billion-context 0.1.27 → 0.1.28

package/README.md

@@ -303,6 +303,7 @@ with no file at all).
 | `ACP_SESSION_HEADER` | `x-acp-session` | Conversation-id header name |
 | `ACP_COMPRESS_TOOL` | `1` | Set `0` to disable injecting the compress tool |
 | `ACP_COMPRESS_NUDGE` | `1` | Set `0` to disable compression nudges |
+| `ACP_REASONING_KEEP` | *(default)* | Responses API only: set `none` to drop all reasoning items. Default routes reasoning through the compression pipeline so it is hidden automatically once its turn is summarized (prevents the unbounded accumulation that broke Codex's prompt-cache prefix). |
 | `ACP_DEBUG` | `0` | Set `1` for verbose logging |
 | `ACP_PASSTHROUGH` | `0` | Set `1` to forward without compression |
 | `ACP_AUTO_UPDATE` | `1` | Set `0` to disable background self-update |

package/README.zh-CN.md

@@ -253,6 +253,7 @@ bili --no-auto-update        # 本次启动禁用自动更新
 | `ACP_SESSION_HEADER` | `x-acp-session` | 会话标识 header 名 |
 | `ACP_COMPRESS_TOOL` | `1` | 设 `0` 禁止注入 compress 工具 |
 | `ACP_COMPRESS_NUDGE` | `1` | 设 `0` 禁止压缩 nudge |
+| `ACP_REASONING_KEEP` | *(默认)* | 仅 Responses API：设 `none` 丢弃全部 reasoning。默认走压缩管线，turn 被压缩时自动隐藏（防止 reasoning 无限累积撑爆 Codex 的 prompt-cache 前缀）。 |
 | `ACP_DEBUG` | `0` | 设 `1` 打开详细日志 |
 | `ACP_PASSTHROUGH` | `0` | 设 `1` 不压缩直接转发 |
 | `ACP_AUTO_UPDATE` | `1` | 设 `0` 禁用后台自动更新 |

