billion-context 0.1.99 → 0.1.100-pr.647.509

package/dist/index.js

@@ -47777,7 +47777,14 @@ async function fetchWithTimeout(url, opts, timeoutMs, externalSignal) {
     const finalOpts = {
       ...opts,
       signal: controller.signal,
-      dispatcher: opts.dispatcher ?? directDispatcher(effective)
+      dispatcher: opts.dispatcher ?? directDispatcher(effective),
+      // Forward-proxy correctness: never silently follow a redirect.
+      // undici's default (follow) downgrades POST→GET and drops the body
+      // on 301/302/303, so a redirecting upstream (CDN/WAF) turns a valid
+      // POST into a 405 at the redirect target. Pass the 3xx through to
+      // the client, which follows it with its own policy. Internal bili
+      // fetches that want to follow (registry, upstream test) opt in.
+      redirect: opts.redirect ?? "manual"
     };
     const raw = await fetch(url, finalOpts);
     if (raw.body) {
@@ -48514,6 +48521,7 @@ function loadOptions(env = process.env) {
     passthrough: passthrough.enabled,
     passthroughSource: passthrough.source,
     autoUpdate: (env.ACP_AUTO_UPDATE ?? (fileConfig.autoUpdate === false ? "0" : "1")) !== "0",
+    hostUsageCredit: parseHostUsageCredit(env.BILI_HOST_USAGE_CREDIT ?? fileConfig.hostUsageCredit),
     updateTag: (env.ACP_UPDATE_TAG ?? fileConfig.updateTag ?? "latest").trim() || "latest",
     logFile: env.ACP_LOG_FILE !== void 0 ? env.ACP_LOG_FILE || void 0 : fileConfig.logFile,
     mitm: {
@@ -48579,6 +48587,7 @@ function parseRouteEntry(v2) {
     if (obj.compress) route.compress = obj.compress;
     const compatRoles = parseCompatRoles(obj.compat?.roles);
     if (compatRoles) route.compat = { roles: compatRoles };
+    if (typeof obj.passthrough === "boolean") route.passthrough = obj.passthrough;
     return route;
   }
   if (v2 === null) return {};
@@ -48590,6 +48599,9 @@ function parsePromptCacheRouting(value) {
 function parseUpstreamProxyMode(value) {
   return value === "manual" || value === "auto" ? value : "direct";
 }
+function parseHostUsageCredit(value) {
+  return value === "off" ? "off" : "auto";
+}
 function parseCompressSettings(v2) {
   if (!v2 || typeof v2 !== "object" || Array.isArray(v2)) return void 0;
   const obj = v2;
@@ -49005,7 +49017,8 @@ async function fetchFresh() {
     try {
       const { response, clearTimer } = await fetchWithTimeout(REGISTRY_URL, {
         ...opts,
-        headers: { Accept: "application/json" }
+        headers: { Accept: "application/json" },
+        redirect: "follow"
       }, 15e3);
       const text = await response.text();
       clearTimer();
@@ -49244,7 +49257,9 @@ function isCodexClient(headers) {
   const ua = headers["user-agent"];
   if (!ua) return false;
   const s3 = Array.isArray(ua) ? ua[0] : ua;
-  return typeof s3 === "string" && CODEX_UA_PREFIXES.some((p2) => s3.startsWith(p2));
+  if (typeof s3 !== "string") return false;
+  if (CODEX_UA_PREFIXES.some((p2) => s3.startsWith(p2))) return true;
+  return s3.includes("codex");
 }
 function hasCompactionTrigger(input) {
   if (!Array.isArray(input)) return false;
@@ -51498,20 +51513,44 @@ ${body}`;
 }
 
 // src/loop/tag-echo-filter.ts
-var PAIRED = /\x3cacp\s[^<>]{0,256}>([^<>]{0,64})<\/acp>/;
-var LONE_OPEN = /\x3cacp(?:\s[^<>]{0,256})?>/;
-var LONE_CLOSE = /<\/acp(?=[\s>])[^<>]{0,32}>/;
-var PARTIAL_TAIL = /(\x3cacp\s[^<>]*|\x3c\/acp(?:\s[^<>]{0,32})?|<\/?a?c?p?)$/;
-var TRUNC_OPEN = /\x3cacp\s[^<>]*$/;
-var TRUNC_CLOSE = /\x3c\/acp(?:\s[^<>]{0,32})?$/;
-var CLOSE_TAG = "</acp";
+function buildAcplikeName() {
+  const cores = ["acp", "apc", "cap", "cpa", "pac", "pca"];
+  const names = new Set(cores);
+  for (const c of cores) {
+    for (const ch of ["a", "c", "p", "i"]) {
+      for (let pos = 0; pos <= c.length; pos++) names.add(c.slice(0, pos) + ch + c.slice(pos));
+    }
+  }
+  return [...names].sort((a, b2) => b2.length - a.length).join("|");
+}
+var ACP_NAME_ALT = `(?:${buildAcplikeName()})`;
+var NAME = ACP_NAME_ALT;
+var PAIRED = new RegExp("<" + NAME + "\\s[^<>]{0,256}>([^<>]{0,64})<\\/" + NAME + ">");
+var LONE_OPEN = new RegExp("<" + NAME + "(?:\\s[^<>]{0,256})?>");
+var LONE_CLOSE = new RegExp("<\\/" + NAME + "(?=[\\s>])[^<>]{0,32}>");
+var PARTIAL_TAIL = new RegExp("(<" + NAME + "\\s[^<>]*|<\\/" + NAME + "(?:\\s[^<>]{0,32})?|<\\/?[acip]*)$");
+var TRUNC_OPEN = new RegExp("<" + NAME + "\\s[^<>]*$");
+var TRUNC_CLOSE = new RegExp("<\\/" + NAME + "(?:\\s[^<>]{0,32})?$");
+var DEFINITE_TAIL = new RegExp("^<" + NAME + "\\s|^<\\/" + NAME);
+var OPEN_WITH_ATTRS = new RegExp("^<" + NAME + "\\s");
+var CLOSE_HEAD = "</";
