fix(daemon): ship baseline OpenCode for Windows and stop retrying AVX2 illegal-instruction crashes (#5733)

* Classify AVX2 illegal-instruction agent crashes as cpu_unsupported and stop retrying them

The Windows opencode.exe bundled with the AMR runtime is a Bun build that
requires AVX2. On CPUs without it (Intel Atom/Celeron/Pentium N-series
through 2021) the binary dies with STATUS_ILLEGAL_INSTRUCTION before
readiness; vela surfaces an ACP fatal, the daemon classified it as a
retryable fatal_rpc_error, and the automatic retry deterministically
failed again. 88 of 105 pre-readiness Windows crash traces in Langfuse
are this single failure.

The crash text (Bun 'Illegal instruction' banner with a no_avx2 CPU
feature line, or the raw 0xC000001D / decimal 3221225501 exit status) now
classifies as process_exit/cpu_unsupported ahead of the fatal_rpc_error
close-reason promotion, is never auto-retried, and renders a dedicated
error card telling the user to update Open Design instead of a dead
Retry button.

* Bump @powerformer/vela-cli to 0.0.23 so packaged Windows builds bundle the baseline (non-AVX2) OpenCode

The 0.0.23 win32-x64 platform package compiles the OpenCode fork with
--target=windows-x64-baseline, so the bundled opencode.exe no longer
requires AVX2 and starts on Intel Atom/Celeron/Pentium N-series CPUs.
Verified in vela CI: the build log shows the baseline target and the new
post-publish smoke job boots the published bundle's opencode.exe on a
real Windows runner (powerformer/vela#838).

* Refresh Nix pnpm deps hashes for the vela-cli 0.0.23 lockfile change

* Add the abort-after-panic (exit status 3) production shape to cpu_unsupported specs

A user on an AVX-but-no-AVX2 CPU (Sandy/Ivy Bridge era) hit the same
illegal-instruction crash, but Bun panicked during its own panic handler
and abort()ed, so vela reported exit status 3 instead of
STATUS_ILLEGAL_INSTRUCTION. The stderr banner (no_avx2 / Illegal
instruction) is the only signal; encode the real Langfuse trace shape so
the classifier keeps catching it.

* Bump @powerformer/vela-cli to 0.0.25 so packaged Windows builds bundle the baseline (non-AVX2) OpenCode

0.0.25 is the first stable cut after powerformer/vela#838 merged: its
win32-x64 package compiles the OpenCode fork with
--target=windows-x64-baseline (verified in the release CI log), the
post-publish smoke boots the published bundle on a real Windows runner,
and the lockfile regenerator now pins the baseline asset so a routine
version bump cannot silently flip back. 0.0.24 predated that merge and
still shipped the AVX2 build.

* Tighten cpu_unsupported to precise signals only (no bare 'Illegal instruction')

Review is right that a bare illegal-instruction line is too broad: an
unrelated SIGILL on an AVX2-capable machine (runtime bug, corrupted
jump) would be mislabeled 'Processor not supported' and lose its retry.
Bun's crash banner always prints the CPU-feature line, so machines
actually missing AVX2 are identified by no_avx2; the bare Windows
STATUS_ILLEGAL_INSTRUCTION renderings (0xC000001D / 3221225501) stay
because those runs carry no banner at all. Generic illegal-instruction
banners without no_avx2 return to process_crashed / fatal_rpc_error,
locked by a new negative spec.

* Refresh Nix pnpm deps hashes for the vela-cli 0.0.25 lockfile change

* Gate raw STATUS_ILLEGAL_INSTRUCTION on the bundled-opencode startup context

Author: lefarcen

apps/daemon/src/run-failure-classification.ts

@@ -495,6 +495,32 @@ function isProcessCrashText(text: string): boolean {
     .test(text);
 }
 
