billion-context 0.1.82-pr.512.336 → 0.1.82

package/README.md

@@ -112,63 +112,6 @@ turn is a semantic mismatch the model tolerates (it is clearly marked
   summaries (their tool call isn't in the agent's history and the agent's view
   skips `acp_summary`) — but that needs the id match above, which doesn't occur.
 
-## Which do I need? — billion-context vs billion-context-pi vs opencode-acp
-
-Three names keep appearing around this project. They are **different layers of
-the same idea** — context compression for coding agents — not three competing
-products. The deciding question is whether your client has a **"context hook"**
-(a place to host an in-process extension that registers the `compress` tool and
-runs compression locally):
-
-| Name | What it is | Where it runs | Pick it when |
-|---|---|---|---|
-| **billion-context-pi** | Standalone **in-process pi extension** (pure-extension mode — acp-kernel integrated directly into pi; the reference implementation). No external proxy needed. | Inside the pi process | **Preferred for pi.** Active, not superseded. |
-| **opencode-acp** | Standalone **in-process opencode extension** (same idea, for opencode). No external proxy needed. | Inside the opencode process | Preferred for opencode when you want pure in-process, no proxy. |
-| **billion-context** (`bili`) | The external **compression proxy** + launchers + built-in client-side plugins. **Mostly for clients that have no context hook.** | A local proxy process | Default for hookless clients (Claude Code, Codex, Cursor, Aider, ZCode…). Also how you drive **omp** (no standalone omp extension exists — it uses `bili`'s built-in plugin). |
-
-```
-Does your client have a "context hook" (a spot for an in-process extension)?
-
-  ├─ yes → prefer an in-process extension (no external proxy running)
-  │    • pi       → billion-context-pi    ← preferred
-  │    • opencode → opencode-acp
-  │    • omp      → billion-context (built-in omp plugin; no standalone omp pkg)
-  │
-  └─ no  → the billion-context proxy (bili)
-       Claude Code / Codex / Cursor / Aider / ZCode …
-       two ways (prefer the launcher):
-         ✅ bili <client>   launcher mode   ← preferred
-         ○  /bili/ prefix   plain proxy mode
-```
-
-**When to pick which:**
-
-- **Client has a context hook** → prefer the in-process extension (tightest
-  integration, no separate proxy process):
-  - **pi** → **`billion-context-pi`** — the reference implementation and the
-    primary way to run pi with compression; acp-kernel runs *inside* pi.
-  - **opencode** → **`opencode-acp`** — the same idea for opencode.
-  - **omp** → **`billion-context`** — omp has no standalone extension of its own,
-    so it rides `billion-context`'s built-in omp plugin (`bili omp`).
-- **Client has no context hook** → **`billion-context`** (`bili`), the external
-  proxy — its main job. Anything that can set a base URL but can't host an
-  in-process extension: Claude Code, Codex, Cursor, Aider, ZCode, etc. Two ways,
-  **prefer the launcher**:
-  - `bili <client>` — **launcher mode** (spawns a proxy and wires the client up
-    automatically; recommended).
-  - `/bili/` base-URL prefix — plain proxy mode (point the client at a running
-    proxy yourself).
-
-> Note: `billion-context` also ships thin built-in plugins for pi / omp /
-> opencode, so those hooked clients *can* optionally route through one shared
-> proxy (`bili pi` / `bili opencode`). In that case any standalone extension
-> (e.g. `billion-context-pi`) self-disables via `BILLION_CONTEXT_PROXY` to avoid
