feat: cache-tracking progress

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

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

.claude.json

@@ -0,0 +1,5 @@
+{
+  "permissions": {
+    "defaultMode": "dontAsk"
+  }
+}

.gitignore

@@ -2,3 +2,6 @@ __pycache__/
 archive/
 .omx/
 .clawd-agents/
+# Claude Code local artifacts
+.claude/settings.local.json
+.claude/sessions/

CLAUDE.md

@@ -0,0 +1,21 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## Detected stack
+- Languages: Rust.
+- Frameworks: none detected from the supported starter markers.
+
+## Verification
+- Run Rust verification from `rust/`: `cargo fmt`, `cargo clippy --workspace --all-targets -- -D warnings`, `cargo test --workspace`
+- `src/` and `tests/` are both present; update both surfaces together when behavior changes.
+
+## Repository shape
+- `rust/` contains the Rust workspace and active CLI/runtime implementation.
+- `src/` contains source files that should stay consistent with generated guidance and tests.
+- `tests/` contains validation surfaces that should be reviewed alongside code changes.
+
+## Working agreement
+- Prefer small, reviewable changes and keep generated bootstrap files aligned with actual repo workflows.
+- Keep shared defaults in `.claude.json`; reserve `.claude/settings.local.json` for machine-local overrides.
+- Do not overwrite existing `CLAUDE.md` content automatically; update it intentionally when repo workflows change.

rust/.claude/sessions/session-1775012674485.json

@@ -0,0 +1 @@
+{"messages":[{"blocks":[{"text":"clear","type":"text"}],"role":"user"},{"blocks":[{"text":"\n\nI've cleared the conversation. How can I help you today?","type":"text"}],"role":"assistant","usage":{"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"input_tokens":4272,"output_tokens":17}}],"version":1}

rust/.claude/sessions/session-1775012687059.json

@@ -0,0 +1 @@
+{"messages":[{"blocks":[{"text":"exit","type":"text"}],"role":"user"},{"blocks":[{"text":"\n\nGoodbye! 👋","type":"text"}],"role":"assistant","usage":{"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"input_tokens":4272,"output_tokens":10}}],"version":1}

rust/.claude/sessions/session-1775013221875.json

@@ -0,0 +1 @@
+{"messages":[],"version":1}

rust/.clawd-todos.json

@@ -1,22 +1,27 @@
 [
   {
-    "content": "Phase 0: Structural Cleanup — spawn 4 agents for 0.1-0.4",
-    "activeForm": "Executing Phase 0: Structural Cleanup via sub-agents",
+    "content": "Architecture & dependency analysis",
+    "activeForm": "Complete",
+    "status": "completed"
+  },
+  {
+    "content": "Runtime crate deep analysis",
+    "activeForm": "Complete",
+    "status": "completed"
+  },
+  {
+    "content": "CLI & Tools analysis",
+    "activeForm": "Complete",
+    "status": "completed"
+  },
+  {
+    "content": "Code quality verification",
+    "activeForm": "Complete",
+    "status": "completed"
+  },
+  {
+    "content": "Synthesize findings into unified report",
+    "activeForm": "Writing report",
     "status": "in_progress"
-  },
-  {
-    "content": "Phase 1.1-1.2: Status bar with live HUD and token counter",
-    "activeForm": "Awaiting Phase 0",
-    "status": "pending"
-  },
-  {
-    "content": "Phase 2.4: Remove artificial 8ms stream delay",
-    "activeForm": "Awaiting Phase 0",
-    "status": "pending"
-  },
-  {
-    "content": "Phase 3.1: Collapsible tool output",
-    "activeForm": "Awaiting Phase 0",
-    "status": "pending"
   }
 ]

rust/Cargo.lock

@@ -1545,10 +1545,12 @@ dependencies = [
 name = "tools"
 version = "0.1.0"
 dependencies = [
+ "api",
  "reqwest",
  "runtime",
  "serde",
  "serde_json",
+ "tokio",
 ]
 
 [[package]]

rust/README.md

@@ -1,230 +1,149 @@
-# Rusty Claude CLI
+# 🦞 Claw Code — Rust Implementation
 
-`rust/` contains the Rust workspace for the integrated `rusty-claude-cli` deliverable.
-It is intended to be something you can clone, build, and run directly.
+A high-performance Rust rewrite of the Claude Code CLI agent harness. Built for speed, safety, and native tool execution.
 