+var CLOSE_NAME_ANCHORED = new RegExp("^" + NAME);
 var HOLD_LIMIT = 128;
 var TAG_OPEN_CAP = 4096;
 var SWALLOW_CAP = 80;
+function looseCloseEnd(s3) {
+  let idx = s3.indexOf(CLOSE_HEAD);
+  while (idx >= 0) {
+    const m2 = CLOSE_NAME_ANCHORED.exec(s3.slice(idx + 2));
+    if (m2 && s3[idx + 2 + m2[0].length] === ">") return idx + 2 + m2[0].length + 1;
+    idx = s3.indexOf(CLOSE_HEAD, idx + 1);
+  }
+  return -1;
+}
 function stripAcpTags(text) {
   return text.replace(new RegExp(PAIRED.source, "g"), "").replace(new RegExp(LONE_OPEN.source, "g"), "").replace(new RegExp(LONE_CLOSE.source, "g"), "").replace(new RegExp(TRUNC_OPEN.source), "").replace(new RegExp(TRUNC_CLOSE.source), "");
 }
-var RENDER_TAG_DETECT = /\x3c\/?acp(?=[\s>])|\\u003c\/?acp(?=[\s>\\])/;
+var RENDER_TAG_DETECT = new RegExp("<\\/?" + NAME + "(?=[\\s>])|\\\\u003c\\/?" + NAME + "(?=[\\s>\\\\])");
 function containsRenderTagText(s3) {
   return RENDER_TAG_DETECT.test(s3);
 }
@@ -51528,6 +51567,8 @@ function createTagEchoFilter(onDrop) {
   let swallowed = "";
   let droppedAny = false;
   let notified = false;
+  let inputChars = 0;
+  let outputChars = 0;
   const drop = (snippet) => {
     droppedAny = true;
     if (onDrop && !notified) {
@@ -51541,10 +51582,9 @@ function createTagEchoFilter(onDrop) {
     for (; ; ) {
       if (swallowUntilClose) {
         const combined = swallowed + buf;
-        const closeIdx = combined.indexOf(CLOSE_TAG);
-        if (closeIdx >= 0 && combined[closeIdx + CLOSE_TAG.length] === ">") {
-          const end = closeIdx + CLOSE_TAG.length + 1;
-          drop(swallowed + combined.slice(0, end));
+        const end = looseCloseEnd(combined);
+        if (end >= 0) {
+          drop(combined.slice(0, end));
           swallowed = "";
           swallowUntilClose = false;
           buf = combined.slice(end);
@@ -51569,7 +51609,7 @@ function createTagEchoFilter(onDrop) {
       if (!m2) {
         const t = PARTIAL_TAIL.exec(buf);
         if (t) {
-          const definite = /^\x3cacp\s/.test(t[0]) || /^\x3c\/acp/.test(t[0]);
+          const definite = DEFINITE_TAIL.test(t[0]);
           const cap = definite ? TAG_OPEN_CAP : HOLD_LIMIT;
           if (t[0].length <= cap) {
             held = t[0];
@@ -51588,7 +51628,7 @@ function createTagEchoFilter(onDrop) {
       drop(m2[0]);
       out += buf.slice(0, m2.index);
       buf = buf.slice(m2.index + m2[0].length);
-      if (m2 === o && /^\x3cacp\s/.test(m2[0])) {
+      if (m2 === o && OPEN_WITH_ATTRS.test(m2[0])) {
         swallowUntilClose = true;
         swallowed = "";
       }
@@ -51597,9 +51637,12 @@ function createTagEchoFilter(onDrop) {
   };
   return {
     push(delta) {
+      inputChars += delta.length;
       const chunk = held + delta;
       held = "";
-      return process2(chunk);
+      const r = process2(chunk);
+      outputChars += r.length;
+      return r;
     },
     flush() {
       const rest = swallowed + held;
@@ -51607,27 +51650,36 @@ function createTagEchoFilter(onDrop) {
       swallowed = "";
       held = "";
       swallowUntilClose = false;
+      let result;
       if (wasSwallowing) {
         if (rest.length > 0) drop(rest);
-        return "";
-      }
-      const t = new RegExp(TRUNC_OPEN.source).exec(rest);
-      if (t) {
-        drop(t[0]);
-        return rest.slice(0, t.index);
-      }
-      const tc = new RegExp(TRUNC_CLOSE.source).exec(rest);
-      if (tc) {
-        drop(tc[0]);
-        return rest.slice(0, tc.index);
+        result = "";
+      } else {
+        const t = new RegExp(TRUNC_OPEN.source).exec(rest);
+        if (t) {
+          drop(t[0]);
+          result = rest.slice(0, t.index);
+        } else {
+          const tc = new RegExp(TRUNC_CLOSE.source).exec(rest);
+          if (tc) {
+            drop(tc[0]);
+            result = rest.slice(0, tc.index);
+          } else {
+            result = rest;
+          }
+        }
       }
-      return rest;
+      outputChars += result.length;
+      return result;
     },
     dropped() {
       return droppedAny;
     },
     pending() {
       return held.length > 0 || swallowUntilClose;
+    },
+    stats() {
+      return { inputChars, outputChars, dropped: droppedAny };
     }
   };
 }
@@ -51932,16 +51984,41 @@ function splitChunks(messages, startIdx, endIdx, budget, minUnits, countText = d
   }
   return chunks;
 }
-function summaryPayload(protocol, model, system, content, stream2) {
+function summaryPayload(protocol, model, system, content, stream2, includeMaxOutputTokens) {
   if (protocol === "anthropic") {
