fix(tools): use resolve_tool_path for consistent path resolution (#3937)

Replace workspace_dir.join(path) with resolve_tool_path(path) in
file_write, file_edit, and pdf_read tools to correctly handle absolute
paths within the workspace directory, preventing path doubling.

Closes #3774

Cargo.lock

@@ -5120,7 +5120,7 @@ version = "3.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e67ba7e9b2b56446f1d419b1d807906278ffa1a658a8a5d8a39dcb1f5a78614f"
 dependencies = [
- "toml_edit 0.25.4+spec-1.1.0",
+ "toml_edit 0.25.5+spec-1.1.0",
 ]
 
 [[package]]
@@ -6415,9 +6415,9 @@ dependencies = [
 
 [[package]]
 name = "serialport"
-version = "4.7.3"
+version = "4.9.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2acaf3f973e8616d7ceac415f53fc60e190b2a686fbcf8d27d0256c741c5007b"
+checksum = "a4d91116f97173694f1642263b2ff837f80d933aa837e2314969f6728f661df3"
 dependencies = [
  "bitflags 2.11.0",
  "cfg-if",
@@ -6428,7 +6428,7 @@ dependencies = [
  "nix 0.26.4",
  "scopeguard",
  "unescaper",
- "winapi",
+ "windows-sys 0.52.0",
 ]
 
 [[package]]
@@ -7047,9 +7047,9 @@ dependencies = [
 
 [[package]]
 name = "tokio-websockets"
-version = "0.13.1"
+version = "0.13.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8b6aa6c8b5a31e06fd3760eb5c1b8d9072e30731f0467ee3795617fe768e7449"
+checksum = "dad543404f98bfc969aeb71994105c592acfc6c43323fddcd016bb208d1c65cb"
 dependencies = [
  "base64",
  "bytes",
@@ -7057,7 +7057,7 @@ dependencies = [
  "futures-sink",
  "http 1.4.0",
  "httparse",
- "rand 0.9.2",
+ "rand 0.10.0",
  "ring",
  "rustls-pki-types",
  "simdutf8",
@@ -7095,17 +7095,17 @@ dependencies = [
 
 [[package]]
 name = "toml"
-version = "1.0.6+spec-1.1.0"
+version = "1.0.7+spec-1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "399b1124a3c9e16766831c6bba21e50192572cdd98706ea114f9502509686ffc"
+checksum = "dd28d57d8a6f6e458bc0b8784f8fdcc4b99a437936056fa122cb234f18656a96"
 dependencies = [
  "indexmap",
  "serde_core",
  "serde_spanned 1.0.4",
- "toml_datetime 1.0.0+spec-1.1.0",
+ "toml_datetime 1.0.1+spec-1.1.0",
  "toml_parser",
  "toml_writer",
- "winnow 0.7.15",
+ "winnow 1.0.0",
 ]
 
 [[package]]
@@ -7128,9 +7128,9 @@ dependencies = [
 
 [[package]]
 name = "toml_datetime"
-version = "1.0.0+spec-1.1.0"
+version = "1.0.1+spec-1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "32c2555c699578a4f59f0cc68e5116c8d7cabbd45e1409b989d4be085b53f13e"
+checksum = "9b320e741db58cac564e26c607d3cc1fdc4a88fd36c879568c07856ed83ff3e9"
 dependencies = [
  "serde_core",
 ]
@@ -7151,23 +7151,23 @@ dependencies = [
 
 [[package]]
 name = "toml_edit"
-version = "0.25.4+spec-1.1.0"
+version = "0.25.5+spec-1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7193cbd0ce53dc966037f54351dbbcf0d5a642c7f0038c382ef9e677ce8c13f2"
+checksum = "8ca1a40644a28bce036923f6a431df0b34236949d111cc07cb6dca830c9ef2e1"
 dependencies = [
  "indexmap",
- "toml_datetime 1.0.0+spec-1.1.0",
+ "toml_datetime 1.0.1+spec-1.1.0",
  "toml_parser",
- "winnow 0.7.15",
+ "winnow 1.0.0",
 ]
 
 [[package]]
 name = "toml_parser"
-version = "1.0.9+spec-1.1.0"
+version = "1.0.10+spec-1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "702d4415e08923e7e1ef96cd5727c0dfed80b4d2fa25db9647fe5eb6f7c5a4c4"
+checksum = "7df25b4befd31c4816df190124375d5a20c6b6921e2cad937316de3fccd63420"
 dependencies = [
- "winnow 0.7.15",
+ "winnow 1.0.0",
 ]
 
 [[package]]
@@ -7178,9 +7178,9 @@ checksum = "5d99f8c9a7727884afe522e9bd5edbfc91a3312b36a77b5fb8926e4c31a41801"
 
 [[package]]
 name = "toml_writer"
-version = "1.0.6+spec-1.1.0"
+version = "1.0.7+spec-1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ab16f14aed21ee8bfd8ec22513f7287cd4a91aa92e44edfe2c17ddd004e92607"
+checksum = "f17aaa1c6e3dc22b1da4b6bba97d066e354c7945cac2f7852d4e4e7ca7a6b56d"
 
 [[package]]
 name = "tonic"
@@ -8894,6 +8894,15 @@ dependencies = [
  "memchr",
 ]
 
+[[package]]
+name = "winnow"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a90e88e4667264a994d34e6d1ab2d26d398dcdca8b7f52bec8668957517fc7d8"
+dependencies = [
+ "memchr",
+]
+
 [[package]]
 name = "winx"
 version = "0.36.4"
@@ -9149,7 +9158,7 @@ dependencies = [
  "thiserror 2.0.18",
  "tokio",
  "tokio-test",
- "toml 1.0.6+spec-1.1.0",
+ "toml 1.0.7+spec-1.1.0",
  "tracing",
 ]
 
@@ -9228,7 +9237,7 @@ dependencies = [
  "tokio-stream",
  "tokio-tungstenite 0.28.0",
  "tokio-util",
- "toml 1.0.6+spec-1.1.0",
+ "toml 1.0.7+spec-1.1.0",
  "tower",
  "tower-http",
  "tracing",

Cargo.toml

@@ -14,15 +14,6 @@ readme = "README.md"
 keywords = ["ai", "agent", "cli", "assistant", "chatbot"]
 categories = ["command-line-utilities", "api-bindings"]
 rust-version = "1.87"
-
-[[bin]]
-name = "zeroclaw"
-path = "src/main.rs"
-
-[lib]
-name = "zeroclaw"
-path = "src/lib.rs"
-
 include = [
   "/src/**/*",
   "/build.rs",
@@ -31,8 +22,17 @@ include = [
   "/LICENSE*",
   "/README.md",
   "/web/dist/**/*",
+  "/tool_descriptions/**/*",
 ]
 
+[[bin]]
+name = "zeroclaw"
+path = "src/main.rs"
+
+[lib]
+name = "zeroclaw"
+path = "src/lib.rs"
+
 [dependencies]
 # CLI - minimal and fast
 clap = { version = "4.5", features = ["derive"] }
@@ -215,7 +215,7 @@ landlock = { version = "0.4", optional = true }
 libc = "0.2"
 
 [features]
-default = ["observability-prometheus", "channel-nostr"]
+default = ["observability-prometheus", "channel-nostr", "skill-creation"]
 channel-nostr = ["dep:nostr-sdk"]
 hardware = ["nusb", "tokio-serial"]
 channel-matrix = ["dep:matrix-sdk"]
@@ -240,6 +240,8 @@ metrics = ["observability-prometheus"]
 probe = ["dep:probe-rs"]
 # rag-pdf = PDF ingestion for datasheet RAG
 rag-pdf = ["dep:pdf-extract"]
+# skill-creation = Autonomous skill creation from successful multi-step tasks
+skill-creation = []
 # whatsapp-web = Native WhatsApp Web client with custom rusqlite storage backend
 whatsapp-web = ["dep:wa-rs", "dep:wa-rs-core", "dep:wa-rs-binary", "dep:wa-rs-proto", "dep:wa-rs-ureq-http", "dep:wa-rs-tokio-transport", "dep:serde-big-array", "dep:prost", "dep:qrcode"]
 # WASM plugin system (extism-based)

docs/reference/api/config-reference.md

@@ -183,6 +183,8 @@ Delegate sub-agent configurations. Each key under `[agents]` defines a named sub
 | `agentic` | `false` | Enable multi-turn tool-call loop mode for the sub-agent |
 | `allowed_tools` | `[]` | Tool allowlist for agentic mode |
 | `max_iterations` | `10` | Max tool-call iterations for agentic mode |
+| `timeout_secs` | `120` | Timeout in seconds for non-agentic provider calls (1–3600) |
+| `agentic_timeout_secs` | `300` | Timeout in seconds for agentic sub-agent loops (1–3600) |
 
 Notes:
 
@@ -199,11 +201,13 @@ max_depth = 2
 agentic = true
 allowed_tools = ["web_search", "http_request", "file_read"]
 max_iterations = 8
+agentic_timeout_secs = 600
 
