ummaya 0.2.4 → 0.2.6

package/tui/src/ink/reconciler.ts

@@ -34,9 +34,8 @@ if (process.env.NODE_ENV === 'development') {
   try {
     // eslint-disable-next-line custom-rules/no-top-level-dynamic-import -- dev-only; NODE_ENV check is DCE'd in production
     void import('./devtools.js')
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-  } catch (error: any) {
-    if (error.code === 'ERR_MODULE_NOT_FOUND') {
+  } catch (error: unknown) {
+    if (error instanceof Error && 'code' in error && error.code === 'ERR_MODULE_NOT_FOUND') {
       // biome-ignore lint/suspicious/noConsole: intentional warning
       console.warn(
         `

package/tui/src/ink/render-to-screen.ts

@@ -45,6 +45,31 @@ let output: Output | undefined
 const timing = { reconcile: 0, yoga: 0, paint: 0, scan: 0, calls: 0 }
 const LOG_EVERY = 20
 
+type SearchRendererContainer = ReturnType<typeof reconciler.createContainer>
+type SearchReconcilerRuntime = typeof reconciler & {
+  createContainer(
+    root: DOMElement,
+    tag: typeof LegacyRoot,
+    hydrationCallbacks: null,
+    isStrictMode: boolean,
+    concurrentUpdatesByDefaultOverride: null,
+    identifierPrefix: string,
+    onUncaughtError: () => void,
+    onCaughtError: () => void,
+    onRecoverableError: () => void,
+    onDefaultTransitionIndicator: () => void,
+  ): SearchRendererContainer
+  updateContainerSync(
+    element: ReactElement | null,
+    container: SearchRendererContainer,
+    parentComponent: null,
+    callback: () => void,
+  ): void
+  flushSyncWork(): void
+}
+
+const searchReconciler = reconciler as SearchReconcilerRuntime
+
 /** Render a React element (wrapped in all contexts the component needs —
  *  caller's job) to an isolated Screen buffer at the given width. Returns
  *  the Screen + natural height (from yoga). Used for search: render ONE
@@ -66,8 +91,7 @@ export function renderToScreen(
     stylePool = new StylePool()
     charPool = new CharPool()
     hyperlinkPool = new HyperlinkPool()
-    // @ts-expect-error react-reconciler 0.33 takes 10 args; @types says 11
-    container = reconciler.createContainer(
+    container = searchReconciler.createContainer(
       root,
       LegacyRoot,
       null,
@@ -82,10 +106,8 @@ export function renderToScreen(
   }
 
   const t0 = performance.now()
-  // @ts-expect-error updateContainerSync exists but not in @types
-  reconciler.updateContainerSync(el, container, null, noop)
-  // @ts-expect-error flushSyncWork exists but not in @types
-  reconciler.flushSyncWork()
+  searchReconciler.updateContainerSync(el, container, null, noop)
+  searchReconciler.flushSyncWork()
   const t1 = performance.now()
 
   // Yoga layout. Root might not have a yogaNode if the tree is empty.
@@ -117,10 +139,8 @@ export function renderToScreen(
   const t3 = performance.now()
 
   // Unmount so next call gets a fresh tree. Leaves root/container/pools.
-  // @ts-expect-error updateContainerSync exists but not in @types
-  reconciler.updateContainerSync(null, container, null, noop)
-  // @ts-expect-error flushSyncWork exists but not in @types
-  reconciler.flushSyncWork()
+  searchReconciler.updateContainerSync(null, container, null, noop)
+  searchReconciler.flushSyncWork()
 