     return { model, max_tokens: MAX_SUMMARY_OUTPUT_TOKENS, system, messages: [{ role: "user", content }], stream: stream2 };
   }
   if (protocol === "openai") {
     return { model, max_tokens: MAX_SUMMARY_OUTPUT_TOKENS, messages: [{ role: "system", content: system }, { role: "user", content }], stream: stream2 };
   }
-  return { model, max_output_tokens: MAX_SUMMARY_OUTPUT_TOKENS, instructions: system, input: [{ role: "user", content }], stream: stream2, store: false };
+  const payload = { model, instructions: system, input: [{ role: "user", content }], stream: stream2, store: false };
+  if (includeMaxOutputTokens) payload.max_output_tokens = MAX_SUMMARY_OUTPUT_TOKENS;
+  return payload;
 }
 var STREAM_REQUIRED_RE = /\bstream\b[^\n]{0,60}\btrue\b/i;
+var MAX_OUTPUT_TOKENS_REJECTED_RE = /\bmax_output_tokens\b/i;
+function noMaxOutputTokensKey(deps) {
+  return `${deps.url}\0${deps.model}`;
+}
+function hasLearnedNoMaxOutputTokens(deps) {
+  const learned = deps.session.metadata.preflightNoMaxOutputTokens;
+  return typeof learned === "object" && learned !== null && learned[noMaxOutputTokensKey(deps)] === true;
+}
+function rememberNoMaxOutputTokens(deps) {
+  const learned = deps.session.metadata.preflightNoMaxOutputTokens;
+  const map = typeof learned === "object" && learned !== null ? learned : {};
+  map[noMaxOutputTokensKey(deps)] = true;
+  deps.session.metadata.preflightNoMaxOutputTokens = map;
+}
+var SUMMARY_STRIP_HEADERS = /* @__PURE__ */ new Set(["x-openai-internal-codex-responses-lite"]);
+function summaryHeaders(deps) {
+  const headers = {};
+  for (const [k2, v2] of Object.entries(deps.headers)) {
+    if (SUMMARY_STRIP_HEADERS.has(k2.toLowerCase())) continue;
+    headers[k2] = v2;
+  }
+  return headers;
+}
 function extractSummaryFromSse(protocol, text) {
   let out = "";
   for (const line of text.split("\n")) {
@@ -52005,25 +52082,39 @@ async function summarizeRange(deps, content, startRef, endRef) {
   const system = buildCompressSystemPrompt(deps.prompts) + `
 
 TASK: The conversation segment below (messages ${startRef}\u2013${endRef}) must be compressed because the session context exceeds the current model's window. Write a tier-1 compression summary of the segment following every rule above. Output ONLY the summary text \u2014 no preamble, no closing remarks, no tool calls.`;
-  const learned = deps.session.metadata.preflightStreamSummary === true;
-  try {
-    return await requestSummary(deps, system, content, learned);
-  } catch (err2) {
-    if (err2 instanceof UpstreamHttpError && err2.status === 400 && !learned && STREAM_REQUIRED_RE.test(err2.body)) {
-      deps.session.metadata.preflightStreamSummary = true;
-      deps.log("info", "[preflight] upstream requires stream for summaries; retrying with SSE (learned for this session)");
-      return await requestSummary(deps, system, content, true);
+  let stream2 = deps.session.metadata.preflightStreamSummary === true;
+  let includeMaxOutputTokens = !(deps.protocol === "responses" && hasLearnedNoMaxOutputTokens(deps));
+  for (; ; ) {
+    try {
+      return await requestSummary(deps, system, content, stream2, includeMaxOutputTokens);
+    } catch (err2) {
+      if (err2 instanceof UpstreamHttpError && err2.status === 400) {
+        let adapted = false;
+        if (!stream2 && STREAM_REQUIRED_RE.test(err2.body)) {
+          deps.session.metadata.preflightStreamSummary = true;
+          stream2 = true;
+          adapted = true;
+          deps.log("info", "[preflight] upstream requires stream for summaries; retrying with SSE (learned for this session)");
+        }
+        if (deps.protocol === "responses" && includeMaxOutputTokens && MAX_OUTPUT_TOKENS_REJECTED_RE.test(err2.body)) {
+          rememberNoMaxOutputTokens(deps);
+          includeMaxOutputTokens = false;
+          adapted = true;
+          deps.log("info", `[preflight] upstream rejects max_output_tokens for summaries (model=${deps.model}); retrying without it (learned for this session+upstream+model)`);
+        }
+        if (adapted) continue;
+      }
+      throw err2;
     }
-    throw err2;
   }
 }
-async function requestSummary(deps, system, content, stream2) {
+async function requestSummary(deps, system, content, stream2, includeMaxOutputTokens) {
   const { response, clearTimer } = await fetchWithRetry(
     deps.url,
     {
       method: "POST",
-      headers: { "content-type": "application/json", ...deps.headers },
-      body: JSON.stringify(summaryPayload(deps.protocol, deps.model, system, content, stream2)),
+      headers: { "content-type": "application/json", ...summaryHeaders(deps) },
