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.

1 files changed, 10 insertions, 2 deletions

diff --git a/libpanto/src/provider_openai_chat.zig b/libpanto/src/provider_openai_chat.zig
index b236a60..3f01e7f 100644
--- a/libpanto/src/provider_openai_chat.zig
+++ b/libpanto/src/provider_openai_chat.zig
@@ -94,6 +94,7 @@ pub const OpenAIChatRequest = struct {
         // Build the request body.
         const body = try json_mod.serializeRequest(self.allocator, self.config, conv, tools);
         defer self.allocator.free(body);
+        std.log.debug("openai_chat => {s}", .{body});
 
         // Auth header
         const auth_value = try std.fmt.allocPrint(
@@ -165,11 +166,19 @@ pub const OpenAIChatRequest = struct {
                 if (err_buf.items.len > 16 * 1024) break;
             }
             const status: u16 = @intFromEnum(rr.response.head.status);
-            std.log.err("openai_chat HTTP {d}: {s}", .{ status, err_buf.items });
             // Classify the status into a retryable/terminal provider error.
             // HTTP 400 with a context marker becomes `ContextOverflow` so the
             // caller can compact and retry rather than hard-fail.
             const classified = provider_mod.classifyHttpStatus(status, err_buf.items);
+            // 401/403 is routinely recovered by the turn-runner's forced token
+            // refresh + reopen; demote it to `.debug` (still in the debug log)
+            // so a transparent refresh doesn't surface a scary error line. The
+            // retry layer raises a hard error only if recovery ultimately fails.
+            if (classified == error.ProviderAuthFailed) {
+                std.log.debug("openai_chat HTTP {d} (recoverable auth): {s}", .{ status, err_buf.items });
+            } else {
+                std.log.err("openai_chat HTTP {d}: {s}", .{ status, err_buf.items });
+            }
             if (self.diag) |d| {
                 d.status_code = status;
                 d.retry_after_ms = retry_after_ms;
@@ -1032,7 +1041,6 @@ test "inbound wire tool name is decoded to dotted form (even split across __)" {
     try testing.expectEqualStrings("std.read", tu.name);
 }
 
-
 test "parallel tool_calls emit one complete start/delta/complete cycle per block" {
     // Regression test: previously, the OpenAI provider deferred ALL
     // tool_use onBlockComplete callbacks to finalize, so a four-tool