-## Workspace layout
+## Quick Start
 
-```text
+```bash
+# Build
+cd rust/
+cargo build --release
+
+# Run interactive REPL
+./target/release/claw
+
+# One-shot prompt
+./target/release/claw prompt "explain this codebase"
+
+# With specific model
+./target/release/claw --model sonnet prompt "fix the bug in main.rs"
+```
+
+## Configuration
+
+Set your API credentials:
+
+```bash
+export ANTHROPIC_API_KEY="sk-ant-..."
+# Or use a proxy
+export ANTHROPIC_BASE_URL="https://your-proxy.com"
+```
+
+Or authenticate via OAuth:
+
+```bash
+claw login
+```
+
+## Features
+
+| Feature | Status |
+|---------|--------|
+| Anthropic API + streaming | ✅ |
+| OAuth login/logout | ✅ |
+| Interactive REPL (rustyline) | ✅ |
+| Tool system (bash, read, write, edit, grep, glob) | ✅ |
+| Web tools (search, fetch) | ✅ |
+| Sub-agent orchestration | ✅ |
+| Todo tracking | ✅ |
+| Notebook editing | ✅ |
+| CLAUDE.md / project memory | ✅ |
+| Config file hierarchy (.claude.json) | ✅ |
+| Permission system | ✅ |
+| MCP server lifecycle | ✅ |
+| Session persistence + resume | ✅ |
+| Extended thinking (thinking blocks) | ✅ |
+| Cost tracking + usage display | ✅ |
+| Git integration | ✅ |
+| Markdown terminal rendering (ANSI) | ✅ |
+| Model aliases (opus/sonnet/haiku) | ✅ |
+| Slash commands (/status, /compact, /clear, etc.) | ✅ |
+| Hooks (PreToolUse/PostToolUse) | 🔧 Config only |
+| Plugin system | 📋 Planned |
+| Skills registry | 📋 Planned |
+
+## Model Aliases
+
+Short names resolve to the latest model versions:
+
+| Alias | Resolves To |
+|-------|------------|
+| `opus` | `claude-opus-4-6` |
+| `sonnet` | `claude-sonnet-4-6` |
+| `haiku` | `claude-haiku-4-5-20251213` |
+
+## CLI Flags
+
+```
+claw [OPTIONS] [COMMAND]
+
+Options:
+  --model MODEL                    Set the model (alias or full name)
+  --dangerously-skip-permissions   Skip all permission checks
+  --permission-mode MODE           Set read-only, workspace-write, or danger-full-access
+  --allowedTools TOOLS             Restrict enabled tools
+  --output-format FORMAT           Output format (text or json)
+  --version, -V                    Print version info
+
+Commands:
+  prompt <text>      One-shot prompt (non-interactive)
+  login              Authenticate via OAuth
+  logout             Clear stored credentials
+  init               Initialize project config
+  doctor             Check environment health
+  self-update        Update to latest version
+```
+
+## Slash Commands (REPL)
+
+| Command | Description |
+|---------|-------------|
+| `/help` | Show help |
+| `/status` | Show session status (model, tokens, cost) |
+| `/cost` | Show cost breakdown |
+| `/compact` | Compact conversation history |
+| `/clear` | Clear conversation |
+| `/model [name]` | Show or switch model |
+| `/permissions` | Show or switch permission mode |
+| `/config [section]` | Show config (env, hooks, model) |
+| `/memory` | Show CLAUDE.md contents |
+| `/diff` | Show git diff |
+| `/export [path]` | Export conversation |
+| `/session [id]` | Resume a previous session |
+| `/version` | Show version |
+
+## Workspace Layout
+
+```
 rust/
-├── Cargo.toml
+├── Cargo.toml              # Workspace root
 ├── Cargo.lock
-├── README.md
 └── crates/
-    ├── api/               # Anthropic API client + SSE streaming support
-    ├── commands/          # Shared slash-command metadata/help surfaces
-    ├── compat-harness/    # Upstream TS manifest extraction harness
-    ├── runtime/           # Session/runtime/config/prompt orchestration
-    ├── rusty-claude-cli/  # Main CLI binary
+    ├── api/                # Anthropic API client + SSE streaming
+    ├── commands/           # Shared slash-command registry
+    ├── compat-harness/     # TS manifest extraction harness
+    ├── runtime/            # Session, config, permissions, MCP, prompts
+    ├── rusty-claude-cli/   # Main CLI binary (`claw`)
     └── tools/              # Built-in tool implementations
 ```
 
-## Prerequisites
+### Crate Responsibilities
 
-- Rust toolchain installed (`rustup`, stable toolchain)
-- Network access and Anthropic credentials for live prompt/REPL usage
+- **api** — HTTP client, SSE stream parser, request/response types, auth (API key + OAuth bearer)
+- **commands** — Slash command definitions and help text generation
+- **compat-harness** — Extracts tool/prompt manifests from upstream TS source
+- **runtime** — `ConversationRuntime` agentic loop, `ConfigLoader` hierarchy, `Session` persistence, permission policy, MCP client, system prompt assembly, usage tracking
+- **rusty-claude-cli** — REPL, one-shot prompt, streaming display, tool call rendering, CLI argument parsing
+- **tools** — Tool specs + execution: Bash, ReadFile, WriteFile, EditFile, GlobSearch, GrepSearch, WebSearch, WebFetch, Agent, TodoWrite, NotebookEdit, Skill, ToolSearch, REPL runtimes
 
-## Build
+## Stats
 
-From the repository root:
+- **~20K lines** of Rust
+- **6 crates** in workspace
+- **Binary name:** `claw`
+- **Default model:** `claude-opus-4-6`
+- **Default permissions:** `danger-full-access`
 
-```bash
-cd rust
-cargo build --release -p rusty-claude-cli
-```
+## License
 
-The optimized binary will be written to:
-
-```bash
-./target/release/rusty-claude-cli
-```
-
-## Test
-
-Run the verified workspace test suite used for release-readiness:
-
-```bash
-cd rust
-cargo test --workspace --exclude compat-harness
-```
-
-## Quick start
-
-### Show help
-
-```bash
-cd rust
-cargo run -p rusty-claude-cli -- --help
-```
-
-### Print version
-
-```bash
-cd rust
-cargo run -p rusty-claude-cli -- --version
-```
-
-### Login with OAuth
-
-Configure `settings.json` with an `oauth` block containing `clientId`, `authorizeUrl`, `tokenUrl`, optional `callbackPort`, and optional `scopes`, then run:
-
-```bash
-cd rust
-cargo run -p rusty-claude-cli -- login
-```
-
-This opens the browser, listens on the configured localhost callback, exchanges the auth code for tokens, and stores OAuth credentials in `~/.claude/credentials.json` (or `$CLAUDE_CONFIG_HOME/credentials.json`).
-
-### Logout
-
-```bash
-cd rust
-cargo run -p rusty-claude-cli -- logout
-```
-
-This removes only the stored OAuth credentials and preserves unrelated JSON fields in `credentials.json`.
-
-### Self-update
-
-```bash
-cd rust
-cargo run -p rusty-claude-cli -- self-update
-```
-
-The command checks the latest GitHub release for `instructkr/clawd-code`, compares it to the current binary version, downloads the matching binary asset plus checksum manifest, verifies SHA-256, replaces the current executable, and prints the release changelog. If no published release or matching asset exists, it exits safely with an explanatory message.
-
-## Usage examples
-
-### 1) Prompt mode
-
-Send one prompt, stream the answer, then exit:
-
-```bash
-cd rust
-cargo run -p rusty-claude-cli -- prompt "Summarize the architecture of this repository"
-```
-
-Use a specific model:
-
-```bash
-cd rust
-cargo run -p rusty-claude-cli -- --model claude-sonnet-4-20250514 prompt "List the key crates in this workspace"
-```
-
-Restrict enabled tools in an interactive session:
-
-```bash
-cd rust
-cargo run -p rusty-claude-cli -- --allowedTools read,glob
-```
-
-Bootstrap Claude project files for the current repo:
-
-```bash
-cd rust
-cargo run -p rusty-claude-cli -- init
-```
-
-### 2) REPL mode
-
-Start the interactive shell:
-
-```bash
-cd rust
-cargo run -p rusty-claude-cli --
-```
-
-Inside the REPL, useful commands include:
-
-```text
-/help
-/status
-/model claude-sonnet-4-20250514
-/permissions workspace-write
-/cost
-/compact
-/memory
-/config
-/init
-/diff
-/version
-/export notes.txt
-/sessions
-/session list
-/exit
-```
-
-### 3) Resume an existing session
-
-Inspect or maintain a saved session file without entering the REPL:
-
-```bash
-cd rust
-cargo run -p rusty-claude-cli -- --resume session-123456 /status /compact /cost
-```
-
-You can also inspect memory/config state for a restored session:
-
-```bash
-cd rust
-cargo run -p rusty-claude-cli -- --resume ~/.claude/sessions/session-123456.json /memory /config
-```
-
-## Available commands
-
-### Top-level CLI commands
-
-- `prompt <text...>` — run one prompt non-interactively
-- `--resume <session-id-or-path> [/commands...]` — inspect or maintain a saved session stored under `~/.claude/sessions/`
-- `dump-manifests` — print extracted upstream manifest counts
-- `bootstrap-plan` — print the current bootstrap skeleton
-- `system-prompt [--cwd PATH] [--date YYYY-MM-DD]` — render the synthesized system prompt
-- `self-update` — update the installed binary from the latest GitHub release when a matching asset is available
-- `--help` / `-h` — show CLI help
-- `--version` / `-V` — print the CLI version and build info locally (no API call)
-- `--output-format text|json` — choose non-interactive prompt output rendering
-- `--allowedTools <tool[,tool...]>` — restrict enabled tools for interactive sessions and prompt-mode tool use
-
-### Interactive slash commands
-
-- `/help` — show command help
-- `/status` — show current session status
-- `/compact` — compact local session history
-- `/model [model]` — inspect or switch the active model
-- `/permissions [read-only|workspace-write|danger-full-access]` — inspect or switch permissions
-- `/clear [--confirm]` — clear the current local session
-- `/cost` — show token usage totals
-- `/resume <session-id-or-path>` — load a saved session into the REPL
-- `/config [env|hooks|model]` — inspect discovered Claude config
-- `/memory` — inspect loaded instruction memory files
-- `/init` — bootstrap `.claude.json`, `.claude/`, `CLAUDE.md`, and local ignore rules
-- `/diff` — show the current git diff for the workspace
-- `/version` — print version and build metadata locally
-- `/export [file]` — export the current conversation transcript
-- `/sessions` — list recent managed local sessions from `~/.claude/sessions/`
-- `/session [list|switch <session-id>]` — inspect or switch managed local sessions
-- `/exit` — leave the REPL
-
-## Environment variables
-
-### Anthropic/API
-
-- `ANTHROPIC_API_KEY` — highest-precedence API credential
-- `ANTHROPIC_AUTH_TOKEN` — bearer-token override used when no API key is set
-- Persisted OAuth credentials in `~/.claude/credentials.json` — used when neither env var is set
-- `ANTHROPIC_BASE_URL` — override the Anthropic API base URL
-- `ANTHROPIC_MODEL` — default model used by selected live integration tests
-
-### CLI/runtime
-
-- `RUSTY_CLAUDE_PERMISSION_MODE` — default REPL permission mode (`read-only`, `workspace-write`, or `danger-full-access`)
-- `CLAUDE_CONFIG_HOME` — override Claude config discovery root
-- `CLAUDE_CODE_REMOTE` — enable remote-session bootstrap handling when supported
-- `CLAUDE_CODE_REMOTE_SESSION_ID` — remote session identifier when using remote mode
-- `CLAUDE_CODE_UPSTREAM` — override the upstream TS source path for compat-harness extraction
-- `CLAWD_WEB_SEARCH_BASE_URL` — override the built-in web search service endpoint used by tooling
-
-## Notes
-
-- `compat-harness` exists to compare the Rust port against the upstream TypeScript codebase and is intentionally excluded from the requested release test run.