+      body: JSON.stringify(summaryPayload(deps.protocol, deps.model, system, content, stream2, includeMaxOutputTokens)),
       dispatcher: proxyDispatcher(deps.proxyUrl)
     },
     void 0,
@@ -53660,6 +53751,18 @@ async function* runCompressLoop(upstream, ctx, requestBody, requestOptions, adap
   }
 }
 
+// src/degenerate-turn.ts
+function degenerateTurnWarning(o) {
+  if ((o.reason ?? o.terminalReason) !== o.terminalReason) return null;
+  if (o.toolCalls > 0) return null;
+  if (o.text.outputChars > 0) return null;
+  const bits = [];
+  if (o.sawThinking) bits.push("thinking present");
+  if (o.text.dropped && o.text.inputChars > 0) bits.push(`${o.text.inputChars} chars of emitted text stripped as render-tag echo`);
+  else bits.push("no visible text emitted");
+  return `[degenerate-turn] ${o.wire}: turn ended ${o.terminalReason} with zero visible text and zero tool calls (${bits.join("; ")}) \u2014 the agent receives an empty turn and may stall until nudged (#673)`;
+}
+
 // src/loop/adapter-responses.ts
 var RESPONSES_ITEM_ID_MAX = 64;
 function normalizeResponsesMessageItems(input) {
@@ -53723,7 +53826,7 @@ function dropWhitespaceResponsesMessages(input) {
   }
   return dropped;
 }
-var RENDER_TAG_RE = /\x3cacp\s[^>]*\x3e[^<]*\x3c\/acp\x3e|\x3cacp\s[^>]*\/\x3e/g;
+var RENDER_TAG_RE = new RegExp("<" + ACP_NAME_ALT + "\\s[^>]*>[^<]*<\\/" + ACP_NAME_ALT + ">|<" + ACP_NAME_ALT + "\\s[^>]*\\/>", "g");
 function stripRenderTags(text) {
   return text.replace(RENDER_TAG_RE, "");
 }
@@ -53938,6 +54041,17 @@ data: ${JSON.stringify({ type: "response.created", response: { id: createdRespId
           yield { kind: "text", delta: tail, ...raw ? { raw } : {} };
         }
       };
+      let sawReasoning = false;
+      let toolCallsEmitted = 0;
+      let degenerateWarned = false;
+      const maybeWarnDegenerate = (reason) => {
+        if (degenerateWarned) return;
+        const msg = degenerateTurnWarning({ reason, terminalReason: "completed", toolCalls: toolCallsEmitted, text: tagFilter.stats(), sawThinking: sawReasoning, wire: "responses" });
+        if (msg) {
+          degenerateWarned = true;
+          log("warn", msg);
+        }
+      };
       for await (const eventStr of iterSseEvents(upstream)) {
         const explicitType = extractEventType(eventStr);
         const dataLine = extractDataLine(eventStr);
@@ -53950,6 +54064,7 @@ data: ${JSON.stringify({ type: "response.created", response: { id: createdRespId
         }
         const type = explicitType ?? (typeof obj.type === "string" ? obj.type : null);
         if (!type) continue;
+        if (type.startsWith("response.reasoning")) sawReasoning = true;
         const rawBuf = Buffer.from(eventStr + "\n\n", "utf8");
         if (round === 1 && typeof obj.output_index === "number") {
           outputIndex = Math.max(outputIndex, obj.output_index + 1);
@@ -54033,6 +54148,7 @@ data: ${JSON.stringify({ type: "response.created", response: { id: createdRespId
             if (fc) {
               if (typeof item.arguments === "string" && item.arguments) fc.arguments = item.arguments;
               pending.delete(itemId);
+              toolCallsEmitted++;
               yield {
                 kind: "tool_call",
                 name: fc.name,
@@ -54041,6 +54157,7 @@ data: ${JSON.stringify({ type: "response.created", response: { id: createdRespId
               };
             } else {
               yield { kind: "meta", chunk: rawBuf, firstRoundOnly: false };
+              toolCallsEmitted++;
               yield {
                 kind: "tool_call",
                 name: typeof item.name === "string" ? item.name : "",
@@ -54077,6 +54194,7 @@ data: ${JSON.stringify({ type: "response.created", response: { id: createdRespId
             outputTokens: typeof respUsage?.output_tokens === "number" ? respUsage.output_tokens : void 0,
             cachedTokens: typeof pd?.cached_tokens === "number" ? pd.cached_tokens : void 0
           };
+          maybeWarnDegenerate("completed");
           yield { kind: "done", finishReason: "completed" };
         } else if (type === "response.incomplete") {
           yield* flushFilter();
@@ -54394,11 +54512,23 @@ function createOpenaiAdapter(requestBody, clientSystem, hostCredit = 0, absorbNa
           yield { kind: "text", delta: tail, raw: buildContent(tail) };
         }
       };
+      let sawReasoning = false;
+      let toolCallsEmitted = 0;
+      let degenerateWarned = false;
+      const maybeWarnDegenerate = (reason) => {
+        if (degenerateWarned) return;