   timing.reconcile += t1 - t0
   timing.yoga += t2 - t1

package/tui/src/ipc/bridge.ts

@@ -255,6 +255,14 @@ class AsyncQueue<T> {
           this._resolve = resolve
         })
       },
+      return: (): Promise<IteratorResult<T>> => {
+        if (this._resolve) {
+          const r = this._resolve
+          this._resolve = null
+          r({ value: undefined as unknown as T, done: true })
+        }
+        return Promise.resolve({ value: undefined as unknown as T, done: true })
+      },
     }
   }
 }
@@ -339,13 +347,19 @@ export function createBridge(opts: BridgeOptions = {}): IPCBridge {
   let _tuiSessionToken: string | null = opts.tuiSessionToken ?? null
   let _lastSeenCorrelationId: string | null = null
   let _lastSeenFrameSeq: number | null = null
-  // Bounded ring of applied (session_id, frame_seq) keys for replay-dedup.
-  // Cap mirrors the backend SessionRingBuffer size (Spec 032: 256 frames);
-  // oldest entries are evicted via FIFO replacement so the Set cannot grow
-  // unbounded across long sessions.
+  // Bounded rings for replay de-duplication.
+  //
+  // `applied_frame_seqs` remains the public debug surface expected by the
+  // resume tests. The actual skip decision uses a fuller frame identity below:
+  // some TUI-owned tool_call result paths are not ring-stamped by the backend
+  // yet and legally arrive with the default frame_seq=0. Dropping by
+  // session:frame_seq alone loses distinct tool_result frames in ordinary
+  // runtime, which is worse than replay duplication.
   const _APPLIED_FRAME_SEQS_CAP = 256
   const _appliedFrameSeqs = new Set<string>()
   const _appliedFrameSeqsOrder: string[] = []
+  const _appliedFrameIdentities = new Set<string>()
+  const _appliedFrameIdentitiesOrder: string[] = []
 
   // ------------------------------------------------------------------
   // Spawn first process
@@ -435,6 +449,40 @@ export function createBridge(opts: BridgeOptions = {}): IPCBridge {
     }
   }
 
+  function _frameReplayIdentity(frame: IPCFrame): string | null {
+    if (frame.frame_seq == null) return null
+    const sessionKey = frame.session_id ?? _sessionId
+    if (!sessionKey) return null
+    return [
+      sessionKey,
+      frame.frame_seq,
+      frame.kind,
+      frame.correlation_id ?? '',
+      frame.transaction_id ?? '',
+    ].join(':')
+  }
+
+  function _rememberAppliedFrameIdentity(identity: string): void {
+    _appliedFrameIdentities.add(identity)
+    _appliedFrameIdentitiesOrder.push(identity)
+    if (_appliedFrameIdentitiesOrder.length > _APPLIED_FRAME_SEQS_CAP) {
+      const evicted = _appliedFrameIdentitiesOrder.shift()
+      if (evicted !== undefined) _appliedFrameIdentities.delete(evicted)
+    }
+  }
+
+  function _rememberAppliedFrameSeq(frame: IPCFrame): void {
+    if (!_sessionId || frame.frame_seq == null) return
+    const key = `${frame.session_id ?? _sessionId}:${frame.frame_seq}`
+    if (_appliedFrameSeqs.has(key)) return
+    _appliedFrameSeqs.add(key)
+    _appliedFrameSeqsOrder.push(key)
+    if (_appliedFrameSeqsOrder.length > _APPLIED_FRAME_SEQS_CAP) {
+      const evicted = _appliedFrameSeqsOrder.shift()
+      if (evicted !== undefined) _appliedFrameSeqs.delete(evicted)
+    }
+  }
+
   async function _doReconnect(): Promise<void> {
     if (_closed || _reconnecting) return
     _reconnecting = true
@@ -536,22 +584,21 @@ export function createBridge(opts: BridgeOptions = {}): IPCBridge {
                 _lastSeenCorrelationId = frame.correlation_id
               }
 
-              // De-duplicate replayed frames (applied_frame_seqs set)
-              if (_sessionId && frame.frame_seq != null) {
-                const key = `${frame.session_id ?? _sessionId}:${frame.frame_seq}`
-                if (_appliedFrameSeqs.has(key)) {
+              // De-duplicate replayed frames by full frame identity. Do not
+              // use session:frame_seq as the skip key: TUI-owned tool_result
+              // frames can currently share the default frame_seq=0 while still
+              // being separate, valid results.
+              const replayIdentity = _frameReplayIdentity(frame)
+              if (replayIdentity !== null) {
+                if (_appliedFrameIdentities.has(replayIdentity)) {
                   _log(
                     'DEBUG',
-                    `Skipping duplicate replay frame session=${frame.session_id} frame_seq=${frame.frame_seq}`,
+                    `Skipping duplicate replay frame identity=${replayIdentity}`,
                   )
                   continue
                 }
-                _appliedFrameSeqs.add(key)
-                _appliedFrameSeqsOrder.push(key)
-                if (_appliedFrameSeqsOrder.length > _APPLIED_FRAME_SEQS_CAP) {
-                  const evicted = _appliedFrameSeqsOrder.shift()
-                  if (evicted !== undefined) _appliedFrameSeqs.delete(evicted)
-                }
+                _rememberAppliedFrameIdentity(replayIdentity)
+                _rememberAppliedFrameSeq(frame)
               }
 