+// The child binary executed an instruction this CPU does not implement — in
+// practice a Bun-compiled agent (bundled opencode) built for AVX2 running on a
+// CPU without it (Intel Atom/Celeron/Pentium N-series through 2021, and
+// AVX-but-not-AVX2 Sandy/Ivy Bridge cores). Matched only on signals that
+// prove the unsupported-CPU case:
+// - `no_avx2`: the CPU-feature line Bun's crash banner prints on such
+//   machines. Unconditional — the feature line itself is the proof.
+// - Windows STATUS_ILLEGAL_INSTRUCTION (hex 0xC000001D or Go/Node's decimal
+//   exit-status rendering 3221225501), but ONLY inside vela's bundled-opencode
+//   startup wrapper text ("start opencode server" / "opencode exited before
+//   readiness"). The raw status code is a generic Windows SIGILL that any
+//   agent binary could die with for unrelated reasons; every bannerless
+//   production trace carries the vela wrapper, so the gate costs no recall.
+// A bare "Illegal instruction" line is deliberately NOT matched: any
+// unrelated SIGILL (a runtime bug on an AVX2-capable machine) would then be
+// mislabeled as a processor limitation and lose its retry. The same binary on
+// the same CPU fails deterministically, so cpu_unsupported must never be
+// auto-retried.
+function isCpuUnsupportedCrashText(text: string): boolean {
+  if (/\bno_avx2\b/i.test(text)) return true;
+  return (
+    /0xc000001d|\b3221225501\b/i.test(text) &&
+    /\bstart opencode server\b|\bopencode exited before readiness\b/i.test(text)
+  );
+}
+
 // The daemon emits a `runtime_close` diagnostic into the run's event stream at
 // finalize time (see `deriveRpcCloseReason` in server.ts) carrying the mechanism
 // that ended the child as `rpc_close_reason`. When the agent-level error code is
@@ -869,6 +895,21 @@ export function classifyRunFailure(
     );
   }
 
+  // Must be checked BEFORE the fatal_rpc_error close-reason promotion below:
+  // when the bundled agent binary dies of an illegal instruction before
+  // readiness, vela surfaces an ACP fatal and the close reason alone would
+  // classify this as a retryable fatal_rpc_error — but the retry re-runs the
+  // same binary on the same CPU and deterministically fails again.
+  if (isCpuUnsupportedCrashText(text)) {
+    return classification(
+      'process_exit',
+      'cpu_unsupported',
+      inferFailureStageFromEvents(input.events, 'session_init'),
+      false,
+      'none',
+    );
+  }
+
   // ACP fatal paths ask the host to terminate the child after the protocol
   // failure. The resulting exit/signal is therefore cleanup, not the cause.
   // Prefer the runtime_close reason once specific text classifiers above have

apps/daemon/tests/run-failure-classification.test.ts

@@ -1141,6 +1141,10 @@ describe('classifyRunFailure — signal and interrupt attribution', () => {
       ),
     ).toMatchObject({
       failure_category: 'process_exit',
+      // A bare Bun illegal-instruction banner WITHOUT the no_avx2 CPU-feature
+      // line stays process_crashed: it may be an unrelated SIGILL on an
+      // AVX2-capable machine, so it must not claim the cpu_unsupported detail
+      // (which shows "Processor not supported" guidance).
       failure_detail: 'process_crashed',
       retryable: false,
       user_action: 'none',
@@ -1152,6 +1156,158 @@ function runtimeCloseEvent(reason: string): RunEventForFailureClassification {
   return { event: 'diagnostic', data: { type: 'runtime_close', rpc_close_reason: reason } };
 }
 
