From 02b4c7a35ac0b714bc045d54fb4bb45d1ce4e490 Mon Sep 17 00:00:00 2001
From: t <t@tjp.lol>
Date: Sat, 13 Jun 2026 23:45:59 -0600
Subject: Add Codex Responses support and session debugging

Teach provider config and auth resolution about the Codex Responses
dialect, including user-facing `style = "openai_responses"` with
`dialect = "codex"`. Serialize and parse Responses traffic with
provider-specific reasoning replay, assistant phase metadata, and robust
function-call assembly keyed by `output_index` so streamed tool inputs
survive proxy quirks and empty terminal payloads.

Also persist thinking origins and message metadata across sessions, add
the Anthropic interleaved-thinking header switch, write per-session
debug logs, and improve the TUI and scripts for inspecting tool output
and session costs.
---
 libpanto/src/openai_responses_json.zig | 319 +++++++++++++++++++++++++++++++--
 1 file changed, 304 insertions(+), 15 deletions(-)

(limited to 'libpanto/src/openai_responses_json.zig')

diff --git a/libpanto/src/openai_responses_json.zig b/libpanto/src/openai_responses_json.zig
index 8ff7712..295e79c 100644
--- a/libpanto/src/openai_responses_json.zig
+++ b/libpanto/src/openai_responses_json.zig
@@ -30,6 +30,11 @@ const tool_registry_mod = @import("tool_registry.zig");
 // Request serialization
 // ===========================================================================
 
+pub const RequestDialect = enum {
+    public,
+    codex,
+};
+
 /// Serialize a Conversation into a `/responses` request body. Caller owns the
 /// returned slice.
 pub fn serializeRequest(
@@ -37,6 +42,7 @@ pub fn serializeRequest(
     cfg: *const config_mod.OpenAIResponsesConfig,
     conv: *const conversation.Conversation,
     tools: *const tool_registry_mod.ToolRegistry,
+    dialect: RequestDialect,
 ) ![]u8 {
     var aw: Writer.Allocating = .init(allocator);
     errdefer aw.deinit();
@@ -54,8 +60,24 @@ pub fn serializeRequest(
     try s.objectField("store");
     try s.write(false);
 
-    try s.objectField("max_output_tokens");
-    try s.write(cfg.max_tokens);
+    if (dialect == .public) {
+        try s.objectField("max_output_tokens");
+        try s.write(cfg.max_tokens);
+    }
+
+    if (dialect == .codex) {
+        try s.objectField("text");
+        try s.beginObject();
+        try s.objectField("verbosity");
+        try s.write("low");
+        try s.endObject();
+
+        try s.objectField("tool_choice");
+        try s.write("auto");
+
+        try s.objectField("parallel_tool_calls");
+        try s.write(true);
+    }
 
     // Carry encrypted reasoning so multi-turn reasoning can continue without
     // server-side storage.
@@ -110,7 +132,7 @@ pub fn serializeRequest(
     try s.beginArray();
     for (conversation.activeMessageWindow(conv.messages.items)) |msg| {
         if (msg.role == .system) continue;
-        try writeInputForMessage(&s, msg, allocator);
+        try writeInputForMessage(&s, msg, allocator, cfg, dialect);
     }
     try s.endArray();
 
@@ -129,7 +151,13 @@ fn writeReasoning(s: *std.json.Stringify, effort: []const u8) !void {
 }
 
