Make agents and skills commands usable beyond placeholder parsing

Wire /agents and /skills through the Rust command stack so they can run as direct CLI subcommands, direct slash invocations, and resume-safe slash commands. The handlers now provide structured usage output, skills discovery also covers legacy /commands markdown entries, and the reporting/tests line up more closely with the original TypeScript behavior where feasible.

Constraint: The Rust port does not yet have the original TypeScript TUI menus or plugin/MCP skill registry, so text reports approximate those views
Rejected: Rebuild the original interactive React menus in Rust now | too large for the current CLI parity slice
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Keep /skills discovery and the Skill tool aligned if command/skill registry parity expands later
Tested: cargo test --workspace
Tested: cargo clippy --workspace --all-targets -- -D warnings
Tested: cargo run -q -p rusty-claude-cli -- agents --help
Tested: cargo run -q -p rusty-claude-cli -- /agents
Not-tested: Live Anthropic-backed REPL execution of /agents or /skills

rust/Cargo.lock

@@ -111,6 +111,7 @@ dependencies = [
 name = "commands"
 version = "0.1.0"
 dependencies = [
+ "plugins",
  "runtime",
 ]
 
@@ -825,6 +826,14 @@ dependencies = [
  "time",
 ]
 
+[[package]]
+name = "plugins"
+version = "0.1.0"
+dependencies = [
+ "serde",
+ "serde_json",
+]
+
 [[package]]
 name = "potential_utf"
 version = "0.1.4"
@@ -1092,6 +1101,7 @@ name = "runtime"
 version = "0.1.0"
 dependencies = [
  "glob",
+ "plugins",
  "regex",
  "serde",
  "serde_json",
@@ -1181,6 +1191,7 @@ dependencies = [
  "commands",
  "compat-harness",
  "crossterm",
+ "plugins",
  "pulldown-cmark",
  "runtime",
  "rustyline",
@@ -1546,6 +1557,7 @@ name = "tools"
 version = "0.1.0"
 dependencies = [
  "api",
+ "plugins",
  "reqwest",
  "runtime",
  "serde",

rust/README.md

@@ -1,6 +1,6 @@
 # 🦞 Claw Code — Rust Implementation
 
-A high-performance Rust rewrite of the Claude Code CLI agent harness. Built for speed, safety, and native tool execution.
+A high-performance Rust rewrite of the Claw Code CLI agent harness. Built for speed, safety, and native tool execution.
 
 ## Quick Start
 

rust/crates/api/src/client.rs

@@ -1,5 +1,4 @@
 use std::collections::VecDeque;
-use std::sync::{Arc, Mutex};
 use std::time::{Duration, SystemTime, UNIX_EPOCH};
 
 use runtime::{
@@ -9,9 +8,8 @@ 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, Usage};
+use crate::types::{MessageRequest, MessageResponse, StreamEvent};
 
 const DEFAULT_BASE_URL: &str = "https://api.anthropic.com";
 const ANTHROPIC_VERSION: &str = "2023-06-01";
@@ -110,8 +108,6 @@ 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 {
@@ -124,8 +120,6 @@ 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)),
         }
     }
 
@@ -138,8 +132,6 @@ 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)),
         }
     }
 
@@ -197,30 +189,6 @@ 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
@@ -230,19 +198,10 @@ 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
@@ -252,10 +211,6 @@ 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)
     }
 
@@ -263,7 +218,6 @@ 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?;
@@ -273,30 +227,9 @@ 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,
@@ -594,7 +527,6 @@ pub struct MessageStream {
     parser: SseParser,
     pending: VecDeque<StreamEvent>,
     done: bool,
-    cache_tracking: Option<StreamCacheTracking>,
 }
 
 impl MessageStream {
@@ -606,9 +538,6 @@ 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));
             }
 
@@ -616,14 +545,8 @@ 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);
             }
 
@@ -639,53 +562,6 @@ 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() {
@@ -730,7 +606,7 @@ mod tests {
     use super::{ALT_REQUEST_ID_HEADER, REQUEST_ID_HEADER};
     use std::io::{Read, Write};
     use std::net::TcpListener;
-    use std::sync::atomic::{AtomicU64, Ordering};
+    use std::sync::{Mutex, OnceLock};
     use std::thread;
     use std::time::{Duration, SystemTime, UNIX_EPOCH};
 
@@ -740,15 +616,19 @@ 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")
@@ -788,7 +668,7 @@ mod tests {
 
     #[test]
     fn read_api_key_requires_presence() {
-        let _guard = test_env_lock();
+        let _guard = env_lock();
         std::env::remove_var("ANTHROPIC_AUTH_TOKEN");
         std::env::remove_var("ANTHROPIC_API_KEY");
         std::env::remove_var("CLAUDE_CONFIG_HOME");
@@ -798,7 +678,7 @@ mod tests {
 
     #[test]
     fn read_api_key_requires_non_empty_value() {
-        let _guard = test_env_lock();
+        let _guard = 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");
@@ -808,7 +688,7 @@ mod tests {
 
     #[test]
     fn read_api_key_prefers_api_key_env() {
-        let _guard = test_env_lock();
+        let _guard = env_lock();
         std::env::set_var("ANTHROPIC_AUTH_TOKEN", "auth-token");
         std::env::set_var("ANTHROPIC_API_KEY", "legacy-key");
         assert_eq!(
@@ -821,7 +701,7 @@ mod tests {
 
     #[test]
     fn read_auth_token_reads_auth_token_env() {
-        let _guard = test_env_lock();
+        let _guard = 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");
@@ -841,7 +721,7 @@ mod tests {
 
     #[test]
     fn auth_source_from_env_combines_api_key_and_bearer_token() {
-        let _guard = test_env_lock();
+        let _guard = 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");
@@ -853,7 +733,7 @@ mod tests {
 
     #[test]
     fn auth_source_from_saved_oauth_when_env_absent() {
-        let _guard = test_env_lock();
+        let _guard = env_lock();
         let config_home = temp_config_home();
         std::env::set_var("CLAUDE_CONFIG_HOME", &config_home);
         std::env::remove_var("ANTHROPIC_AUTH_TOKEN");
@@ -892,7 +772,7 @@ mod tests {
 
     #[test]
     fn resolve_saved_oauth_token_refreshes_expired_credentials() {
-        let _guard = test_env_lock();
+        let _guard = env_lock();
         let config_home = temp_config_home();
         std::env::set_var("CLAUDE_CONFIG_HOME", &config_home);
         std::env::remove_var("ANTHROPIC_AUTH_TOKEN");
@@ -924,7 +804,7 @@ mod tests {
 
     #[test]
     fn resolve_startup_auth_source_uses_saved_oauth_without_loading_config() {
-        let _guard = test_env_lock();
+        let _guard = env_lock();
         let config_home = temp_config_home();
         std::env::set_var("CLAUDE_CONFIG_HOME", &config_home);
         std::env::remove_var("ANTHROPIC_AUTH_TOKEN");
@@ -948,7 +828,7 @@ mod tests {
 
     #[test]
     fn resolve_startup_auth_source_errors_when_refreshable_token_lacks_config() {
-        let _guard = test_env_lock();
+        let _guard = env_lock();
         let config_home = temp_config_home();
         std::env::set_var("CLAUDE_CONFIG_HOME", &config_home);
         std::env::remove_var("ANTHROPIC_AUTH_TOKEN");
@@ -980,7 +860,7 @@ mod tests {
 
     #[test]
     fn resolve_saved_oauth_token_preserves_refresh_token_when_refresh_response_omits_it() {
-        let _guard = test_env_lock();
+        let _guard = 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

@@ -1,6 +1,5 @@
 mod client;
 mod error;
-mod prompt_cache;
 mod sse;
 mod types;
 
@@ -9,10 +8,6 @@ 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,
@@ -20,11 +15,3 @@ 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

@@ -1,727 +0,0 @@
-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/src/sse.rs

@@ -216,4 +216,64 @@ mod tests {
             ))
         );
     }
+
+    #[test]
+    fn parses_thinking_content_block_start() {
+        let frame = concat!(
+            "event: content_block_start\n",
+            "data: {\"type\":\"content_block_start\",\"index\":0,\"content_block\":{\"type\":\"thinking\",\"thinking\":\"\",\"signature\":null}}\n\n"
+        );
+
+        let event = parse_frame(frame).expect("frame should parse");
+        assert_eq!(
+            event,
+            Some(StreamEvent::ContentBlockStart(
+                crate::types::ContentBlockStartEvent {
+                    index: 0,
+                    content_block: OutputContentBlock::Thinking {
+                        thinking: String::new(),
+                        signature: None,
+                    },
+                },
+            ))
+        );
+    }
+
+    #[test]
+    fn parses_thinking_related_deltas() {
+        let thinking = concat!(
+            "event: content_block_delta\n",
+            "data: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"thinking_delta\",\"thinking\":\"step 1\"}}\n\n"
+        );
+        let signature = concat!(
+            "event: content_block_delta\n",
+            "data: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"signature_delta\",\"signature\":\"sig_123\"}}\n\n"
+        );
+
+        let thinking_event = parse_frame(thinking).expect("thinking delta should parse");
+        let signature_event = parse_frame(signature).expect("signature delta should parse");
+
+        assert_eq!(
+            thinking_event,
+            Some(StreamEvent::ContentBlockDelta(
+                crate::types::ContentBlockDeltaEvent {
+                    index: 0,
+                    delta: ContentBlockDelta::ThinkingDelta {
+                        thinking: "step 1".to_string(),
+                    },
+                }
+            ))
+        );
+        assert_eq!(
+            signature_event,
+            Some(StreamEvent::ContentBlockDelta(
+                crate::types::ContentBlockDeltaEvent {
+                    index: 0,
+                    delta: ContentBlockDelta::SignatureDelta {
+                        signature: "sig_123".to_string(),
+                    },
+                }
+            ))
+        );
+    }
 }

rust/crates/api/src/types.rs

@@ -135,6 +135,15 @@ pub enum OutputContentBlock {
         name: String,
         input: Value,
     },
+    Thinking {
+        #[serde(default)]
+        thinking: String,
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        signature: Option<String>,
+    },
+    RedactedThinking {
+        data: Value,
+    },
 }
 
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
@@ -190,6 +199,8 @@ pub struct ContentBlockDeltaEvent {
 pub enum ContentBlockDelta {
     TextDelta { text: String },
     InputJsonDelta { partial_json: String },
+    ThinkingDelta { thinking: String },
+    SignatureDelta { signature: String },
 }
 
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]

rust/crates/api/tests/client_integration.rs

@@ -1,25 +1,17 @@
 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,
-    PromptCache, StreamEvent, ToolChoice, ToolDefinition,
+    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()));
@@ -53,8 +45,6 @@ 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 {
@@ -86,18 +76,21 @@ async fn send_message_posts_json_and_parses_response() {
 }
 
 #[tokio::test]
-async fn send_message_parses_prompt_cache_token_usage_from_response() {
+async fn send_message_parses_response_with_thinking_blocks() {
     let state = Arc::new(Mutex::new(Vec::<CapturedRequest>::new()));
     let body = concat!(
         "{",
-        "\"id\":\"msg_cache_tokens\",",
+        "\"id\":\"msg_thinking\",",
         "\"type\":\"message\",",
         "\"role\":\"assistant\",",
-        "\"content\":[{\"type\":\"text\",\"text\":\"Cache tokens\"}],",
+        "\"content\":[",
+        "{\"type\":\"thinking\",\"thinking\":\"step 1\",\"signature\":\"sig_123\"},",
+        "{\"type\":\"text\",\"text\":\"Final answer\"}",
+        "],",
         "\"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}",
+        "\"usage\":{\"input_tokens\":12,\"output_tokens\":4}",
         "}"
     );
     let server = spawn_server(
@@ -112,29 +105,24 @@ async fn send_message_parses_prompt_cache_token_usage_from_response() {
         .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);
+    assert_eq!(response.content.len(), 2);
+    assert!(matches!(
+        &response.content[0],
+        OutputContentBlock::Thinking { thinking, signature }
+            if thinking == "step 1" && signature.as_deref() == Some("sig_123")
+    ));
+    assert!(matches!(
+        &response.content[1],
+        OutputContentBlock::Text { text } if text == "Final answer"
+    ));
 }
 
