feat: cache-tracking progress

This threads typed hook settings through runtime config, adds a shell-based hook runner, and executes PreToolUse/PostToolUse around each tool call in the conversation loop. The CLI now rebuilds runtimes with settings-derived hook configuration so user-defined Claude hook commands actually run before and after tools.

Constraint: Hook behavior needed to match Claude-style settings.json hooks without broad plugin/MCP parity work in this change
Rejected: Delay hook loading to the tool executor layer | would miss denied tool calls and duplicate runtime policy plumbing
Confidence: medium
Scope-risk: moderate
Reversibility: clean
Directive: Keep hook execution in the runtime loop so permission decisions and tool results remain wrapped by the same conversation semantics
Tested: cargo test; cargo build --release
Not-tested: Real user hook scripts outside the test harness; broader plugin/skills parity

rust/crates/api/src/client.rs

@@ -1,4 +1,5 @@
 use std::collections::VecDeque;
+use std::sync::{Arc, Mutex};
 use std::time::{Duration, SystemTime, UNIX_EPOCH};
 
 use runtime::{
@@ -8,8 +9,9 @@ use runtime::{
 use serde::Deserialize;
 
 use crate::error::ApiError;
+use crate::prompt_cache::{PromptCache, PromptCacheRecord, PromptCacheStats};
 use crate::sse::SseParser;
-use crate::types::{MessageRequest, MessageResponse, StreamEvent};
+use crate::types::{MessageRequest, MessageResponse, StreamEvent, Usage};
 
 const DEFAULT_BASE_URL: &str = "https://api.anthropic.com";
 const ANTHROPIC_VERSION: &str = "2023-06-01";
@@ -108,6 +110,8 @@ pub struct AnthropicClient {
     max_retries: u32,
     initial_backoff: Duration,
     max_backoff: Duration,
+    prompt_cache: Option<PromptCache>,
+    last_prompt_cache_record: Arc<Mutex<Option<PromptCacheRecord>>>,
 }
 
 impl AnthropicClient {
@@ -120,6 +124,8 @@ impl AnthropicClient {
             max_retries: DEFAULT_MAX_RETRIES,
             initial_backoff: DEFAULT_INITIAL_BACKOFF,
             max_backoff: DEFAULT_MAX_BACKOFF,
+            prompt_cache: None,
+            last_prompt_cache_record: Arc::new(Mutex::new(None)),
         }
     }
 
@@ -132,6 +138,8 @@ impl AnthropicClient {
             max_retries: DEFAULT_MAX_RETRIES,
             initial_backoff: DEFAULT_INITIAL_BACKOFF,
             max_backoff: DEFAULT_MAX_BACKOFF,
+            prompt_cache: None,
+            last_prompt_cache_record: Arc::new(Mutex::new(None)),
         }
     }
 
@@ -189,6 +197,30 @@ impl AnthropicClient {
         self
     }
 
+    #[must_use]
+    pub fn with_prompt_cache(mut self, prompt_cache: PromptCache) -> Self {
+        self.prompt_cache = Some(prompt_cache);
+        self
+    }
+
+    #[must_use]
+    pub fn prompt_cache(&self) -> Option<&PromptCache> {
+        self.prompt_cache.as_ref()
+    }
+
+    #[must_use]
+    pub fn prompt_cache_stats(&self) -> Option<PromptCacheStats> {
+        self.prompt_cache.as_ref().map(PromptCache::stats)
+    }
+
+    #[must_use]
+    pub fn take_last_prompt_cache_record(&self) -> Option<PromptCacheRecord> {
+        self.last_prompt_cache_record()
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+            .take()
+    }
+
     #[must_use]
     pub fn auth_source(&self) -> &AuthSource {
         &self.auth
@@ -198,10 +230,19 @@ impl AnthropicClient {
         &self,
         request: &MessageRequest,
     ) -> Result<MessageResponse, ApiError> {
+        self.store_last_prompt_cache_record(None);
         let request = MessageRequest {
             stream: false,
             ..request.clone()
         };
+        if let Some(prompt_cache) = &self.prompt_cache {
+            if let Some(response) = prompt_cache.lookup_completion(&request) {
+                self.store_last_prompt_cache_record(Some(prompt_cache_record_from_stats(
+                    prompt_cache.stats(),
+                )));
+                return Ok(response);
+            }
+        }
         let response = self.send_with_retry(&request).await?;
         let request_id = request_id_from_headers(response.headers());
         let mut response = response
@@ -211,6 +252,10 @@ impl AnthropicClient {
         if response.request_id.is_none() {
             response.request_id = request_id;
         }
+        if let Some(prompt_cache) = &self.prompt_cache {
+            let record = prompt_cache.record_response(&request, &response);
+            self.store_last_prompt_cache_record(Some(record));
+        }
         Ok(response)
     }
 
@@ -218,6 +263,7 @@ impl AnthropicClient {
         &self,
         request: &MessageRequest,
     ) -> Result<MessageStream, ApiError> {
+        self.store_last_prompt_cache_record(None);
         let response = self
             .send_with_retry(&request.clone().with_streaming())
             .await?;
@@ -227,9 +273,30 @@ impl AnthropicClient {
             parser: SseParser::new(),
             pending: VecDeque::new(),
             done: false,
+            cache_tracking: self
+                .prompt_cache
+                .as_ref()
+                .map(|prompt_cache| StreamCacheTracking {
+                    prompt_cache: prompt_cache.clone(),
+                    request: request.clone().with_streaming(),
+                    last_usage: None,
+                    finalized: false,
+                    last_record: self.last_prompt_cache_record.clone(),
+                }),
         })
     }
 
+    fn store_last_prompt_cache_record(&self, record: Option<PromptCacheRecord>) {
+        *self
+            .last_prompt_cache_record()
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner) = record;
+    }
+
+    fn last_prompt_cache_record(&self) -> &Arc<Mutex<Option<PromptCacheRecord>>> {
+        &self.last_prompt_cache_record
+    }
+
     pub async fn exchange_oauth_code(
         &self,
         config: &OAuthConfig,
@@ -527,6 +594,7 @@ pub struct MessageStream {
     parser: SseParser,
     pending: VecDeque<StreamEvent>,
     done: bool,
+    cache_tracking: Option<StreamCacheTracking>,
 }
 
 impl MessageStream {
@@ -538,6 +606,9 @@ impl MessageStream {
     pub async fn next_event(&mut self) -> Result<Option<StreamEvent>, ApiError> {
         loop {
             if let Some(event) = self.pending.pop_front() {
+                if let Some(cache_tracking) = &mut self.cache_tracking {
+                    cache_tracking.observe(&event);
+                }
                 return Ok(Some(event));
             }
 
@@ -545,8 +616,14 @@ impl MessageStream {
                 let remaining = self.parser.finish()?;
                 self.pending.extend(remaining);
                 if let Some(event) = self.pending.pop_front() {
+                    if let Some(cache_tracking) = &mut self.cache_tracking {
+                        cache_tracking.observe(&event);
+                    }
                     return Ok(Some(event));
                 }
+                if let Some(cache_tracking) = &mut self.cache_tracking {
+                    cache_tracking.finalize();
+                }
                 return Ok(None);
             }
 
@@ -562,6 +639,53 @@ impl MessageStream {
     }
 }
 
+#[derive(Debug, Clone)]
+struct StreamCacheTracking {
+    prompt_cache: PromptCache,
+    request: MessageRequest,
+    last_usage: Option<Usage>,
+    finalized: bool,
+    last_record: Arc<Mutex<Option<PromptCacheRecord>>>,
+}
+
+impl StreamCacheTracking {
+    fn observe(&mut self, event: &StreamEvent) {
+        match event {
+            StreamEvent::MessageStart(event) => {
+                self.last_usage = Some(event.message.usage.clone());
+            }
+            StreamEvent::MessageDelta(event) => {
+                self.last_usage = Some(event.usage.clone());
+            }
+            StreamEvent::ContentBlockStart(_)
+            | StreamEvent::ContentBlockDelta(_)
+            | StreamEvent::ContentBlockStop(_)
+            | StreamEvent::MessageStop(_) => {}
+        }
+    }
+
+    fn finalize(&mut self) {
+        if self.finalized {
+            return;
+        }
+        if let Some(usage) = &self.last_usage {
+            let record = self.prompt_cache.record_usage(&self.request, usage);
+            *self
+                .last_record
+                .lock()
+                .unwrap_or_else(std::sync::PoisonError::into_inner) = Some(record);
+        }
+        self.finalized = true;
+    }
+}
+
+fn prompt_cache_record_from_stats(stats: PromptCacheStats) -> PromptCacheRecord {
+    PromptCacheRecord {
+        cache_break: None,
+        stats,
+    }
+}
+
 async fn expect_success(response: reqwest::Response) -> Result<reqwest::Response, ApiError> {
     let status = response.status();
     if status.is_success() {
@@ -606,7 +730,7 @@ mod tests {
     use super::{ALT_REQUEST_ID_HEADER, REQUEST_ID_HEADER};
     use std::io::{Read, Write};
     use std::net::TcpListener;
-    use std::sync::{Mutex, OnceLock};
+    use std::sync::atomic::{AtomicU64, Ordering};
     use std::thread;
     use std::time::{Duration, SystemTime, UNIX_EPOCH};
 
@@ -616,19 +740,15 @@ mod tests {
         now_unix_timestamp, oauth_token_is_expired, resolve_saved_oauth_token,
         resolve_startup_auth_source, AnthropicClient, AuthSource, OAuthTokenSet,
     };
+    use crate::test_env_lock;
     use crate::types::{ContentBlockDelta, MessageRequest};
 
-    fn env_lock() -> std::sync::MutexGuard<'static, ()> {
-        static LOCK: OnceLock<Mutex<()>> = OnceLock::new();
-        LOCK.get_or_init(|| Mutex::new(()))
-            .lock()
-            .expect("env lock")
-    }
-
     fn temp_config_home() -> std::path::PathBuf {
+        static NEXT_ID: AtomicU64 = AtomicU64::new(0);
         std::env::temp_dir().join(format!(
-            "api-oauth-test-{}-{}",
+            "api-oauth-test-{}-{}-{}",
             std::process::id(),
+            NEXT_ID.fetch_add(1, Ordering::Relaxed),
             SystemTime::now()
                 .duration_since(UNIX_EPOCH)
                 .expect("time")
@@ -668,7 +788,7 @@ mod tests {
 
     #[test]
     fn read_api_key_requires_presence() {
-        let _guard = env_lock();
+        let _guard = test_env_lock();
         std::env::remove_var("ANTHROPIC_AUTH_TOKEN");
         std::env::remove_var("ANTHROPIC_API_KEY");
         std::env::remove_var("CLAUDE_CONFIG_HOME");
@@ -678,7 +798,7 @@ mod tests {
 
     #[test]
     fn read_api_key_requires_non_empty_value() {
-        let _guard = env_lock();
+        let _guard = test_env_lock();
         std::env::set_var("ANTHROPIC_AUTH_TOKEN", "");
         std::env::remove_var("ANTHROPIC_API_KEY");
         let error = super::read_api_key().expect_err("empty key should error");
@@ -688,7 +808,7 @@ mod tests {
 
     #[test]
     fn read_api_key_prefers_api_key_env() {
-        let _guard = env_lock();
+        let _guard = test_env_lock();
         std::env::set_var("ANTHROPIC_AUTH_TOKEN", "auth-token");
         std::env::set_var("ANTHROPIC_API_KEY", "legacy-key");
         assert_eq!(
@@ -701,7 +821,7 @@ mod tests {
 
     #[test]
     fn read_auth_token_reads_auth_token_env() {
-        let _guard = env_lock();
+        let _guard = test_env_lock();
         std::env::set_var("ANTHROPIC_AUTH_TOKEN", "auth-token");
         assert_eq!(super::read_auth_token().as_deref(), Some("auth-token"));
         std::env::remove_var("ANTHROPIC_AUTH_TOKEN");
@@ -721,7 +841,7 @@ mod tests {
 
     #[test]
     fn auth_source_from_env_combines_api_key_and_bearer_token() {
-        let _guard = env_lock();
+        let _guard = test_env_lock();
         std::env::set_var("ANTHROPIC_AUTH_TOKEN", "auth-token");
         std::env::set_var("ANTHROPIC_API_KEY", "legacy-key");
         let auth = AuthSource::from_env().expect("env auth");
@@ -733,7 +853,7 @@ mod tests {
 
     #[test]
     fn auth_source_from_saved_oauth_when_env_absent() {
-        let _guard = env_lock();
+        let _guard = test_env_lock();
         let config_home = temp_config_home();
         std::env::set_var("CLAUDE_CONFIG_HOME", &config_home);
         std::env::remove_var("ANTHROPIC_AUTH_TOKEN");
@@ -772,7 +892,7 @@ mod tests {
 
     #[test]
     fn resolve_saved_oauth_token_refreshes_expired_credentials() {
-        let _guard = env_lock();
+        let _guard = test_env_lock();
         let config_home = temp_config_home();
         std::env::set_var("CLAUDE_CONFIG_HOME", &config_home);
         std::env::remove_var("ANTHROPIC_AUTH_TOKEN");
@@ -804,7 +924,7 @@ mod tests {
 
     #[test]
     fn resolve_startup_auth_source_uses_saved_oauth_without_loading_config() {
-        let _guard = env_lock();
+        let _guard = test_env_lock();
         let config_home = temp_config_home();
         std::env::set_var("CLAUDE_CONFIG_HOME", &config_home);
         std::env::remove_var("ANTHROPIC_AUTH_TOKEN");
@@ -828,7 +948,7 @@ mod tests {
 
     #[test]
     fn resolve_startup_auth_source_errors_when_refreshable_token_lacks_config() {
-        let _guard = env_lock();
+        let _guard = test_env_lock();
         let config_home = temp_config_home();
         std::env::set_var("CLAUDE_CONFIG_HOME", &config_home);
         std::env::remove_var("ANTHROPIC_AUTH_TOKEN");
@@ -860,7 +980,7 @@ mod tests {
 
     #[test]
     fn resolve_saved_oauth_token_preserves_refresh_token_when_refresh_response_omits_it() {
-        let _guard = env_lock();
+        let _guard = test_env_lock();
         let config_home = temp_config_home();
         std::env::set_var("CLAUDE_CONFIG_HOME", &config_home);
         std::env::remove_var("ANTHROPIC_AUTH_TOKEN");

rust/crates/api/src/lib.rs

@@ -1,5 +1,6 @@
 mod client;
 mod error;
+mod prompt_cache;
 mod sse;
 mod types;
 
