billion-context 0.1.92 → 0.1.93-pr.631.452

package/dist/index.js

@@ -5632,7 +5632,7 @@ var require_util2 = __commonJS({
         callback();
       }
     };
-    function createInflate(zlibOptions) {
+    function createInflate2(zlibOptions) {
       return new InflateStream(zlibOptions);
     }
     function extractMimeType(headers) {
@@ -5764,7 +5764,7 @@ var require_util2 = __commonJS({
       readAllBytes,
       simpleRangeHeaderValue,
       buildContentRange,
-      createInflate,
+      createInflate: createInflate2,
       extractMimeType,
       getDecodeSplit,
       environmentSettingsObject,
@@ -13122,7 +13122,7 @@ var require_mock_interceptor = __commonJS({
 var require_mock_client = __commonJS({
   "node_modules/undici/lib/mock/mock-client.js"(exports, module) {
     "use strict";
-    var { promisify: promisify2 } = __require("util");
+    var { promisify } = __require("util");
     var Client = require_client();
     var { buildMockDispatch } = require_mock_utils();
     var {
@@ -13170,7 +13170,7 @@ var require_mock_client = __commonJS({
         this[kDispatches] = [];
       }
       async [kClose]() {
-        await promisify2(this[kOriginalClose])();
+        await promisify(this[kOriginalClose])();
         this[kConnected] = 0;
         this[kMockAgent][Symbols.kClients].delete(this[kOrigin]);
       }
@@ -13383,7 +13383,7 @@ var require_mock_call_history = __commonJS({
 var require_mock_pool = __commonJS({
   "node_modules/undici/lib/mock/mock-pool.js"(exports, module) {
     "use strict";
-    var { promisify: promisify2 } = __require("util");
+    var { promisify } = __require("util");
     var Pool2 = require_pool();
     var { buildMockDispatch } = require_mock_utils();
     var {
@@ -13431,7 +13431,7 @@ var require_mock_pool = __commonJS({
         this[kDispatches] = [];
       }
       async [kClose]() {
-        await promisify2(this[kOriginalClose])();
+        await promisify(this[kOriginalClose])();
         this[kConnected] = 0;
         this[kMockAgent][Symbols.kClients].delete(this[kOrigin]);
       }
@@ -17545,17 +17545,17 @@ var require_cache2 = __commonJS({
 var require_decompress = __commonJS({
   "node_modules/undici/lib/interceptor/decompress.js"(exports, module) {
     "use strict";
-    var { createInflate, createGunzip, createBrotliDecompress, createZstdDecompress } = __require("zlib");
+    var { createInflate: createInflate2, createGunzip: createGunzip2, createBrotliDecompress: createBrotliDecompress2, createZstdDecompress } = __require("zlib");
     var { pipeline } = __require("stream");
     var DecoratorHandler = require_decorator_handler();
     var { runtimeFeatures } = require_runtime_features();
     var supportedEncodings = {
-      gzip: createGunzip,
-      "x-gzip": createGunzip,
-      br: createBrotliDecompress,
-      deflate: createInflate,
-      compress: createInflate,
-      "x-compress": createInflate,
+      gzip: createGunzip2,
+      "x-gzip": createGunzip2,
+      br: createBrotliDecompress2,
+      deflate: createInflate2,
+      compress: createInflate2,
+      "x-compress": createInflate2,
       ...runtimeFeatures.has("zstd") ? { zstd: createZstdDecompress } : {}
     };
     var defaultSkipStatusCodes = (
@@ -20384,7 +20384,7 @@ var require_fetch = __commonJS({
       clampAndCoarsenConnectionTimingInfo,
       simpleRangeHeaderValue,
       buildContentRange,
-      createInflate,
+      createInflate: createInflate2,
       extractMimeType,
       hasAuthenticationEntry,
       includesCredentials,
@@ -21366,7 +21366,7 @@ var require_fetch = __commonJS({
                       finishFlush: zlib.constants.Z_SYNC_FLUSH
                     }));
                   } else if (coding === "deflate") {
-                    decoders.push(createInflate({
+                    decoders.push(createInflate2({
                       flush: zlib.constants.Z_SYNC_FLUSH,