 /// Emit the `input` item(s) for one conversation message.
-fn writeInputForMessage(s: *std.json.Stringify, msg: conversation.Message, allocator: Allocator) !void {
+fn writeInputForMessage(
+    s: *std.json.Stringify,
+    msg: conversation.Message,
+    allocator: Allocator,
+    cfg: *const config_mod.OpenAIResponsesConfig,
+    dialect: RequestDialect,
+) !void {
     switch (msg.role) {
         .system => {},
         .user => {
@@ -161,17 +189,33 @@ fn writeInputForMessage(s: *std.json.Stringify, msg: conversation.Message, alloc
             defer text_buf.deinit(allocator);
             try concatTextBlocks(msg.content.items, &text_buf, allocator);
             if (text_buf.items.len > 0) {
-                try writeRoleMessage(s, "user", "input_text", text_buf.items);
+                try writeRoleMessage(s, "user", "input_text", text_buf.items, null);
             }
         },
         .assistant => {
+            // Replay opaque reasoning items first so stateless follow-up turns
+            // preserve encrypted reasoning continuity.
+            for (msg.content.items) |block| {
+                if (block != .Thinking) continue;
+                const tb = block.Thinking;
+                const sig = tb.signature orelse continue;
+                const origin = tb.signature_origin orelse continue;
+                if (!origin.matches(
+                    if (dialect == .codex) .openai_codex_responses else .openai_responses,
+                    cfg.base_url,
+                    cfg.model,
+                )) continue;
+                if (sig.len == 0 or sig[0] != '{') continue;
+                try writeRawJson(s, sig);
+            }
+
             // Assistant text first (as an output_text message), then each
             // tool call as a `function_call` item.
             var text_buf: std.ArrayList(u8) = .empty;
             defer text_buf.deinit(allocator);
             try concatTextBlocks(msg.content.items, &text_buf, allocator);
             if (text_buf.items.len > 0) {
-                try writeRoleMessage(s, "assistant", "output_text", text_buf.items);
+                try writeRoleMessage(s, "assistant", "output_text", text_buf.items, openAIPhaseFromMetadata(msg.metadata));
             }
             for (msg.content.items) |block| {
                 if (block != .ToolUse) continue;
@@ -192,10 +236,20 @@ fn writeInputForMessage(s: *std.json.Stringify, msg: conversation.Message, alloc
     }
 }
 
-fn writeRoleMessage(s: *std.json.Stringify, role: []const u8, content_type: []const u8, text: []const u8) !void {
+fn writeRoleMessage(
+    s: *std.json.Stringify,
+    role: []const u8,
+    content_type: []const u8,
+    text: []const u8,
+    phase: ?[]const u8,
+) !void {
     try s.beginObject();
     try s.objectField("role");
     try s.write(role);
+    if (phase) |p| {
+        try s.objectField("phase");
+        try s.write(p);
+    }
     try s.objectField("content");
     try s.beginArray();
     try s.beginObject();
@@ -208,6 +262,17 @@ fn writeRoleMessage(s: *std.json.Stringify, role: []const u8, content_type: []co
     try s.endObject();
 }
 
+fn openAIPhaseFromMetadata(metadata: ?[]const u8) ?[]const u8 {
+    const md = metadata orelse return null;
+    var parsed = std.json.parseFromSlice(std.json.Value, std.heap.page_allocator, md, .{}) catch return null;
+    defer parsed.deinit();
+    if (parsed.value != .object) return null;
+    const phase = strField(parsed.value.object, "openai_responses_phase") orelse return null;
+    if (std.mem.eql(u8, phase, "commentary")) return "commentary";
+    if (std.mem.eql(u8, phase, "final_answer")) return "final_answer";
+    return null;
+}
+
 fn concatTextBlocks(
     blocks: []const conversation.ContentBlock,
     out: *std.ArrayList(u8),
@@ -244,6 +309,7 @@ pub const EventKind = enum {
     output_text_delta,
     reasoning_summary_delta,
     function_call_arguments_delta,
+    function_call_arguments_done,
     output_item_done,
     completed,
     failed,
@@ -261,8 +327,16 @@ pub const StreamEvent = struct {
     /// Item id (`response.output_item.added/done`, function-call argument
     /// deltas reference `item_id`).
     item_id: ?[]const u8 = null,
+    /// Output array index. Some function-call argument events identify the
+    /// item by this instead of repeating `item_id`.
+    output_index: ?usize = null,
     /// Item type on add/done: "message" | "function_call" | "reasoning".
     item_type: ?[]const u8 = null,
+    /// Assistant message phase on message output items.
+    item_phase: ?[]const u8 = null,
+    /// Raw reasoning output item JSON, used for stateless encrypted reasoning
+    /// replay on follow-up turns.
+    reasoning_item_json: ?[]const u8 = null,
     /// Function-call identity (on `output_item.added`/`done`).
     call_id: ?[]const u8 = null,
     name: ?[]const u8 = null,
@@ -272,6 +346,8 @@ pub const StreamEvent = struct {
     error_message: ?[]const u8 = null,
     /// Usage on `response.completed`.
     usage: ?Usage = null,
+    /// Function-call output items included in a terminal `response.completed`.
+    completed_items: []const OutputItem = &.{},
 
     pub const Usage = struct {
         input_tokens: u64 = 0,
@@ -280,6 +356,15 @@ pub const StreamEvent = struct {
         reasoning_tokens: u64 = 0,
     };
 