@@ -8,6 +9,10 @@ pub use client::{
     AnthropicClient, AuthSource, MessageStream, OAuthTokenSet,
 };
 pub use error::ApiError;
+pub use prompt_cache::{
+    CacheBreakEvent, PromptCache, PromptCacheConfig, PromptCachePaths, PromptCacheRecord,
+    PromptCacheStats,
+};
 pub use sse::{parse_frame, SseParser};
 pub use types::{
     ContentBlockDelta, ContentBlockDeltaEvent, ContentBlockStartEvent, ContentBlockStopEvent,
@@ -15,3 +20,11 @@ pub use types::{
     MessageResponse, MessageStartEvent, MessageStopEvent, OutputContentBlock, StreamEvent,
     ToolChoice, ToolDefinition, ToolResultContentBlock, Usage,
 };
+
+#[cfg(test)]
+pub(crate) fn test_env_lock() -> std::sync::MutexGuard<'static, ()> {
+    static LOCK: std::sync::OnceLock<std::sync::Mutex<()>> = std::sync::OnceLock::new();
+    LOCK.get_or_init(|| std::sync::Mutex::new(()))
+        .lock()
+        .unwrap_or_else(std::sync::PoisonError::into_inner)
+}

rust/crates/api/src/prompt_cache.rs

@@ -0,0 +1,727 @@
+use std::fs;
+use std::path::{Path, PathBuf};
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, SystemTime, UNIX_EPOCH};
+
+use serde::{Deserialize, Serialize};
+
+use crate::types::{MessageRequest, MessageResponse, Usage};
+
+const DEFAULT_COMPLETION_TTL_SECS: u64 = 30;
+const DEFAULT_PROMPT_TTL_SECS: u64 = 5 * 60;
+const DEFAULT_BREAK_MIN_DROP: u32 = 2_000;
+const MAX_SANITIZED_LENGTH: usize = 80;
+const REQUEST_FINGERPRINT_VERSION: u32 = 1;
+const REQUEST_FINGERPRINT_PREFIX: &str = "v1";
+const FNV_OFFSET_BASIS: u64 = 0xcbf2_9ce4_8422_2325;
+const FNV_PRIME: u64 = 0x0000_0100_0000_01b3;
+
+#[derive(Debug, Clone)]
+pub struct PromptCacheConfig {
+    pub session_id: String,
+    pub completion_ttl: Duration,
+    pub prompt_ttl: Duration,
+    pub cache_break_min_drop: u32,
+}
+
+impl PromptCacheConfig {
+    #[must_use]
+    pub fn new(session_id: impl Into<String>) -> Self {
+        Self {
+            session_id: session_id.into(),
+            completion_ttl: Duration::from_secs(DEFAULT_COMPLETION_TTL_SECS),
+            prompt_ttl: Duration::from_secs(DEFAULT_PROMPT_TTL_SECS),
+            cache_break_min_drop: DEFAULT_BREAK_MIN_DROP,
+        }
+    }
+}
+
+impl Default for PromptCacheConfig {
+    fn default() -> Self {
+        Self::new("default")
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct PromptCachePaths {
+    pub root: PathBuf,
+    pub session_dir: PathBuf,
+    pub completion_dir: PathBuf,
+    pub session_state_path: PathBuf,
+    pub stats_path: PathBuf,
+}
+
+impl PromptCachePaths {
+    #[must_use]
+    pub fn for_session(session_id: &str) -> Self {
+        let root = base_cache_root();
+        let session_dir = root.join(sanitize_path_segment(session_id));
+        let completion_dir = session_dir.join("completions");
+        Self {
+            root,
+            session_state_path: session_dir.join("session-state.json"),
+            stats_path: session_dir.join("stats.json"),
+            session_dir,
+            completion_dir,
+        }
+    }
+
+    #[must_use]
+    pub fn completion_entry_path(&self, request_hash: &str) -> PathBuf {
+        self.completion_dir.join(format!("{request_hash}.json"))
+    }
+}
+
+#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize, Deserialize)]
+pub struct PromptCacheStats {
+    pub tracked_requests: u64,
+    pub completion_cache_hits: u64,
+    pub completion_cache_misses: u64,
+    pub completion_cache_writes: u64,
+    pub expected_invalidations: u64,
+    pub unexpected_cache_breaks: u64,
+    pub total_cache_creation_input_tokens: u64,
+    pub total_cache_read_input_tokens: u64,
+    pub last_cache_creation_input_tokens: Option<u32>,
+    pub last_cache_read_input_tokens: Option<u32>,
+    pub last_request_hash: Option<String>,
+    pub last_completion_cache_key: Option<String>,
+    pub last_break_reason: Option<String>,
+    pub last_cache_source: Option<String>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct CacheBreakEvent {
+    pub unexpected: bool,
+    pub reason: String,
+    pub previous_cache_read_input_tokens: u32,
+    pub current_cache_read_input_tokens: u32,
+    pub token_drop: u32,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct PromptCacheRecord {
+    pub cache_break: Option<CacheBreakEvent>,
+    pub stats: PromptCacheStats,
+}
+
+#[derive(Debug, Clone)]
+pub struct PromptCache {
+    inner: Arc<Mutex<PromptCacheInner>>,
+}
+
+impl PromptCache {
+    #[must_use]
+    pub fn new(session_id: impl Into<String>) -> Self {
+        Self::with_config(PromptCacheConfig::new(session_id))
+    }
+
+    #[must_use]
+    pub fn with_config(config: PromptCacheConfig) -> Self {
+        let paths = PromptCachePaths::for_session(&config.session_id);
+        let stats = read_json::<PromptCacheStats>(&paths.stats_path).unwrap_or_default();
+        let previous = read_json::<TrackedPromptState>(&paths.session_state_path);
+        Self {
+            inner: Arc::new(Mutex::new(PromptCacheInner {
+                config,
+                paths,
+                stats,
+                previous,
+            })),
+        }
+    }
+
+    #[must_use]
+    pub fn paths(&self) -> PromptCachePaths {
+        self.lock().paths.clone()
+    }
+
+    #[must_use]
+    pub fn stats(&self) -> PromptCacheStats {
+        self.lock().stats.clone()
+    }
+
+    #[must_use]
+    pub fn lookup_completion(&self, request: &MessageRequest) -> Option<MessageResponse> {
+        let request_hash = request_hash_hex(request);
+        let (paths, ttl) = {
+            let inner = self.lock();
+            (inner.paths.clone(), inner.config.completion_ttl)
+        };
+        let entry_path = paths.completion_entry_path(&request_hash);
+        let entry = read_json::<CompletionCacheEntry>(&entry_path);
+        let Some(entry) = entry else {
+            let mut inner = self.lock();
+            inner.stats.completion_cache_misses += 1;
+            inner.stats.last_completion_cache_key = Some(request_hash);
+            persist_state(&inner);
+            return None;
+        };
+
+        if entry.fingerprint_version != current_fingerprint_version() {
+            let mut inner = self.lock();
+            inner.stats.completion_cache_misses += 1;
+            inner.stats.last_completion_cache_key = Some(request_hash.clone());
+            let _ = fs::remove_file(entry_path);
+            persist_state(&inner);
+            return None;
+        }
+
+        let expired = now_unix_secs().saturating_sub(entry.cached_at_unix_secs) >= ttl.as_secs();
+        let mut inner = self.lock();
+        inner.stats.last_completion_cache_key = Some(request_hash.clone());
+        if expired {
+            inner.stats.completion_cache_misses += 1;
+            let _ = fs::remove_file(entry_path);
+            persist_state(&inner);
+            return None;
+        }
+
+        inner.stats.completion_cache_hits += 1;
+        apply_usage_to_stats(
+            &mut inner.stats,
+            &entry.response.usage,
+            &request_hash,
+            "completion-cache",
+        );
+        inner.previous = Some(TrackedPromptState::from_usage(
+            request,
+            &entry.response.usage,
+        ));
+        persist_state(&inner);
+        Some(entry.response)
+    }
+
+    #[must_use]
+    pub fn record_response(
+        &self,
+        request: &MessageRequest,
+        response: &MessageResponse,
+    ) -> PromptCacheRecord {
+        self.record_usage_internal(request, &response.usage, Some(response))
+    }
+
+    #[must_use]
+    pub fn record_usage(&self, request: &MessageRequest, usage: &Usage) -> PromptCacheRecord {
+        self.record_usage_internal(request, usage, None)
+    }
+
+    fn record_usage_internal(
+        &self,
+        request: &MessageRequest,
+        usage: &Usage,
+        response: Option<&MessageResponse>,
+    ) -> PromptCacheRecord {
+        let request_hash = request_hash_hex(request);
+        let mut inner = self.lock();
+        let previous = inner.previous.clone();
+        let current = TrackedPromptState::from_usage(request, usage);
+        let cache_break = detect_cache_break(&inner.config, previous.as_ref(), &current);
+
+        inner.stats.tracked_requests += 1;
+        apply_usage_to_stats(&mut inner.stats, usage, &request_hash, "api-response");
+        if let Some(event) = &cache_break {
+            if event.unexpected {
+                inner.stats.unexpected_cache_breaks += 1;
+            } else {
+                inner.stats.expected_invalidations += 1;
+            }
+            inner.stats.last_break_reason = Some(event.reason.clone());
+        }
+
+        inner.previous = Some(current);
+        if let Some(response) = response {
+            write_completion_entry(&inner.paths, &request_hash, response);
+            inner.stats.completion_cache_writes += 1;
+        }
+        persist_state(&inner);
+
+        PromptCacheRecord {
+            cache_break,
+            stats: inner.stats.clone(),
+        }
+    }
+
+    fn lock(&self) -> std::sync::MutexGuard<'_, PromptCacheInner> {
+        self.inner
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+    }
+}
+
+#[derive(Debug)]
+struct PromptCacheInner {
+    config: PromptCacheConfig,
+    paths: PromptCachePaths,
+    stats: PromptCacheStats,
+    previous: Option<TrackedPromptState>,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+struct CompletionCacheEntry {
+    cached_at_unix_secs: u64,
+    #[serde(default = "current_fingerprint_version")]
+    fingerprint_version: u32,
+    response: MessageResponse,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+struct TrackedPromptState {
+    observed_at_unix_secs: u64,
+    #[serde(default = "current_fingerprint_version")]
+    fingerprint_version: u32,
+    model_hash: u64,
+    system_hash: u64,
+    tools_hash: u64,
+    messages_hash: u64,
+    cache_read_input_tokens: u32,
+}
+
+impl TrackedPromptState {
+    fn from_usage(request: &MessageRequest, usage: &Usage) -> Self {
+        let hashes = RequestFingerprints::from_request(request);
+        Self {
+            observed_at_unix_secs: now_unix_secs(),
+            fingerprint_version: current_fingerprint_version(),
+            model_hash: hashes.model,
+            system_hash: hashes.system,
+            tools_hash: hashes.tools,
+            messages_hash: hashes.messages,
+            cache_read_input_tokens: usage.cache_read_input_tokens,
+        }
+    }
+}
+
+#[derive(Debug, Clone, Copy)]
+struct RequestFingerprints {
+    model: u64,
+    system: u64,
+    tools: u64,
+    messages: u64,
+}
+
+impl RequestFingerprints {
+    fn from_request(request: &MessageRequest) -> Self {
+        Self {
+            model: hash_serializable(&request.model),
+            system: hash_serializable(&request.system),
+            tools: hash_serializable(&request.tools),
+            messages: hash_serializable(&request.messages),
+        }
+    }
+}
+
+fn detect_cache_break(
+    config: &PromptCacheConfig,
+    previous: Option<&TrackedPromptState>,
+    current: &TrackedPromptState,
+) -> Option<CacheBreakEvent> {
+    let previous = previous?;
+    if previous.fingerprint_version != current.fingerprint_version {
+        return Some(CacheBreakEvent {
+            unexpected: false,
+            reason: format!(
+                "fingerprint version changed (v{} -> v{})",
+                previous.fingerprint_version, current.fingerprint_version
+            ),
+            previous_cache_read_input_tokens: previous.cache_read_input_tokens,
+            current_cache_read_input_tokens: current.cache_read_input_tokens,
+            token_drop: previous
+                .cache_read_input_tokens
+                .saturating_sub(current.cache_read_input_tokens),
+        });
+    }
+    let token_drop = previous
+        .cache_read_input_tokens
+        .saturating_sub(current.cache_read_input_tokens);
+    if token_drop < config.cache_break_min_drop {
+        return None;
+    }
+
+    let mut reasons = Vec::new();
+    if previous.model_hash != current.model_hash {
+        reasons.push("model changed");
+    }
+    if previous.system_hash != current.system_hash {
+        reasons.push("system prompt changed");
+    }
+    if previous.tools_hash != current.tools_hash {
+        reasons.push("tool definitions changed");
+    }
+    if previous.messages_hash != current.messages_hash {
+        reasons.push("message payload changed");
+    }
+
+    let elapsed = current
+        .observed_at_unix_secs
+        .saturating_sub(previous.observed_at_unix_secs);
+
+    let (unexpected, reason) = if reasons.is_empty() {
+        if elapsed > config.prompt_ttl.as_secs() {
+            (
+                false,
+                format!("possible prompt cache TTL expiry after {elapsed}s"),
+            )
+        } else {
+            (
+                true,
+                "cache read tokens dropped while prompt fingerprint remained stable".to_string(),
+            )
+        }
+    } else {
+        (false, reasons.join(", "))
+    };
+
+    Some(CacheBreakEvent {
+        unexpected,
+        reason,
+        previous_cache_read_input_tokens: previous.cache_read_input_tokens,
+        current_cache_read_input_tokens: current.cache_read_input_tokens,
+        token_drop,
+    })
+}
+
+fn apply_usage_to_stats(
+    stats: &mut PromptCacheStats,
+    usage: &Usage,
+    request_hash: &str,
+    source: &str,
+) {
+    stats.total_cache_creation_input_tokens += u64::from(usage.cache_creation_input_tokens);
+    stats.total_cache_read_input_tokens += u64::from(usage.cache_read_input_tokens);
+    stats.last_cache_creation_input_tokens = Some(usage.cache_creation_input_tokens);
+    stats.last_cache_read_input_tokens = Some(usage.cache_read_input_tokens);
+    stats.last_request_hash = Some(request_hash.to_string());
+    stats.last_cache_source = Some(source.to_string());
+}
+
+fn persist_state(inner: &PromptCacheInner) {
+    let _ = ensure_cache_dirs(&inner.paths);
+    let _ = write_json(&inner.paths.stats_path, &inner.stats);
+    if let Some(previous) = &inner.previous {
+        let _ = write_json(&inner.paths.session_state_path, previous);
+    }
+}
+
+fn write_completion_entry(
+    paths: &PromptCachePaths,
+    request_hash: &str,
+    response: &MessageResponse,
+) {
+    let _ = ensure_cache_dirs(paths);
+    let entry = CompletionCacheEntry {
+        cached_at_unix_secs: now_unix_secs(),
+        fingerprint_version: current_fingerprint_version(),
+        response: response.clone(),
+    };
+    let _ = write_json(&paths.completion_entry_path(request_hash), &entry);
+}
+
+fn ensure_cache_dirs(paths: &PromptCachePaths) -> std::io::Result<()> {
+    fs::create_dir_all(&paths.completion_dir)
+}
+
+fn write_json<T: Serialize>(path: &Path, value: &T) -> std::io::Result<()> {
+    let json = serde_json::to_vec_pretty(value)
+        .map_err(|error| std::io::Error::new(std::io::ErrorKind::InvalidData, error))?;
+    fs::write(path, json)
+}
+
+fn read_json<T: for<'de> Deserialize<'de>>(path: &Path) -> Option<T> {
+    let bytes = fs::read(path).ok()?;
+    serde_json::from_slice(&bytes).ok()
+}
+
+fn request_hash_hex(request: &MessageRequest) -> String {
+    format!(
+        "{REQUEST_FINGERPRINT_PREFIX}-{:016x}",
+        hash_serializable(request)
+    )
+}
+
+fn hash_serializable<T: Serialize>(value: &T) -> u64 {
+    let json = serde_json::to_vec(value).unwrap_or_default();
+    stable_hash_bytes(&json)
+}
+
+fn sanitize_path_segment(value: &str) -> String {
+    let sanitized: String = value
+        .chars()
+        .map(|ch| if ch.is_ascii_alphanumeric() { ch } else { '-' })
+        .collect();
+    if sanitized.len() <= MAX_SANITIZED_LENGTH {
+        return sanitized;
+    }
+    let suffix = format!("-{:x}", hash_string(value));
+    format!(
+        "{}{}",
+        &sanitized[..MAX_SANITIZED_LENGTH.saturating_sub(suffix.len())],
+        suffix
+    )
+}
+
+fn hash_string(value: &str) -> u64 {
+    stable_hash_bytes(value.as_bytes())
+}
+
+fn base_cache_root() -> PathBuf {
+    if let Some(config_home) = std::env::var_os("CLAUDE_CONFIG_HOME") {
+        return PathBuf::from(config_home)
+            .join("cache")
+            .join("prompt-cache");
+    }
+    if let Some(home) = std::env::var_os("HOME") {
+        return PathBuf::from(home)
+            .join(".claude")
+            .join("cache")
+            .join("prompt-cache");
+    }
+    std::env::temp_dir().join("claude-prompt-cache")
+}
+
+fn now_unix_secs() -> u64 {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map_or(0, |duration| duration.as_secs())
+}
+
+const fn current_fingerprint_version() -> u32 {
+    REQUEST_FINGERPRINT_VERSION
+}
+
+fn stable_hash_bytes(bytes: &[u8]) -> u64 {
+    let mut hash = FNV_OFFSET_BASIS;
+    for byte in bytes {
+        hash ^= u64::from(*byte);
+        hash = hash.wrapping_mul(FNV_PRIME);
+    }
+    hash
+}
+
+#[cfg(test)]
+mod tests {
+    use std::time::{Duration, SystemTime, UNIX_EPOCH};
+
+    use super::{
+        detect_cache_break, read_json, request_hash_hex, sanitize_path_segment, PromptCache,
+        PromptCacheConfig, PromptCachePaths, TrackedPromptState, REQUEST_FINGERPRINT_PREFIX,
+    };
+    use crate::test_env_lock;
+    use crate::types::{InputMessage, MessageRequest, MessageResponse, OutputContentBlock, Usage};
+
+    #[test]
+    fn path_builder_sanitizes_session_identifier() {
+        let paths = PromptCachePaths::for_session("session:/with spaces");
+        let session_dir = paths
+            .session_dir
+            .file_name()
+            .and_then(|value| value.to_str())
+            .expect("session dir name");
+        assert_eq!(session_dir, "session--with-spaces");
+        assert!(paths.completion_dir.ends_with("completions"));
+        assert!(paths.stats_path.ends_with("stats.json"));
+        assert!(paths.session_state_path.ends_with("session-state.json"));
+    }
+
+    #[test]
+    fn request_fingerprint_drives_unexpected_break_detection() {
+        let request = sample_request("same");
+        let previous = TrackedPromptState::from_usage(
+            &request,
+            &Usage {
+                input_tokens: 0,
+                cache_creation_input_tokens: 0,
+                cache_read_input_tokens: 6_000,
+                output_tokens: 0,
+            },
+        );
+        let current = TrackedPromptState::from_usage(
+            &request,
+            &Usage {
+                input_tokens: 0,
+                cache_creation_input_tokens: 0,
+                cache_read_input_tokens: 1_000,
+                output_tokens: 0,
+            },
+        );
+        let event = detect_cache_break(&PromptCacheConfig::default(), Some(&previous), &current)
+            .expect("break should be detected");
+        assert!(event.unexpected);
+        assert!(event.reason.contains("stable"));
+    }
+
+    #[test]
+    fn changed_prompt_marks_break_as_expected() {
+        let previous_request = sample_request("first");
+        let current_request = sample_request("second");
+        let previous = TrackedPromptState::from_usage(
+            &previous_request,
+            &Usage {
+                input_tokens: 0,
+                cache_creation_input_tokens: 0,
+                cache_read_input_tokens: 6_000,
+                output_tokens: 0,
+            },
+        );
+        let current = TrackedPromptState::from_usage(
+            &current_request,
+            &Usage {
+                input_tokens: 0,
+                cache_creation_input_tokens: 0,
+                cache_read_input_tokens: 1_000,
+                output_tokens: 0,
+            },
+        );
+        let event = detect_cache_break(&PromptCacheConfig::default(), Some(&previous), &current)
+            .expect("break should be detected");
+        assert!(!event.unexpected);
+        assert!(event.reason.contains("message payload changed"));
+    }
+
+    #[test]
+    fn completion_cache_round_trip_persists_recent_response() {
+        let _guard = test_env_lock();
+        let temp_root = std::env::temp_dir().join(format!(
+            "prompt-cache-test-{}-{}",
+            std::process::id(),
+            SystemTime::now()
+                .duration_since(UNIX_EPOCH)
+                .expect("time")
+                .as_nanos()
+        ));
+        std::env::set_var("CLAUDE_CONFIG_HOME", &temp_root);
+        let cache = PromptCache::new("unit-test-session");
+        let request = sample_request("cache me");
+        let response = sample_response(42, 12, "cached");
+
+        assert!(cache.lookup_completion(&request).is_none());
+        let record = cache.record_response(&request, &response);
+        assert!(record.cache_break.is_none());
+
+        let cached = cache
+            .lookup_completion(&request)
+            .expect("cached response should load");
+        assert_eq!(cached.content, response.content);
+
+        let stats = cache.stats();
+        assert_eq!(stats.completion_cache_hits, 1);
+        assert_eq!(stats.completion_cache_misses, 1);
+        assert_eq!(stats.completion_cache_writes, 1);
+
+        let persisted = read_json::<super::PromptCacheStats>(&cache.paths().stats_path)
+            .expect("stats should persist");
+        assert_eq!(persisted.completion_cache_hits, 1);
+
+        std::fs::remove_dir_all(temp_root).expect("cleanup temp root");
+        std::env::remove_var("CLAUDE_CONFIG_HOME");
+    }
+
+    #[test]
+    fn distinct_requests_do_not_collide_in_completion_cache() {
+        let _guard = test_env_lock();
+        let temp_root = std::env::temp_dir().join(format!(
+            "prompt-cache-distinct-{}-{}",
+            std::process::id(),
+            SystemTime::now()
+                .duration_since(UNIX_EPOCH)
+                .expect("time")
+                .as_nanos()
+        ));
+        std::env::set_var("CLAUDE_CONFIG_HOME", &temp_root);
+        let cache = PromptCache::new("distinct-request-session");
+        let first_request = sample_request("first");
+        let second_request = sample_request("second");
+
+        let response = sample_response(42, 12, "cached");
+        let _ = cache.record_response(&first_request, &response);
+
+        assert!(cache.lookup_completion(&second_request).is_none());
+
+        std::fs::remove_dir_all(temp_root).expect("cleanup temp root");
+        std::env::remove_var("CLAUDE_CONFIG_HOME");
+    }
+
+    #[test]
+    fn expired_completion_entries_are_not_reused() {
+        let _guard = test_env_lock();
+        let temp_root = std::env::temp_dir().join(format!(
+            "prompt-cache-expired-{}-{}",
+            std::process::id(),
+            SystemTime::now()
+                .duration_since(UNIX_EPOCH)
+                .expect("time")
+                .as_nanos()
+        ));
+        std::env::set_var("CLAUDE_CONFIG_HOME", &temp_root);
+        let cache = PromptCache::with_config(PromptCacheConfig {
+            session_id: "expired-session".to_string(),
+            completion_ttl: Duration::ZERO,
+            ..PromptCacheConfig::default()
+        });
+        let request = sample_request("expire me");
+        let response = sample_response(7, 3, "stale");
+
+        let _ = cache.record_response(&request, &response);
+
+        assert!(cache.lookup_completion(&request).is_none());
+        let stats = cache.stats();
+        assert_eq!(stats.completion_cache_hits, 0);
+        assert_eq!(stats.completion_cache_misses, 1);
+
+        std::fs::remove_dir_all(temp_root).expect("cleanup temp root");
+        std::env::remove_var("CLAUDE_CONFIG_HOME");
+    }
+
+    #[test]
+    fn sanitize_path_caps_long_values() {
+        let long_value = "x".repeat(200);
+        let sanitized = sanitize_path_segment(&long_value);
+        assert!(sanitized.len() <= 80);
+    }
+
+    #[test]
+    fn request_hashes_are_versioned_and_stable() {
+        let request = sample_request("stable");
+        let first = request_hash_hex(&request);
+        let second = request_hash_hex(&request);
+        assert_eq!(first, second);
+        assert!(first.starts_with(REQUEST_FINGERPRINT_PREFIX));
+    }
+
+    fn sample_request(text: &str) -> MessageRequest {
+        MessageRequest {
+            model: "claude-3-7-sonnet-latest".to_string(),
+            max_tokens: 64,
+            messages: vec![InputMessage::user_text(text)],
+            system: Some("system".to_string()),
+            tools: None,
+            tool_choice: None,
+            stream: false,
+        }
+    }
+
+    fn sample_response(
+        cache_read_input_tokens: u32,
+        output_tokens: u32,
+        text: &str,
+    ) -> MessageResponse {
+        MessageResponse {
+            id: "msg_test".to_string(),
+            kind: "message".to_string(),
+            role: "assistant".to_string(),
+            content: vec![OutputContentBlock::Text {
+                text: text.to_string(),
+            }],
+            model: "claude-3-7-sonnet-latest".to_string(),
+            stop_reason: Some("end_turn".to_string()),
+            stop_sequence: None,
+            usage: Usage {
+                input_tokens: 10,
+                cache_creation_input_tokens: 5,
+                cache_read_input_tokens,
+                output_tokens,
+            },
+            request_id: Some("req_test".to_string()),
+        }
+    }
+}