                       finishFlush: zlib.constants.Z_SYNC_FLUSH
                     }));
@@ -23432,7 +23432,7 @@ var require_connection = __commonJS({
 var require_permessage_deflate = __commonJS({
   "node_modules/undici/lib/web/websocket/permessage-deflate.js"(exports, module) {
     "use strict";
-    var { createInflateRaw, Z_DEFAULT_WINDOWBITS } = __require("zlib");
+    var { createInflateRaw: createInflateRaw2, Z_DEFAULT_WINDOWBITS } = __require("zlib");
     var { isValidClientWindowBits } = require_util5();
     var { MessageSizeExceededError } = require_errors();
     var tail = Buffer.from([0, 0, 255, 255]);
@@ -23468,7 +23468,7 @@ var require_permessage_deflate = __commonJS({
             windowBits = Number.parseInt(this.#options.serverMaxWindowBits);
           }
           try {
-            this.#inflate = createInflateRaw({ windowBits });
+            this.#inflate = createInflateRaw2({ windowBits });
           } catch (err2) {
             callback(err2);
             return;
@@ -38339,7 +38339,7 @@ var require_tls = __commonJS({
       }
       return rval;
     };
-    var inflate2 = function(c, record, s3) {
+    var inflate = function(c, record, s3) {
       var rval = false;
       try {
         var bytes = c.inflate(record.fragment.getBytes());
@@ -39484,7 +39484,7 @@ var require_tls = __commonJS({
           case tls4.CompressionMethod.none:
             break;
           case tls4.CompressionMethod.deflate:
-            state.read.compressFunction = inflate2;
+            state.read.compressFunction = inflate;
             state.write.compressFunction = deflate;
             break;
           default:
@@ -46740,6 +46740,74 @@ function renderCompressedDrilldown(blocks, state, sort, limit, countTokens) {
   }
   return lines.join("\n");
 }
+function renderMessage2(m2) {
+  const parts = [];
+  switch (m2.contentType) {
+    case "text":
+      parts.push(m2.text ?? "");
+      break;
+    case "tool-call":
+      parts.push(`\`${m2.toolName ?? "?"}(${m2.toolCallId ?? ""})\` args: ${m2.text ?? ""}`);
+      break;
+    case "tool-result":
+      parts.push(`\`${m2.toolName ?? "?"}(${m2.toolCallId ?? ""})\` \u2192 ${m2.text ?? ""}`);
+      break;
+    case "reasoning":
+      parts.push(`_reasoning_: ${m2.text ?? ""}`);
+      break;
+  }
+  const body = parts.join("\n").trim();
+  return body === "" ? "_(empty)_" : body + "\n";
+}
+function renderHandoff(input) {
+  const { coreMessages, state, full, meta } = input;
+  const lines = [];
+  lines.push("# billion-context session handoff");
+  lines.push("");
+  lines.push(`- title: ${meta.title ?? "(untitled)"}`);
+  if (meta.label) lines.push(`- label: ${meta.label}`);
+  lines.push(`- session id: ${meta.sessionId}`);
+  for (const bullet of meta.extraBullets ?? []) lines.push(bullet);
+  if (meta.contextTokens) lines.push(`- last context tokens: ~${meta.contextTokens}`);
+  lines.push(`- compression blocks: ${state.blocks.length} (active ${state.blocks.filter((b2) => b2.active).length})`);
+  lines.push("");
+  const folded = input.folded === true;
+  const view = full || folded ? coreMessages : prune(coreMessages, state);
+  lines.push(full && !folded ? `## Full conversation (${coreMessages.length} messages)` : folded ? `## Conversation (persisted folded snapshot, ${coreMessages.length} messages)` : `## Conversation (folded view as the model saw it, ${coreMessages.length} client messages)`);
+  lines.push("");
+  if (view.length === 0) {
+    lines.push("No conversation messages to export.");
+    lines.push("");
+  }
+  let lastRole = "";
+  for (const m2 of view) {
+    if (m2.role !== lastRole) {
+      lines.push(`### ${m2.role}`);
+      lines.push("");
+      lastRole = m2.role;
+    }
+    lines.push(renderMessage2(m2));
+  }
+  lines.push("");
+  if (full && folded) {
+    for (const b2 of input.blocksFull ?? []) {