-> double compression. For pi, the standalone `billion-context-pi` remains the
-> preferred path; the proxy is the fallback / unification option.
-
-**TL;DR:** pi → `billion-context-pi`. opencode → `opencode-acp`. omp and every
-hookless client → `billion-context` (`bili`), launcher mode first.
-
 ## Install
 
 ```bash
@@ -439,7 +382,7 @@ pass an explicit `x-acp-session` header per conversation to avoid collisions.
 
 Early. Protocol handling and compression work against mock tests (500+ passing). Real-model integration testing is the next milestone. Expect rough edges.
 
-Client-side plugins for pi / omp / opencode ship inside `billion-context` (`dist/agent/*.js`) for the cooperative-proxy path. See the **"Which do I need?"** section above for how `billion-context`, the standalone `billion-context-pi`, and `opencode-acp` relate.
+See [billion-context-pi](https://github.com/ranxianglei/billion-context-pi) for the pi-extension mode (in-process, tighter integration, the reference implementation).
 
 ## License
 

package/README.zh-CN.md

@@ -50,45 +50,6 @@ AI 编程助手的<strong>通用上下文压缩代理</strong>
 
 代理向对话注入四个上下文管理工具(`compress`、`decompress`、`search_context`、`acp_status`)。模型在对话增长时调用 `compress`,代理在服务端执行 —— 压缩后的范围在下一轮之前折叠进对话历史。
 
-## 该选哪个?—— billion-context / billion-context-pi / opencode-acp
-
-围绕这个项目经常会出现三个名字。它们是<strong>同一件事(编程助手的上下文压缩)的不同层次</strong>,<em>不是三选一的对立产品</em>。关键看你的客户端有没有<strong>「上下文 hook」</strong>(能装一个进程内扩展、把 `compress` 工具注册进去并在本地跑压缩的地方):
-
-| 名字 | 是什么 | 跑在哪 | 什么时候选它 |
-|---|---|---|---|
-| **billion-context-pi** | 独立的<strong>进程内 pi 扩展</strong>(纯扩展模式 —— acp-kernel 直接集成进 pi;参考实现)。不需要外部代理。 | pi 进程内 | <strong>pi 首选。</strong>活跃,并未被取代。 |
-| **opencode-acp** | 独立的<strong>进程内 opencode 扩展</strong>(同样思路,面向 opencode)。不需要外部代理。 | opencode 进程内 | 想要纯进程内、不跑代理时,opencode 首选。 |
-| **billion-context**(`bili`) | 外部<strong>压缩代理</strong> + 启动器 + 内置客户端插件。<strong>主要面向没有上下文 hook 的客户端。</strong> | 本地代理进程 | 无 hook 客户端(Claude Code、Codex、Cursor、Aider、ZCode…)的默认选择;也是驱动 **omp** 的方式(没有独立的 omp 扩展包 —— 用 `bili` 内置插件)。 |
-
-```
-你的客户端有没有「上下文 hook」(能装进程内扩展的地方)?
-
-  ├─ 有 → 优先进程内扩展(不跑外部代理)
-  │    • pi       → billion-context-pi    ← 首选
-  │    • opencode → opencode-acp
-  │    • omp      → billion-context(内置 omp 插件;无独立 omp 包)
-  │
-  └─ 没有 → billion-context 外部代理(bili)
-       Claude Code / Codex / Cursor / Aider / ZCode …
-       两种方式(优先启动器):
-         ✅ bili <client>   启动器模式   ← 首选
-         ○  /bili/ 前缀      纯代理模式
-```
-
-<strong>什么时候选哪个:</strong>
-
-- <strong>客户端有上下文 hook</strong> → 优先用进程内扩展(集成最紧、不额外跑一个代理进程):
-  - <strong>pi</strong> → <strong>`billion-context-pi`</strong> —— 参考实现,也是 pi 跑压缩的主要方式,acp-kernel 直接在 pi 进程里跑。
-  - <strong>opencode</strong> → <strong>`opencode-acp`</strong> —— opencode 的同款思路。
-  - <strong>omp</strong> → <strong>`billion-context`</strong> —— omp 没有自己的独立扩展包,所以走 `billion-context` 内置的 omp 插件(`bili omp`)。
-- <strong>客户端没有上下文 hook</strong> → <strong>`billion-context`</strong>(`bili`)外部代理 —— 这是它的主场。任何能设 base URL、但装不了进程内扩展的客户端:Claude Code、Codex、Cursor、Aider、ZCode 等。两种方式,<strong>优先启动器</strong>:
-  - `bili <client>` —— <strong>启动器模式</strong>(自动拉起代理并接好客户端;推荐)。
-  - `/bili/` base-URL 前缀 —— 纯代理模式(你自己把客户端指向已运行的代理)。
-
-> 说明:`billion-context` 也自带针对 pi / omp / opencode 的轻量内置插件,让这些有 hook 的客户端<strong>可以</strong>选择性走同一个共享代理(`bili pi` / `bili opencode`)。此时任何独立扩展(如 `billion-context-pi`)会经 `BILLION_CONTEXT_PROXY` 自禁用,避免双层压缩。对 pi 而言,独立的 `billion-context-pi` 仍是首选路径;代理是兜底 / 统一入口的选项。
-
-<strong>一句话总结:</strong>pi → `billion-context-pi`;opencode → `opencode-acp`;omp 和所有无 hook 客户端 → `billion-context`(`bili`),优先启动器模式。
-
 ## 安装
 