rust/crates/api/tests/client_integration.rs

@@ -1,17 +1,25 @@
 use std::collections::HashMap;
 use std::sync::Arc;
+use std::sync::{Mutex as StdMutex, OnceLock};
 use std::time::Duration;
 
 use api::{
     AnthropicClient, ApiError, ContentBlockDelta, ContentBlockDeltaEvent, ContentBlockStartEvent,
     InputContentBlock, InputMessage, MessageDeltaEvent, MessageRequest, OutputContentBlock,
-    StreamEvent, ToolChoice, ToolDefinition,
+    PromptCache, StreamEvent, ToolChoice, ToolDefinition,
 };
 use serde_json::json;
 use tokio::io::{AsyncReadExt, AsyncWriteExt};
 use tokio::net::TcpListener;
 use tokio::sync::Mutex;
 
+fn env_lock() -> std::sync::MutexGuard<'static, ()> {
+    static LOCK: OnceLock<StdMutex<()>> = OnceLock::new();
+    LOCK.get_or_init(|| StdMutex::new(()))
+        .lock()
+        .unwrap_or_else(std::sync::PoisonError::into_inner)
+}
+
 #[tokio::test]
 async fn send_message_posts_json_and_parses_response() {
     let state = Arc::new(Mutex::new(Vec::<CapturedRequest>::new()));
@@ -45,6 +53,8 @@ async fn send_message_posts_json_and_parses_response() {
     assert_eq!(response.id, "msg_test");
     assert_eq!(response.total_tokens(), 16);
     assert_eq!(response.request_id.as_deref(), Some("req_body_123"));
+    assert_eq!(response.usage.cache_creation_input_tokens, 0);
+    assert_eq!(response.usage.cache_read_input_tokens, 0);
     assert_eq!(
         response.content,
         vec![OutputContentBlock::Text {
@@ -76,11 +86,55 @@ async fn send_message_posts_json_and_parses_response() {
 }
 
 #[tokio::test]
+async fn send_message_parses_prompt_cache_token_usage_from_response() {
+    let state = Arc::new(Mutex::new(Vec::<CapturedRequest>::new()));
+    let body = concat!(
+        "{",
+        "\"id\":\"msg_cache_tokens\",",
+        "\"type\":\"message\",",
+        "\"role\":\"assistant\",",
+        "\"content\":[{\"type\":\"text\",\"text\":\"Cache tokens\"}],",
+        "\"model\":\"claude-3-7-sonnet-latest\",",
+        "\"stop_reason\":\"end_turn\",",
+        "\"stop_sequence\":null,",
+        "\"usage\":{\"input_tokens\":12,\"cache_creation_input_tokens\":321,\"cache_read_input_tokens\":654,\"output_tokens\":4}",
+        "}"
+    );
+    let server = spawn_server(
+        state,
+        vec![http_response("200 OK", "application/json", body)],
+    )
+    .await;
+
+    let client = AnthropicClient::new("test-key").with_base_url(server.base_url());
+    let response = client
+        .send_message(&sample_request(false))
+        .await
+        .expect("request should succeed");
+
+    assert_eq!(response.usage.input_tokens, 12);
+    assert_eq!(response.usage.cache_creation_input_tokens, 321);
+    assert_eq!(response.usage.cache_read_input_tokens, 654);
+    assert_eq!(response.usage.output_tokens, 4);
+}
+
+#[tokio::test]
+#[allow(clippy::await_holding_lock)]
 async fn stream_message_parses_sse_events_with_tool_use() {
+    let _guard = env_lock();
+    let temp_root = std::env::temp_dir().join(format!(
+        "api-stream-cache-{}-{}",
+        std::process::id(),
+        std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .expect("time")
+            .as_nanos()
+    ));
+    std::env::set_var("CLAUDE_CONFIG_HOME", &temp_root);
     let state = Arc::new(Mutex::new(Vec::<CapturedRequest>::new()));
     let sse = concat!(
         "event: message_start\n",
-        "data: {\"type\":\"message_start\",\"message\":{\"id\":\"msg_stream\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[],\"model\":\"claude-3-7-sonnet-latest\",\"stop_reason\":null,\"stop_sequence\":null,\"usage\":{\"input_tokens\":8,\"output_tokens\":0}}}\n\n",
+        "data: {\"type\":\"message_start\",\"message\":{\"id\":\"msg_stream\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[],\"model\":\"claude-3-7-sonnet-latest\",\"stop_reason\":null,\"stop_sequence\":null,\"usage\":{\"input_tokens\":8,\"cache_creation_input_tokens\":13,\"cache_read_input_tokens\":21,\"output_tokens\":0}}}\n\n",
         "event: content_block_start\n",
         "data: {\"type\":\"content_block_start\",\"index\":0,\"content_block\":{\"type\":\"tool_use\",\"id\":\"toolu_123\",\"name\":\"get_weather\",\"input\":{}}}\n\n",
         "event: content_block_delta\n",
@@ -88,7 +142,7 @@ async fn stream_message_parses_sse_events_with_tool_use() {
         "event: content_block_stop\n",
         "data: {\"type\":\"content_block_stop\",\"index\":0}\n\n",
         "event: message_delta\n",
-        "data: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":\"tool_use\",\"stop_sequence\":null},\"usage\":{\"input_tokens\":8,\"output_tokens\":1}}\n\n",
+        "data: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":\"tool_use\",\"stop_sequence\":null},\"usage\":{\"input_tokens\":8,\"cache_creation_input_tokens\":34,\"cache_read_input_tokens\":55,\"output_tokens\":1}}\n\n",
         "event: message_stop\n",
         "data: {\"type\":\"message_stop\"}\n\n",
         "data: [DONE]\n\n"
@@ -106,7 +160,8 @@ async fn stream_message_parses_sse_events_with_tool_use() {
 
     let client = AnthropicClient::new("test-key")
         .with_auth_token(Some("proxy-token".to_string()))
-        .with_base_url(server.base_url());
+        .with_base_url(server.base_url())
+        .with_prompt_cache(PromptCache::new("stream-session"));
     let mut stream = client
         .stream_message(&sample_request(false))
         .await
@@ -160,6 +215,20 @@ async fn stream_message_parses_sse_events_with_tool_use() {
     let captured = state.lock().await;
     let request = captured.first().expect("server should capture request");
     assert!(request.body.contains("\"stream\":true"));
+
+    let cache_stats = client
+        .prompt_cache_stats()
+        .expect("prompt cache stats should exist");
+    assert_eq!(cache_stats.tracked_requests, 1);
+    assert_eq!(cache_stats.last_cache_creation_input_tokens, Some(34));
+    assert_eq!(cache_stats.last_cache_read_input_tokens, Some(55));
+    assert_eq!(
+        cache_stats.last_cache_source.as_deref(),
+        Some("api-response")
+    );
+
+    std::fs::remove_dir_all(temp_root).expect("cleanup temp root");
+    std::env::remove_var("CLAUDE_CONFIG_HOME");
 }
 
 #[tokio::test]
@@ -243,6 +312,121 @@ async fn surfaces_retry_exhaustion_for_persistent_retryable_errors() {
     }
 }
 