+    pub const OutputItem = struct {
+        output_index: usize,
+        item_id: ?[]const u8 = null,
+        item_type: ?[]const u8 = null,
+        call_id: ?[]const u8 = null,
+        name: ?[]const u8 = null,
+        arguments: ?[]const u8 = null,
+    };
+
     pub fn deinit(self: *StreamEvent) void {
         self.parsed.deinit();
     }
@@ -305,23 +390,36 @@ pub fn parseStreamEvent(allocator: Allocator, payload: []const u8) !StreamEvent
         ev.kind = .output_text_delta;
         ev.delta = strField(obj, "delta");
         ev.item_id = strField(obj, "item_id");
+        ev.output_index = usizeField(obj, "output_index");
     } else if (std.mem.eql(u8, type_str, "response.reasoning_summary_text.delta")) {
         ev.kind = .reasoning_summary_delta;
         ev.delta = strField(obj, "delta");
         ev.item_id = strField(obj, "item_id");
+        ev.output_index = usizeField(obj, "output_index");
     } else if (std.mem.eql(u8, type_str, "response.function_call_arguments.delta")) {
         ev.kind = .function_call_arguments_delta;
-        ev.delta = strField(obj, "delta");
+        ev.delta = strField(obj, "delta") orelse
+            strField(obj, "arguments_delta") orelse
+            strField(obj, "arguments");
         ev.item_id = strField(obj, "item_id");
+        ev.output_index = usizeField(obj, "output_index");
+    } else if (std.mem.eql(u8, type_str, "response.function_call_arguments.done")) {
+        ev.kind = .function_call_arguments_done;
+        ev.arguments = strField(obj, "arguments");
+        ev.item_id = strField(obj, "item_id");
+        ev.output_index = usizeField(obj, "output_index");
     } else if (std.mem.eql(u8, type_str, "response.output_item.added")) {
         ev.kind = .output_item_added;
-        readItem(obj, &ev);
+        ev.output_index = usizeField(obj, "output_index");
+        try readItem(parsed.arena.allocator(), obj, &ev);
     } else if (std.mem.eql(u8, type_str, "response.output_item.done")) {
         ev.kind = .output_item_done;
-        readItem(obj, &ev);
-    } else if (std.mem.eql(u8, type_str, "response.completed")) {
+        ev.output_index = usizeField(obj, "output_index");
+        try readItem(parsed.arena.allocator(), obj, &ev);
+    } else if (std.mem.eql(u8, type_str, "response.completed") or std.mem.eql(u8, type_str, "response.done")) {
         ev.kind = .completed;
         readUsage(obj, &ev);
+        try readCompletedOutput(parsed.arena.allocator(), obj, &ev);
     } else if (std.mem.eql(u8, type_str, "response.failed") or std.mem.eql(u8, type_str, "response.incomplete")) {
         ev.kind = .failed;
         ev.error_message = readResponseError(obj);
@@ -334,15 +432,31 @@ pub fn parseStreamEvent(allocator: Allocator, payload: []const u8) !StreamEvent
     return ev;
 }
 
-fn readItem(obj: std.json.ObjectMap, ev: *StreamEvent) void {
+fn readItem(allocator: Allocator, obj: std.json.ObjectMap, ev: *StreamEvent) !void {
     const item = obj.get("item") orelse return;
     if (item != .object) return;
     const io = item.object;
     ev.item_id = strField(io, "id");
     ev.item_type = strField(io, "type");
+    ev.item_phase = strField(io, "phase");
     ev.call_id = strField(io, "call_id");
     ev.name = strField(io, "name");
     ev.arguments = strField(io, "arguments");
+    if (ev.item_type) |it| {
+        if (std.mem.eql(u8, it, "reasoning")) {
+            if (io.get("encrypted_content") != null) {
+                ev.reasoning_item_json = try stringifyValue(allocator, item);
+            }
+        }
+    }
+}
+
+fn stringifyValue(allocator: Allocator, value: std.json.Value) ![]const u8 {
+    var aw: Writer.Allocating = .init(allocator);
+    errdefer aw.deinit();
+    var s: std.json.Stringify = .{ .writer = &aw.writer };
+    try s.write(value);
+    return try aw.toOwnedSlice();
 }
 
 fn readUsage(obj: std.json.ObjectMap, ev: *StreamEvent) void {
@@ -366,6 +480,32 @@ fn readUsage(obj: std.json.ObjectMap, ev: *StreamEvent) void {
     ev.usage = usage;
 }
 