 [agents.coder]
 provider = "ollama"
 model = "qwen2.5-coder:32b"
 temperature = 0.2
+timeout_secs = 60
 ```
 
 ## `[runtime]`

install.sh

@@ -1359,6 +1359,12 @@ if [[ -n "$TARGET_VERSION" ]]; then
   step_dot "Installing ZeroClaw v${TARGET_VERSION}"
 fi
 if [[ "$SKIP_BUILD" == false ]]; then
+  # Clean stale build artifacts on upgrade to prevent bindgen/build-script
+  # cache mismatches (e.g. libsqlite3-sys bindgen.rs not found).
+  if [[ "$INSTALL_MODE" == "upgrade" && -d "$WORK_DIR/target/release/build" ]]; then
+    step_dot "Cleaning stale build cache (upgrade detected)"
+    cargo clean --release 2>/dev/null || true
+  fi
   step_dot "Building release binary"
   cargo build --release --locked
   step_ok "Release binary built"

src/agent/agent.rs

@@ -4,6 +4,7 @@ use crate::agent::dispatcher::{
 use crate::agent::memory_loader::{DefaultMemoryLoader, MemoryLoader};
 use crate::agent::prompt::{PromptContext, SystemPromptBuilder};
 use crate::config::Config;
+use crate::i18n::ToolDescriptions;
 use crate::memory::{self, Memory, MemoryCategory};
 use crate::observability::{self, Observer, ObserverEvent};
 use crate::providers::{self, ChatMessage, ChatRequest, ConversationMessage, Provider};
@@ -40,6 +41,7 @@ pub struct Agent {
     route_model_by_hint: HashMap<String, String>,
     allowed_tools: Option<Vec<String>>,
     response_cache: Option<Arc<crate::memory::response_cache::ResponseCache>>,
+    tool_descriptions: Option<ToolDescriptions>,
 }
 
 pub struct AgentBuilder {
@@ -64,6 +66,7 @@ pub struct AgentBuilder {
     route_model_by_hint: Option<HashMap<String, String>>,
     allowed_tools: Option<Vec<String>>,
     response_cache: Option<Arc<crate::memory::response_cache::ResponseCache>>,
+    tool_descriptions: Option<ToolDescriptions>,
 }
 
 impl AgentBuilder {
@@ -90,6 +93,7 @@ impl AgentBuilder {
             route_model_by_hint: None,
             allowed_tools: None,
             response_cache: None,
+            tool_descriptions: None,
         }
     }
 
@@ -207,6 +211,11 @@ impl AgentBuilder {
         self
     }
 
+    pub fn tool_descriptions(mut self, tool_descriptions: Option<ToolDescriptions>) -> Self {
+        self.tool_descriptions = tool_descriptions;
+        self
+    }
+
     pub fn build(self) -> Result<Agent> {
         let mut tools = self
             .tools
@@ -257,6 +266,7 @@ impl AgentBuilder {
             route_model_by_hint: self.route_model_by_hint.unwrap_or_default(),
             allowed_tools: allowed,
             response_cache: self.response_cache,
+            tool_descriptions: self.tool_descriptions,
         })
     }
 }
@@ -456,6 +466,7 @@ impl Agent {
             skills_prompt_mode: self.skills_prompt_mode,
             identity_config: Some(&self.identity_config),
             dispatcher_instructions: &instructions,
+            tool_descriptions: self.tool_descriptions.as_ref(),
         };
         self.prompt_builder.build(&ctx)
     }

src/agent/loop_.rs

@@ -1,5 +1,6 @@
 use crate::approval::{ApprovalManager, ApprovalRequest, ApprovalResponse};
 use crate::config::Config;
+use crate::i18n::ToolDescriptions;
 use crate::memory::{self, Memory, MemoryCategory};
 use crate::multimodal;
 use crate::observability::{self, runtime_trace, Observer, ObserverEvent};
@@ -2421,9 +2422,10 @@ pub(crate) async fn run_tool_call_loop(
     };
 
     let turn_id = Uuid::new_v4().to_string();
-    let mut seen_tool_signatures: HashSet<(String, String)> = HashSet::new();
 
     for iteration in 0..max_iterations {
+        let mut seen_tool_signatures: HashSet<(String, String)> = HashSet::new();
+
         if cancellation_token
             .as_ref()
             .is_some_and(CancellationToken::is_cancelled)
@@ -3077,7 +3079,10 @@ pub(crate) async fn run_tool_call_loop(
 
 /// Build the tool instruction block for the system prompt so the LLM knows
 /// how to invoke tools.
-pub(crate) fn build_tool_instructions(tools_registry: &[Box<dyn Tool>]) -> String {
+pub(crate) fn build_tool_instructions(
+    tools_registry: &[Box<dyn Tool>],
+    tool_descriptions: Option<&ToolDescriptions>,
+) -> String {
     let mut instructions = String::new();
     instructions.push_str("\n## Tool Use Protocol\n\n");
     instructions.push_str("To use a tool, wrap a JSON object in <tool_call></tool_call> tags:\n\n");
@@ -3093,11 +3098,14 @@ pub(crate) fn build_tool_instructions(tools_registry: &[Box<dyn Tool>]) -> Strin
     instructions.push_str("### Available Tools\n\n");
 
     for tool in tools_registry {
+        let desc = tool_descriptions
+            .and_then(|td| td.get(tool.name()))
+            .unwrap_or_else(|| tool.description());
         let _ = writeln!(
             instructions,
             "**{}**: {}\nParameters: `{}`\n",
             tool.name(),
-            tool.description(),
+            desc,
             tool.parameters_schema()
         );
     }
@@ -3323,6 +3331,16 @@ pub async fn run(
         .map(|b| b.board.clone())
         .collect();
 
+    // ── Load locale-aware tool descriptions ────────────────────────
+    let i18n_locale = config
+        .locale
+        .as_deref()
+        .filter(|s| !s.is_empty())
+        .map(ToString::to_string)
+        .unwrap_or_else(crate::i18n::detect_locale);
+    let i18n_search_dirs = crate::i18n::default_search_dirs(&config.workspace_dir);
+    let i18n_descs = crate::i18n::ToolDescriptions::load(&i18n_locale, &i18n_search_dirs);
+
     // ── Build system prompt from workspace MD files (OpenClaw framework) ──
     let skills = crate::skills::load_skills_with_config(&config.workspace_dir, &config);
     let mut tool_descs: Vec<(&str, &str)> = vec![
@@ -3452,7 +3470,7 @@ pub async fn run(
 
     // Append structured tool-use instructions with schemas (only for non-native providers)
     if !native_tools {
-        system_prompt.push_str(&build_tool_instructions(&tools_registry));
+        system_prompt.push_str(&build_tool_instructions(&tools_registry, Some(&i18n_descs)));
     }
 
     // Append deferred MCP tool names so the LLM knows what is available
@@ -3590,6 +3608,27 @@ pub async fn run(
                 }
             }
         }
+
+        // After successful multi-step execution, attempt autonomous skill creation.
+        #[cfg(feature = "skill-creation")]
+        if config.skills.skill_creation.enabled {
+            let tool_calls = crate::skills::creator::extract_tool_calls_from_history(&history);
+            if tool_calls.len() >= 2 {
+                let creator = crate::skills::creator::SkillCreator::new(
+                    config.workspace_dir.clone(),
+                    config.skills.skill_creation.clone(),
+                );
+                match creator.create_from_execution(&msg, &tool_calls, None).await {
+                    Ok(Some(slug)) => {
+                        tracing::info!(slug, "Auto-created skill from execution");
+                    }
+                    Ok(None) => {
+                        tracing::debug!("Skill creation skipped (duplicate or disabled)");
+                    }
+                    Err(e) => tracing::warn!("Skill creation failed: {e}"),
+                }
+            }
+        }
         final_output = response.clone();
         println!("{response}");
         observer.record_event(&ObserverEvent::TurnComplete);
@@ -3988,6 +4027,16 @@ pub async fn process_message(
         .map(|b| b.board.clone())
         .collect();
 
+    // ── Load locale-aware tool descriptions ────────────────────────
+    let i18n_locale = config
+        .locale
+        .as_deref()
+        .filter(|s| !s.is_empty())
+        .map(ToString::to_string)
+        .unwrap_or_else(crate::i18n::detect_locale);
+    let i18n_search_dirs = crate::i18n::default_search_dirs(&config.workspace_dir);
+    let i18n_descs = crate::i18n::ToolDescriptions::load(&i18n_locale, &i18n_search_dirs);
+
     let skills = crate::skills::load_skills_with_config(&config.workspace_dir, &config);
     let mut tool_descs: Vec<(&str, &str)> = vec![
         ("shell", "Execute terminal commands."),
@@ -4053,7 +4102,7 @@ pub async fn process_message(
         config.skills.prompt_injection_mode,
     );
     if !native_tools {
-        system_prompt.push_str(&build_tool_instructions(&tools_registry));
+        system_prompt.push_str(&build_tool_instructions(&tools_registry, Some(&i18n_descs)));
     }
     if !deferred_section.is_empty() {
         system_prompt.push('\n');
@@ -5763,7 +5812,7 @@ Tail"#;
             std::path::Path::new("/tmp"),
         ));
         let tools = tools::default_tools(security);
-        let instructions = build_tool_instructions(&tools);
+        let instructions = build_tool_instructions(&tools, None);
 
         assert!(instructions.contains("## Tool Use Protocol"));
         assert!(instructions.contains("<tool_call>"));

src/agent/prompt.rs

@@ -1,4 +1,5 @@
 use crate::config::IdentityConfig;
+use crate::i18n::ToolDescriptions;
 use crate::identity;
 use crate::skills::Skill;
 use crate::tools::Tool;
@@ -17,6 +18,9 @@ pub struct PromptContext<'a> {
     pub skills_prompt_mode: crate::config::SkillsPromptInjectionMode,
     pub identity_config: Option<&'a IdentityConfig>,
     pub dispatcher_instructions: &'a str,
+    /// Locale-aware tool descriptions. When present, tool descriptions in
+    /// prompts are resolved from the locale file instead of hardcoded values.
+    pub tool_descriptions: Option<&'a ToolDescriptions>,
 }
 
 pub trait PromptSection: Send + Sync {
@@ -124,11 +128,15 @@ impl PromptSection for ToolsSection {
     fn build(&self, ctx: &PromptContext<'_>) -> Result<String> {
         let mut out = String::from("## Tools\n\n");
         for tool in ctx.tools {
+            let desc = ctx
+                .tool_descriptions
+                .and_then(|td: &ToolDescriptions| td.get(tool.name()))
+                .unwrap_or_else(|| tool.description());
             let _ = writeln!(
                 out,
                 "- **{}**: {}\n  Parameters: `{}`",
                 tool.name(),
-                tool.description(),
+                desc,
                 tool.parameters_schema()
             );
         }
@@ -317,6 +325,7 @@ mod tests {
             skills_prompt_mode: crate::config::SkillsPromptInjectionMode::Full,
             identity_config: Some(&identity_config),
             dispatcher_instructions: "",
+            tool_descriptions: None,
         };
 
         let section = IdentitySection;
@@ -345,6 +354,7 @@ mod tests {
             skills_prompt_mode: crate::config::SkillsPromptInjectionMode::Full,
             identity_config: None,
             dispatcher_instructions: "instr",
+            tool_descriptions: None,
         };
         let prompt = SystemPromptBuilder::with_defaults().build(&ctx).unwrap();
         assert!(prompt.contains("## Tools"));
@@ -380,6 +390,7 @@ mod tests {
             skills_prompt_mode: crate::config::SkillsPromptInjectionMode::Full,
             identity_config: None,
             dispatcher_instructions: "",
+            tool_descriptions: None,
         };
 
         let output = SkillsSection.build(&ctx).unwrap();
@@ -418,6 +429,7 @@ mod tests {
             skills_prompt_mode: crate::config::SkillsPromptInjectionMode::Compact,
             identity_config: None,
             dispatcher_instructions: "",
+            tool_descriptions: None,
         };
 
         let output = SkillsSection.build(&ctx).unwrap();
@@ -439,6 +451,7 @@ mod tests {
             skills_prompt_mode: crate::config::SkillsPromptInjectionMode::Full,
             identity_config: None,
             dispatcher_instructions: "instr",
+            tool_descriptions: None,
         };
 
         let rendered = DateTimeSection.build(&ctx).unwrap();
@@ -477,6 +490,7 @@ mod tests {
             skills_prompt_mode: crate::config::SkillsPromptInjectionMode::Full,
             identity_config: None,
             dispatcher_instructions: "",
+            tool_descriptions: None,
         };
 
         let prompt = SystemPromptBuilder::with_defaults().build(&ctx).unwrap();

src/channels/mod.rs

@@ -3229,12 +3229,16 @@ fn build_channel_by_id(config: &Config, channel_id: &str) -> Result<Arc<dyn Chan
                 .telegram
                 .as_ref()
                 .context("Telegram channel is not configured")?;
+            let ack = tg
+                .ack_reactions
+                .unwrap_or(config.channels_config.ack_reactions);
             Ok(Arc::new(
                 TelegramChannel::new(
                     tg.bot_token.clone(),
                     tg.allowed_users.clone(),
                     tg.mention_only,
                 )
+                .with_ack_reactions(ack)
                 .with_streaming(tg.stream_mode, tg.draft_update_interval_ms)
                 .with_transcription(config.transcription.clone())
                 .with_workspace_dir(config.workspace_dir.clone()),
@@ -3322,6 +3326,9 @@ fn collect_configured_channels(
     let mut channels = Vec::new();
 
     if let Some(ref tg) = config.channels_config.telegram {
+        let ack = tg
+            .ack_reactions
+            .unwrap_or(config.channels_config.ack_reactions);
         channels.push(ConfiguredChannel {
             display_name: "Telegram",
             channel: Arc::new(
@@ -3330,6 +3337,7 @@ fn collect_configured_channels(
                     tg.allowed_users.clone(),
                     tg.mention_only,
                 )
+                .with_ack_reactions(ack)
                 .with_streaming(tg.stream_mode, tg.draft_update_interval_ms)
                 .with_transcription(config.transcription.clone())
                 .with_workspace_dir(config.workspace_dir.clone()),
@@ -3931,6 +3939,16 @@ pub async fn start_channels(config: Config) -> Result<()> {
 
     let skills = crate::skills::load_skills_with_config(&workspace, &config);
 
+    // ── Load locale-aware tool descriptions ────────────────────────
+    let i18n_locale = config
+        .locale
+        .as_deref()
+        .filter(|s| !s.is_empty())
+        .map(ToString::to_string)
+        .unwrap_or_else(crate::i18n::detect_locale);
+    let i18n_search_dirs = crate::i18n::default_search_dirs(&workspace);
+    let i18n_descs = crate::i18n::ToolDescriptions::load(&i18n_locale, &i18n_search_dirs);
+
     // Collect tool descriptions for the prompt
     let mut tool_descs: Vec<(&str, &str)> = vec![
         (
@@ -4010,7 +4028,10 @@ pub async fn start_channels(config: Config) -> Result<()> {
         config.skills.prompt_injection_mode,
     );
     if !native_tools {