+        const msg = degenerateTurnWarning({ reason, terminalReason: "stop", toolCalls: toolCallsEmitted, text: tagFilter.stats(), sawThinking: sawReasoning, wire: "openai" });
+        if (msg) {
+          degenerateWarned = true;
+          log("warn", msg);
+        }
+      };
       const rawToolChunks = [];
       let sawRealToolCall = false;
       const flushPendingAsStructured = function* () {
         for (const [, tc] of pending) {
           if (tc.name.length > 0 || tc.id.length > 0) {
+            toolCallsEmitted++;
             yield {
               kind: "tool_call",
               name: tc.name,
@@ -54421,6 +54551,7 @@ function createOpenaiAdapter(requestBody, clientSystem, hostCredit = 0, absorbNa
         }
         for (const [idx, tc] of pending) {
           if (realIndexes.has(idx) && (tc.name.length > 0 || tc.id.length > 0)) {
+            toolCallsEmitted++;
             yield {
               kind: "tool_call",
               name: tc.name,
@@ -54440,6 +54571,7 @@ function createOpenaiAdapter(requestBody, clientSystem, hostCredit = 0, absorbNa
         }
         for (const [idx, tc] of pending) {
           if (!realIndexes.has(idx) && tc.name.length > 0) {
+            toolCallsEmitted++;
             yield {
               kind: "tool_call",
               name: tc.name,
@@ -54464,6 +54596,7 @@ function createOpenaiAdapter(requestBody, clientSystem, hostCredit = 0, absorbNa
           if (sawRealToolCall) {
             yield { kind: "meta", chunk: Buffer.from(eventStr + "\n\n", "utf8") };
           }
+          maybeWarnDegenerate("stop");
           yield { kind: "done", finishReason: "stop", ...sawRealToolCall ? { suppressCompletion: true } : {} };
           continue;
         }
@@ -54509,9 +54642,11 @@ function createOpenaiAdapter(requestBody, clientSystem, hostCredit = 0, absorbNa
           if (sawRealToolCall) {
             const chunk = patchUsageChunk(eventStr, parsed, u2 ?? {}, hostCredit);
             yield { kind: "meta", chunk };
+            maybeWarnDegenerate(finishReason);
             yield { kind: "done", finishReason, suppressCompletion: true };
             continue;
           } else {
+            maybeWarnDegenerate(finishReason);
             yield {
               kind: "done",
               finishReason: hadToolCalls && finishReason === "stop" ? "tool_calls" : finishReason
@@ -54520,6 +54655,7 @@ function createOpenaiAdapter(requestBody, clientSystem, hostCredit = 0, absorbNa
         }
         if (!delta) continue;
         if (typeof delta.reasoning_content === "string" && delta.reasoning_content.length > 0) {
+          sawReasoning = true;
           yield { kind: "reasoning", delta: delta.reasoning_content, raw: finishReason ? stripFinishReasonChunk(rawBuf) : rawBuf };
         }
         if (delta.tool_calls) {
@@ -54797,6 +54933,17 @@ ${systemPrompt}` : systemPrompt;
         log("warn", `[tag-echo] stripped model-emitted render tag: ${snippet.slice(0, 80).replace(/\n/g, " ")}`);
       });
       let lastTextIndex = null;
+      let sawThinking = false;
+      let toolCallsEmitted = 0;
+      let degenerateWarned = false;
+      const maybeWarnDegenerate = (reason) => {
+        if (degenerateWarned) return;
+        const msg = degenerateTurnWarning({ reason, terminalReason: "end_turn", toolCalls: toolCallsEmitted, text: tagFilter.stats(), sawThinking, wire: "anthropic" });
+        if (msg) {
+          degenerateWarned = true;
+          log("warn", msg);
+        }
+      };
       for await (const eventStr of iterSseEvents2(upstream)) {
         const parsed = parseAnthropicSse(eventStr);
         if (!parsed) continue;
@@ -54833,7 +54980,10 @@ ${systemPrompt}` : systemPrompt;
             const id = typeof block.id === "string" ? block.id : `toolu_${upstreamIndex}`;
             pending.set(upstreamIndex, { id, name, json: "" });
           } else {
-            if (block.type === "thinking" || block.type === "redacted_thinking") thinkingIndexes.add(upstreamIndex);
+            if (block.type === "thinking" || block.type === "redacted_thinking") {
+              thinkingIndexes.add(upstreamIndex);
+              sawThinking = true;
+            }
             const ci2 = clientIndex++;
             indexMap.set(upstreamIndex, ci2);
             openBlocks.push(ci2);
@@ -54878,6 +55028,7 @@ ${systemPrompt}` : systemPrompt;
           const tb = pending.get(upstreamIndex);
           if (tb) {
             pending.delete(upstreamIndex);
+            toolCallsEmitted++;
             yield {
               kind: "tool_call",
               name: tb.name,
@@ -54924,6 +55075,7 @@ ${systemPrompt}` : systemPrompt;