 #[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,\"cache_creation_input_tokens\":13,\"cache_read_input_tokens\":21,\"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,\"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",
@@ -142,7 +130,7 @@ 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,\"cache_creation_input_tokens\":34,\"cache_read_input_tokens\":55,\"output_tokens\":1}}\n\n",
+        "data: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":\"tool_use\",\"stop_sequence\":null},\"usage\":{\"input_tokens\":8,\"output_tokens\":1}}\n\n",
         "event: message_stop\n",
         "data: {\"type\":\"message_stop\"}\n\n",
         "data: [DONE]\n\n"
@@ -160,8 +148,7 @@ 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_prompt_cache(PromptCache::new("stream-session"));
+        .with_base_url(server.base_url());
     let mut stream = client
         .stream_message(&sample_request(false))
         .await
@@ -215,20 +202,85 @@ 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")
+#[tokio::test]
+async fn stream_message_parses_sse_events_with_thinking_blocks() {
+    let state = Arc::new(Mutex::new(Vec::<CapturedRequest>::new()));
+    let sse = concat!(
+        "event: message_start\n",
+        "data: {\"type\":\"message_start\",\"message\":{\"id\":\"msg_stream_thinking\",\"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",
+        "event: content_block_start\n",
+        "data: {\"type\":\"content_block_start\",\"index\":0,\"content_block\":{\"type\":\"thinking\",\"thinking\":\"\"}}\n\n",
+        "event: content_block_delta\n",
+        "data: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"thinking_delta\",\"thinking\":\"step 1\"}}\n\n",
+        "event: content_block_delta\n",
+        "data: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"signature_delta\",\"signature\":\"sig_123\"}}\n\n",
+        "event: content_block_stop\n",
+        "data: {\"type\":\"content_block_stop\",\"index\":0}\n\n",
+        "event: content_block_start\n",
+        "data: {\"type\":\"content_block_start\",\"index\":1,\"content_block\":{\"type\":\"text\",\"text\":\"Final answer\"}}\n\n",
+        "event: content_block_stop\n",
+        "data: {\"type\":\"content_block_stop\",\"index\":1}\n\n",
+        "event: message_delta\n",
+        "data: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":\"end_turn\",\"stop_sequence\":null},\"usage\":{\"input_tokens\":8,\"output_tokens\":1}}\n\n",
+        "event: message_stop\n",
+        "data: {\"type\":\"message_stop\"}\n\n",
+        "data: [DONE]\n\n"
     );
+    let server = spawn_server(
+        state,
+        vec![http_response("200 OK", "text/event-stream", sse)],
+    )
+    .await;
 
-    std::fs::remove_dir_all(temp_root).expect("cleanup temp root");
-    std::env::remove_var("CLAUDE_CONFIG_HOME");
+    let client = AnthropicClient::new("test-key").with_base_url(server.base_url());
+    let mut stream = client
+        .stream_message(&sample_request(false))
+        .await
+        .expect("stream should start");
+
+    let mut events = Vec::new();
+    while let Some(event) = stream
+        .next_event()
+        .await
+        .expect("stream event should parse")
+    {
+        events.push(event);
+    }
+
+    assert_eq!(events.len(), 9);
+    assert!(matches!(
+        &events[1],
+        StreamEvent::ContentBlockStart(ContentBlockStartEvent {
+            content_block: OutputContentBlock::Thinking { thinking, signature },
+            ..
+        }) if thinking.is_empty() && signature.is_none()
+    ));
+    assert!(matches!(
+        &events[2],
+        StreamEvent::ContentBlockDelta(ContentBlockDeltaEvent {
+            delta: ContentBlockDelta::ThinkingDelta { thinking },
+            ..
+        }) if thinking == "step 1"
+    ));
+    assert!(matches!(
+        &events[3],
+        StreamEvent::ContentBlockDelta(ContentBlockDeltaEvent {
+            delta: ContentBlockDelta::SignatureDelta { signature },
+            ..
+        }) if signature == "sig_123"
+    ));
+    assert!(matches!(
+        &events[5],
+        StreamEvent::ContentBlockStart(ContentBlockStartEvent {
+            content_block: OutputContentBlock::Text { text },
+            ..
+        }) if text == "Final answer"
+    ));
+    assert!(matches!(events[6], StreamEvent::ContentBlockStop(_)));
+    assert!(matches!(events[7], StreamEvent::MessageDelta(_)));
+    assert!(matches!(events[8], StreamEvent::MessageStop(_)));
 }
 
 #[tokio::test]
@@ -312,121 +364,6 @@ 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/Cargo.toml

@@ -9,4 +9,5 @@ publish.workspace = true
 workspace = true
 
 [dependencies]
+plugins = { path = "../plugins" }
 runtime = { path = "../runtime" }

rust/crates/compat-harness/src/lib.rs

@@ -70,16 +70,12 @@ fn upstream_repo_candidates(primary_repo_root: &Path) -> Vec<PathBuf> {
     }
 
     for ancestor in primary_repo_root.ancestors().take(4) {
-        candidates.push(ancestor.join("claude-code"));
+        candidates.push(ancestor.join("claw-code"));
         candidates.push(ancestor.join("clawd-code"));
     }
 
-    candidates.push(
-        primary_repo_root
-            .join("reference-source")
-            .join("claude-code"),
-    );
-    candidates.push(primary_repo_root.join("vendor").join("claude-code"));
+    candidates.push(primary_repo_root.join("reference-source").join("claw-code"));
+    candidates.push(primary_repo_root.join("vendor").join("claw-code"));
 
     let mut deduped = Vec::new();
     for candidate in candidates {

rust/crates/plugins/Cargo.toml

@@ -0,0 +1,13 @@
+[package]
+name = "plugins"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+publish.workspace = true
+
+[dependencies]
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
+
+[lints]
+workspace = true

rust/crates/plugins/bundled/example-bundled/.claude-plugin/plugin.json

@@ -0,0 +1,10 @@
+{
+  "name": "example-bundled",
+  "version": "0.1.0",
+  "description": "Example bundled plugin scaffold for the Rust plugin system",
+  "defaultEnabled": false,
+  "hooks": {
+    "PreToolUse": ["./hooks/pre.sh"],
+    "PostToolUse": ["./hooks/post.sh"]
+  }
+}

rust/crates/plugins/bundled/example-bundled/hooks/post.sh

@@ -0,0 +1,2 @@
+#!/bin/sh
+printf '%s\n' 'example bundled post hook'

rust/crates/plugins/bundled/example-bundled/hooks/pre.sh

@@ -0,0 +1,2 @@
+#!/bin/sh
+printf '%s\n' 'example bundled pre hook'

rust/crates/plugins/bundled/sample-hooks/.claude-plugin/plugin.json

@@ -0,0 +1,10 @@
+{
+  "name": "sample-hooks",
+  "version": "0.1.0",
+  "description": "Bundled sample plugin scaffold for hook integration tests.",
+  "defaultEnabled": false,
+  "hooks": {
+    "PreToolUse": ["./hooks/pre.sh"],
+    "PostToolUse": ["./hooks/post.sh"]
+  }
+}

rust/crates/plugins/bundled/sample-hooks/hooks/post.sh

@@ -0,0 +1,2 @@
+#!/bin/sh
+printf 'sample bundled post hook'

rust/crates/plugins/bundled/sample-hooks/hooks/pre.sh

@@ -0,0 +1,2 @@
+#!/bin/sh
+printf 'sample bundled pre hook'

rust/crates/plugins/src/hooks.rs

@@ -0,0 +1,395 @@
+use std::ffi::OsStr;
+use std::path::Path;
+use std::process::Command;
+
+use serde_json::json;
+
+use crate::{PluginError, PluginHooks, PluginRegistry};
+
+#[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 {
+    hooks: PluginHooks,
+}
+
+impl HookRunner {
+    #[must_use]
+    pub fn new(hooks: PluginHooks) -> Self {
+        Self { hooks }
+    }
+
+    pub fn from_registry(plugin_registry: &PluginRegistry) -> Result<Self, PluginError> {
+        Ok(Self::new(plugin_registry.aggregated_hooks()?))
+    }
+
+    #[must_use]
+    pub fn run_pre_tool_use(&self, tool_name: &str, tool_input: &str) -> HookRunResult {
+        self.run_commands(
+            HookEvent::PreToolUse,
+            &self.hooks.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.hooks.post_tool_use,
+            tool_name,
+            tool_input,
+            Some(tool_output),
+            is_error,
+        )
+    }
+
+    fn run_commands(
+        &self,
+        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 } => {
+                    messages.push(message.unwrap_or_else(|| {
+                        format!("{} hook denied tool `{tool_name}`", event.as_str())
+                    }));
+                    return HookRunResult {
+                        denied: true,
+                        messages,
+                    };
+                }
+                HookCommandOutcome::Warn { message } => messages.push(message),
+            }
+        }
+
+        HookRunResult::allow(messages)
+    }
+
+    #[allow(clippy::too_many_arguments, clippy::unused_self)]
+    fn run_command(
+        &self,
+        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 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 = if Path::new(command).exists() {
+        let mut command_builder = Command::new("sh");
+        command_builder.arg(command);
+        CommandWithStdin::new(command_builder)
+    } else {
+        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 as _;
+            child_stdin.write_all(stdin)?;
+        }
+        child.wait_with_output()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{HookRunResult, HookRunner};
+    use crate::{PluginManager, PluginManagerConfig};
+    use std::fs;
+    use std::path::{Path, PathBuf};
+    use std::time::{SystemTime, UNIX_EPOCH};
+
+    fn temp_dir(label: &str) -> PathBuf {
+        let nanos = SystemTime::now()
+            .duration_since(UNIX_EPOCH)
+            .expect("time should be after epoch")
+            .as_nanos();
+        std::env::temp_dir().join(format!("plugins-hook-runner-{label}-{nanos}"))
+    }
+
+    fn write_hook_plugin(root: &Path, name: &str, pre_message: &str, post_message: &str) {
+        fs::create_dir_all(root.join(".claude-plugin")).expect("manifest dir");
+        fs::create_dir_all(root.join("hooks")).expect("hooks dir");
+        fs::write(
+            root.join("hooks").join("pre.sh"),
+            format!("#!/bin/sh\nprintf '%s\\n' '{pre_message}'\n"),
+        )
+        .expect("write pre hook");
+        fs::write(
+            root.join("hooks").join("post.sh"),
+            format!("#!/bin/sh\nprintf '%s\\n' '{post_message}'\n"),
+        )
+        .expect("write post hook");
+        fs::write(
+            root.join(".claude-plugin").join("plugin.json"),
+            format!(
+                "{{\n  \"name\": \"{name}\",\n  \"version\": \"1.0.0\",\n  \"description\": \"hook plugin\",\n  \"hooks\": {{\n    \"PreToolUse\": [\"./hooks/pre.sh\"],\n    \"PostToolUse\": [\"./hooks/post.sh\"]\n  }}\n}}"
+            ),
+        )
+        .expect("write plugin manifest");
+    }
+
+    #[test]
+    fn collects_and_runs_hooks_from_enabled_plugins() {
+        let config_home = temp_dir("config");
+        let first_source_root = temp_dir("source-a");
+        let second_source_root = temp_dir("source-b");
+        write_hook_plugin(
+            &first_source_root,
+            "first",
+            "plugin pre one",
+            "plugin post one",
+        );
+        write_hook_plugin(
+            &second_source_root,
+            "second",
+            "plugin pre two",
+            "plugin post two",
+        );
+
+        let mut manager = PluginManager::new(PluginManagerConfig::new(&config_home));
+        manager
+            .install(first_source_root.to_str().expect("utf8 path"))
+            .expect("first plugin install should succeed");
+        manager
+            .install(second_source_root.to_str().expect("utf8 path"))
+            .expect("second plugin install should succeed");
+        let registry = manager.plugin_registry().expect("registry should build");
+
+        let runner = HookRunner::from_registry(&registry).expect("plugin hooks should load");
+
+        assert_eq!(
+            runner.run_pre_tool_use("Read", r#"{"path":"README.md"}"#),
+            HookRunResult::allow(vec![
+                "plugin pre one".to_string(),
+                "plugin pre two".to_string(),
+            ])
+        );
+        assert_eq!(
+            runner.run_post_tool_use("Read", r#"{"path":"README.md"}"#, "ok", false),
+            HookRunResult::allow(vec![
+                "plugin post one".to_string(),
+                "plugin post two".to_string(),
+            ])
+        );
+
+        let _ = fs::remove_dir_all(config_home);
+        let _ = fs::remove_dir_all(first_source_root);
+        let _ = fs::remove_dir_all(second_source_root);
+    }
+
+    #[test]
+    fn pre_tool_use_denies_when_plugin_hook_exits_two() {
+        let runner = HookRunner::new(crate::PluginHooks {
+            pre_tool_use: vec!["printf 'blocked by plugin'; exit 2".to_string()],
+            post_tool_use: Vec::new(),
+        });
+
+        let result = runner.run_pre_tool_use("Bash", r#"{"command":"pwd"}"#);
+
+        assert!(result.is_denied());
+        assert_eq!(result.messages(), &["blocked by plugin".to_string()]);
+    }
+}

rust/crates/runtime/Cargo.toml

@@ -8,6 +8,7 @@ publish.workspace = true
 [dependencies]
 sha2 = "0.10"
 glob = "0.3"
+plugins = { path = "../plugins" }
 regex = "1"
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"

rust/crates/runtime/src/compact.rs

@@ -1,5 +1,10 @@
 use crate::session::{ContentBlock, ConversationMessage, MessageRole, Session};
 
+const COMPACT_CONTINUATION_PREAMBLE: &str =
+    "This session is being continued from a previous conversation that ran out of context. The summary below covers the earlier portion of the conversation.\n\n";
+const COMPACT_RECENT_MESSAGES_NOTE: &str = "Recent messages are preserved verbatim.";
+const COMPACT_DIRECT_RESUME_INSTRUCTION: &str = "Continue the conversation from where it left off without asking the user any further questions. Resume directly — do not acknowledge the summary, do not recap what was happening, and do not preface with continuation text.";
+
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub struct CompactionConfig {
     pub preserve_recent_messages: usize,
@@ -30,8 +35,15 @@ pub fn estimate_session_tokens(session: &Session) -> usize {
 