-        system_prompt.push_str(&build_tool_instructions(tools_registry.as_ref()));
+        system_prompt.push_str(&build_tool_instructions(
+            tools_registry.as_ref(),
+            Some(&i18n_descs),
+        ));
     }
 
     // Append deferred MCP tool names so the LLM knows what is available
@@ -6760,7 +6781,7 @@ BTC is currently around $65,000 based on latest tool output."#
             "build_system_prompt should not emit protocol block directly"
         );
 
-        prompt.push_str(&build_tool_instructions(&[]));
+        prompt.push_str(&build_tool_instructions(&[], None));
 
         assert_eq!(
             prompt.matches("## Tool Use Protocol").count(),
@@ -8668,6 +8689,7 @@ This is an example JSON object for profile settings."#;
             draft_update_interval_ms: 1000,
             interrupt_on_new_message: false,
             mention_only: false,
+            ack_reactions: None,
         });
         match build_channel_by_id(&config, "telegram") {
             Ok(channel) => assert_eq!(channel.name(), "telegram"),

src/channels/telegram.rs

@@ -332,6 +332,7 @@ pub struct TelegramChannel {
     transcription: Option<crate::config::TranscriptionConfig>,
     voice_transcriptions: Mutex<std::collections::HashMap<String, String>>,
     workspace_dir: Option<std::path::PathBuf>,
+    ack_reactions: bool,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
@@ -370,9 +371,16 @@ impl TelegramChannel {
             transcription: None,
             voice_transcriptions: Mutex::new(std::collections::HashMap::new()),
             workspace_dir: None,
+            ack_reactions: true,
         }
     }
 
+    /// Configure whether Telegram-native acknowledgement reactions are sent.
+    pub fn with_ack_reactions(mut self, enabled: bool) -> Self {
+        self.ack_reactions = enabled;
+        self
+    }
+
     /// Configure workspace directory for saving downloaded attachments.
     pub fn with_workspace_dir(mut self, dir: std::path::PathBuf) -> Self {
         self.workspace_dir = Some(dir);
@@ -2689,13 +2697,15 @@ Ensure only one `zeroclaw` process is using this bot token."
                         continue;
                     };
 
-                    if let Some((reaction_chat_id, reaction_message_id)) =
-                        Self::extract_update_message_target(update)
-                    {
-                        self.try_add_ack_reaction_nonblocking(
-                            reaction_chat_id,
-                            reaction_message_id,
-                        );
+                    if self.ack_reactions {
+                        if let Some((reaction_chat_id, reaction_message_id)) =
+                            Self::extract_update_message_target(update)
+                        {
+                            self.try_add_ack_reaction_nonblocking(
+                                reaction_chat_id,
+                                reaction_message_id,
+                            );
+                        }
                     }
 
                     // Send "typing" indicator immediately when we receive a message
@@ -4681,4 +4691,24 @@ mod tests {
         // the agent loop will return ProviderCapabilityError before calling
         // the provider, and the channel will send "⚠️ Error: ..." to the user.
     }
+
+    #[test]
+    fn ack_reactions_defaults_to_true() {
+        let ch = TelegramChannel::new("token".into(), vec!["*".into()], false);
+        assert!(ch.ack_reactions);
+    }
+
+    #[test]
+    fn with_ack_reactions_false_disables_reactions() {
+        let ch =
+            TelegramChannel::new("token".into(), vec!["*".into()], false).with_ack_reactions(false);
+        assert!(!ch.ack_reactions);
+    }
+
+    #[test]
+    fn with_ack_reactions_true_keeps_reactions() {
+        let ch =
+            TelegramChannel::new("token".into(), vec!["*".into()], false).with_ack_reactions(true);
+        assert!(ch.ack_reactions);
+    }
 }

src/config/mod.rs

@@ -22,11 +22,11 @@ pub use schema::{
     OtpMethod, PeripheralBoardConfig, PeripheralsConfig, PluginsConfig, ProjectIntelConfig,
     ProxyConfig, ProxyScope, QdrantConfig, QueryClassificationConfig, ReliabilityConfig,
     ResourceLimitsConfig, RuntimeConfig, SandboxBackend, SandboxConfig, SchedulerConfig,
-    SecretsConfig, SecurityConfig, SecurityOpsConfig, SkillsConfig, SkillsPromptInjectionMode,
-    SlackConfig, StorageConfig, StorageProviderConfig, StorageProviderSection, StreamMode,
-    SwarmConfig, SwarmStrategy, TelegramConfig, ToolFilterGroup, ToolFilterGroupMode,
-    TranscriptionConfig, TtsConfig, TunnelConfig, WebFetchConfig, WebSearchConfig, WebhookConfig,
-    WorkspaceConfig,
+    SecretsConfig, SecurityConfig, SecurityOpsConfig, SkillCreationConfig, SkillsConfig,
+    SkillsPromptInjectionMode, SlackConfig, StorageConfig, StorageProviderConfig,
+    StorageProviderSection, StreamMode, SwarmConfig, SwarmStrategy, TelegramConfig,
+    ToolFilterGroup, ToolFilterGroupMode, TranscriptionConfig, TtsConfig, TunnelConfig,
+    WebFetchConfig, WebSearchConfig, WebhookConfig, WorkspaceConfig,
 };
 
 pub fn name_and_presence<T: traits::ChannelConfig>(channel: Option<&T>) -> (&'static str, bool) {
@@ -55,6 +55,7 @@ mod tests {
             draft_update_interval_ms: 1000,
             interrupt_on_new_message: false,
             mention_only: false,
+            ack_reactions: None,
         };
 
         let discord = DiscordConfig {

src/config/schema.rs

@@ -339,6 +339,17 @@ pub struct Config {
     /// Plugin system configuration (`[plugins]`).
     #[serde(default)]
     pub plugins: PluginsConfig,
+
+    /// Locale for tool descriptions (e.g. `"en"`, `"zh-CN"`).
+    ///
+    /// When set, tool descriptions shown in system prompts are loaded from
+    /// `tool_descriptions/<locale>.toml`. Falls back to English, then to
+    /// hardcoded descriptions.
+    ///
+    /// If omitted or empty, the locale is auto-detected from `ZEROCLAW_LOCALE`,
+    /// `LANG`, or `LC_ALL` environment variables (defaulting to `"en"`).
+    #[serde(default)]
+    pub locale: Option<String>,
 }
 
 /// Multi-client workspace isolation configuration.
@@ -449,6 +460,14 @@ pub struct DelegateAgentConfig {
     /// Maximum tool-call iterations in agentic mode.
     #[serde(default = "default_max_tool_iterations")]
     pub max_iterations: usize,
+    /// Timeout in seconds for non-agentic provider calls.
+    /// Defaults to 120 when unset. Must be between 1 and 3600.
+    #[serde(default)]
+    pub timeout_secs: Option<u64>,
+    /// Timeout in seconds for agentic sub-agent loops.
+    /// Defaults to 300 when unset. Must be between 1 and 3600.
+    #[serde(default)]
+    pub agentic_timeout_secs: Option<u64>,
 }
 
 // ── Swarms ──────────────────────────────────────────────────────
@@ -1163,6 +1182,34 @@ pub struct SkillsConfig {
     /// `full` preserves legacy behavior. `compact` keeps context small and loads skills on demand.
     #[serde(default)]
     pub prompt_injection_mode: SkillsPromptInjectionMode,
+    /// Autonomous skill creation from successful multi-step task executions.
+    #[serde(default)]
+    pub skill_creation: SkillCreationConfig,
+}
+
+/// Autonomous skill creation configuration (`[skills.skill_creation]` section).
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(default)]
+pub struct SkillCreationConfig {
+    /// Enable automatic skill creation after successful multi-step tasks.
+    /// Default: `false`.
+    pub enabled: bool,
+    /// Maximum number of auto-generated skills to keep.
+    /// When exceeded, the oldest auto-generated skill is removed (LRU eviction).
+    pub max_skills: usize,
+    /// Embedding similarity threshold for deduplication.
+    /// Skills with descriptions more similar than this value are skipped.
+    pub similarity_threshold: f64,
+}
+
+impl Default for SkillCreationConfig {
+    fn default() -> Self {
+        Self {
+            enabled: false,
+            max_skills: 500,
+            similarity_threshold: 0.85,
+        }
+    }
 }
 
 /// Multimodal (image) handling configuration (`[multimodal]` section).
@@ -4511,6 +4558,11 @@ pub struct TelegramConfig {
     /// Direct messages are always processed.
     #[serde(default)]
     pub mention_only: bool,
+    /// Override for the top-level `ack_reactions` setting. When `None`, the
+    /// channel falls back to `[channels_config].ack_reactions`. When set
+    /// explicitly, it takes precedence.
+    #[serde(default)]
+    pub ack_reactions: Option<bool>,
 }
 
 impl ChannelConfig for TelegramConfig {
@@ -5083,6 +5135,10 @@ pub struct OtpConfig {
     /// Domain-category presets expanded into `gated_domains`.
     #[serde(default)]
     pub gated_domain_categories: Vec<String>,
+
+    /// Maximum number of OTP challenge attempts before lockout.
+    #[serde(default = "default_otp_challenge_max_attempts")]
+    pub challenge_max_attempts: u32,
 }
 
 fn default_otp_token_ttl_secs() -> u64 {
@@ -5093,6 +5149,10 @@ fn default_otp_cache_valid_secs() -> u64 {
     300
 }
 
+fn default_otp_challenge_max_attempts() -> u32 {
+    3
+}
+
 fn default_otp_gated_actions() -> Vec<String> {
     vec![
         "shell".to_string(),
@@ -5113,6 +5173,7 @@ impl Default for OtpConfig {
             gated_actions: default_otp_gated_actions(),
             gated_domains: Vec::new(),
             gated_domain_categories: Vec::new(),
+            challenge_max_attempts: default_otp_challenge_max_attempts(),
         }
     }
 }
@@ -5991,6 +6052,7 @@ impl Default for Config {
             knowledge: KnowledgeConfig::default(),
             linkedin: LinkedInConfig::default(),
             plugins: PluginsConfig::default(),
+            locale: None,
         }
     }
 }
@@ -6400,6 +6462,45 @@ fn read_codex_openai_api_key() -> Option<String> {
         .map(ToString::to_string)
 }
 
+/// Ensure that essential bootstrap files exist in the workspace directory.
+///
+/// When the workspace is created outside of `zeroclaw onboard` (e.g., non-tty
+/// daemon/cron sessions), these files would otherwise be missing. This function
+/// creates sensible defaults that allow the agent to operate with a basic identity.
+async fn ensure_bootstrap_files(workspace_dir: &Path) -> Result<()> {
+    let defaults: &[(&str, &str)] = &[
+        (
+            "IDENTITY.md",
+            "# IDENTITY.md — Who Am I?\n\n\
+             I am ZeroClaw, an autonomous AI agent.\n\n\
+             ## Traits\n\
+             - Helpful, precise, and safety-conscious\n\
+             - I prioritize clarity and correctness\n",
+        ),
+        (
+            "SOUL.md",
+            "# SOUL.md — Who You Are\n\n\
+             You are ZeroClaw, an autonomous AI agent.\n\n\
+             ## Core Principles\n\
+             - Be helpful and accurate\n\
+             - Respect user intent and boundaries\n\
+             - Ask before taking destructive actions\n\
+             - Prefer safe, reversible operations\n",
+        ),
+    ];
+
+    for (filename, content) in defaults {
+        let path = workspace_dir.join(filename);
+        if !path.exists() {
+            fs::write(&path, content)
+                .await
+                .with_context(|| format!("Failed to create default {filename} in workspace"))?;
+        }
+    }
+
+    Ok(())
+}
+
 impl Config {
     pub async fn load_or_init() -> Result<Self> {
         let (default_zeroclaw_dir, default_workspace_dir) = default_config_and_workspace_dirs()?;
@@ -6416,6 +6517,8 @@ impl Config {
             .await
             .context("Failed to create workspace directory")?;
 
+        ensure_bootstrap_files(&workspace_dir).await?;
+
         if config_path.exists() {
             // Warn if config file is world-readable (may contain API keys)
             #[cfg(unix)]
@@ -6942,6 +7045,9 @@ impl Config {
         }
 
         // Security OTP / estop
+        if self.security.otp.challenge_max_attempts == 0 {
+            anyhow::bail!("security.otp.challenge_max_attempts must be greater than 0");
+        }
         if self.security.otp.token_ttl_secs == 0 {
             anyhow::bail!("security.otp.token_ttl_secs must be greater than 0");
         }
@@ -7252,6 +7358,31 @@ impl Config {
             anyhow::bail!("security.nevis: {msg}");
         }
 
+        // Delegate agent timeouts
+        const MAX_DELEGATE_TIMEOUT_SECS: u64 = 3600;
+        for (name, agent) in &self.agents {
+            if let Some(timeout) = agent.timeout_secs {
+                if timeout == 0 {
+                    anyhow::bail!("agents.{name}.timeout_secs must be greater than 0");
+                }
+                if timeout > MAX_DELEGATE_TIMEOUT_SECS {
+                    anyhow::bail!(
+                        "agents.{name}.timeout_secs exceeds max {MAX_DELEGATE_TIMEOUT_SECS}"
+                    );
+                }
+            }
+            if let Some(timeout) = agent.agentic_timeout_secs {
+                if timeout == 0 {
+                    anyhow::bail!("agents.{name}.agentic_timeout_secs must be greater than 0");
+                }
+                if timeout > MAX_DELEGATE_TIMEOUT_SECS {
+                    anyhow::bail!(
+                        "agents.{name}.agentic_timeout_secs exceeds max {MAX_DELEGATE_TIMEOUT_SECS}"
+                    );
+                }
+            }
+        }
+
         // Transcription
         {
             let dp = self.transcription.default_provider.trim();
@@ -8360,6 +8491,7 @@ default_temperature = 0.7
                     draft_update_interval_ms: default_draft_update_interval_ms(),
                     interrupt_on_new_message: false,
                     mention_only: false,
+                    ack_reactions: None,
                 }),
                 discord: None,
                 slack: None,
@@ -8427,6 +8559,7 @@ default_temperature = 0.7
             knowledge: KnowledgeConfig::default(),
             linkedin: LinkedInConfig::default(),
             plugins: PluginsConfig::default(),
+            locale: None,
         };
 
         let toml_str = toml::to_string_pretty(&config).unwrap();
@@ -8760,6 +8893,7 @@ tool_dispatcher = "xml"
             knowledge: KnowledgeConfig::default(),
             linkedin: LinkedInConfig::default(),
             plugins: PluginsConfig::default(),
+            locale: None,
         };
 
         config.save().await.unwrap();
@@ -8818,6 +8952,8 @@ tool_dispatcher = "xml"
                 agentic: false,
                 allowed_tools: Vec::new(),
                 max_iterations: 10,
+                timeout_secs: None,