               _log('DEBUG', `recv kind=${frame.kind} session=${frame.session_id}`)

package/tui/src/ipc/bridgeSingleton.ts

@@ -11,6 +11,7 @@
 // prompt is actually typed. Subsequent turns reuse the running child.
 
 import { createBridge, type IPCBridge } from './bridge.js'
+import { getSessionId } from '../bootstrap/state.js'
 import {
   ingestManifestFrame,
   waitForManifestSync,
@@ -36,7 +37,9 @@ let _pluginsModifiedThisSession = false
 
 export function getOrCreateUmmayaBridge(): IPCBridge {
   if (_bridge !== null) return _bridge
+  const sessionId = getUmmayaBridgeSessionId()
   _bridge = createBridge({
+    sessionId,
     // Epic ε #2296 T010 — route adapter_manifest_sync frames to the TS-side
     // manifest cache on receipt (before any LLM turn, at backend boot time).
     // FR-015: fires once per backend lifecycle; FR-016: replace-on-frame
@@ -69,7 +72,7 @@ export async function ensureUmmayaAdapterManifest(
 
 export function getUmmayaBridgeSessionId(): string {
   if (_sessionId === null) {
-    _sessionId = crypto.randomUUID()
+    _sessionId = getSessionId()
   }
   return _sessionId
 }
@@ -104,4 +107,5 @@ export async function closeUmmayaBridge(): Promise<void> {
     await b.close()
   }
   _pluginsModifiedThisSession = false
+  _sessionId = null
 }

package/tui/src/ipc/codec.ts

@@ -158,7 +158,7 @@ const ToolCallFrameSchema = BaseFrame.extend({
 })
 
 const ToolResultEnvelopeSchema = z.object({
-  kind: z.enum(['find', 'locate', 'send', 'check']),
+  kind: z.enum(['find', 'locate', 'send', 'check', 'document']),
 }).passthrough()
 
 const ToolResultFrameSchema = BaseFrame.extend({
@@ -176,7 +176,7 @@ const WorkerStatusFrameSchema = BaseFrame.extend({
   kind: z.literal('worker_status'),
   worker_id: z.string(),
   role_id: z.string(),
-  current_primitive: z.enum(['find', 'locate', 'send', 'check']),
+  current_primitive: z.enum(['find', 'locate', 'send', 'check', 'document']),
   status: z.enum(['idle', 'running', 'waiting_permission', 'error']),
 })
 
@@ -184,7 +184,7 @@ const PermissionRequestFrameSchema = BaseFrame.extend({
   kind: z.literal('permission_request'),
   request_id: z.string(),
   worker_id: z.string(),
-  primitive_kind: z.enum(['find', 'locate', 'send', 'check']),
+  primitive_kind: z.enum(['find', 'locate', 'send', 'check', 'document']),
   description_ko: z.string(),
   description_en: z.string(),
   risk_level: z.enum(['low', 'medium', 'high']),