+fn readCompletedOutput(allocator: Allocator, obj: std.json.ObjectMap, ev: *StreamEvent) !void {
+    const resp = obj.get("response") orelse return;
+    if (resp != .object) return;
+    const output = resp.object.get("output") orelse return;
+    if (output != .array) return;
+
+    var items: std.ArrayList(StreamEvent.OutputItem) = .empty;
+    errdefer items.deinit(allocator);
+    for (output.array.items, 0..) |v, i| {
+        if (v != .object) continue;
+        const item_type = strField(v.object, "type") orelse continue;
+        if (!std.mem.eql(u8, item_type, "function_call")) continue;
+        try items.append(allocator, .{
+            .output_index = i,
+            .item_id = strField(v.object, "id"),
+            .item_type = item_type,
+            .call_id = strField(v.object, "call_id"),
+            .name = strField(v.object, "name"),
+            .arguments = strField(v.object, "arguments"),
+        });
+    }
+    if (items.items.len == 0) return;
+    const buf = try items.toOwnedSlice(allocator);
+    ev.completed_items = buf;
+}
+
 fn readResponseError(obj: std.json.ObjectMap) ?[]const u8 {
     const resp = obj.get("response") orelse return null;
     if (resp != .object) return null;
@@ -389,6 +529,12 @@ fn u64Field(obj: std.json.ObjectMap, name: []const u8) u64 {
     return @intCast(v.integer);
 }
 
+fn usizeField(obj: std.json.ObjectMap, name: []const u8) ?usize {
+    const v = obj.get(name) orelse return null;
+    if (v != .integer or v.integer < 0) return null;
+    return @intCast(v.integer);
+}
+
 // ===========================================================================
 // Tests
 // ===========================================================================
@@ -421,7 +567,7 @@ test "responses serializeRequest - instructions, input, store/include" {
     cfg.reasoning = .high;
     var tools = emptyTools();
     defer tools.deinit();
-    const body = try serializeRequest(allocator, &cfg, &conv, &tools);
+    const body = try serializeRequest(allocator, &cfg, &conv, &tools, .public);
     defer allocator.free(body);
 
     var parsed = try std.json.parseFromSlice(std.json.Value, allocator, body, .{});
@@ -458,7 +604,7 @@ test "responses serializeRequest - tools are flat function items" {
     });
 
     const cfg = testConfig("gpt-5.1-codex");
-    const body = try serializeRequest(allocator, &cfg, &conv, &tools);
+    const body = try serializeRequest(allocator, &cfg, &conv, &tools, .public);
     defer allocator.free(body);
     var parsed = try std.json.parseFromSlice(std.json.Value, allocator, body, .{});
     defer parsed.deinit();
@@ -471,6 +617,87 @@ test "responses serializeRequest - tools are flat function items" {
     try testing.expect(tool0.get("parameters").? == .object);
 }
 