+                agentic_timeout_secs: None,
             },
         );
 
@@ -8942,6 +9078,7 @@ tool_dispatcher = "xml"
             draft_update_interval_ms: 500,
             interrupt_on_new_message: true,
             mention_only: false,
+            ack_reactions: None,
         };
         let json = serde_json::to_string(&tc).unwrap();
         let parsed: TelegramConfig = serde_json::from_str(&json).unwrap();
@@ -11256,6 +11393,7 @@ require_otp_to_resume = true
             draft_update_interval_ms: default_draft_update_interval_ms(),
             interrupt_on_new_message: false,
             mention_only: false,
+            ack_reactions: None,
         });
 
         // Save (triggers encryption)
@@ -11811,4 +11949,109 @@ require_otp_to_resume = true
             "Debug output must show [REDACTED] for client_secret"
         );
     }
+
+    #[test]
+    async fn telegram_config_ack_reactions_false_deserializes() {
+        let toml_str = r#"
+            bot_token = "123:ABC"
+            allowed_users = ["alice"]
+            ack_reactions = false
+        "#;
+        let cfg: TelegramConfig = toml::from_str(toml_str).unwrap();
+        assert_eq!(cfg.ack_reactions, Some(false));
+    }
+
+    #[test]
+    async fn telegram_config_ack_reactions_true_deserializes() {
+        let toml_str = r#"
+            bot_token = "123:ABC"
+            allowed_users = ["alice"]
+            ack_reactions = true
+        "#;
+        let cfg: TelegramConfig = toml::from_str(toml_str).unwrap();
+        assert_eq!(cfg.ack_reactions, Some(true));
+    }
+
+    #[test]
+    async fn telegram_config_ack_reactions_missing_defaults_to_none() {
+        let toml_str = r#"
+            bot_token = "123:ABC"
+            allowed_users = ["alice"]
+        "#;
+        let cfg: TelegramConfig = toml::from_str(toml_str).unwrap();
+        assert_eq!(cfg.ack_reactions, None);
+    }
+
+    #[test]
+    async fn telegram_config_ack_reactions_channel_overrides_top_level() {
+        let tg_toml = r#"
+            bot_token = "123:ABC"
+            allowed_users = ["alice"]
+            ack_reactions = false
+        "#;
+        let tg: TelegramConfig = toml::from_str(tg_toml).unwrap();
+        let top_level_ack = true;
+        let effective = tg.ack_reactions.unwrap_or(top_level_ack);
+        assert!(
+            !effective,
+            "channel-level false must override top-level true"
+        );
+    }
+
+    #[test]
+    async fn telegram_config_ack_reactions_falls_back_to_top_level() {
+        let tg_toml = r#"
+            bot_token = "123:ABC"
+            allowed_users = ["alice"]
+        "#;
+        let tg: TelegramConfig = toml::from_str(tg_toml).unwrap();
+        let top_level_ack = false;
+        let effective = tg.ack_reactions.unwrap_or(top_level_ack);
+        assert!(
+            !effective,
+            "must fall back to top-level false when channel omits field"
+        );
+    }
+
+    // ── Bootstrap files ─────────────────────────────────────
+
+    #[test]
+    async fn ensure_bootstrap_files_creates_missing_files() {
+        let tmp = TempDir::new().unwrap();
+        let ws = tmp.path().join("workspace");
+        tokio::fs::create_dir_all(&ws).await.unwrap();
+
+        ensure_bootstrap_files(&ws).await.unwrap();
+
+        let soul = tokio::fs::read_to_string(ws.join("SOUL.md")).await.unwrap();
+        let identity = tokio::fs::read_to_string(ws.join("IDENTITY.md"))
+            .await
+            .unwrap();
+        assert!(soul.contains("SOUL.md"));
+        assert!(identity.contains("IDENTITY.md"));
+    }
+
+    #[test]
+    async fn ensure_bootstrap_files_does_not_overwrite_existing() {
+        let tmp = TempDir::new().unwrap();
+        let ws = tmp.path().join("workspace");
+        tokio::fs::create_dir_all(&ws).await.unwrap();
+
+        let custom = "# My custom SOUL";
+        tokio::fs::write(ws.join("SOUL.md"), custom).await.unwrap();
+
+        ensure_bootstrap_files(&ws).await.unwrap();
+
+        let soul = tokio::fs::read_to_string(ws.join("SOUL.md")).await.unwrap();
+        assert_eq!(
+            soul, custom,
+            "ensure_bootstrap_files must not overwrite existing files"
+        );
+
+        // IDENTITY.md should still be created since it was missing
+        let identity = tokio::fs::read_to_string(ws.join("IDENTITY.md"))
+            .await
+            .unwrap();
+        assert!(identity.contains("IDENTITY.md"));
+    }
 }

src/daemon/mod.rs

@@ -642,6 +642,7 @@ mod tests {
             draft_update_interval_ms: 1000,
             interrupt_on_new_message: false,
             mention_only: false,
+            ack_reactions: None,
         });
         assert!(has_supervised_channels(&config));
     }
@@ -755,6 +756,7 @@ mod tests {
             draft_update_interval_ms: 1000,
             interrupt_on_new_message: false,
             mention_only: false,
+            ack_reactions: None,
         });
 
         let target = resolve_heartbeat_delivery(&config).unwrap();
@@ -771,6 +773,7 @@ mod tests {
             draft_update_interval_ms: 1000,
             interrupt_on_new_message: false,
             mention_only: false,
+            ack_reactions: None,
         });
 
         let target = resolve_heartbeat_delivery(&config).unwrap();

src/doctor/mod.rs

@@ -1281,6 +1281,8 @@ mod tests {
                 agentic: false,
                 allowed_tools: Vec::new(),
                 max_iterations: 10,
+                timeout_secs: None,
+                agentic_timeout_secs: None,
             },
         );
         config.agents.insert(
@@ -1295,6 +1297,8 @@ mod tests {
                 agentic: false,
                 allowed_tools: Vec::new(),
                 max_iterations: 10,
+                timeout_secs: None,
+                agentic_timeout_secs: None,
             },
         );
 

src/gateway/mod.rs

@@ -633,6 +633,21 @@ pub async fn run_gateway(host: &str, port: u16, config: Config) -> Result<()> {
         println!("  🌐 Public URL: {url}");
     }
     println!("  🌐 Web Dashboard: http://{display_addr}/");
+    if let Some(code) = pairing.pairing_code() {
+        println!();
+        println!("  🔐 PAIRING REQUIRED — use this one-time code:");
+        println!("     ┌──────────────┐");
+        println!("     │  {code}  │");
+        println!("     └──────────────┘");
+        println!();
+    } else if pairing.require_pairing() {
+        println!("  🔒 Pairing: ACTIVE (bearer token required)");
+        println!("     To pair a new device: zeroclaw gateway get-paircode --new");
+        println!();
+    } else {
+        println!("  ⚠️  Pairing: DISABLED (all requests accepted)");
+        println!();
+    }
     println!("  POST /pair      — pair a new client (X-Pairing-Code header)");
     println!("  POST /webhook   — {{\"message\": \"your prompt\"}}");
     if whatsapp_channel.is_some() {
@@ -656,19 +671,6 @@ pub async fn run_gateway(host: &str, port: u16, config: Config) -> Result<()> {
     }
     println!("  GET  /health    — health check");
     println!("  GET  /metrics   — Prometheus metrics");
-    if let Some(code) = pairing.pairing_code() {
-        println!();
-        println!("  🔐 PAIRING REQUIRED — use this one-time code:");
-        println!("     ┌──────────────┐");
-        println!("     │  {code}  │");
-        println!("     └──────────────┘");
-        println!("     Send: POST /pair with header X-Pairing-Code: {code}");
-    } else if pairing.require_pairing() {
-        println!("  🔒 Pairing: ACTIVE (bearer token required)");
-        println!("     To pair a new device: zeroclaw gateway get-paircode --new");
-    } else {
-        println!("  ⚠️  Pairing: DISABLED (all requests accepted)");
-    }
     println!("  Press Ctrl+C to stop.\n");
 
     crate::health::mark_component_ok("gateway");

src/i18n.rs

@@ -0,0 +1,311 @@
+//! Internationalization support for tool descriptions.
+//!
+//! Loads tool descriptions from TOML locale files in `tool_descriptions/`.
+//! Falls back to English when a locale file or specific key is missing,
+//! and ultimately falls back to the hardcoded `tool.description()` value
+//! if no file-based description exists.
+
+use std::collections::HashMap;
+use std::path::{Path, PathBuf};
+use tracing::debug;
+
+/// Container for locale-specific tool descriptions loaded from TOML files.
+#[derive(Debug, Clone)]
+pub struct ToolDescriptions {
+    /// Descriptions from the requested locale (may be empty if file missing).
+    locale_descriptions: HashMap<String, String>,
+    /// English fallback descriptions (always loaded when locale != "en").
+    english_fallback: HashMap<String, String>,
+    /// The resolved locale tag (e.g. "en", "zh-CN").
+    locale: String,
+}
+
+/// TOML structure: `[tools]` table mapping tool name -> description string.
+#[derive(Debug, serde::Deserialize)]
+struct DescriptionFile {
+    #[serde(default)]
+    tools: HashMap<String, String>,
+}
+
+impl ToolDescriptions {
+    /// Load descriptions for the given locale.
+    ///
+    /// `search_dirs` lists directories to probe for `tool_descriptions/<locale>.toml`.
+    /// The first directory containing a matching file wins.
+    ///
+    /// Resolution:
+    /// 1. Look up tool name in the locale file.
+    /// 2. If missing (or locale file absent), look up in `en.toml`.
+    /// 3. If still missing, callers fall back to `tool.description()`.
+    pub fn load(locale: &str, search_dirs: &[PathBuf]) -> Self {
+        let locale_descriptions = load_locale_file(locale, search_dirs);
+
+        let english_fallback = if locale == "en" {
+            HashMap::new()
+        } else {
+            load_locale_file("en", search_dirs)
+        };
+
+        debug!(
+            locale = locale,
+            locale_keys = locale_descriptions.len(),
+            english_keys = english_fallback.len(),
+            "tool descriptions loaded"
+        );
+
+        Self {
+            locale_descriptions,
+            english_fallback,
+            locale: locale.to_string(),
+        }
+    }
+
+    /// Get the description for a tool by name.
+    ///
+    /// Returns `Some(description)` if found in the locale file or English fallback.
+    /// Returns `None` if neither file contains the key (caller should use hardcoded).
+    pub fn get(&self, tool_name: &str) -> Option<&str> {
+        self.locale_descriptions
+            .get(tool_name)
+            .or_else(|| self.english_fallback.get(tool_name))
+            .map(String::as_str)
+    }
+
+    /// The resolved locale tag.
+    pub fn locale(&self) -> &str {
+        &self.locale
+    }
+
+    /// Create an empty instance that always returns `None` (hardcoded fallback).
+    pub fn empty() -> Self {
+        Self {
+            locale_descriptions: HashMap::new(),
+            english_fallback: HashMap::new(),
+            locale: "en".to_string(),
+        }
+    }
+}
+
+/// Detect the user's preferred locale from environment variables.
+///
+/// Checks `ZEROCLAW_LOCALE`, then `LANG`, then `LC_ALL`.
+/// Returns "en" if none are set or parseable.
+pub fn detect_locale() -> String {
+    if let Ok(val) = std::env::var("ZEROCLAW_LOCALE") {
+        let val = val.trim().to_string();
+        if !val.is_empty() {
+            return normalize_locale(&val);
+        }
+    }
+    for var in &["LANG", "LC_ALL"] {
+        if let Ok(val) = std::env::var(var) {
+            let locale = normalize_locale(&val);
+            if locale != "C" && locale != "POSIX" && !locale.is_empty() {
+                return locale;
+            }
+        }
+    }
+    "en".to_string()
+}
+
+/// Normalize a raw locale string (e.g. "zh_CN.UTF-8") to a tag we use
+/// for file lookup (e.g. "zh-CN").
+fn normalize_locale(raw: &str) -> String {
+    // Strip encoding suffix (.UTF-8, .utf8, etc.)
+    let base = raw.split('.').next().unwrap_or(raw);
+    // Replace underscores with hyphens for BCP-47-ish consistency
+    base.replace('_', "-")
+}
+
+/// Build the default set of search directories for locale files.
+///
+/// 1. The workspace directory itself (for project-local overrides).
+/// 2. The binary's parent directory (for installed distributions).
+/// 3. The compile-time `CARGO_MANIFEST_DIR` as a final fallback during dev.
+pub fn default_search_dirs(workspace_dir: &Path) -> Vec<PathBuf> {
+    let mut dirs = vec![workspace_dir.to_path_buf()];
+
+    if let Ok(exe) = std::env::current_exe() {
+        if let Some(parent) = exe.parent() {
+            dirs.push(parent.to_path_buf());
+        }
+    }
+
+    // During development, also check the project root (where Cargo.toml lives).
+    let manifest_dir = PathBuf::from(env!("CARGO_MANIFEST_DIR"));
+    if !dirs.contains(&manifest_dir) {
+        dirs.push(manifest_dir);
+    }
+
+    dirs
+}
+
+/// Try to load and parse a locale TOML file from the first matching search dir.
+fn load_locale_file(locale: &str, search_dirs: &[PathBuf]) -> HashMap<String, String> {
+    let filename = format!("tool_descriptions/{locale}.toml");
+
+    for dir in search_dirs {
+        let path = dir.join(&filename);
+        match std::fs::read_to_string(&path) {
+            Ok(contents) => match toml::from_str::<DescriptionFile>(&contents) {
+                Ok(parsed) => {
+                    debug!(path = %path.display(), keys = parsed.tools.len(), "loaded locale file");
+                    return parsed.tools;
+                }
+                Err(e) => {
+                    debug!(path = %path.display(), error = %e, "failed to parse locale file");
+                }
+            },
+            Err(_) => {
+                // File not found in this directory, try next.
+            }
+        }
+    }
+
+    debug!(
+        locale = locale,
+        "no locale file found in any search directory"
+    );
+    HashMap::new()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::fs;
+
+    /// Helper: create a temp dir with a `tool_descriptions/<locale>.toml` file.
+    fn write_locale_file(dir: &Path, locale: &str, content: &str) {
+        let td = dir.join("tool_descriptions");
+        fs::create_dir_all(&td).unwrap();
+        fs::write(td.join(format!("{locale}.toml")), content).unwrap();
+    }
+
+    #[test]
+    fn load_english_descriptions() {
+        let tmp = tempfile::tempdir().unwrap();
+        write_locale_file(
+            tmp.path(),
+            "en",
+            r#"[tools]
+shell = "Execute a shell command"
+file_read = "Read file contents"
+"#,
+        );
+        let descs = ToolDescriptions::load("en", &[tmp.path().to_path_buf()]);