package/tui/src/ipc/frames.generated.ts

@@ -381,7 +381,7 @@ export type CallId2 = string;
 /**
  * Primitive kind discriminator per Spec 031.
  */
-export type Kind6 = 'find' | 'locate' | 'send' | 'check';
+export type Kind6 = 'find' | 'locate' | 'send' | 'check' | 'document';
 /**
  * Opaque session identifier.
  */
@@ -461,7 +461,7 @@ export type RoleId = string;
 /**
  * Primitive currently being invoked by this worker.
  */
-export type CurrentPrimitive = 'find' | 'locate' | 'send' | 'check';
+export type CurrentPrimitive = 'find' | 'locate' | 'send' | 'check' | 'document';
 /**
  * Worker execution status.
  */
@@ -509,7 +509,7 @@ export type WorkerId1 = string;
 /**
  * The primitive the worker wants to invoke.
  */
-export type PrimitiveKind = 'find' | 'locate' | 'send' | 'check';
+export type PrimitiveKind = 'find' | 'locate' | 'send' | 'check' | 'document';
 /**
  * Korean-language description shown to the citizen.
  */
@@ -573,7 +573,7 @@ export type ReceiptId = string | null;
 /**
  * The primitive that was authorised. The TUI feeds this into `aalToLayer(primitive, isIrreversible)` to recompute the gauntlet layer (1=green / 2=orange / 3=red) for the receipt row. None on deny / timeout / legacy backends.
  */
-export type PrimitiveKind1 = ('find' | 'locate' | 'send' | 'check') | null;
+export type PrimitiveKind1 = ('find' | 'locate' | 'send' | 'check' | 'document') | null;
 /**
  * The fully-qualified adapter id (e.g. `mock_verify_mobile_id`, `mock_submit_welfare_grant`) the citizen authorised. The TUI uses this to render the human-readable Korean adapter name in /consent list and the modal title. None for non-adapter primitives (rare) or legacy backends.
  */
@@ -1200,7 +1200,7 @@ export type Name5 = string;
 /**
  * Primitive verb the adapter is registered under (I6).
  */
-export type Primitive = 'find' | 'send' | 'check' | 'locate';
+export type Primitive = 'find' | 'send' | 'check' | 'locate' | 'document';
 /**
  * Agency-published policy URL (HTTPS). None only when source_mode == 'internal' (I4/I5).
  */
@@ -1442,7 +1442,7 @@ export interface ToolDefinitionFunction {
  * JSON Schema (Draft 2020-12) for the tool input. Pydantic accepts any dict shape; deeper schema validation is delegated to LLMClient.
  */
 export interface Parameters {
-  [k: string]: any;
+  [k: string]: unknown;
 }
 /**
  * backend -> TUI: streaming assistant text delta.
@@ -1527,7 +1527,7 @@ export interface ToolCallFrame {
  * Tool-specific arguments. Concrete adapter calls carry adapter schema fields directly.
  */
 export interface Arguments1 {
-  [k: string]: any;
+  [k: string]: unknown;
 }
 /**
  * backend -> TUI (render): the output of a tool invocation.
@@ -1553,7 +1553,7 @@ export interface ToolResultFrame {
  */
 export interface ToolResultEnvelope {
   kind: Kind6;
-  [k: string]: any;
+  [k: string]: unknown;
 }
 /**
  * backend -> TUI: Spec 027 coordinator phase update.
@@ -1663,7 +1663,7 @@ export interface SessionEventFrame {
  * Event-specific payload. For list: {sessions: [{id, created_at, turn_count}]}. For resume: {id: str}. For others: {}.
  */
 export interface Payload {
-  [k: string]: any;
+  [k: string]: unknown;
 }
 /**
  * backend -> TUI: a backend error surfaced to the TUI for rendering.
@@ -1689,7 +1689,7 @@ export interface ErrorFrame {
  * Structured error details. UMMAYA_* env var values MUST be redacted.
  */
 export interface Details {
-  [k: string]: any;
+  [k: string]: unknown;
 }
 /**
  * Begins a streamed payload (assistant output, tool result chunking).
@@ -2020,13 +2020,13 @@ export interface AdapterManifestEntry {
  * Full backend Pydantic JSON Schema for adapter input parameters. Credential-only fields such as authKey must be omitted by adapters.
  */
 export interface InputSchemaJson {
-  [k: string]: any;
+  [k: string]: unknown;
 }
 /**
  * Full backend Pydantic JSON Schema for adapter output data.
  */
 export interface OutputSchemaJson {
-  [k: string]: any;
+  [k: string]: unknown;
 }
 /**
  * TUI -> backend: citizen requests revocation of a prior consent receipt.

package/tui/src/ipc/llmClient.ts

@@ -29,8 +29,13 @@ import type {
   UmmayaContentBlockParam,
   UmmayaStopReason,
 } from './llmTypes.js'
+import {
+  extractTextualToolCallProposals,
+  parseTrailingRawJsonToolCallProposal,
+} from '../utils/rawJsonToolCall.js'
 
 export const UMMAYA_DEFAULT_MODEL = 'LGAI-EXAONE/K-EXAONE-236B-A23B'
+const TEXTUAL_TOOL_CALL_OPEN = '<tool_call>'
 
 export type LLMClientErrorClass = 'llm' | 'tool' | 'network'
 