+      lines.push(`## Block ${b2.blockId}${b2.topic ? ` \u2014 ${b2.topic}` : ""}`);
+      lines.push("");
+      lines.push(`### Original messages (${b2.count})`);
+      lines.push("");
+      lines.push(b2.fullText.trim());
+      lines.push("");
+    }
+  }
+  return lines.join("\n");
+}
+function matchSession(sessions2, selector, labelOf) {
+  const exact = sessions2.filter((s3) => s3.id === selector);
+  if (exact.length > 0) return exact;
+  const byLabel = sessions2.filter((s3) => labelOf(s3) === selector);
+  if (byLabel.length > 0) return byLabel;
+  return sessions2.filter((s3) => s3.id.startsWith(selector) || (labelOf(s3) ?? "").startsWith(selector));
+}
 function parseCompressArgs(input, opts) {
   const callId = opts?.callId;
   const diag = { ok: false, kind: "ok", invalidItems: 0 };
@@ -48293,6 +48361,7 @@ function loadOptions(env = process.env) {
     passthrough: passthrough.enabled,
     passthroughSource: passthrough.source,
     autoUpdate: (env.ACP_AUTO_UPDATE ?? (fileConfig.autoUpdate === false ? "0" : "1")) !== "0",
+    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: {
       enabled: (env.BILI_MITM ?? (fileConfig.mitm?.enabled === false ? "0" : "1")) !== "0",
@@ -50329,6 +50398,7 @@ function buildAlertMessage(dir, count) {
 // src/persist.ts
 var PERSIST_VERSION = 3;
 var MIGRATION_MARKER = ".bili-migration-286.done";
+var STALE_WARN_THROTTLE_MS = 6e4;
 function mergeState(parsed) {
   const fresh = createInitialState();
   return {
@@ -50362,11 +50432,14 @@ var SessionStore = class {
   enabled;
   dir;
   store;
+  log;
+  staleWarnAt = /* @__PURE__ */ new Map();
   constructor(opts) {
     const debounceMs = opts?.debounceMs ?? defaultDebounce();
     this.enabled = (opts?.enabled ?? true) && debounceMs >= 0;
     this.dir = opts?.dir ?? defaultDir();
     const baseLog = opts?.log ?? defaultLogger;
+    this.log = baseLog;
     const epermAlert = new PersistEpermAlert({
       dir: this.dir,
       threshold: epermAlertThreshold(),
@@ -50516,40 +50589,83 @@ var SessionStore = class {
       });
     }
   }
+  /** Shared envelope probe for the sync read paths: namespaced path
+   *  first, then the _unknown/ location, letting the kernel store layer
+   *  also check its discovered map and the flat legacy name. */
+  loadEnvelope(id, meta) {
+    const envelopes = [
+      this.store.loadSync(id, relPathFor(id, meta?.protocol, meta?.upstreamOrigin)),
+      meta?.protocol ? this.store.loadSync(id, relPathFor(id)) : null
+    ];
+    for (const envelope of envelopes) {
+      if (envelope) return envelope;
+    }
+    return null;
+  }
   /** Synchronous reload of a single session. Used on a memory miss (after
    *  LRU eviction). Sync fs is acceptable here because a miss is rare and
    *  reads a single small file (~1ms). Returns null if missing/corrupt or the
    *  body id does not match what we asked for. */
   loadSync(id, meta) {
     if (!this.enabled) return null;
-    const envelopes = [
-      this.store.loadSync(id, relPathFor(id, meta?.protocol, meta?.upstreamOrigin)),
-      meta?.protocol ? this.store.loadSync(id, relPathFor(id)) : null
-    ];
-    for (const envelope of envelopes) {
-      if (envelope) {
-        const session = buildSession(envelope.payload);
-        if (hasNegativePersistedTokens(envelope.payload)) {
-          this.scheduleSave(session);
-          log("info", `[persist] clamped negative token stats on reload for ${id} (#408)`);
-        }
-        return session;
-      }
+    const envelope = this.loadEnvelope(id, meta);
+    if (!envelope) return null;
+    const session = buildSession(envelope.payload);
+    if (hasNegativePersistedTokens(envelope.payload)) {
+      this.scheduleSave(session);
+      log("info", `[persist] clamped negative token stats on reload for ${id} (#408)`);
+    }
+    return session;
+  }
+  /** #405 fix #4: dual-instance rollback guard. When two proxy processes
