From 8009df42aae1ce4ac947d0ad3ae3addb5546e821 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Christian=20Schr=C3=B6der?=
 <1188270+cwschroeder@users.noreply.github.com>
Date: Sun, 18 Oct 2026 02:54:40 -0700
Subject: [PATCH] fix(server): keep late research context refusals streamed

A streamed research turn can start, prepare a compressed context and then
refuse at the final context-window check before the response sink opens. The
run's catch block treated every unopened failure as a whole-reply error, so
that refusal came back as a JSON status even though the request was a stream
and the same refusal on the opened path is streamed.

Route error.Refused with a context-window depth through the streamed context
error when the sink has not opened. Every other unopened failure keeps the
existing whole-reply path.

The regression drives both research windows that prepare a compressed prompt
against the stub engine: the search runs, the response opens, no whole-reply
status or body follows, the stream carries context_length_exceeded, and the
preparing count drains to zero. With the openai.zig hunk removed the test
fails at cortheon_test.zig:369 on `try std.testing.expect(got.opened)`, so it
covers the failure it corrects.
---
 zig/src/server/cortheon/cortheon_test.zig | 23 +++++++++++++++++++++++
 zig/src/server/openai.zig                 |  5 ++++-
 2 files changed, 27 insertions(+), 1 deletion(-)

diff --git a/zig/src/server/cortheon/cortheon_test.zig b/zig/src/server/cortheon/cortheon_test.zig
index 15413efef1..250126f2ff 100644
--- a/zig/src/server/cortheon/cortheon_test.zig
+++ b/zig/src/server/cortheon/cortheon_test.zig
@@ -354,3 +354,26 @@ test "the web budget ends the research and the reply still comes back without Co
     try std.testing.expect(m.get("tool_calls") == null);
     try std.testing.expect(std.mem.indexOf(u8, m.strField("content").?, "cortheon_search") == null);
 }
+
+test "research expansion keeps a context-window refusal streamed" {
+    for ([_]u32{ 2000, 2400 }) |window| {
+        var rig: Rig = .{ .eng = .{ .script = &.{"Zig latest release notes"} } };
+        try rig.start(true);
+        defer rig.stop();
+        rig.srv.info.context_window = window;
+        var arena: std.heap.ArenaAllocator = .init(gpa);
+        defer arena.deinit();
+        const got = try rig.ask(arena.allocator(), "{\"stream\":true,\"max_tokens\":1,\"messages\":[{\"role\":\"user\",\"content\":\"What is the latest release of Zig?\"}]}");
+        try std.testing.expect(rig.web.searches > 0);
+        try std.testing.expectEqual(@as(usize, 1), rig.eng.prompts.items.len);
+        try std.testing.expect(got.opened);
+        try std.testing.expectEqual(@as(?u16, null), got.status);
+        try std.testing.expect(got.body == null);
+        var context_error = false;
+        for (got.events.items) |event| if (event.get("error")) |err| {
+            if (err.strField("code")) |code| context_error = std.mem.eql(u8, code, "context_length_exceeded");
+        };
+        try std.testing.expect(context_error);
+        try std.testing.expectEqual(@as(i64, 0), rig.srv.preparing.load(.acquire));
+    }
+}
diff --git a/zig/src/server/openai.zig b/zig/src/server/openai.zig
index 5b73d55f2c..60f6b2b843 100644
--- a/zig/src/server/openai.zig
+++ b/zig/src/server/openai.zig
@@ -521,7 +521,10 @@ const Run = struct {
         live.out = r.out;
         live.every_ns = @intFromFloat(r.srv.config.sse_keepalive_s * std.time.ns_per_s);
         var reply = (if (r.turn) |t| cortheon.loop.stream(t, r.srv, &cx, &r.plan.input, prepared, sink, live) else chat.generate(r.srv, &cx, prepared, sink, live)) catch |e| {
-            if (!sink_state.opened) return r.unsent(&cx, e, field);
+            if (!sink_state.opened) {
+                if (e == error.Refused and cx.kind == .context_length) return r.streamedContext(&cx, tools, field);
+                return r.unsent(&cx, e, field);
+            }
             switch (e) {
                 error.Cancelled => return,
                 error.Refused => if (cx.kind != .other and cx.kind != .server) {