 #[must_use]
 pub fn should_compact(session: &Session, config: CompactionConfig) -> bool {
-    session.messages.len() > config.preserve_recent_messages
-        && estimate_session_tokens(session) >= config.max_estimated_tokens
+    let start = compacted_summary_prefix_len(session);
+    let compactable = &session.messages[start..];
+
+    compactable.len() > config.preserve_recent_messages
+        && compactable
+            .iter()
+            .map(estimate_message_tokens)
+            .sum::<usize>()
+            >= config.max_estimated_tokens
 }
 
 #[must_use]
@@ -56,16 +68,18 @@ pub fn get_compact_continuation_message(
     recent_messages_preserved: bool,
 ) -> String {
     let mut base = format!(
-        "This session is being continued from a previous conversation that ran out of context. The summary below covers the earlier portion of the conversation.\n\n{}",
+        "{COMPACT_CONTINUATION_PREAMBLE}{}",
         format_compact_summary(summary)
     );
 
     if recent_messages_preserved {
-        base.push_str("\n\nRecent messages are preserved verbatim.");
+        base.push_str("\n\n");
+        base.push_str(COMPACT_RECENT_MESSAGES_NOTE);
     }
 
     if suppress_follow_up_questions {
-        base.push_str("\nContinue the conversation from where it left off without asking the user any further questions. Resume directly — do not acknowledge the summary, do not recap what was happening, and do not preface with continuation text.");
+        base.push('\n');
+        base.push_str(COMPACT_DIRECT_RESUME_INSTRUCTION);
     }
 
     base
@@ -82,13 +96,19 @@ pub fn compact_session(session: &Session, config: CompactionConfig) -> Compactio
         };
     }
 
+    let existing_summary = session
+        .messages
+        .first()
+        .and_then(extract_existing_compacted_summary);
+    let compacted_prefix_len = usize::from(existing_summary.is_some());
     let keep_from = session
         .messages
         .len()
         .saturating_sub(config.preserve_recent_messages);
-    let removed = &session.messages[..keep_from];
+    let removed = &session.messages[compacted_prefix_len..keep_from];
     let preserved = session.messages[keep_from..].to_vec();
-    let summary = summarize_messages(removed);
+    let summary =
+        merge_compact_summaries(existing_summary.as_deref(), &summarize_messages(removed));
     let formatted_summary = format_compact_summary(&summary);
     let continuation = get_compact_continuation_message(&summary, true, !preserved.is_empty());
 
@@ -110,6 +130,16 @@ pub fn compact_session(session: &Session, config: CompactionConfig) -> Compactio
     }
 }
 
+fn compacted_summary_prefix_len(session: &Session) -> usize {
+    usize::from(
+        session
+            .messages
+            .first()
+            .and_then(extract_existing_compacted_summary)
+            .is_some(),
+    )
+}
+
 fn summarize_messages(messages: &[ConversationMessage]) -> String {
     let user_messages = messages
         .iter()
@@ -197,6 +227,41 @@ fn summarize_messages(messages: &[ConversationMessage]) -> String {
     lines.join("\n")
 }
 
+fn merge_compact_summaries(existing_summary: Option<&str>, new_summary: &str) -> String {
+    let Some(existing_summary) = existing_summary else {
+        return new_summary.to_string();
+    };
+
+    let previous_highlights = extract_summary_highlights(existing_summary);
+    let new_formatted_summary = format_compact_summary(new_summary);
+    let new_highlights = extract_summary_highlights(&new_formatted_summary);
+    let new_timeline = extract_summary_timeline(&new_formatted_summary);
+
+    let mut lines = vec!["<summary>".to_string(), "Conversation summary:".to_string()];
+
+    if !previous_highlights.is_empty() {
+        lines.push("- Previously compacted context:".to_string());
+        lines.extend(
+            previous_highlights
+                .into_iter()
+                .map(|line| format!("  {line}")),
+        );
+    }
+
+    if !new_highlights.is_empty() {
+        lines.push("- Newly compacted context:".to_string());
+        lines.extend(new_highlights.into_iter().map(|line| format!("  {line}")));
+    }
+
+    if !new_timeline.is_empty() {
+        lines.push("- Key timeline:".to_string());
+        lines.extend(new_timeline.into_iter().map(|line| format!("  {line}")));
+    }
+
+    lines.push("</summary>".to_string());
+    lines.join("\n")
+}
+
 fn summarize_block(block: &ContentBlock) -> String {
     let raw = match block {
         ContentBlock::Text { text } => text.clone(),
@@ -374,11 +439,71 @@ fn collapse_blank_lines(content: &str) -> String {
     result
 }
 
+fn extract_existing_compacted_summary(message: &ConversationMessage) -> Option<String> {
+    if message.role != MessageRole::System {
+        return None;
+    }
+
+    let text = first_text_block(message)?;
+    let summary = text.strip_prefix(COMPACT_CONTINUATION_PREAMBLE)?;
+    let summary = summary
+        .split_once(&format!("\n\n{COMPACT_RECENT_MESSAGES_NOTE}"))
+        .map_or(summary, |(value, _)| value);
+    let summary = summary
+        .split_once(&format!("\n{COMPACT_DIRECT_RESUME_INSTRUCTION}"))
+        .map_or(summary, |(value, _)| value);
+    Some(summary.trim().to_string())
+}
+
+fn extract_summary_highlights(summary: &str) -> Vec<String> {
+    let mut lines = Vec::new();
+    let mut in_timeline = false;
+
+    for line in format_compact_summary(summary).lines() {
+        let trimmed = line.trim_end();
+        if trimmed.is_empty() || trimmed == "Summary:" || trimmed == "Conversation summary:" {
+            continue;
+        }
+        if trimmed == "- Key timeline:" {
+            in_timeline = true;
+            continue;
+        }
+        if in_timeline {
+            continue;
+        }
+        lines.push(trimmed.to_string());
+    }
+
+    lines
+}
+
+fn extract_summary_timeline(summary: &str) -> Vec<String> {
+    let mut lines = Vec::new();
+    let mut in_timeline = false;
+
+    for line in format_compact_summary(summary).lines() {
+        let trimmed = line.trim_end();
+        if trimmed == "- Key timeline:" {
+            in_timeline = true;
+            continue;
+        }
+        if !in_timeline {
+            continue;
+        }
+        if trimmed.is_empty() {
+            break;
+        }
+        lines.push(trimmed.to_string());
+    }
+
+    lines
+}
+
 #[cfg(test)]
 mod tests {
     use super::{
         collect_key_files, compact_session, estimate_session_tokens, format_compact_summary,
-        infer_pending_work, should_compact, CompactionConfig,
+        get_compact_continuation_message, infer_pending_work, should_compact, CompactionConfig,
     };
     use crate::session::{ContentBlock, ConversationMessage, MessageRole, Session};
 
@@ -453,6 +578,98 @@ mod tests {
         );
     }
 
+    #[test]
+    fn keeps_previous_compacted_context_when_compacting_again() {
+        let initial_session = Session {
+            version: 1,
+            messages: vec![
+                ConversationMessage::user_text("Investigate rust/crates/runtime/src/compact.rs"),
+                ConversationMessage::assistant(vec![ContentBlock::Text {
+                    text: "I will inspect the compact flow.".to_string(),
+                }]),
+                ConversationMessage::user_text(
+                    "Also update rust/crates/runtime/src/conversation.rs",
+                ),
+                ConversationMessage::assistant(vec![ContentBlock::Text {
+                    text: "Next: preserve prior summary context during auto compact.".to_string(),
+                }]),
+            ],
+        };
+        let config = CompactionConfig {
+            preserve_recent_messages: 2,
+            max_estimated_tokens: 1,
+        };
+
+        let first = compact_session(&initial_session, config);
+        let mut follow_up_messages = first.compacted_session.messages.clone();
+        follow_up_messages.extend([
+            ConversationMessage::user_text("Please add regression tests for compaction."),
+            ConversationMessage::assistant(vec![ContentBlock::Text {
+                text: "Working on regression coverage now.".to_string(),
+            }]),
+        ]);
+
+        let second = compact_session(
+            &Session {
+                version: 1,
+                messages: follow_up_messages,
+            },
+            config,
+        );
+
+        assert!(second
+            .formatted_summary
+            .contains("Previously compacted context:"));
+        assert!(second
+            .formatted_summary
+            .contains("Scope: 2 earlier messages compacted"));
+        assert!(second
+            .formatted_summary
+            .contains("Newly compacted context:"));
+        assert!(second
+            .formatted_summary
+            .contains("Also update rust/crates/runtime/src/conversation.rs"));
+        assert!(matches!(
+            &second.compacted_session.messages[0].blocks[0],
+            ContentBlock::Text { text }
+                if text.contains("Previously compacted context:")
+                    && text.contains("Newly compacted context:")
+        ));
+        assert!(matches!(
+            &second.compacted_session.messages[1].blocks[0],
+            ContentBlock::Text { text } if text.contains("Please add regression tests for compaction.")
+        ));
+    }
+
+    #[test]
+    fn ignores_existing_compacted_summary_when_deciding_to_recompact() {
+        let summary = "<summary>Conversation summary:\n- Scope: earlier work preserved.\n- Key timeline:\n  - user: large preserved context\n</summary>";
+        let session = Session {
+            version: 1,
+            messages: vec![
+                ConversationMessage {
+                    role: MessageRole::System,
+                    blocks: vec![ContentBlock::Text {
+                        text: get_compact_continuation_message(summary, true, true),
+                    }],
+                    usage: None,
+                },
+                ConversationMessage::user_text("tiny"),
+                ConversationMessage::assistant(vec![ContentBlock::Text {
+                    text: "recent".to_string(),
+                }]),
+            ],
+        };
+
+        assert!(!should_compact(
+            &session,
+            CompactionConfig {
+                preserve_recent_messages: 2,
+                max_estimated_tokens: 1,
+            }
+        ));
+    }
+
     #[test]
     fn truncates_long_blocks_in_summary() {
         let summary = super::summarize_block(&ContentBlock::Text {

rust/crates/runtime/src/config.rs

@@ -35,9 +35,19 @@ pub struct RuntimeConfig {
     feature_config: RuntimeFeatureConfig,
 }
 
+#[derive(Debug, Clone, PartialEq, Eq, Default)]
+pub struct RuntimePluginConfig {
+    enabled_plugins: BTreeMap<String, bool>,
+    external_directories: Vec<String>,
+    install_root: Option<String>,
+    registry_path: Option<String>,
+    bundled_root: Option<String>,
+}
+
 #[derive(Debug, Clone, PartialEq, Eq, Default)]
 pub struct RuntimeFeatureConfig {
     hooks: RuntimeHookConfig,
+    plugins: RuntimePluginConfig,
     mcp: McpConfigCollection,
     oauth: Option<OAuthConfig>,
     model: Option<String>,
@@ -174,13 +184,15 @@ impl ConfigLoader {
     #[must_use]
     pub fn default_for(cwd: impl Into<PathBuf>) -> Self {
         let cwd = cwd.into();
-        let config_home = std::env::var_os("CLAUDE_CONFIG_HOME")
-            .map(PathBuf::from)
-            .or_else(|| std::env::var_os("HOME").map(|home| PathBuf::from(home).join(".claude")))
-            .unwrap_or_else(|| PathBuf::from(".claude"));
+        let config_home = default_config_home();
         Self { cwd, config_home }
     }
 
+    #[must_use]
+    pub fn config_home(&self) -> &Path {
+        &self.config_home
+    }
+
     #[must_use]
     pub fn discover(&self) -> Vec<ConfigEntry> {
         let user_legacy_path = self.config_home.parent().map_or_else(
@@ -229,6 +241,7 @@ impl ConfigLoader {
 
         let feature_config = RuntimeFeatureConfig {
             hooks: parse_optional_hooks_config(&merged_value)?,
+            plugins: parse_optional_plugin_config(&merged_value)?,
             mcp: McpConfigCollection {
                 servers: mcp_servers,
             },
@@ -291,6 +304,11 @@ impl RuntimeConfig {
         &self.feature_config.hooks
     }
 
+    #[must_use]
+    pub fn plugins(&self) -> &RuntimePluginConfig {
+        &self.feature_config.plugins
+    }
+
     #[must_use]
     pub fn oauth(&self) -> Option<&OAuthConfig> {
         self.feature_config.oauth.as_ref()
@@ -319,11 +337,22 @@ impl RuntimeFeatureConfig {
         self
     }
 
+    #[must_use]
+    pub fn with_plugins(mut self, plugins: RuntimePluginConfig) -> Self {
+        self.plugins = plugins;
+        self
+    }
+
     #[must_use]
     pub fn hooks(&self) -> &RuntimeHookConfig {
         &self.hooks
     }
 
+    #[must_use]
+    pub fn plugins(&self) -> &RuntimePluginConfig {
+        &self.plugins
+    }
+
     #[must_use]
     pub fn mcp(&self) -> &McpConfigCollection {
         &self.mcp
@@ -350,6 +379,53 @@ impl RuntimeFeatureConfig {
     }
 }
 