 ```bash
@@ -297,7 +258,7 @@ Windows 下会自动发现常见 Clash/Mihomo 静态系统代理;Web UI 会显
 
 早期。协议处理和压缩已通过 mock 测试(500+ 项通过)。真实模型集成测试是下一里程碑。预期会有粗糙的地方。
 
-针对 pi / omp / opencode 的客户端插件随 `billion-context` 一起发布(`dist/agent/*.js`),用于协作代理路径。三者(`billion-context`、独立的 `billion-context-pi`、`opencode-acp`)如何取舍,见上文「该选哪个?」一节。
+pi 扩展模式(进程内、更紧密集成、参考实现)见 [billion-context-pi](https://github.com/ranxianglei/billion-context-pi)。
 
 ## 许可证
 

package/dist/index.js

@@ -48360,64 +48360,33 @@ async function contextFromRegistry(model, host) {
 function peekRegistryContext(model, host) {
   return registryLookup(cache2, model, host);
 }
-var VARIANT_SUFFIXES = [
-  "-interleaved-thinking",
-  "-thinking",
-  "-reasoning",
-  "-extended",
-  "-fast",
-  "-high",
-  "-medium",
-  "-low"
-];
-function modelVariants(name) {
-  const variants = [name];
-  let current = name;
-  for (; ; ) {
-    let stripped = false;
-    for (const suffix of VARIANT_SUFFIXES) {
-      if (current.length > suffix.length && current.endsWith(suffix)) {
-        current = current.slice(0, -suffix.length);
-        variants.push(current);
-        stripped = true;
-        break;
-      }
-    }
-    if (!stripped) break;
-  }
-  return variants;
-}
 function registryLookup(reg, model, host) {
   if (!reg || !model) return void 0;
   const provider = host ? providerFromHost(host) : void 0;
-  for (const name of modelVariants(model)) {
-    const candidates = provider ? [`${provider}/${name}`, name] : [name];
-    for (const key of candidates) {
-      const entry = reg[key];
-      const ctx = entry?.limit?.context;
-      if (typeof ctx === "number" && ctx > 0) return ctx;
-    }
-    if (provider === void 0) {
-      const suffix = `/${name}`;
-      let max;
-      const distinct = /* @__PURE__ */ new Set();
-      const parts = [];
-      for (const key of Object.keys(reg)) {
-        if (!key.endsWith(suffix)) continue;
-        const ctx = reg[key].limit?.context;
-        if (typeof ctx !== "number" || ctx <= 0) continue;
-        if (max === void 0 || ctx > max) max = ctx;
-        distinct.add(ctx);
-        parts.push(`${key}=${ctx}`);
-      }
-      if (max !== void 0) {
-        if (distinct.size > 1 && !warnedConflicts.has(name)) {
-          warnedConflicts.add(name);
-          log("warn", `[acp-registry] conflicting context windows for "${name}" (${parts.join(", ")}) \u2014 using max ${max}`);
-        }
-        return max;
-      }
+  const candidates = provider ? [`${provider}/${model}`, model] : [model];
+  for (const key of candidates) {
+    const entry = reg[key];
+    const ctx = entry?.limit?.context;
+    if (typeof ctx === "number" && ctx > 0) return ctx;
+  }
+  if (provider === void 0) {
+    const suffix = `/${model}`;
+    let max;
+    const distinct = /* @__PURE__ */ new Set();
+    const parts = [];
+    for (const key of Object.keys(reg)) {
+      if (!key.endsWith(suffix)) continue;
+      const ctx = reg[key].limit?.context;
+      if (typeof ctx !== "number" || ctx <= 0) continue;
+      if (max === void 0 || ctx > max) max = ctx;
+      distinct.add(ctx);
+      parts.push(`${key}=${ctx}`);
+    }
+    if (max !== void 0 && distinct.size > 1 && !warnedConflicts.has(model)) {
+      warnedConflicts.add(model);
+      log("warn", `[acp-registry] conflicting context windows for "${model}" (${parts.join(", ")}) \u2014 using max ${max}`);
     }
+    return max;
   }
   return void 0;
 }
@@ -54736,7 +54705,7 @@ function handlePluginStatus(conversationId2, res, fallbackLatest = false) {
   if (entry) entry.lastSeen = Date.now();
   const limit = session.metadata.effectiveContextLimit;
   const mem = remembered.get(session.id);
-  const modelContextLimit = typeof limit === "number" && limit > 0 ? limit : 0;
+  const modelContextLimit = typeof limit === "number" && limit > 0 ? limit : 2e5;
   let panel;
   try {
     panel = buildStatusPanel({
@@ -56760,9 +56729,6 @@ function logUnrecognizedPath(log2, url) {