+   *  share BILI_SESSIONS_DIR, whoever saves last used to win — an instance
+   *  holding a STALE in-memory copy would roll counters back (requests:3 →
+   *  2). Both signals below are monotonic per session id, so either being
+   *  strictly smaller proves staleness. Returns the fresher-on-disk payload
+   *  when the incoming record is stale, else null. */
+  staleDiskPayload(incoming) {
+    const envelope = this.loadEnvelope(incoming.id, incoming.meta);
+    if (!envelope) return null;
+    const disk = envelope.payload;
+    const diskRequests = disk.stats?.requests ?? disk.requests ?? 0;
+    const incRequests = incoming.stats?.requests ?? incoming.requests ?? 0;
+    if (incRequests < diskRequests) return disk;
+    if (incRequests === diskRequests) {
+      const diskBlocks = disk.state?.nextBlockId ?? 0;
+      const incBlocks = incoming.state?.nextBlockId ?? 0;
+      if (incBlocks < diskBlocks) return disk;
     }
     return null;
   }
+  guardedBuild(session) {
+    return () => {
+      const record = buildRecord(session);
+      const disk = this.staleDiskPayload(record);
+      if (disk === null) return record;
+      const now = Date.now();
+      const last = this.staleWarnAt.get(record.id) ?? 0;
+      if (now - last >= STALE_WARN_THROTTLE_MS) {
+        this.staleWarnAt.set(record.id, now);
+        this.log(
+          "warn",
+          `[persist] rejected stale snapshot for session ${record.id}: in-memory copy is older than the on-disk one (another bili instance holds newer state) \u2014 keeping disk state, no rollback (#405)`
+        );
+      }
+      return disk;
+    };
+  }
   /** Schedule a debounced write for a session. Multiple calls within the
    *  window coalesce; the record is built at WRITE time, so the freshest
    *  session state is persisted. Safe to call on the hot path. No-op if
    *  disabled. */
   scheduleSave(session) {
-    this.store.scheduleSave(session.id, () => buildRecord(session));
+    this.store.scheduleSave(session.id, this.guardedBuild(session));
   }
   /** Asynchronously persist a session right now (skips the debounce). Throws
    *  on write failure so callers can react (e.g. avoid evicting). Serialized
    *  per-session by the kernel store's write chains. */
   async writeNow(session) {
-    await this.store.writeNow(session.id, () => buildRecord(session));
+    await this.store.writeNow(session.id, this.guardedBuild(session));
   }
   /** Synchronous flush for a single session. Used on memory eviction so a
    *  dirty evicted session is not lost. Sync because eviction runs in the
@@ -50557,7 +50673,7 @@ var SessionStore = class {
    *  Returns true on success, false on failure (caller must NOT evict on
    *  failure for a never-persisted session or it is lost permanently). */
   flushSync(session) {
-    return this.store.flushSync(session.id, () => buildRecord(session));
+    return this.store.flushSync(session.id, this.guardedBuild(session));
   }
   /** Flush all dirty sessions with a pending debounce timer. Called on
    *  SIGTERM/SIGINT for graceful shutdown. The kernel store flushes its own
@@ -57416,8 +57532,7 @@ function tunnelAllowlistFromEnv(env = process.env) {
 }
 
 // src/content-encoding.ts
-import { promisify } from "util";
-import { brotliDecompress, gunzip, inflate, inflateRaw } from "zlib";
+import { createBrotliDecompress, createGunzip, createInflate, createInflateRaw } from "zlib";
 
 // node_modules/fzstd/esm/index.mjs
 var ab = ArrayBuffer;
@@ -58090,31 +58205,54 @@ var Decompress = /* @__PURE__ */ (function() {
 })();
 
 // src/content-encoding.ts
-var gunzipAsync = promisify(gunzip);
-var inflateAsync = promisify(inflate);
-var inflateRawAsync = promisify(inflateRaw);
-var brotliAsync = promisify(brotliDecompress);
-async function decodeZstd(input, maxOutputBytes) {