-- The CLI currently focuses on a practical integrated workflow: prompt execution, REPL operation, session inspection/resume, config discovery, and tool/runtime plumbing.
+See repository root.

rust/crates/api/src/client.rs

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

rust/crates/api/src/lib.rs

@@ -1,13 +1,18 @@
 mod client;
 mod error;
+mod prompt_cache;
 mod sse;
 mod types;
 
 pub use client::{
-    oauth_token_is_expired, read_base_url, resolve_saved_oauth_token,
-    resolve_startup_auth_source, AnthropicClient, AuthSource, MessageStream, OAuthTokenSet,
+    oauth_token_is_expired, read_base_url, resolve_saved_oauth_token, resolve_startup_auth_source,
+    AnthropicClient, AuthSource, MessageStream, OAuthTokenSet,
 };
 pub use error::ApiError;
+pub use prompt_cache::{
+    CacheBreakEvent, PromptCache, PromptCacheConfig, PromptCachePaths, PromptCacheRecord,
+    PromptCacheStats,
+};
 pub use sse::{parse_frame, SseParser};
 pub use types::{
     ContentBlockDelta, ContentBlockDeltaEvent, ContentBlockStartEvent, ContentBlockStopEvent,
@@ -15,3 +20,11 @@ pub use types::{
     MessageResponse, MessageStartEvent, MessageStopEvent, OutputContentBlock, StreamEvent,
     ToolChoice, ToolDefinition, ToolResultContentBlock, Usage,
 };
+
+#[cfg(test)]
+pub(crate) fn test_env_lock() -> std::sync::MutexGuard<'static, ()> {
+    static LOCK: std::sync::OnceLock<std::sync::Mutex<()>> = std::sync::OnceLock::new();
+    LOCK.get_or_init(|| std::sync::Mutex::new(()))
+        .lock()
+        .unwrap_or_else(std::sync::PoisonError::into_inner)
+}

rust/crates/api/src/prompt_cache.rs

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

rust/crates/api/tests/client_integration.rs

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

rust/crates/runtime/src/config.rs

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

rust/crates/runtime/src/conversation.rs

@@ -4,6 +4,8 @@ use std::fmt::{Display, Formatter};
 use crate::compact::{
     compact_session, estimate_session_tokens, CompactionConfig, CompactionResult,
 };
+use crate::config::RuntimeFeatureConfig;
+use crate::hooks::{HookRunResult, HookRunner};
 use crate::permissions::{PermissionOutcome, PermissionPolicy, PermissionPrompter};
 use crate::session::{ContentBlock, ConversationMessage, Session};
 use crate::usage::{TokenUsage, UsageTracker};
@@ -23,9 +25,19 @@ pub enum AssistantEvent {
         input: String,
     },
     Usage(TokenUsage),
+    PromptCache(PromptCacheEvent),
     MessageStop,
 }
 
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct PromptCacheEvent {
+    pub unexpected: bool,
+    pub reason: String,
+    pub previous_cache_read_input_tokens: u32,
+    pub current_cache_read_input_tokens: u32,
+    pub token_drop: u32,
+}
+
 pub trait ApiClient {
     fn stream(&mut self, request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError>;
 }
@@ -82,6 +94,7 @@ 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,
 }
@@ -94,6 +107,7 @@ pub struct ConversationRuntime<C, T> {
     system_prompt: Vec<String>,
     max_iterations: usize,
     usage_tracker: UsageTracker,
+    hook_runner: HookRunner,
 }
 
 impl<C, T> ConversationRuntime<C, T>
@@ -108,6 +122,25 @@ where
         tool_executor: T,
         permission_policy: PermissionPolicy,
         system_prompt: Vec<String>,
+    ) -> Self {
+        Self::new_with_features(
+            session,
+            api_client,
+            tool_executor,
+            permission_policy,
+            system_prompt,
+            &RuntimeFeatureConfig::default(),
+        )
+    }
+
+    #[must_use]
+    pub fn new_with_features(
+        session: Session,
+        api_client: C,
+        tool_executor: T,
+        permission_policy: PermissionPolicy,
+        system_prompt: Vec<String>,
+        feature_config: &RuntimeFeatureConfig,
     ) -> Self {
         let usage_tracker = UsageTracker::from_session(&session);
         Self {
@@ -118,6 +151,7 @@ where
             system_prompt,
             max_iterations: usize::MAX,
             usage_tracker,
+            hook_runner: HookRunner::from_feature_config(feature_config),
         }
     }
 
@@ -138,6 +172,7 @@ where
 
         let mut assistant_messages = Vec::new();
         let mut tool_results = Vec::new();
+        let mut prompt_cache_events = Vec::new();
         let mut iterations = 0;
 
         loop {
@@ -153,10 +188,12 @@ where
                 messages: self.session.messages.clone(),
             };
             let events = self.api_client.stream(request)?;
-            let (assistant_message, usage) = build_assistant_message(events)?;
+            let (assistant_message, usage, turn_prompt_cache_events) =
+                build_assistant_message(events)?;
             if let Some(usage) = usage {
                 self.usage_tracker.record(usage);
             }
+            prompt_cache_events.extend(turn_prompt_cache_events);
             let pending_tool_uses = assistant_message
                 .blocks
                 .iter()
@@ -185,19 +222,41 @@ where
 
                 let result_message = match permission_outcome {
                     PermissionOutcome::Allow => {
+                        let pre_hook_result = self.hook_runner.run_pre_tool_use(&tool_name, &input);
+                        if pre_hook_result.is_denied() {
+                            let deny_message = format!("PreToolUse hook denied tool `{tool_name}`");
+                            ConversationMessage::tool_result(
+                                tool_use_id,
+                                tool_name,
+                                format_hook_message(&pre_hook_result, &deny_message),
+                                true,
+                            )
+                        } else {
+                            let (mut output, mut is_error) =
                                 match self.tool_executor.execute(&tool_name, &input) {
-                            Ok(output) => ConversationMessage::tool_result(
+                                    Ok(output) => (output, false),
+                                    Err(error) => (error.to_string(), true),
+                                };
+                            output = merge_hook_feedback(pre_hook_result.messages(), output, false);
+
+                            let post_hook_result = self
+                                .hook_runner
+                                .run_post_tool_use(&tool_name, &input, &output, is_error);
+                            if post_hook_result.is_denied() {
+                                is_error = true;
+                            }
+                            output = merge_hook_feedback(
+                                post_hook_result.messages(),
+                                output,
+                                post_hook_result.is_denied(),
+                            );
+
+                            ConversationMessage::tool_result(
                                 tool_use_id,
                                 tool_name,
                                 output,
-                                false,
-                            ),
-                            Err(error) => ConversationMessage::tool_result(
-                                tool_use_id,
-                                tool_name,
-                                error.to_string(),
-                                true,
-                            ),
+                                is_error,
+                            )
                         }
                     }
                     PermissionOutcome::Deny { reason } => {
@@ -212,6 +271,7 @@ where
         Ok(TurnSummary {
             assistant_messages,
             tool_results,
+            prompt_cache_events,
             iterations,
             usage: self.usage_tracker.cumulative_usage(),
         })
@@ -245,9 +305,17 @@ where
 
 fn build_assistant_message(
     events: Vec<AssistantEvent>,
-) -> Result<(ConversationMessage, Option<TokenUsage>), RuntimeError> {
+) -> Result<
+    (
+        ConversationMessage,
+        Option<TokenUsage>,
+        Vec<PromptCacheEvent>,
+    ),
+    RuntimeError,
+> {
     let mut text = String::new();
     let mut blocks = Vec::new();
+    let mut prompt_cache_events = Vec::new();
     let mut finished = false;
     let mut usage = None;
 
@@ -259,6 +327,7 @@ fn build_assistant_message(
                 blocks.push(ContentBlock::ToolUse { id, name, input });
             }
             AssistantEvent::Usage(value) => usage = Some(value),
+            AssistantEvent::PromptCache(event) => prompt_cache_events.push(event),
             AssistantEvent::MessageStop => {
                 finished = true;
             }
@@ -279,6 +348,7 @@ fn build_assistant_message(
     Ok((
         ConversationMessage::assistant_with_usage(blocks, usage),
         usage,
+        prompt_cache_events,
     ))
 }
 
@@ -290,6 +360,32 @@ fn flush_text_block(text: &mut String, blocks: &mut Vec<ContentBlock>) {
     }
 }
 
+fn format_hook_message(result: &HookRunResult, fallback: &str) -> String {
+    if result.messages().is_empty() {
+        fallback.to_string()
+    } else {
+        result.messages().join("\n")
+    }
+}
+
+fn merge_hook_feedback(messages: &[String], output: String, denied: bool) -> String {
+    if messages.is_empty() {
+        return output;
+    }
+
+    let mut sections = Vec::new();
+    if !output.trim().is_empty() {
+        sections.push(output);
+    }
+    let label = if denied {
+        "Hook feedback (denied)"
+    } else {
+        "Hook feedback"
+    };
+    sections.push(format!("{label}:\n{}", messages.join("\n")));
+    sections.join("\n\n")
+}
+
 type ToolHandler = Box<dyn FnMut(&str) -> Result<String, ToolError>>;
 
 #[derive(Default)]
@@ -325,10 +421,11 @@ impl ToolExecutor for StaticToolExecutor {
 #[cfg(test)]
 mod tests {
     use super::{
-        ApiClient, ApiRequest, AssistantEvent, ConversationRuntime, RuntimeError,
+        ApiClient, ApiRequest, AssistantEvent, ConversationRuntime, PromptCacheEvent, RuntimeError,
         StaticToolExecutor,
     };
     use crate::compact::CompactionConfig;
+    use crate::config::{RuntimeFeatureConfig, RuntimeHookConfig};
     use crate::permissions::{
         PermissionMode, PermissionPolicy, PermissionPromptDecision, PermissionPrompter,
         PermissionRequest,
@@ -381,6 +478,15 @@ mod tests {
                             cache_creation_input_tokens: 1,
                             cache_read_input_tokens: 3,
                         }),
+                        AssistantEvent::PromptCache(PromptCacheEvent {
+                            unexpected: true,
+                            reason:
+                                "cache read tokens dropped while prompt fingerprint remained stable"
+                                    .to_string(),
+                            previous_cache_read_input_tokens: 6_000,
+                            current_cache_read_input_tokens: 1_000,
+                            token_drop: 5_000,
+                        }),
                         AssistantEvent::MessageStop,
                     ])