+impl RuntimePluginConfig {
+    #[must_use]
+    pub fn enabled_plugins(&self) -> &BTreeMap<String, bool> {
+        &self.enabled_plugins
+    }
+
+    #[must_use]
+    pub fn external_directories(&self) -> &[String] {
+        &self.external_directories
+    }
+
+    #[must_use]
+    pub fn install_root(&self) -> Option<&str> {
+        self.install_root.as_deref()
+    }
+
+    #[must_use]
+    pub fn registry_path(&self) -> Option<&str> {
+        self.registry_path.as_deref()
+    }
+
+    #[must_use]
+    pub fn bundled_root(&self) -> Option<&str> {
+        self.bundled_root.as_deref()
+    }
+
+    pub fn set_plugin_state(&mut self, plugin_id: String, enabled: bool) {
+        self.enabled_plugins.insert(plugin_id, enabled);
+    }
+
+    #[must_use]
+    pub fn state_for(&self, plugin_id: &str, default_enabled: bool) -> bool {
+        self.enabled_plugins
+            .get(plugin_id)
+            .copied()
+            .unwrap_or(default_enabled)
+    }
+}
+
+#[must_use]
+pub fn default_config_home() -> PathBuf {
+    std::env::var_os("CLAUDE_CONFIG_HOME")
+        .map(PathBuf::from)
+        .or_else(|| std::env::var_os("HOME").map(|home| PathBuf::from(home).join(".claude")))
+        .unwrap_or_else(|| PathBuf::from(".claude"))
+}
+
 impl RuntimeHookConfig {
     #[must_use]
     pub fn new(pre_tool_use: Vec<String>, post_tool_use: Vec<String>) -> Self {
@@ -368,6 +444,18 @@ impl RuntimeHookConfig {
     pub fn post_tool_use(&self) -> &[String] {
         &self.post_tool_use
     }
+
+    #[must_use]
+    pub fn merged(&self, other: &Self) -> Self {
+        let mut merged = self.clone();
+        merged.extend(other);
+        merged
+    }
+
+    pub fn extend(&mut self, other: &Self) {
+        extend_unique(&mut self.pre_tool_use, other.pre_tool_use());
+        extend_unique(&mut self.post_tool_use, other.post_tool_use());
+    }
 }
 
 impl McpConfigCollection {
@@ -484,6 +572,36 @@ fn parse_optional_hooks_config(root: &JsonValue) -> Result<RuntimeHookConfig, Co
     })
 }
 
+fn parse_optional_plugin_config(root: &JsonValue) -> Result<RuntimePluginConfig, ConfigError> {
+    let Some(object) = root.as_object() else {
+        return Ok(RuntimePluginConfig::default());
+    };
+
+    let mut config = RuntimePluginConfig::default();
+    if let Some(enabled_plugins) = object.get("enabledPlugins") {
+        config.enabled_plugins = parse_bool_map(enabled_plugins, "merged settings.enabledPlugins")?;
+    }
+
+    let Some(plugins_value) = object.get("plugins") else {
+        return Ok(config);
+    };
+    let plugins = expect_object(plugins_value, "merged settings.plugins")?;
+
+    if let Some(enabled_value) = plugins.get("enabled") {
+        config.enabled_plugins = parse_bool_map(enabled_value, "merged settings.plugins.enabled")?;
+    }
+    config.external_directories =
+        optional_string_array(plugins, "externalDirectories", "merged settings.plugins")?
+            .unwrap_or_default();
+    config.install_root =
+        optional_string(plugins, "installRoot", "merged settings.plugins")?.map(str::to_string);
+    config.registry_path =
+        optional_string(plugins, "registryPath", "merged settings.plugins")?.map(str::to_string);
+    config.bundled_root =
+        optional_string(plugins, "bundledRoot", "merged settings.plugins")?.map(str::to_string);
+    Ok(config)
+}
+
 fn parse_optional_permission_mode(
     root: &JsonValue,
 ) -> Result<Option<ResolvedPermissionMode>, ConfigError> {
@@ -716,6 +834,24 @@ fn optional_u16(
     }
 }
 
+fn parse_bool_map(value: &JsonValue, context: &str) -> Result<BTreeMap<String, bool>, ConfigError> {
+    let Some(map) = value.as_object() else {
+        return Err(ConfigError::Parse(format!(
+            "{context}: expected JSON object"
+        )));
+    };
+    map.iter()
+        .map(|(key, value)| {
+            value
+                .as_bool()
+                .map(|enabled| (key.clone(), enabled))
+                .ok_or_else(|| {
+                    ConfigError::Parse(format!("{context}: field {key} must be a boolean"))
+                })
+        })
+        .collect()
+}
+
 fn optional_string_array(
     object: &BTreeMap<String, JsonValue>,
     key: &str,
@@ -790,6 +926,18 @@ fn deep_merge_objects(
     }
 }
 
+fn extend_unique(target: &mut Vec<String>, values: &[String]) {
+    for value in values {
+        push_unique(target, value.clone());
+    }
+}
+
+fn push_unique(target: &mut Vec<String>, value: String) {
+    if !target.iter().any(|existing| existing == &value) {
+        target.push(value);
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::{
@@ -1033,6 +1181,96 @@ mod tests {
         fs::remove_dir_all(root).expect("cleanup temp dir");
     }
 
+    #[test]
+    fn parses_plugin_config_from_enabled_plugins() {
+        let root = temp_dir();
+        let cwd = root.join("project");
+        let home = root.join("home").join(".claude");
+        fs::create_dir_all(cwd.join(".claude")).expect("project config dir");
+        fs::create_dir_all(&home).expect("home config dir");
+
+        fs::write(
+            home.join("settings.json"),
+            r#"{
+              "enabledPlugins": {
+                "tool-guard@builtin": true,
+                "sample-plugin@external": false
+              }
+            }"#,
+        )
+        .expect("write user settings");
+
+        let loaded = ConfigLoader::new(&cwd, &home)
+            .load()
+            .expect("config should load");
+
+        assert_eq!(
+            loaded.plugins().enabled_plugins().get("tool-guard@builtin"),
+            Some(&true)
+        );
+        assert_eq!(
+            loaded
+                .plugins()
+                .enabled_plugins()
+                .get("sample-plugin@external"),
+            Some(&false)
+        );
+
+        fs::remove_dir_all(root).expect("cleanup temp dir");
+    }
+
+    #[test]
+    fn parses_plugin_config() {
+        let root = temp_dir();
+        let cwd = root.join("project");
+        let home = root.join("home").join(".claude");
+        fs::create_dir_all(cwd.join(".claude")).expect("project config dir");
+        fs::create_dir_all(&home).expect("home config dir");
+
+        fs::write(
+            home.join("settings.json"),
+            r#"{
+              "enabledPlugins": {
+                "core-helpers@builtin": true
+              },
+              "plugins": {
+                "externalDirectories": ["./external-plugins"],
+                "installRoot": "plugin-cache/installed",
+                "registryPath": "plugin-cache/installed.json",
+                "bundledRoot": "./bundled-plugins"
+              }
+            }"#,
+        )
+        .expect("write plugin settings");
+
+        let loaded = ConfigLoader::new(&cwd, &home)
+            .load()
+            .expect("config should load");
+
+        assert_eq!(
+            loaded
+                .plugins()
+                .enabled_plugins()
+                .get("core-helpers@builtin"),
+            Some(&true)
+        );
+        assert_eq!(
+            loaded.plugins().external_directories(),
+            &["./external-plugins".to_string()]
+        );
+        assert_eq!(
+            loaded.plugins().install_root(),
+            Some("plugin-cache/installed")
+        );
+        assert_eq!(
+            loaded.plugins().registry_path(),
+            Some("plugin-cache/installed.json")
+        );
+        assert_eq!(loaded.plugins().bundled_root(), Some("./bundled-plugins"));
+
+        fs::remove_dir_all(root).expect("cleanup temp dir");
+    }
+
     #[test]
     fn rejects_invalid_mcp_server_shapes() {
         let root = temp_dir();

rust/crates/runtime/src/conversation.rs

@@ -1,15 +1,20 @@
 use std::collections::BTreeMap;
 use std::fmt::{Display, Formatter};
 
+use plugins::{HookRunner as PluginHookRunner, PluginRegistry};
+
 use crate::compact::{
     compact_session, estimate_session_tokens, CompactionConfig, CompactionResult,
 };
 use crate::config::RuntimeFeatureConfig;
-use crate::hooks::{HookRunResult, HookRunner};
+use crate::hooks::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 = 200_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>,
@@ -25,19 +30,9 @@ 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>;
 }
@@ -94,9 +89,14 @@ 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> {
@@ -108,6 +108,25 @@ pub struct ConversationRuntime<C, T> {
     max_iterations: usize,
     usage_tracker: UsageTracker,
     hook_runner: HookRunner,
+    auto_compaction_input_tokens_threshold: u32,
+    plugin_hook_runner: Option<PluginHookRunner>,
+    plugin_registry: Option<PluginRegistry>,
+    plugins_shutdown: bool,
+}
+
+impl<C, T> ConversationRuntime<C, T> {
+    fn shutdown_registered_plugins(&mut self) -> Result<(), RuntimeError> {
+        if self.plugins_shutdown {
+            return Ok(());
+        }
+        if let Some(registry) = &self.plugin_registry {
+            registry
+                .shutdown()
+                .map_err(|error| RuntimeError::new(format!("plugin shutdown failed: {error}")))?;
+        }
+        self.plugins_shutdown = true;
+        Ok(())
+    }
 }
 
 impl<C, T> ConversationRuntime<C, T>
@@ -129,18 +148,19 @@ where
             tool_executor,
             permission_policy,
             system_prompt,
-            &RuntimeFeatureConfig::default(),
+            RuntimeFeatureConfig::default(),
         )
     }
 
     #[must_use]
+    #[allow(clippy::needless_pass_by_value)]
     pub fn new_with_features(
         session: Session,
         api_client: C,
         tool_executor: T,
         permission_policy: PermissionPolicy,
         system_prompt: Vec<String>,
-        feature_config: &RuntimeFeatureConfig,
+        feature_config: RuntimeFeatureConfig,
     ) -> Self {
         let usage_tracker = UsageTracker::from_session(&session);
         Self {
@@ -151,16 +171,57 @@ where
             system_prompt,
             max_iterations: usize::MAX,
             usage_tracker,
-            hook_runner: HookRunner::from_feature_config(feature_config),
+            hook_runner: HookRunner::from_feature_config(&feature_config),
+            auto_compaction_input_tokens_threshold: auto_compaction_threshold_from_env(),
+            plugin_hook_runner: None,
+            plugin_registry: None,
+            plugins_shutdown: false,
         }
     }
 
+    #[allow(clippy::needless_pass_by_value)]
+    pub fn new_with_plugins(
+        session: Session,
+        api_client: C,
+        tool_executor: T,
+        permission_policy: PermissionPolicy,
+        system_prompt: Vec<String>,
+        feature_config: RuntimeFeatureConfig,
+        plugin_registry: PluginRegistry,
+    ) -> Result<Self, RuntimeError> {
+        let plugin_hook_runner =
+            PluginHookRunner::from_registry(&plugin_registry).map_err(|error| {
+                RuntimeError::new(format!("plugin hook registration failed: {error}"))
+            })?;
+        plugin_registry
+            .initialize()
+            .map_err(|error| RuntimeError::new(format!("plugin initialization failed: {error}")))?;
+        let mut runtime = Self::new_with_features(
+            session,
+            api_client,
+            tool_executor,
+            permission_policy,
+            system_prompt,
+            feature_config,
+        );
+        runtime.plugin_hook_runner = Some(plugin_hook_runner);
+        runtime.plugin_registry = Some(plugin_registry);
+        Ok(runtime)
+    }
+
     #[must_use]
     pub fn with_max_iterations(mut self, max_iterations: usize) -> Self {
         self.max_iterations = max_iterations;
         self
     }
 
+    #[must_use]
+    pub fn with_auto_compaction_input_tokens_threshold(mut self, threshold: u32) -> Self {
+        self.auto_compaction_input_tokens_threshold = threshold;
+        self
+    }
+
+    #[allow(clippy::too_many_lines)]
     pub fn run_turn(
         &mut self,
         user_input: impl Into<String>,
@@ -172,8 +233,8 @@ where
 
         let mut assistant_messages = Vec::new();
         let mut tool_results = Vec::new();
-        let mut prompt_cache_events = Vec::new();
         let mut iterations = 0;
+        let mut max_turn_input_tokens = 0;
 
         loop {
             iterations += 1;
@@ -188,12 +249,11 @@ where
                 messages: self.session.messages.clone(),
             };
             let events = self.api_client.stream(request)?;
-            let (assistant_message, usage, turn_prompt_cache_events) =
-                build_assistant_message(events)?;
+            let (assistant_message, usage) = build_assistant_message(events)?;
             if let Some(usage) = usage {
+                max_turn_input_tokens = max_turn_input_tokens.max(usage.input_tokens);
                 self.usage_tracker.record(usage);
             }
-            prompt_cache_events.extend(turn_prompt_cache_events);
             let pending_tool_uses = assistant_message
                 .blocks
                 .iter()
@@ -228,7 +288,21 @@ where
                             ConversationMessage::tool_result(
                                 tool_use_id,
                                 tool_name,
-                                format_hook_message(&pre_hook_result, &deny_message),
+                                format_hook_message(pre_hook_result.messages(), &deny_message),
+                                true,
+                            )
+                        } else {
+                            let plugin_pre_hook_result =
+                                self.run_plugin_pre_tool_use(&tool_name, &input);
+                            if plugin_pre_hook_result.is_denied() {
+                                let deny_message =
+                                    format!("PreToolUse hook denied tool `{tool_name}`");
+                                let mut messages = pre_hook_result.messages().to_vec();
+                                messages.extend(plugin_pre_hook_result.messages().iter().cloned());
+                                ConversationMessage::tool_result(
+                                    tool_use_id,
+                                    tool_name,
+                                    format_hook_message(&messages, &deny_message),
                                     true,
                                 )
                             } else {
@@ -237,19 +311,43 @@ where
                                         Ok(output) => (output, false),
                                         Err(error) => (error.to_string(), true),
                                     };
-                            output = merge_hook_feedback(pre_hook_result.messages(), output, false);
+                                output =
+                                    merge_hook_feedback(pre_hook_result.messages(), output, false);
+                                output = merge_hook_feedback(
+                                    plugin_pre_hook_result.messages(),
+                                    output,
+                                    false,
+                                );
 
