compaction

1 files changed, 31 insertions, 0 deletions

diff --git a/libpanto/src/provider.zig b/libpanto/src/provider.zig
index 1a8ee4e..6ccbefe 100644
--- a/libpanto/src/provider.zig
+++ b/libpanto/src/provider.zig
@@ -14,6 +14,28 @@ pub const ContentBlockType = enum {
     ToolResult,
 };
 
+/// Heuristic detector for provider context-overflow rejections, applied to
+/// an HTTP 400 error response body. Both OpenAI-compatible and Anthropic
+/// APIs reject oversized requests on the input side with HTTP 400 and a
+/// recognizable message; matching it lets the agent compact and retry
+/// instead of surfacing a hard error.
+///
+/// Markers (case-sensitive substrings, as the wire emits them):
+///   - OpenAI:    `context_length_exceeded`, `maximum context length`
+///   - Anthropic: `prompt is too long`
+pub fn isContextOverflowBody(body: []const u8) bool {
+    const markers = [_][]const u8{
+        "context_length_exceeded",
+        "maximum context length",
+        "prompt is too long",
+        "context window",
+    };
+    for (markers) |m| {
+        if (std.mem.indexOf(u8, body, m) != null) return true;
+    }
+    return false;
+}
+
 /// Vtable for receiving streaming events from a Provider.
 ///
 /// The lifecycle callbacks (`onMessageStart` ... `onMessageComplete`) return
@@ -143,3 +165,12 @@ pub const StreamFn = *const fn (
     conv: *conversation.Conversation,
     receiver: *Receiver,
 ) anyerror!void;
+
+test "isContextOverflowBody - matches known markers, rejects others" {
+    const t2 = std.testing;
+    try t2.expect(isContextOverflowBody("{\"error\":{\"code\":\"context_length_exceeded\"}}"));
+    try t2.expect(isContextOverflowBody("This model's maximum context length is 8192 tokens"));
+    try t2.expect(isContextOverflowBody("prompt is too long: 250000 tokens > 200000 maximum"));
+    try t2.expect(!isContextOverflowBody("{\"error\":{\"code\":\"invalid_api_key\"}}"));
+    try t2.expect(!isContextOverflowBody("rate limit exceeded"));
+}