+#[tokio::test]
+#[allow(clippy::await_holding_lock)]
+async fn send_message_reuses_recent_completion_cache_entries() {
+    let _guard = env_lock();
+    let temp_root = std::env::temp_dir().join(format!(
+        "api-prompt-cache-{}-{}",
+        std::process::id(),
+        std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .expect("time")
+            .as_nanos()
+    ));
+    std::env::set_var("CLAUDE_CONFIG_HOME", &temp_root);
+
+    let state = Arc::new(Mutex::new(Vec::<CapturedRequest>::new()));
+    let server = spawn_server(
+        state.clone(),
+        vec![http_response(
+            "200 OK",
+            "application/json",
+            "{\"id\":\"msg_cached\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[{\"type\":\"text\",\"text\":\"Cached once\"}],\"model\":\"claude-3-7-sonnet-latest\",\"stop_reason\":\"end_turn\",\"stop_sequence\":null,\"usage\":{\"input_tokens\":3,\"cache_creation_input_tokens\":5,\"cache_read_input_tokens\":4000,\"output_tokens\":2}}",
+        )],
+    )
+    .await;
+
+    let client = AnthropicClient::new("test-key")
+        .with_base_url(server.base_url())
+        .with_prompt_cache(PromptCache::new("integration-session"));
+
+    let first = client
+        .send_message(&sample_request(false))
+        .await
+        .expect("first request should succeed");
+    let second = client
+        .send_message(&sample_request(false))
+        .await
+        .expect("second request should reuse cache");
+
+    assert_eq!(first.content, second.content);
+    assert_eq!(state.lock().await.len(), 1);
+
+    let cache_stats = client
+        .prompt_cache_stats()
+        .expect("prompt cache stats should exist");
+    assert_eq!(cache_stats.completion_cache_hits, 1);
+    assert_eq!(cache_stats.completion_cache_misses, 1);
+    assert_eq!(cache_stats.completion_cache_writes, 1);
+
+    std::fs::remove_dir_all(temp_root).expect("cleanup temp root");
+    std::env::remove_var("CLAUDE_CONFIG_HOME");
+}
+
+#[tokio::test]
+#[allow(clippy::await_holding_lock)]
+async fn send_message_tracks_unexpected_prompt_cache_breaks() {
+    let _guard = env_lock();
+    let temp_root = std::env::temp_dir().join(format!(
+        "api-prompt-break-{}-{}",
+        std::process::id(),
+        std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .expect("time")
+            .as_nanos()
+    ));
+    std::env::set_var("CLAUDE_CONFIG_HOME", &temp_root);
+
+    let state = Arc::new(Mutex::new(Vec::<CapturedRequest>::new()));
+    let server = spawn_server(
+        state,
+        vec![
+            http_response(
+                "200 OK",
+                "application/json",
+                "{\"id\":\"msg_one\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[{\"type\":\"text\",\"text\":\"One\"}],\"model\":\"claude-3-7-sonnet-latest\",\"stop_reason\":\"end_turn\",\"stop_sequence\":null,\"usage\":{\"input_tokens\":3,\"cache_creation_input_tokens\":5,\"cache_read_input_tokens\":6000,\"output_tokens\":2}}",
+            ),
+            http_response(
+                "200 OK",
+                "application/json",
+                "{\"id\":\"msg_two\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[{\"type\":\"text\",\"text\":\"Two\"}],\"model\":\"claude-3-7-sonnet-latest\",\"stop_reason\":\"end_turn\",\"stop_sequence\":null,\"usage\":{\"input_tokens\":3,\"cache_creation_input_tokens\":0,\"cache_read_input_tokens\":1000,\"output_tokens\":2}}",
+            ),
+        ],
+    )
+    .await;
+
+    let request = sample_request(false);
+    let client = AnthropicClient::new("test-key")
+        .with_base_url(server.base_url())
+        .with_prompt_cache(PromptCache::with_config(api::PromptCacheConfig {
+            session_id: "break-session".to_string(),
+            completion_ttl: Duration::from_secs(0),
+            ..api::PromptCacheConfig::default()
+        }));
+
+    client
+        .send_message(&request)
+        .await
+        .expect("first response should succeed");
+    client
+        .send_message(&request)
+        .await
+        .expect("second response should succeed");
+
+    let cache_stats = client
+        .prompt_cache_stats()
+        .expect("prompt cache stats should exist");
+    assert_eq!(cache_stats.unexpected_cache_breaks, 1);
+    assert_eq!(
+        cache_stats.last_break_reason.as_deref(),
+        Some("cache read tokens dropped while prompt fingerprint remained stable")
+    );
+
+    std::fs::remove_dir_all(temp_root).expect("cleanup temp root");
+    std::env::remove_var("CLAUDE_CONFIG_HOME");
+}
+
 #[tokio::test]
 #[ignore = "requires ANTHROPIC_API_KEY and network access"]
 async fn live_stream_smoke_test() {

rust/crates/commands/src/lib.rs

@@ -117,48 +117,6 @@ const SLASH_COMMAND_SPECS: &[SlashCommandSpec] = &[
         argument_hint: None,
         resume_supported: true,
     },
-    SlashCommandSpec {
-        name: "bughunter",
-        summary: "Inspect the codebase for likely bugs",
-        argument_hint: Some("[scope]"),
-        resume_supported: false,
-    },
-    SlashCommandSpec {
-        name: "commit",
-        summary: "Generate a commit message and create a git commit",
-        argument_hint: None,
-        resume_supported: false,
-    },
-    SlashCommandSpec {
-        name: "pr",
-        summary: "Draft or create a pull request from the conversation",
-        argument_hint: Some("[context]"),
-        resume_supported: false,
-    },
-    SlashCommandSpec {
-        name: "issue",
-        summary: "Draft or create a GitHub issue from the conversation",
-        argument_hint: Some("[context]"),
-        resume_supported: false,
-    },
-    SlashCommandSpec {
-        name: "ultraplan",
-        summary: "Run a deep planning prompt with multi-step reasoning",
-        argument_hint: Some("[task]"),
-        resume_supported: false,
-    },
-    SlashCommandSpec {
-        name: "teleport",
-        summary: "Jump to a file or symbol by searching the workspace",
-        argument_hint: Some("<symbol-or-path>"),
-        resume_supported: false,
-    },
-    SlashCommandSpec {
-        name: "debug-tool-call",
-        summary: "Replay the last tool call with debug details",
-        argument_hint: None,
-        resume_supported: false,
-    },
     SlashCommandSpec {
         name: "export",
         summary: "Export the current conversation to a file",
@@ -178,23 +136,6 @@ pub enum SlashCommand {
     Help,
     Status,
     Compact,
-    Bughunter {
-        scope: Option<String>,
-    },
-    Commit,
-    Pr {
-        context: Option<String>,
-    },
-    Issue {
-        context: Option<String>,
-    },
-    Ultraplan {
-        task: Option<String>,
-    },
-    Teleport {
-        target: Option<String>,
-    },
-    DebugToolCall,
     Model {
         model: Option<String>,
     },
@@ -239,23 +180,6 @@ impl SlashCommand {
             "help" => Self::Help,
             "status" => Self::Status,
             "compact" => Self::Compact,
-            "bughunter" => Self::Bughunter {
-                scope: remainder_after_command(trimmed, command),
-            },
-            "commit" => Self::Commit,
-            "pr" => Self::Pr {
-                context: remainder_after_command(trimmed, command),
-            },
-            "issue" => Self::Issue {
-                context: remainder_after_command(trimmed, command),
-            },
-            "ultraplan" => Self::Ultraplan {
-                task: remainder_after_command(trimmed, command),
-            },
-            "teleport" => Self::Teleport {
-                target: remainder_after_command(trimmed, command),
-            },
-            "debug-tool-call" => Self::DebugToolCall,
             "model" => Self::Model {
                 model: parts.next().map(ToOwned::to_owned),
             },
@@ -288,15 +212,6 @@ impl SlashCommand {
     }
 }
 
-fn remainder_after_command(input: &str, command: &str) -> Option<String> {
-    input
-        .trim()
-        .strip_prefix(&format!("/{command}"))
-        .map(str::trim)
-        .filter(|value| !value.is_empty())
-        .map(ToOwned::to_owned)
-}
-
 #[must_use]
 pub fn slash_command_specs() -> &'static [SlashCommandSpec] {
     SLASH_COMMAND_SPECS
@@ -364,13 +279,6 @@ pub fn handle_slash_command(
             session: session.clone(),
         }),
         SlashCommand::Status
-        | SlashCommand::Bughunter { .. }
-        | SlashCommand::Commit
-        | SlashCommand::Pr { .. }
-        | SlashCommand::Issue { .. }
-        | SlashCommand::Ultraplan { .. }
-        | SlashCommand::Teleport { .. }
-        | SlashCommand::DebugToolCall
         | SlashCommand::Model { .. }
         | SlashCommand::Permissions { .. }
         | SlashCommand::Clear { .. }
@@ -399,41 +307,6 @@ mod tests {
     fn parses_supported_slash_commands() {
         assert_eq!(SlashCommand::parse("/help"), Some(SlashCommand::Help));
         assert_eq!(SlashCommand::parse(" /status "), Some(SlashCommand::Status));
-        assert_eq!(
-            SlashCommand::parse("/bughunter runtime"),
-            Some(SlashCommand::Bughunter {
-                scope: Some("runtime".to_string())
-            })
-        );
-        assert_eq!(SlashCommand::parse("/commit"), Some(SlashCommand::Commit));
-        assert_eq!(
-            SlashCommand::parse("/pr ready for review"),
-            Some(SlashCommand::Pr {
-                context: Some("ready for review".to_string())
-            })
-        );
-        assert_eq!(
-            SlashCommand::parse("/issue flaky test"),
-            Some(SlashCommand::Issue {
-                context: Some("flaky test".to_string())
-            })
-        );
-        assert_eq!(
-            SlashCommand::parse("/ultraplan ship both features"),
-            Some(SlashCommand::Ultraplan {
-                task: Some("ship both features".to_string())
-            })
-        );
-        assert_eq!(
-            SlashCommand::parse("/teleport conversation.rs"),
-            Some(SlashCommand::Teleport {
-                target: Some("conversation.rs".to_string())
-            })
-        );
-        assert_eq!(
-            SlashCommand::parse("/debug-tool-call"),
-            Some(SlashCommand::DebugToolCall)
-        );
         assert_eq!(
             SlashCommand::parse("/model claude-opus"),
             Some(SlashCommand::Model {
@@ -501,13 +374,6 @@ mod tests {
         assert!(help.contains("/help"));
         assert!(help.contains("/status"));
         assert!(help.contains("/compact"));
-        assert!(help.contains("/bughunter [scope]"));
-        assert!(help.contains("/commit"));
-        assert!(help.contains("/pr [context]"));
-        assert!(help.contains("/issue [context]"));
-        assert!(help.contains("/ultraplan [task]"));
-        assert!(help.contains("/teleport <symbol-or-path>"));
-        assert!(help.contains("/debug-tool-call"));
         assert!(help.contains("/model [model]"));
         assert!(help.contains("/permissions [read-only|workspace-write|danger-full-access]"));
         assert!(help.contains("/clear [--confirm]"));
@@ -520,7 +386,7 @@ mod tests {
         assert!(help.contains("/version"));
         assert!(help.contains("/export [file]"));
         assert!(help.contains("/session [list|switch <session-id>]"));
-        assert_eq!(slash_command_specs().len(), 22);
+        assert_eq!(slash_command_specs().len(), 15);
         assert_eq!(resume_supported_slash_commands().len(), 11);
     }
 
@@ -568,22 +434,6 @@ mod tests {
         let session = Session::new();
         assert!(handle_slash_command("/unknown", &session, CompactionConfig::default()).is_none());
         assert!(handle_slash_command("/status", &session, CompactionConfig::default()).is_none());
-        assert!(
-            handle_slash_command("/bughunter", &session, CompactionConfig::default()).is_none()
-        );
-        assert!(handle_slash_command("/commit", &session, CompactionConfig::default()).is_none());
-        assert!(handle_slash_command("/pr", &session, CompactionConfig::default()).is_none());
-        assert!(handle_slash_command("/issue", &session, CompactionConfig::default()).is_none());
-        assert!(
-            handle_slash_command("/ultraplan", &session, CompactionConfig::default()).is_none()
-        );
-        assert!(
-            handle_slash_command("/teleport foo", &session, CompactionConfig::default()).is_none()
-        );
-        assert!(
-            handle_slash_command("/debug-tool-call", &session, CompactionConfig::default())
-                .is_none()
-        );
         assert!(
             handle_slash_command("/model claude", &session, CompactionConfig::default()).is_none()
         );

rust/crates/runtime/src/config.rs

@@ -37,6 +37,7 @@ pub struct RuntimeConfig {
 
 #[derive(Debug, Clone, PartialEq, Eq, Default)]
 pub struct RuntimeFeatureConfig {
+    hooks: RuntimeHookConfig,
     mcp: McpConfigCollection,
     oauth: Option<OAuthConfig>,
     model: Option<String>,
@@ -44,6 +45,12 @@ pub struct RuntimeFeatureConfig {
     sandbox: SandboxConfig,
 }
 
+#[derive(Debug, Clone, PartialEq, Eq, Default)]
+pub struct RuntimeHookConfig {
+    pre_tool_use: Vec<String>,
+    post_tool_use: Vec<String>,
+}
+
 #[derive(Debug, Clone, PartialEq, Eq, Default)]
 pub struct McpConfigCollection {
     servers: BTreeMap<String, ScopedMcpServerConfig>,
@@ -221,6 +228,7 @@ impl ConfigLoader {
         let merged_value = JsonValue::Object(merged.clone());
 
         let feature_config = RuntimeFeatureConfig {
+            hooks: parse_optional_hooks_config(&merged_value)?,
             mcp: McpConfigCollection {
                 servers: mcp_servers,
             },
@@ -278,6 +286,11 @@ impl RuntimeConfig {
         &self.feature_config.mcp
     }
 
+    #[must_use]
+    pub fn hooks(&self) -> &RuntimeHookConfig {
+        &self.feature_config.hooks
+    }
+
     #[must_use]
     pub fn oauth(&self) -> Option<&OAuthConfig> {
         self.feature_config.oauth.as_ref()
@@ -300,6 +313,17 @@ impl RuntimeConfig {
 }
 
 impl RuntimeFeatureConfig {
+    #[must_use]
+    pub fn with_hooks(mut self, hooks: RuntimeHookConfig) -> Self {
+        self.hooks = hooks;
+        self
+    }
+
+    #[must_use]
+    pub fn hooks(&self) -> &RuntimeHookConfig {
+        &self.hooks
+    }
+
     #[must_use]
     pub fn mcp(&self) -> &McpConfigCollection {
         &self.mcp
@@ -326,6 +350,26 @@ impl RuntimeFeatureConfig {
     }
 }
 
+impl RuntimeHookConfig {
+    #[must_use]
+    pub fn new(pre_tool_use: Vec<String>, post_tool_use: Vec<String>) -> Self {
+        Self {
+            pre_tool_use,
+            post_tool_use,
+        }
+    }
+
+    #[must_use]
+    pub fn pre_tool_use(&self) -> &[String] {
+        &self.pre_tool_use
+    }
+
+    #[must_use]
+    pub fn post_tool_use(&self) -> &[String] {
+        &self.post_tool_use
+    }
+}
+
 impl McpConfigCollection {
     #[must_use]
     pub fn servers(&self) -> &BTreeMap<String, ScopedMcpServerConfig> {
@@ -424,6 +468,22 @@ fn parse_optional_model(root: &JsonValue) -> Option<String> {
         .map(ToOwned::to_owned)
 }
 
+fn parse_optional_hooks_config(root: &JsonValue) -> Result<RuntimeHookConfig, ConfigError> {
+    let Some(object) = root.as_object() else {
+        return Ok(RuntimeHookConfig::default());
+    };
+    let Some(hooks_value) = object.get("hooks") else {
+        return Ok(RuntimeHookConfig::default());
+    };
+    let hooks = expect_object(hooks_value, "merged settings.hooks")?;
+    Ok(RuntimeHookConfig {
+        pre_tool_use: optional_string_array(hooks, "PreToolUse", "merged settings.hooks")?
+            .unwrap_or_default(),
+        post_tool_use: optional_string_array(hooks, "PostToolUse", "merged settings.hooks")?
+            .unwrap_or_default(),
+    })
+}
+
 fn parse_optional_permission_mode(
     root: &JsonValue,
 ) -> Result<Option<ResolvedPermissionMode>, ConfigError> {
@@ -836,6 +896,8 @@ mod tests {
             .and_then(JsonValue::as_object)
             .expect("hooks object")
             .contains_key("PostToolUse"));
+        assert_eq!(loaded.hooks().pre_tool_use(), &["base".to_string()]);
+        assert_eq!(loaded.hooks().post_tool_use(), &["project".to_string()]);
         assert!(loaded.mcp().get("home").is_some());
         assert!(loaded.mcp().get("project").is_some());
 

rust/crates/runtime/src/conversation.rs

@@ -4,13 +4,12 @@ use std::fmt::{Display, Formatter};
 use crate::compact::{
     compact_session, estimate_session_tokens, CompactionConfig, CompactionResult,
 };
+use crate::config::RuntimeFeatureConfig;
+use crate::hooks::{HookRunResult, HookRunner};
 use crate::permissions::{PermissionOutcome, PermissionPolicy, PermissionPrompter};
 use crate::session::{ContentBlock, ConversationMessage, Session};
 use crate::usage::{TokenUsage, UsageTracker};
 
-const DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD: u32 = 100_000;
-const AUTO_COMPACTION_THRESHOLD_ENV_VAR: &str = "CLAUDE_CODE_AUTO_COMPACT_INPUT_TOKENS";
-
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct ApiRequest {
     pub system_prompt: Vec<String>,
@@ -26,9 +25,19 @@ pub enum AssistantEvent {
         input: String,
     },
     Usage(TokenUsage),
+    PromptCache(PromptCacheEvent),
     MessageStop,
 }
 
