@ruelya/grok-build 1.0.3 → 1.0.5-rev

package/README.md

@@ -177,23 +177,21 @@ grok-build-orchestrator = true
 
 ### Prompt cache key & recap model
 
-#### `auto_prompt_cache_key` (per model)
+#### `auto_prompt_cache_key` (absorbed)
 
-**Effect:** On **Responses** backends (`responses` / `openai_responses`), each main turn automatically sends a **session-stable** `prompt_cache_key`. Providers that support prompt caching can reuse KV for the stable prefix of the conversation → lower latency and cost on long sessions. **Off by default.** Chat Completions / Messages ignore this flag.
+**Status:** Absorbed by official Grok Build (product 1.0.3). The per-model flag is kept for config compatibility and does **not** change the wire.
 
-**Config** (on a `[model.*]` entry):
+**Expected behavior:** On Responses backends (`responses` / `openai_responses`), every request sends `prompt_cache_key`. If unset, it defaults to `x_grok_conv_id` (main turns: the session id). Recap and `/btw` send the **parent session id** so they share that prefix cache. Chat Completions / Messages still do not forward the field.
 
 ```toml
-[model.sol]
-api_backend = "openai_responses"
-auto_prompt_cache_key = true
+# No longer required — Responses already send prompt_cache_key.
+# [model.sol]
+# auto_prompt_cache_key = true
 ```
 
-Keys are derived from the session (main vs recap use different derivations so caches do not collide).
-
 #### Recap model (`[models] recap`)
 
-**Effect:** Session **recap** (`/recap` and automatic return-from-away recap) can run on a **cheaper / smaller** model than the main chat model. When unset, recap follows the OAuth-preferred official built-in. When set, your override wins.
+**Effect:** Session **recap** (`/recap` and automatic return-from-away recap) can run on a **cheaper / smaller** model than the main chat model. **Unset** → current session model (official default). **Set** → that catalog key or model id (own endpoint / credentials when the `[model.*]` entry has them). If the override cannot be resolved, recap falls back to the session model.
 
 **Config:**
 
@@ -252,7 +250,7 @@ auto_sync_catalog = true     # refresh models.dev on sync
 
 | File | Purpose |
 |------|---------|
-| `examples/config.api-backends.toml` | Loose backends + `auto_prompt_cache_key` |
+| `examples/config.api-backends.toml` | Loose backends + `[models] recap` |
 | `examples/config.usage-pricing.toml` | Cost modes / live display |
 | `examples/prices.custom.toml` | Custom $/1M rates |
 | `examples/usage-sync.toml` | WebDAV sync |
@@ -433,23 +431,21 @@ grok-build-orchestrator = true
 
 ### Prompt cache key 与 recap 模型
 
-#### `auto_prompt_cache_key`（按模型）
+#### `auto_prompt_cache_key`（已被官方吸收）
 
-**效用：** 在 **Responses** 后端（`responses` / `openai_responses`）上，每一主轮自动发送 **会话稳定** 的 `prompt_cache_key`。支持 prompt 缓存的提供商可复用对话稳定前缀的 KV → 长会话更低延迟与费用。**默认关闭。** Chat Completions / Messages 会忽略该标志。
+**状态：** 已被官方 Grok Build（product 1.0.3）吸收。按模型开关仍可写在配置里，**不再改变线上结果**。
 
-**配置**（写在某个 `[model.*]` 上）：
+**预期行为：** 在 Responses 后端（`responses` / `openai_responses`）上，每个请求都会带 `prompt_cache_key`。未显式设置时回退为 `x_grok_conv_id`（主轮即 session id）。Recap 与 `/btw` 发送 **父 session id**，以便共享该前缀缓存。Chat Completions / Messages 仍不转发该字段。
 
 ```toml
-[model.sol]
-api_backend = "openai_responses"
-auto_prompt_cache_key = true
+# 已不必设置 — Responses 默认就会带 prompt_cache_key。
+# [model.sol]
+# auto_prompt_cache_key = true
 ```
 