+        assert_eq!(descs.get("shell"), Some("Execute a shell command"));
+        assert_eq!(descs.get("file_read"), Some("Read file contents"));
+        assert_eq!(descs.get("nonexistent"), None);
+        assert_eq!(descs.locale(), "en");
+    }
+
+    #[test]
+    fn fallback_to_english_when_locale_key_missing() {
+        let tmp = tempfile::tempdir().unwrap();
+        write_locale_file(
+            tmp.path(),
+            "en",
+            r#"[tools]
+shell = "Execute a shell command"
+file_read = "Read file contents"
+"#,
+        );
+        write_locale_file(
+            tmp.path(),
+            "zh-CN",
+            r#"[tools]
+shell = "在工作区目录中执行 shell 命令"
+"#,
+        );
+        let descs = ToolDescriptions::load("zh-CN", &[tmp.path().to_path_buf()]);
+        // Translated key returns Chinese.
+        assert_eq!(descs.get("shell"), Some("在工作区目录中执行 shell 命令"));
+        // Missing key falls back to English.
+        assert_eq!(descs.get("file_read"), Some("Read file contents"));
+        assert_eq!(descs.locale(), "zh-CN");
+    }
+
+    #[test]
+    fn fallback_when_locale_file_missing() {
+        let tmp = tempfile::tempdir().unwrap();
+        write_locale_file(
+            tmp.path(),
+            "en",
+            r#"[tools]
+shell = "Execute a shell command"
+"#,
+        );
+        // Request a locale that has no file.
+        let descs = ToolDescriptions::load("fr", &[tmp.path().to_path_buf()]);
+        // Falls back to English.
+        assert_eq!(descs.get("shell"), Some("Execute a shell command"));
+        assert_eq!(descs.locale(), "fr");
+    }
+
+    #[test]
+    fn fallback_when_no_files_exist() {
+        let tmp = tempfile::tempdir().unwrap();
+        let descs = ToolDescriptions::load("en", &[tmp.path().to_path_buf()]);
+        assert_eq!(descs.get("shell"), None);
+    }
+
+    #[test]
+    fn empty_always_returns_none() {
+        let descs = ToolDescriptions::empty();
+        assert_eq!(descs.get("shell"), None);
+        assert_eq!(descs.locale(), "en");
+    }
+
+    #[test]
+    fn detect_locale_from_env() {
+        // Save and restore env.
+        let saved = std::env::var("ZEROCLAW_LOCALE").ok();
+        let saved_lang = std::env::var("LANG").ok();
+
+        std::env::set_var("ZEROCLAW_LOCALE", "ja-JP");
+        assert_eq!(detect_locale(), "ja-JP");
+
+        std::env::remove_var("ZEROCLAW_LOCALE");
+        std::env::set_var("LANG", "zh_CN.UTF-8");
+        assert_eq!(detect_locale(), "zh-CN");
+
+        // Restore.
+        match saved {
+            Some(v) => std::env::set_var("ZEROCLAW_LOCALE", v),
+            None => std::env::remove_var("ZEROCLAW_LOCALE"),
+        }
+        match saved_lang {
+            Some(v) => std::env::set_var("LANG", v),
+            None => std::env::remove_var("LANG"),
+        }
+    }
+
+    #[test]
+    fn normalize_locale_strips_encoding() {
+        assert_eq!(normalize_locale("en_US.UTF-8"), "en-US");
+        assert_eq!(normalize_locale("zh_CN.utf8"), "zh-CN");
+        assert_eq!(normalize_locale("fr"), "fr");
+        assert_eq!(normalize_locale("pt_BR"), "pt-BR");
+    }
+
+    #[test]
+    fn config_locale_overrides_env() {
+        // This tests the precedence logic: if config provides a locale,
+        // it should be used instead of detect_locale().
+        // The actual override happens at the call site in prompt.rs / loop_.rs,
+        // so here we just verify ToolDescriptions works with an explicit locale.
+        let tmp = tempfile::tempdir().unwrap();
+        write_locale_file(
+            tmp.path(),
+            "de",
+            r#"[tools]
+shell = "Einen Shell-Befehl im Arbeitsverzeichnis ausführen"
+"#,
+        );
+        let descs = ToolDescriptions::load("de", &[tmp.path().to_path_buf()]);
+        assert_eq!(
+            descs.get("shell"),
+            Some("Einen Shell-Befehl im Arbeitsverzeichnis ausführen")
+        );
+    }
+}

src/integrations/registry.rs

@@ -840,6 +840,7 @@ mod tests {
             draft_update_interval_ms: 1000,
             interrupt_on_new_message: false,
             mention_only: false,
+            ack_reactions: None,
         });
         let entries = all_integrations();
         let tg = entries.iter().find(|e| e.name == "Telegram").unwrap();

src/lib.rs

@@ -54,6 +54,7 @@ pub(crate) mod hardware;
 pub(crate) mod health;
 pub(crate) mod heartbeat;
 pub mod hooks;
+pub mod i18n;
 pub(crate) mod identity;
 pub(crate) mod integrations;
 pub mod memory;

src/main.rs

@@ -89,6 +89,7 @@ mod hardware;
 mod health;
 mod heartbeat;
 mod hooks;
+mod i18n;
 mod identity;
 mod integrations;
 mod memory;

src/onboard/wizard.rs

@@ -193,6 +193,7 @@ pub async fn run_wizard(force: bool) -> Result<Config> {
         knowledge: crate::config::KnowledgeConfig::default(),
         linkedin: crate::config::LinkedInConfig::default(),
         plugins: crate::config::PluginsConfig::default(),
+        locale: None,
     };
 
     println!(
@@ -567,6 +568,7 @@ async fn run_quick_setup_with_home(
         knowledge: crate::config::KnowledgeConfig::default(),
         linkedin: crate::config::LinkedInConfig::default(),
         plugins: crate::config::PluginsConfig::default(),
+        locale: None,
     };
 
     config.save().await?;
@@ -3683,6 +3685,7 @@ fn setup_channels() -> Result<ChannelsConfig> {
                     draft_update_interval_ms: 1000,
                     interrupt_on_new_message: false,
                     mention_only: false,
+                    ack_reactions: None,
                 });
             }
             ChannelMenuChoice::Discord => {

src/providers/claude_code.rs

@@ -387,7 +387,7 @@ mod tests {
             use std::io::Write;
             let dir = std::env::temp_dir().join("zeroclaw_test_claude_code");
             std::fs::create_dir_all(&dir).unwrap();
-            let path = dir.join("fake_claude.sh");
+            let path = dir.join(format!("fake_claude_{}.sh", std::process::id()));
             let mut f = std::fs::File::create(&path).unwrap();
             writeln!(f, "#!/bin/sh\ncat /dev/stdin").unwrap();
             drop(f);

src/skills/creator.rs

@@ -0,0 +1,897 @@
+// Autonomous skill creation from successful multi-step task executions.
+//
+// After the agent completes a multi-step tool-call sequence, this module
+// can persist the execution as a reusable skill definition (SKILL.toml)
+// under `~/.zeroclaw/workspace/skills/<slug>/`.
+
+use crate::config::SkillCreationConfig;
+use crate::memory::embeddings::EmbeddingProvider;
+use crate::memory::vector::cosine_similarity;
+use anyhow::{Context, Result};
+use std::path::PathBuf;
+
+/// A record of a single tool call executed during a task.
+#[derive(Debug, Clone)]
+pub struct ToolCallRecord {
+    pub name: String,
+    pub args: serde_json::Value,
+}
+
+/// Creates reusable skill definitions from successful multi-step executions.
+pub struct SkillCreator {
+    workspace_dir: PathBuf,
+    config: SkillCreationConfig,
+}
+
+impl SkillCreator {
+    pub fn new(workspace_dir: PathBuf, config: SkillCreationConfig) -> Self {
+        Self {
+            workspace_dir,
+            config,
+        }
+    }
+
+    /// Attempt to create a skill from a successful multi-step task execution.
+    /// Returns `Ok(Some(slug))` if a skill was created, `Ok(None)` if skipped
+    /// (disabled, duplicate, or insufficient tool calls).
+    pub async fn create_from_execution(
+        &self,
+        task_description: &str,
+        tool_calls: &[ToolCallRecord],
+        embedding_provider: Option<&dyn EmbeddingProvider>,
+    ) -> Result<Option<String>> {
+        if !self.config.enabled {
+            return Ok(None);
+        }
+
+        if tool_calls.len() < 2 {
+            return Ok(None);
+        }
+
+        // Deduplicate via embeddings when an embedding provider is available.
+        if let Some(provider) = embedding_provider {
+            if provider.name() != "none" && self.is_duplicate(task_description, provider).await? {
+                return Ok(None);
+            }
+        }
+
+        let slug = Self::generate_slug(task_description);
+        if !Self::validate_slug(&slug) {
+            return Ok(None);
+        }
+
+        // Enforce LRU limit before writing a new skill.
+        self.enforce_lru_limit().await?;
+
+        let skill_dir = self.skills_dir().join(&slug);
+        tokio::fs::create_dir_all(&skill_dir)
+            .await
+            .with_context(|| {
+                format!("Failed to create skill directory: {}", skill_dir.display())
+            })?;
+
+        let toml_content = Self::generate_skill_toml(&slug, task_description, tool_calls);
+        let toml_path = skill_dir.join("SKILL.toml");
+        tokio::fs::write(&toml_path, toml_content.as_bytes())
+            .await
+            .with_context(|| format!("Failed to write {}", toml_path.display()))?;
+
+        Ok(Some(slug))
+    }
+
+    /// Generate a URL-safe slug from a task description.
+    /// Alphanumeric and hyphens only, max 64 characters.
+    fn generate_slug(description: &str) -> String {
+        let slug: String = description
+            .to_lowercase()
+            .chars()
+            .map(|c| if c.is_alphanumeric() { c } else { '-' })
+            .collect();
+
+        // Collapse consecutive hyphens.
+        let mut collapsed = String::with_capacity(slug.len());
+        let mut prev_hyphen = false;
+        for c in slug.chars() {
+            if c == '-' {
+                if !prev_hyphen {
+                    collapsed.push('-');
+                }
+                prev_hyphen = true;
+            } else {
+                collapsed.push(c);
+                prev_hyphen = false;
+            }
+        }
+
+        // Trim leading/trailing hyphens, then truncate.
+        let trimmed = collapsed.trim_matches('-');
+        if trimmed.len() > 64 {
+            // Truncate at a hyphen boundary if possible.
+            let truncated = &trimmed[..64];
+            truncated.trim_end_matches('-').to_string()
+        } else {
+            trimmed.to_string()
+        }
+    }
+
+    /// Validate that a slug is non-empty, alphanumeric + hyphens, max 64 chars.
+    fn validate_slug(slug: &str) -> bool {
+        !slug.is_empty()
+            && slug.len() <= 64
+            && slug.chars().all(|c| c.is_ascii_alphanumeric() || c == '-')
+            && !slug.starts_with('-')
+            && !slug.ends_with('-')
+    }
+
+    /// Generate SKILL.toml content from task execution data.
+    fn generate_skill_toml(slug: &str, description: &str, tool_calls: &[ToolCallRecord]) -> String {
+        use std::fmt::Write;
+        let mut toml = String::new();
+        toml.push_str("[skill]\n");
+        let _ = writeln!(toml, "name = {}", toml_escape(slug));
+        let _ = writeln!(
+            toml,
+            "description = {}",
+            toml_escape(&format!("Auto-generated: {description}"))
+        );
+        toml.push_str("version = \"0.1.0\"\n");
+        toml.push_str("author = \"zeroclaw-auto\"\n");
+        toml.push_str("tags = [\"auto-generated\"]\n");
+
+        for call in tool_calls {
+            toml.push('\n');
+            toml.push_str("[[tools]]\n");
+            let _ = writeln!(toml, "name = {}", toml_escape(&call.name));
+            let _ = writeln!(
+                toml,
+                "description = {}",
+                toml_escape(&format!("Tool used in task: {}", call.name))
+            );
+            toml.push_str("kind = \"shell\"\n");
+
+            // Extract the command from args if available, otherwise use the tool name.
+            let command = call
+                .args
+                .get("command")
+                .and_then(serde_json::Value::as_str)
+                .unwrap_or(&call.name);
+            let _ = writeln!(toml, "command = {}", toml_escape(command));
+        }
+
+        toml
+    }
+
+    /// Check if a skill with a similar description already exists.
+    async fn is_duplicate(
+        &self,
+        description: &str,
+        embedding_provider: &dyn EmbeddingProvider,
+    ) -> Result<bool> {
+        let new_embedding = embedding_provider.embed_one(description).await?;
+        if new_embedding.is_empty() {
+            return Ok(false);
+        }
+
+        let skills_dir = self.skills_dir();
+        if !skills_dir.exists() {
+            return Ok(false);
+        }
+
+        let mut entries = tokio::fs::read_dir(&skills_dir).await?;
+        while let Some(entry) = entries.next_entry().await? {
+            let toml_path = entry.path().join("SKILL.toml");
+            if !toml_path.exists() {
+                continue;
+            }
+
+            let content = tokio::fs::read_to_string(&toml_path).await?;
+            // Extract description from the TOML to compare.
+            if let Some(desc) = extract_description_from_toml(&content) {
+                let existing_embedding = embedding_provider.embed_one(&desc).await?;
+                if !existing_embedding.is_empty() {
+                    #[allow(clippy::cast_possible_truncation)]
+                    let similarity =
+                        f64::from(cosine_similarity(&new_embedding, &existing_embedding));
+                    if similarity > self.config.similarity_threshold {
+                        return Ok(true);
+                    }
+                }
+            }
+        }
+
+        Ok(false)
+    }
+
+    /// Remove the oldest auto-generated skill when we exceed `max_skills`.
+    async fn enforce_lru_limit(&self) -> Result<()> {
+        let skills_dir = self.skills_dir();
+        if !skills_dir.exists() {
+            return Ok(());
+        }
+
+        let mut auto_skills: Vec<(PathBuf, std::time::SystemTime)> = Vec::new();
+
+        let mut entries = tokio::fs::read_dir(&skills_dir).await?;
+        while let Some(entry) = entries.next_entry().await? {
+            let toml_path = entry.path().join("SKILL.toml");
+            if !toml_path.exists() {
+                continue;
+            }
+
+            let content = tokio::fs::read_to_string(&toml_path).await?;
+            if content.contains("\"zeroclaw-auto\"") || content.contains("\"auto-generated\"") {
+                let modified = tokio::fs::metadata(&toml_path)
+                    .await?
+                    .modified()
+                    .unwrap_or(std::time::UNIX_EPOCH);
+                auto_skills.push((entry.path(), modified));
+            }
+        }
+
+        // If at or above the limit, remove the oldest.