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct PromptCacheEvent {
+    pub unexpected: bool,
+    pub reason: String,
+    pub previous_cache_read_input_tokens: u32,
+    pub current_cache_read_input_tokens: u32,
+    pub token_drop: u32,
+}
+
 pub trait ApiClient {
     fn stream(&mut self, request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError>;
 }
@@ -85,14 +94,9 @@ impl std::error::Error for RuntimeError {}
 pub struct TurnSummary {
     pub assistant_messages: Vec<ConversationMessage>,
     pub tool_results: Vec<ConversationMessage>,
+    pub prompt_cache_events: Vec<PromptCacheEvent>,
     pub iterations: usize,
     pub usage: TokenUsage,
-    pub auto_compaction: Option<AutoCompactionEvent>,
-}
-
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
-pub struct AutoCompactionEvent {
-    pub removed_message_count: usize,
 }
 
 pub struct ConversationRuntime<C, T> {
@@ -103,7 +107,7 @@ pub struct ConversationRuntime<C, T> {
     system_prompt: Vec<String>,
     max_iterations: usize,
     usage_tracker: UsageTracker,
-    auto_compaction_input_tokens_threshold: u32,
+    hook_runner: HookRunner,
 }
 
 impl<C, T> ConversationRuntime<C, T>
@@ -118,6 +122,25 @@ where
         tool_executor: T,
         permission_policy: PermissionPolicy,
         system_prompt: Vec<String>,
+    ) -> Self {
+        Self::new_with_features(
+            session,
+            api_client,
+            tool_executor,
+            permission_policy,
+            system_prompt,
+            &RuntimeFeatureConfig::default(),
+        )
+    }
+
+    #[must_use]
+    pub fn new_with_features(
+        session: Session,
+        api_client: C,
+        tool_executor: T,
+        permission_policy: PermissionPolicy,
+        system_prompt: Vec<String>,
+        feature_config: &RuntimeFeatureConfig,
     ) -> Self {
         let usage_tracker = UsageTracker::from_session(&session);
         Self {
@@ -128,7 +151,7 @@ where
             system_prompt,
             max_iterations: usize::MAX,
             usage_tracker,
-            auto_compaction_input_tokens_threshold: auto_compaction_threshold_from_env(),
+            hook_runner: HookRunner::from_feature_config(feature_config),
         }
     }
 
@@ -138,12 +161,6 @@ where
         self
     }
 
-    #[must_use]
-    pub fn with_auto_compaction_input_tokens_threshold(mut self, threshold: u32) -> Self {
-        self.auto_compaction_input_tokens_threshold = threshold;
-        self
-    }
-
     pub fn run_turn(
         &mut self,
         user_input: impl Into<String>,
@@ -155,6 +172,7 @@ where
 
         let mut assistant_messages = Vec::new();
         let mut tool_results = Vec::new();
+        let mut prompt_cache_events = Vec::new();
         let mut iterations = 0;
 
         loop {
@@ -170,10 +188,12 @@ where
                 messages: self.session.messages.clone(),
             };
             let events = self.api_client.stream(request)?;
-            let (assistant_message, usage) = build_assistant_message(events)?;
+            let (assistant_message, usage, turn_prompt_cache_events) =
+                build_assistant_message(events)?;
             if let Some(usage) = usage {
                 self.usage_tracker.record(usage);
             }
+            prompt_cache_events.extend(turn_prompt_cache_events);
             let pending_tool_uses = assistant_message
                 .blocks
                 .iter()
@@ -202,19 +222,41 @@ where
 
                 let result_message = match permission_outcome {
                     PermissionOutcome::Allow => {
+                        let pre_hook_result = self.hook_runner.run_pre_tool_use(&tool_name, &input);
+                        if pre_hook_result.is_denied() {
+                            let deny_message = format!("PreToolUse hook denied tool `{tool_name}`");
+                            ConversationMessage::tool_result(
+                                tool_use_id,
+                                tool_name,
+                                format_hook_message(&pre_hook_result, &deny_message),
+                                true,
+                            )
+                        } else {
+                            let (mut output, mut is_error) =
                                 match self.tool_executor.execute(&tool_name, &input) {
-                            Ok(output) => ConversationMessage::tool_result(
+                                    Ok(output) => (output, false),
+                                    Err(error) => (error.to_string(), true),
+                                };
+                            output = merge_hook_feedback(pre_hook_result.messages(), output, false);
+
+                            let post_hook_result = self
+                                .hook_runner
+                                .run_post_tool_use(&tool_name, &input, &output, is_error);
+                            if post_hook_result.is_denied() {
+                                is_error = true;
+                            }
+                            output = merge_hook_feedback(
+                                post_hook_result.messages(),
+                                output,
+                                post_hook_result.is_denied(),
+                            );
+
+                            ConversationMessage::tool_result(
                                 tool_use_id,
                                 tool_name,
                                 output,
-                                false,
-                            ),
-                            Err(error) => ConversationMessage::tool_result(
-                                tool_use_id,
-                                tool_name,
-                                error.to_string(),
-                                true,
-                            ),
+                                is_error,
+                            )
                         }
                     }
                     PermissionOutcome::Deny { reason } => {
@@ -226,14 +268,12 @@ where
             }
         }
 
-        let auto_compaction = self.maybe_auto_compact();
-
         Ok(TurnSummary {
             assistant_messages,
             tool_results,
+            prompt_cache_events,
             iterations,
             usage: self.usage_tracker.cumulative_usage(),
-            auto_compaction,
         })
     }
 
@@ -261,55 +301,21 @@ where
     pub fn into_session(self) -> Session {
         self.session
     }
-
-    fn maybe_auto_compact(&mut self) -> Option<AutoCompactionEvent> {
-        if self.usage_tracker.cumulative_usage().input_tokens
-            < self.auto_compaction_input_tokens_threshold
-        {
-            return None;
-        }
-
-        let result = compact_session(
-            &self.session,
-            CompactionConfig {
-                max_estimated_tokens: 0,
-                ..CompactionConfig::default()
-            },
-        );
-
-        if result.removed_message_count == 0 {
-            return None;
-        }
-
-        self.session = result.compacted_session;
-        Some(AutoCompactionEvent {
-            removed_message_count: result.removed_message_count,
-        })
-    }
-}
-
-#[must_use]
-pub fn auto_compaction_threshold_from_env() -> u32 {
-    parse_auto_compaction_threshold(
-        std::env::var(AUTO_COMPACTION_THRESHOLD_ENV_VAR)
-            .ok()
-            .as_deref(),
-    )
-}
-
-#[must_use]
-fn parse_auto_compaction_threshold(value: Option<&str>) -> u32 {
-    value
-        .and_then(|raw| raw.trim().parse::<u32>().ok())
-        .filter(|threshold| *threshold > 0)
-        .unwrap_or(DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD)
 }
 
 fn build_assistant_message(
     events: Vec<AssistantEvent>,
-) -> Result<(ConversationMessage, Option<TokenUsage>), RuntimeError> {
+) -> Result<
+    (
+        ConversationMessage,
+        Option<TokenUsage>,
+        Vec<PromptCacheEvent>,
+    ),
+    RuntimeError,
+> {
     let mut text = String::new();
     let mut blocks = Vec::new();
+    let mut prompt_cache_events = Vec::new();
     let mut finished = false;
     let mut usage = None;
 
@@ -321,6 +327,7 @@ fn build_assistant_message(
                 blocks.push(ContentBlock::ToolUse { id, name, input });
             }
             AssistantEvent::Usage(value) => usage = Some(value),
+            AssistantEvent::PromptCache(event) => prompt_cache_events.push(event),
             AssistantEvent::MessageStop => {
                 finished = true;
             }
@@ -341,6 +348,7 @@ fn build_assistant_message(
     Ok((
         ConversationMessage::assistant_with_usage(blocks, usage),
         usage,
+        prompt_cache_events,
     ))
 }
 
@@ -352,6 +360,32 @@ fn flush_text_block(text: &mut String, blocks: &mut Vec<ContentBlock>) {
     }
 }
 
+fn format_hook_message(result: &HookRunResult, fallback: &str) -> String {
+    if result.messages().is_empty() {
+        fallback.to_string()
+    } else {
+        result.messages().join("\n")
+    }
+}
+
+fn merge_hook_feedback(messages: &[String], output: String, denied: bool) -> String {
+    if messages.is_empty() {
+        return output;
+    }
+
+    let mut sections = Vec::new();
+    if !output.trim().is_empty() {
+        sections.push(output);
+    }
+    let label = if denied {
+        "Hook feedback (denied)"
+    } else {
+        "Hook feedback"
+    };
+    sections.push(format!("{label}:\n{}", messages.join("\n")));
+    sections.join("\n\n")
+}
+
 type ToolHandler = Box<dyn FnMut(&str) -> Result<String, ToolError>>;
 
 #[derive(Default)]
@@ -387,11 +421,11 @@ impl ToolExecutor for StaticToolExecutor {
 #[cfg(test)]
 mod tests {
     use super::{
-        parse_auto_compaction_threshold, ApiClient, ApiRequest, AssistantEvent,
-        AutoCompactionEvent, ConversationRuntime, RuntimeError, StaticToolExecutor,
-        DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD,
+        ApiClient, ApiRequest, AssistantEvent, ConversationRuntime, PromptCacheEvent, RuntimeError,
+        StaticToolExecutor,
     };
     use crate::compact::CompactionConfig;
+    use crate::config::{RuntimeFeatureConfig, RuntimeHookConfig};
     use crate::permissions::{
         PermissionMode, PermissionPolicy, PermissionPromptDecision, PermissionPrompter,
         PermissionRequest,
@@ -444,6 +478,15 @@ mod tests {
                             cache_creation_input_tokens: 1,
                             cache_read_input_tokens: 3,
                         }),
+                        AssistantEvent::PromptCache(PromptCacheEvent {
+                            unexpected: true,
+                            reason:
+                                "cache read tokens dropped while prompt fingerprint remained stable"
+                                    .to_string(),
+                            previous_cache_read_input_tokens: 6_000,
+                            current_cache_read_input_tokens: 1_000,
+                            token_drop: 5_000,
+                        }),
                         AssistantEvent::MessageStop,
                     ])
                 }
@@ -497,9 +540,10 @@ mod tests {
         assert_eq!(summary.iterations, 2);
         assert_eq!(summary.assistant_messages.len(), 2);
         assert_eq!(summary.tool_results.len(), 1);
+        assert_eq!(summary.prompt_cache_events.len(), 1);
         assert_eq!(runtime.session().messages.len(), 4);
         assert_eq!(summary.usage.output_tokens, 10);
-        assert_eq!(summary.auto_compaction, None);
+        assert!(summary.prompt_cache_events[0].unexpected);
         assert!(matches!(
             runtime.session().messages[1].blocks[1],
             ContentBlock::ToolUse { .. }
@@ -567,6 +611,141 @@ mod tests {
         ));
     }
 
+    #[test]
+    fn denies_tool_use_when_pre_tool_hook_blocks() {
+        struct SingleCallApiClient;
+        impl ApiClient for SingleCallApiClient {
+            fn stream(&mut self, request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError> {
+                if request
+                    .messages
+                    .iter()
+                    .any(|message| message.role == MessageRole::Tool)
+                {
+                    return Ok(vec![
+                        AssistantEvent::TextDelta("blocked".to_string()),
+                        AssistantEvent::MessageStop,
+                    ]);
+                }
+                Ok(vec![
+                    AssistantEvent::ToolUse {
+                        id: "tool-1".to_string(),
+                        name: "blocked".to_string(),
+                        input: r#"{"path":"secret.txt"}"#.to_string(),
+                    },
+                    AssistantEvent::MessageStop,
+                ])
+            }
+        }
+
+        let mut runtime = ConversationRuntime::new_with_features(
+            Session::new(),
+            SingleCallApiClient,
+            StaticToolExecutor::new().register("blocked", |_input| {
+                panic!("tool should not execute when hook denies")
+            }),
+            PermissionPolicy::new(PermissionMode::DangerFullAccess),
+            vec!["system".to_string()],
+            &RuntimeFeatureConfig::default().with_hooks(RuntimeHookConfig::new(
+                vec![shell_snippet("printf 'blocked by hook'; exit 2")],
+                Vec::new(),
+            )),
+        );
+
+        let summary = runtime
+            .run_turn("use the tool", None)
+            .expect("conversation should continue after hook denial");
+
+        assert_eq!(summary.tool_results.len(), 1);
+        let ContentBlock::ToolResult {
+            is_error, output, ..
+        } = &summary.tool_results[0].blocks[0]
+        else {
+            panic!("expected tool result block");
+        };
+        assert!(
+            *is_error,
+            "hook denial should produce an error result: {output}"
+        );
+        assert!(
+            output.contains("denied tool") || output.contains("blocked by hook"),
+            "unexpected hook denial output: {output:?}"
+        );
+    }
+
+    #[test]
+    fn appends_post_tool_hook_feedback_to_tool_result() {
+        struct TwoCallApiClient {
+            calls: usize,
+        }
+
+        impl ApiClient for TwoCallApiClient {
+            fn stream(&mut self, request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError> {
+                self.calls += 1;
+                match self.calls {
+                    1 => Ok(vec![
+                        AssistantEvent::ToolUse {
+                            id: "tool-1".to_string(),
+                            name: "add".to_string(),
+                            input: r#"{"lhs":2,"rhs":2}"#.to_string(),
+                        },
+                        AssistantEvent::MessageStop,
+                    ]),
+                    2 => {
+                        assert!(request
+                            .messages
+                            .iter()
+                            .any(|message| message.role == MessageRole::Tool));
+                        Ok(vec![
+                            AssistantEvent::TextDelta("done".to_string()),
+                            AssistantEvent::MessageStop,
+                        ])
+                    }
+                    _ => Err(RuntimeError::new("unexpected extra API call")),
+                }
+            }
+        }
+
+        let mut runtime = ConversationRuntime::new_with_features(
+            Session::new(),
+            TwoCallApiClient { calls: 0 },
+            StaticToolExecutor::new().register("add", |_input| Ok("4".to_string())),
+            PermissionPolicy::new(PermissionMode::DangerFullAccess),
+            vec!["system".to_string()],
+            &RuntimeFeatureConfig::default().with_hooks(RuntimeHookConfig::new(
+                vec![shell_snippet("printf 'pre hook ran'")],
+                vec![shell_snippet("printf 'post hook ran'")],
+            )),
+        );
+
+        let summary = runtime
+            .run_turn("use add", None)
+            .expect("tool loop succeeds");
+
+        assert_eq!(summary.tool_results.len(), 1);
+        let ContentBlock::ToolResult {
+            is_error, output, ..
+        } = &summary.tool_results[0].blocks[0]
+        else {
+            panic!("expected tool result block");
+        };
+        assert!(
+            !*is_error,
+            "post hook should preserve non-error result: {output:?}"
+        );
+        assert!(
+            output.contains('4'),
+            "tool output missing value: {output:?}"
+        );
+        assert!(
+            output.contains("pre hook ran"),
+            "tool output missing pre hook feedback: {output:?}"
+        );
+        assert!(
+            output.contains("post hook ran"),
+            "tool output missing post hook feedback: {output:?}"
+        );
+    }
+
     #[test]
     fn reconstructs_usage_tracker_from_restored_session() {
         struct SimpleApi;
@@ -646,110 +825,13 @@ mod tests {
         );
     }
 
-    #[test]
-    fn auto_compacts_when_cumulative_input_threshold_is_crossed() {
-        struct SimpleApi;
-        impl ApiClient for SimpleApi {
-            fn stream(
-                &mut self,
-                _request: ApiRequest,
-            ) -> Result<Vec<AssistantEvent>, RuntimeError> {
-                Ok(vec![
-                    AssistantEvent::TextDelta("done".to_string()),
-                    AssistantEvent::Usage(TokenUsage {
-                        input_tokens: 120_000,
-                        output_tokens: 4,
-                        cache_creation_input_tokens: 0,
-                        cache_read_input_tokens: 0,
-                    }),
-                    AssistantEvent::MessageStop,
-                ])
-            }
+    #[cfg(windows)]
+    fn shell_snippet(script: &str) -> String {
+        script.replace('\'', "\"")
     }
 