+describe('cpu_unsupported (AVX2) crash classification', () => {
+  // Windows AMR failure shape from Langfuse: the bundled opencode.exe is a Bun
+  // build requiring AVX2; on CPUs without it the child dies with an illegal
+  // instruction BEFORE readiness, vela surfaces an ACP fatal, and the daemon
+  // stamps runtime_close: fatal_rpc_error. The crash text must win over the
+  // fatal_rpc_error close-reason promotion — retrying the same binary on the
+  // same CPU deterministically fails again.
+  it('classifies a Bun illegal-instruction crash under an ACP fatal close as cpu_unsupported', () => {
+    const stderr = [
+      '============================================================',
+      'Bun v1.3.10 (30e609e0) Windows x64',
+      'CPU: sse42 popcnt no_avx no_avx2',
+      'panic(main thread): Illegal instruction',
+      'oh no: Bun has crashed. This indicates a bug in Bun, not your code.',
+    ].join('\n');
+    expect(
+      classify('AGENT_EXECUTION_FAILED', '', [
+        { event: 'stderr', data: { chunk: stderr } },
+        errorEvent('AGENT_EXECUTION_FAILED', ''),
+        runtimeCloseEvent('fatal_rpc_error'),
+      ]),
+    ).toMatchObject({
+      failure_category: 'process_exit',
+      failure_detail: 'cpu_unsupported',
+      retryable: false,
+      user_action: 'none',
+    });
+  });
+
+  it('classifies the abort-after-panic shape (exit status 3) via its stderr banner', () => {
+    // Production shape (Langfuse trace 266a5706, 0.15.0 stable): on a CPU with
+    // AVX but not AVX2 (Sandy/Ivy Bridge era), Bun panics on an illegal
+    // instruction, panics again during the panic, and abort()s — so the exit
+    // status vela reports is 3, not STATUS_ILLEGAL_INSTRUCTION. Only the
+    // stderr banner carries the truth.
+    const stderr = [
+      '============================================================',
+      'Bun v1.3.14 (0d9b296a) Windows x64',
+      'Windows v.win10_cu',
+      'CPU: sse42 avx',
+      'Args: ',
+      'Features: no_avx2 ',
+      '',
+      'panic: Illegal instruction at address 0x7FF6C08DF82C',
+      'panicked during a panic. Aborting.',
+    ].join('\n');
+    expect(
+      classify(
+        'AGENT_EXECUTION_FAILED',
+        'json-rpc id 2: start opencode server: opencode exited before readiness: exit status 3',
+        [
+          { event: 'stderr', data: { chunk: stderr } },
+          errorEvent(
+            'AGENT_EXECUTION_FAILED',
+            'json-rpc id 2: start opencode server: opencode exited before readiness: exit status 3',
+          ),
+          runtimeCloseEvent('fatal_rpc_error'),
+        ],
+      ),
+    ).toMatchObject({
+      failure_category: 'process_exit',
+      failure_detail: 'cpu_unsupported',
+      retryable: false,
+      user_action: 'none',
+    });
+  });
+
+  it('classifies a bare STATUS_ILLEGAL_INSTRUCTION exit under an ACP fatal close as cpu_unsupported', () => {
+    // No Bun crash banner — vela only reports the raw Windows exit status
+    // (0xC000001D, decimal 3221225501 in Go/Node exit-status text).
+    expect(
+      classify(
+        'AGENT_EXECUTION_FAILED',