+        if auto_skills.len() >= self.config.max_skills {
+            auto_skills.sort_by_key(|(_, modified)| *modified);
+            if let Some((oldest_dir, _)) = auto_skills.first() {
+                tokio::fs::remove_dir_all(oldest_dir)
+                    .await
+                    .with_context(|| {
+                        format!(
+                            "Failed to remove oldest auto-generated skill: {}",
+                            oldest_dir.display()
+                        )
+                    })?;
+            }
+        }
+
+        Ok(())
+    }
+
+    fn skills_dir(&self) -> PathBuf {
+        self.workspace_dir.join("skills")
+    }
+}
+
+/// Escape a string for TOML value (double-quoted).
+fn toml_escape(s: &str) -> String {
+    let escaped = s
+        .replace('\\', "\\\\")
+        .replace('"', "\\\"")
+        .replace('\n', "\\n")
+        .replace('\r', "\\r")
+        .replace('\t', "\\t");
+    format!("\"{escaped}\"")
+}
+
+/// Extract the description field from a SKILL.toml string.
+fn extract_description_from_toml(content: &str) -> Option<String> {
+    #[derive(serde::Deserialize)]
+    struct Partial {
+        skill: PartialSkill,
+    }
+    #[derive(serde::Deserialize)]
+    struct PartialSkill {
+        description: Option<String>,
+    }
+    toml::from_str::<Partial>(content)
+        .ok()
+        .and_then(|p| p.skill.description)
+}
+
+/// Extract `ToolCallRecord`s from the agent conversation history.
+///
+/// Scans assistant messages for tool call patterns (both JSON and XML formats)
+/// and returns records for each unique tool invocation.
+pub fn extract_tool_calls_from_history(
+    history: &[crate::providers::ChatMessage],
+) -> Vec<ToolCallRecord> {
+    let mut records = Vec::new();
+
+    for msg in history {
+        if msg.role != "assistant" {
+            continue;
+        }
+
+        // Try parsing as JSON (native tool_calls format).
+        if let Ok(value) = serde_json::from_str::<serde_json::Value>(&msg.content) {
+            if let Some(tool_calls) = value.get("tool_calls").and_then(|v| v.as_array()) {
+                for call in tool_calls {
+                    if let Some(function) = call.get("function") {
+                        let name = function
+                            .get("name")
+                            .and_then(serde_json::Value::as_str)
+                            .unwrap_or("")
+                            .to_string();
+                        let args_str = function
+                            .get("arguments")
+                            .and_then(serde_json::Value::as_str)
+                            .unwrap_or("{}");
+                        let args = serde_json::from_str(args_str).unwrap_or_default();
+                        if !name.is_empty() {
+                            records.push(ToolCallRecord { name, args });
+                        }
+                    }
+                }
+            }
+        }
+
+        // Also try XML tool call format: <tool_name>...</tool_name>
+        // Simple extraction for `<shell>{"command":"..."}</shell>` style tags.
+        let content = &msg.content;
+        let mut pos = 0;
+        while pos < content.len() {
+            if let Some(start) = content[pos..].find('<') {
+                let abs_start = pos + start;
+                if let Some(end) = content[abs_start..].find('>') {
+                    let tag = &content[abs_start + 1..abs_start + end];
+                    // Skip closing tags and meta tags.
+                    if tag.starts_with('/') || tag.starts_with('!') || tag.starts_with('?') {
+                        pos = abs_start + end + 1;
+                        continue;
+                    }
+                    let tag_name = tag.split_whitespace().next().unwrap_or(tag);
+                    let close_tag = format!("</{tag_name}>");
+                    if let Some(close_pos) = content[abs_start + end + 1..].find(&close_tag) {
+                        let inner = &content[abs_start + end + 1..abs_start + end + 1 + close_pos];
+                        let args: serde_json::Value =
+                            serde_json::from_str(inner.trim()).unwrap_or_default();
+                        // Only add if it looks like a tool call (not HTML/formatting tags).
+                        if tag_name != "tool_result"
+                            && tag_name != "tool_results"
+                            && !tag_name.contains(':')
+                            && args.is_object()
+                            && !args.as_object().map_or(true, |o| o.is_empty())
+                        {
+                            records.push(ToolCallRecord {
+                                name: tag_name.to_string(),
+                                args,
+                            });
+                        }
+                        pos = abs_start + end + 1 + close_pos + close_tag.len();
+                    } else {
+                        pos = abs_start + end + 1;
+                    }
+                } else {
+                    break;
+                }
+            } else {
+                break;
+            }
+        }
+    }
+
+    records
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::memory::embeddings::{EmbeddingProvider, NoopEmbedding};
+    use async_trait::async_trait;
+
+    // ── Slug generation ──────────────────────────────────────────
+
+    #[test]
+    fn slug_basic() {
+        assert_eq!(
+            SkillCreator::generate_slug("Deploy to production"),
+            "deploy-to-production"
+        );
+    }
+
+    #[test]
+    fn slug_special_characters() {
+        assert_eq!(
+            SkillCreator::generate_slug("Build & test (CI/CD) pipeline!"),
+            "build-test-ci-cd-pipeline"
+        );
+    }
+
+    #[test]
+    fn slug_max_length() {
+        let long_desc = "a".repeat(100);
+        let slug = SkillCreator::generate_slug(&long_desc);
+        assert!(slug.len() <= 64);
+    }
+
+    #[test]
+    fn slug_leading_trailing_hyphens() {
+        let slug = SkillCreator::generate_slug("---hello world---");
+        assert!(!slug.starts_with('-'));
+        assert!(!slug.ends_with('-'));
+    }
+
+    #[test]
+    fn slug_consecutive_spaces() {
+        assert_eq!(SkillCreator::generate_slug("hello    world"), "hello-world");
+    }
+
+    #[test]
+    fn slug_empty_input() {
+        let slug = SkillCreator::generate_slug("");
+        assert!(slug.is_empty());
+    }
+
+    #[test]
+    fn slug_only_symbols() {
+        let slug = SkillCreator::generate_slug("!@#$%^&*()");
+        assert!(slug.is_empty());
+    }
+
+    #[test]
+    fn slug_unicode() {
+        let slug = SkillCreator::generate_slug("Deploy cafe app");
+        assert_eq!(slug, "deploy-cafe-app");
+    }
+
+    // ── Slug validation ──────────────────────────────────────────
+
+    #[test]
+    fn validate_slug_valid() {
+        assert!(SkillCreator::validate_slug("deploy-to-production"));
+        assert!(SkillCreator::validate_slug("a"));
+        assert!(SkillCreator::validate_slug("abc123"));
+    }
+
+    #[test]
+    fn validate_slug_invalid() {
+        assert!(!SkillCreator::validate_slug(""));
+        assert!(!SkillCreator::validate_slug("-starts-with-hyphen"));
+        assert!(!SkillCreator::validate_slug("ends-with-hyphen-"));
+        assert!(!SkillCreator::validate_slug("has spaces"));
+        assert!(!SkillCreator::validate_slug("has_underscores"));
+        assert!(!SkillCreator::validate_slug(&"a".repeat(65)));
+    }
+
+    // ── TOML generation ──────────────────────────────────────────
+
+    #[test]
+    fn toml_generation_valid_format() {
+        let calls = vec![
+            ToolCallRecord {
+                name: "shell".into(),
+                args: serde_json::json!({"command": "cargo build"}),
+            },
+            ToolCallRecord {
+                name: "shell".into(),
+                args: serde_json::json!({"command": "cargo test"}),
+            },
+        ];
+        let toml_str = SkillCreator::generate_skill_toml(
+            "build-and-test",
+            "Build and test the project",
+            &calls,
+        );
+
+        // Should parse as valid TOML.
+        let parsed: toml::Value =
+            toml::from_str(&toml_str).expect("Generated TOML should be valid");
+        let skill = parsed.get("skill").expect("Should have [skill] section");
+        assert_eq!(
+            skill.get("name").and_then(toml::Value::as_str),
+            Some("build-and-test")
+        );
+        assert_eq!(
+            skill.get("author").and_then(toml::Value::as_str),
+            Some("zeroclaw-auto")
+        );
+        assert_eq!(
+            skill.get("version").and_then(toml::Value::as_str),
+            Some("0.1.0")
+        );
+
+        let tools = parsed.get("tools").and_then(toml::Value::as_array).unwrap();
+        assert_eq!(tools.len(), 2);
+        assert_eq!(
+            tools[0].get("command").and_then(toml::Value::as_str),
+            Some("cargo build")
+        );
+    }
+
+    #[test]
+    fn toml_generation_escapes_quotes() {
+        let calls = vec![ToolCallRecord {
+            name: "shell".into(),
+            args: serde_json::json!({"command": "echo \"hello\""}),
+        }];
+        let toml_str =
+            SkillCreator::generate_skill_toml("echo-test", "Test \"quoted\" description", &calls);
+        let parsed: toml::Value =
+            toml::from_str(&toml_str).expect("TOML with quotes should be valid");
+        let desc = parsed
+            .get("skill")
+            .and_then(|s| s.get("description"))
+            .and_then(toml::Value::as_str)
+            .unwrap();
+        assert!(desc.contains("quoted"));
+    }
+
+    #[test]
+    fn toml_generation_no_command_arg() {
+        let calls = vec![ToolCallRecord {
+            name: "memory_store".into(),
+            args: serde_json::json!({"key": "foo", "value": "bar"}),
+        }];
+        let toml_str = SkillCreator::generate_skill_toml("memory-op", "Store to memory", &calls);
+        let parsed: toml::Value = toml::from_str(&toml_str).expect("TOML should be valid");
+        let tools = parsed.get("tools").and_then(toml::Value::as_array).unwrap();
+        // When no "command" arg exists, falls back to tool name.
+        assert_eq!(
+            tools[0].get("command").and_then(toml::Value::as_str),
+            Some("memory_store")
+        );
+    }
+
+    // ── TOML description extraction ──────────────────────────────
+
+    #[test]
+    fn extract_description_from_valid_toml() {
+        let content = r#"
+[skill]
+name = "test"
+description = "Auto-generated: Build project"
+version = "0.1.0"
+"#;
+        assert_eq!(
+            extract_description_from_toml(content),
+            Some("Auto-generated: Build project".into())
+        );
+    }
+
+    #[test]
+    fn extract_description_from_invalid_toml() {
+        assert_eq!(extract_description_from_toml("not valid toml {{"), None);
+    }
+
+    // ── Deduplication ────────────────────────────────────────────
+
+    /// A mock embedding provider that returns deterministic embeddings.
+    ///
+    /// The "new" description (first text embedded) always gets `[1, 0, 0]`.
+    /// The "existing" skill description (second text embedded) gets a vector
+    /// whose cosine similarity with `[1, 0, 0]` equals `self.similarity`.
+    struct MockEmbeddingProvider {
+        similarity: f32,
+        call_count: std::sync::atomic::AtomicUsize,
+    }
+
+    impl MockEmbeddingProvider {
+        fn new(similarity: f32) -> Self {
+            Self {
+                similarity,
+                call_count: std::sync::atomic::AtomicUsize::new(0),
+            }
+        }
+    }
+
+    #[async_trait]
+    impl EmbeddingProvider for MockEmbeddingProvider {
+        fn name(&self) -> &str {
+            "mock"
+        }
+        fn dimensions(&self) -> usize {
+            3
+        }
+        async fn embed(&self, texts: &[&str]) -> anyhow::Result<Vec<Vec<f32>>> {
+            Ok(texts
+                .iter()
+                .map(|_| {
+                    let call = self
+                        .call_count
+                        .fetch_add(1, std::sync::atomic::Ordering::Relaxed);
+                    if call == 0 {
+                        // First call: the "new" description.
+                        vec![1.0, 0.0, 0.0]
+                    } else {
+                        // Subsequent calls: existing skill descriptions.
+                        // Produce a vector with the configured cosine similarity to [1,0,0].
+                        vec![
+                            self.similarity,
+                            (1.0 - self.similarity * self.similarity).sqrt(),
+                            0.0,
+                        ]
+                    }
+                })
+                .collect())
+        }
+    }
+
+    #[tokio::test]
+    async fn dedup_skips_similar_descriptions() {
+        let dir = tempfile::tempdir().unwrap();
+        let skills_dir = dir.path().join("skills").join("existing-skill");
+        tokio::fs::create_dir_all(&skills_dir).await.unwrap();
+        tokio::fs::write(
+            skills_dir.join("SKILL.toml"),
+            r#"
+[skill]
+name = "existing-skill"
+description = "Auto-generated: Build the project"
+version = "0.1.0"
+author = "zeroclaw-auto"
+tags = ["auto-generated"]
+"#,
+        )
+        .await
+        .unwrap();
+
+        let config = SkillCreationConfig {
+            enabled: true,
+            max_skills: 500,
+            similarity_threshold: 0.85,
+        };
+
+        // High similarity provider -> should detect as duplicate.
+        let provider = MockEmbeddingProvider::new(0.95);
+        let creator = SkillCreator::new(dir.path().to_path_buf(), config.clone());
+        assert!(creator
+            .is_duplicate("Build the project", &provider)
+            .await
+            .unwrap());
+
+        // Low similarity provider -> not a duplicate.
+        let provider_low = MockEmbeddingProvider::new(0.3);
+        let creator2 = SkillCreator::new(dir.path().to_path_buf(), config);
+        assert!(!creator2
+            .is_duplicate("Completely different task", &provider_low)
+            .await
+            .unwrap());
+    }
+
+    // ── LRU eviction ─────────────────────────────────────────────
+
+    #[tokio::test]
+    async fn lru_eviction_removes_oldest() {
+        let dir = tempfile::tempdir().unwrap();
+        let config = SkillCreationConfig {