+test "responses serializeRequest - codex dialect omits max tokens and sends codex defaults" {
+    const allocator = testing.allocator;
+    var conv = conversation.Conversation.init(allocator);
+    defer conv.deinit();
+    try addUserText(&conv, "Hello!");
+
+    var tools = emptyTools();
+    defer tools.deinit();
+    const cfg = testConfig("gpt-5.5");
+    const body = try serializeRequest(allocator, &cfg, &conv, &tools, .codex);
+    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("max_output_tokens") == null);
+    try testing.expectEqualStrings("low", root.get("text").?.object.get("verbosity").?.string);
+    try testing.expectEqualStrings("auto", root.get("tool_choice").?.string);
+    try testing.expect(root.get("parallel_tool_calls").?.bool);
+}
+
+test "responses serializeRequest - assistant phase metadata and reasoning signature replay" {
+    const allocator = testing.allocator;
+    var conv = conversation.Conversation.init(allocator);
+    defer conv.deinit();
+
+    const sig = try allocator.dupe(u8,
+        "{\"type\":\"reasoning\",\"id\":\"rs_1\",\"encrypted_content\":\"sealed\"}"
+    );
+    const thinking = try conversation.textualBlockFromSlice(allocator, "thinking");
+    const text = try conversation.textualBlockFromSlice(allocator, "answer");
+    try conv.addAssistantMessage(&.{
+        .{ .Thinking = .{ .text = thinking, .signature = sig } },
+        .{ .Text = text },
+    }, null);
+    try conversation.setThinkingOrigins(allocator, conv.messages.items[0].content.items, .openai_codex_responses, "u", "gpt-5.5");
+    conv.messages.items[0].metadata = try allocator.dupe(u8, "{\"openai_responses_phase\":\"final_answer\"}");
+
+    var tools = emptyTools();
+    defer tools.deinit();
+    const cfg = testConfig("gpt-5.5");
+    const body = try serializeRequest(allocator, &cfg, &conv, &tools, .codex);
+    defer allocator.free(body);
+    var parsed = try std.json.parseFromSlice(std.json.Value, allocator, body, .{});
+    defer parsed.deinit();
+
+    const input = parsed.value.object.get("input").?.array.items;
+    try testing.expectEqual(@as(usize, 2), input.len);
+    try testing.expectEqualStrings("reasoning", input[0].object.get("type").?.string);
+    try testing.expectEqualStrings("sealed", input[0].object.get("encrypted_content").?.string);
+    try testing.expectEqualStrings("assistant", input[1].object.get("role").?.string);
+    try testing.expectEqualStrings("final_answer", input[1].object.get("phase").?.string);
+}
+
+test "responses serializeRequest - mismatched signature origin skips reasoning replay" {
+    const allocator = testing.allocator;
+    var conv = conversation.Conversation.init(allocator);
+    defer conv.deinit();
+
+    const sig = try allocator.dupe(u8,
+        "{\"type\":\"reasoning\",\"id\":\"rs_1\",\"encrypted_content\":\"sealed\"}"
+    );
+    try conv.addAssistantMessage(&.{
+        .{ .Thinking = .{ .text = try conversation.textualBlockFromSlice(allocator, "thinking"), .signature = sig } },
+        .{ .Text = try conversation.textualBlockFromSlice(allocator, "answer") },
+    }, null);
+    try conversation.setThinkingOrigins(allocator, conv.messages.items[0].content.items, .openai_responses, "https://api.individual.githubcopilot.com", "gpt-5.4-mini");
+
+    var tools = emptyTools();
+    defer tools.deinit();
+    const cfg = testConfig("gpt-5.5");
+    const body = try serializeRequest(allocator, &cfg, &conv, &tools, .codex);
+    defer allocator.free(body);
+    var parsed = try std.json.parseFromSlice(std.json.Value, allocator, body, .{});
+    defer parsed.deinit();
+
+    const input = parsed.value.object.get("input").?.array.items;
+    try testing.expectEqual(@as(usize, 1), input.len);
+    try testing.expectEqualStrings("assistant", input[0].object.get("role").?.string);
+}
+
 test "responses serializeRequest - assistant tool_use + tool result round-trip" {
     const allocator = testing.allocator;
     var conv = conversation.Conversation.init(allocator);
@@ -495,7 +722,7 @@ test "responses serializeRequest - assistant tool_use + tool result round-trip"
     var tools = emptyTools();
     defer tools.deinit();
     const cfg = testConfig("gpt-5.1-codex");
-    const body = try serializeRequest(allocator, &cfg, &conv, &tools);
+    const body = try serializeRequest(allocator, &cfg, &conv, &tools, .public);
     defer allocator.free(body);
     var parsed = try std.json.parseFromSlice(std.json.Value, allocator, body, .{});
     defer parsed.deinit();
@@ -540,6 +767,29 @@ test "responses parseStreamEvent - function_call item added/done + args delta" {
     try testing.expectEqual(EventKind.function_call_arguments_delta, d.kind);
     try testing.expectEqualStrings("fc_1", d.item_id.?);
     try testing.expectEqualStrings("{\"x\":1}", d.delta.?);
+
+    var alt = try parseStreamEvent(allocator,
+        \\{"type":"response.function_call_arguments.delta","item_id":"fc_1","arguments_delta":"{\"x\":1}"}
+    );
+    defer alt.deinit();
+    try testing.expectEqual(EventKind.function_call_arguments_delta, alt.kind);
+    try testing.expectEqualStrings("{\"x\":1}", alt.delta.?);
+
+    var by_index = try parseStreamEvent(allocator,
+        \\{"type":"response.function_call_arguments.delta","output_index":0,"delta":"{\"x\":1}"}
+    );
+    defer by_index.deinit();
+    try testing.expectEqual(EventKind.function_call_arguments_delta, by_index.kind);
+    try testing.expectEqual(@as(usize, 0), by_index.output_index.?);
+    try testing.expectEqualStrings("{\"x\":1}", by_index.delta.?);
+
+    var done = try parseStreamEvent(allocator,
+        \\{"type":"response.function_call_arguments.done","output_index":0,"arguments":"{\"x\":1}"}
+    );
+    defer done.deinit();
+    try testing.expectEqual(EventKind.function_call_arguments_done, done.kind);
+    try testing.expectEqual(@as(usize, 0), done.output_index.?);
+    try testing.expectEqualStrings("{\"x\":1}", done.arguments.?);
 }
 