+        'start opencode server: exit status 3221225501',
+        [
+          errorEvent('AGENT_EXECUTION_FAILED', 'start opencode server: exit status 3221225501'),
+          runtimeCloseEvent('fatal_rpc_error'),
+        ],
+      ),
+    ).toMatchObject({
+      failure_category: 'process_exit',
+      failure_detail: 'cpu_unsupported',
+      retryable: false,
+      user_action: 'none',
+    });
+  });
+
+  it('classifies the hex STATUS_ILLEGAL_INSTRUCTION form as cpu_unsupported', () => {
+    const message = 'start opencode server: opencode exited before readiness: exit status 0xC000001D';
+    expect(
+      classify('AGENT_EXECUTION_FAILED', message, [
+        errorEvent('AGENT_EXECUTION_FAILED', message),
+        runtimeCloseEvent('fatal_rpc_error'),
+      ]),
+    ).toMatchObject({
+      failure_detail: 'cpu_unsupported',
+      retryable: false,
+    });
+  });
+
+  it('keeps a STATUS_ILLEGAL_INSTRUCTION exit outside the opencode startup context retryable', () => {
+    // The raw status code is generic Windows SIGILL — any agent binary can die
+    // with it for reasons that have nothing to do with AVX2. Without vela's
+    // bundled-opencode startup wrapper text it must stay on the existing
+    // fatal_rpc_error path instead of surfacing the processor-support card.
+    const message = 'codex acp bridge exited: exit status 3221225501';
+    expect(
+      classify('AGENT_EXECUTION_FAILED', message, [
+        errorEvent('AGENT_EXECUTION_FAILED', message),
+        runtimeCloseEvent('fatal_rpc_error'),
+      ]),
+    ).toMatchObject({
+      failure_detail: 'fatal_rpc_error',
+      retryable: true,
+    });
+  });
+
+  it('does not claim an illegal-instruction crash without the no_avx2 feature line', () => {
+    // A SIGILL on an AVX2-capable machine (runtime bug, corrupted jump) prints
+    // the same "Illegal instruction" panic but a CPU-feature line WITHOUT
+    // no_avx2. That must keep the retryable fatal_rpc_error path — labeling it
+    // "Processor not supported" would mislead the user and drop the retry.
+    const stderr = [
+      'Bun v1.3.14 (0d9b296a) Windows x64',
+      'CPU: sse42 avx avx2',
+      'panic(main thread): Illegal instruction at address 0x7FF6C08DF82C',
+    ].join('\n');
+    expect(
+      classify('AGENT_EXECUTION_FAILED', '', [
+        { event: 'stderr', data: { chunk: stderr } },
+        errorEvent('AGENT_EXECUTION_FAILED', ''),
+        runtimeCloseEvent('fatal_rpc_error'),
+      ]),
+    ).toMatchObject({
+      failure_detail: 'fatal_rpc_error',
+      retryable: true,
+    });
+  });
+
+  it('keeps plain ACP fatal closes without crash text on fatal_rpc_error', () => {
+    expect(
+      classify('AGENT_EXECUTION_FAILED', '', [
+        errorEvent('AGENT_EXECUTION_FAILED', ''),
+        runtimeCloseEvent('fatal_rpc_error'),
+      ]),
+    ).toMatchObject({
+      failure_detail: 'fatal_rpc_error',
+      retryable: true,
+    });
+  });
+})
+
 describe('execution_failed close-reason refinement', () => {
   // A generic AGENT_EXECUTION_FAILED whose text matched no pattern, plus the
   // runtime_close diagnostic the daemon stamps at finalize time.