+            enabled: true,
+            max_skills: 2,
+            similarity_threshold: 0.85,
+        };
+
+        let skills_dir = dir.path().join("skills");
+
+        // Create two auto-generated skills with different timestamps.
+        for (i, name) in ["old-skill", "new-skill"].iter().enumerate() {
+            let skill_dir = skills_dir.join(name);
+            tokio::fs::create_dir_all(&skill_dir).await.unwrap();
+            tokio::fs::write(
+                skill_dir.join("SKILL.toml"),
+                format!(
+                    r#"[skill]
+name = "{name}"
+description = "Auto-generated: Skill {i}"
+version = "0.1.0"
+author = "zeroclaw-auto"
+tags = ["auto-generated"]
+"#
+                ),
+            )
+            .await
+            .unwrap();
+            // Small delay to ensure different timestamps.
+            tokio::time::sleep(std::time::Duration::from_millis(50)).await;
+        }
+
+        let creator = SkillCreator::new(dir.path().to_path_buf(), config);
+        creator.enforce_lru_limit().await.unwrap();
+
+        // The oldest skill should have been removed.
+        assert!(!skills_dir.join("old-skill").exists());
+        assert!(skills_dir.join("new-skill").exists());
+    }
+
+    // ── End-to-end: create_from_execution ────────────────────────
+
+    #[tokio::test]
+    async fn create_from_execution_disabled() {
+        let dir = tempfile::tempdir().unwrap();
+        let config = SkillCreationConfig {
+            enabled: false,
+            ..Default::default()
+        };
+        let creator = SkillCreator::new(dir.path().to_path_buf(), config);
+        let calls = vec![
+            ToolCallRecord {
+                name: "shell".into(),
+                args: serde_json::json!({"command": "ls"}),
+            },
+            ToolCallRecord {
+                name: "shell".into(),
+                args: serde_json::json!({"command": "pwd"}),
+            },
+        ];
+        let result = creator
+            .create_from_execution("List files", &calls, None)
+            .await
+            .unwrap();
+        assert!(result.is_none());
+    }
+
+    #[tokio::test]
+    async fn create_from_execution_insufficient_steps() {
+        let dir = tempfile::tempdir().unwrap();
+        let config = SkillCreationConfig {
+            enabled: true,
+            ..Default::default()
+        };
+        let creator = SkillCreator::new(dir.path().to_path_buf(), config);
+        let calls = vec![ToolCallRecord {
+            name: "shell".into(),
+            args: serde_json::json!({"command": "ls"}),
+        }];
+        let result = creator
+            .create_from_execution("List files", &calls, None)
+            .await
+            .unwrap();
+        assert!(result.is_none());
+    }
+
+    #[tokio::test]
+    async fn create_from_execution_success() {
+        let dir = tempfile::tempdir().unwrap();
+        let config = SkillCreationConfig {
+            enabled: true,
+            max_skills: 500,
+            similarity_threshold: 0.85,
+        };
+        let creator = SkillCreator::new(dir.path().to_path_buf(), config);
+        let calls = vec![
+            ToolCallRecord {
+                name: "shell".into(),
+                args: serde_json::json!({"command": "cargo build"}),
+            },
+            ToolCallRecord {
+                name: "shell".into(),
+                args: serde_json::json!({"command": "cargo test"}),
+            },
+        ];
+
+        // Use noop embedding (no deduplication).
+        let noop = NoopEmbedding;
+        let result = creator
+            .create_from_execution("Build and test", &calls, Some(&noop))
+            .await
+            .unwrap();
+        assert_eq!(result, Some("build-and-test".into()));
+
+        // Verify the skill directory and TOML were created.
+        let skill_dir = dir.path().join("skills").join("build-and-test");
+        assert!(skill_dir.exists());
+        let toml_content = tokio::fs::read_to_string(skill_dir.join("SKILL.toml"))
+            .await
+            .unwrap();
+        assert!(toml_content.contains("build-and-test"));
+        assert!(toml_content.contains("zeroclaw-auto"));
+    }
+
+    #[tokio::test]
+    async fn create_from_execution_with_dedup() {
+        let dir = tempfile::tempdir().unwrap();
+        let config = SkillCreationConfig {
+            enabled: true,
+            max_skills: 500,
+            similarity_threshold: 0.85,
+        };
+
+        // First, create an existing skill.
+        let skills_dir = dir.path().join("skills").join("existing");
+        tokio::fs::create_dir_all(&skills_dir).await.unwrap();
+        tokio::fs::write(
+            skills_dir.join("SKILL.toml"),
+            r#"[skill]
+name = "existing"
+description = "Auto-generated: Build and test"
+version = "0.1.0"
+author = "zeroclaw-auto"
+tags = ["auto-generated"]
+"#,
+        )
+        .await
+        .unwrap();
+
+        // High similarity provider -> should skip.
+        let provider = MockEmbeddingProvider::new(0.95);
+        let creator = SkillCreator::new(dir.path().to_path_buf(), config);
+        let calls = vec![
+            ToolCallRecord {
+                name: "shell".into(),
+                args: serde_json::json!({"command": "cargo build"}),
+            },
+            ToolCallRecord {
+                name: "shell".into(),
+                args: serde_json::json!({"command": "cargo test"}),
+            },
+        ];
+        let result = creator
+            .create_from_execution("Build and test", &calls, Some(&provider))
+            .await
+            .unwrap();
+        assert!(result.is_none());
+    }
+
+    // ── Tool call extraction from history ────────────────────────
+
+    #[test]
+    fn extract_from_empty_history() {
+        let history = vec![];
+        let records = extract_tool_calls_from_history(&history);
+        assert!(records.is_empty());
+    }
+
+    #[test]
+    fn extract_from_user_messages_only() {
+        use crate::providers::ChatMessage;
+        let history = vec![ChatMessage::user("hello"), ChatMessage::user("world")];
+        let records = extract_tool_calls_from_history(&history);
+        assert!(records.is_empty());
+    }
+
+    // ── Fuzz-like tests for slug ─────────────────────────────────
+
+    #[test]
+    fn slug_fuzz_various_inputs() {
+        let inputs = [
+            "",
+            " ",
+            "---",
+            "a",
+            "hello world!",
+            "UPPER CASE",
+            "with-hyphens-already",
+            "with__underscores",
+            "123 numbers 456",
+            "emoji: cafe",
+            &"x".repeat(200),
+            "a-b-c-d-e-f-g-h-i-j-k-l-m-n-o-p-q-r-s-t-u-v-w-x-y-z-0-1-2-3-4-5",
+        ];
+
+        for input in &inputs {
+            let slug = SkillCreator::generate_slug(input);
+            // Slug should always pass validation (or be empty for degenerate input).
+            if !slug.is_empty() {
+                assert!(
+                    SkillCreator::validate_slug(&slug),
+                    "Generated slug '{slug}' from '{input}' failed validation"
+                );
+            }
+        }
+    }
+
+    // ── Fuzz-like tests for TOML generation ──────────────────────
+
+    #[test]
+    fn toml_fuzz_various_inputs() {
+        let descriptions = [
+            "simple task",
+            "task with \"quotes\" and \\ backslashes",
+            "task with\nnewlines\r\nand tabs\there",
+            "",
+            &"long ".repeat(100),
+        ];
+
+        let args_variants = [
+            serde_json::json!({}),
+            serde_json::json!({"command": "echo hello"}),
+            serde_json::json!({"command": "echo \"hello world\"", "extra": 42}),
+        ];
+
+        for desc in &descriptions {
+            for args in &args_variants {
+                let calls = vec![
+                    ToolCallRecord {
+                        name: "tool1".into(),
+                        args: args.clone(),
+                    },
+                    ToolCallRecord {
+                        name: "tool2".into(),
+                        args: args.clone(),
+                    },
+                ];
+                let toml_str = SkillCreator::generate_skill_toml("test-slug", desc, &calls);
+                // Must always produce valid TOML.
+                let _parsed: toml::Value = toml::from_str(&toml_str)
+                    .unwrap_or_else(|e| panic!("Invalid TOML for desc '{desc}': {e}\n{toml_str}"));
+            }
+        }
+    }
+}

src/skills/mod.rs

@@ -7,6 +7,8 @@ use std::process::Command;
 use std::time::{Duration, SystemTime};
 
 mod audit;
+#[cfg(feature = "skill-creation")]
+pub mod creator;
 
 const OPEN_SKILLS_REPO_URL: &str = "https://github.com/besoeasy/open-skills";
 const OPEN_SKILLS_SYNC_MARKER: &str = ".zeroclaw-open-skills-sync";

src/tools/delegate.rs

@@ -296,8 +296,9 @@ impl Tool for DelegateTool {
         }
 
         // Wrap the provider call in a timeout to prevent indefinite blocking
+        let timeout_secs = agent_config.timeout_secs.unwrap_or(DELEGATE_TIMEOUT_SECS);
         let result = tokio::time::timeout(
-            Duration::from_secs(DELEGATE_TIMEOUT_SECS),
+            Duration::from_secs(timeout_secs),
             provider.chat_with_system(
                 agent_config.system_prompt.as_deref(),
                 &full_prompt,
@@ -314,7 +315,7 @@ impl Tool for DelegateTool {
                     success: false,
                     output: String::new(),
                     error: Some(format!(
-                        "Agent '{agent_name}' timed out after {DELEGATE_TIMEOUT_SECS}s"
+                        "Agent '{agent_name}' timed out after {timeout_secs}s"
                     )),
                 });
             }
@@ -401,8 +402,11 @@ impl DelegateTool {
 
         let noop_observer = NoopObserver;
 
+        let agentic_timeout_secs = agent_config
+            .agentic_timeout_secs
+            .unwrap_or(DELEGATE_AGENTIC_TIMEOUT_SECS);
         let result = tokio::time::timeout(
-            Duration::from_secs(DELEGATE_AGENTIC_TIMEOUT_SECS),
+            Duration::from_secs(agentic_timeout_secs),
             run_tool_call_loop(
                 provider,
                 &mut history,
@@ -454,7 +458,7 @@ impl DelegateTool {
                 success: false,
                 output: String::new(),
                 error: Some(format!(
-                    "Agent '{agent_name}' timed out after {DELEGATE_AGENTIC_TIMEOUT_SECS}s"
+                    "Agent '{agent_name}' timed out after {agentic_timeout_secs}s"
                 )),
             }),
         }
@@ -531,6 +535,8 @@ mod tests {
                 agentic: false,
                 allowed_tools: Vec::new(),
                 max_iterations: 10,
+                timeout_secs: None,
+                agentic_timeout_secs: None,
             },
         );
         agents.insert(
@@ -545,6 +551,8 @@ mod tests {
                 agentic: false,
                 allowed_tools: Vec::new(),
                 max_iterations: 10,
+                timeout_secs: None,
+                agentic_timeout_secs: None,
             },
         );
         agents
@@ -698,6 +706,8 @@ mod tests {
             agentic: true,
             allowed_tools,
             max_iterations,
+            timeout_secs: None,
+            agentic_timeout_secs: None,
         }
     }
 
@@ -806,6 +816,8 @@ mod tests {
                 agentic: false,
                 allowed_tools: Vec::new(),
                 max_iterations: 10,
+                timeout_secs: None,
+                agentic_timeout_secs: None,
             },
         );
         let tool = DelegateTool::new(agents, None, test_security());
@@ -912,6 +924,8 @@ mod tests {
                 agentic: false,
                 allowed_tools: Vec::new(),
                 max_iterations: 10,
+                timeout_secs: None,
+                agentic_timeout_secs: None,
             },
         );
         let tool = DelegateTool::new(agents, None, test_security());
@@ -947,6 +961,8 @@ mod tests {
                 agentic: false,
                 allowed_tools: Vec::new(),
                 max_iterations: 10,
+                timeout_secs: None,
+                agentic_timeout_secs: None,
             },
         );
         let tool = DelegateTool::new(agents, None, test_security());
@@ -1221,4 +1237,226 @@ mod tests {
         handle.write().push(Arc::new(FakeMcpTool));
         assert_eq!(handle.read().len(), 2);
     }
+
+    // ── Configurable timeout tests ──────────────────────────────────
+
+    #[test]
+    fn default_timeout_values_used_when_config_unset() {
+        let config = DelegateAgentConfig {
+            provider: "ollama".to_string(),
+            model: "llama3".to_string(),
+            system_prompt: None,
+            api_key: None,
+            temperature: None,
+            max_depth: 3,
+            agentic: false,
+            allowed_tools: Vec::new(),
+            max_iterations: 10,
+            timeout_secs: None,
+            agentic_timeout_secs: None,
+        };
+        assert_eq!(config.timeout_secs.unwrap_or(DELEGATE_TIMEOUT_SECS), 120);
+        assert_eq!(
+            config
+                .agentic_timeout_secs
+                .unwrap_or(DELEGATE_AGENTIC_TIMEOUT_SECS),
+            300
+        );
+    }
+
+    #[test]
+    fn custom_timeout_values_are_respected() {
+        let config = DelegateAgentConfig {
+            provider: "ollama".to_string(),
+            model: "llama3".to_string(),
+            system_prompt: None,
+            api_key: None,
+            temperature: None,
+            max_depth: 3,
+            agentic: false,
+            allowed_tools: Vec::new(),
+            max_iterations: 10,
+            timeout_secs: Some(60),
+            agentic_timeout_secs: Some(600),
+        };
+        assert_eq!(config.timeout_secs.unwrap_or(DELEGATE_TIMEOUT_SECS), 60);
+        assert_eq!(
+            config
+                .agentic_timeout_secs
+                .unwrap_or(DELEGATE_AGENTIC_TIMEOUT_SECS),
+            600
+        );
+    }
+
+    #[test]
+    fn timeout_deserialization_defaults_to_none() {
+        let toml_str = r#"
+            provider = "ollama"
+            model = "llama3"
+        "#;
+        let config: DelegateAgentConfig = toml::from_str(toml_str).unwrap();
+        assert!(config.timeout_secs.is_none());
+        assert!(config.agentic_timeout_secs.is_none());
+    }
+
+    #[test]