                 }
@@ -434,8 +540,10 @@ mod tests {
         assert_eq!(summary.iterations, 2);
         assert_eq!(summary.assistant_messages.len(), 2);
         assert_eq!(summary.tool_results.len(), 1);
+        assert_eq!(summary.prompt_cache_events.len(), 1);
         assert_eq!(runtime.session().messages.len(), 4);
         assert_eq!(summary.usage.output_tokens, 10);
+        assert!(summary.prompt_cache_events[0].unexpected);
         assert!(matches!(
             runtime.session().messages[1].blocks[1],
             ContentBlock::ToolUse { .. }
@@ -503,6 +611,141 @@ mod tests {
         ));
     }
 
+    #[test]
+    fn denies_tool_use_when_pre_tool_hook_blocks() {
+        struct SingleCallApiClient;
+        impl ApiClient for SingleCallApiClient {
+            fn stream(&mut self, request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError> {
+                if request
+                    .messages
+                    .iter()
+                    .any(|message| message.role == MessageRole::Tool)
+                {
+                    return Ok(vec![
+                        AssistantEvent::TextDelta("blocked".to_string()),
+                        AssistantEvent::MessageStop,
+                    ]);
+                }
+                Ok(vec![
+                    AssistantEvent::ToolUse {
+                        id: "tool-1".to_string(),
+                        name: "blocked".to_string(),
+                        input: r#"{"path":"secret.txt"}"#.to_string(),
+                    },
+                    AssistantEvent::MessageStop,
+                ])
+            }
+        }
+
+        let mut runtime = ConversationRuntime::new_with_features(
+            Session::new(),
+            SingleCallApiClient,
+            StaticToolExecutor::new().register("blocked", |_input| {
+                panic!("tool should not execute when hook denies")
+            }),
+            PermissionPolicy::new(PermissionMode::DangerFullAccess),
+            vec!["system".to_string()],
+            &RuntimeFeatureConfig::default().with_hooks(RuntimeHookConfig::new(
+                vec![shell_snippet("printf 'blocked by hook'; exit 2")],
+                Vec::new(),
+            )),
+        );
+
+        let summary = runtime
+            .run_turn("use the tool", None)
+            .expect("conversation should continue after hook denial");
+
+        assert_eq!(summary.tool_results.len(), 1);
+        let ContentBlock::ToolResult {
+            is_error, output, ..
+        } = &summary.tool_results[0].blocks[0]
+        else {
+            panic!("expected tool result block");
+        };
+        assert!(
+            *is_error,
+            "hook denial should produce an error result: {output}"
+        );
+        assert!(
+            output.contains("denied tool") || output.contains("blocked by hook"),
+            "unexpected hook denial output: {output:?}"
+        );
+    }
+
+    #[test]
+    fn appends_post_tool_hook_feedback_to_tool_result() {
+        struct TwoCallApiClient {
+            calls: usize,
+        }
+
+        impl ApiClient for TwoCallApiClient {
+            fn stream(&mut self, request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError> {
+                self.calls += 1;
+                match self.calls {
+                    1 => Ok(vec![
+                        AssistantEvent::ToolUse {
+                            id: "tool-1".to_string(),
+                            name: "add".to_string(),
+                            input: r#"{"lhs":2,"rhs":2}"#.to_string(),
+                        },
+                        AssistantEvent::MessageStop,
+                    ]),
+                    2 => {
+                        assert!(request
+                            .messages
+                            .iter()
+                            .any(|message| message.role == MessageRole::Tool));
+                        Ok(vec![
+                            AssistantEvent::TextDelta("done".to_string()),
+                            AssistantEvent::MessageStop,
+                        ])
+                    }
+                    _ => Err(RuntimeError::new("unexpected extra API call")),
+                }
+            }
+        }
+
+        let mut runtime = ConversationRuntime::new_with_features(
+            Session::new(),
+            TwoCallApiClient { calls: 0 },
+            StaticToolExecutor::new().register("add", |_input| Ok("4".to_string())),
+            PermissionPolicy::new(PermissionMode::DangerFullAccess),
+            vec!["system".to_string()],
+            &RuntimeFeatureConfig::default().with_hooks(RuntimeHookConfig::new(
+                vec![shell_snippet("printf 'pre hook ran'")],
+                vec![shell_snippet("printf 'post hook ran'")],
+            )),
+        );
+
+        let summary = runtime
+            .run_turn("use add", None)
+            .expect("tool loop succeeds");
+
+        assert_eq!(summary.tool_results.len(), 1);
+        let ContentBlock::ToolResult {
+            is_error, output, ..
+        } = &summary.tool_results[0].blocks[0]
+        else {
+            panic!("expected tool result block");
+        };
+        assert!(
+            !*is_error,
+            "post hook should preserve non-error result: {output:?}"
+        );
+        assert!(
+            output.contains('4'),
+            "tool output missing value: {output:?}"
+        );
+        assert!(
+            output.contains("pre hook ran"),
+            "tool output missing pre hook feedback: {output:?}"
+        );
+        assert!(
+            output.contains("post hook ran"),
+            "tool output missing post hook feedback: {output:?}"
+        );
+    }
+
     #[test]
     fn reconstructs_usage_tracker_from_restored_session() {
         struct SimpleApi;
@@ -581,4 +824,14 @@ mod tests {
             MessageRole::System
         );
     }
+
+    #[cfg(windows)]
+    fn shell_snippet(script: &str) -> String {
+        script.replace('\'', "\"")
+    }
+
+    #[cfg(not(windows))]
+    fn shell_snippet(script: &str) -> String {
+        script.to_string()
+    }
 }

rust/crates/runtime/src/hooks.rs

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

rust/crates/runtime/src/lib.rs

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

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

@@ -13,8 +13,9 @@ use std::time::{SystemTime, UNIX_EPOCH};
 
 use api::{
     resolve_startup_auth_source, AnthropicClient, AuthSource, ContentBlockDelta, InputContentBlock,
-    InputMessage, MessageRequest, MessageResponse, OutputContentBlock,
-    StreamEvent as ApiStreamEvent, ToolChoice, ToolDefinition, ToolResultContentBlock,
+    InputMessage, MessageRequest, MessageResponse, OutputContentBlock, PromptCache,
+    PromptCacheRecord, StreamEvent as ApiStreamEvent, ToolChoice, ToolDefinition,
+    ToolResultContentBlock,
 };
 
 use commands::{
@@ -27,9 +28,9 @@ use runtime::{
     clear_oauth_credentials, generate_pkce_pair, generate_state, load_system_prompt,
     parse_oauth_callback_request_target, save_oauth_credentials, ApiClient, ApiRequest,
     AssistantEvent, CompactionConfig, ConfigLoader, ConfigSource, ContentBlock,
-    ConversationMessage, ConversationRuntime, MessageRole, OAuthAuthorizationRequest,
-    OAuthTokenExchangeRequest, PermissionMode, PermissionPolicy, ProjectContext, RuntimeError,
-    Session, TokenUsage, ToolError, ToolExecutor, UsageTracker,
+    ConversationMessage, ConversationRuntime, MessageRole, OAuthAuthorizationRequest, OAuthConfig,
+    OAuthTokenExchangeRequest, PermissionMode, PermissionPolicy, ProjectContext, PromptCacheEvent,
+    RuntimeError, Session, TokenUsage, ToolError, ToolExecutor, UsageTracker,
 };
 use serde_json::json;
 use tools::{execute_tool, mvp_tool_specs, ToolSpec};
@@ -196,6 +197,25 @@ fn parse_args(args: &[String]) -> Result<CliAction, String> {
                 permission_mode = PermissionMode::DangerFullAccess;
                 index += 1;
             }
+            "-p" => {
+                // Claude Code compat: -p "prompt" = one-shot prompt
+                let prompt = args[index + 1..].join(" ");
+                if prompt.trim().is_empty() {
+                    return Err("-p requires a prompt string".to_string());
+                }
+                return Ok(CliAction::Prompt {
+                    prompt,
+                    model: resolve_model_alias(&model).to_string(),
+                    output_format,
+                    allowed_tools: normalize_allowed_tools(&allowed_tool_values)?,
+                    permission_mode,
+                });
+            }
+            "--print" => {
+                // Claude Code compat: --print makes output non-interactive
+                output_format = CliOutputFormat::Text;
+                index += 1;
+            }
             "--allowedTools" | "--allowed-tools" => {
                 let value = args
                     .get(index + 1)
@@ -428,15 +448,26 @@ fn print_bootstrap_plan() {
     }
 }
 
+fn default_oauth_config() -> OAuthConfig {
+    OAuthConfig {
+        client_id: String::from("9d1c250a-e61b-44d9-88ed-5944d1962f5e"),
+        authorize_url: String::from("https://platform.claude.com/oauth/authorize"),
+        token_url: String::from("https://platform.claude.com/v1/oauth/token"),
+        callback_port: None,
+        manual_redirect_url: None,
+        scopes: vec![
+            String::from("user:profile"),
+            String::from("user:inference"),
+            String::from("user:sessions:claude_code"),
+        ],
+    }
+}
+
 fn run_login() -> Result<(), Box<dyn std::error::Error>> {
     let cwd = env::current_dir()?;
     let config = ConfigLoader::default_for(&cwd).load()?;
-    let oauth = config.oauth().ok_or_else(|| {
-        io::Error::new(
-            io::ErrorKind::NotFound,
-            "OAuth config is missing. Add settings.oauth.clientId/authorizeUrl/tokenUrl first.",
-        )
-    })?;
+    let default_oauth = default_oauth_config();
+    let oauth = config.oauth().unwrap_or(&default_oauth);
     let callback_port = oauth.callback_port.unwrap_or(DEFAULT_OAUTH_CALLBACK_PORT);
     let redirect_uri = runtime::loopback_redirect_uri(callback_port);
     let pkce = generate_pkce_pair()?;
@@ -965,6 +996,7 @@ impl LiveCli {
         let session = create_managed_session_handle()?;
         let runtime = build_runtime(
             Session::new(),
+            session.id.clone(),
             model.clone(),
             system_prompt.clone(),
             enable_tools,
@@ -1020,13 +1052,14 @@ impl LiveCli {
         let mut permission_prompter = CliPermissionPrompter::new(self.permission_mode);
         let result = self.runtime.run_turn(input, Some(&mut permission_prompter));
         match result {
-            Ok(_) => {
+            Ok(summary) => {
                 spinner.finish(
                     "✨ Done",
                     TerminalRenderer::new().color_theme(),
                     &mut stdout,
                 )?;
                 println!();
+                print_prompt_cache_events(&summary);
