anthropic thinking

1 files changed, 81 insertions, 3 deletions

diff --git a/libpanto/src/config.zig b/libpanto/src/config.zig
index 5e01bb1..587f00a 100644
--- a/libpanto/src/config.zig
+++ b/libpanto/src/config.zig
@@ -41,6 +41,28 @@ pub const ReasoningEffort = enum {
     high,
 };
 
+/// Anthropic extended-thinking mode.
+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.
+    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+.
+    adaptive,
+};
+
+/// Effort level sent with Anthropic adaptive thinking (`thinking = .adaptive`).
+/// Ignored when `thinking` is `.enabled` or `.disabled`.
+pub const Effort = enum {
+    low,
+    medium,
+    high,
+    xhigh,
+    max,
+};
+
 pub const OpenAIChatConfig = struct {
     api_key: []const u8,
     base_url: []const u8,
@@ -57,6 +79,22 @@ pub const AnthropicMessagesConfig = struct {
     api_version: []const u8 = "2023-06-01",
     /// Required by Anthropic's Messages API.
     max_tokens: u32 = 64_000,
+    /// 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,
+    /// Effort level for adaptive thinking. Only emitted on the wire when
+    /// `thinking == .adaptive`; ignored otherwise.
+    effort: Effort = .medium,
+    /// Maximum tokens Claude may spend on internal reasoning when
+    /// `thinking == .enabled`. `null` falls back to `max_tokens - 1`.
+    /// Ignored when `thinking == .adaptive` or `.disabled`.
+    thinking_budget_tokens: ?u32 = 32_000,
+    /// When true and `thinking == .enabled`, sends the
+    /// `interleaved-thinking-2025-05-14` beta header so Claude can think
+    /// between tool calls. Ignored when `thinking == .adaptive` (interleaving
+    /// is automatic there) or `.disabled`.
+    thinking_interleaved: bool = false,
 };
 
 /// Per-provider transport/auth/model configuration. Tagged by `APIStyle`.
@@ -70,8 +108,8 @@ pub const ProviderConfig = union(APIStyle) {
 
     /// The wire-format provider identity for this config: the ground-truth
     /// `{api_style, base_url, model, reasoning}` that a turn is sent with.
-    /// Anthropic has no reasoning-effort knob, so it reports `.default`.
-    /// Borrowed slices; valid as long as the config is.
+    /// Anthropic carries thinking/effort/budget/interleaved instead of
+    /// reasoning. Borrowed slices; valid as long as the config is.
     pub fn wireIdentity(self: ProviderConfig) WireIdentity {
         return switch (self) {
             .openai_chat => |c| .{
@@ -84,7 +122,10 @@ pub const ProviderConfig = union(APIStyle) {
                 .api_style = .anthropic_messages,
                 .base_url = c.base_url,
                 .model = c.model,
-                .reasoning = .default,
+                .thinking = c.thinking,
+                .effort = c.effort,
+                .thinking_budget_tokens = c.thinking_budget_tokens,
+                .thinking_interleaved = c.thinking_interleaved,
             },
         };
     }
@@ -93,11 +134,24 @@ pub const ProviderConfig = union(APIStyle) {
 /// Wire-format provider identity (see `ProviderConfig.wireIdentity`). This
 /// is the same shape as `session_store.WireIdentity`; defined here to avoid
 /// a module cycle (config must not import session_store).
+///
+/// OpenAI uses `reasoning`; Anthropic uses `thinking`/`effort`/
+/// `thinking_budget_tokens`/`thinking_interleaved`. Fields unused by the
+/// active provider carry their zero/default values.
 pub const WireIdentity = struct {
     api_style: APIStyle,
     base_url: []const u8,
     model: []const u8,
+    /// OpenAI only.
     reasoning: ReasoningEffort = .default,
+    /// Anthropic only.
+    thinking: Thinking = .enabled,
+    /// Anthropic only; only meaningful when `thinking == .adaptive`.
+    effort: Effort = .medium,
+    /// Anthropic only; only meaningful when `thinking == .enabled`.
+    thinking_budget_tokens: ?u32 = 32_000,
+    /// Anthropic only; only meaningful when `thinking == .enabled`.
+    thinking_interleaved: bool = false,
 };
 
 /// Compaction settings the agent consults when summarizing old history.
@@ -220,6 +274,30 @@ 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(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);
+}
+
+test "ProviderConfig - anthropic_messages wireIdentity carries thinking fields" {
+    const cfg: ProviderConfig = .{ .anthropic_messages = .{
+        .api_key = "k",
+        .base_url = "https://api.anthropic.com",
+        .model = "claude-opus-4-8",
+        .thinking = .adaptive,
+        .effort = .high,
+        .thinking_budget_tokens = null,
+        .thinking_interleaved = false,
+    } };
+    const id = cfg.wireIdentity();
+    try t.expectEqual(APIStyle.anthropic_messages, id.api_style);
+    try t.expectEqual(Thinking.adaptive, id.thinking);
+    try t.expectEqual(Effort.high, id.effort);
+    try t.expectEqual(@as(?u32, null), id.thinking_budget_tokens);
+    try t.expectEqual(false, id.thinking_interleaved);
+    // reasoning field carries its zero default; not meaningful for Anthropic
+    try t.expectEqual(ReasoningEffort.default, id.reasoning);
 }
 
 test "ProviderConfig - openai_chat reasoning defaults to .default" {