kaizen-agent 1.7.0 → 1.7.2

package/README.md

@@ -305,7 +305,7 @@ the OS can find — Spotlight on macOS, the app menu on Linux, a Desktop shortcu
 edit, a reject reason, a note on a run — has a mic. Press it, talk, press it again; the
 words land at the caret next to whatever you had typed. Recognition is Whisper running
 inside the page, so audio never leaves your machine. The first press downloads the model
-from huggingface.co (about 165 MB on a WebGPU browser, 80 MB on the CPU fallback) and
+from huggingface.co (about 165 MB on a WebGPU browser, up to 205 MB on a GPU without fp16, 80 MB on the CPU fallback) and
 caches it in the browser. The transformers.js library itself is fetched from jsDelivr the
 first time you press the mic in a session, so that first press needs the network; the
 model does not download again. Chromium and Edge qualify (Firefox falls back to the

package/cli/install.ts

@@ -221,6 +221,7 @@ if (wantWeb) {
     // The pid file only knows a --daemon; a board started from the shortcut or a
     // terminal has none, so also ask the OS what listens on the port.
     const recorded = existsSync(pidFile) ? Number(readFileSync(pidFile, "utf8")) : NaN;
+    const { pidOnPort } = await import("./state.ts");
     const listening = await pidOnPort(port ?? 7420);
     const pids = [...new Set([recorded, listening].filter((p): p is number => Number.isFinite(p) && p > 0))];
     if (!pids.length) { console.log("  nothing running on the kaizen web port"); process.exit(0); }
@@ -797,19 +798,6 @@ if (blocked) process.exit(1);
 
 // ---------------------------------------------------------------- helpers
 
-// Whoever holds the board's port, by the OS's own accounting.
-async function pidOnPort(port: number): Promise<number | null> {
-  try {
-    if (process.platform === "win32") {
-      const out = await Bun.$`netstat -ano -p tcp`.text();
-      const m = new RegExp(`127\\.0\\.0\\.1:${port}\\s+\\S+\\s+LISTENING\\s+(\\d+)`).exec(out);
-      return m ? Number(m[1]) : null;
-    }
-    const out = await Bun.$`lsof -t -iTCP:${port} -sTCP:LISTEN`.quiet().nothrow().text();
-    return Number(out.trim().split("\n")[0]) || null;
-  } catch { return null; }
-}
-
 function step(msg: string) { console.log(`  ${c.green("+")} ${msg}`); }
 
 // Runs the upgrade and reports what it actually did, version to version, so the

package/cli/state.ts

@@ -689,3 +689,16 @@ export function launchRun(it: Item, from: string): Launch {
     return { ok: false, agent, prompt, projectDir, manual, why: err?.message ?? String(err) };
   }
 }
+
+// Whoever holds the board's port, by the OS's own accounting.
+export async function pidOnPort(port: number): Promise<number | null> {
+  try {
+    if (process.platform === "win32") {
+      const out = await Bun.$`netstat -ano -p tcp`.text();
+      const m = new RegExp(`127\\.0\\.0\\.1:${port}\\s+\\S+\\s+LISTENING\\s+(\\d+)`).exec(out);
+      return m ? Number(m[1]) : null;
+    }
+    const out = await Bun.$`lsof -t -iTCP:${port} -sTCP:LISTEN`.quiet().nothrow().text();
+    return Number(out.trim().split("\n")[0]) || null;
+  } catch { return null; }
+}

package/cli/web.ts

@@ -5,7 +5,7 @@ import { join, dirname } from "node:path";
 import {
   home, type Item, boardCards, knownProjects, remember, locate, label, tilde,
   readInbox, writeInbox, replaceIdea, abandonRun, launchRun, parseItem, short, setArchived, writeNotes,
-  appendNote, saveAttachment,
+  appendNote, saveAttachment, pidOnPort,
 } from "./state.ts";
 
 const PAGE = join(dirname(import.meta.dir), "web", "board.html");
@@ -108,6 +108,13 @@ export async function web(repoDir: string, opts: Opts = {}) {
       const url = `http://127.0.0.1:${port}/`;
       const up = await fetch(url, { signal: AbortSignal.timeout(1500) }).then((r) => r.ok, () => false);
       if (up) { console.log(`\n  kaizen web already at ${url}\n`); if (opts.open !== false) await openApp(url); return; }
+      // A listener that does not answer is a board that hung; it is ours to stop.
+      const pid = await pidOnPort(port);
+      if (pid && pid !== process.pid) {
+        try { process.kill(pid, "SIGTERM"); console.log(`  stopped a hung board on ${port} (pid ${pid})`); } catch { /* not ours to kill */ }
+        await Bun.sleep(300);
+        try { server = serve(port); continue; } catch { /* still held */ }
+      }
       console.log(`  port ${port} refused by the OS and nothing answers there; trying ${port + 1}`);
     }
   }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "kaizen-agent",
-  "version": "1.7.0",
+  "version": "1.7.2",
   "description": "Staged workflow for AI agents — code, writing, research, operations: plan, human approval, build, independent review, bounded fix loop.",
   "type": "module",
   "bin": {