                 self.persist_session()?;
                 Ok(())
             }
@@ -1056,6 +1089,7 @@ impl LiveCli {
         let session = self.runtime.session().clone();
         let mut runtime = build_runtime(
             session,
+            self.session.id.clone(),
             self.model.clone(),
             self.system_prompt.clone(),
             true,
@@ -1075,6 +1109,7 @@ impl LiveCli {
                 "iterations": summary.iterations,
                 "tool_uses": collect_tool_uses(&summary),
                 "tool_results": collect_tool_results(&summary),
+                "prompt_cache_events": collect_prompt_cache_events(&summary),
                 "usage": {
                     "input_tokens": summary.usage.input_tokens,
                     "output_tokens": summary.usage.output_tokens,
@@ -1202,6 +1237,7 @@ impl LiveCli {
         let message_count = session.messages.len();
         self.runtime = build_runtime(
             session,
+            self.session.id.clone(),
             model.clone(),
             self.system_prompt.clone(),
             true,
@@ -1245,6 +1281,7 @@ impl LiveCli {
         self.permission_mode = permission_mode_from_label(normalized);
         self.runtime = build_runtime(
             session,
+            self.session.id.clone(),
             self.model.clone(),
             self.system_prompt.clone(),
             true,
@@ -1270,6 +1307,7 @@ impl LiveCli {
         self.session = create_managed_session_handle()?;
         self.runtime = build_runtime(
             Session::new(),
+            self.session.id.clone(),
             self.model.clone(),
             self.system_prompt.clone(),
             true,
@@ -1305,6 +1343,7 @@ impl LiveCli {
         let message_count = session.messages.len();
         self.runtime = build_runtime(
             session,
+            handle.id.clone(),
             self.model.clone(),
             self.system_prompt.clone(),
             true,
@@ -1377,6 +1416,7 @@ impl LiveCli {
                 let message_count = session.messages.len();
                 self.runtime = build_runtime(
                     session,
+                    handle.id.clone(),
                     self.model.clone(),
                     self.system_prompt.clone(),
                     true,
@@ -1407,6 +1447,7 @@ impl LiveCli {
         let skipped = removed == 0;
         self.runtime = build_runtime(
             result.compacted_session,
+            self.session.id.clone(),
             self.model.clone(),
             self.system_prompt.clone(),
             true,
@@ -1873,8 +1914,19 @@ fn build_system_prompt() -> Result<Vec<String>, Box<dyn std::error::Error>> {
     )?)
 }
 
+fn build_runtime_feature_config(
+) -> Result<runtime::RuntimeFeatureConfig, Box<dyn std::error::Error>> {
+    let cwd = env::current_dir()?;
+    Ok(ConfigLoader::default_for(cwd)
+        .load()?
+        .feature_config()
+        .clone())
+}
+
+#[allow(clippy::too_many_arguments)]
 fn build_runtime(
     session: Session,
+    session_id: String,
     model: String,
     system_prompt: Vec<String>,
     enable_tools: bool,
@@ -1883,12 +1935,19 @@ fn build_runtime(
     permission_mode: PermissionMode,
 ) -> Result<ConversationRuntime<AnthropicRuntimeClient, CliToolExecutor>, Box<dyn std::error::Error>>
 {
-    Ok(ConversationRuntime::new(
+    Ok(ConversationRuntime::new_with_features(
         session,
-        AnthropicRuntimeClient::new(model, enable_tools, emit_output, allowed_tools.clone())?,
+        AnthropicRuntimeClient::new(
+            model,
+            enable_tools,
+            emit_output,
+            allowed_tools.clone(),
+            session_id,
+        )?,
         CliToolExecutor::new(allowed_tools, emit_output),
         permission_policy(permission_mode),
         system_prompt,
+        &build_runtime_feature_config()?,
     ))
 }
 
@@ -1953,11 +2012,13 @@ impl AnthropicRuntimeClient {
         enable_tools: bool,
         emit_output: bool,
         allowed_tools: Option<AllowedToolSet>,
+        session_id: impl Into<String>,
     ) -> Result<Self, Box<dyn std::error::Error>> {
         Ok(Self {
             runtime: tokio::runtime::Runtime::new()?,
             client: AnthropicClient::from_auth(resolve_cli_auth_source()?)
-                .with_base_url(api::read_base_url()),
+                .with_base_url(api::read_base_url())
+                .with_prompt_cache(PromptCache::new(session_id)),
             model,
             enable_tools,
             emit_output,
@@ -2072,8 +2133,8 @@ impl ApiClient for AnthropicRuntimeClient {
                         events.push(AssistantEvent::Usage(TokenUsage {
                             input_tokens: delta.usage.input_tokens,
                             output_tokens: delta.usage.output_tokens,
-                            cache_creation_input_tokens: 0,
-                            cache_read_input_tokens: 0,
+                            cache_creation_input_tokens: delta.usage.cache_creation_input_tokens,
+                            cache_read_input_tokens: delta.usage.cache_read_input_tokens,
                         }));
                     }
                     ApiStreamEvent::MessageStop(_) => {
@@ -2088,6 +2149,8 @@ impl ApiClient for AnthropicRuntimeClient {
                 }
             }
 
+            push_prompt_cache_record(&self.client, &mut events);
+
             if !saw_stop
                 && events.iter().any(|event| {
                     matches!(event, AssistantEvent::TextDelta(text) if !text.is_empty())
@@ -2112,7 +2175,9 @@ impl ApiClient for AnthropicRuntimeClient {
                 })
                 .await
                 .map_err(|error| RuntimeError::new(error.to_string()))?;
-            response_to_events(response, out)
+            let mut events = response_to_events(response, out)?;
+            push_prompt_cache_record(&self.client, &mut events);
+            Ok(events)
         })
     }
 }
@@ -2173,6 +2238,39 @@ fn collect_tool_results(summary: &runtime::TurnSummary) -> Vec<serde_json::Value
         .collect()
 }
 
+fn collect_prompt_cache_events(summary: &runtime::TurnSummary) -> Vec<serde_json::Value> {
+    summary
+        .prompt_cache_events
+        .iter()
+        .map(|event| {
+            json!({
+                "unexpected": event.unexpected,
+                "reason": event.reason,
+                "previous_cache_read_input_tokens": event.previous_cache_read_input_tokens,
+                "current_cache_read_input_tokens": event.current_cache_read_input_tokens,
+                "token_drop": event.token_drop,
+            })
+        })
+        .collect()
+}
+
+fn print_prompt_cache_events(summary: &runtime::TurnSummary) {
+    for event in &summary.prompt_cache_events {
+        let label = if event.unexpected {
+            "Prompt cache break"
+        } else {
+            "Prompt cache invalidation"
+        };
+        println!(
+            "{label}: {} (cache read {} -> {}, drop {})",
+            event.reason,
+            event.previous_cache_read_input_tokens,
+            event.current_cache_read_input_tokens,
+            event.token_drop,
+        );
+    }
+}
+
 fn slash_command_completion_candidates() -> Vec<String> {
     slash_command_specs()
         .iter()
@@ -2319,18 +2417,20 @@ fn first_visible_line(text: &str) -> &str {
 }
 
 fn format_bash_result(icon: &str, parsed: &serde_json::Value) -> String {
+    use std::fmt::Write as _;
+
     let mut lines = vec![format!("{icon} \x1b[38;5;245mbash\x1b[0m")];
     if let Some(task_id) = parsed
         .get("backgroundTaskId")
         .and_then(|value| value.as_str())
     {
-        lines[0].push_str(&format!(" backgrounded ({task_id})"));
+        let _ = write!(lines[0], " backgrounded ({task_id})");
     } else if let Some(status) = parsed
         .get("returnCodeInterpretation")
         .and_then(|value| value.as_str())
         .filter(|status| !status.is_empty())
     {
-        lines[0].push_str(&format!(" {status}"));
+        let _ = write!(lines[0], " {status}");
     }
 
     if let Some(stdout) = parsed.get("stdout").and_then(|value| value.as_str()) {
@@ -2352,15 +2452,15 @@ fn format_read_result(icon: &str, parsed: &serde_json::Value) -> String {
     let path = extract_tool_path(file);
     let start_line = file
         .get("startLine")
-        .and_then(|value| value.as_u64())
+        .and_then(serde_json::Value::as_u64)
         .unwrap_or(1);
     let num_lines = file
         .get("numLines")
-        .and_then(|value| value.as_u64())
+        .and_then(serde_json::Value::as_u64)
         .unwrap_or(0);
     let total_lines = file
         .get("totalLines")
-        .and_then(|value| value.as_u64())
+        .and_then(serde_json::Value::as_u64)
         .unwrap_or(num_lines);
     let content = file
         .get("content")
@@ -2386,8 +2486,7 @@ fn format_write_result(icon: &str, parsed: &serde_json::Value) -> String {
     let line_count = parsed
         .get("content")
         .and_then(|value| value.as_str())
-        .map(|content| content.lines().count())
-        .unwrap_or(0);
+        .map_or(0, |content| content.lines().count());
     format!(
         "{icon} \x1b[1;32m✏️ {} {path}\x1b[0m \x1b[2m({line_count} lines)\x1b[0m",
         if kind == "create" { "Wrote" } else { "Updated" },
@@ -2418,7 +2517,7 @@ fn format_edit_result(icon: &str, parsed: &serde_json::Value) -> String {
     let path = extract_tool_path(parsed);
     let suffix = if parsed
         .get("replaceAll")
-        .and_then(|value| value.as_bool())
+        .and_then(serde_json::Value::as_bool)
         .unwrap_or(false)
     {
         " (replace all)"
@@ -2446,7 +2545,7 @@ fn format_edit_result(icon: &str, parsed: &serde_json::Value) -> String {
 fn format_glob_result(icon: &str, parsed: &serde_json::Value) -> String {
     let num_files = parsed
         .get("numFiles")
-        .and_then(|value| value.as_u64())
+        .and_then(serde_json::Value::as_u64)
         .unwrap_or(0);
     let filenames = parsed
         .get("filenames")
@@ -2470,11 +2569,11 @@ fn format_glob_result(icon: &str, parsed: &serde_json::Value) -> String {
 fn format_grep_result(icon: &str, parsed: &serde_json::Value) -> String {
     let num_matches = parsed
         .get("numMatches")
-        .and_then(|value| value.as_u64())
+        .and_then(serde_json::Value::as_u64)
         .unwrap_or(0);
     let num_files = parsed
         .get("numFiles")
-        .and_then(|value| value.as_u64())
+        .and_then(serde_json::Value::as_u64)
         .unwrap_or(0);
     let content = parsed
         .get("content")
@@ -2581,6 +2680,26 @@ fn response_to_events(
     Ok(events)
 }
 
+fn push_prompt_cache_record(client: &AnthropicClient, events: &mut Vec<AssistantEvent>) {
+    if let Some(event) = client
+        .take_last_prompt_cache_record()
+        .and_then(prompt_cache_record_to_runtime_event)
+    {
+        events.push(AssistantEvent::PromptCache(event));
+    }
+}
+
+fn prompt_cache_record_to_runtime_event(record: PromptCacheRecord) -> Option<PromptCacheEvent> {
+    let cache_break = record.cache_break?;
