speechflow 2.4.0 → 2.4.2

package/speechflow-cli/src/speechflow-node-t2t-proofread.ts

@@ -119,12 +119,14 @@ export default class SpeechFlowNodeT2TProofread extends SpeechFlowNode {
             const cfg = this.setup[this.params.lang]
             if (!cfg)
                 throw new Error(`unsupported language: ${this.params.lang}`)
-            this.log("info", `PROMPT "${cfg.systemPrompt}", chat: ${JSON.stringify(cfg.chat)}`)
-            return llm.complete({
+            this.log("info", `input: "${text}"`)
+            const output = await llm.complete({
                 system:   cfg.systemPrompt,
                 messages: cfg.chat,
                 prompt:   text
             })
+            this.log("info", `output: "${output}"`)
+            return output
         }
 
         /*  establish a transform stream and connect it to LLM  */
@@ -141,14 +143,20 @@ export default class SpeechFlowNodeT2TProofread extends SpeechFlowNode {
                     callback()
                 }
                 else {
-                    proofread(chunk.payload).then((payload) => {
-                        const chunkNew = chunk.clone()
-                        chunkNew.payload = payload
-                        this.push(chunkNew)
+                    if (chunk.kind === "final") {
+                        proofread(chunk.payload).then((payload) => {
+                            const chunkNew = chunk.clone()
+                            chunkNew.payload = payload
+                            this.push(chunkNew)
+                            callback()
+                        }).catch((error: unknown) => {
+                            callback(util.ensureError(error))
+                        })
+                    }
+                    else {
+                        this.push(chunk)
                         callback()
-                    }).catch((error: unknown) => {
-                        callback(util.ensureError(error))
-                    })
+                    }
                 }
             },
             final (callback) {

package/speechflow-cli/src/speechflow-node-t2t-sentence.ts

@@ -324,12 +324,30 @@ export default class SpeechFlowNodeT2TSentence extends SpeechFlowNode {
                                 return
                             }
                             if (element.chunk.kind === "intermediate") {
-                                self.queueRecv.walk(-1)
-                                self.queueRecv.delete()
+                                /*  the trailing element is a (speculative) intermediate chunk:
+                                    if the newly arrived chunk carries the very same intermediate
+                                    payload (Deepgram re-sends identical intermediates), treat it
+                                    as a no-op to avoid churning the queue (delete+append) and
+                                    re-emitting an unchanged preview  */
+                                if (chunk.kind === "intermediate"
+                                    && (element.chunk.payload as string) === (chunk.payload as string)) {
+                                    self.lastChunkTime = Date.now()
+                                    callback()
+                                    return
+                                }
+
+                                /*  remove the trailing intermediate silently, so the
+                                    transient half-mutated queue (intermediate deleted, but
+                                    replacement not yet appended) does NOT emit a "write"
+                                    event and trigger a premature flush that would emit a
+                                    truncated preview and oscillate the dashboard preview  */
+                                self.queue.silently(() => {
+                                    self.queueRecv.walk(-1)
+                                    self.queueRecv.delete()
+                                })
                             }
                         }
                     }
-                    previewedPayload = ""
                     self.queueRecv.append({ type: "text-frame", chunk, complete: false })
                     self.lastChunkTime = Date.now()
                     callback()

package/speechflow-cli/src/speechflow-node-xio-exec.ts

@@ -93,7 +93,7 @@ export default class SpeechFlowNodeXIOExec extends SpeechFlowNode {
         })
 
         /*  handle subprocess errors  */
-        this.subprocess.on("error", (err) => {
+        this.subprocess.nodeChildProcess.on("error", (err: Error) => {
             this.log("error", `subprocess error: ${err.message}`)
             /*  NOTICE: do not emit("error") on the node itself, since nothing
                 listens for it and it would become an uncaughtException tearing
@@ -104,7 +104,7 @@ export default class SpeechFlowNodeXIOExec extends SpeechFlowNode {
         })
 
         /*  handle subprocess exit  */
-        this.subprocess.on("exit", (code, signal) => {
+        this.subprocess.nodeChildProcess.on("exit", (code: number | null, signal: string | null) => {
             if (code !== 0 && code !== null)
                 this.log("warning", `subprocess exited with code ${code}`)
             else if (signal)
@@ -126,10 +126,12 @@ export default class SpeechFlowNodeXIOExec extends SpeechFlowNode {
         /*  dispatch according to mode  */
         if (this.params.mode === "rw") {
             /*  bidirectional mode: both stdin and stdout  */
+            /*  NOTICE: the object form of Stream.Duplex.from() accepts Node streams at
+                run-time, but @types/node types it for Web streams only, so we have to cast  */
             this.stream = Stream.Duplex.from({
                 readable: this.subprocess.stdout,
                 writable: this.subprocess.stdin
-            })
+            } as unknown as Parameters<typeof Stream.Duplex.from>[0])
             const wrapper1 = util.createTransformStreamForWritableSide(this.params.type, highWaterMark)
             const wrapper2 = util.createTransformStreamForReadableSide(
                 this.params.type, () => this.timeZero, highWaterMark,
@@ -176,8 +178,8 @@ export default class SpeechFlowNodeXIOExec extends SpeechFlowNode {
             }
 
             /*  remove event listeners to prevent errors during kill sequence  */
-            this.subprocess.removeAllListeners("error")
-            this.subprocess.removeAllListeners("exit")
+            this.subprocess.nodeChildProcess.removeAllListeners("error")
+            this.subprocess.nodeChildProcess.removeAllListeners("exit")
 
             /*  wait for subprocess to exit gracefully  */
             const ac2 = new AbortController()

package/speechflow-cli/src/speechflow-util-llm.ts

@@ -61,7 +61,7 @@ export class LLM extends EventEmitter {
             api:           "",
             model:         "",
             key:           "",
-            timeout:       30 * 1000,
+            timeout:       240 * 1000,
             temperature:   0.7,
             maxTokens:     1024,
             cacheDir:      "",
@@ -333,10 +333,18 @@ export class LLM extends EventEmitter {
                 throw new Error(`failed to perform HuggingFace Transformers text generation: ${err}`, { cause: err })
             })
             const single = Array.isArray(result) ? result[0] : result
-            const generatedText = (single as Transformers.TextGenerationSingle).generated_text
-            const content = typeof generatedText === "string" ?
+            const generatedText = (single as Transformers.TextGenerationOutput[number]).generated_text
+            const messageContent = typeof generatedText === "string" ?
                 generatedText :
                 generatedText.at(-1)?.content
+
+            /*  flatten the potentially multi-modal message content into plain text  */
+            const content = typeof messageContent === "string" ?
+                messageContent :
+                messageContent
+                    ?.filter((part) => part.type === "text")
+                    .map((part) => part.text as string)
+                    .join("")
             if (!content)
                 throw new Error("HuggingFace Transformers API returned empty content")
             return content