update: Consolidationの不要なToken上限の削除

crates/manifest/src/config.rs

@@ -218,9 +218,6 @@ impl MemoryConfig {
                 .extract_worker_max_input_tokens
                 .or(self.extract_worker_max_input_tokens),
             consolidation_model: upper.consolidation_model.or(self.consolidation_model),
-            consolidation_worker_max_input_tokens: upper
-                .consolidation_worker_max_input_tokens
-                .or(self.consolidation_worker_max_input_tokens),
             consolidation_threshold_files: upper
                 .consolidation_threshold_files
                 .or(self.consolidation_threshold_files),

crates/manifest/src/defaults.rs

@@ -50,8 +50,3 @@ pub const COMPACT_DEFAULT_REFERENCE_COUNT: usize = 5;
 /// own LLM calls. Exceeding this aborts the extract run.
 /// See [`crate::MemoryConfig::extract_worker_max_input_tokens`].
 pub const MEMORY_EXTRACT_WORKER_MAX_INPUT_TOKENS: u64 = 30_000;
-
-/// Cumulative input-token cap for the memory Phase 2 (consolidation)
-/// worker's own LLM calls. Exceeding this aborts the consolidation run.
-/// See [`crate::MemoryConfig::consolidation_worker_max_input_tokens`].
-pub const MEMORY_CONSOLIDATION_WORKER_MAX_INPUT_TOKENS: u64 = 80_000;

crates/manifest/src/lib.rs

@@ -90,11 +90,6 @@ pub struct MemoryConfig {
     /// Reasoning-class models are recommended.
     #[serde(default)]
     pub consolidation_model: Option<ModelManifest>,
-    /// Cumulative input-token cap for the consolidation worker's own
-    /// LLM calls. Exceeding this aborts the consolidation run. `None` ⇒
-    /// [`defaults::MEMORY_CONSOLIDATION_WORKER_MAX_INPUT_TOKENS`].
-    #[serde(default)]
-    pub consolidation_worker_max_input_tokens: Option<u64>,
     /// Phase 2 trigger: file-count threshold of `_staging/`. Phase 2
     /// fires when the staging directory has at least this many entries.
     /// Either threshold reaching its limit fires Phase 2 (logical OR).

crates/pod/src/pod.rs

@@ -1735,9 +1735,6 @@ impl<C: LlmClient, St: Store> Pod<C, St> {
             Err(e) => return Err(PodError::ConsolidationLock(e)),
         };
 
-        let cap = memory_cfg
-            .consolidation_worker_max_input_tokens
-            .unwrap_or(manifest::defaults::MEMORY_CONSOLIDATION_WORKER_MAX_INPUT_TOKENS);
         let client = match self.build_consolidator_client(memory_cfg) {
             Ok(c) => c,
             Err(e) => {
@@ -1749,20 +1746,6 @@ impl<C: LlmClient, St: Store> Pod<C, St> {
             Worker::new(client).system_prompt(consolidate::CONSOLIDATION_SYSTEM_PROMPT);
         worker.set_cache_key(Some(self.session_id.to_string()));
 
-        let input_so_far = Arc::new(std::sync::atomic::AtomicU64::new(0));
-        {
-            let acc = input_so_far.clone();
-            worker.on_usage(move |event| {
-                if let Some(tokens) = event.input_tokens {
-                    acc.fetch_add(tokens, Ordering::Relaxed);
-                }
-            });
-        }
-        worker.set_interceptor(MemoryConsolidationWorkerInterceptor {
-            input_so_far: input_so_far.clone(),
-            max_input_tokens: cap,
-        });
-
         // Memory tools are self-contained — they bypass ScopedFs and write
         // directly under the workspace via WorkspaceLayout. Resident
         // knowledge injection (`Pod::set_resident_knowledge_injection`) is
@@ -1843,30 +1826,6 @@ enum ConsolidateDecision {
     Completed,
 }
 
-/// Pre-request interceptor for the Phase 2 consolidation worker. Same
-/// shape as the extract interceptor; kept separate so the abort message
-/// names the right subsystem.
-struct MemoryConsolidationWorkerInterceptor {
-    input_so_far: Arc<std::sync::atomic::AtomicU64>,
-    max_input_tokens: u64,
-}
-
-#[async_trait]
-impl llm_worker::interceptor::Interceptor for MemoryConsolidationWorkerInterceptor {
-    async fn pre_llm_request(
-        &self,
-        _context: &mut Vec<Item>,
-    ) -> llm_worker::interceptor::PreRequestAction {
-        if self.input_so_far.load(Ordering::Relaxed) > self.max_input_tokens {
-            return llm_worker::interceptor::PreRequestAction::Cancel(format!(
-                "Phase 2 consolidation worker input exceeded {} tokens",
-                self.max_input_tokens
-            ));
-        }
-        llm_worker::interceptor::PreRequestAction::Continue
-    }
-}
-
 impl<St: Store> Pod<Box<dyn LlmClient>, St> {
     /// Create a Pod entirely from a validated manifest.
     ///