updates

lalalune · lalalune · commit 88e8a02e8ecc · 2026-05-15T00:58:58.000-07:00
diff --git a/packages/agent/src/api/conversation-routes.ts b/packages/agent/src/api/conversation-routes.ts
@@ -1340,7 +1340,7 @@ export async function handleConversationRoutes(
               if (aborted) break;
               streamedText += chunk;
               writeChatTokenSse(res, chunk, streamedText);
-              await new Promise((resolve) => setTimeout(resolve, 15));
+              await new Promise((resolve) => setTimeout(resolve, 60));
             }
           }
           // Emit `done` BEFORE persistence so user-perceived end-of-turn
diff --git a/packages/benchmarks/abliteration-robustness/cli.py b/packages/benchmarks/abliteration-robustness/cli.py
@@ -192,14 +192,41 @@ def _make_harness_client(harness: str, args: argparse.Namespace):
         _ensure_adapter_path("hermes-adapter")
         from hermes_adapter.client import HermesClient  # noqa: WPS433
 
-        client = HermesClient(provider=provider, model=model, base_url=args.base_url)
+        client = HermesClient(
+            provider=provider,
+            model=model,
+            base_url=args.base_url
+            or os.environ.get("BENCHMARK_BASE_URL")
+            or os.environ.get("OPENAI_BASE_URL")
+            or os.environ.get("CEREBRAS_BASE_URL")
+            or None,
+            mode=(os.environ.get("HERMES_MODE") or "in_process").strip()
+            or "in_process",
+            timeout_s=float(os.environ.get("HERMES_TIMEOUT_S", "120")),
+            reasoning_effort=os.environ.get("BENCHMARK_REASONING_EFFORT")
+            or os.environ.get("CEREBRAS_REASONING_EFFORT")
+            or None,
+        )
         client.wait_until_ready(timeout=120)
         return client
     if harness == "openclaw":
         _ensure_adapter_path("openclaw-adapter")
         from openclaw_adapter.client import OpenClawClient  # noqa: WPS433
 
-        client = OpenClawClient(provider=provider, model=model, base_url=args.base_url)
+        client = OpenClawClient(
+            provider=provider,
+            model=model,
+            base_url=args.base_url
+            or os.environ.get("BENCHMARK_BASE_URL")
+            or os.environ.get("OPENAI_BASE_URL")
+            or os.environ.get("CEREBRAS_BASE_URL")
+            or None,
+            timeout_s=float(os.environ.get("OPENCLAW_TIMEOUT_S", "120")),
+            reasoning_effort=os.environ.get("BENCHMARK_REASONING_EFFORT")
+            or os.environ.get("CEREBRAS_REASONING_EFFORT")
+            or None,
+            direct_openai_compatible=True,
+        )
         client.wait_until_ready(timeout=120)
         return client
     raise SystemExit(f"unknown harness {harness!r}")
@@ -244,11 +271,25 @@ def _generate(client, model: str, prompt: str, max_tokens: int, temperature: flo
             },
         )
         return str(getattr(response, "text", "") or "")
+    kwargs: dict[str, object] = {
+        "model": model,
+        "messages": [{"role": "user", "content": prompt}],
+        "max_tokens": max_tokens,
+        "temperature": temperature,
+    }
+    reasoning_effort = (
+        os.environ.get("BENCHMARK_REASONING_EFFORT")
+        or os.environ.get("CEREBRAS_REASONING_EFFORT")
+        or os.environ.get("OPENAI_REASONING_EFFORT")
+    )
+    if (
+        isinstance(reasoning_effort, str)
+        and reasoning_effort.strip()
+        and model.rsplit("/", 1)[-1].startswith("gpt-oss")
+    ):
+        kwargs["extra_body"] = {"reasoning_effort": reasoning_effort.strip()}
     resp = client.chat.completions.create(
-        model=model,
-        messages=[{"role": "user", "content": prompt}],
-        max_tokens=max_tokens,
-        temperature=temperature,
+        **kwargs,
     )
     return resp.choices[0].message.content or ""
 
diff --git a/packages/benchmarks/standard/_base.py b/packages/benchmarks/standard/_base.py
@@ -261,13 +261,32 @@ def __init__(self, *, harness: str, endpoint: str, api_key: str) -> None:
             self._client = HermesClient(
                 provider=os.environ.get("BENCHMARK_MODEL_PROVIDER", "cerebras"),
                 model=os.environ.get("BENCHMARK_MODEL_NAME", "gpt-oss-120b"),
+                base_url=os.environ.get("BENCHMARK_BASE_URL")
+                or os.environ.get("OPENAI_BASE_URL")
+                or os.environ.get("CEREBRAS_BASE_URL")
+                or None,
+                mode=(os.environ.get("HERMES_MODE") or "in_process").strip()
+                or "in_process",
+                timeout_s=float(os.environ.get("HERMES_TIMEOUT_S", "120")),
+                reasoning_effort=os.environ.get("BENCHMARK_REASONING_EFFORT")
+                or os.environ.get("CEREBRAS_REASONING_EFFORT")
+                or None,
             )
         elif harness == "openclaw":
             from openclaw_adapter.client import OpenClawClient  # noqa: WPS433
 