-                            let post_hook_result = self
-                                .hook_runner
-                                .run_post_tool_use(&tool_name, &input, &output, is_error);
+                                let hook_output = output.clone();
+                                let post_hook_result = self.hook_runner.run_post_tool_use(
+                                    &tool_name,
+                                    &input,
+                                    &hook_output,
+                                    is_error,
+                                );
+                                let plugin_post_hook_result = self.run_plugin_post_tool_use(
+                                    &tool_name,
+                                    &input,
+                                    &hook_output,
+                                    is_error,
+                                );
                                 if post_hook_result.is_denied() {
                                     is_error = true;
                                 }
+                                if plugin_post_hook_result.is_denied() {
+                                    is_error = true;
+                                }
                                 output = merge_hook_feedback(
                                     post_hook_result.messages(),
                                     output,
                                     post_hook_result.is_denied(),
                                 );
+                                output = merge_hook_feedback(
+                                    plugin_post_hook_result.messages(),
+                                    output,
+                                    plugin_post_hook_result.is_denied(),
+                                );
 
                                 ConversationMessage::tool_result(
                                     tool_use_id,
@@ -259,6 +357,7 @@ where
                                 )
                             }
                         }
+                    }
                     PermissionOutcome::Deny { reason } => {
                         ConversationMessage::tool_result(tool_use_id, tool_name, reason, true)
                     }
@@ -268,12 +367,14 @@ where
             }
         }
 
+        let auto_compaction = self.maybe_auto_compact(max_turn_input_tokens);
+
         Ok(TurnSummary {
             assistant_messages,
             tool_results,
-            prompt_cache_events,
             iterations,
             usage: self.usage_tracker.cumulative_usage(),
+            auto_compaction,
         })
     }
 
@@ -298,24 +399,88 @@ where
     }
 
     #[must_use]
-    pub fn into_session(self) -> Session {
-        self.session
+    pub fn into_session(mut self) -> Session {
+        let _ = self.shutdown_registered_plugins();
+        std::mem::take(&mut self.session)
     }
+
+    pub fn shutdown_plugins(&mut self) -> Result<(), RuntimeError> {
+        self.shutdown_registered_plugins()
+    }
+
+    fn run_plugin_pre_tool_use(&self, tool_name: &str, input: &str) -> plugins::HookRunResult {
+        self.plugin_hook_runner.as_ref().map_or_else(
+            || plugins::HookRunResult::allow(Vec::new()),
+            |runner| runner.run_pre_tool_use(tool_name, input),
+        )
+    }
+
+    fn run_plugin_post_tool_use(
+        &self,
+        tool_name: &str,
+        input: &str,
+        output: &str,
+        is_error: bool,
+    ) -> plugins::HookRunResult {
+        self.plugin_hook_runner.as_ref().map_or_else(
+            || plugins::HookRunResult::allow(Vec::new()),
+            |runner| runner.run_post_tool_use(tool_name, input, output, is_error),
+        )
+    }
+
+    fn maybe_auto_compact(&mut self, turn_input_tokens: u32) -> Option<AutoCompactionEvent> {
+        if turn_input_tokens < self.auto_compaction_input_tokens_threshold {
+            return None;
+        }
+
+        let result = compact_session(
+            &self.session,
+            CompactionConfig {
+                max_estimated_tokens: usize::try_from(self.auto_compaction_input_tokens_threshold)
+                    .unwrap_or(usize::MAX),
+                ..CompactionConfig::default()
+            },
+        );
+
+        if result.removed_message_count == 0 {
+            return None;
+        }
+
+        self.session = result.compacted_session;
+        Some(AutoCompactionEvent {
+            removed_message_count: result.removed_message_count,
+        })
+    }
+}
+
+impl<C, T> Drop for ConversationRuntime<C, T> {
+    fn drop(&mut self) {
+        let _ = self.shutdown_registered_plugins();
+    }
+}
+
+#[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>,
-        Vec<PromptCacheEvent>,
-    ),
-    RuntimeError,
-> {
+) -> Result<(ConversationMessage, Option<TokenUsage>), 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;
 
@@ -327,7 +492,6 @@ 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;
             }
@@ -348,7 +512,6 @@ fn build_assistant_message(
     Ok((
         ConversationMessage::assistant_with_usage(blocks, usage),
         usage,
-        prompt_cache_events,
     ))
 }
 
@@ -360,11 +523,11 @@ 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() {
+fn format_hook_message(messages: &[String], fallback: &str) -> String {
+    if messages.is_empty() {
         fallback.to_string()
     } else {
-        result.messages().join("\n")
+        messages.join("\n")
     }
 }
 
@@ -421,8 +584,9 @@ impl ToolExecutor for StaticToolExecutor {
 #[cfg(test)]
 mod tests {
     use super::{
-        ApiClient, ApiRequest, AssistantEvent, ConversationRuntime, PromptCacheEvent, RuntimeError,
-        StaticToolExecutor,
+        parse_auto_compaction_threshold, ApiClient, ApiRequest, AssistantEvent,
+        AutoCompactionEvent, ConversationRuntime, RuntimeError, StaticToolExecutor,
+        DEFAULT_AUTO_COMPACTION_INPUT_TOKENS_THRESHOLD,
     };
     use crate::compact::CompactionConfig;
     use crate::config::{RuntimeFeatureConfig, RuntimeHookConfig};
@@ -433,7 +597,13 @@ mod tests {
     use crate::prompt::{ProjectContext, SystemPromptBuilder};
     use crate::session::{ContentBlock, MessageRole, Session};
     use crate::usage::TokenUsage;
+    use plugins::{PluginManager, PluginManagerConfig};
+    use std::fs;
+    #[cfg(unix)]
+    use std::os::unix::fs::PermissionsExt;
+    use std::path::Path;
     use std::path::PathBuf;
+    use std::time::{SystemTime, UNIX_EPOCH};
 
     struct ScriptedApiClient {
         call_count: usize,
@@ -478,15 +648,6 @@ 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,
                     ])
                 }
@@ -504,6 +665,68 @@ mod tests {
         }
     }
 
+    fn temp_dir(label: &str) -> PathBuf {
+        let nanos = SystemTime::now()
+            .duration_since(UNIX_EPOCH)
+            .expect("time should be after epoch")
+            .as_nanos();
+        std::env::temp_dir().join(format!("runtime-plugin-{label}-{nanos}"))
+    }
+
+    fn write_lifecycle_plugin(root: &Path, name: &str) -> PathBuf {
+        fs::create_dir_all(root.join(".claude-plugin")).expect("manifest dir");
+        fs::create_dir_all(root.join("lifecycle")).expect("lifecycle dir");
+        let log_path = root.join("lifecycle.log");
+        fs::write(
+            root.join("lifecycle").join("init.sh"),
+            "#!/bin/sh\nprintf 'init\\n' >> lifecycle.log\n",
+        )
+        .expect("write init script");
+        fs::write(
+            root.join("lifecycle").join("shutdown.sh"),
+            "#!/bin/sh\nprintf 'shutdown\\n' >> lifecycle.log\n",
+        )
+        .expect("write shutdown script");
+        fs::write(
+            root.join(".claude-plugin").join("plugin.json"),
+            format!(
+                "{{\n  \"name\": \"{name}\",\n  \"version\": \"1.0.0\",\n  \"description\": \"runtime lifecycle plugin\",\n  \"lifecycle\": {{\n    \"Init\": [\"./lifecycle/init.sh\"],\n    \"Shutdown\": [\"./lifecycle/shutdown.sh\"]\n  }}\n}}"
+            ),
+        )
+        .expect("write plugin manifest");
+        log_path
+    }
+
+    fn write_hook_plugin(root: &Path, name: &str, pre_message: &str, post_message: &str) {
+        fs::create_dir_all(root.join(".claude-plugin")).expect("manifest dir");
+        fs::create_dir_all(root.join("hooks")).expect("hooks dir");
+        fs::write(
+            root.join("hooks").join("pre.sh"),
+            format!("#!/bin/sh\nprintf '%s\\n' '{pre_message}'\n"),
+        )
+        .expect("write pre hook");
+        fs::write(
+            root.join("hooks").join("post.sh"),
+            format!("#!/bin/sh\nprintf '%s\\n' '{post_message}'\n"),
+        )
+        .expect("write post hook");
+        #[cfg(unix)]
+        {
+            let exec_mode = fs::Permissions::from_mode(0o755);
+            fs::set_permissions(root.join("hooks").join("pre.sh"), exec_mode.clone())
+                .expect("chmod pre hook");
+            fs::set_permissions(root.join("hooks").join("post.sh"), exec_mode)
+                .expect("chmod post hook");
+        }
+        fs::write(
+            root.join(".claude-plugin").join("plugin.json"),
+            format!(
+                "{{\n  \"name\": \"{name}\",\n  \"version\": \"1.0.0\",\n  \"description\": \"runtime hook plugin\",\n  \"hooks\": {{\n    \"PreToolUse\": [\"./hooks/pre.sh\"],\n    \"PostToolUse\": [\"./hooks/post.sh\"]\n  }}\n}}"
+            ),
+        )
+        .expect("write plugin manifest");
+    }
+
     #[test]
     fn runs_user_to_tool_to_result_loop_end_to_end_and_tracks_usage() {
         let api_client = ScriptedApiClient { call_count: 0 };
@@ -540,10 +763,9 @@ 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!(summary.prompt_cache_events[0].unexpected);
+        assert_eq!(summary.auto_compaction, None);
         assert!(matches!(
             runtime.session().messages[1].blocks[1],
             ContentBlock::ToolUse { .. }
@@ -645,7 +867,7 @@ mod tests {
             }),
             PermissionPolicy::new(PermissionMode::DangerFullAccess),
             vec!["system".to_string()],
-            &RuntimeFeatureConfig::default().with_hooks(RuntimeHookConfig::new(
+            RuntimeFeatureConfig::default().with_hooks(RuntimeHookConfig::new(
                 vec![shell_snippet("printf 'blocked by hook'; exit 2")],
                 Vec::new(),
             )),
@@ -711,7 +933,7 @@ mod tests {
             StaticToolExecutor::new().register("add", |_input| Ok("4".to_string())),
             PermissionPolicy::new(PermissionMode::DangerFullAccess),
             vec!["system".to_string()],
-            &RuntimeFeatureConfig::default().with_hooks(RuntimeHookConfig::new(
+            RuntimeFeatureConfig::default().with_hooks(RuntimeHookConfig::new(
                 vec![shell_snippet("printf 'pre hook ran'")],
                 vec![shell_snippet("printf 'post hook ran'")],
             )),
@@ -746,6 +968,153 @@ mod tests {
         );
     }
 