               cachedTokens: roundCached
             };
           }
+          maybeWarnDegenerate(stopReason);
           yield { kind: "done", finishReason: stopReason };
         } else if (type === "message_stop") {
           if (lastTextIndex !== null) {
@@ -54942,6 +55094,7 @@ ${systemPrompt}` : systemPrompt;
               cachedTokens: roundCached
             };
           }
+          maybeWarnDegenerate(stopReason);
           yield { kind: "done", finishReason: stopReason ?? "end_turn" };
         } else if (round === 1) {
           yield { kind: "meta", chunk: rawBuf, firstRoundOnly: true };
@@ -56552,11 +56705,18 @@ data: ${JSON.stringify({ type: "content_block_delta", index, delta: { type: delt
 
 `;
   };
+  let sawToolUse = false;
+  let sawThinking = false;
+  let visibleTextChars = 0;
+  let finalFinishReason;
   const flushTails = () => {
     let out = "";
     for (const s3 of streams.values()) {
       const tail = s3.filter.flush();
-      if (tail.length > 0) out += syntheticTail(s3.field, s3.index, tail);
+      if (tail.length > 0) {
+        out += syntheticTail(s3.field, s3.index, tail);
+        if (s3.field === "content" || s3.field === "text") visibleTextChars += tail.length;
+      }
     }
     return out;
   };
@@ -56579,6 +56739,28 @@ data: ${JSON.stringify({ type: "content_block_delta", index, delta: { type: delt
       reason: "plugin chat passthrough stream ended without a completion event"
     });
   };
+  const maybeWarnDegenerate = () => {
+    if (!sawTerminal || res.destroyed || res.writableEnded) return;
+    let inputChars = 0;
+    let dropped = false;
+    for (const s3 of streams.values()) {
+      const st2 = s3.filter.stats();
+      inputChars += st2.inputChars;
+      dropped = dropped || st2.dropped;
+    }
+    const msg = degenerateTurnWarning({
+      reason: finalFinishReason,
+      terminalReason: protocol === "anthropic" ? "end_turn" : "stop",
+      toolCalls: sawToolUse ? 1 : 0,
+      text: { inputChars, outputChars: visibleTextChars, dropped },
+      sawThinking,
+      wire: `plugin-passthrough-${protocol}`
+    });
+    if (msg) {
+      log("warn", msg);
+      log2?.(msg);
+    }
+  };
   const pushField = (field, index, text) => {
     const s3 = filterFor(field, index);
     const clean = s3.filter.push(text);
@@ -56598,21 +56780,28 @@ data: ${JSON.stringify({ type: "content_block_delta", index, delta: { type: delt
     let hadText = false;
     for (let ci2 = 0; ci2 < choices.length; ci2++) {
       const ch = choices[ci2];
+      if (ch && typeof ch["finish_reason"] === "string") finalFinishReason = ch["finish_reason"];
       const d = ch?.["delta"];
       if (!d || typeof d !== "object") continue;
       const dd = d;
+      if (dd["tool_calls"] !== void 0) sawToolUse = true;
       for (const field of ["content", "reasoning_content", "reasoning"]) {
         const v2 = dd[field];
         if (typeof v2 !== "string") continue;
         hadText = true;
+        if (field !== "content" && v2.length > 0) sawThinking = true;
         if (!mayStartRenderTag(v2) && !anyPending()) {
           if (v2.length > 0) keptText = true;
+          if (field === "content") visibleTextChars += v2.length;
           continue;
         }
         const index = typeof ch?.["index"] === "number" ? ch["index"] : ci2;
         const [clean, changed] = pushField(field, index, v2);
         if (clean.length === 0) droppedText = true;
-        else keptText = true;
+        else {
+          keptText = true;
+          if (field === "content") visibleTextChars += clean.length;
+        }
         if (changed) {
           if (!rebuilt) {
             rebuilt = { ...ev, choices: choices.map((c) => ({ ...c, delta: { ...c["delta"] } })) };
@@ -56631,6 +56820,13 @@ data: ${JSON.stringify({ type: "content_block_delta", index, delta: { type: delt
     return rawEvent + "\n\n";
   };
   const processAnthropic = (ev, rawEvent) => {
+    if (ev["type"] === "content_block_start") {
+      const cb = ev["content_block"];
+      const bt2 = cb && typeof cb === "object" ? cb["type"] : void 0;
+      if (bt2 === "tool_use") sawToolUse = true;
+      else if (bt2 === "thinking" || bt2 === "redacted_thinking") sawThinking = true;
+      return anyPending() ? flushTails() + rawEvent + "\n\n" : rawEvent + "\n\n";
+    }
     if (ev["type"] !== "content_block_delta") {
       return anyPending() ? flushTails() + rawEvent + "\n\n" : rawEvent + "\n\n";
     }
@@ -56641,11 +56837,17 @@ data: ${JSON.stringify({ type: "content_block_delta", index, delta: { type: delt
       return rawEvent + "\n\n";
     }