-key 由会话派生（主轮与 recap 使用不同派生，避免缓存互相污染）。
-
 #### Recap 模型（`[models] recap`）
 
-**效用：** 会话 **recap**（`/recap` 与离开再回时的自动 recap）可以跑在比主对话 **更便宜/更小** 的模型上。未设置时跟随 OAuth 偏好的官方内置；设置后以你的覆盖为准。
+**效用：** 会话 **recap**（`/recap` 与离开再回时的自动 recap）可以跑在比主对话 **更便宜/更小** 的模型上。**未设置** → 当前会话模型（与官方一致）。**设置** → 该目录键或模型 id（`[model.*]` 上的 endpoint / 凭据会一并使用）。覆盖无法解析时回退到会话模型。
 
 **配置：**
 
@@ -508,7 +504,7 @@ auto_sync_catalog = true     # 同步时刷新 models.dev
 
 | 文件 | 用途 |
 |------|------|
-| `examples/config.api-backends.toml` | 宽松 backend + `auto_prompt_cache_key` |
+| `examples/config.api-backends.toml` | 宽松 backend + `[models] recap` |
 | `examples/config.usage-pricing.toml` | 计费模式 / 实时显示 |
 | `examples/prices.custom.toml` | 自定义 $/1M |
 | `examples/usage-sync.toml` | WebDAV 同步 |

package/artifacts/BUILD_INFO.txt

@@ -1,2 +1,2 @@
-version=1.0.3
-release=https://github.com/Ruelya/grok-build-rev/releases/download/v1.0.3
+version=1.0.5-rev
+release=https://github.com/Ruelya/grok-build-rev/releases/download/v1.0.5-rev

package/artifacts/config-fork-notes.toml

@@ -48,16 +48,15 @@
 #
 # [model."proxy"]
 # api_backend = "openai_responses"
-# auto_prompt_cache_key = true
 #
-# --- Prompt cache key (opt-in; Responses family only) -------------------------
+# --- Prompt cache key (absorbed by official; Responses always send it) --------
 # [model."your-model-id"]
-# # auto_prompt_cache_key = true
+# # auto_prompt_cache_key = true   # no-op; kept for config compatibility
 #
 # --- Recap model (optional) ---------------------------------------------------
-# Override > OAuth official default > current session model
+# Unset = current session model. Set = catalog key / model id.
 # [models]
-# # recap = "grok-4.5"
+# # recap = "cheap-summary"
 #
 # <<< fork-fork-config-notes
 # =============================================================================

package/bin/install.mjs

@@ -73,10 +73,21 @@ function runVersion(exe) {
     shell: false,
     windowsHide: true,
   });
+  // SIGILL / crash → status null + signal; treat as unusable (e.g. neoverse-v2
+  // binary on Neoverse-N1 Ampere).
+  if (r.error || r.signal || (typeof r.status === "number" && r.status !== 0)) {
+    return null;
+  }
   const out = ((r.stdout || "") + (r.stderr || "")).trim();
   return out.split(/\r?\n/)[0] || null;
 }
 
+/** True when an on-disk binary exists but cannot even print --version. */
+function isBrokenBinary(exe) {
+  if (!existsSync(exe) || (fileSize(exe) || 0) < 1_000_000) return false;
+  return runVersion(exe) == null;
+}
+
 function fileSize(p) {
   try {
     return statSync(p).size;
@@ -273,14 +284,31 @@ async function resolveSourceBinary() {
     );
   }
 