+var DecompressedTooLargeError = class extends Error {
+  constructor(limit) {
+    super(`decompressed request exceeds ${limit} bytes`);
+    this.limit = limit;
+    this.name = "DecompressedTooLargeError";
+  }
+  limit;
+};
+function streamDecode(factory, input, max) {
+  return new Promise((resolve, reject) => {
+    const chunks = [];
+    let size = 0;
+    const decoder = factory();
+    decoder.on("data", (chunk) => {
+      size += chunk.length;
+      if (size > max) {
+        decoder.destroy();
+        reject(new DecompressedTooLargeError(max));
+        return;
+      }
+      chunks.push(chunk);
+    });
+    decoder.on("end", () => resolve(Buffer.concat(chunks, size)));
+    decoder.on("error", (err2) => reject(err2));
+    decoder.end(input);
+  });
+}
+async function decodeZstd(input, max) {
   const chunks = [];
   let size = 0;
   const decoder = new Decompress((chunk) => {
     size += chunk.byteLength;
-    if (size > maxOutputBytes) throw new Error(`decompressed request exceeds ${maxOutputBytes} bytes`);
+    if (size > max) throw new DecompressedTooLargeError(max);
     chunks.push(Buffer.from(chunk));
   });
   decoder.push(input, true);
   return Buffer.concat(chunks, size);
 }
-async function decodeOne(coding, input, maxOutputBytes) {
-  const options = { maxOutputLength: maxOutputBytes };
-  if (coding === "gzip" || coding === "x-gzip") return gunzipAsync(input, options);
-  if (coding === "br") return brotliAsync(input, options);
-  if (coding === "zstd") return decodeZstd(input, maxOutputBytes);
+async function decodeOne(coding, input, max) {
+  if (coding === "gzip" || coding === "x-gzip") return streamDecode(createGunzip, input, max);
+  if (coding === "br") return streamDecode(createBrotliDecompress, input, max);
+  if (coding === "zstd") return decodeZstd(input, max);
   if (coding === "deflate") {
     try {
-      return await inflateAsync(input, options);
-    } catch {
-      return inflateRawAsync(input, options);
+      return await streamDecode(createInflate, input, max);
+    } catch (err2) {
+      if (err2 instanceof DecompressedTooLargeError) throw err2;
+      return streamDecode(createInflateRaw, input, max);
     }
   }
   throw new Error(`unsupported request content-encoding: ${coding}`);
@@ -58125,9 +58263,7 @@ async function decodeRequestBody(contentEncoding, input, maxOutputBytes) {
   let body = input;
   for (const coding of codings.reverse()) {
     body = await decodeOne(coding, body, maxOutputBytes);
-    if (body.byteLength > maxOutputBytes) {
-      throw new Error(`decompressed request exceeds ${maxOutputBytes} bytes`);
-    }
+    if (body.byteLength > maxOutputBytes) throw new DecompressedTooLargeError(maxOutputBytes);
   }
   return { body, decoded: true };
 }
@@ -58707,18 +58843,24 @@ async function handle(req, res, opts, core, config, log2, instanceId, instanceSt
     upstreamOrigin = route ? route.upstream : /^https?:\/\//i.test(url) ? new URL(url).origin : opts.upstream;
     protocol = route?.explicitProtocol ?? (req.method === "POST" && bodyBuffer.length > 0 ? urlPath.endsWith("/chat/completions") ? "openai" : urlPath.endsWith("/v1/messages") || urlPath.endsWith("/messages") ? "anthropic" : urlPath.endsWith("/responses") || responsesCompact ? "responses" : null : null);
     if (protocol !== null && bodyBuffer.length > 0) {
-      const decoded = await decodeRequestBody(headerValue2(req, "content-encoding"), bodyBuffer, MAX_REQUEST_BYTES);
-      bodyBuffer = decoded.body;
-      if (decoded.decoded) delete req.headers["content-encoding"];
+      try {
+        const decoded = await decodeRequestBody(headerValue2(req, "content-encoding"), bodyBuffer, MAX_REQUEST_BYTES);
+        bodyBuffer = decoded.body;
+        if (decoded.decoded) delete req.headers["content-encoding"];
+      } catch (decErr) {
+        if (decErr instanceof DecompressedTooLargeError) throw decErr;