     log2("info", `unrecognized path ${key}: forwarding unchanged; further occurrences suppressed`);
   }
 }
-function isModelDiscoveryPath(path18) {
-  return path18.replace(/\/+$/, "").endsWith("/models");
-}
 var BILI_HOP_HEADER = "x-bili-hop";
 function parseLauncherModelWindows(raw) {
   if (!raw) return {};
@@ -57308,11 +57274,11 @@ ${bodyBuffer.toString("utf8")}`);
       const pluginWindow = pluginReportedContextWindow(req.headers);
       const launcherWindow = launcherContextWindow(model);
       const configuredWindow = resolveConfiguredContextLimit(opts.routes, embeddedUrl, model);
-      const peekWindow = peekRegistryContext(model, host);
+      const peekWindow = host ? peekRegistryContext(model, host) : void 0;
       let native = betaWindow ?? pluginWindow ?? launcherWindow ?? configuredWindow ?? peekWindow ?? lookupContextLimit(model);
       const operatorWindowTuned = resolveCompress(opts.routes, embeddedUrl, model, opts.compress).modelContextLimit !== void 0;
       nativeFromFallback = !betaWindow && !pluginWindow && !launcherWindow && !peekWindow && !configuredWindow && !operatorWindowTuned;
-      if (!native) {
+      if (!native && host) {
         native = await contextFromRegistry(model, host);
         if (native) nativeFromFallback = false;
       }
@@ -57413,6 +57379,7 @@ ${bodyBuffer.toString("utf8")}`);
     if (pluginAgent && !pluginConversation) pluginConversation = conversation;
     if (pluginAgent) {
       if (session.metadata.pluginAgent !== pluginAgent) session.metadata.pluginAgent = pluginAgent;
+      session.metadata.effectiveContextLimit = reqConfig.modelContextLimit;
       recordPluginSession(pluginConversation ?? conversation, session.id);
     }
     if (protocol === "responses" || protocol === "openai" || protocol === "anthropic") {
@@ -57430,22 +57397,6 @@ ${bodyBuffer.toString("utf8")}`);
       reqConfig = { ...reqConfig, modelContextLimit: learnedLimit };
       nativeFromFallback = false;
       log2("info", `[${session.id}] self-healed context window: ${resolved} \u2192 ${learnedLimit} (learned from an upstream overflow)`);
-    } else if (nativeFromFallback && reqModel) {
-      const prevInput = session.stats.lastInputTokens ?? 0;
-      const prevWindow = session.metadata.lastTurnWindow;
-      const resolved = reqConfig.modelContextLimit;
-      if (prevWindow !== void 0 && prevInput > prevWindow && prevInput > resolved && prevInput >= 1e3) {
-        const map = session.metadata.learnedContextLimits ?? {};
-        const prev = map[reqModel];
-        if (prev === void 0 || prevInput > prev) {
-          map[reqModel] = prevInput;
-          session.metadata.learnedContextLimits = map;
-          markDirty(session);
-        }
-        reqConfig = { ...reqConfig, modelContextLimit: prevInput };
-        nativeFromFallback = false;
-        log2("info", `[${session.id}] self-healed context window upward: ${resolved} \u2192 ${prevInput} (a prior turn used ${prevInput} input tokens, exceeding the fallback window)`);
-      }
     }
     const nativeWindow = reqConfig.modelContextLimit;
     if (shouldReserveOutputHeadroom(protocol)) {
@@ -57459,8 +57410,6 @@ ${bodyBuffer.toString("utf8")}`);
       }
       if (reserved !== reqConfig.modelContextLimit) reqConfig = { ...reqConfig, modelContextLimit: reserved };
     }
-    session.metadata.effectiveContextLimit = reqConfig.modelContextLimit;
-    session.metadata.lastTurnWindow = reqConfig.modelContextLimit;
     acquireInFlight(session);
     try {
       await withSessionLock(session, async () => {
@@ -57525,7 +57474,7 @@ ${bodyBuffer.toString("utf8")}`);
     }
   }
   if (!prepared) {
-    if (protocol === null && !opts.passthrough && !isModelDiscoveryPath(urlPath)) {
+    if (protocol === null && !opts.passthrough) {
       logUnrecognizedPath(log2, req.url ?? "");
     }
     await forward(req, res, opts, bodyBuffer, null, core, reqConfig, log2, route, instanceId, void 0);