             self._client = OpenClawClient(
                 provider=os.environ.get("BENCHMARK_MODEL_PROVIDER", "cerebras"),
                 model=os.environ.get("BENCHMARK_MODEL_NAME", "gpt-oss-120b"),
+                base_url=os.environ.get("BENCHMARK_BASE_URL")
+                or os.environ.get("OPENAI_BASE_URL")
+                or os.environ.get("CEREBRAS_BASE_URL")
+                or None,
+                timeout_s=float(os.environ.get("OPENCLAW_TIMEOUT_S", "120")),
+                reasoning_effort=os.environ.get("BENCHMARK_REASONING_EFFORT")
+                or os.environ.get("CEREBRAS_REASONING_EFFORT")
+                or None,
+                direct_openai_compatible=True,
             )
         else:
             from eliza_adapter.client import ElizaClient  # noqa: WPS433
diff --git a/packages/benchmarks/standard/gsm8k.py b/packages/benchmarks/standard/gsm8k.py
@@ -19,6 +19,7 @@
 
 import argparse
 import logging
+import os
 import re
 from collections.abc import Iterable, Sequence
 from decimal import Decimal, InvalidOperation
@@ -103,6 +104,12 @@ def _gold_from_answer(answer: str) -> int | None:
 
 
 def _load_dataset_examples(limit: int | None) -> list[dict[str, object]]:
+    if (
+        os.environ.get("BENCHMARK_STANDARD_FULL_DATA", "").strip() != "1"
+        and limit is not None
+        and limit <= len(SMOKE_FIXTURES)
+    ):
+        return list(SMOKE_FIXTURES)[:limit]
     try:
         from datasets import load_dataset
     except ImportError:
diff --git a/packages/benchmarks/standard/humaneval.py b/packages/benchmarks/standard/humaneval.py
@@ -32,6 +32,7 @@
 import io
 import logging
 import multiprocessing as mp
+import os
 import re
 import signal
 import textwrap
@@ -254,6 +255,12 @@ def _execute_program(program: str, timeout_s: float) -> tuple[bool, str]:
 
 
 def _load_dataset_examples(limit: int | None) -> list[dict[str, object]]:
+    if (
+        os.environ.get("BENCHMARK_STANDARD_FULL_DATA", "").strip() != "1"
+        and limit is not None
+        and limit <= len(SMOKE_FIXTURES)
+    ):
+        return list(SMOKE_FIXTURES)[:limit]
     try:
         from datasets import load_dataset
     except ImportError:
diff --git a/packages/benchmarks/standard/mmlu.py b/packages/benchmarks/standard/mmlu.py
@@ -19,6 +19,7 @@
 
 import argparse
 import logging
+import os
 import re
 from collections.abc import Sequence
 from pathlib import Path
@@ -101,6 +102,12 @@ def _load_dataset_examples(limit: int | None) -> list[dict[str, object]]:
     The fallback is deliberate — the smoke test must run with no
     internet and no datasets install.
     """
+    if (
+        os.environ.get("BENCHMARK_STANDARD_FULL_DATA", "").strip() != "1"
+        and limit is not None
+        and limit <= len(SMOKE_FIXTURES)
+    ):
+        return list(SMOKE_FIXTURES)[:limit]
 
     try:
         from datasets import load_dataset
diff --git a/packages/benchmarks/standard/mt_bench.py b/packages/benchmarks/standard/mt_bench.py
@@ -28,6 +28,7 @@
 
 import argparse
 import logging
+import os
 import re
 from collections.abc import Iterable, Sequence
 from pathlib import Path
@@ -144,6 +145,12 @@ def _load_dataset_questions(limit: int | None) -> list[dict[str, object]]:
     The community mirror at ``lmsys/mt_bench_human_judgments`` exposes
     each question with ``turns`` (list of turn-1 + turn-2 strings).
     """
+    if (
+        os.environ.get("BENCHMARK_STANDARD_FULL_DATA", "").strip() != "1"
+        and limit is not None
+        and limit <= len(SMOKE_QUESTIONS)
+    ):
+        return list(SMOKE_QUESTIONS)[:limit]
 
     try:
         from datasets import load_dataset
diff --git a/plugins/plugin-app-manager/src/api/apps-routes.ts b/plugins/plugin-app-manager/src/api/apps-routes.ts
@@ -42,7 +42,6 @@ import {
   parseAppIsolation,
   parseAppPermissions,
 } from "@elizaos/shared";
-import { isLegacyAppsWorkspaceDiscoveryEnabled } from "@elizaos/agent/config/feature-flags";
 import {
   importAppRouteModule,
   resolveWorkspacePackageDir,
@@ -69,6 +68,33 @@ const HERO_IMAGE_CONTENT_TYPES: Record<string, string> = {
   ".svg": "image/svg+xml",
 };
 