package/web/board.html

@@ -229,7 +229,12 @@ const TFJS="https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.3.0",ASR_MO
 // Inference lives in a module Worker so a slow chunk never freezes the board.
 const WORKER_SRC=`import{pipeline}from"${TFJS}";let asr=null;
 onmessage=async e=>{const m=e.data;try{
- if(m.type==="load"){asr=await pipeline("automatic-speech-recognition","${ASR_MODEL}",{device:m.webgpu?"webgpu":"wasm",dtype:m.webgpu?{encoder_model:"fp16",decoder_model_merged:"q4"}:"q8",progress_callback:p=>{if(p.status==="progress")postMessage({type:"progress",file:p.file,pct:p.progress||0})}});postMessage({type:"ready"})}
+ if(m.type==="load"){const prog=p=>{if(p.status==="progress")postMessage({type:"progress",file:p.file,pct:p.progress||0})};
+  // WebGPU first, with the encoder in fp16 only where the adapter has shader-f16; any GPU failure falls back to the CPU.
+  const gpu=m.webgpu?{device:"webgpu",dtype:{encoder_model:m.f16?"fp16":"fp32",decoder_model_merged:"q4"}}:null,cpu={device:"wasm",dtype:"q8"};
+  try{if(!gpu)throw 0;asr=await pipeline("automatic-speech-recognition","${ASR_MODEL}",{...gpu,progress_callback:prog})}
+  catch(e){if(gpu)postMessage({type:"fallback",why:String(e&&e.message||e)});asr=await pipeline("automatic-speech-recognition","${ASR_MODEL}",{...cpu,progress_callback:prog})}
+  postMessage({type:"ready"})}
  if(m.type==="run"){const r=await asr(m.audio,{chunk_length_s:30,return_timestamps:false});postMessage({type:"text",sid:m.sid,seq:m.seq,text:(r.text||"").trim()})}
 }catch(err){postMessage({type:"error",sid:m.sid,seq:m.seq,error:String(err&&err.message||err)})}}`;
 // The worklet only forwards 128-frame blocks; chunking and silence detection stay on the main thread.
@@ -238,11 +243,13 @@ const HAS_MIC=!!(navigator.mediaDevices&&navigator.mediaDevices.getUserMedia&&wi
 // `active` is the one session listening; `sessions` also holds stopped ones still waiting on a chunk.
 let asrWorker=null,asrReady=null,asrProgress=()=>{},active=null,starting=false;const sessions=new Set();
 // One worker for the page; loading is awaited by whoever pressed first, later presses join.
-function loadAsr(){if(asrReady)return asrReady;asrReady=new Promise((res,rej)=>{
+function loadAsr(){if(asrReady)return asrReady;asrReady=new Promise(async(res,rej)=>{
  asrWorker=new Worker(URL.createObjectURL(new Blob([WORKER_SRC],{type:"text/javascript"})),{type:"module"});
- const files={};asrWorker.onmessage=e=>{const m=e.data;if(m.type==="progress"){files[m.file]=m.pct;const v=Object.values(files);asrProgress(Math.round(v.reduce((a,b)=>a+b,0)/v.length))}else if(m.type==="ready")res();else if(m.type==="error"&&m.seq==null)rej(new Error(m.error));for(const s of sessions)s.onmsg(m)};
+ const files={};asrWorker.onmessage=e=>{const m=e.data;if(m.type==="progress"){files[m.file]=m.pct;const v=Object.values(files);asrProgress(Math.round(v.reduce((a,b)=>a+b,0)/v.length))}else if(m.type==="ready")res();else if(m.type==="fallback")toast("GPU could not run the model, using the CPU: "+m.why);else if(m.type==="error"&&m.seq==null)rej(new Error(m.error));for(const s of sessions)s.onmsg(m)};
  asrWorker.onerror=e=>rej(new Error(e.message||"worker failed"));
- const webgpu=!!navigator.gpu;if(!webgpu)toast("No WebGPU: dictation will run on the CPU and be slower");asrWorker.postMessage({type:"load",webgpu})});
+ // Ask the adapter what it can do: no adapter means the CPU path; no shader-f16 means an fp32 encoder.
+ let adapter=null;try{adapter=navigator.gpu&&await navigator.gpu.requestAdapter()}catch(e){}
+ if(!adapter)toast("No WebGPU: dictation will run on the CPU and be slower");asrWorker.postMessage({type:"load",webgpu:!!adapter,f16:!!(adapter&&adapter.features.has("shader-f16"))})});
  asrReady.catch(()=>{asrReady=null;try{asrWorker.terminate()}catch(e){}asrWorker=null});return asrReady}
 // After a re-render only a listener whose field is gone is stopped; closing the panel stops any.
 function stopDictation(all){if(active&&(all||!active.field.isConnected))active.stop()}