-  const cacheDir = join(PKG_ROOT, "artifacts", "cache", t.key);
+  // Versioned cache so upgrading 1.0.2 → 1.0.3 never reuses a bad binary
+  // (e.g. neoverse-v2 SIGILL on Ampere Neoverse-N1).
+  const ver = String(pkg.version || "unknown").replace(/[^\w.-]+/g, "_");
+  const cacheDir = join(PKG_ROOT, "artifacts", "cache", ver, t.key);
   mkdirSync(cacheDir, { recursive: true });
   const dest = join(cacheDir, t.exeName);
-  // Always re-download when missing; keep cache across reinstalls of same version.
-  if (!existsSync(dest) || fileSize(dest) < 1_000_000) {
+  const needFetch =
+    !existsSync(dest) ||
+    (fileSize(dest) || 0) < 1_000_000 ||
+    isBrokenBinary(dest);
+  if (needFetch) {
     console.log(`Downloading ${url} …`);
     await download(url, dest);
   }
+  // Refuse to install a still-broken download (wrong arch / CPU baseline).
+  if (isBrokenBinary(dest)) {
+    throw new Error(
+      [
+        `Downloaded binary cannot run on this CPU (Illegal instruction / crash).`,
+        `  url: ${url}`,
+        `  tip: need a portable linux-arm64 build (target-cpu=generic, not neoverse-v2).`,
+        `  reinstall: npm i -g @ruelya/grok-build@latest && npx grok-build install`,
+      ].join("\n")
+    );
+  }
   return { path: dest, source: "download", url };
 }
 

package/config_example.toml

@@ -2,7 +2,7 @@
 # Grok Build — config_example.toml
 # Copy fields you need into:  %USERPROFILE%\.grok\config.toml  (or $GROK_HOME)
 # All keys are optional unless noted. Comments document accepted values.
-# Fork extras (openai_responses, toolset.style, auto_prompt_cache_key, recap)
+# Fork extras (openai_responses, toolset.style, recap model)
 # are marked  # [fork]
 # =============================================================================
 
@@ -92,7 +92,7 @@ combine_queued_prompts = false
 [models]
 default = "grok-4.5"
 default_reasoning_effort = "high"   # none | low | medium | high | …
-# recap = "grok-4.5"                 # [fork] dedicated /recap model
+# recap = "cheap-summary"            # [fork] /recap model; unset = session model
 # web_search = "…"
 # session_summary = "…"
 # image_description = "…"
@@ -129,7 +129,7 @@ default_reasoning_effort = "high"   # none | low | medium | high | …
 #              "openai_chat_completions" | "openai_responses" | "anthropic_messages"
 #   See docs/api-backends.md · examples/config.api-backends.toml
 # api_backend = "openai_responses"
-# auto_prompt_cache_key = true        # [fork] main-turn pck (Responses family only)
+# auto_prompt_cache_key = true        # absorbed: Responses already send pck; no-op
 # temperature = 1.0
 # top_p = 1.0
 # max_completion_tokens = 16384
@@ -163,7 +163,6 @@ default_reasoning_effort = "high"   # none | low | medium | high | …
 # base_url = "https://your-gateway/v1"
 # api_key = "…"
 # api_backend = "openai_responses"    # [fork] loose Responses parse
-# auto_prompt_cache_key = true
 # context_window = 200000
 # reasoning_effort = "high"
 #

package/examples/config.api-backends.toml

@@ -19,7 +19,6 @@ base_url = "https://your-gateway/v1"
 # api_key = "…"
 # env_key = "OPENAI_API_KEY"
 api_backend = "openai_responses"
-auto_prompt_cache_key = true
 context_window = 200000
 reasoning_effort = "high"
 
@@ -60,6 +59,7 @@ anthropic-version = "2023-06-01"
 # env_key = "OPENAI_API_KEY"
 # api_backend = "chat_completions"
 
-# Optional catalog default for picker:
+# Optional catalog default + cheaper recap model:
 # [models]
 # default = "sol"
+# recap = "cheap-chat"

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@ruelya/grok-build",
-  "version": "1.0.3",
-  "description": "Grok Build fork (rev). Thin npm installer — downloads native binary from GitHub Releases.",
+  "version": "1.0.5-rev",
+  "description": "Grok Build fork. Product version tracks upstream (currently 1.0.5-rev). Thin installer downloads from GitHub Releases.",
   "type": "module",
   "bin": {
     "grok": "bin/grok.mjs",
@@ -48,5 +48,5 @@
     "access": "public",
     "registry": "https://registry.npmjs.org/"
   },
-  "forkReleaseBase": "https://github.com/Ruelya/grok-build-rev/releases/download/v1.0.3"
+  "forkReleaseBase": "https://github.com/Ruelya/grok-build-rev/releases/download/v1.0.5-rev"
 }