+    Some(PromptCacheEvent {
+        unexpected: cache_break.unexpected,
+        reason: cache_break.reason,
+        previous_cache_read_input_tokens: cache_break.previous_cache_read_input_tokens,
+        current_cache_read_input_tokens: cache_break.current_cache_read_input_tokens,
+        token_drop: cache_break.token_drop,
+    })
+}
+
 struct CliToolExecutor {
     renderer: TerminalRenderer,
     emit_output: bool,

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

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

rust/crates/tools/Cargo.toml

@@ -6,10 +6,12 @@ license.workspace = true
 publish.workspace = true
 
 [dependencies]
+api = { path = "../api" }
 runtime = { path = "../runtime" }
 reqwest = { version = "0.12", default-features = false, features = ["blocking", "rustls-tls"] }
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
+tokio = { version = "1", features = ["rt-multi-thread"] }
 
 [lints]
 workspace = true

rust/crates/tools/src/lib.rs

@@ -3,10 +3,17 @@ use std::path::{Path, PathBuf};
 use std::process::Command;
 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,
+};
 use reqwest::blocking::Client;
 use runtime::{
-    edit_file, execute_bash, glob_search, grep_search, read_file, write_file, BashCommandInput,
-    GrepSearchInput, PermissionMode,
+    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,
 };
 use serde::{Deserialize, Serialize};
 use serde_json::{json, Value};
@@ -702,7 +709,7 @@ struct SkillOutput {
     prompt: String,
 }
 
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Debug, Clone, Serialize, Deserialize)]
 struct AgentOutput {
     #[serde(rename = "agentId")]
     agent_id: String,
@@ -718,6 +725,20 @@ struct AgentOutput {
     manifest_file: String,
     #[serde(rename = "createdAt")]
     created_at: String,
+    #[serde(rename = "startedAt", skip_serializing_if = "Option::is_none")]
+    started_at: Option<String>,
+    #[serde(rename = "completedAt", skip_serializing_if = "Option::is_none")]
+    completed_at: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    error: Option<String>,
+}
+
+#[derive(Debug, Clone)]
+struct AgentJob {
+    manifest: AgentOutput,
+    prompt: String,
+    system_prompt: Vec<String>,
+    allowed_tools: BTreeSet<String>,
 }
 
 #[derive(Debug, Serialize)]
@@ -1315,7 +1336,18 @@ fn resolve_skill_path(skill: &str) -> Result<std::path::PathBuf, String> {
     Err(format!("unknown skill: {requested}"))
 }
 
+const DEFAULT_AGENT_MODEL: &str = "claude-opus-4-6";
+const DEFAULT_AGENT_SYSTEM_DATE: &str = "2026-03-31";
+const DEFAULT_AGENT_MAX_ITERATIONS: usize = 32;
+
 fn execute_agent(input: AgentInput) -> Result<AgentOutput, String> {
+    execute_agent_with_spawn(input, spawn_agent_job)
+}
+
+fn execute_agent_with_spawn<F>(input: AgentInput, spawn_fn: F) -> Result<AgentOutput, String>
+where
+    F: FnOnce(AgentJob) -> Result<(), String>,
+{
     if input.description.trim().is_empty() {
         return Err(String::from("description must not be empty"));
     }
@@ -1329,6 +1361,7 @@ fn execute_agent(input: AgentInput) -> Result<AgentOutput, String> {
     let output_file = output_dir.join(format!("{agent_id}.md"));
     let manifest_file = output_dir.join(format!("{agent_id}.json"));
     let normalized_subagent_type = normalize_subagent_type(input.subagent_type.as_deref());
+    let model = resolve_agent_model(input.model.as_deref());
     let agent_name = input
         .name
         .as_deref()
@@ -1336,6 +1369,8 @@ fn execute_agent(input: AgentInput) -> Result<AgentOutput, String> {
         .filter(|name| !name.is_empty())
         .unwrap_or_else(|| slugify_agent_name(&input.description));
     let created_at = iso8601_now();
+    let system_prompt = build_agent_system_prompt(&normalized_subagent_type)?;
+    let allowed_tools = allowed_tools_for_subagent(&normalized_subagent_type);
 
     let output_contents = format!(
         "# Agent Task
@@ -1359,21 +1394,545 @@ fn execute_agent(input: AgentInput) -> Result<AgentOutput, String> {
         name: agent_name,
         description: input.description,
         subagent_type: Some(normalized_subagent_type),
-        model: input.model,
-        status: String::from("queued"),
+        model: Some(model),
+        status: String::from("running"),
         output_file: output_file.display().to_string(),
         manifest_file: manifest_file.display().to_string(),
-        created_at,
+        created_at: created_at.clone(),
+        started_at: Some(created_at),
+        completed_at: None,
+        error: None,
     };
-    std::fs::write(
-        &manifest_file,
-        serde_json::to_string_pretty(&manifest).map_err(|error| error.to_string())?,
-    )
-    .map_err(|error| error.to_string())?;
+    write_agent_manifest(&manifest)?;
+
+    let manifest_for_spawn = manifest.clone();
+    let job = AgentJob {
+        manifest: manifest_for_spawn,
+        prompt: input.prompt,
+        system_prompt,
+        allowed_tools,
+    };
+    if let Err(error) = spawn_fn(job) {
+        let error = format!("failed to spawn sub-agent: {error}");
+        persist_agent_terminal_state(&manifest, "failed", None, Some(error.clone()))?;
+        return Err(error);
+    }
 
     Ok(manifest)
 }
 
+fn spawn_agent_job(job: AgentJob) -> Result<(), String> {
+    let thread_name = format!("clawd-agent-{}", job.manifest.agent_id);
+    std::thread::Builder::new()
+        .name(thread_name)
+        .spawn(move || {
+            let result =
+                std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| run_agent_job(&job)));
+            match result {
+                Ok(Ok(())) => {}
+                Ok(Err(error)) => {
+                    let _ =
+                        persist_agent_terminal_state(&job.manifest, "failed", None, Some(error));
+                }
+                Err(_) => {
+                    let _ = persist_agent_terminal_state(
+                        &job.manifest,
+                        "failed",
+                        None,
+                        Some(String::from("sub-agent thread panicked")),
+                    );
+                }
+            }
+        })
+        .map(|_| ())
+        .map_err(|error| error.to_string())
+}
+
+fn run_agent_job(job: &AgentJob) -> Result<(), String> {
+    let mut runtime = build_agent_runtime(job)?.with_max_iterations(DEFAULT_AGENT_MAX_ITERATIONS);
+    let summary = runtime
+        .run_turn(job.prompt.clone(), None)
+        .map_err(|error| error.to_string())?;
+    let final_text = final_assistant_text(&summary);
+    persist_agent_terminal_state(&job.manifest, "completed", Some(final_text.as_str()), None)
+}
+
+fn build_agent_runtime(
+    job: &AgentJob,
+) -> Result<ConversationRuntime<AnthropicRuntimeClient, SubagentToolExecutor>, String> {
+    let model = job
+        .manifest
+        .model
+        .clone()
+        .unwrap_or_else(|| DEFAULT_AGENT_MODEL.to_string());
+    let allowed_tools = job.allowed_tools.clone();
+    let api_client =
+        AnthropicRuntimeClient::new(model, allowed_tools.clone(), job.manifest.agent_id.clone())?;
+    let tool_executor = SubagentToolExecutor::new(allowed_tools);
+    Ok(ConversationRuntime::new(
+        Session::new(),
+        api_client,
+        tool_executor,
+        agent_permission_policy(),
+        job.system_prompt.clone(),
+    ))
+}
+
+fn build_agent_system_prompt(subagent_type: &str) -> Result<Vec<String>, String> {
+    let cwd = std::env::current_dir().map_err(|error| error.to_string())?;
+    let mut prompt = load_system_prompt(
+        cwd,
+        DEFAULT_AGENT_SYSTEM_DATE.to_string(),
+        std::env::consts::OS,
+        "unknown",
+    )
+    .map_err(|error| error.to_string())?;
+    prompt.push(format!(
+        "You are a background sub-agent of type `{subagent_type}`. Work only on the delegated task, use only the tools available to you, do not ask the user questions, and finish with a concise result."
+    ));
+    Ok(prompt)
+}
+
+fn resolve_agent_model(model: Option<&str>) -> String {
+    model
+        .map(str::trim)
+        .filter(|model| !model.is_empty())
+        .unwrap_or(DEFAULT_AGENT_MODEL)
+        .to_string()
+}
+
+fn allowed_tools_for_subagent(subagent_type: &str) -> BTreeSet<String> {
+    let tools = match subagent_type {
+        "Explore" => vec![
+            "read_file",
+            "glob_search",
+            "grep_search",
+            "WebFetch",
+            "WebSearch",
+            "ToolSearch",
+            "Skill",
+            "StructuredOutput",
+        ],
+        "Plan" => vec![
+            "read_file",
+            "glob_search",
+            "grep_search",
+            "WebFetch",
+            "WebSearch",
+            "ToolSearch",
+            "Skill",
+            "TodoWrite",
+            "StructuredOutput",
+            "SendUserMessage",
+        ],
+        "Verification" => vec![
+            "bash",
+            "read_file",
+            "glob_search",
+            "grep_search",
+            "WebFetch",
+            "WebSearch",
+            "ToolSearch",
+            "TodoWrite",
+            "StructuredOutput",
+            "SendUserMessage",
+            "PowerShell",
+        ],
+        "claude-code-guide" => vec![
+            "read_file",
+            "glob_search",
+            "grep_search",
+            "WebFetch",
+            "WebSearch",
+            "ToolSearch",
+            "Skill",
+            "StructuredOutput",
+            "SendUserMessage",
+        ],
+        "statusline-setup" => vec![
+            "bash",
+            "read_file",
+            "write_file",
+            "edit_file",
+            "glob_search",
+            "grep_search",
+            "ToolSearch",
+        ],
+        _ => vec![
+            "bash",
+            "read_file",
+            "write_file",
+            "edit_file",
+            "glob_search",
+            "grep_search",
+            "WebFetch",
+            "WebSearch",
+            "TodoWrite",
+            "Skill",
+            "ToolSearch",
+            "NotebookEdit",
+            "Sleep",
+            "SendUserMessage",
+            "Config",
+            "StructuredOutput",
+            "REPL",
+            "PowerShell",
+        ],
+    };
+    tools.into_iter().map(str::to_string).collect()
+}
+
+fn agent_permission_policy() -> PermissionPolicy {
+    mvp_tool_specs().into_iter().fold(
+        PermissionPolicy::new(PermissionMode::DangerFullAccess),
+        |policy, spec| policy.with_tool_requirement(spec.name, spec.required_permission),
+    )
+}
+
+fn write_agent_manifest(manifest: &AgentOutput) -> Result<(), String> {
+    std::fs::write(
+        &manifest.manifest_file,
+        serde_json::to_string_pretty(manifest).map_err(|error| error.to_string())?,
+    )
+    .map_err(|error| error.to_string())
+}
+
+fn persist_agent_terminal_state(
+    manifest: &AgentOutput,
+    status: &str,
+    result: Option<&str>,
+    error: Option<String>,
+) -> Result<(), String> {
+    append_agent_output(
+        &manifest.output_file,
+        &format_agent_terminal_output(status, result, error.as_deref()),
+    )?;
+    let mut next_manifest = manifest.clone();