-        let session = Session {
-            version: 1,
-            messages: vec![
-                crate::session::ConversationMessage::user_text("one"),
-                crate::session::ConversationMessage::assistant(vec![ContentBlock::Text {
-                    text: "two".to_string(),
-                }]),
-                crate::session::ConversationMessage::user_text("three"),
-                crate::session::ConversationMessage::assistant(vec![ContentBlock::Text {
-                    text: "four".to_string(),
-                }]),
-            ],
-        };
-
-        let mut runtime = ConversationRuntime::new(
-            session,
-            SimpleApi,
-            StaticToolExecutor::new(),
-            PermissionPolicy::new(PermissionMode::DangerFullAccess),
-            vec!["system".to_string()],
-        )
-        .with_auto_compaction_input_tokens_threshold(100_000);
-
-        let summary = runtime
-            .run_turn("trigger", None)
-            .expect("turn should succeed");
-
-        assert_eq!(
-            summary.auto_compaction,
-            Some(AutoCompactionEvent {
-                removed_message_count: 2,
-            })
-        );
-        assert_eq!(runtime.session().messages[0].role, MessageRole::System);
-    }
-
-    #[test]
-    fn skips_auto_compaction_below_threshold() {
-        struct SimpleApi;
-        impl ApiClient for SimpleApi {
-            fn stream(
-                &mut self,
-                _request: ApiRequest,
-            ) -> Result<Vec<AssistantEvent>, RuntimeError> {
-                Ok(vec![
-                    AssistantEvent::TextDelta("done".to_string()),
-                    AssistantEvent::Usage(TokenUsage {
-                        input_tokens: 99_999,
-                        output_tokens: 4,
-                        cache_creation_input_tokens: 0,
-                        cache_read_input_tokens: 0,
-                    }),
-                    AssistantEvent::MessageStop,
-                ])
-            }
-        }
-
-        let mut runtime = ConversationRuntime::new(
-            Session::new(),
-            SimpleApi,
-            StaticToolExecutor::new(),
-            PermissionPolicy::new(PermissionMode::DangerFullAccess),
-            vec!["system".to_string()],
-        )
-        .with_auto_compaction_input_tokens_threshold(100_000);
-
-        let summary = runtime
-            .run_turn("trigger", None)
-            .expect("turn should succeed");
-        assert_eq!(summary.auto_compaction, None);
-        assert_eq!(runtime.session().messages.len(), 2);
-    }
-
-    #[test]
-    fn auto_compaction_threshold_defaults_and_parses_values() {
-        assert_eq!(
-            parse_auto_compaction_threshold(None),
-            DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD
-        );
-        assert_eq!(parse_auto_compaction_threshold(Some("4321")), 4321);
-        assert_eq!(
-            parse_auto_compaction_threshold(Some("not-a-number")),
-            DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD
-        );
+    #[cfg(not(windows))]
+    fn shell_snippet(script: &str) -> String {
+        script.to_string()
     }
 }

rust/crates/runtime/src/hooks.rs

@@ -0,0 +1,347 @@
+use std::ffi::OsStr;
+use std::process::Command;
+
+use serde_json::json;
+
+use crate::config::{RuntimeFeatureConfig, RuntimeHookConfig};
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum HookEvent {
+    PreToolUse,
+    PostToolUse,
+}
+
+impl HookEvent {
+    fn as_str(self) -> &'static str {
+        match self {
+            Self::PreToolUse => "PreToolUse",
+            Self::PostToolUse => "PostToolUse",
+        }
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct HookRunResult {
+    denied: bool,
+    messages: Vec<String>,
+}
+
+impl HookRunResult {
+    #[must_use]
+    pub fn allow(messages: Vec<String>) -> Self {
+        Self {
+            denied: false,
+            messages,
+        }
+    }
+
+    #[must_use]
+    pub fn is_denied(&self) -> bool {
+        self.denied
+    }
+
+    #[must_use]
+    pub fn messages(&self) -> &[String] {
+        &self.messages
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Default)]
+pub struct HookRunner {
+    config: RuntimeHookConfig,
+}
+
+impl HookRunner {
+    #[must_use]
+    pub fn new(config: RuntimeHookConfig) -> Self {
+        Self { config }
+    }
+
+    #[must_use]
+    pub fn from_feature_config(feature_config: &RuntimeFeatureConfig) -> Self {
+        Self::new(feature_config.hooks().clone())
+    }
+
+    #[must_use]
+    pub fn run_pre_tool_use(&self, tool_name: &str, tool_input: &str) -> HookRunResult {
+        Self::run_commands(
+            HookEvent::PreToolUse,
+            self.config.pre_tool_use(),
+            tool_name,
+            tool_input,
+            None,
+            false,
+        )
+    }
+
+    #[must_use]
+    pub fn run_post_tool_use(
+        &self,
+        tool_name: &str,
+        tool_input: &str,
+        tool_output: &str,
+        is_error: bool,
+    ) -> HookRunResult {
+        Self::run_commands(
+            HookEvent::PostToolUse,
+            self.config.post_tool_use(),
+            tool_name,
+            tool_input,
+            Some(tool_output),
+            is_error,
+        )
+    }
+
+    fn run_commands(
+        event: HookEvent,
+        commands: &[String],
+        tool_name: &str,
+        tool_input: &str,
+        tool_output: Option<&str>,
+        is_error: bool,
+    ) -> HookRunResult {
+        if commands.is_empty() {
+            return HookRunResult::allow(Vec::new());
+        }
+
+        let payload = json!({
+            "hook_event_name": event.as_str(),
+            "tool_name": tool_name,
+            "tool_input": parse_tool_input(tool_input),
+            "tool_input_json": tool_input,
+            "tool_output": tool_output,
+            "tool_result_is_error": is_error,
+        })
+        .to_string();
+
+        let mut messages = Vec::new();
+
+        for command in commands {
+            match Self::run_command(
+                command,
+                event,
+                tool_name,
+                tool_input,
+                tool_output,
+                is_error,
+                &payload,
+            ) {
+                HookCommandOutcome::Allow { message } => {
+                    if let Some(message) = message {
+                        messages.push(message);
+                    }
+                }
+                HookCommandOutcome::Deny { message } => {
+                    let message = message.unwrap_or_else(|| {
+                        format!("{} hook denied tool `{tool_name}`", event.as_str())
+                    });
+                    messages.push(message);
+                    return HookRunResult {
+                        denied: true,
+                        messages,
+                    };
+                }
+                HookCommandOutcome::Warn { message } => messages.push(message),
+            }
+        }
+
+        HookRunResult::allow(messages)
+    }
+
+    fn run_command(
+        command: &str,
+        event: HookEvent,
+        tool_name: &str,
+        tool_input: &str,
+        tool_output: Option<&str>,
+        is_error: bool,
+        payload: &str,
+    ) -> HookCommandOutcome {
+        let mut child = shell_command(command);
+        child.stdin(std::process::Stdio::piped());
+        child.stdout(std::process::Stdio::piped());
+        child.stderr(std::process::Stdio::piped());
+        child.env("HOOK_EVENT", event.as_str());
+        child.env("HOOK_TOOL_NAME", tool_name);
+        child.env("HOOK_TOOL_INPUT", tool_input);
+        child.env("HOOK_TOOL_IS_ERROR", if is_error { "1" } else { "0" });
+        if let Some(tool_output) = tool_output {
+            child.env("HOOK_TOOL_OUTPUT", tool_output);
+        }
+
+        match child.output_with_stdin(payload.as_bytes()) {
+            Ok(output) => {
+                let stdout = String::from_utf8_lossy(&output.stdout).trim().to_string();
+                let stderr = String::from_utf8_lossy(&output.stderr).trim().to_string();
+                let message = (!stdout.is_empty()).then_some(stdout);
+                match output.status.code() {
+                    Some(0) => HookCommandOutcome::Allow { message },
+                    Some(2) => HookCommandOutcome::Deny { message },
+                    Some(code) => HookCommandOutcome::Warn {
+                        message: format_hook_warning(
+                            command,
+                            code,
+                            message.as_deref(),
+                            stderr.as_str(),
+                        ),
+                    },
+                    None => HookCommandOutcome::Warn {
+                        message: format!(
+                            "{} hook `{command}` terminated by signal while handling `{tool_name}`",
+                            event.as_str()
+                        ),
+                    },
+                }
+            }
+            Err(error) => HookCommandOutcome::Warn {
+                message: format!(
+                    "{} hook `{command}` failed to start for `{tool_name}`: {error}",
+                    event.as_str()
+                ),
+            },
+        }
+    }
+}
+
+enum HookCommandOutcome {
+    Allow { message: Option<String> },
+    Deny { message: Option<String> },
+    Warn { message: String },
+}
+
+fn parse_tool_input(tool_input: &str) -> serde_json::Value {
+    serde_json::from_str(tool_input).unwrap_or_else(|_| json!({ "raw": tool_input }))
+}
+
+fn format_hook_warning(command: &str, code: i32, stdout: Option<&str>, stderr: &str) -> String {
+    let mut message =
+        format!("Hook `{command}` exited with status {code}; allowing tool execution to continue");
+    if let Some(stdout) = stdout.filter(|stdout| !stdout.is_empty()) {
+        message.push_str(": ");
+        message.push_str(stdout);
+    } else if !stderr.is_empty() {
+        message.push_str(": ");
+        message.push_str(stderr);
+    }
+    message
+}
+
+fn shell_command(command: &str) -> CommandWithStdin {
+    #[cfg(windows)]
+    let mut command_builder = {
+        let mut command_builder = Command::new("cmd");
+        command_builder.arg("/C").arg(command);
+        CommandWithStdin::new(command_builder)
+    };
+
+    #[cfg(not(windows))]
+    let command_builder = {
+        let mut command_builder = Command::new("sh");
+        command_builder.arg("-lc").arg(command);
+        CommandWithStdin::new(command_builder)
+    };
+
+    command_builder
+}
+
+struct CommandWithStdin {
+    command: Command,
+}
+
+impl CommandWithStdin {
+    fn new(command: Command) -> Self {
+        Self { command }
+    }
+
+    fn stdin(&mut self, cfg: std::process::Stdio) -> &mut Self {
+        self.command.stdin(cfg);
+        self
+    }
+
+    fn stdout(&mut self, cfg: std::process::Stdio) -> &mut Self {
+        self.command.stdout(cfg);
+        self
+    }
+
+    fn stderr(&mut self, cfg: std::process::Stdio) -> &mut Self {
+        self.command.stderr(cfg);
+        self
+    }
+
+    fn env<K, V>(&mut self, key: K, value: V) -> &mut Self
+    where
+        K: AsRef<OsStr>,
+        V: AsRef<OsStr>,
+    {
+        self.command.env(key, value);
+        self
+    }
+
+    fn output_with_stdin(&mut self, stdin: &[u8]) -> std::io::Result<std::process::Output> {
+        let mut child = self.command.spawn()?;
+        if let Some(mut child_stdin) = child.stdin.take() {
+            use std::io::Write;
+            child_stdin.write_all(stdin)?;
+        }
+        child.wait_with_output()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{HookRunResult, HookRunner};
+    use crate::config::{RuntimeFeatureConfig, RuntimeHookConfig};
+
+    #[test]
+    fn allows_exit_code_zero_and_captures_stdout() {
+        let runner = HookRunner::new(RuntimeHookConfig::new(
+            vec![shell_snippet("printf 'pre ok'")],
+            Vec::new(),
+        ));
+
+        let result = runner.run_pre_tool_use("Read", r#"{"path":"README.md"}"#);
+
+        assert_eq!(result, HookRunResult::allow(vec!["pre ok".to_string()]));
+    }
+
+    #[test]
+    fn denies_exit_code_two() {
+        let runner = HookRunner::new(RuntimeHookConfig::new(
+            vec![shell_snippet("printf 'blocked by hook'; exit 2")],
+            Vec::new(),
+        ));
+
+        let result = runner.run_pre_tool_use("Bash", r#"{"command":"pwd"}"#);
+
+        assert!(result.is_denied());
+        assert_eq!(result.messages(), &["blocked by hook".to_string()]);
+    }
+
+    #[test]
+    fn warns_for_other_non_zero_statuses() {
+        let runner = HookRunner::from_feature_config(&RuntimeFeatureConfig::default().with_hooks(
+            RuntimeHookConfig::new(
+                vec![shell_snippet("printf 'warning hook'; exit 1")],
+                Vec::new(),
+            ),
+        ));
+
+        let result = runner.run_pre_tool_use("Edit", r#"{"file":"src/lib.rs"}"#);
+
+        assert!(!result.is_denied());
+        assert!(result
+            .messages()
+            .iter()
+            .any(|message| message.contains("allowing tool execution to continue")));
+    }
+
+    #[cfg(windows)]
+    fn shell_snippet(script: &str) -> String {
+        script.replace('\'', "\"")
+    }
+
+    #[cfg(not(windows))]
+    fn shell_snippet(script: &str) -> String {
+        script.to_string()
+    }
+}

rust/crates/runtime/src/lib.rs

@@ -4,6 +4,7 @@ mod compact;
 mod config;
 mod conversation;
 mod file_ops;
+mod hooks;
 mod json;
 mod mcp;
 mod mcp_client;
@@ -26,18 +27,19 @@ pub use config::{
     ConfigEntry, ConfigError, ConfigLoader, ConfigSource, McpClaudeAiProxyServerConfig,
     McpConfigCollection, McpOAuthConfig, McpRemoteServerConfig, McpSdkServerConfig,
     McpServerConfig, McpStdioServerConfig, McpTransport, McpWebSocketServerConfig, OAuthConfig,
-    ResolvedPermissionMode, RuntimeConfig, RuntimeFeatureConfig, ScopedMcpServerConfig,
-    CLAUDE_CODE_SETTINGS_SCHEMA_NAME,
+    ResolvedPermissionMode, RuntimeConfig, RuntimeFeatureConfig, RuntimeHookConfig,
+    ScopedMcpServerConfig, CLAUDE_CODE_SETTINGS_SCHEMA_NAME,
 };
 pub use conversation::{
-    auto_compaction_threshold_from_env, ApiClient, ApiRequest, AssistantEvent, AutoCompactionEvent,
-    ConversationRuntime, RuntimeError, StaticToolExecutor, ToolError, ToolExecutor, TurnSummary,
+    ApiClient, ApiRequest, AssistantEvent, ConversationRuntime, PromptCacheEvent, RuntimeError,
+    StaticToolExecutor, ToolError, ToolExecutor, TurnSummary,
 };
 pub use file_ops::{
     edit_file, glob_search, grep_search, read_file, write_file, EditFileOutput, GlobSearchOutput,
     GrepSearchInput, GrepSearchOutput, ReadFileOutput, StructuredPatchHunk, TextFilePayload,
     WriteFileOutput,
 };
+pub use hooks::{HookEvent, HookRunResult, HookRunner};
 pub use mcp::{
     mcp_server_signature, mcp_tool_name, mcp_tool_prefix, normalize_name_for_mcp,
     scoped_mcp_config_hash, unwrap_ccr_proxy_url,

rust/crates/rusty-claude-cli/src/main.rs

@@ -13,8 +13,9 @@ use std::time::{SystemTime, UNIX_EPOCH};
 
 use api::{
     resolve_startup_auth_source, AnthropicClient, AuthSource, ContentBlockDelta, InputContentBlock,
-    InputMessage, MessageRequest, MessageResponse, OutputContentBlock,
-    StreamEvent as ApiStreamEvent, ToolChoice, ToolDefinition, ToolResultContentBlock,
+    InputMessage, MessageRequest, MessageResponse, OutputContentBlock, PromptCache,
+    PromptCacheRecord, StreamEvent as ApiStreamEvent, ToolChoice, ToolDefinition,
+    ToolResultContentBlock,
 };
 
 use commands::{
@@ -28,8 +29,8 @@ use runtime::{
     parse_oauth_callback_request_target, save_oauth_credentials, ApiClient, ApiRequest,
     AssistantEvent, CompactionConfig, ConfigLoader, ConfigSource, ContentBlock,
     ConversationMessage, ConversationRuntime, MessageRole, OAuthAuthorizationRequest, OAuthConfig,
-    OAuthTokenExchangeRequest, PermissionMode, PermissionPolicy, ProjectContext, RuntimeError,
-    Session, TokenUsage, ToolError, ToolExecutor, UsageTracker,
+    OAuthTokenExchangeRequest, PermissionMode, PermissionPolicy, ProjectContext, PromptCacheEvent,
+    RuntimeError, Session, TokenUsage, ToolError, ToolExecutor, UsageTracker,
 };
 use serde_json::json;
 use tools::{execute_tool, mvp_tool_specs, ToolSpec};
@@ -775,10 +776,6 @@ fn format_compact_report(removed: usize, resulting_messages: usize, skipped: boo
     }
 }
 