+    #[test]
+    fn initializes_and_shuts_down_plugins_with_runtime_lifecycle() {
+        let config_home = temp_dir("config");
+        let source_root = temp_dir("source");
+        let _ = write_lifecycle_plugin(&source_root, "runtime-lifecycle");
+
+        let mut manager = PluginManager::new(PluginManagerConfig::new(&config_home));
+        let install = manager
+            .install(source_root.to_str().expect("utf8 path"))
+            .expect("install should succeed");
+        let log_path = install.install_path.join("lifecycle.log");
+        let registry = manager.plugin_registry().expect("registry should load");
+
+        {
+            let runtime = ConversationRuntime::new_with_plugins(
+                Session::new(),
+                ScriptedApiClient { call_count: 0 },
+                StaticToolExecutor::new().register("add", |_input| Ok("4".to_string())),
+                PermissionPolicy::new(PermissionMode::WorkspaceWrite),
+                vec!["system".to_string()],
+                RuntimeFeatureConfig::default(),
+                registry,
+            )
+            .expect("runtime should initialize plugins");
+
+            let log = fs::read_to_string(&log_path).expect("init log should exist");
+            assert_eq!(log, "init\n");
+            drop(runtime);
+        }
+
+        let log = fs::read_to_string(&log_path).expect("shutdown log should exist");
+        assert_eq!(log, "init\nshutdown\n");
+
+        let _ = fs::remove_dir_all(config_home);
+        let _ = fs::remove_dir_all(source_root);
+    }
+
+    #[test]
+    fn executes_hooks_from_installed_plugins_during_tool_use() {
+        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 config_home = temp_dir("hook-config");
+        let first_source_root = temp_dir("hook-source-a");
+        let second_source_root = temp_dir("hook-source-b");
+        write_hook_plugin(
+            &first_source_root,
+            "first",
+            "plugin pre one",
+            "plugin post one",
+        );
+        write_hook_plugin(
+            &second_source_root,
+            "second",
+            "plugin pre two",
+            "plugin post two",
+        );
+
+        let mut manager = PluginManager::new(PluginManagerConfig::new(&config_home));
+        manager
+            .install(first_source_root.to_str().expect("utf8 path"))
+            .expect("first plugin install should succeed");
+        manager
+            .install(second_source_root.to_str().expect("utf8 path"))
+            .expect("second plugin install should succeed");
+        let registry = manager.plugin_registry().expect("registry should load");
+
+        let mut runtime = ConversationRuntime::new_with_plugins(
+            Session::new(),
+            TwoCallApiClient { calls: 0 },
+            StaticToolExecutor::new().register("add", |_input| Ok("4".to_string())),
+            PermissionPolicy::new(PermissionMode::DangerFullAccess),
+            vec!["system".to_string()],
+            RuntimeFeatureConfig::default(),
+            registry,
+        )
+        .expect("runtime should load plugin hooks");
+
+        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,
+            "plugin hooks should not force an error: {output:?}"
+        );
+        assert!(
+            output.contains('4'),
+            "tool output missing value: {output:?}"
+        );
+        assert!(
+            output.contains("plugin pre one"),
+            "tool output missing first pre hook feedback: {output:?}"
+        );
+        assert!(
+            output.contains("plugin pre two"),
+            "tool output missing second pre hook feedback: {output:?}"
+        );
+        assert!(
+            output.contains("plugin post one"),
+            "tool output missing first post hook feedback: {output:?}"
+        );
+        assert!(
+            output.contains("plugin post two"),
+            "tool output missing second post hook feedback: {output:?}"
+        );
+
+        let _ = fs::remove_dir_all(config_home);
+        let _ = fs::remove_dir_all(first_source_root);
+        let _ = fs::remove_dir_all(second_source_root);
+    }
+
     #[test]
     fn reconstructs_usage_tracker_from_restored_session() {
         struct SimpleApi;
@@ -834,4 +1203,177 @@ mod tests {
     fn shell_snippet(script: &str) -> String {
         script.to_string()
     }
+
+    #[test]
+    fn auto_compacts_when_turn_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,
+                ])
+            }
+        }
+
+        let session = Session {
+            version: 1,
+            messages: vec![
+                crate::session::ConversationMessage::user_text("one ".repeat(30_000)),
+                crate::session::ConversationMessage::assistant(vec![ContentBlock::Text {
+                    text: "two ".repeat(30_000),
+                }]),
+                crate::session::ConversationMessage::user_text("three ".repeat(30_000)),
+                crate::session::ConversationMessage::assistant(vec![ContentBlock::Text {
+                    text: "four ".repeat(30_000),
+                }]),
+            ],
+        };
+
+        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 auto_compaction_does_not_repeat_after_context_is_already_compacted() {
+        struct SequentialUsageApi {
+            call_count: usize,
+        }
+
+        impl ApiClient for SequentialUsageApi {
+            fn stream(
+                &mut self,
+                _request: ApiRequest,
+            ) -> Result<Vec<AssistantEvent>, RuntimeError> {
+                self.call_count += 1;
+                let input_tokens = if self.call_count == 1 { 120_000 } else { 64 };
+                Ok(vec![
+                    AssistantEvent::TextDelta("done".to_string()),
+                    AssistantEvent::Usage(TokenUsage {
+                        input_tokens,
+                        output_tokens: 4,
+                        cache_creation_input_tokens: 0,
+                        cache_read_input_tokens: 0,
+                    }),
+                    AssistantEvent::MessageStop,
+                ])
+            }
+        }
+
+        let session = Session {
+            version: 1,
+            messages: vec![
+                crate::session::ConversationMessage::user_text("one ".repeat(30_000)),
+                crate::session::ConversationMessage::assistant(vec![ContentBlock::Text {
+                    text: "two ".repeat(30_000),
+                }]),
+                crate::session::ConversationMessage::user_text("three ".repeat(30_000)),
+                crate::session::ConversationMessage::assistant(vec![ContentBlock::Text {
+                    text: "four ".repeat(30_000),
+                }]),
+            ],
+        };
+
+        let mut runtime = ConversationRuntime::new(
+            session,
+            SequentialUsageApi { call_count: 0 },
+            StaticToolExecutor::new(),
+            PermissionPolicy::new(PermissionMode::DangerFullAccess),
+            vec!["system".to_string()],
+        )
+        .with_auto_compaction_input_tokens_threshold(100_000);
+
+        let first = runtime
+            .run_turn("trigger", None)
+            .expect("first turn should succeed");
+        assert_eq!(
+            first.auto_compaction,
+            Some(AutoCompactionEvent {
+                removed_message_count: 2,
+            })
+        );
+
+        let second = runtime
+            .run_turn("continue", None)
+            .expect("second turn should succeed");
+        assert_eq!(second.auto_compaction, None);
+        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
+        );
+    }
 }

rust/crates/runtime/src/hooks.rs

@@ -1,4 +1,5 @@
 use std::ffi::OsStr;
+use std::path::Path;
 use std::process::Command;
 
 use serde_json::json;
@@ -64,7 +65,7 @@ impl HookRunner {
 
     #[must_use]
     pub fn run_pre_tool_use(&self, tool_name: &str, tool_input: &str) -> HookRunResult {
-        Self::run_commands(
+        self.run_commands(
             HookEvent::PreToolUse,
             self.config.pre_tool_use(),
             tool_name,
@@ -82,7 +83,7 @@ impl HookRunner {
         tool_output: &str,
         is_error: bool,
     ) -> HookRunResult {
-        Self::run_commands(
+        self.run_commands(
             HookEvent::PostToolUse,
             self.config.post_tool_use(),
             tool_name,
@@ -93,6 +94,7 @@ impl HookRunner {
     }
 
     fn run_commands(
+        &self,
         event: HookEvent,
         commands: &[String],
         tool_name: &str,
@@ -117,7 +119,7 @@ impl HookRunner {
         let mut messages = Vec::new();
 
         for command in commands {
-            match Self::run_command(
+            match self.run_command(
                 command,
                 event,
                 tool_name,
@@ -148,7 +150,9 @@ impl HookRunner {
         HookRunResult::allow(messages)
     }
 
+    #[allow(clippy::too_many_arguments, clippy::unused_self)]
     fn run_command(
+        &self,
         command: &str,
         event: HookEvent,
         tool_name: &str,
@@ -235,7 +239,11 @@ fn shell_command(command: &str) -> CommandWithStdin {
     };
 
     #[cfg(not(windows))]
-    let command_builder = {
+    let command_builder = if Path::new(command).exists() {
+        let mut command_builder = Command::new("sh");
+        command_builder.arg(command);
+        CommandWithStdin::new(command_builder)
+    } else {
         let mut command_builder = Command::new("sh");
         command_builder.arg("-lc").arg(command);
         CommandWithStdin::new(command_builder)

rust/crates/runtime/src/lib.rs

@@ -28,11 +28,11 @@ pub use config::{
     McpConfigCollection, McpOAuthConfig, McpRemoteServerConfig, McpSdkServerConfig,
     McpServerConfig, McpStdioServerConfig, McpTransport, McpWebSocketServerConfig, OAuthConfig,
     ResolvedPermissionMode, RuntimeConfig, RuntimeFeatureConfig, RuntimeHookConfig,
-    ScopedMcpServerConfig, CLAUDE_CODE_SETTINGS_SCHEMA_NAME,
+    RuntimePluginConfig, ScopedMcpServerConfig, CLAUDE_CODE_SETTINGS_SCHEMA_NAME,
 };
 pub use conversation::{
-    ApiClient, ApiRequest, AssistantEvent, ConversationRuntime, PromptCacheEvent, RuntimeError,
-    StaticToolExecutor, ToolError, ToolExecutor, TurnSummary,
+    auto_compaction_threshold_from_env, ApiClient, ApiRequest, AssistantEvent, AutoCompactionEvent,
+    ConversationRuntime, RuntimeError, StaticToolExecutor, ToolError, ToolExecutor, TurnSummary,
 };
 pub use file_ops::{
     edit_file, glob_search, grep_search, read_file, write_file, EditFileOutput, GlobSearchOutput,

rust/crates/runtime/src/prompt.rs

@@ -421,7 +421,7 @@ fn render_config_section(config: &RuntimeConfig) -> String {
     let mut lines = vec!["# Runtime config".to_string()];
     if config.loaded_entries().is_empty() {
         lines.extend(prepend_bullets(vec![
-            "No Claude Code settings files loaded.".to_string(),
+            "No Claw Code settings files loaded.".to_string()
         ]));
         return lines.join("\n");
     }

rust/crates/rusty-claude-cli/Cargo.toml

@@ -17,6 +17,7 @@ crossterm = "0.28"
 pulldown-cmark = "0.13"
 rustyline = "15"
 runtime = { path = "../runtime" }
+plugins = { path = "../plugins" }
 serde_json = "1"
 syntect = "5"
 tokio = { version = "1", features = ["rt-multi-thread", "time"] }

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

@@ -8,7 +8,7 @@ const STARTER_CLAUDE_JSON: &str = concat!(
     "  }\n",
     "}\n",
 );
-const GITIGNORE_COMMENT: &str = "# Claude Code local artifacts";
+const GITIGNORE_COMMENT: &str = "# Claw Code local artifacts";
 const GITIGNORE_ENTRIES: [&str; 2] = [".claude/settings.local.json", ".claude/sessions/"];
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
@@ -164,7 +164,7 @@ pub(crate) fn render_init_claude_md(cwd: &Path) -> String {
     let mut lines = vec![
         "# CLAUDE.md".to_string(),
         String::new(),
-        "This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.".to_string(),
+        "This file provides guidance to Claw Code (clawcode.dev) when working with code in this repository.".to_string(),
         String::new(),
     ];
 

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 +426,8 @@ impl TerminalRenderer {
         }
     }
 
-    fn start_heading(state: &mut RenderState, level: u8, output: &mut String) {
+    #[allow(clippy::unused_self)]
+    fn start_heading(&self, state: &mut RenderState, level: u8, output: &mut String) {
         state.heading_level = Some(level);
         if !output.is_empty() {
             output.push('\n');

rust/crates/tools/Cargo.toml

@@ -7,6 +7,7 @@ publish.workspace = true
 
 [dependencies]
 api = { path = "../api" }
+plugins = { path = "../plugins" }
 runtime = { path = "../runtime" }
 reqwest = { version = "0.12", default-features = false, features = ["blocking", "rustls-tls"] }
 serde = { version = "1", features = ["derive"] }

rust/crates/tools/src/lib.rs

@@ -5,15 +5,16 @@ use std::time::{Duration, Instant};
 
 use api::{
     read_base_url, AnthropicClient, ContentBlockDelta, InputContentBlock, InputMessage,
-    MessageRequest, MessageResponse, OutputContentBlock, PromptCache, PromptCacheRecord,
-    StreamEvent as ApiStreamEvent, ToolChoice, ToolDefinition, ToolResultContentBlock,
+    MessageRequest, MessageResponse, OutputContentBlock, StreamEvent as ApiStreamEvent, ToolChoice,
+    ToolDefinition, ToolResultContentBlock,
 };
+use plugins::{PluginManager, PluginManagerConfig, PluginTool};
 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,
-    PromptCacheEvent, RuntimeError, Session, TokenUsage, ToolError, ToolExecutor,
+    ApiClient, ApiRequest, AssistantEvent, BashCommandInput, ConfigLoader, ContentBlock,
+    ConversationMessage, ConversationRuntime, GrepSearchInput, MessageRole, PermissionMode,
+    PermissionPolicy, RuntimeConfig, RuntimeError, Session, TokenUsage, ToolError, ToolExecutor,
 };
 use serde::{Deserialize, Serialize};
 use serde_json::{json, Value};
@@ -55,6 +56,239 @@ pub struct ToolSpec {
     pub required_permission: PermissionMode,
 }
 
