Pruneの実装

crates/pod/Cargo.toml

@@ -17,6 +17,7 @@ serde_json = "1.0.149"
 thiserror = "2.0"
 tokio = { version = "1.49", features = ["fs", "io-util", "macros", "net", "rt-multi-thread", "signal", "sync"] }
 toml = "1.1.2"
+tracing = "0.1.44"
 
 [dev-dependencies]
 async-trait = "0.1.89"

crates/pod/src/lib.rs

@@ -4,12 +4,15 @@ pub mod runtime_dir;
 pub mod shared_state;
 pub mod socket_server;
 
+pub mod prune_hook;
+
 mod hook_interceptor;
 mod pod;
 
 pub use controller::{PodController, PodHandle};
 pub use manifest::{PodManifest, ProviderConfig, ProviderKind, Scope};
 pub use hook::{Hook, HookEventKind, HookRegistryBuilder};
+pub use prune_hook::PruneHook;
 pub use pod::{Pod, PodError, PodRunResult, apply_worker_manifest};
 pub use protocol::{ErrorCode, Event, Method, TurnResult};
 pub use provider::{ProviderError, build_client};

crates/pod/src/prune_hook.rs

@@ -0,0 +1,38 @@
+//! PruneHook — applies conditional pruning before each LLM request.
+//!
+//! Wraps [`llm_worker::prune::prune()`] as a [`Hook<PreLlmRequest>`] so
+//! that Pod can register it in the hook pipeline.
+
+use async_trait::async_trait;
+use llm_worker::interceptor::PreRequestAction;
+use llm_worker::prune::{PruneConfig, prune};
+use llm_worker::Item;
+use tracing::debug;
+
+use crate::hook::{Hook, PreLlmRequest};
+
+/// Hook that conditionally prunes old tool-result content before each
+/// LLM request, reclaiming context-window tokens.
+pub struct PruneHook {
+    config: PruneConfig,
+}
+
+impl PruneHook {
+    pub fn new(config: PruneConfig) -> Self {
+        Self { config }
+    }
+}
+
+#[async_trait]
+impl Hook<PreLlmRequest> for PruneHook {
+    async fn call(&self, context: &mut Vec<Item>) -> PreRequestAction {
+        if let Some(result) = prune(context, &self.config) {
+            debug!(
+                pruned = result.pruned_count,
+                estimated_savings = result.estimated_savings,
+                "Pruned old tool-result content"
+            );
+        }
+        PreRequestAction::Continue
+    }
+}