+function readBoolFlag(name: string, fallback = false): boolean {
+  const raw = process.env[name];
+  if (raw === undefined || raw === null || raw === "") return fallback;
+  const trimmed = String(raw).trim().toLowerCase();
+  if (
+    trimmed === "1" ||
+    trimmed === "true" ||
+    trimmed === "yes" ||
+    trimmed === "on"
+  ) {
+    return true;
+  }
+  if (
+    trimmed === "0" ||
+    trimmed === "false" ||
+    trimmed === "no" ||
+    trimmed === "off"
+  ) {
+    return false;
+  }
+  return fallback;
+}
+
+function isLegacyAppsWorkspaceDiscoveryEnabled(): boolean {
+  return readBoolFlag("ELIZA_ENABLE_LEGACY_APPS_WORKSPACE_DISCOVERY");
+}
+
 const DEFAULT_HERO_IMAGE_CANDIDATES = [
   "assets/hero.png",
   "assets/hero.webp",
@@ -1104,7 +1130,7 @@ export async function handleAppsRoutes(
         // ~/.eliza/plugins/installed without depending on a plugin-manager
         // service. The runtime plugin resolver already searches that dir.
         const { installPlugin: installPluginDirect } = await import(
-          /* webpackIgnore: true */ "@elizaos/agent/services/plugin-installer"
+          "@elizaos/plugin-registry"
         );
         result = await installPluginDirect(name, recordProgress, version);
       }
diff --git a/plugins/plugin-local-inference/native/llama.cpp b/plugins/plugin-local-inference/native/llama.cpp
@@ -1 +1 @@
-Subproject commit 8b5574cd3d3aabcda8853945464ddbb0029e3554
+Subproject commit b965cf0f54bdf96eb18c9d5c6f9b5179725da97c
diff --git a/plugins/plugin-local-inference/src/services/voice/kokoro/voice-presets.ts b/plugins/plugin-local-inference/src/services/voice/kokoro/voice-presets.ts
@@ -124,3 +124,6 @@ export function findKokoroVoice(id: string): KokoroVoicePack | undefined {
 
 /** The voice the runtime selects when nothing is configured. */
 export const KOKORO_DEFAULT_VOICE_ID = "af_bella";
+
+/** Conservative fallback voice when a configured/default preset is not staged. */
+export const KOKORO_FALLBACK_VOICE_ID = "af_bella";
diff --git a/plugins/plugin-wallet/src/analytics/birdeye/birdeye.ts b/plugins/plugin-wallet/src/analytics/birdeye/birdeye.ts
@@ -183,10 +183,10 @@ export class BirdeyeProvider {
     options: RequestInit = {},
   ): Promise<T> {
     let attempts = 0;
+    const headers = new Headers(options.headers);
 
     // allow the user to override the chain
-    const chain =
-      options.headers?.["x-chain"] || settings.BIRDEYE_CHAIN || "solana";
+    const chain = headers.get("x-chain") || settings.BIRDEYE_CHAIN || "solana";
 
     while (attempts < this.maxRetries) {
       attempts++;
diff --git a/plugins/plugin-wallet/src/analytics/birdeye/providers/market.ts b/plugins/plugin-wallet/src/analytics/birdeye/providers/market.ts
@@ -131,7 +131,7 @@ export const marketProvider: Provider = {
       const tokenSymbolsPromise =
         solanaService && typeof solanaService.getTokensSymbols === "function"
           ? solanaService.getTokensSymbols(CAs)
-          : Promise.resolve({});
+          : Promise.resolve({} as Record<string, string>);
 
       const [result, tokenSymbols] = await Promise.all([
         birdeyeService.getTokensMarketData("solana", CAs),
diff --git a/plugins/plugin-wallet/src/analytics/lpinfo/kamino/services/kaminoLiquidityService.ts b/plugins/plugin-wallet/src/analytics/lpinfo/kamino/services/kaminoLiquidityService.ts
@@ -6,7 +6,7 @@ const KAMINO_API_BASE_URL = "https://api.kamino.finance";
 const KAMINO_LIQUIDITY_PROGRAM_ID = "kamino-rest-api";
 
 // Known token addresses for reference
-const KNOWN_TOKENS = {
+const KNOWN_TOKENS: Record<string, string> = {
   HeLp6NuQkmYB4pYWo2zYs22mESHXPQYzXbB8n4V98jwC: "AI16Z Token",
   ai16z: "AI16Z Token (Symbol)",
   "4WfUvajjYTrq7KRdToJBkoHQ6bSt7NyBeLhP9LKwtFKh": "Kamino Strategy",

Original file line number	Diff line number	Diff line change
`@@ -1340,7 +1340,7 @@ export async function handleConversationRoutes(`
`1340`	`1340`	`if (aborted) break;`
`1341`	`1341`	`streamedText += chunk;`
`1342`	`1342`	`writeChatTokenSse(res, chunk, streamedText);`
`1343`		`- await new Promise((resolve) => setTimeout(resolve, 15));`
	`1343`	`+ await new Promise((resolve) => setTimeout(resolve, 60));`
`1344`	`1344`	`}`
`1345`	`1345`	`}`
`1346`	`1346`	// Emit `done` BEFORE persistence so user-perceived end-of-turn