compaction

1 files changed, 19 insertions, 0 deletions

diff --git a/libpanto/src/config.zig b/libpanto/src/config.zig
index 6b8d9a8..114280c 100644
--- a/libpanto/src/config.zig
+++ b/libpanto/src/config.zig
@@ -69,6 +69,24 @@ pub const ProviderConfig = union(APIStyle) {
     }
 };
 
+/// Compaction settings the agent consults when summarizing old history.
+///
+/// `keep_verbatim` is the budget (in tokens) governing how much recent
+/// conversation is kept verbatim after compaction. The retention walk
+/// accumulates whole turns backward and stops once the running total
+/// exceeds this value, so `keep_verbatim` is an *upper bound* on the size
+/// of the kept suffix (the turn that crosses the threshold is summarized,
+/// not kept).
+///
+/// `model`, when set, is the provider/model used to run the compaction
+/// request itself. On failure (e.g. the compaction model rejects the
+/// transcript for context length) the agent falls back to the active chat
+/// model. When null, compaction uses the active chat model directly.
+pub const CompactionConfig = struct {
+    keep_verbatim: u32 = 20_000,
+    model: ?ProviderConfig = null,
+};
+
 /// An immutable snapshot of everything the agent consults per turn: which
 /// provider/model to talk to, and which tools to expose. The agent holds a
 /// `*const Config` and re-reads it each turn; replacing the pointer swaps
@@ -80,6 +98,7 @@ pub const ProviderConfig = union(APIStyle) {
 pub const Config = struct {
     provider: ProviderConfig,
     registry: *const ToolRegistry,
+    compaction: CompactionConfig = .{},
 
     pub fn style(self: Config) APIStyle {
         return self.provider.style();