fix anthropic thinking effort json placement

4 files changed, 28 insertions, 8 deletions

diff --git a/libpanto/src/anthropic_messages_json.zig b/libpanto/src/anthropic_messages_json.zig
index 3ad6069..99d89eb 100644
--- a/libpanto/src/anthropic_messages_json.zig
+++ b/libpanto/src/anthropic_messages_json.zig
@@ -55,7 +55,7 @@ pub fn serializeRequest(
     // Extended thinking configuration.
     // `.disabled`  — omit the field entirely.
     // `.enabled`   — manual budget: { type, budget_tokens, display }.
-    // `.adaptive`  — adaptive mode: { type, display } + top-level effort.
+    // `.adaptive`  — adaptive mode: { type, display } + output_config.effort.
     switch (cfg.thinking) {
         .disabled => {},
         .enabled => {
@@ -82,8 +82,11 @@ pub fn serializeRequest(
             try s.objectField("display");
             try s.write("summarized");
             try s.endObject();
+            try s.objectField("output_config");
+            try s.beginObject();
             try s.objectField("effort");
             try s.write(@tagName(cfg.effort));
+            try s.endObject();
         },
     }
 
@@ -1285,7 +1288,9 @@ test "serializeRequest - thinking adaptive default effort" {
     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);
+    try testing.expect(root.get("effort") == null);
+    const oc = root.get("output_config").?.object;
+    try testing.expectEqualStrings("medium", oc.get("effort").?.string);
 }
 
 test "serializeRequest - thinking adaptive explicit effort levels" {
@@ -1312,7 +1317,8 @@ test "serializeRequest - thinking adaptive explicit effort levels" {
 
         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);
+        const oc = parsed.value.object.get("output_config").?.object;
+        try testing.expectEqualStrings(entry.name, oc.get("effort").?.string);
     }
 }
 
diff --git a/libpanto/src/config.zig b/libpanto/src/config.zig
index 587f00a..4aae81f 100644
--- a/libpanto/src/config.zig
+++ b/libpanto/src/config.zig
@@ -46,7 +46,8 @@ pub const Thinking = enum {
     /// Do not request extended thinking.
     disabled,
     /// Manual extended thinking with an explicit token budget (`thinking_budget_tokens`).
-    /// Supported on all current models including Haiku 4.5. Default.
+    /// Supported on Haiku 4.5 and older models, but NOT on Opus 4.8+ (which
+    /// only accepts adaptive). Not safe as a cross-model default.
     enabled,
     /// Adaptive thinking: Claude decides when and how much to think.
     /// Requires Opus 4.6 / Sonnet 4.6 or newer; automatic on Opus 4.7+.
@@ -82,7 +83,12 @@ pub const AnthropicMessagesConfig = struct {
     /// Extended-thinking mode. `.enabled` sends a manual thinking block with
     /// `thinking_budget_tokens`; `.adaptive` lets Claude decide and uses
     /// `effort` instead. `.disabled` omits thinking entirely.
-    thinking: Thinking = .enabled,
+    ///
+    /// Defaults to `.disabled`: it is the only mode accepted by every current
+    /// model. `.enabled` fails on Opus 4.8+ (adaptive-only) and `.adaptive`
+    /// fails on Haiku 4.5 (no adaptive support), so neither is a safe default
+    /// without parsing model strings.
+    thinking: Thinking = .disabled,
     /// Effort level for adaptive thinking. Only emitted on the wire when
     /// `thinking == .adaptive`; ignored otherwise.
     effort: Effort = .medium,
@@ -274,7 +280,7 @@ test "ProviderConfig - anthropic_messages variant" {
     try t.expectEqual(APIStyle.anthropic_messages, cfg.style());
     try t.expectEqualStrings("2023-06-01", cfg.anthropic_messages.api_version);
     try t.expectEqual(@as(u32, 64_000), cfg.anthropic_messages.max_tokens);
-    try t.expectEqual(Thinking.enabled, cfg.anthropic_messages.thinking);
+    try t.expectEqual(Thinking.disabled, cfg.anthropic_messages.thinking);
     try t.expectEqual(Effort.medium, cfg.anthropic_messages.effort);
     try t.expectEqual(@as(?u32, 32_000), cfg.anthropic_messages.thinking_budget_tokens);
     try t.expectEqual(false, cfg.anthropic_messages.thinking_interleaved);
diff --git a/libpanto/src/provider_anthropic_messages.zig b/libpanto/src/provider_anthropic_messages.zig
index 37ce818..9740c0e 100644
--- a/libpanto/src/provider_anthropic_messages.zig
+++ b/libpanto/src/provider_anthropic_messages.zig
@@ -124,6 +124,10 @@ pub const AnthropicMessagesRequest = struct {
         rr.response = try rr.req.receiveHead(&redirect_buf);
 
         if (@intFromEnum(rr.response.head.status) >= 400) {
+            // `head.bytes` (which `iterateHeaders` walks) points into the
+            // connection read buffer and is invalidated the moment the body
+            // stream is initialized below. Capture Retry-After first.
+            const retry_after_ms = provider_mod.retryAfterFromHead(rr.response.head);
             const body_reader = rr.response.reader(&rr.transfer_buf);
             var err_buf: std.ArrayList(u8) = .empty;
             defer err_buf.deinit(self.allocator);
@@ -143,7 +147,7 @@ pub const AnthropicMessagesRequest = struct {
             const classified = provider_mod.classifyHttpStatus(status, err_buf.items);
             if (self.diag) |d| {
                 d.status_code = status;
-                d.retry_after_ms = provider_mod.retryAfterFromHead(rr.response.head);
+                d.retry_after_ms = retry_after_ms;
             }
             return classified;
         }
diff --git a/libpanto/src/provider_openai_chat.zig b/libpanto/src/provider_openai_chat.zig
index 8319d9d..5152d3e 100644
--- a/libpanto/src/provider_openai_chat.zig
+++ b/libpanto/src/provider_openai_chat.zig
@@ -140,6 +140,10 @@ pub const OpenAIChatRequest = struct {
         rr.response = try rr.req.receiveHead(&redirect_buf);
 
         if (@intFromEnum(rr.response.head.status) >= 400) {
+            // `head.bytes` (which `iterateHeaders` walks) points into the
+            // connection read buffer and is invalidated the moment the body
+            // stream is initialized below. Capture Retry-After first.
+            const retry_after_ms = provider_mod.retryAfterFromHead(rr.response.head);
             // Drain body for diagnostics.
             const body_reader = rr.response.reader(&rr.transfer_buf);
             var err_buf: std.ArrayList(u8) = .empty;
@@ -159,7 +163,7 @@ pub const OpenAIChatRequest = struct {
             const classified = provider_mod.classifyHttpStatus(status, err_buf.items);
             if (self.diag) |d| {
                 d.status_code = status;
-                d.retry_after_ms = provider_mod.retryAfterFromHead(rr.response.head);
+                d.retry_after_ms = retry_after_ms;
             }
             return classified;
         }