+#[derive(Debug, Clone, PartialEq)]
+pub struct RegisteredTool {
+    pub definition: ToolDefinition,
+    pub required_permission: PermissionMode,
+    handler: RegisteredToolHandler,
+}
+
+#[allow(clippy::large_enum_variant)]
+#[derive(Debug, Clone, PartialEq)]
+enum RegisteredToolHandler {
+    Builtin,
+    Plugin(PluginTool),
+}
+
+#[derive(Debug, Clone, PartialEq)]
+pub struct GlobalToolRegistry {
+    entries: Vec<RegisteredTool>,
+}
+
+impl GlobalToolRegistry {
+    #[must_use]
+    pub fn builtin() -> Self {
+        Self {
+            entries: mvp_tool_specs()
+                .into_iter()
+                .map(|spec| RegisteredTool {
+                    definition: ToolDefinition {
+                        name: spec.name.to_string(),
+                        description: Some(spec.description.to_string()),
+                        input_schema: spec.input_schema,
+                    },
+                    required_permission: spec.required_permission,
+                    handler: RegisteredToolHandler::Builtin,
+                })
+                .collect(),
+        }
+    }
+
+    pub fn with_plugin_tools(plugin_tools: Vec<PluginTool>) -> Result<Self, String> {
+        let mut registry = Self::builtin();
+        let mut seen = registry
+            .entries
+            .iter()
+            .map(|entry| {
+                (
+                    normalize_registry_tool_name(&entry.definition.name),
+                    entry.definition.name.clone(),
+                )
+            })
+            .collect::<BTreeMap<_, _>>();
+
+        for tool in plugin_tools {
+            let normalized = normalize_registry_tool_name(&tool.definition().name);
+            if let Some(existing) = seen.get(&normalized) {
+                return Err(format!(
+                    "plugin tool `{}` from `{}` conflicts with already-registered tool `{existing}`",
+                    tool.definition().name,
+                    tool.plugin_id()
+                ));
+            }
+            seen.insert(normalized, tool.definition().name.clone());
+            registry.entries.push(RegisteredTool {
+                definition: ToolDefinition {
+                    name: tool.definition().name.clone(),
+                    description: tool.definition().description.clone(),
+                    input_schema: tool.definition().input_schema.clone(),
+                },
+                required_permission: permission_mode_from_plugin_tool(tool.required_permission())?,
+                handler: RegisteredToolHandler::Plugin(tool),
+            });
+        }
+
+        Ok(registry)
+    }
+
+    #[must_use]
+    pub fn entries(&self) -> &[RegisteredTool] {
+        &self.entries
+    }
+
+    fn find_entry(&self, name: &str) -> Option<&RegisteredTool> {
+        let normalized = normalize_registry_tool_name(name);
+        self.entries.iter().find(|entry| {
+            normalize_registry_tool_name(entry.definition.name.as_str()) == normalized
+        })
+    }
+
+    #[must_use]
+    pub fn definitions(&self, allowed_tools: Option<&BTreeSet<String>>) -> Vec<ToolDefinition> {
+        self.entries
+            .iter()
+            .filter(|entry| {
+                allowed_tools.is_none_or(|allowed| allowed.contains(entry.definition.name.as_str()))
+            })
+            .map(|entry| entry.definition.clone())
+            .collect()
+    }
+
+    #[must_use]
+    pub fn permission_specs(
+        &self,
+        allowed_tools: Option<&BTreeSet<String>>,
+    ) -> Vec<(String, PermissionMode)> {
+        self.entries
+            .iter()
+            .filter(|entry| {
+                allowed_tools.is_none_or(|allowed| allowed.contains(entry.definition.name.as_str()))
+            })
+            .map(|entry| (entry.definition.name.clone(), entry.required_permission))
+            .collect()
+    }
+
+    pub fn normalize_allowed_tools(
+        &self,
+        values: &[String],
+    ) -> Result<Option<BTreeSet<String>>, String> {
+        if values.is_empty() {
+            return Ok(None);
+        }
+
+        let canonical_names = self
+            .entries
+            .iter()
+            .map(|entry| entry.definition.name.clone())
+            .collect::<Vec<_>>();
+        let mut name_map = canonical_names
+            .iter()
+            .map(|name| (normalize_registry_tool_name(name), name.clone()))
+            .collect::<BTreeMap<_, _>>();
+
+        for (alias, canonical) in [
+            ("read", "read_file"),
+            ("write", "write_file"),
+            ("edit", "edit_file"),
+            ("glob", "glob_search"),
+            ("grep", "grep_search"),
+        ] {
+            if canonical_names.iter().any(|name| name == canonical) {
+                name_map.insert(alias.to_string(), canonical.to_string());
+            }
+        }
+
+        let mut allowed = BTreeSet::new();
+        for value in values {
+            for token in value
+                .split(|ch: char| ch == ',' || ch.is_whitespace())
+                .filter(|token| !token.is_empty())
+            {
+                let normalized = normalize_registry_tool_name(token);
+                let canonical = name_map.get(&normalized).ok_or_else(|| {
+                    format!(
+                        "unsupported tool in --allowedTools: {token} (expected one of: {})",
+                        canonical_names.join(", ")
+                    )
+                })?;
+                allowed.insert(canonical.clone());
+            }
+        }
+
+        Ok(Some(allowed))
+    }
+
+    pub fn execute(&self, name: &str, input: &Value) -> Result<String, String> {
+        let entry = self
+            .find_entry(name)
+            .ok_or_else(|| format!("unsupported tool: {name}"))?;
+        match &entry.handler {
+            RegisteredToolHandler::Builtin => execute_tool(&entry.definition.name, input),
+            RegisteredToolHandler::Plugin(tool) => {
+                tool.execute(input).map_err(|error| error.to_string())
+            }
+        }
+    }
+}
+
+impl Default for GlobalToolRegistry {
+    fn default() -> Self {
+        Self::builtin()
+    }
+}
+
+fn normalize_registry_tool_name(value: &str) -> String {
+    let trimmed = value.trim();
+    let chars = trimmed.chars().collect::<Vec<_>>();
+    let mut normalized = String::new();
+
+    for (index, ch) in chars.iter().copied().enumerate() {
+        if matches!(ch, '-' | ' ' | '\t' | '\n') {
+            if !normalized.ends_with('_') {
+                normalized.push('_');
+            }
+            continue;
+        }
+
+        if ch == '_' {
+            if !normalized.ends_with('_') {
+                normalized.push('_');
+            }
+            continue;
+        }
+
+        if ch.is_uppercase() {
+            let prev = chars.get(index.wrapping_sub(1)).copied();
+            let next = chars.get(index + 1).copied();
+            let needs_separator = index > 0
+                && !normalized.ends_with('_')
+                && (prev.is_some_and(|prev| prev.is_lowercase() || prev.is_ascii_digit())
+                    || (prev.is_some_and(char::is_uppercase)
+                        && next.is_some_and(char::is_lowercase)));
+            if needs_separator {
+                normalized.push('_');
+            }
+            normalized.extend(ch.to_lowercase());
+            continue;
+        }
+
+        normalized.push(ch.to_ascii_lowercase());
+    }
+
+    normalized.trim_matches('_').to_string()
+}
+
+fn permission_mode_from_plugin_tool(value: &str) -> Result<PermissionMode, String> {
+    match value {
+        "read-only" => Ok(PermissionMode::ReadOnly),
+        "workspace-write" => Ok(PermissionMode::WorkspaceWrite),
+        "danger-full-access" => Ok(PermissionMode::DangerFullAccess),
+        other => Err(format!(
+            "unsupported plugin tool permission `{other}` (expected read-only, workspace-write, or danger-full-access)"
+        )),
+    }
+}
+
 #[must_use]
 #[allow(clippy::too_many_lines)]
 pub fn mvp_tool_specs() -> Vec<ToolSpec> {
@@ -323,7 +557,7 @@ pub fn mvp_tool_specs() -> Vec<ToolSpec> {
         },
         ToolSpec {
             name: "Config",
-            description: "Get or set Claude Code settings.",
+            description: "Get or set Claw Code settings.",
             input_schema: json!({
                 "type": "object",
                 "properties": {
@@ -1466,14 +1700,15 @@ fn build_agent_runtime(
         .clone()
         .unwrap_or_else(|| DEFAULT_AGENT_MODEL.to_string());
     let allowed_tools = job.allowed_tools.clone();
+    let tool_registry = current_tool_registry()?;
     let api_client =
-        AnthropicRuntimeClient::new(model, allowed_tools.clone(), job.manifest.agent_id.clone())?;
-    let tool_executor = SubagentToolExecutor::new(allowed_tools);
+        AnthropicRuntimeClient::new(model, allowed_tools.clone(), tool_registry.clone())?;
+    let tool_executor = SubagentToolExecutor::new(allowed_tools, tool_registry.clone());
     Ok(ConversationRuntime::new(
         Session::new(),
         api_client,
         tool_executor,
-        agent_permission_policy(),
+        agent_permission_policy(&tool_registry),
         job.system_prompt.clone(),
     ))
 }
@@ -1538,7 +1773,7 @@ fn allowed_tools_for_subagent(subagent_type: &str) -> BTreeSet<String> {
             "SendUserMessage",
             "PowerShell",
         ],
-        "claude-code-guide" => vec![
+        "claw-code-guide" => vec![
             "read_file",
             "glob_search",
             "grep_search",
@@ -1582,10 +1817,12 @@ fn allowed_tools_for_subagent(subagent_type: &str) -> BTreeSet<String> {
     tools.into_iter().map(str::to_string).collect()
 }
 
-fn agent_permission_policy() -> PermissionPolicy {
-    mvp_tool_specs().into_iter().fold(
+fn agent_permission_policy(tool_registry: &GlobalToolRegistry) -> PermissionPolicy {
+    tool_registry.permission_specs(None).into_iter().fold(
         PermissionPolicy::new(PermissionMode::DangerFullAccess),
-        |policy, spec| policy.with_tool_requirement(spec.name, spec.required_permission),
+        |policy, (name, required_permission)| {
+            policy.with_tool_requirement(name, required_permission)
+        },
     )
 }
 
@@ -1641,38 +1878,31 @@ struct AnthropicRuntimeClient {
     client: AnthropicClient,
     model: String,
     allowed_tools: BTreeSet<String>,
+    tool_registry: GlobalToolRegistry,
 }
 
 impl AnthropicRuntimeClient {
     fn new(
         model: String,
         allowed_tools: BTreeSet<String>,
-        session_id: impl Into<String>,
+        tool_registry: GlobalToolRegistry,
     ) -> Result<Self, String> {
         let client = AnthropicClient::from_env()
             .map_err(|error| error.to_string())?
-            .with_base_url(read_base_url())
-            .with_prompt_cache(PromptCache::new(session_id));
+            .with_base_url(read_base_url());
         Ok(Self {
             runtime: tokio::runtime::Runtime::new().map_err(|error| error.to_string())?,
             client,
             model,
             allowed_tools,
+            tool_registry,
         })
     }
 }
 
 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()
-            .map(|spec| ToolDefinition {
-                name: spec.name.to_string(),
-                description: Some(spec.description.to_string()),
-                input_schema: spec.input_schema,
-            })
-            .collect::<Vec<_>>();
+        let tools = self.tool_registry.definitions(Some(&self.allowed_tools));
         let message_request = MessageRequest {
             model: self.model.clone(),
             max_tokens: 32_000,
@@ -1723,6 +1953,8 @@ impl ApiClient for AnthropicRuntimeClient {
                                 input.push_str(&partial_json);
                             }
                         }
+                        ContentBlockDelta::ThinkingDelta { .. }
+                        | ContentBlockDelta::SignatureDelta { .. } => {}
                     },
                     ApiStreamEvent::ContentBlockStop(_) => {
                         if let Some((id, name, input)) = pending_tool.take() {
@@ -1733,8 +1965,8 @@ 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: delta.usage.cache_creation_input_tokens,
-                            cache_read_input_tokens: delta.usage.cache_read_input_tokens,
+                            cache_creation_input_tokens: 0,
+                            cache_read_input_tokens: 0,
                         }));
                     }
                     ApiStreamEvent::MessageStop(_) => {
@@ -1744,8 +1976,6 @@ 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())
@@ -1770,41 +2000,89 @@ impl ApiClient for AnthropicRuntimeClient {
                 })
                 .await
                 .map_err(|error| RuntimeError::new(error.to_string()))?;
-            let mut events = response_to_events(response);
-            push_prompt_cache_record(&self.client, &mut events);
-            Ok(events)
+            Ok(response_to_events(response))
         })
     }
 }
 
 struct SubagentToolExecutor {
     allowed_tools: BTreeSet<String>,
+    tool_registry: GlobalToolRegistry,
 }
 
 impl SubagentToolExecutor {
-    fn new(allowed_tools: BTreeSet<String>) -> Self {
-        Self { allowed_tools }
+    fn new(allowed_tools: BTreeSet<String>, tool_registry: GlobalToolRegistry) -> Self {
+        Self {
+            allowed_tools,
+            tool_registry,
+        }
     }
 }
 
 impl ToolExecutor for SubagentToolExecutor {
     fn execute(&mut self, tool_name: &str, input: &str) -> Result<String, ToolError> {
-        if !self.allowed_tools.contains(tool_name) {
+        let entry = self
+            .tool_registry
+            .find_entry(tool_name)
+            .ok_or_else(|| ToolError::new(format!("unsupported tool: {tool_name}")))?;
+        if !self.allowed_tools.contains(entry.definition.name.as_str()) {
             return Err(ToolError::new(format!(
                 "tool `{tool_name}` is not enabled for this sub-agent"
             )));
         }
         let value = serde_json::from_str(input)
             .map_err(|error| ToolError::new(format!("invalid tool input JSON: {error}")))?;
-        execute_tool(tool_name, &value).map_err(ToolError::new)
+        self.tool_registry
+            .execute(tool_name, &value)
+            .map_err(ToolError::new)
     }
 }
 