 test "responses parseStreamEvent - completed usage" {
@@ -555,6 +805,45 @@ test "responses parseStreamEvent - completed usage" {
     try testing.expectEqual(@as(u64, 8), ev.usage.?.reasoning_tokens);
 }
 
+test "responses parseStreamEvent - done alias, phase, encrypted reasoning item" {
+    const allocator = testing.allocator;
+    var done = try parseStreamEvent(allocator,
+        \\{"type":"response.done","response":{"usage":{"input_tokens":1,"output_tokens":2}}}
+    );
+    defer done.deinit();
+    try testing.expectEqual(EventKind.completed, done.kind);
+    try testing.expectEqual(@as(u64, 1), done.usage.?.input_tokens);
+
+    var msg = try parseStreamEvent(allocator,
+        \\{"type":"response.output_item.done","item":{"type":"message","id":"msg_1","phase":"commentary"}}
+    );
+    defer msg.deinit();
+    try testing.expectEqual(EventKind.output_item_done, msg.kind);
+    try testing.expectEqualStrings("commentary", msg.item_phase.?);
+
+    var reasoning = try parseStreamEvent(allocator,
+        \\{"type":"response.output_item.done","item":{"type":"reasoning","id":"rs_1","encrypted_content":"sealed"}}
+    );
+    defer reasoning.deinit();
+    try testing.expectEqual(EventKind.output_item_done, reasoning.kind);
+    try testing.expect(reasoning.reasoning_item_json != null);
+}
+
+test "responses parseStreamEvent - completed output function calls" {
+    const allocator = testing.allocator;
+    var ev = try parseStreamEvent(allocator,
+        \\{"type":"response.completed","response":{"output":[{"type":"function_call","id":"fc_1","call_id":"call_9","name":"std__read","arguments":"{\"path\":\"a\"}","status":"completed"}],"usage":{"input_tokens":100,"output_tokens":20}}}
+    );
+    defer ev.deinit();
+    try testing.expectEqual(EventKind.completed, ev.kind);
+    try testing.expectEqual(@as(usize, 1), ev.completed_items.len);
+    try testing.expectEqual(@as(usize, 0), ev.completed_items[0].output_index);
+    try testing.expectEqualStrings("fc_1", ev.completed_items[0].item_id.?);
+    try testing.expectEqualStrings("call_9", ev.completed_items[0].call_id.?);
+    try testing.expectEqualStrings("std__read", ev.completed_items[0].name.?);
+    try testing.expectEqualStrings("{\"path\":\"a\"}", ev.completed_items[0].arguments.?);
+}
+
 test "responses parseStreamEvent - error + failed" {
     const allocator = testing.allocator;
     var e = try parseStreamEvent(allocator,
-- 
cgit v1.3