     const raw = d[field];
+    if (field === "thinking" && raw.length > 0) sawThinking = true;
     if (!mayStartRenderTag(raw) && !anyPending()) {
+      if (field === "text" && raw.length > 0) visibleTextChars += raw.length;
       return rawEvent + "\n\n";
     }
     const [clean, changed] = pushField(field, index, raw);
-    if (!changed) return rawEvent + "\n\n";
+    if (!changed) {
+      if (field === "text" && raw.length > 0) visibleTextChars += raw.length;
+      return rawEvent + "\n\n";
+    }
+    if (field === "text" && clean.length > 0) visibleTextChars += clean.length;
     if (clean.length === 0 && Object.keys(d ?? {}).length <= 2) return "";
     return rebuildEvent(rawEvent, { ...ev, delta: { ...d, [field]: clean } });
   };
@@ -56676,6 +56878,10 @@ data: ${JSON.stringify({ type: "content_block_delta", index, delta: { type: delt
             continue;
           }
           if (ev["type"] === "message_stop") sawTerminal = true;
+          if (ev["type"] === "message_delta") {
+            const d = ev["delta"];
+            if (d && typeof d["stop_reason"] === "string") finalFinishReason = d["stop_reason"];
+          }
           const sample = usageFromSseEvent(ev);
           if (sample) mergeUsageSample(acc, sample);
           let backfilled = false;
@@ -56699,6 +56905,7 @@ data: ${JSON.stringify({ type: "content_block_delta", index, delta: { type: delt
     if (rest.length > 0 && !res.destroyed && !res.writableEnded) await write(rest);
     settleUsage();
     maybeNoteTruncated();
+    maybeWarnDegenerate();
   } catch (e) {
     settleUsage();
     maybeNoteTruncated();
@@ -56735,6 +56942,9 @@ async function pipePluginResponsesWithStrip(stream2, res, session, log2) {
     log("warn", `[tag-echo] stripped model-emitted render tag (plugin passthrough): ${snippet.slice(0, 80).replace(/\n/g, " ")}`);
     log2?.(`[tag-echo] stripped model-emitted render tag from plugin passthrough text`);
   });
+  let sawFunctionCall = false;
+  let sawReasoning = false;
+  let responseStatus;
   const write = (s3) => {
     if (!res.write(Buffer.from(s3, "utf8"))) {
       return new Promise((r) => res.once("drain", () => r()));
@@ -56755,6 +56965,22 @@ async function pipePluginResponsesWithStrip(stream2, res, session, log2) {
       reason: "plugin responses passthrough stream ended without a completion event"
     });
   };
+  const maybeWarnDegenerate = () => {
+    if (!sawTerminal || res.destroyed || res.writableEnded) return;
+    const st2 = tagFilter.stats();
+    const msg = degenerateTurnWarning({
+      reason: responseStatus,
+      terminalReason: "completed",
+      toolCalls: sawFunctionCall ? 1 : 0,
+      text: st2,
+      sawThinking: sawReasoning,
+      wire: "plugin-passthrough-responses"
+    });
+    if (msg) {
+      log("warn", msg);
+      log2?.(msg);
+    }
+  };
   let lastDeltaMeta = null;
   const flushTail = (after) => {
     const tail = tagFilter.flush();
@@ -56794,6 +57020,16 @@ async function pipePluginResponsesWithStrip(stream2, res, session, log2) {
           const sample = usageFromSseEvent(ev);
           if (sample) mergeUsageSample(acc, sample);
           const type = ev["type"];
+          if (typeof type === "string") {
+            if (type.startsWith("response.reasoning")) sawReasoning = true;
+            if (type === "response.output_item.added" || type === "response.output_item.done") {
+              const item = ev["item"];
+              const it2 = item?.["type"];
+              if (it2 === "function_call" || it2 === "custom_tool_call") sawFunctionCall = true;
+            }
+            const resp = ev["response"];
+            if (resp && typeof resp["status"] === "string") responseStatus = resp["status"];
+          }
           if (type === "response.output_text.done" || type === "response.content_part.done" || type === "response.output_item.done" || type === "response.completed" || type === "response.failed" || type === "response.incomplete") {
             if (type === "response.completed" || type === "response.failed" || type === "response.incomplete") sawTerminal = true;
             let evOut = ev;
@@ -56839,6 +57075,7 @@ async function pipePluginResponsesWithStrip(stream2, res, session, log2) {
       const rest = flushTail("");
       if (rest.length > 0) await write(rest);
     }
+    maybeWarnDegenerate();
     settleUsage();
     maybeNoteTruncated();
   } catch (e) {
@@ -58914,6 +59151,7 @@ async function handle(req, res, opts, core, config, log2, instanceId, instanceSt
     try {
       const result = await fetchWithTimeout(targetUrl2, {