package/dist/index.js

@@ -46647,7 +46647,6 @@ function firstImagePart(content) {
 // src/responses.ts
 var OPAQUE_ITEM_TYPES = /* @__PURE__ */ new Set([
   "additional_tools",
-  "reasoning",
   "computer_call",
   "computer_call_output",
   "file_search_call",
@@ -46660,6 +46659,9 @@ var OPAQUE_ITEM_TYPES = /* @__PURE__ */ new Set([
 function isOpaqueItem(it2) {
   return OPAQUE_ITEM_TYPES.has(it2.type);
 }
+function shouldDropAllReasoning() {
+  return (process.env.ACP_REASONING_KEEP ?? "").trim().toLowerCase() === "none";
+}
 function partText(p2) {
   if (p2.type === "input_text" || p2.type === "output_text") {
     const t = p2.text;
@@ -46691,6 +46693,7 @@ function responsesToCore(body) {
   const systemParts = [];
   const preamble = [];
   const customToolCallIds = /* @__PURE__ */ new Set();
+  let droppedReasoning = 0;
   if (typeof body.instructions === "string" && body.instructions.trim()) {
     systemParts.push(body.instructions);
   }
@@ -46701,7 +46704,7 @@ function responsesToCore(body) {
     const base = deriveMessageId("user", "text", body.input);
     msgs.push({ id: clusters.next(base), role: "user", contentType: "text", text: body.input });
     idx++;
-    return { msgs, systemParts, preamble, customToolCallIds };
+    return { msgs, systemParts, preamble, customToolCallIds, droppedReasoning };
   }
   for (const it2 of items) {
     if (isOpaqueItem(it2)) {
@@ -46709,6 +46712,23 @@ function responsesToCore(body) {
       continue;
     }
     switch (it2.type) {
+      case "reasoning": {
+        if (shouldDropAllReasoning()) {
+          droppedReasoning++;
+          break;
+        }
+        const rid = typeof it2.id === "string" ? String(it2.id) : hashId(JSON.stringify(it2));
+        const base = deriveMessageId("assistant", "reasoning", rid);
+        msgs.push({
+          id: clusters.next(base),
+          role: "assistant",
+          contentType: "reasoning",
+          text: rid,
+          rawResponsesItem: it2
+        });
+        idx++;
+        break;
+      }
       case "message": {
         const m2 = it2;
         const text = messageContent(m2.content);
@@ -46809,7 +46829,7 @@ function responsesToCore(body) {
         break;
     }
   }
-  return { msgs, systemParts, preamble, customToolCallIds };
+  return { msgs, systemParts, preamble, customToolCallIds, droppedReasoning };
 }
 function coreToResponses(messages, customToolCallIds = /* @__PURE__ */ new Set()) {
   const out = [];
@@ -46843,6 +46863,10 @@ function coreToResponses(messages, customToolCallIds = /* @__PURE__ */ new Set()
             arguments: m2.text ?? ""
           });
         }
+      } else if (m2.contentType === "reasoning") {
+        if (m2.rawResponsesItem) {
+          out.push(m2.rawResponsesItem);
+        }
       }
     } else if (m2.role === "tool") {
       const callId = m2.toolCallId ?? "";
@@ -49796,7 +49820,8 @@ async function handle(req, res, opts, core, config, log2) {
   }
   const url = req.url ?? "";
   const urlPath = url.split("?", 2)[0];
-  const protocol = req.method === "POST" && bodyBuffer.length > 0 ? urlPath.endsWith("/chat/completions") ? "openai" : urlPath.endsWith("/v1/messages") || urlPath.endsWith("/messages") ? "anthropic" : urlPath.endsWith("/responses") ? "responses" : null : null;
+  const countTokens = isCountTokensRequest(req.method ?? "GET", urlPath, bodyBuffer.length > 0);
+  const protocol = req.method === "POST" && bodyBuffer.length > 0 ? urlPath.endsWith("/chat/completions") ? "openai" : urlPath.endsWith("/v1/messages") || urlPath.endsWith("/messages") ? "anthropic" : urlPath.endsWith("/responses") ? "responses" : countTokens ? "anthropic" : null : null;
   const route = resolveUpstream(opts, req.url ?? "", req);
   const upstreamOrigin = route ? route.upstream : opts.upstream;
   let parsed = null;
@@ -49839,7 +49864,7 @@ async function handle(req, res, opts, core, config, log2) {
     const clientLabel = clientConversationHeader(req.headers);
     const session = getSession(sessionId, { protocol, upstreamOrigin, label: clientLabel ?? void 0 });
     await withSessionLock(session, async () => {
-      prepared = protocol === "anthropic" ? prepareAnthropic(parsed, req, opts, core, reqConfig, log2, session) : protocol === "openai" ? prepareOpenai(parsed, req, opts, core, reqConfig, log2, session) : prepareResponses(parsed, req, opts, core, reqConfig, log2, session);
+      prepared = countTokens ? prepareCountTokens(parsed, core, reqConfig, log2, session) : protocol === "anthropic" ? prepareAnthropic(parsed, req, opts, core, reqConfig, log2, session) : protocol === "openai" ? prepareOpenai(parsed, req, opts, core, reqConfig, log2, session) : prepareResponses(parsed, req, opts, core, reqConfig, log2, session);
       acquireInFlight(session);
       try {
         await forward(req, res, opts, prepared.body, prepared, core, reqConfig, log2, route, affinity);
@@ -49880,6 +49905,42 @@ function diagNudge(turn, sessionId, tokenCount, limit) {
   const inject = n.shouldInject ? `INJECT T${n.tier ?? "?"}` : "idle";
   return `[${sessionId}] nudge ${inject}: usage=${pct2} (${tokenCount}/${limit}), growth=${growth}/${floor} (ref=${ref}, interval=${interval}), pendingT1=${pendingT1}/${interval}, reason="${n.reason.slice(0, 120)}"`;
 }
+function isCountTokensRequest(method, urlPath, hasBody) {
+  return method === "POST" && hasBody && process.env.ACP_COUNT_TOKENS_PASSTHROUGH !== "1" && urlPath.endsWith("/messages/count_tokens");
+}
+function prepareCountTokens(parsed, core, config, log2, session) {
+  const sessionId = session.id;
+  try {
+    const { msgs, cacheControls } = anthropicToCore(parsed);
+    const turn = core.processTurn({ messages: msgs, state: session.state, config, tokenCount: session.stats.lastInputTokens, renderTags: "text-only" });
+    const rebuiltMessages = coreToAnthropic(turn.messages, cacheControls);
+    log2("info", `[${sessionId}] count_tokens pruned: ${msgs.length} \u2192 ${turn.messages.length} msgs`);
+    return {
+      body: JSON.stringify({ ...parsed, messages: rebuiltMessages }),
+      session,
+      processedMessages: [],
+      originalMessages: msgs,
+      protocol: "anthropic",
+      stream: false,
+      // MUST stay false + empty processedMessages: keeps forward() on
+      // plain-pipeThrough so usage-capture/markDirty stays unreachable
+      // (else it poisons lastInputTokens with the compressed count and
+      // suppresses the proxy's own nudges).
+      compressInjected: false
+    };
+  } catch (err) {
+    log2("warn", `[${sessionId}] count_tokens prune failed, forwarding unchanged: ${String(err)}`);
+    return {
+      body: JSON.stringify(parsed),
+      session,
+      processedMessages: [],
+      originalMessages: [],
+      protocol: "anthropic",
+      stream: false,
+      compressInjected: false
+    };
+  }
+}
 function prepareAnthropic(parsed, req, opts, core, config, log2, session) {
   const sessionId = session.id;
   const stream2 = parsed.stream === true;