+    next_manifest.status = status.to_string();
+    next_manifest.completed_at = Some(iso8601_now());
+    next_manifest.error = error;
+    write_agent_manifest(&next_manifest)
+}
+
+fn append_agent_output(path: &str, suffix: &str) -> Result<(), String> {
+    use std::io::Write as _;
+
+    let mut file = std::fs::OpenOptions::new()
+        .append(true)
+        .open(path)
+        .map_err(|error| error.to_string())?;
+    file.write_all(suffix.as_bytes())
+        .map_err(|error| error.to_string())
+}
+
+fn format_agent_terminal_output(status: &str, result: Option<&str>, error: Option<&str>) -> String {
+    let mut sections = vec![format!("\n## Result\n\n- status: {status}\n")];
+    if let Some(result) = result.filter(|value| !value.trim().is_empty()) {
+        sections.push(format!("\n### Final response\n\n{}\n", result.trim()));
+    }
+    if let Some(error) = error.filter(|value| !value.trim().is_empty()) {
+        sections.push(format!("\n### Error\n\n{}\n", error.trim()));
+    }
+    sections.join("")
+}
+
+struct AnthropicRuntimeClient {
+    runtime: tokio::runtime::Runtime,
+    client: AnthropicClient,
+    model: String,
+    allowed_tools: BTreeSet<String>,
+}
+
+impl AnthropicRuntimeClient {
+    fn new(
+        model: String,
+        allowed_tools: BTreeSet<String>,
+        session_id: impl Into<String>,
+    ) -> Result<Self, String> {
+        let client = AnthropicClient::from_env()
+            .map_err(|error| error.to_string())?
+            .with_base_url(read_base_url())
+            .with_prompt_cache(PromptCache::new(session_id));
+        Ok(Self {
+            runtime: tokio::runtime::Runtime::new().map_err(|error| error.to_string())?,
+            client,
+            model,
+            allowed_tools,
+        })
+    }
+}
+
+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 message_request = MessageRequest {
+            model: self.model.clone(),
+            max_tokens: 32_000,
+            messages: convert_messages(&request.messages),
+            system: (!request.system_prompt.is_empty()).then(|| request.system_prompt.join("\n\n")),
+            tools: (!tools.is_empty()).then_some(tools),
+            tool_choice: (!self.allowed_tools.is_empty()).then_some(ToolChoice::Auto),
+            stream: true,
+        };
+
+        self.runtime.block_on(async {
+            let mut stream = self
+                .client
+                .stream_message(&message_request)
+                .await
+                .map_err(|error| RuntimeError::new(error.to_string()))?;
+            let mut events = Vec::new();
+            let mut pending_tool: Option<(String, String, String)> = None;
+            let mut saw_stop = false;
+
+            while let Some(event) = stream
+                .next_event()
+                .await
+                .map_err(|error| RuntimeError::new(error.to_string()))?
+            {
+                match event {
+                    ApiStreamEvent::MessageStart(start) => {
+                        for block in start.message.content {
+                            push_output_block(block, &mut events, &mut pending_tool, true);
+                        }
+                    }
+                    ApiStreamEvent::ContentBlockStart(start) => {
+                        push_output_block(
+                            start.content_block,
+                            &mut events,
+                            &mut pending_tool,
+                            true,
+                        );
+                    }
+                    ApiStreamEvent::ContentBlockDelta(delta) => match delta.delta {
+                        ContentBlockDelta::TextDelta { text } => {
+                            if !text.is_empty() {
+                                events.push(AssistantEvent::TextDelta(text));
+                            }
+                        }
+                        ContentBlockDelta::InputJsonDelta { partial_json } => {
+                            if let Some((_, _, input)) = &mut pending_tool {
+                                input.push_str(&partial_json);
+                            }
+                        }
+                    },
+                    ApiStreamEvent::ContentBlockStop(_) => {
+                        if let Some((id, name, input)) = pending_tool.take() {
+                            events.push(AssistantEvent::ToolUse { id, name, input });
+                        }
+                    }
+                    ApiStreamEvent::MessageDelta(delta) => {
+                        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,
+                        }));
+                    }
+                    ApiStreamEvent::MessageStop(_) => {
+                        saw_stop = true;
+                        events.push(AssistantEvent::MessageStop);
+                    }
+                }
+            }
+
+            push_prompt_cache_record(&self.client, &mut events);
+
+            if !saw_stop
+                && events.iter().any(|event| {
+                    matches!(event, AssistantEvent::TextDelta(text) if !text.is_empty())
+                        || matches!(event, AssistantEvent::ToolUse { .. })
+                })
+            {
+                events.push(AssistantEvent::MessageStop);
+            }
+
+            if events
+                .iter()
+                .any(|event| matches!(event, AssistantEvent::MessageStop))
+            {
+                return Ok(events);
+            }
+
+            let response = self
+                .client
+                .send_message(&MessageRequest {
+                    stream: false,
+                    ..message_request.clone()
+                })
+                .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)
+        })
+    }
+}
+
+struct SubagentToolExecutor {
+    allowed_tools: BTreeSet<String>,
+}
+
+impl SubagentToolExecutor {
+    fn new(allowed_tools: BTreeSet<String>) -> Self {
+        Self { allowed_tools }
+    }
+}
+
+impl ToolExecutor for SubagentToolExecutor {
+    fn execute(&mut self, tool_name: &str, input: &str) -> Result<String, ToolError> {
+        if !self.allowed_tools.contains(tool_name) {
+            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)
+    }
+}
+
+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 convert_messages(messages: &[ConversationMessage]) -> Vec<InputMessage> {
+    messages
+        .iter()
+        .filter_map(|message| {
+            let role = match message.role {
+                MessageRole::System | MessageRole::User | MessageRole::Tool => "user",
+                MessageRole::Assistant => "assistant",
+            };
+            let content = message
+                .blocks
+                .iter()
+                .map(|block| match block {
+                    ContentBlock::Text { text } => InputContentBlock::Text { text: text.clone() },
+                    ContentBlock::ToolUse { id, name, input } => InputContentBlock::ToolUse {
+                        id: id.clone(),
+                        name: name.clone(),
+                        input: serde_json::from_str(input)
+                            .unwrap_or_else(|_| serde_json::json!({ "raw": input })),
+                    },
+                    ContentBlock::ToolResult {
+                        tool_use_id,
+                        output,
+                        is_error,
+                        ..
+                    } => InputContentBlock::ToolResult {
+                        tool_use_id: tool_use_id.clone(),
+                        content: vec![ToolResultContentBlock::Text {
+                            text: output.clone(),
+                        }],
+                        is_error: *is_error,
+                    },
+                })
+                .collect::<Vec<_>>();
+            (!content.is_empty()).then(|| InputMessage {
+                role: role.to_string(),
+                content,
+            })
+        })
+        .collect()
+}
+
+fn push_output_block(
+    block: OutputContentBlock,
+    events: &mut Vec<AssistantEvent>,
+    pending_tool: &mut Option<(String, String, String)>,
+    streaming_tool_input: bool,
+) {
+    match block {
+        OutputContentBlock::Text { text } => {
+            if !text.is_empty() {
+                events.push(AssistantEvent::TextDelta(text));
+            }
+        }
+        OutputContentBlock::ToolUse { id, name, input } => {
+            let initial_input = if streaming_tool_input
+                && input.is_object()
+                && input.as_object().is_some_and(serde_json::Map::is_empty)
+            {
+                String::new()
+            } else {
+                input.to_string()
+            };
+            *pending_tool = Some((id, name, initial_input));
+        }
+    }
+}
+
+fn response_to_events(response: MessageResponse) -> Vec<AssistantEvent> {
+    let mut events = Vec::new();
+    let mut pending_tool = None;
+
+    for block in response.content {
+        push_output_block(block, &mut events, &mut pending_tool, false);
+        if let Some((id, name, input)) = pending_tool.take() {
+            events.push(AssistantEvent::ToolUse { id, name, input });
+        }
+    }
+
+    events.push(AssistantEvent::Usage(TokenUsage {
+        input_tokens: response.usage.input_tokens,
+        output_tokens: response.usage.output_tokens,
+        cache_creation_input_tokens: response.usage.cache_creation_input_tokens,
+        cache_read_input_tokens: response.usage.cache_read_input_tokens,
+    }));
+    events.push(AssistantEvent::MessageStop);
+    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
+        .last()
+        .map(|message| {
+            message
+                .blocks
+                .iter()
+                .filter_map(|block| match block {
+                    ContentBlock::Text { text } => Some(text.as_str()),
+                    _ => None,
+                })
+                .collect::<Vec<_>>()
+                .join("")
+        })
+        .unwrap_or_default()
+}
+
 #[allow(clippy::needless_pass_by_value)]
 fn execute_tool_search(input: ToolSearchInput) -> ToolSearchOutput {
     let deferred = deferred_tool_specs();
@@ -2207,7 +2766,7 @@ fn execute_shell_command(
             persisted_output_path: None,
             persisted_output_size: None,
             sandbox_status: None,
-});
+        });
     }
 
     let mut process = std::process::Command::new(shell);
@@ -2276,7 +2835,7 @@ Command exceeded timeout of {timeout_ms} ms",
                     persisted_output_path: None,
                     persisted_output_size: None,
                     sandbox_status: None,
-});
+                });
             }
             std::thread::sleep(Duration::from_millis(10));
         }
@@ -2365,6 +2924,7 @@ fn parse_skill_description(contents: &str) -> Option<String> {
 
 #[cfg(test)]
 mod tests {
+    use std::collections::BTreeSet;
     use std::fs;
     use std::io::{Read, Write};
     use std::net::{SocketAddr, TcpListener};
@@ -2373,7 +2933,12 @@ mod tests {
     use std::thread;
     use std::time::Duration;
 