-fn format_auto_compaction_notice(removed: usize) -> String {
-    format!("[auto-compacted: removed {removed} messages]")
-}
-
 fn parse_git_status_metadata(status: Option<&str>) -> (Option<PathBuf>, Option<String>) {
     let Some(status) = status else {
         return (None, None);
@@ -917,14 +914,7 @@ fn run_resume_command(
                 )),
             })
         }
-        SlashCommand::Bughunter { .. }
-        | SlashCommand::Commit
-        | SlashCommand::Pr { .. }
-        | SlashCommand::Issue { .. }
-        | SlashCommand::Ultraplan { .. }
-        | SlashCommand::Teleport { .. }
-        | SlashCommand::DebugToolCall
-        | SlashCommand::Resume { .. }
+        SlashCommand::Resume { .. }
         | SlashCommand::Model { .. }
         | SlashCommand::Permissions { .. }
         | SlashCommand::Session { .. }
@@ -1006,6 +996,7 @@ impl LiveCli {
         let session = create_managed_session_handle()?;
         let runtime = build_runtime(
             Session::new(),
+            session.id.clone(),
             model.clone(),
             system_prompt.clone(),
             enable_tools,
@@ -1068,12 +1059,7 @@ impl LiveCli {
                     &mut stdout,
                 )?;
                 println!();
-                if let Some(event) = summary.auto_compaction {
-                    println!(
-                        "{}",
-                        format_auto_compaction_notice(event.removed_message_count)
-                    );
-                }
+                print_prompt_cache_events(&summary);
                 self.persist_session()?;
                 Ok(())
             }
@@ -1103,6 +1089,7 @@ impl LiveCli {
         let session = self.runtime.session().clone();
         let mut runtime = build_runtime(
             session,
+            self.session.id.clone(),
             self.model.clone(),
             self.system_prompt.clone(),
             true,
@@ -1120,12 +1107,9 @@ impl LiveCli {
                 "message": final_assistant_text(&summary),
                 "model": self.model,
                 "iterations": summary.iterations,
-                "auto_compaction": summary.auto_compaction.map(|event| json!({
-                    "removed_messages": event.removed_message_count,
-                    "notice": format_auto_compaction_notice(event.removed_message_count),
-                })),
                 "tool_uses": collect_tool_uses(&summary),
                 "tool_results": collect_tool_results(&summary),
+                "prompt_cache_events": collect_prompt_cache_events(&summary),
                 "usage": {
                     "input_tokens": summary.usage.input_tokens,
                     "output_tokens": summary.usage.output_tokens,
@@ -1150,34 +1134,6 @@ impl LiveCli {
                 self.print_status();
                 false
             }
-            SlashCommand::Bughunter { scope } => {
-                self.run_bughunter(scope.as_deref())?;
-                false
-            }
-            SlashCommand::Commit => {
-                self.run_commit()?;
-                true
-            }
-            SlashCommand::Pr { context } => {
-                self.run_pr(context.as_deref())?;
-                false
-            }
-            SlashCommand::Issue { context } => {
-                self.run_issue(context.as_deref())?;
-                false
-            }
-            SlashCommand::Ultraplan { task } => {
-                self.run_ultraplan(task.as_deref())?;
-                false
-            }
-            SlashCommand::Teleport { target } => {
-                self.run_teleport(target.as_deref())?;
-                false
-            }
-            SlashCommand::DebugToolCall => {
-                self.run_debug_tool_call()?;
-                false
-            }
             SlashCommand::Compact => {
                 self.compact()?;
                 false
@@ -1281,6 +1237,7 @@ impl LiveCli {
         let message_count = session.messages.len();
         self.runtime = build_runtime(
             session,
+            self.session.id.clone(),
             model.clone(),
             self.system_prompt.clone(),
             true,
@@ -1324,6 +1281,7 @@ impl LiveCli {
         self.permission_mode = permission_mode_from_label(normalized);
         self.runtime = build_runtime(
             session,
+            self.session.id.clone(),
             self.model.clone(),
             self.system_prompt.clone(),
             true,
@@ -1349,6 +1307,7 @@ impl LiveCli {
         self.session = create_managed_session_handle()?;
         self.runtime = build_runtime(
             Session::new(),
+            self.session.id.clone(),
             self.model.clone(),
             self.system_prompt.clone(),
             true,
@@ -1384,6 +1343,7 @@ impl LiveCli {
         let message_count = session.messages.len();
         self.runtime = build_runtime(
             session,
+            handle.id.clone(),
             self.model.clone(),
             self.system_prompt.clone(),
             true,
@@ -1456,6 +1416,7 @@ impl LiveCli {
                 let message_count = session.messages.len();
                 self.runtime = build_runtime(
                     session,
+                    handle.id.clone(),
                     self.model.clone(),
                     self.system_prompt.clone(),
                     true,
@@ -1486,6 +1447,7 @@ impl LiveCli {
         let skipped = removed == 0;
         self.runtime = build_runtime(
             result.compacted_session,
+            self.session.id.clone(),
             self.model.clone(),
             self.system_prompt.clone(),
             true,
@@ -1497,160 +1459,6 @@ impl LiveCli {
         println!("{}", format_compact_report(removed, kept, skipped));
         Ok(())
     }
-
-    fn run_internal_prompt_text(
-        &self,
-        prompt: &str,
-        enable_tools: bool,
-    ) -> Result<String, Box<dyn std::error::Error>> {
-        let session = self.runtime.session().clone();
-        let mut runtime = build_runtime(
-            session,
-            self.model.clone(),
-            self.system_prompt.clone(),
-            enable_tools,
-            false,
-            self.allowed_tools.clone(),
-            self.permission_mode,
-        )?;
-        let mut permission_prompter = CliPermissionPrompter::new(self.permission_mode);
-        let summary = runtime.run_turn(prompt, Some(&mut permission_prompter))?;
-        Ok(final_assistant_text(&summary).trim().to_string())
-    }
-
-    fn run_bughunter(&self, scope: Option<&str>) -> Result<(), Box<dyn std::error::Error>> {
-        let scope = scope.unwrap_or("the current repository");
-        let prompt = format!(
-            "You are /bughunter. Inspect {scope} and identify the most likely bugs or correctness issues. Prioritize concrete findings with file paths, severity, and suggested fixes. Use tools if needed."
-        );
-        println!("{}", self.run_internal_prompt_text(&prompt, true)?);
-        Ok(())
-    }
-
-    fn run_ultraplan(&self, task: Option<&str>) -> Result<(), Box<dyn std::error::Error>> {
-        let task = task.unwrap_or("the current repo work");
-        let prompt = format!(
-            "You are /ultraplan. Produce a deep multi-step execution plan for {task}. Include goals, risks, implementation sequence, verification steps, and rollback considerations. Use tools if needed."
-        );
-        println!("{}", self.run_internal_prompt_text(&prompt, true)?);
-        Ok(())
-    }
-
-    fn run_teleport(&self, target: Option<&str>) -> Result<(), Box<dyn std::error::Error>> {
-        let Some(target) = target.map(str::trim).filter(|value| !value.is_empty()) else {
-            println!("Usage: /teleport <symbol-or-path>");
-            return Ok(());
-        };
-
-        println!("{}", render_teleport_report(target)?);
-        Ok(())
-    }
-
-    fn run_debug_tool_call(&self) -> Result<(), Box<dyn std::error::Error>> {
-        println!("{}", render_last_tool_debug_report(self.runtime.session())?);
-        Ok(())
-    }
-
-    fn run_commit(&mut self) -> Result<(), Box<dyn std::error::Error>> {
-        let status = git_output(&["status", "--short"])?;
-        if status.trim().is_empty() {
-            println!("Commit\n  Result           skipped\n  Reason           no workspace changes");
-            return Ok(());
-        }
-
-        git_status_ok(&["add", "-A"])?;
-        let staged_stat = git_output(&["diff", "--cached", "--stat"])?;
-        let prompt = format!(
-            "Generate a git commit message in plain text Lore format only. Base it on this staged diff summary:\n\n{}\n\nRecent conversation context:\n{}",
-            truncate_for_prompt(&staged_stat, 8_000),
-            recent_user_context(self.runtime.session(), 6)
-        );
-        let message = sanitize_generated_message(&self.run_internal_prompt_text(&prompt, false)?);
-        if message.trim().is_empty() {
-            return Err("generated commit message was empty".into());
-        }
-
-        let path = write_temp_text_file("claw-commit-message.txt", &message)?;
-        let output = Command::new("git")
-            .args(["commit", "--file"])
-            .arg(&path)
-            .current_dir(env::current_dir()?)
-            .output()?;
-        if !output.status.success() {
-            let stderr = String::from_utf8_lossy(&output.stderr).trim().to_string();
-            return Err(format!("git commit failed: {stderr}").into());
-        }
-
-        println!(
-            "Commit\n  Result           created\n  Message file     {}\n\n{}",
-            path.display(),
-            message.trim()
-        );
-        Ok(())
-    }
-
-    fn run_pr(&self, context: Option<&str>) -> Result<(), Box<dyn std::error::Error>> {
-        let staged = git_output(&["diff", "--stat"])?;
-        let prompt = format!(
-            "Generate a pull request title and body from this conversation and diff summary. Output plain text in this format exactly:\nTITLE: <title>\nBODY:\n<body markdown>\n\nContext hint: {}\n\nDiff summary:\n{}",
-            context.unwrap_or("none"),
-            truncate_for_prompt(&staged, 10_000)
-        );
-        let draft = sanitize_generated_message(&self.run_internal_prompt_text(&prompt, false)?);
-        let (title, body) = parse_titled_body(&draft)
-            .ok_or_else(|| "failed to parse generated PR title/body".to_string())?;
-
-        if command_exists("gh") {
-            let body_path = write_temp_text_file("claw-pr-body.md", &body)?;
-            let output = Command::new("gh")
-                .args(["pr", "create", "--title", &title, "--body-file"])
-                .arg(&body_path)
-                .current_dir(env::current_dir()?)
-                .output()?;
-            if output.status.success() {
-                let stdout = String::from_utf8_lossy(&output.stdout).trim().to_string();
-                println!(
-                    "PR\n  Result           created\n  Title            {title}\n  URL              {}",
-                    if stdout.is_empty() { "<unknown>" } else { &stdout }
-                );
-                return Ok(());
-            }
-        }
-
-        println!("PR draft\n  Title            {title}\n\n{body}");
-        Ok(())
-    }
-
-    fn run_issue(&self, context: Option<&str>) -> Result<(), Box<dyn std::error::Error>> {
-        let prompt = format!(
-            "Generate a GitHub issue title and body from this conversation. Output plain text in this format exactly:\nTITLE: <title>\nBODY:\n<body markdown>\n\nContext hint: {}\n\nConversation context:\n{}",
-            context.unwrap_or("none"),
-            truncate_for_prompt(&recent_user_context(self.runtime.session(), 10), 10_000)
-        );
-        let draft = sanitize_generated_message(&self.run_internal_prompt_text(&prompt, false)?);
-        let (title, body) = parse_titled_body(&draft)
-            .ok_or_else(|| "failed to parse generated issue title/body".to_string())?;
-
-        if command_exists("gh") {
-            let body_path = write_temp_text_file("claw-issue-body.md", &body)?;
-            let output = Command::new("gh")
-                .args(["issue", "create", "--title", &title, "--body-file"])
-                .arg(&body_path)
-                .current_dir(env::current_dir()?)
-                .output()?;
-            if output.status.success() {
-                let stdout = String::from_utf8_lossy(&output.stdout).trim().to_string();
-                println!(
-                    "Issue\n  Result           created\n  Title            {title}\n  URL              {}",
-                    if stdout.is_empty() { "<unknown>" } else { &stdout }
-                );
-                return Ok(());
-            }
-        }
-
-        println!("Issue draft\n  Title            {title}\n\n{body}");
-        Ok(())
-    }
 }
 
 fn sessions_dir() -> Result<PathBuf, Box<dyn std::error::Error>> {
@@ -2002,206 +1810,6 @@ fn render_diff_report() -> Result<String, Box<dyn std::error::Error>> {
     Ok(format!("Diff\n\n{}", diff.trim_end()))
 }
 
-fn render_teleport_report(target: &str) -> Result<String, Box<dyn std::error::Error>> {
-    let cwd = env::current_dir()?;
-
-    let file_list = Command::new("rg")
-        .args(["--files"])
-        .current_dir(&cwd)
-        .output()?;
-    let file_matches = if file_list.status.success() {
-        String::from_utf8(file_list.stdout)?
-            .lines()
-            .filter(|line| line.contains(target))
-            .take(10)
-            .map(ToOwned::to_owned)
-            .collect::<Vec<_>>()
-    } else {
-        Vec::new()
-    };
-
-    let content_output = Command::new("rg")
-        .args(["-n", "-S", "--color", "never", target, "."])
-        .current_dir(&cwd)
-        .output()?;
-
-    let mut lines = vec![format!("Teleport\n  Target           {target}")];
-    if !file_matches.is_empty() {
-        lines.push(String::new());
-        lines.push("File matches".to_string());
-        lines.extend(file_matches.into_iter().map(|path| format!("  {path}")));
-    }
-
-    if content_output.status.success() {
-        let matches = String::from_utf8(content_output.stdout)?;
-        if !matches.trim().is_empty() {
-            lines.push(String::new());
-            lines.push("Content matches".to_string());
-            lines.push(truncate_for_prompt(&matches, 4_000));
-        }
-    }
-
-    if lines.len() == 1 {
-        lines.push("  Result           no matches found".to_string());
-    }
-
-    Ok(lines.join("\n"))
-}
-
-fn render_last_tool_debug_report(session: &Session) -> Result<String, Box<dyn std::error::Error>> {
-    let last_tool_use = session
-        .messages
-        .iter()
-        .rev()
-        .find_map(|message| {
-            message.blocks.iter().rev().find_map(|block| match block {
-                ContentBlock::ToolUse { id, name, input } => {
-                    Some((id.clone(), name.clone(), input.clone()))
-                }
-                _ => None,
-            })
-        })
-        .ok_or_else(|| "no prior tool call found in session".to_string())?;
-
-    let tool_result = session.messages.iter().rev().find_map(|message| {
-        message.blocks.iter().rev().find_map(|block| match block {
-            ContentBlock::ToolResult {
-                tool_use_id,
-                tool_name,
-                output,
-                is_error,
-            } if tool_use_id == &last_tool_use.0 => {
-                Some((tool_name.clone(), output.clone(), *is_error))
-            }
-            _ => None,
-        })
-    });
-
-    let mut lines = vec![
-        "Debug tool call".to_string(),
-        format!("  Tool id          {}", last_tool_use.0),
-        format!("  Tool name        {}", last_tool_use.1),
-        "  Input".to_string(),
-        indent_block(&last_tool_use.2, 4),
-    ];
-
-    match tool_result {
-        Some((tool_name, output, is_error)) => {
-            lines.push("  Result".to_string());
-            lines.push(format!("    name           {tool_name}"));
-            lines.push(format!(
-                "    status         {}",
-                if is_error { "error" } else { "ok" }
-            ));
-            lines.push(indent_block(&output, 4));
-        }
-        None => lines.push("  Result           missing tool result".to_string()),
-    }
-
-    Ok(lines.join("\n"))
-}
-
-fn indent_block(value: &str, spaces: usize) -> String {
-    let indent = " ".repeat(spaces);
-    value
-        .lines()
-        .map(|line| format!("{indent}{line}"))
-        .collect::<Vec<_>>()
-        .join("\n")
-}
-
-fn git_output(args: &[&str]) -> Result<String, Box<dyn std::error::Error>> {
-    let output = Command::new("git")
-        .args(args)
-        .current_dir(env::current_dir()?)
-        .output()?;
-    if !output.status.success() {
-        let stderr = String::from_utf8_lossy(&output.stderr).trim().to_string();
-        return Err(format!("git {} failed: {stderr}", args.join(" ")).into());
-    }
-    Ok(String::from_utf8(output.stdout)?)
-}
-
-fn git_status_ok(args: &[&str]) -> Result<(), Box<dyn std::error::Error>> {
-    let output = Command::new("git")
-        .args(args)
-        .current_dir(env::current_dir()?)
-        .output()?;
-    if !output.status.success() {
-        let stderr = String::from_utf8_lossy(&output.stderr).trim().to_string();
-        return Err(format!("git {} failed: {stderr}", args.join(" ")).into());
-    }
-    Ok(())
-}
-
-fn command_exists(name: &str) -> bool {
-    Command::new("which")
-        .arg(name)
-        .output()
-        .map(|output| output.status.success())
-        .unwrap_or(false)
-}
-
-fn write_temp_text_file(
-    filename: &str,
-    contents: &str,
-) -> Result<PathBuf, Box<dyn std::error::Error>> {
-    let path = env::temp_dir().join(filename);
-    fs::write(&path, contents)?;
-    Ok(path)
-}
-
-fn recent_user_context(session: &Session, limit: usize) -> String {
-    let requests = session
-        .messages
-        .iter()
-        .filter(|message| message.role == MessageRole::User)
-        .filter_map(|message| {
-            message.blocks.iter().find_map(|block| match block {
-                ContentBlock::Text { text } => Some(text.trim().to_string()),
-                _ => None,
-            })
-        })
-        .rev()
-        .take(limit)
-        .collect::<Vec<_>>();
-
-    if requests.is_empty() {
-        "<no prior user messages>".to_string()
-    } else {
-        requests
-            .into_iter()
-            .rev()
-            .enumerate()
-            .map(|(index, text)| format!("{}. {}", index + 1, text))
-            .collect::<Vec<_>>()
-            .join("\n")
-    }
-}
-
-fn truncate_for_prompt(value: &str, limit: usize) -> String {
-    if value.chars().count() <= limit {
-        value.trim().to_string()
-    } else {
-        let truncated = value.chars().take(limit).collect::<String>();
-        format!("{}\n…[truncated]", truncated.trim_end())
-    }
-}
-
-fn sanitize_generated_message(value: &str) -> String {
-    value.trim().trim_matches('`').trim().replace("\r\n", "\n")
-}
-
-fn parse_titled_body(value: &str) -> Option<(String, String)> {
-    let normalized = sanitize_generated_message(value);
-    let title = normalized
-        .lines()
-        .find_map(|line| line.strip_prefix("TITLE:").map(str::trim))?;
-    let body_start = normalized.find("BODY:")?;
-    let body = normalized[body_start + "BODY:".len()..].trim();
-    Some((title.to_string(), body.to_string()))
-}
-
 fn render_version_report() -> String {
     let git_sha = GIT_SHA.unwrap_or("unknown");
     let target = BUILD_TARGET.unwrap_or("unknown");
@@ -2306,8 +1914,19 @@ fn build_system_prompt() -> Result<Vec<String>, Box<dyn std::error::Error>> {
     )?)
 }
 