+        protocol = null;
+        log2("warn", `decode body failed (${String(decErr)}) - forwarding raw body verbatim to ${upstreamOrigin}`);
+      }
     }
   } catch (err2) {
-    if (err2 instanceof BodyTooLargeError) {
+    if (err2 instanceof BodyTooLargeError || err2 instanceof DecompressedTooLargeError) {
       log2("warn", `413: request body exceeds ${err2.limit} bytes`);
       res.writeHead(413, { "content-type": "application/json" });
       res.end(JSON.stringify({ error: { type: "request_too_large", message: err2.message } }));
       return;
     }
-    log2("warn", `read/decode body failed: ${String(err2)}`);
+    log2("warn", `failed to prepare inbound request (${String(err2)}) - 400`);
     res.writeHead(400, { "content-type": "application/json" });
     res.end(JSON.stringify({ error: { type: "invalid_request", message: String(err2) } }));
     return;
@@ -63583,6 +63725,12 @@ var To = (s3) => {
 import path12 from "path";
 import { fileURLToPath as fileURLToPath3 } from "url";
 var REGISTRY_BASE = "https://registry.npmjs.org";
+function normalizeUpdateTag(tag) {
+  return (tag ?? "latest").trim() || "latest";
+}
+function registryUrlFor(packageName, tag) {
+  return `${REGISTRY_BASE}/${packageName}/${encodeURIComponent(tag)}`;
+}
 var CHECK_INTERVAL_MS = 3 * 60 * 1e3;
 var THROTTLE_FILE = path12.join(cacheDir(), ".update-check");
 var LOCK_FILE = path12.join(cacheDir(), ".update-lock");
@@ -63593,22 +63741,42 @@ function shouldStealLock(holderAlive, ageMs) {
 var timer2;
 var inFlight = false;
 var firstCheckDone = false;
-function parseVersion(v2) {
-  return v2.replace(/^v/, "").split(".").map((n) => parseInt(n, 10) || 0);
-}
-function isNewer(latest, current) {
-  const l = parseVersion(latest);
-  const c = parseVersion(current);
+function isVersionNewer(latest, current) {
+  const next = parseSemVer(latest);
+  const prev = parseSemVer(current);
+  if (!next || !prev) return false;
   for (let i = 0; i < 3; i++) {
-    const lv = l[i] ?? 0;
-    const cv = c[i] ?? 0;
-    if (lv > cv) return true;
-    if (lv < cv) return false;
+    const a = next.parts[i] ?? 0;
+    const b2 = prev.parts[i] ?? 0;
+    if (a !== b2) return a > b2;
+  }
+  if (!next.pre.length && prev.pre.length) return true;
+  if (next.pre.length && !prev.pre.length) return false;
+  for (let i = 0; i < Math.max(next.pre.length, prev.pre.length); i++) {
+    const a = next.pre[i];
+    const b2 = prev.pre[i];
+    if (a === void 0) return false;
+    if (b2 === void 0) return true;
+    if (a === b2) continue;
+    const aNumber = /^\d+$/.test(a) ? Number(a) : void 0;
+    const bNumber = /^\d+$/.test(b2) ? Number(b2) : void 0;
+    if (aNumber !== void 0 && bNumber !== void 0) return aNumber > bNumber;
+    if (aNumber !== void 0) return false;
+    if (bNumber !== void 0) return true;
+    return a > b2;
   }
   return false;
 }
+function parseSemVer(version2) {
+  const match = version2.match(/^v?(\d+)\.(\d+)\.(\d+)(?:-([0-9A-Za-z.-]+))?(?:\+.+)?$/);
+  if (!match) return void 0;
+  return {
+    parts: [Number(match[1]), Number(match[2]), Number(match[3])],
+    pre: match[4]?.split(".") ?? []
+  };
+}
 function staleInstallStatus(diskVersion, runningVersion) {
-  return diskVersion !== void 0 && isNewer(diskVersion, runningVersion) ? "restart" : "current";
+  return diskVersion !== void 0 && isVersionNewer(diskVersion, runningVersion) ? "restart" : "current";
 }
 async function readLastCheck() {
   try {
@@ -63814,7 +63982,8 @@ async function checkForUpdate(opts, force = false) {
       return;
     }
     log("info", `[update] checking npm registry for ${opts.packageName}${sinceLastSec < 0 ? " (startup check)" : sinceLastSec === 0 ? "" : ` (last check ${sinceLastSec}s ago)`}\u2026`);
-    const url = `${REGISTRY_BASE}/${opts.packageName}/latest`;
+    const tag = normalizeUpdateTag(opts.updateTag);
+    const url = registryUrlFor(opts.packageName, tag);
