scratch.
+ * @param acc - the live replay accumulator.
+ * @returns whether any tail text was discarded.
+ */
+export function clearAssistantStream(acc: ReplayAccumulator): boolean {
+  const streamed = acc.streaming !== '' || acc.streamingReasoning !== ''
+  acc.streaming = ''
+  acc.streamingReasoning = ''
+  return streamed
+}
+
+/**
+ * The append-only flush boundary for a transcript view: the count of entries
+ * no later event can remove. Entries at or beyond this index are mutable and
+ * must stay in the live tree.
+ *
+ * `pending` rows are excluded even though they are not a running tool/retry:
+ * the inbox claims or cancels them durably (`agent/inbox/spliced` removals,
+ * `user/message` retirement), and an append-only `<Static>` flush cannot
+ * erase a row that vanishes from the view — the retired row would ghost on
+ * screen until the next source-backed replay. Running commands join the
+ * mutable boundary for the same reason in reverse: `command/done` mutates the
+ * row's state/summary, so a flushed row would keep its stale running mark
+ * until a resize-triggered replay. Everything else (including a completed
+ * tail) is final: later events only APPEND new rows.
+ * @param entries - the view's transcript entries in order.
+ * @returns the count of entries safe to flush (0 for an empty transcript).
+ */
+export function settledEntryCount(entries: readonly TranscriptEntry[]): number {
+  for (let index = 0; index < entries.length; index++) {
+    const entry = entries[index]
+    if (entry.kind === 'pending') return index
+    if (entry.kind === 'tool' && entry.state === 'running') return index
+    if (entry.kind === 'retry' && entry.state === 'running') return index
+    if (entry.kind === 'command' && entry.state === 'running') return index
+  }
+  return entries.length
+}