@@ -78,6 +83,9 @@ const _tracer = trace.getTracer('ummaya.tui.llm', '0.1.0')
 interface _TurnAccumulator {
   messageId: string | null
   contentBlocks: UmmayaContentBlockParam[]
+  accumulatedText: string
+  emittedTextLength: number
+  shouldBufferTextDeltas: boolean
   usage: UmmayaUsage
   stopReason: UmmayaStopReason
   /** Index of the current text block. It is opened lazily on the first
@@ -105,6 +113,9 @@ function _defaultAccumulator(): _TurnAccumulator {
   return {
     messageId: null,
     contentBlocks: [],
+    accumulatedText: '',
+    emittedTextLength: 0,
+    shouldBufferTextDeltas: false,
     usage: { input_tokens: 0, output_tokens: 0 },
     stopReason: 'end_turn',
     blockIndex: -1,
@@ -115,13 +126,21 @@ function _defaultAccumulator(): _TurnAccumulator {
   }
 }
 
+function firstToolCallTextOffset(text: string): number {
+  const braceOffset = text.indexOf('{')
+  const tagOffset = text.indexOf(TEXTUAL_TOOL_CALL_OPEN)
+  if (braceOffset < 0) return tagOffset
+  if (tagOffset < 0) return braceOffset
+  return Math.min(braceOffset, tagOffset)
+}
+
 // ---------------------------------------------------------------------------
 // Helper: extract usage from an AssistantChunkFrame done-trailer.
 //
 // Spec 032 does not currently define a typed usage payload on the frame itself;
 // the Python backend is expected to embed usage counts in the frame's `trailer`
 // extra fields or in a sibling ephemeral dict.  Until the backend-side contract
-// is finalised in Spec 032 US4, we read from `(frame as any).usage` first
+// is finalised in Spec 032 US4, we read from the typed legacy `usage` field first
 // (a forward-compatible extension field the backend may include) and fall back
 // to zeros.  A TODO(T091) marks the binding point for the actual backend wiring.
 // ---------------------------------------------------------------------------
@@ -417,6 +436,103 @@ export class LLMClient {
         } satisfies UmmayaRawMessageStreamEvent
       }
 
+      const appendVisibleTextDelta = function* (
+        text: string,
+        previousTextLength: number,
+      ): Generator<UmmayaRawMessageStreamEvent, void, unknown> {
+        if (text.length === 0) return
+        yield* ensureTextBlock()
+        yield {
+          type: 'content_block_delta',
+          index: acc.blockIndex,
+          delta: { type: 'text_delta', text },
+        } satisfies UmmayaRawMessageStreamEvent
+
+        const existing = acc.contentBlocks[acc.blockIndex]
+        if (existing && existing.type === 'text') {
+          existing.text += text
+        } else {
+          acc.contentBlocks[acc.blockIndex] = { type: 'text', text }
+        }
+        acc.emittedTextLength = previousTextLength + text.length
+      }
+
+      const appendAssistantText = function* (
+        text: string,
+      ): Generator<UmmayaRawMessageStreamEvent, void, unknown> {
+        if (text.length === 0) return
+        const previousTextLength = acc.accumulatedText.length
+        const bufferOffset = firstToolCallTextOffset(text)
+        const startsBuffering = !acc.shouldBufferTextDeltas && bufferOffset >= 0
+        if (startsBuffering) acc.shouldBufferTextDeltas = true
+        acc.accumulatedText += text
+        if (startsBuffering) {
+          if (bufferOffset > 0) {
+            yield* appendVisibleTextDelta(
+              text.slice(0, bufferOffset),
+              previousTextLength,
+            )
+          }
+          return
+        }
+        if (!acc.shouldBufferTextDeltas) {
+          yield* appendVisibleTextDelta(text, previousTextLength)
+        }
+      }
+
+      const flushBufferedText = function* (): Generator<
+        UmmayaRawMessageStreamEvent,
+        void,
+        unknown
+      > {
+        if (acc.accumulatedText.length <= acc.emittedTextLength) return
+        yield* appendVisibleTextDelta(
+          acc.accumulatedText.slice(acc.emittedTextLength),
+          acc.emittedTextLength,
+        )
+      }
+
+      const flushPreludeBeforeRawJsonToolUse = function* (
+        prelude: string,
+      ): Generator<UmmayaRawMessageStreamEvent, void, unknown> {
+        if (prelude.length <= acc.emittedTextLength) return
+        yield* appendVisibleTextDelta(
+          prelude.slice(acc.emittedTextLength),
+          acc.emittedTextLength,
+        )
+      }
+
+      const appendRawJsonToolUseBlock = function* (params: {
+        readonly name: string
+        readonly input: Record<string, unknown>
+        readonly id?: string
+      }): Generator<UmmayaRawMessageStreamEvent, void, unknown> {
+        const toolBlockIndex = acc.contentBlocks.length
+        acc.toolBlockCounter += 1
+        const toolUseId = params.id ?? 'call_raw_json_tool_0'
+        yield {
+          type: 'content_block_start',
+          index: toolBlockIndex,
+          content_block: {
+            type: 'tool_use',
+            id: toolUseId,
+            name: params.name,
+            input: params.input,
+          },
+        } satisfies UmmayaRawMessageStreamEvent
+        yield {
+          type: 'content_block_stop',
+          index: toolBlockIndex,
+        } satisfies UmmayaRawMessageStreamEvent
+        acc.contentBlocks[toolBlockIndex] = {
+          type: 'tool_use',
+          id: toolUseId,
+          name: params.name,
+          input: params.input,
+        }
+        acc.stopReason = 'tool_use'
+      }
+
       for await (const frame of this.bridge.frames()) {
         // Filter: only process frames for this turn's correlation_id.
         if (frame.correlation_id !== correlationId) continue
@@ -483,21 +599,7 @@ export class LLMClient {
 