+fn build_runtime_feature_config(
+) -> Result<runtime::RuntimeFeatureConfig, Box<dyn std::error::Error>> {
+    let cwd = env::current_dir()?;
+    Ok(ConfigLoader::default_for(cwd)
+        .load()?
+        .feature_config()
+        .clone())
+}
+
+#[allow(clippy::too_many_arguments)]
 fn build_runtime(
     session: Session,
+    session_id: String,
     model: String,
     system_prompt: Vec<String>,
     enable_tools: bool,
@@ -2316,12 +1935,19 @@ fn build_runtime(
     permission_mode: PermissionMode,
 ) -> Result<ConversationRuntime<AnthropicRuntimeClient, CliToolExecutor>, Box<dyn std::error::Error>>
 {
-    Ok(ConversationRuntime::new(
+    Ok(ConversationRuntime::new_with_features(
         session,
-        AnthropicRuntimeClient::new(model, enable_tools, emit_output, allowed_tools.clone())?,
+        AnthropicRuntimeClient::new(
+            model,
+            enable_tools,
+            emit_output,
+            allowed_tools.clone(),
+            session_id,
+        )?,
         CliToolExecutor::new(allowed_tools, emit_output),
         permission_policy(permission_mode),
         system_prompt,
+        &build_runtime_feature_config()?,
     ))
 }
 
@@ -2386,11 +2012,13 @@ impl AnthropicRuntimeClient {
         enable_tools: bool,
         emit_output: bool,
         allowed_tools: Option<AllowedToolSet>,
+        session_id: impl Into<String>,
     ) -> Result<Self, Box<dyn std::error::Error>> {
         Ok(Self {
             runtime: tokio::runtime::Runtime::new()?,
             client: AnthropicClient::from_auth(resolve_cli_auth_source()?)
-                .with_base_url(api::read_base_url()),
+                .with_base_url(api::read_base_url())
+                .with_prompt_cache(PromptCache::new(session_id)),
             model,
             enable_tools,
             emit_output,
@@ -2505,8 +2133,8 @@ impl ApiClient for AnthropicRuntimeClient {
                         events.push(AssistantEvent::Usage(TokenUsage {
                             input_tokens: delta.usage.input_tokens,
                             output_tokens: delta.usage.output_tokens,
-                            cache_creation_input_tokens: 0,
-                            cache_read_input_tokens: 0,
+                            cache_creation_input_tokens: delta.usage.cache_creation_input_tokens,
+                            cache_read_input_tokens: delta.usage.cache_read_input_tokens,
                         }));
                     }
                     ApiStreamEvent::MessageStop(_) => {
@@ -2521,6 +2149,8 @@ impl ApiClient for AnthropicRuntimeClient {
                 }
             }
 
+            push_prompt_cache_record(&self.client, &mut events);
+
             if !saw_stop
                 && events.iter().any(|event| {
                     matches!(event, AssistantEvent::TextDelta(text) if !text.is_empty())
@@ -2545,7 +2175,9 @@ impl ApiClient for AnthropicRuntimeClient {
                 })
                 .await
                 .map_err(|error| RuntimeError::new(error.to_string()))?;
-            response_to_events(response, out)
+            let mut events = response_to_events(response, out)?;
+            push_prompt_cache_record(&self.client, &mut events);
+            Ok(events)
         })
     }
 }
@@ -2606,6 +2238,39 @@ fn collect_tool_results(summary: &runtime::TurnSummary) -> Vec<serde_json::Value
         .collect()
 }
 
+fn collect_prompt_cache_events(summary: &runtime::TurnSummary) -> Vec<serde_json::Value> {
+    summary
+        .prompt_cache_events
+        .iter()
+        .map(|event| {
+            json!({
+                "unexpected": event.unexpected,
+                "reason": event.reason,
+                "previous_cache_read_input_tokens": event.previous_cache_read_input_tokens,
+                "current_cache_read_input_tokens": event.current_cache_read_input_tokens,
+                "token_drop": event.token_drop,
+            })
+        })
+        .collect()
+}
+
+fn print_prompt_cache_events(summary: &runtime::TurnSummary) {
+    for event in &summary.prompt_cache_events {
+        let label = if event.unexpected {
+            "Prompt cache break"
+        } else {
+            "Prompt cache invalidation"
+        };
+        println!(
+            "{label}: {} (cache read {} -> {}, drop {})",
+            event.reason,
+            event.previous_cache_read_input_tokens,
+            event.current_cache_read_input_tokens,
+            event.token_drop,
+        );
+    }
+}
+
 fn slash_command_completion_candidates() -> Vec<String> {
     slash_command_specs()
         .iter()
@@ -2752,18 +2417,20 @@ fn first_visible_line(text: &str) -> &str {
 }
 
 fn format_bash_result(icon: &str, parsed: &serde_json::Value) -> String {
+    use std::fmt::Write as _;
+
     let mut lines = vec![format!("{icon} \x1b[38;5;245mbash\x1b[0m")];
     if let Some(task_id) = parsed
         .get("backgroundTaskId")
         .and_then(|value| value.as_str())
     {
-        lines[0].push_str(&format!(" backgrounded ({task_id})"));
+        let _ = write!(lines[0], " backgrounded ({task_id})");
     } else if let Some(status) = parsed
         .get("returnCodeInterpretation")
         .and_then(|value| value.as_str())
         .filter(|status| !status.is_empty())
     {
-        lines[0].push_str(&format!(" {status}"));
+        let _ = write!(lines[0], " {status}");
     }
 
     if let Some(stdout) = parsed.get("stdout").and_then(|value| value.as_str()) {
@@ -2785,15 +2452,15 @@ fn format_read_result(icon: &str, parsed: &serde_json::Value) -> String {
     let path = extract_tool_path(file);
     let start_line = file
         .get("startLine")
-        .and_then(|value| value.as_u64())
+        .and_then(serde_json::Value::as_u64)
         .unwrap_or(1);
     let num_lines = file
         .get("numLines")
-        .and_then(|value| value.as_u64())
+        .and_then(serde_json::Value::as_u64)
         .unwrap_or(0);
     let total_lines = file
         .get("totalLines")
-        .and_then(|value| value.as_u64())
+        .and_then(serde_json::Value::as_u64)
         .unwrap_or(num_lines);
     let content = file
         .get("content")
@@ -2819,8 +2486,7 @@ fn format_write_result(icon: &str, parsed: &serde_json::Value) -> String {
     let line_count = parsed
         .get("content")
         .and_then(|value| value.as_str())
-        .map(|content| content.lines().count())
-        .unwrap_or(0);
+        .map_or(0, |content| content.lines().count());
     format!(
         "{icon} \x1b[1;32m✏️ {} {path}\x1b[0m \x1b[2m({line_count} lines)\x1b[0m",
         if kind == "create" { "Wrote" } else { "Updated" },
@@ -2851,7 +2517,7 @@ fn format_edit_result(icon: &str, parsed: &serde_json::Value) -> String {
     let path = extract_tool_path(parsed);
     let suffix = if parsed
         .get("replaceAll")
-        .and_then(|value| value.as_bool())
+        .and_then(serde_json::Value::as_bool)
         .unwrap_or(false)
     {
         " (replace all)"
@@ -2879,7 +2545,7 @@ fn format_edit_result(icon: &str, parsed: &serde_json::Value) -> String {
 fn format_glob_result(icon: &str, parsed: &serde_json::Value) -> String {
     let num_files = parsed
         .get("numFiles")
-        .and_then(|value| value.as_u64())
+        .and_then(serde_json::Value::as_u64)
         .unwrap_or(0);
     let filenames = parsed
         .get("filenames")
@@ -2903,11 +2569,11 @@ fn format_glob_result(icon: &str, parsed: &serde_json::Value) -> String {
 fn format_grep_result(icon: &str, parsed: &serde_json::Value) -> String {
     let num_matches = parsed
         .get("numMatches")
-        .and_then(|value| value.as_u64())
+        .and_then(serde_json::Value::as_u64)
         .unwrap_or(0);
     let num_files = parsed
         .get("numFiles")
-        .and_then(|value| value.as_u64())
+        .and_then(serde_json::Value::as_u64)
         .unwrap_or(0);
     let content = parsed
         .get("content")
@@ -3014,6 +2680,26 @@ fn response_to_events(
     Ok(events)
 }
 
+fn push_prompt_cache_record(client: &AnthropicClient, events: &mut Vec<AssistantEvent>) {
+    if let Some(event) = client
+        .take_last_prompt_cache_record()
+        .and_then(prompt_cache_record_to_runtime_event)
+    {
+        events.push(AssistantEvent::PromptCache(event));
+    }
+}
+
+fn prompt_cache_record_to_runtime_event(record: PromptCacheRecord) -> Option<PromptCacheEvent> {
+    let cache_break = record.cache_break?;
+    Some(PromptCacheEvent {
+        unexpected: cache_break.unexpected,
+        reason: cache_break.reason,
+        previous_cache_read_input_tokens: cache_break.previous_cache_read_input_tokens,
+        current_cache_read_input_tokens: cache_break.current_cache_read_input_tokens,
+        token_drop: cache_break.token_drop,
+    })
+}
+
 struct CliToolExecutor {
     renderer: TerminalRenderer,
     emit_output: bool,

rust/crates/rusty-claude-cli/src/render.rs

@@ -286,7 +286,7 @@ impl TerminalRenderer {
     ) {
         match event {
             Event::Start(Tag::Heading { level, .. }) => {
-                self.start_heading(state, level as u8, output)
+                Self::start_heading(state, level as u8, output);
             }
             Event::End(TagEnd::Paragraph) => output.push_str("\n\n"),
             Event::Start(Tag::BlockQuote(..)) => self.start_quote(state, output),
@@ -426,7 +426,7 @@ impl TerminalRenderer {
         }
     }
 
-    fn start_heading(&self, state: &mut RenderState, level: u8, output: &mut String) {
+    fn start_heading(state: &mut RenderState, level: u8, output: &mut String) {
         state.heading_level = Some(level);
         if !output.is_empty() {
             output.push('\n');

rust/crates/tools/src/lib.rs

@@ -5,15 +5,15 @@ use std::time::{Duration, Instant};
 
 use api::{
     read_base_url, AnthropicClient, ContentBlockDelta, InputContentBlock, InputMessage,
-    MessageRequest, MessageResponse, OutputContentBlock, StreamEvent as ApiStreamEvent, ToolChoice,
-    ToolDefinition, ToolResultContentBlock,
+    MessageRequest, MessageResponse, OutputContentBlock, PromptCache, PromptCacheRecord,
+    StreamEvent as ApiStreamEvent, ToolChoice, ToolDefinition, ToolResultContentBlock,
 };
 use reqwest::blocking::Client;
 use runtime::{
     edit_file, execute_bash, glob_search, grep_search, load_system_prompt, read_file, write_file,
     ApiClient, ApiRequest, AssistantEvent, BashCommandInput, ContentBlock, ConversationMessage,
     ConversationRuntime, GrepSearchInput, MessageRole, PermissionMode, PermissionPolicy,
-    RuntimeError, Session, TokenUsage, ToolError, ToolExecutor,
+    PromptCacheEvent, RuntimeError, Session, TokenUsage, ToolError, ToolExecutor,
 };
 use serde::{Deserialize, Serialize};
 use serde_json::{json, Value};
@@ -1466,7 +1466,8 @@ fn build_agent_runtime(
         .clone()
         .unwrap_or_else(|| DEFAULT_AGENT_MODEL.to_string());
     let allowed_tools = job.allowed_tools.clone();
-    let api_client = AnthropicRuntimeClient::new(model, allowed_tools.clone())?;
+    let api_client =
+        AnthropicRuntimeClient::new(model, allowed_tools.clone(), job.manifest.agent_id.clone())?;
     let tool_executor = SubagentToolExecutor::new(allowed_tools);
     Ok(ConversationRuntime::new(
         Session::new(),
@@ -1643,10 +1644,15 @@ struct AnthropicRuntimeClient {
 }
 
 impl AnthropicRuntimeClient {
-    fn new(model: String, allowed_tools: BTreeSet<String>) -> Result<Self, String> {
+    fn new(
+        model: String,
+        allowed_tools: BTreeSet<String>,
+        session_id: impl Into<String>,
+    ) -> Result<Self, String> {
         let client = AnthropicClient::from_env()
             .map_err(|error| error.to_string())?
-            .with_base_url(read_base_url());
+            .with_base_url(read_base_url())
+            .with_prompt_cache(PromptCache::new(session_id));
         Ok(Self {
             runtime: tokio::runtime::Runtime::new().map_err(|error| error.to_string())?,
             client,
@@ -1657,6 +1663,7 @@ impl AnthropicRuntimeClient {
 }
 
 impl ApiClient for AnthropicRuntimeClient {
+    #[allow(clippy::too_many_lines)]
     fn stream(&mut self, request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError> {
         let tools = tool_specs_for_allowed_tools(Some(&self.allowed_tools))
             .into_iter()
@@ -1726,8 +1733,8 @@ impl ApiClient for AnthropicRuntimeClient {
                         events.push(AssistantEvent::Usage(TokenUsage {
                             input_tokens: delta.usage.input_tokens,
                             output_tokens: delta.usage.output_tokens,
-                            cache_creation_input_tokens: 0,
-                            cache_read_input_tokens: 0,
+                            cache_creation_input_tokens: delta.usage.cache_creation_input_tokens,
+                            cache_read_input_tokens: delta.usage.cache_read_input_tokens,
                         }));
                     }
                     ApiStreamEvent::MessageStop(_) => {
@@ -1737,6 +1744,8 @@ impl ApiClient for AnthropicRuntimeClient {
                 }
             }
 
+            push_prompt_cache_record(&self.client, &mut events);
+
             if !saw_stop
                 && events.iter().any(|event| {
                     matches!(event, AssistantEvent::TextDelta(text) if !text.is_empty())
@@ -1761,7 +1770,9 @@ impl ApiClient for AnthropicRuntimeClient {
                 })
                 .await
                 .map_err(|error| RuntimeError::new(error.to_string()))?;
-            Ok(response_to_events(response))
+            let mut events = response_to_events(response);
+            push_prompt_cache_record(&self.client, &mut events);
+            Ok(events)
         })
     }
 }
@@ -1884,6 +1895,26 @@ fn response_to_events(response: MessageResponse) -> Vec<AssistantEvent> {
     events
 }
 
+fn push_prompt_cache_record(client: &AnthropicClient, events: &mut Vec<AssistantEvent>) {
+    if let Some(event) = client
+        .take_last_prompt_cache_record()
+        .and_then(prompt_cache_record_to_runtime_event)
+    {
+        events.push(AssistantEvent::PromptCache(event));
+    }
+}
+
+fn prompt_cache_record_to_runtime_event(record: PromptCacheRecord) -> Option<PromptCacheEvent> {
+    let cache_break = record.cache_break?;
+    Some(PromptCacheEvent {
+        unexpected: cache_break.unexpected,
+        reason: cache_break.reason,
+        previous_cache_read_input_tokens: cache_break.previous_cache_read_input_tokens,
+        current_cache_read_input_tokens: cache_break.current_cache_read_input_tokens,
+        token_drop: cache_break.token_drop,
+    })
+}
+
 fn final_assistant_text(summary: &runtime::TurnSummary) -> String {
     summary
         .assistant_messages