-fn tool_specs_for_allowed_tools(allowed_tools: Option<&BTreeSet<String>>) -> Vec<ToolSpec> {
-    mvp_tool_specs()
-        .into_iter()
-        .filter(|spec| allowed_tools.is_none_or(|allowed| allowed.contains(spec.name)))
-        .collect()
+fn current_tool_registry() -> Result<GlobalToolRegistry, String> {
+    let cwd = std::env::current_dir().map_err(|error| error.to_string())?;
+    let loader = ConfigLoader::default_for(&cwd);
+    let runtime_config = loader.load().map_err(|error| error.to_string())?;
+    let plugin_manager = build_plugin_manager(&cwd, &loader, &runtime_config);
+    let plugin_tools = plugin_manager
+        .aggregated_tools()
+        .map_err(|error| error.to_string())?;
+    GlobalToolRegistry::with_plugin_tools(plugin_tools)
+}
+
+fn build_plugin_manager(
+    cwd: &Path,
+    loader: &ConfigLoader,
+    runtime_config: &RuntimeConfig,
+) -> PluginManager {
+    let plugin_settings = runtime_config.plugins();
+    let mut plugin_config = PluginManagerConfig::new(loader.config_home().to_path_buf());
+    plugin_config.enabled_plugins = plugin_settings.enabled_plugins().clone();
+    plugin_config.external_dirs = plugin_settings
+        .external_directories()
+        .iter()
+        .map(|path| resolve_plugin_path(cwd, loader.config_home(), path))
+        .collect();
+    plugin_config.install_root = plugin_settings
+        .install_root()
+        .map(|path| resolve_plugin_path(cwd, loader.config_home(), path));
+    plugin_config.registry_path = plugin_settings
+        .registry_path()
+        .map(|path| resolve_plugin_path(cwd, loader.config_home(), path));
+    plugin_config.bundled_root = plugin_settings
+        .bundled_root()
+        .map(|path| resolve_plugin_path(cwd, loader.config_home(), path));
+    PluginManager::new(plugin_config)
+}
+
+fn resolve_plugin_path(cwd: &Path, config_home: &Path, value: &str) -> PathBuf {
+    let path = PathBuf::from(value);
+    if path.is_absolute() {
+        path
+    } else if value.starts_with('.') {
+        cwd.join(path)
+    } else {
+        config_home.join(path)
+    }
 }
 
 fn convert_messages(messages: &[ConversationMessage]) -> Vec<InputMessage> {
@@ -1871,6 +2149,7 @@ fn push_output_block(
             };
             *pending_tool = Some((id, name, initial_input));
         }
+        OutputContentBlock::Thinking { .. } | OutputContentBlock::RedactedThinking { .. } => {}
     }
 }
 
@@ -1895,26 +2174,6 @@ 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
@@ -2118,7 +2377,7 @@ fn normalize_subagent_type(subagent_type: Option<&str>) -> String {
         "verification" | "verificationagent" | "verify" | "verifier" => {
             String::from("Verification")
         }
-        "claudecodeguide" | "claudecodeguideagent" | "guide" => String::from("claude-code-guide"),
+        "claudecodeguide" | "claudecodeguideagent" | "guide" => String::from("claw-code-guide"),
         "statusline" | "statuslinesetup" => String::from("statusline-setup"),
         _ => trimmed.to_string(),
     }
@@ -2936,9 +3195,13 @@ mod tests {
     use super::{
         agent_permission_policy, allowed_tools_for_subagent, execute_agent_with_spawn,
         execute_tool, final_assistant_text, mvp_tool_specs, persist_agent_terminal_state,
-        AgentInput, AgentJob, SubagentToolExecutor,
+        response_to_events, AgentInput, AgentJob, GlobalToolRegistry, SubagentToolExecutor,
+    };
+    use api::{MessageResponse, OutputContentBlock, Usage};
+    use plugins::{PluginTool, PluginToolDefinition, PluginToolPermission};
+    use runtime::{
+        ApiRequest, AssistantEvent, ConversationRuntime, RuntimeError, Session, ToolExecutor,
     };
-    use runtime::{ApiRequest, AssistantEvent, ConversationRuntime, RuntimeError, Session};
     use serde_json::json;
 
     fn env_lock() -> &'static Mutex<()> {
@@ -2954,6 +3217,17 @@ mod tests {
         std::env::temp_dir().join(format!("clawd-tools-{unique}-{name}"))
     }
 
+    fn make_executable(path: &PathBuf) {
+        #[cfg(unix)]
+        {
+            use std::os::unix::fs::PermissionsExt;
+
+            let mut permissions = std::fs::metadata(path).expect("metadata").permissions();
+            permissions.set_mode(0o755);
+            std::fs::set_permissions(path, permissions).expect("chmod");
+        }
+    }
+
     #[test]
     fn exposes_mvp_tools() {
         let names = mvp_tool_specs()
@@ -2983,6 +3257,170 @@ mod tests {
         assert!(error.contains("unsupported tool"));
     }
 
+    #[test]
+    fn global_registry_registers_and_executes_plugin_tools() {
+        let script = temp_path("plugin-tool.sh");
+        std::fs::write(
+            &script,
+            "#!/bin/sh\nINPUT=$(cat)\nprintf '{\"plugin\":\"%s\",\"tool\":\"%s\",\"input\":%s}\\n' \"$CLAWD_PLUGIN_ID\" \"$CLAWD_TOOL_NAME\" \"$INPUT\"\n",
+        )
+        .expect("write script");
+        make_executable(&script);
+
+        let registry = GlobalToolRegistry::with_plugin_tools(vec![PluginTool::new(
+            "demo@external",
+            "demo",
+            PluginToolDefinition {
+                name: "plugin_echo".to_string(),
+                description: Some("Echo plugin input".to_string()),
+                input_schema: json!({
+                    "type": "object",
+                    "properties": { "message": { "type": "string" } },
+                    "required": ["message"],
+                    "additionalProperties": false
+                }),
+            },
+            "sh".to_string(),
+            vec![script.display().to_string()],
+            PluginToolPermission::WorkspaceWrite,
+            script.parent().map(PathBuf::from),
+        )])
+        .expect("registry should build");
+
+        let names = registry
+            .definitions(None)
+            .into_iter()
+            .map(|definition| definition.name)
+            .collect::<Vec<_>>();
+        assert!(names.contains(&"bash".to_string()));
+        assert!(names.contains(&"plugin_echo".to_string()));
+
+        let output = registry
+            .execute("plugin_echo", &json!({ "message": "hello" }))
+            .expect("plugin tool should execute");
+        let payload: serde_json::Value = serde_json::from_str(&output).expect("valid json");
+        assert_eq!(payload["plugin"], "demo@external");
+        assert_eq!(payload["tool"], "plugin_echo");
+        assert_eq!(payload["input"]["message"], "hello");
+
+        let _ = std::fs::remove_file(script);
+    }
+
+    #[test]
+    fn global_registry_normalizes_plugin_tool_names_for_allowlists_and_execution() {
+        let script = temp_path("plugin-tool-normalized.sh");
+        std::fs::write(
+            &script,
+            "#!/bin/sh\nINPUT=$(cat)\nprintf '{\"tool\":\"%s\",\"input\":%s}\\n' \"$CLAWD_TOOL_NAME\" \"$INPUT\"\n",
+        )
+        .expect("write script");
+        make_executable(&script);
+
+        let registry = GlobalToolRegistry::with_plugin_tools(vec![PluginTool::new(
+            "demo@external",
+            "demo",
+            PluginToolDefinition {
+                name: "plugin_echo".to_string(),
+                description: Some("Echo plugin input".to_string()),
+                input_schema: json!({
+                    "type": "object",
+                    "properties": { "message": { "type": "string" } },
+                    "required": ["message"],
+                    "additionalProperties": false
+                }),
+            },
+            script.display().to_string(),
+            Vec::new(),
+            PluginToolPermission::WorkspaceWrite,
+            script.parent().map(PathBuf::from),
+        )])
+        .expect("registry should build");
+
+        let allowed = registry
+            .normalize_allowed_tools(&[String::from("PLUGIN-ECHO")])
+            .expect("plugin tool allowlist should normalize")
+            .expect("allowlist should be present");
+        assert!(allowed.contains("plugin_echo"));
+
+        let output = registry
+            .execute("plugin-echo", &json!({ "message": "hello" }))
+            .expect("normalized plugin tool name should execute");
+        let payload: serde_json::Value = serde_json::from_str(&output).expect("valid json");
+        assert_eq!(payload["tool"], "plugin_echo");
+        assert_eq!(payload["input"]["message"], "hello");
+
+        let builtin_output = GlobalToolRegistry::builtin()
+            .execute("structured-output", &json!({ "ok": true }))
+            .expect("normalized builtin tool name should execute");
+        let builtin_payload: serde_json::Value =
+            serde_json::from_str(&builtin_output).expect("valid json");
+        assert_eq!(builtin_payload["structured_output"]["ok"], true);
+
+        let _ = std::fs::remove_file(script);
+    }
+
+    #[test]
+    fn subagent_executor_executes_allowed_plugin_tools() {
+        let script = temp_path("subagent-plugin-tool.sh");
+        std::fs::write(
+            &script,
+            "#!/bin/sh\nINPUT=$(cat)\nprintf '{\"tool\":\"%s\",\"input\":%s}\\n' \"$CLAWD_TOOL_NAME\" \"$INPUT\"\n",
+        )
+        .expect("write script");
+        make_executable(&script);
+
+        let registry = GlobalToolRegistry::with_plugin_tools(vec![PluginTool::new(
+            "demo@external",
+            "demo",
+            PluginToolDefinition {
+                name: "plugin_echo".to_string(),
+                description: Some("Echo plugin input".to_string()),
+                input_schema: json!({
+                    "type": "object",
+                    "properties": { "message": { "type": "string" } },
+                    "required": ["message"],
+                    "additionalProperties": false
+                }),
+            },
+            script.display().to_string(),
+            Vec::new(),
+            PluginToolPermission::WorkspaceWrite,
+            script.parent().map(PathBuf::from),
+        )])
+        .expect("registry should build");
+
+        let mut executor =
+            SubagentToolExecutor::new(BTreeSet::from([String::from("plugin_echo")]), registry);
+        let output = executor
+            .execute("plugin-echo", r#"{"message":"hello"}"#)
+            .expect("plugin tool should execute for subagent");
+        let payload: serde_json::Value = serde_json::from_str(&output).expect("valid json");
+        assert_eq!(payload["tool"], "plugin_echo");
+        assert_eq!(payload["input"]["message"], "hello");
+
+        let _ = std::fs::remove_file(script);
+    }
+
+    #[test]
+    fn global_registry_rejects_conflicting_plugin_tool_names() {
+        let error = GlobalToolRegistry::with_plugin_tools(vec![PluginTool::new(
+            "demo@external",
+            "demo",
+            PluginToolDefinition {
+                name: "read-file".to_string(),
+                description: Some("Conflicts with builtin".to_string()),
+                input_schema: json!({ "type": "object" }),
+            },
+            "echo".to_string(),
+            Vec::new(),
+            PluginToolPermission::ReadOnly,
+            None,
+        )])
+        .expect_err("conflicting plugin tool should fail");
+
+        assert!(error.contains("conflicts with already-registered tool `read_file`"));
+    }
+
     #[test]
     fn web_fetch_returns_prompt_aware_summary() {
         let server = TestServer::spawn(Arc::new(|request_line: &str| {
@@ -3562,8 +4000,11 @@ mod tests {
                 calls: 0,
                 input_path: path.display().to_string(),
             },
-            SubagentToolExecutor::new(BTreeSet::from([String::from("read_file")])),
-            agent_permission_policy(),
+            SubagentToolExecutor::new(
+                BTreeSet::from([String::from("read_file")]),
+                GlobalToolRegistry::builtin(),
+            ),
+            agent_permission_policy(&GlobalToolRegistry::builtin()),
             vec![String::from("system prompt")],
         );
 
@@ -3589,6 +4030,42 @@ mod tests {
         let _ = std::fs::remove_file(path);
     }
 
+    #[test]
+    fn response_to_events_ignores_thinking_blocks() {
+        let events = response_to_events(MessageResponse {
+            id: "msg-1".to_string(),
+            kind: "message".to_string(),
+            model: "claude-opus-4-6".to_string(),
+            role: "assistant".to_string(),
+            content: vec![
+                OutputContentBlock::Thinking {
+                    thinking: "step 1".to_string(),
+                    signature: Some("sig_123".to_string()),
+                },
+                OutputContentBlock::Text {
+                    text: "Final answer".to_string(),
+                },
+            ],
+            stop_reason: Some("end_turn".to_string()),
+            stop_sequence: None,
+            usage: Usage {
+                input_tokens: 1,
+                output_tokens: 1,
+                cache_creation_input_tokens: 0,
+                cache_read_input_tokens: 0,
+            },
+            request_id: None,
+        });
+
+        assert!(matches!(
+            &events[0],
+            AssistantEvent::TextDelta(text) if text == "Final answer"
+        ));
+        assert!(!events
+            .iter()
+            .any(|event| matches!(event, AssistantEvent::ToolUse { .. })));
+    }
+
     #[test]
     fn agent_rejects_blank_required_fields() {
         let missing_description = execute_tool(