1 files changed, 192 insertions, 0 deletions
diff --git a/libpanto/src/anthropic_messages_json.zig b/libpanto/src/anthropic_messages_json.zig
index bccf692..78f4b5f 100644
--- a/libpanto/src/anthropic_messages_json.zig
+++ b/libpanto/src/anthropic_messages_json.zig
@@ -52,6 +52,41 @@ pub fn serializeRequest(
     try s.objectField("stream");
     try s.write(true);
 
+    // Extended thinking configuration.
+    // `.disabled`  — omit the field entirely.
+    // `.enabled`   — manual budget: { type, budget_tokens, display }.
+    // `.adaptive`  — adaptive mode: { type, display } + top-level effort.
+    switch (cfg.thinking) {
+        .disabled => {},
+        .enabled => {
+            // Resolve budget: explicit value or fall back to max_tokens - 1.
+            // Clamp so budget is always strictly less than max_tokens (required
+            // by Anthropic when not using interleaved thinking).
+            const raw_budget: u32 = cfg.thinking_budget_tokens orelse (cfg.max_tokens -| 1);
+            const budget: u32 = if (raw_budget >= cfg.max_tokens) cfg.max_tokens -| 1 else raw_budget;
+            try s.objectField("thinking");
+            try s.beginObject();
+            try s.objectField("type");
+            try s.write("enabled");
+            try s.objectField("budget_tokens");
+            try s.write(budget);
+            try s.objectField("display");
+            try s.write("summarized");
+            try s.endObject();
+        },
+        .adaptive => {
+            try s.objectField("thinking");
+            try s.beginObject();
+            try s.objectField("type");
+            try s.write("adaptive");
+            try s.objectField("display");
+            try s.write("summarized");
+            try s.endObject();
+            try s.objectField("effort");
+            try s.write(@tagName(cfg.effort));
+        },
+    }
+
     // Build and emit the concatenated system prompt, if any.
     var system_buf: std.ArrayList(u8) = .empty;
     defer system_buf.deinit(allocator);
@@ -1134,6 +1169,163 @@ test "serializeRequest - tool result with image part emits image block" {
     try testing.expectEqualStrings("iVBOR==", src.get("data").?.string);
 }
 
+test "serializeRequest - thinking disabled omits thinking and effort fields" {
+    const allocator = testing.allocator;
+    var conv = conversation.Conversation.init(allocator);
+    defer conv.deinit();
+    try conv.addUserMessage("hi");
+
+    var cfg = testConfig("claude-x");
+    cfg.thinking = .disabled;
+    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 root = parsed.value.object;
+    try testing.expect(root.get("thinking") == null);
+    try testing.expect(root.get("effort") == null);
+}
+
+test "serializeRequest - thinking enabled explicit budget" {
+    const allocator = testing.allocator;
+    var conv = conversation.Conversation.init(allocator);
+    defer conv.deinit();
+    try conv.addUserMessage("hi");
+
+    var cfg = testConfig("claude-x"); // max_tokens = 1024
+    cfg.thinking = .enabled;
+    cfg.thinking_budget_tokens = 500; // within max_tokens
+    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 root = parsed.value.object;
+    const th = root.get("thinking").?.object;
+    try testing.expectEqualStrings("enabled", th.get("type").?.string);
+    try testing.expectEqual(@as(i64, 500), th.get("budget_tokens").?.integer);
+    try testing.expectEqualStrings("summarized", th.get("display").?.string);
+    try testing.expect(root.get("effort") == null);
+}
+
+test "serializeRequest - thinking enabled null budget falls back to max_tokens - 1" {
+    const allocator = testing.allocator;
+    var conv = conversation.Conversation.init(allocator);
+    defer conv.deinit();
+    try conv.addUserMessage("hi");
+
+    var cfg = testConfig("claude-x"); // max_tokens = 1024
+    cfg.thinking = .enabled;
+    cfg.thinking_budget_tokens = null;
+    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 th = parsed.value.object.get("thinking").?.object;
+    try testing.expectEqual(@as(i64, 1023), th.get("budget_tokens").?.integer);
+}
+
+test "serializeRequest - thinking enabled budget clamped when >= max_tokens" {
+    const allocator = testing.allocator;
+    var conv = conversation.Conversation.init(allocator);
+    defer conv.deinit();
+    try conv.addUserMessage("hi");
+
+    var cfg = testConfig("claude-x"); // max_tokens = 1024
+    cfg.thinking = .enabled;
+    cfg.thinking_budget_tokens = 2_000; // > max_tokens
+    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 th = parsed.value.object.get("thinking").?.object;
+    // Clamped to max_tokens - 1 = 1023
+    try testing.expectEqual(@as(i64, 1023), th.get("budget_tokens").?.integer);
+}
+
+test "serializeRequest - thinking adaptive default effort" {
+    const allocator = testing.allocator;
+    var conv = conversation.Conversation.init(allocator);
+    defer conv.deinit();
+    try conv.addUserMessage("hi");
+
+    var cfg = testConfig("claude-x");
+    cfg.thinking = .adaptive;
+    // effort defaults to .medium
+    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 root = parsed.value.object;
+    const th = root.get("thinking").?.object;
+    try testing.expectEqualStrings("adaptive", th.get("type").?.string);
+    try testing.expectEqualStrings("summarized", th.get("display").?.string);
+    try testing.expect(th.get("budget_tokens") == null);
+    try testing.expectEqualStrings("medium", root.get("effort").?.string);
+}
+
+test "serializeRequest - thinking adaptive explicit effort levels" {
+    const allocator = testing.allocator;
+    const efforts = [_]struct { e: config_mod.Effort, name: []const u8 }{
+        .{ .e = .low, .name = "low" },
+        .{ .e = .medium, .name = "medium" },
+        .{ .e = .high, .name = "high" },
+        .{ .e = .xhigh, .name = "xhigh" },
+        .{ .e = .max, .name = "max" },
+    };
+    for (efforts) |entry| {
+        var conv = conversation.Conversation.init(allocator);
+        defer conv.deinit();
+        try conv.addUserMessage("hi");
+
+        var cfg = testConfig("claude-x");
+        cfg.thinking = .adaptive;
+        cfg.effort = entry.e;
+        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();
+        try testing.expectEqualStrings(entry.name, parsed.value.object.get("effort").?.string);
+    }
+}
+
+test "serializeRequest - thinking adaptive ignores budget_tokens" {
+    const allocator = testing.allocator;
+    var conv = conversation.Conversation.init(allocator);
+    defer conv.deinit();
+    try conv.addUserMessage("hi");
+
+    var cfg = testConfig("claude-x");
+    cfg.thinking = .adaptive;
+    cfg.thinking_budget_tokens = 9_999; // should be ignored
+    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 th = parsed.value.object.get("thinking").?.object;
+    try testing.expect(th.get("budget_tokens") == null);
+}
+
 test "serializeRequest - empty ToolUse input becomes {} not invalid JSON" {
     const allocator = testing.allocator;