diff --git a/.agents/docs/mcppls-0.0.17-release.md b/.agents/docs/mcppls-0.0.17-release.md
index c4d52a8a..7b4f4709 100644
--- a/.agents/docs/mcppls-0.0.17-release.md
+++ b/.agents/docs/mcppls-0.0.17-release.md
@@ -358,3 +358,9 @@ U13 和 two-core 也走同样的方式：U13 每边 3 次（只跑 long 阶段
   - termux、Zed（默认、推荐设置）、self-hosting 4 项、三个平台的 startup timing、最新版 mcpp 的 fixture 都通过。
 
   这次 U13 能过，是因为窗口内发生了 trim；分组现象没有变，所以 Known issues 保持不变。
+- **第一次 Release 运行（`40022574908`）在四个平台的 payload job 上都失败**：
+  - 原因：GitHub 对 release 下载的 302 响应里，现在带一行 9243 字节的 `content-security-policy`；packaging 用的 tinyhttps 0.3 一行 header 超过 8192 字节就放弃（`Incomplete headers: line exceeds 8192 bytes`）；
+  - PR CI 没有暴露：它的 payload 缓存里已经有这些包，不需要下载；
+  - 修复：PR #70，遇到这一种错误时改用 curl 下载，校验方式不变；server 不使用 tinyhttps，所以只影响发布流水线；
+  - 根本修复应在 tinyhttps（放宽行长限制），之后去掉这个回退；
+  - 这次失败的运行没有创建 tag，也没有创建 release。
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 86bc2455..2947de3f 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -133,6 +133,9 @@ what an idle or a crashed clangd costs you, and typing imports with autosave on.
   is not available with Clang at all (Clang has no coroutines on 32-bit x86 MSVC, so the MSVC STL's `std` module
   cannot be built), and such a project is read with the semantic kit (docs/20-projects.md). A conformance fixture that
   needs a newer mcpp than CI's pinned one declares it (`requires-mcpp`).
+- Packaging downloads a GitHub release asset with curl when its HTTP client (tinyhttps 0.3) cannot read GitHub's
+  answer: since 2026-10 the redirect carries a 9.2 KB header line, over the client's 8 KB limit. The download is
+  verified against the lock the same way. This is the release pipeline only; the server makes no such downloads.
 - **A change can be measured without being built again.** A manual nightly can test the payloads a pull request's CI
   already built (`payload-run`), run only the ux rounds (`scope`), run their edits stage several times in new server
   sessions (`ux-repeat`), and run a released server beside them on the same machines (`ux-baseline`), compared by
diff --git a/modules/pack/src/fetch.cpp b/modules/pack/src/fetch.cpp
index 78d62c47..8fa62a91 100644
--- a/modules/pack/src/fetch.cpp
+++ b/modules/pack/src/fetch.cpp
@@ -10,7 +10,9 @@ import mcppls.base.error;
 import mcppls.base.log;
 import mcppls.base.path;
 import mcppls.base.sha256;
+import mcppls.platform.env;
 import mcppls.platform.fs;
+import mcppls.platform.process;
 
 namespace mcppls::pack::fetch {
 namespace fs = mcppls::platform::fs;
@@ -30,6 +32,27 @@ mcpplibs::tinyhttps::HttpClientConfig client_config() {
 
 std::string partial_name(std::string_view target) { return std::string { target } + ".partial"; }
 
+// GitHub's answer to a release download -- a redirect to its asset host -- carries a content-security-policy
+// line of 9.2 KB (measured 2026-10-24); tinyhttps 0.3 gives up on any header line over 8 KB. curl, on every CI
+// runner and developer machine, takes such a download instead, verified the same way. Remove when tinyhttps
+// takes longer lines.
+bool header_line_too_long(std::string_view error) { return error.contains("line exceeds"); }
+
+// The download with curl, into `partial`; empty when it worked, else why not.
+std::string download_with_curl(const std::string& url, const std::string& partial) {
+    const auto curl = platform::env::find_executable("curl");
+    if (!curl) return "curl is not on PATH";
+    platform::SpawnOptions spawn;
+    spawn.program = *curl;
+    spawn.arguments = { "--fail", "--silent", "--show-error", "--location", "--connect-timeout", "30", "--output", partial, url };
+    spawn.pipeError = true;
+    const auto ran = platform::run(std::move(spawn), std::chrono::minutes { 30 });
+    if (!ran) return ran.error().message;
+    if (ran->timedOut) return "curl did not finish in time";
+    if (ran->exitCode != 0) return std::format("curl exited {}: {}", ran->exitCode, platform::last_lines(ran->error, 3));
+    return {};
+}
+
 void remove_quietly(const std::string& path) {
     std::error_code ignored;
     std::filesystem::remove(path, ignored);
@@ -75,16 +98,29 @@ base::Result<std::string> get(const Entry& entry, std::string_view cacheDir, int
         remove_quietly(partial);
 
         const auto answer = client.download_to_file(entry.url, std::filesystem::path { partial });
-        if (!answer.ok()) {
-            lastError = answer.error.empty() ? std::format("HTTP {}", answer.statusCode) : answer.error;
+        std::uint64_t written { static_cast<std::uint64_t>(answer.bytesWritten) };
+        std::string failed;
+        if (!answer.ok()) failed = answer.error.empty() ? std::format("HTTP {}", answer.statusCode) : answer.error;
+        if (!failed.empty() && header_line_too_long(failed)) {
+            if (!quiet) base::log::info("fetch: {} ({}); downloading it with curl", entry.url, failed);
+            remove_quietly(partial);
+            if (const auto why = download_with_curl(entry.url, partial); why.empty()) {
+                failed.clear();
+                written = fs::stamp(partial) ? fs::stamp(partial)->size : 0;
+            } else {
+                failed = std::format("{}; with curl: {}", failed, why);
+            }
+        }
+        if (!failed.empty()) {
+            lastError = failed;
         } else if (auto got = digest_of(partial); !got) {
             lastError = got.error().message;
         } else if (*got != entry.sha256) {
             lastError = std::format("sha256 mismatch for {}: the lock says {}, the download is {}",
                                     entry.file, entry.sha256, *got);
-        } else if (entry.size && static_cast<std::uint64_t>(answer.bytesWritten) != *entry.size) {
+        } else if (entry.size && written != *entry.size) {
             lastError = std::format("size mismatch for {}: the lock says {} bytes, the download is {}",
-                                    entry.file, *entry.size, answer.bytesWritten);
+                                    entry.file, *entry.size, written);
         } else {
             std::error_code failed;
             std::filesystem::rename(partial, target, failed);