apps/daemon/tests/run-retry-policy.test.ts

@@ -300,6 +300,25 @@ describe('decideSafeRunRetry', () => {
     });
   });
 
+  it('never auto-retries a cpu_unsupported crash even when marked retryable', () => {
+    // An AVX2-requiring binary on a CPU without AVX2 crashes deterministically;
+    // the process_exit allowlist must keep cpu_unsupported out even if an
+    // upstream retryable hint leaks in as true.
+    expect(
+      decide({
+        failure: {
+          failure_category: 'process_exit',
+          failure_detail: 'cpu_unsupported',
+          failure_stage: 'session_init',
+          retryable: true,
+        },
+      }),
+    ).toMatchObject({
+      shouldRetry: false,
+      retrySuppressedReason: 'non_retryable_category',
+    });
+  });
+
   it('never auto-retries process kills, crashes, or interruptions', () => {
     for (const failure_detail of [
       'signal_killed',

apps/web/src/i18n/locales/ar.ts

@@ -78,13 +78,15 @@ export const ar: Dict = {
   'chat.runError.title.emptyOutput': "لا يوجد ناتج",
   'chat.runError.title.sessionExpired': "انتهت الجلسة",
   'chat.runError.title.gitBashMissing': "Git Bash غير موجود",
+  'chat.runError.title.cpuUnsupported': "المعالج غير مدعوم",
   'chat.runError.quotaExhaustedMessage': "انتهت حصة خدمة النموذج أو حد الفوترة، لذا لن تفيد إعادة المحاولة. اشحن رصيدك لدى المزوّد أو بدّل إلى نموذج أو خدمة أخرى.",
   'chat.runError.workspaceCreditsMessage': "نفدت أرصدة مساحة عملك. أضف أرصدة (أو اطلب من مالك المساحة إعادة الشحن)، أو بدّل إلى نموذج أو خدمة أخرى.",
   'chat.runError.timedOutMessage': "استغرق هذا التشغيل وقتًا طويلاً وتم إيقافه. أعد المحاولة، أو قلّص المهمة ثم أعد المحاولة.",
   'chat.runError.inactivityTimeoutMessage': "توقّف الوكيل عن إنتاج ناتج جديد لفترة طويلة فتم إيقافه بسبب انتهاء المهلة. عادةً ما تعيد المحاولة تشغيله من جديد.",
   'chat.runError.emptyOutputMessage': "انتهى الوكيل دون إنتاج أي ناتج. هذا مؤقت غالبًا، فأعد المحاولة.",
   'chat.runError.sessionExpiredMessage': "انتهت صلاحية الجلسة المستأنفة. تمت إعادة ضبطها، لذا أعد المحاولة لبدء تشغيل جديد.",
   'chat.runError.gitBashMissingMessage': "يلزم Git Bash لتشغيل هذا الوكيل على Windows لكنه غير موجود. ثبّت Git for Windows ثم أعد المحاولة.",
+  'chat.runError.cpuUnsupportedMessage': "تتطلب بيئة تشغيل هذا الوكيل مجموعة تعليمات للمعالج (AVX2) غير متوفرة في هذا الجهاز، لذا يتعذر تشغيلها. حدِّث Open Design إلى أحدث إصدار الذي يتضمن بيئة تشغيل متوافقة.",
   'common.cancel': 'إلغاء',
   'chat.selectFromLibrary': 'استيراد من المكتبة',
   'chat.importFigma': 'استيراد من Figma',

apps/web/src/i18n/locales/de.ts

@@ -78,13 +78,15 @@ export const de: Dict = {
   'chat.runError.title.emptyOutput': "Keine Ausgabe",
   'chat.runError.title.sessionExpired': "Sitzung abgelaufen",
   'chat.runError.title.gitBashMissing': "Git Bash fehlt",
+  'chat.runError.title.cpuUnsupported': "Prozessor nicht unterstützt",
   'chat.runError.quotaExhaustedMessage': "Das Kontingent oder Abrechnungslimit deines Modelldienstes ist aufgebraucht – ein erneuter Versuch hilft nicht. Lade bei deinem Anbieter auf oder wechsle zu einem anderen Modell oder Dienst.",
   'chat.runError.workspaceCreditsMessage': "Dein Workspace hat keine Credits mehr. Füge Credits hinzu (oder bitte den Workspace-Eigentümer aufzuladen) oder wechsle zu einem anderen Modell oder Dienst.",
   'chat.runError.timedOutMessage': "Dieser Lauf hat zu lange gedauert und wurde abgebrochen. Versuche es erneut oder verkleinere die Aufgabe und wiederhole.",
   'chat.runError.inactivityTimeoutMessage': "Der Agent hat zu lange keine neue Ausgabe geliefert und wurde als Zeitüberschreitung abgebrochen. Ein erneuter Versuch bringt ihn meist wieder in Gang.",
   'chat.runError.emptyOutputMessage': "Der Agent wurde beendet, ohne Ausgabe zu erzeugen. Das ist meist vorübergehend, versuche es erneut.",
   'chat.runError.sessionExpiredMessage': "Die fortgesetzte Sitzung war abgelaufen. Sie wurde zurückgesetzt, wiederhole, um einen neuen Lauf zu starten.",
   'chat.runError.gitBashMissingMessage': "Für diesen Agenten unter Windows wird Git Bash benötigt, es wurde aber nicht gefunden. Installiere Git für Windows und versuche es erneut.",
+  'chat.runError.cpuUnsupportedMessage': "Die Laufzeitumgebung dieses Agenten benötigt einen CPU-Befehlssatz (AVX2), den dieses Gerät nicht besitzt, und kann daher nicht starten. Aktualisiere Open Design auf die neueste Version, die eine kompatible Laufzeitumgebung mitliefert.",
   'common.cancel': 'Abbrechen',
   'chat.selectFromLibrary': 'Aus Bibliothek importieren',
   'chat.importFigma': 'Aus Figma importieren',