     const registryDispatcher = egressDispatcher(opts, url);
     const res = await fetchWithEgress(url, {
       signal: AbortSignal.timeout(5e3),
@@ -63833,11 +64002,11 @@ async function checkForUpdate(opts, force = false) {
     }
     const diskVersion = installDir ? await readDiskVersion(installDir) : void 0;
     const currentVersion = diskVersion ?? opts.currentVersion;
-    if (!isNewer(latest, currentVersion)) {
+    if (!isVersionNewer(latest, currentVersion)) {
       if (staleInstallStatus(diskVersion, opts.currentVersion) === "restart") {
         log("warn", `[update] running v${opts.currentVersion} but v${diskVersion} is installed \u2014 restart bili to activate (auto-update replaced the on-disk install; this process is still on the old code)`);
       } else {
-        log("info", `[update] current=${currentVersion} latest=${latest} (up to date)`);
+        log("info", `[update] current=${currentVersion} latest=${latest} tag=${tag} (up to date)`);
       }
       return;
     }
@@ -65902,7 +66071,33 @@ function latestBlockTime(s3) {
   for (const b2 of s3.state.blocks) if (b2.createdAt > latest) latest = b2.createdAt;
   return latest;
 }
-function renderHandoff(s3, full) {
+function renderHandoff2(s3, full) {
+  const messages = s3.lastMessages;
+  if (messages && messages.length > 0) {
+    const folded = s3.lastMessagesFolded === true;
+    const blocksFull = full && folded ? s3.state.blocks.filter((b2) => b2.active).map((b2) => {
+      const content = s3.blockContents.get(b2.blockId);
+      return content ? { blockId: b2.blockId, topic: b2.topic, count: content.full.count, fullText: content.full.text } : void 0;
+    }).filter((x) => x !== void 0) : void 0;
+    return renderHandoff({
+      coreMessages: messages,
+      state: s3.state,
+      full,
+      folded,
+      blocksFull,
+      meta: {
+        title: s3.meta.title,
+        label: s3.meta.label,
+        sessionId: s3.id,
+        contextTokens: s3.stats.contextTokens,
+        extraBullets: [
+          ...s3.meta.protocol ? [`- protocol: ${s3.meta.protocol}`] : [],
+          ...s3.meta.upstreamOrigin ? [`- upstream: ${s3.meta.upstreamOrigin}`] : [],
+          `- requests: ${s3.stats.requests}`
+        ]
+      }
+    });
+  }
   const lines = [];
   lines.push(`# billion-context session handoff`);
   lines.push("");
@@ -65915,38 +66110,6 @@ function renderHandoff(s3, full) {
   if (s3.stats.contextTokens) lines.push(`- last context tokens: ~${s3.stats.contextTokens}`);
   lines.push(`- compression blocks: ${s3.state.blocks.length} (active ${s3.state.blocks.filter((b2) => b2.active).length})`);
   lines.push("");
-  const messages = s3.lastMessages;
-  if (messages && messages.length > 0) {
-    const folded = s3.lastMessagesFolded === true;
-    const view = full || folded ? messages : prune(messages, s3.state);
-    lines.push(
-      full && !folded ? `## Full conversation (${messages.length} messages)` : folded ? `## Conversation (persisted folded snapshot, ${messages.length} messages)` : `## Conversation (folded view as the model saw it, ${messages.length} client messages)`
-    );
-    lines.push("");
-    let lastRole = "";
-    for (const m2 of view) {
-      if (m2.role !== lastRole) {
-        lines.push(`### ${m2.role}`);
-        lines.push("");
-        lastRole = m2.role;
-      }
-      lines.push(renderMessage2(m2));
-    }
-    lines.push("");
-    if (full && folded) {
-      for (const b2 of s3.state.blocks.filter((x) => x.active)) {
-        const content = s3.blockContents.get(b2.blockId);
-        if (!content) continue;
-        lines.push(`## Block ${b2.blockId}${b2.topic ? ` \u2014 ${b2.topic}` : ""}`);
-        lines.push("");
-        lines.push(`### Original messages (${content.full.count})`);
-        lines.push("");
-        lines.push(content.full.text.trim());
-        lines.push("");