             // Emit text delta (even if delta is empty — forward compat).
             if (chunk.delta.length > 0) {
-              yield* ensureTextBlock()
-              yield {
-                type: 'content_block_delta',
-                index: acc.blockIndex,
-                delta: { type: 'text_delta', text: chunk.delta },
-              } satisfies UmmayaRawMessageStreamEvent
-
-              // Accumulate text into the content block for the final object.
-              const existing = acc.contentBlocks[acc.blockIndex]
-              if (existing && existing.type === 'text') {
-                existing.text += chunk.delta
-              } else {
-                // First delta for this block index.
-                acc.contentBlocks[acc.blockIndex] = { type: 'text', text: chunk.delta }
-              }
+              yield* appendAssistantText(chunk.delta)
             }
           } else {
             // done=true: terminal chunk (V2 invariant satisfied here).
@@ -510,18 +612,37 @@ export class LLMClient {
 
             // Emit any final delta text if present.
             if (chunk.delta.length > 0) {
-              yield* ensureTextBlock()
-              yield {
-                type: 'content_block_delta',
-                index: acc.blockIndex,
-                delta: { type: 'text_delta', text: chunk.delta },
-              } satisfies UmmayaRawMessageStreamEvent
-              const existing = acc.contentBlocks[acc.blockIndex]
-              if (existing && existing.type === 'text') {
-                existing.text += chunk.delta
-              } else {
-                acc.contentBlocks[acc.blockIndex] = { type: 'text', text: chunk.delta }
+              yield* appendAssistantText(chunk.delta)
+            }
+
+            const textualToolUse = extractTextualToolCallProposals({
+              text: acc.accumulatedText,
+            })
+            const rawJsonToolUse = textualToolUse === undefined
+              ? parseTrailingRawJsonToolCallProposal({
+                  text: acc.accumulatedText,
+                })
+              : undefined
+            if (textualToolUse !== undefined || rawJsonToolUse !== undefined) {
+              const prelude = textualToolUse?.text ?? rawJsonToolUse?.prelude ?? ''
+              yield* flushPreludeBeforeRawJsonToolUse(prelude)
+              yield* closeOpenBlock()
+              const proposals = textualToolUse?.proposals ??
+                (rawJsonToolUse !== undefined ? [rawJsonToolUse.proposal] : [])
+              for (let index = 0; index < proposals.length; index += 1) {
+                const proposal = proposals[index]
+                if (!proposal) continue
+                yield* appendRawJsonToolUseBlock({
+                  name: proposal.name,
+                  input: proposal.input,
+                  id: textualToolUse !== undefined
+                    ? `call_textual_tool_${index}`
+                    : 'call_raw_json_tool_0',
+                })
               }
+              acc.accumulatedText = prelude
+            } else if (acc.shouldBufferTextDeltas) {
+              yield* flushBufferedText()
             }
 
             // Extract usage from the done-frame.
@@ -557,6 +678,9 @@ export class LLMClient {
         else if (frame.kind === 'tool_call') {
           const toolFrame = frame as ToolCallFrame
           yield* ensureMessageStart(`tool-${correlationId}`)
+          if (acc.shouldBufferTextDeltas) {
+            yield* flushBufferedText()
+          }
           yield* closeOpenBlock()
           // tool_call frames may arrive interleaved with text streaming AND
           // thinking streaming in a multi-turn / parallel-tool / K-EXAONE