apps/web/src/i18n/locales/en.ts

@@ -78,13 +78,15 @@ export const en: Dict = {
   'chat.runError.title.emptyOutput': "No output produced",
   'chat.runError.title.sessionExpired': "Session expired",
   'chat.runError.title.gitBashMissing': "Git Bash missing",
+  'chat.runError.title.cpuUnsupported': "Processor not supported",
   'chat.runError.quotaExhaustedMessage': "Your model service's quota or billing limit is used up, so retrying won't help. Top up with your provider, or switch to another model or service.",
   'chat.runError.workspaceCreditsMessage': "Your workspace is out of credits. Add credits (or ask your workspace owner to refill), or switch to another model or service.",
   'chat.runError.timedOutMessage': "This run took too long and was stopped. Try again, or narrow the task and retry.",
   'chat.runError.inactivityTimeoutMessage': "The agent went quiet for too long and was stopped as a timeout. Retrying usually gets it moving again.",
   'chat.runError.emptyOutputMessage': "The agent finished without producing any output. This is usually temporary, so retry to run it again.",
   'chat.runError.sessionExpiredMessage': "The resumed session had expired. It has been reset, so retry to start a fresh run.",
   'chat.runError.gitBashMissingMessage': "Git Bash is required to run this agent on Windows but wasn't found. Install Git for Windows, then retry.",
+  'chat.runError.cpuUnsupportedMessage': "This agent's runtime needs a CPU instruction set (AVX2) that this device doesn't have, so it can't start. Update Open Design to the latest version, which ships a compatible runtime.",
   'common.cancel': 'Cancel',
   'chat.selectFromLibrary': 'Import from library',
   'chat.importFigma': 'Import from Figma',

apps/web/src/i18n/locales/es-ES.ts

@@ -78,13 +78,15 @@ export const esES: Dict = {
   'chat.runError.title.emptyOutput': "Sin salida",
   'chat.runError.title.sessionExpired': "Sesión caducada",
   'chat.runError.title.gitBashMissing': "Falta Git Bash",
+  'chat.runError.title.cpuUnsupported': "Procesador no compatible",
   'chat.runError.quotaExhaustedMessage': "La cuota o el límite de facturación de tu servicio de modelo se ha agotado, así que reintentar no servirá. Recarga con tu proveedor o cambia a otro modelo o servicio.",
   'chat.runError.workspaceCreditsMessage': "Tu espacio de trabajo se ha quedado sin créditos. Añade créditos (o pide al propietario del espacio que recargue) o cambia a otro modelo o servicio.",
   'chat.runError.timedOutMessage': "Esta ejecución tardó demasiado y se detuvo. Inténtalo de nuevo o reduce la tarea y reintenta.",
   'chat.runError.inactivityTimeoutMessage': "El agente estuvo demasiado tiempo sin nueva salida y se detuvo por tiempo agotado. Reintentar suele reanudarlo.",
   'chat.runError.emptyOutputMessage': "El agente terminó sin producir ninguna salida. Suele ser temporal: reintenta para ejecutarlo de nuevo.",
   'chat.runError.sessionExpiredMessage': "La sesión reanudada había caducado. Se ha restablecido, así que reintenta para iniciar una ejecución nueva.",
   'chat.runError.gitBashMissingMessage': "Se necesita Git Bash para ejecutar este agente en Windows, pero no se encontró. Instala Git para Windows y reintenta.",
+  'chat.runError.cpuUnsupportedMessage': "El entorno de ejecución de este agente necesita un conjunto de instrucciones de CPU (AVX2) que este dispositivo no tiene, por lo que no puede iniciarse. Actualiza Open Design a la última versión, que incluye un entorno de ejecución compatible.",
   'common.cancel': 'Cancelar',
   'chat.selectFromLibrary': 'Importar de la biblioteca',
   'chat.importFigma': 'Importar de Figma',