         method: "HEAD",
+        redirect: "follow",
         ...proxyUrl ? { dispatcher: proxyDispatcher(proxyUrl, 15e3) } : {}
       }, 15e3);
       result.clearTimer();
@@ -59110,7 +59348,11 @@ ${bodyBuffer.toString("utf8")}`);
     }
   }
   let prepared = null;
-  if (!opts.passthrough && hopMarker === void 0 && protocol && parsed && typeof parsed === "object") {
+  const routePassthrough = !opts.passthrough && findRoute(opts.routes, route?.rewrittenUrl)?.passthrough === true;
+  if (routePassthrough && hopMarker === void 0 && protocol && parsed) {
+    log2("info", `[route-passthrough] ${maskUrlsInText(route?.rewrittenUrl ?? "")} matches a passthrough route \u2014 forwarding verbatim, kernel bypassed`);
+  }
+  if (!opts.passthrough && !routePassthrough && hopMarker === void 0 && protocol && parsed && typeof parsed === "object") {
     const sessionHeader = headerValue2(req, opts.sessionHeader);
     let pluginAgent = pluginAgentHeader(req.headers);
     let pluginConversation = pluginConversationHeader(req.headers);
@@ -59371,7 +59613,7 @@ ${bodyBuffer.toString("utf8")}`);
     }
   }
   if (!prepared) {
-    if (protocol === null && !opts.passthrough && !isModelDiscoveryPath(urlPath)) {
+    if (protocol === null && !opts.passthrough && !routePassthrough && !isModelDiscoveryPath(urlPath)) {
       logUnrecognizedPath(log2, req.url ?? "");
     }
     await forward(req, res, opts, bodyBuffer, null, core, reqConfig, log2, route, instanceId, void 0);
@@ -59473,9 +59715,11 @@ function diagNudge(turn, sessionId, tokenCount, limit, model, willInject) {
   const modelTag = model ? ` model=${model}` : "";
   return `[${sessionId}] nudge ${inject}: usage=${pct2} (${tokenCount}/${limit}), growth=${growth}/${floor} (ref=${ref}, interval=${interval}), pendingT1=${pendingT1}/${interval}${modelTag}, reason="${n.reason.slice(0, 120)}"`;
 }
-function armHostUsageCredit(session, originalMessages, processedMessages, log2) {
+function armHostUsageCredit(session, originalMessages, processedMessages, headers, hostUsageCredit, log2) {
   session.hostCreditTokens = 0;
+  if (hostUsageCredit === "off") return;
   if (session.metadata.pluginAgent === "pi" || session.metadata.pluginAgent === "omp") return;
+  if (isCodexClient(headers)) return;
   session.hostCreditTokens = processedMessages.length > 0 ? Math.max(0, estimateCoreMessages(originalMessages) - estimateCoreMessages(processedMessages)) : 0;
   if (session.hostCreditTokens > 0) {
     log2("info", `[${session.id}] host usage backfill armed: +${session.hostCreditTokens} tok (forwarded view is folded); host usage will report the uncompressed baseline`);
@@ -59556,7 +59800,7 @@ function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, sessio
   markDirty(session);
   const rebuilt = { ...parsed, messages: rebuiltMessages, system: systemOut, tools: toolsOut };
   delete rebuilt.prompt_cache_key;
-  armHostUsageCredit(session, originalMessages, processedMessages, log2);
+  armHostUsageCredit(session, originalMessages, processedMessages, req.headers, opts.hostUsageCredit, log2);
   return { body: JSON.stringify(rebuilt), session, processedMessages, originalMessages, anthropicSystem: parsed.system, protocol: "anthropic", stream: stream2, compressInjected: injectTools, pluginMode, nudge, prompts, renderTags: "text-only" };
 }
 var OUTPUT_CLAMP_MARGIN_PCT = 0.05;
@@ -59688,7 +59932,7 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
   if (stream2 && rebuilt.stream_options === void 0) {
     rebuilt.stream_options = { include_usage: true };
   }
-  armHostUsageCredit(session, originalMessages, processedMessages, log2);
+  armHostUsageCredit(session, originalMessages, processedMessages, req.headers, opts.hostUsageCredit, log2);
   if (!isTitleGen && openaiOutboundSystem !== void 0) {
     session.metadata.systemPromptTokens = countSystemAndToolsTokens(openaiOutboundSystem, toolsOut);
   }
@@ -59840,7 +60084,7 @@ function prepareResponses(parsed, req, opts, core, config, prompts, log2, sessio
     });
     log2("info", `[${sessionId}] responses forward tools=[${fwdTools.join(",")}] injectTool=${injectTools}${pluginMode ? " (plugin mode: wire injection suppressed)" : ""} NO_INJECT_TOOL=${!!process.env.ACP_NO_INJECT_TOOL} NO_COMPRESS_PROMPT=${!!process.env.ACP_NO_COMPRESS_PROMPT}`);
   }
-  armHostUsageCredit(session, originalMessages, processedMessages, log2);
+  armHostUsageCredit(session, originalMessages, processedMessages, req.headers, opts.hostUsageCredit, log2);
   if (transformOk) {
     session.metadata.systemPromptTokens = countSystemAndToolsTokens(responsesDevContent ?? "", toolsOut);
   }