-      }
-    }
-    return lines.join("\n");
-  }
   const active = s3.state.blocks.filter((b2) => b2.active);
   if (active.length === 0) {
     lines.push("No active compression blocks and no persisted conversation snapshot (v2 session file). Original messages are only persisted when they are compressed into a block, so this session's conversation content is not available for export.");
@@ -65975,32 +66138,8 @@ function renderHandoff(s3, full) {
   }
   return lines.join("\n");
 }
-function renderMessage2(m2) {
-  const parts = [];
-  switch (m2.contentType) {
-    case "text":
-      parts.push(m2.text ?? "");
-      break;
-    case "tool-call":
-      parts.push(`\`${m2.toolName ?? "?"}(${m2.toolCallId ?? ""})\` args: ${m2.text ?? ""}`);
-      break;
-    case "tool-result":
-      parts.push(`\`${m2.toolName ?? "?"}(${m2.toolCallId ?? ""})\` \u2192 ${m2.text ?? ""}`);
-      break;
-    case "reasoning":
-      parts.push(`_reasoning_: ${m2.text ?? ""}`);
-      break;
-  }
-  const body = parts.join("\n").trim();
-  return body === "" ? "_(empty)_" : body + "\n";
-}
-function matchSession(sessions2, selector) {
-  const exact = sessions2.filter((s3) => s3.id === selector);
-  if (exact.length > 0) return exact;
-  const byLabel = sessions2.filter((s3) => s3.meta.label === selector);
-  if (byLabel.length > 0) return byLabel;
-  const byPrefix = sessions2.filter((s3) => s3.id.startsWith(selector) || (s3.meta.label ?? "").startsWith(selector));
-  return byPrefix;
+function matchSession2(sessions2, selector) {
+  return matchSession(sessions2, selector, (s3) => s3.meta.label);
 }
 async function exportSession(selector, opts = {}) {
   const store = new SessionStore({ dir: opts.dir, enabled: true });
@@ -66015,7 +66154,7 @@ async function exportSession(selector, opts = {}) {
     );
     return ["Persisted sessions:", "", ...rows.map((r) => `  ${r}`), "", "Usage: bili export <session-id|label> [--output handoff.md] [--full]"].join("\n");
   }
-  const matches = matchSession(all, selector);
+  const matches = matchSession2(all, selector);
   if (matches.length === 0) {
     throw new Error(`no session matches "${selector}" (run "bili export" to list sessions)`);
   }
@@ -66023,7 +66162,7 @@ async function exportSession(selector, opts = {}) {
     const ids = matches.map((s3) => s3.id).join(", ");
     throw new Error(`selector "${selector}" matches ${matches.length} sessions (${ids}); use the full session id`);
   }
-  const markdown = renderHandoff(matches[0], opts.full ?? false);
+  const markdown = renderHandoff2(matches[0], opts.full ?? false);
   if (opts.output) {
     mkdirSync6(path16.dirname(path16.resolve(opts.output)), { recursive: true });
     writeFileSync5(opts.output, markdown, "utf8");
@@ -66348,14 +66487,16 @@ async function main() {
       if (v2 !== void 0) process.env[k2] = v2;
     }
     let updaterResolveProxy;
+    let updateTag;
     try {
       const o = loadOptions();
       updaterResolveProxy = (url) => resolveProxy(o.routes, o.proxy, url, o.proxyFallback);
+      updateTag = o.updateTag;
     } catch (e) {
       console.error(`bili update: config load failed (${String(e)}); updater egress goes direct`);
     }
     await checkForUpdate(
-      { packageName: PACKAGE_NAME, currentVersion: VERSION3, autoUpdate: true, resolveProxy: updaterResolveProxy },
+      { packageName: PACKAGE_NAME, currentVersion: VERSION3, autoUpdate: true, resolveProxy: updaterResolveProxy, updateTag },
       true
     );
     return;
@@ -66383,7 +66524,8 @@ async function main() {
       packageName: PACKAGE_NAME,
       currentVersion: VERSION3,
       autoUpdate: true,
-      resolveProxy: (url) => resolveProxy(opts.routes, opts.proxy, url, opts.proxyFallback)
+      resolveProxy: (url) => resolveProxy(opts.routes, opts.proxy, url, opts.proxyFallback),
+      updateTag: opts.updateTag
     });
   }
 }