preserve signature origins across compactions

1 files changed, 76 insertions, 3 deletions

diff --git a/libpanto/src/anthropic_messages_json.zig b/libpanto/src/anthropic_messages_json.zig
index 1b1cf6e..5f119f3 100644
--- a/libpanto/src/anthropic_messages_json.zig
+++ b/libpanto/src/anthropic_messages_json.zig
@@ -262,7 +262,7 @@ fn writeMessage(
     try s.beginArray();
     for (msg.content.items, 0..) |block, bi| {
         const mark = if (cache_block) |cb| cb == bi else false;
-        try writeBlock(s, block, mark, cfg);
+        try writeBlock(s, block, mark, cfg, msg.identity);
     }
     try s.endArray();
 
@@ -284,6 +284,7 @@ fn writeBlock(
     block: conversation.ContentBlock,
     mark_cache: bool,
     cfg: *const config_mod.AnthropicMessagesConfig,
+    msg_identity: ?config_mod.WireIdentity,
 ) !void {
     switch (block) {
         .Text => |tb| {
@@ -304,8 +305,7 @@ fn writeBlock(
             // messages, and opaque signatures are not portable across
             // provider/model boundaries.
             const sig = tb.signature orelse return;
-            const origin = tb.signature_origin orelse return;
-            if (!origin.matches(.anthropic_messages, cfg.base_url, cfg.model)) return;
+            if (!conversation.thinkingSignatureMatches(tb, msg_identity, .anthropic_messages, cfg.base_url, cfg.model)) return;
             try s.beginObject();
             try s.objectField("type");
             try s.write("thinking");
@@ -819,6 +819,79 @@ test "serializeRequest - cache breakpoint skips a trailing thinking block" {
     try testing.expect(content[1].object.get("cache_control") == null);
 }
 
+test "serializeRequest - thinking signature replays from message identity when block origin is absent" {
+    // A reloaded thinking block need not carry its own signature_origin: the
+    // enclosing message's identity is the per-message source of truth, so a
+    // host can rely on it alone (no per-block identity copy required).
+    const allocator = testing.allocator;
+
+    var conv = conversation.Conversation.init(allocator);
+    defer conv.deinit();
+    try conv.addAssistantMessage(&.{
+        .{
+            .Thinking = .{
+                .text = try conversation.textualBlockFromSlice(allocator, "reasoned"),
+                .signature = try allocator.dupe(u8, "EqQBCgIYAhIM1gbcDa9GJwZA"),
+                // No signature_origin on the block.
+            },
+        },
+        .{ .Text = try conversation.textualBlockFromSlice(allocator, "answer") },
+    }, null);
+    // Per-message identity matching the request.
+    conv.messages.items[0].identity = try conversation.dupeWireIdentity(allocator, .{
+        .api_style = .anthropic_messages,
+        .base_url = "u",
+        .model = "claude-x",
+    });
+
+    const cfg = testConfig("claude-x");
+    var empty_tools = tool_registry_mod.ToolRegistry.init(allocator);
+    defer empty_tools.deinit();
+    const body = try serializeRequest(allocator, &cfg, &conv, &empty_tools);
+    defer allocator.free(body);
+
+    var parsed = try std.json.parseFromSlice(std.json.Value, allocator, body, .{});
+    defer parsed.deinit();
+    const content = parsed.value.object.get("messages").?.array.items[0]
+        .object.get("content").?.array.items;
+    try testing.expectEqualStrings("thinking", content[0].object.get("type").?.string);
+    try testing.expectEqualStrings("EqQBCgIYAhIM1gbcDa9GJwZA", content[0].object.get("signature").?.string);
+}
+
+test "serializeRequest - thinking dropped when message identity targets a different model" {
+    const allocator = testing.allocator;
+
+    var conv = conversation.Conversation.init(allocator);
+    defer conv.deinit();
+    try conv.addAssistantMessage(&.{
+        .{ .Thinking = .{
+            .text = try conversation.textualBlockFromSlice(allocator, "reasoned"),
+            .signature = try allocator.dupe(u8, "EqQBCgIYAhIM1gbcDa9GJwZA"),
+        } },
+        .{ .Text = try conversation.textualBlockFromSlice(allocator, "answer") },
+    }, null);
+    // Identity points at a *different* model than the request.
+    conv.messages.items[0].identity = try conversation.dupeWireIdentity(allocator, .{
+        .api_style = .anthropic_messages,
+        .base_url = "u",
+        .model = "some-other-model",
+    });
+
+    const cfg = testConfig("claude-x");
+    var empty_tools = tool_registry_mod.ToolRegistry.init(allocator);
+    defer empty_tools.deinit();
+    const body = try serializeRequest(allocator, &cfg, &conv, &empty_tools);
+    defer allocator.free(body);
+
+    var parsed = try std.json.parseFromSlice(std.json.Value, allocator, body, .{});
+    defer parsed.deinit();
+    const content = parsed.value.object.get("messages").?.array.items[0]
+        .object.get("content").?.array.items;
+    // Only the text block survives; the unportable thinking block is dropped.
+    try testing.expectEqual(@as(usize, 1), content.len);
+    try testing.expectEqualStrings("text", content[0].object.get("type").?.string);
+}
+
 test "serializeRequest - prompt_cache disabled omits all cache_control" {
     const allocator = testing.allocator;