-    use super::{execute_tool, mvp_tool_specs};
+    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,
+    };
+    use runtime::{ApiRequest, AssistantEvent, ConversationRuntime, RuntimeError, Session};
     use serde_json::json;
 
     fn env_lock() -> &'static Mutex<()> {
@@ -2765,32 +3330,48 @@ mod tests {
             .unwrap_or_else(std::sync::PoisonError::into_inner);
         let dir = temp_path("agent-store");
         std::env::set_var("CLAWD_AGENT_STORE", &dir);
+        let captured = Arc::new(Mutex::new(None::<AgentJob>));
+        let captured_for_spawn = Arc::clone(&captured);
 
-        let result = execute_tool(
-            "Agent",
-            &json!({
-                "description": "Audit the branch",
-                "prompt": "Check tests and outstanding work.",
-                "subagent_type": "Explore",
-                "name": "ship-audit"
-            }),
+        let manifest = execute_agent_with_spawn(
+            AgentInput {
+                description: "Audit the branch".to_string(),
+                prompt: "Check tests and outstanding work.".to_string(),
+                subagent_type: Some("Explore".to_string()),
+                name: Some("ship-audit".to_string()),
+                model: None,
+            },
+            move |job| {
+                *captured_for_spawn
+                    .lock()
+                    .unwrap_or_else(std::sync::PoisonError::into_inner) = Some(job);
+                Ok(())
+            },
         )
         .expect("Agent should succeed");
         std::env::remove_var("CLAWD_AGENT_STORE");
 
-        let output: serde_json::Value = serde_json::from_str(&result).expect("valid json");
-        assert_eq!(output["name"], "ship-audit");
-        assert_eq!(output["subagentType"], "Explore");
-        assert_eq!(output["status"], "queued");
-        assert!(output["createdAt"].as_str().is_some());
-        let manifest_file = output["manifestFile"].as_str().expect("manifest file");
-        let output_file = output["outputFile"].as_str().expect("output file");
-        let contents = std::fs::read_to_string(output_file).expect("agent file exists");
+        assert_eq!(manifest.name, "ship-audit");
+        assert_eq!(manifest.subagent_type.as_deref(), Some("Explore"));
+        assert_eq!(manifest.status, "running");
+        assert!(!manifest.created_at.is_empty());
+        assert!(manifest.started_at.is_some());
+        assert!(manifest.completed_at.is_none());
+        let contents = std::fs::read_to_string(&manifest.output_file).expect("agent file exists");
         let manifest_contents =
-            std::fs::read_to_string(manifest_file).expect("manifest file exists");
+            std::fs::read_to_string(&manifest.manifest_file).expect("manifest file exists");
         assert!(contents.contains("Audit the branch"));
         assert!(contents.contains("Check tests and outstanding work."));
         assert!(manifest_contents.contains("\"subagentType\": \"Explore\""));
+        assert!(manifest_contents.contains("\"status\": \"running\""));
+        let captured_job = captured
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+            .clone()
+            .expect("spawn job should be captured");
+        assert_eq!(captured_job.prompt, "Check tests and outstanding work.");
+        assert!(captured_job.allowed_tools.contains("read_file"));
+        assert!(!captured_job.allowed_tools.contains("Agent"));
 
         let normalized = execute_tool(
             "Agent",
@@ -2819,6 +3400,195 @@ mod tests {
         let _ = std::fs::remove_dir_all(dir);
     }
 
+    #[test]
+    fn agent_fake_runner_can_persist_completion_and_failure() {
+        let _guard = env_lock()
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        let dir = temp_path("agent-runner");
+        std::env::set_var("CLAWD_AGENT_STORE", &dir);
+
+        let completed = execute_agent_with_spawn(
+            AgentInput {
+                description: "Complete the task".to_string(),
+                prompt: "Do the work".to_string(),
+                subagent_type: Some("Explore".to_string()),
+                name: Some("complete-task".to_string()),
+                model: Some("claude-sonnet-4-6".to_string()),
+            },
+            |job| {
+                persist_agent_terminal_state(
+                    &job.manifest,
+                    "completed",
+                    Some("Finished successfully"),
+                    None,
+                )
+            },
+        )
+        .expect("completed agent should succeed");
+
+        let completed_manifest = std::fs::read_to_string(&completed.manifest_file)
+            .expect("completed manifest should exist");
+        let completed_output =
+            std::fs::read_to_string(&completed.output_file).expect("completed output should exist");
+        assert!(completed_manifest.contains("\"status\": \"completed\""));
+        assert!(completed_output.contains("Finished successfully"));
+
+        let failed = execute_agent_with_spawn(
+            AgentInput {
+                description: "Fail the task".to_string(),
+                prompt: "Do the failing work".to_string(),
+                subagent_type: Some("Verification".to_string()),
+                name: Some("fail-task".to_string()),
+                model: None,
+            },
+            |job| {
+                persist_agent_terminal_state(
+                    &job.manifest,
+                    "failed",
+                    None,
+                    Some(String::from("simulated failure")),
+                )
+            },
+        )
+        .expect("failed agent should still spawn");
+
+        let failed_manifest =
+            std::fs::read_to_string(&failed.manifest_file).expect("failed manifest should exist");
+        let failed_output =
+            std::fs::read_to_string(&failed.output_file).expect("failed output should exist");
+        assert!(failed_manifest.contains("\"status\": \"failed\""));
+        assert!(failed_manifest.contains("simulated failure"));
+        assert!(failed_output.contains("simulated failure"));
+
+        let spawn_error = execute_agent_with_spawn(
+            AgentInput {
+                description: "Spawn error task".to_string(),
+                prompt: "Never starts".to_string(),
+                subagent_type: None,
+                name: Some("spawn-error".to_string()),
+                model: None,
+            },
+            |_| Err(String::from("thread creation failed")),
+        )
+        .expect_err("spawn errors should surface");
+        assert!(spawn_error.contains("failed to spawn sub-agent"));
+        let spawn_error_manifest = std::fs::read_dir(&dir)
+            .expect("agent dir should exist")
+            .filter_map(Result::ok)
+            .map(|entry| entry.path())
+            .filter(|path| path.extension().and_then(|ext| ext.to_str()) == Some("json"))
+            .find_map(|path| {
+                let contents = std::fs::read_to_string(&path).ok()?;
+                contents
+                    .contains("\"name\": \"spawn-error\"")
+                    .then_some(contents)
+            })
+            .expect("failed manifest should still be written");
+        assert!(spawn_error_manifest.contains("\"status\": \"failed\""));
+        assert!(spawn_error_manifest.contains("thread creation failed"));
+
+        std::env::remove_var("CLAWD_AGENT_STORE");
+        let _ = std::fs::remove_dir_all(dir);
+    }
+
+    #[test]
+    fn agent_tool_subset_mapping_is_expected() {
+        let general = allowed_tools_for_subagent("general-purpose");
+        assert!(general.contains("bash"));
+        assert!(general.contains("write_file"));
+        assert!(!general.contains("Agent"));
+
+        let explore = allowed_tools_for_subagent("Explore");
+        assert!(explore.contains("read_file"));
+        assert!(explore.contains("grep_search"));
+        assert!(!explore.contains("bash"));
+
+        let plan = allowed_tools_for_subagent("Plan");
+        assert!(plan.contains("TodoWrite"));
+        assert!(plan.contains("StructuredOutput"));
+        assert!(!plan.contains("Agent"));
+
+        let verification = allowed_tools_for_subagent("Verification");
+        assert!(verification.contains("bash"));
+        assert!(verification.contains("PowerShell"));
+        assert!(!verification.contains("write_file"));
+    }
+
+    #[derive(Debug)]
+    struct MockSubagentApiClient {
+        calls: usize,
+        input_path: String,
+    }
+
+    impl runtime::ApiClient for MockSubagentApiClient {
+        fn stream(&mut self, request: ApiRequest) -> Result<Vec<AssistantEvent>, RuntimeError> {
+            self.calls += 1;
+            match self.calls {
+                1 => {
+                    assert_eq!(request.messages.len(), 1);
+                    Ok(vec![
+                        AssistantEvent::ToolUse {
+                            id: "tool-1".to_string(),
+                            name: "read_file".to_string(),
+                            input: json!({ "path": self.input_path }).to_string(),
+                        },
+                        AssistantEvent::MessageStop,
+                    ])
+                }
+                2 => {
+                    assert!(request.messages.len() >= 3);
+                    Ok(vec![
+                        AssistantEvent::TextDelta("Scope: completed mock review".to_string()),
+                        AssistantEvent::MessageStop,
+                    ])
+                }
+                _ => panic!("unexpected mock stream call"),
+            }
+        }
+    }
+
+    #[test]
+    fn subagent_runtime_executes_tool_loop_with_isolated_session() {
+        let _guard = env_lock()
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        let path = temp_path("subagent-input.txt");
+        std::fs::write(&path, "hello from child").expect("write input file");
+
+        let mut runtime = ConversationRuntime::new(
+            Session::new(),
+            MockSubagentApiClient {
+                calls: 0,
+                input_path: path.display().to_string(),
+            },
+            SubagentToolExecutor::new(BTreeSet::from([String::from("read_file")])),
+            agent_permission_policy(),
+            vec![String::from("system prompt")],
+        );
+
+        let summary = runtime
+            .run_turn("Inspect the delegated file", None)
+            .expect("subagent loop should succeed");
+
+        assert_eq!(
+            final_assistant_text(&summary),
+            "Scope: completed mock review"
+        );
+        assert!(runtime
+            .session()
+            .messages
+            .iter()
+            .flat_map(|message| message.blocks.iter())
+            .any(|block| matches!(
+                block,
+                runtime::ContentBlock::ToolResult { output, .. }
+                    if output.contains("hello from child")
+            )));
+
+        let _ = std::fs::remove_file(path);
+    }
+
     #[test]
     fn agent_rejects_blank_required_fields() {
         let missing_description = execute_tool(