+    fn timeout_deserialization_with_custom_values() {
+        let toml_str = r#"
+            provider = "ollama"
+            model = "llama3"
+            timeout_secs = 45
+            agentic_timeout_secs = 900
+        "#;
+        let config: DelegateAgentConfig = toml::from_str(toml_str).unwrap();
+        assert_eq!(config.timeout_secs, Some(45));
+        assert_eq!(config.agentic_timeout_secs, Some(900));
+    }
+
+    #[test]
+    fn config_validation_rejects_zero_timeout() {
+        let mut config = crate::config::Config::default();
+        config.agents.insert(
+            "bad".into(),
+            DelegateAgentConfig {
+                provider: "ollama".into(),
+                model: "llama3".into(),
+                system_prompt: None,
+                api_key: None,
+                temperature: None,
+                max_depth: 3,
+                agentic: false,
+                allowed_tools: Vec::new(),
+                max_iterations: 10,
+                timeout_secs: Some(0),
+                agentic_timeout_secs: None,
+            },
+        );
+        let err = config.validate().unwrap_err();
+        assert!(
+            format!("{err}").contains("timeout_secs must be greater than 0"),
+            "unexpected error: {err}"
+        );
+    }
+
+    #[test]
+    fn config_validation_rejects_zero_agentic_timeout() {
+        let mut config = crate::config::Config::default();
+        config.agents.insert(
+            "bad".into(),
+            DelegateAgentConfig {
+                provider: "ollama".into(),
+                model: "llama3".into(),
+                system_prompt: None,
+                api_key: None,
+                temperature: None,
+                max_depth: 3,
+                agentic: false,
+                allowed_tools: Vec::new(),
+                max_iterations: 10,
+                timeout_secs: None,
+                agentic_timeout_secs: Some(0),
+            },
+        );
+        let err = config.validate().unwrap_err();
+        assert!(
+            format!("{err}").contains("agentic_timeout_secs must be greater than 0"),
+            "unexpected error: {err}"
+        );
+    }
+
+    #[test]
+    fn config_validation_rejects_excessive_timeout() {
+        let mut config = crate::config::Config::default();
+        config.agents.insert(
+            "bad".into(),
+            DelegateAgentConfig {
+                provider: "ollama".into(),
+                model: "llama3".into(),
+                system_prompt: None,
+                api_key: None,
+                temperature: None,
+                max_depth: 3,
+                agentic: false,
+                allowed_tools: Vec::new(),
+                max_iterations: 10,
+                timeout_secs: Some(7200),
+                agentic_timeout_secs: None,
+            },
+        );
+        let err = config.validate().unwrap_err();
+        assert!(
+            format!("{err}").contains("exceeds max 3600"),
+            "unexpected error: {err}"
+        );
+    }
+
+    #[test]
+    fn config_validation_rejects_excessive_agentic_timeout() {
+        let mut config = crate::config::Config::default();
+        config.agents.insert(
+            "bad".into(),
+            DelegateAgentConfig {
+                provider: "ollama".into(),
+                model: "llama3".into(),
+                system_prompt: None,
+                api_key: None,
+                temperature: None,
+                max_depth: 3,
+                agentic: false,
+                allowed_tools: Vec::new(),
+                max_iterations: 10,
+                timeout_secs: None,
+                agentic_timeout_secs: Some(5000),
+            },
+        );
+        let err = config.validate().unwrap_err();
+        assert!(
+            format!("{err}").contains("exceeds max 3600"),
+            "unexpected error: {err}"
+        );
+    }
+
+    #[test]
+    fn config_validation_accepts_max_boundary_timeout() {
+        let mut config = crate::config::Config::default();
+        config.agents.insert(
+            "ok".into(),
+            DelegateAgentConfig {
+                provider: "ollama".into(),
+                model: "llama3".into(),
+                system_prompt: None,
+                api_key: None,
+                temperature: None,
+                max_depth: 3,
+                agentic: false,
+                allowed_tools: Vec::new(),
+                max_iterations: 10,
+                timeout_secs: Some(3600),
+                agentic_timeout_secs: Some(3600),
+            },
+        );
+        assert!(config.validate().is_ok());
+    }
+
+    #[test]
+    fn config_validation_accepts_none_timeouts() {
+        let mut config = crate::config::Config::default();
+        config.agents.insert(
+            "ok".into(),
+            DelegateAgentConfig {
+                provider: "ollama".into(),
+                model: "llama3".into(),
+                system_prompt: None,
+                api_key: None,
+                temperature: None,
+                max_depth: 3,
+                agentic: false,
+                allowed_tools: Vec::new(),
+                max_iterations: 10,
+                timeout_secs: None,
+                agentic_timeout_secs: None,
+            },
+        );
+        assert!(config.validate().is_ok());
+    }
 }

src/tools/file_edit.rs

@@ -103,7 +103,7 @@ impl Tool for FileEditTool {
             });
         }
 
-        let full_path = self.security.workspace_dir.join(path);
+        let full_path = self.security.resolve_tool_path(path);
 
         // ── 5. Canonicalize parent ─────────────────────────────────
         let Some(parent) = full_path.parent() else {
@@ -666,6 +666,46 @@ mod tests {
         let _ = tokio::fs::remove_dir_all(&dir).await;
     }
 
+    #[tokio::test]
+    async fn file_edit_absolute_path_in_workspace() {
+        let dir = std::env::temp_dir().join("zeroclaw_test_file_edit_abs_path");
+        let _ = tokio::fs::remove_dir_all(&dir).await;
+        tokio::fs::create_dir_all(&dir).await.unwrap();
+
+        // Canonicalize so the workspace dir matches resolved paths on macOS (/private/var/…)
+        let dir = tokio::fs::canonicalize(&dir).await.unwrap();
+
+        tokio::fs::write(dir.join("target.txt"), "old content")
+            .await
+            .unwrap();
+
+        let tool = FileEditTool::new(test_security(dir.clone()));
+
+        // Pass an absolute path that is within the workspace
+        let abs_path = dir.join("target.txt");
+        let result = tool
+            .execute(json!({
+                "path": abs_path.to_string_lossy().to_string(),
+                "old_string": "old content",
+                "new_string": "new content"
+            }))
+            .await
+            .unwrap();
+
+        assert!(
+            result.success,
+            "editing via absolute workspace path should succeed, error: {:?}",
+            result.error
+        );
+
+        let content = tokio::fs::read_to_string(dir.join("target.txt"))
+            .await
+            .unwrap();
+        assert_eq!(content, "new content");
+
+        let _ = tokio::fs::remove_dir_all(&dir).await;
+    }
+
     #[tokio::test]
     async fn file_edit_blocks_null_byte_in_path() {
         let dir = std::env::temp_dir().join("zeroclaw_test_file_edit_null_byte");

src/tools/file_write.rs

@@ -78,7 +78,7 @@ impl Tool for FileWriteTool {
             });
         }
 
-        let full_path = self.security.workspace_dir.join(path);
+        let full_path = self.security.resolve_tool_path(path);
 
         let Some(parent) = full_path.parent() else {
             return Ok(ToolResult {
@@ -450,6 +450,40 @@ mod tests {
         let _ = tokio::fs::remove_dir_all(&root).await;
     }
 
+    #[tokio::test]
+    async fn file_write_absolute_path_in_workspace() {
+        let dir = std::env::temp_dir().join("zeroclaw_test_file_write_abs_path");
+        let _ = tokio::fs::remove_dir_all(&dir).await;
+        tokio::fs::create_dir_all(&dir).await.unwrap();
+
+        // Canonicalize so the workspace dir matches resolved paths on macOS (/private/var/…)
+        let dir = tokio::fs::canonicalize(&dir).await.unwrap();
+
+        let tool = FileWriteTool::new(test_security(dir.clone()));
+
+        // Pass an absolute path that is within the workspace
+        let abs_path = dir.join("abs_test.txt");
+        let result = tool
+            .execute(
+                json!({"path": abs_path.to_string_lossy().to_string(), "content": "absolute!"}),
+            )
+            .await
+            .unwrap();
+
+        assert!(
+            result.success,
+            "writing via absolute workspace path should succeed, error: {:?}",
+            result.error
+        );
+
+        let content = tokio::fs::read_to_string(dir.join("abs_test.txt"))
+            .await
+            .unwrap();
+        assert_eq!(content, "absolute!");
+
+        let _ = tokio::fs::remove_dir_all(&dir).await;
+    }
+
     #[tokio::test]
     async fn file_write_blocks_null_byte_in_path() {
         let dir = std::env::temp_dir().join("zeroclaw_test_file_write_null");

src/tools/mod.rs

@@ -917,6 +917,8 @@ mod tests {
                 agentic: false,
                 allowed_tools: Vec::new(),
                 max_iterations: 10,
+                timeout_secs: None,
+                agentic_timeout_secs: None,
             },
         );
 

src/tools/model_routing_config.rs

@@ -705,6 +705,8 @@ impl ModelRoutingConfigTool {
                 agentic: false,
                 allowed_tools: Vec::new(),
                 max_iterations: DEFAULT_AGENT_MAX_ITERATIONS,
+                timeout_secs: None,
+                agentic_timeout_secs: None,
             });
 
         next_agent.provider = provider;

src/tools/pdf_read.rs

@@ -100,7 +100,7 @@ impl Tool for PdfReadTool {
             });
         }
 
-        let full_path = self.security.workspace_dir.join(path);
+        let full_path = self.security.resolve_tool_path(path);
 
         let resolved_path = match tokio::fs::canonicalize(&full_path).await {
             Ok(p) => p,

src/tools/swarm.rs

@@ -566,6 +566,8 @@ mod tests {
                 agentic: false,
                 allowed_tools: Vec::new(),
                 max_iterations: 10,
+                timeout_secs: None,
+                agentic_timeout_secs: None,
             },
         );
         agents.insert(
@@ -580,6 +582,8 @@ mod tests {
                 agentic: false,
                 allowed_tools: Vec::new(),
                 max_iterations: 10,
+                timeout_secs: None,
+                agentic_timeout_secs: None,
             },
         );
         agents

tool_descriptions/en.toml

@@ -0,0 +1,59 @@
+# English tool descriptions (default locale)
+#
+# Each key under [tools] matches the tool's name() return value.
+# Values are the human-readable descriptions shown in system prompts.
+
+[tools]
+backup = "Create, list, verify, and restore workspace backups"
+browser = "Web/browser automation with pluggable backends (agent-browser, rust-native, computer_use). Supports DOM actions plus optional OS-level actions (mouse_move, mouse_click, mouse_drag, key_type, key_press, screen_capture) through a computer-use sidecar. Use 'snapshot' to map interactive elements to refs (@e1, @e2). Enforces browser.allowed_domains for open actions."
+browser_delegate = "Delegate browser-based tasks to a browser-capable CLI for interacting with web applications like Teams, Outlook, Jira, Confluence"
+browser_open = "Open an approved HTTPS URL in the system browser. Security constraints: allowlist-only domains, no local/private hosts, no scraping."
+cloud_ops = "Cloud transformation advisory tool. Analyzes IaC plans, assesses migration paths, reviews costs, and checks architecture against Well-Architected Framework pillars. Read-only: does not create or modify cloud resources."
+cloud_patterns = "Cloud pattern library. Given a workload description, suggests applicable cloud-native architectural patterns (containerization, serverless, database modernization, etc.)."
+composio = "Execute actions on 1000+ apps via Composio (Gmail, Notion, GitHub, Slack, etc.). Use action='list' to see available actions (includes parameter names). action='execute' with action_name/tool_slug and params to run an action. If you are unsure of the exact params, pass 'text' instead with a natural-language description of what you want (Composio will resolve the correct parameters via NLP). action='list_accounts' or action='connected_accounts' to list OAuth-connected accounts. action='connect' with app/auth_config_id to get OAuth URL. connected_account_id is auto-resolved when omitted."
+content_search = "Search file contents by regex pattern within the workspace. Supports ripgrep (rg) with grep fallback. Output modes: 'content' (matching lines with context), 'files_with_matches' (file paths only), 'count' (match counts per file). Example: pattern='fn main', include='*.rs', output_mode='content'."
+cron_add = """Create a scheduled cron job (shell or agent) with cron/at/every schedules. Use job_type='agent' with a prompt to run the AI agent on schedule. To deliver output to a channel (Discord, Telegram, Slack, Mattermost, Matrix), set delivery={"mode":"announce","channel":"discord","to":"<channel_id_or_chat_id>"}. This is the preferred tool for sending scheduled/delayed messages to users via channels."""
+cron_list = "List all scheduled cron jobs"
+cron_remove = "Remove a cron job by id"
+cron_run = "Force-run a cron job immediately and record run history"
+cron_runs = "List recent run history for a cron job"
+cron_update = "Patch an existing cron job (schedule, command, prompt, enabled, delivery, model, etc.)"
+data_management = "Workspace data retention, purge, and storage statistics"
+delegate = "Delegate a subtask to a specialized agent. Use when: a task benefits from a different model (e.g. fast summarization, deep reasoning, code generation). The sub-agent runs a single prompt by default; with agentic=true it can iterate with a filtered tool-call loop."
+file_edit = "Edit a file by replacing an exact string match with new content"
+file_read = "Read file contents with line numbers. Supports partial reading via offset and limit. Extracts text from PDF; other binary files are read with lossy UTF-8 conversion."
+file_write = "Write contents to a file in the workspace"
+git_operations = "Perform structured Git operations (status, diff, log, branch, commit, add, checkout, stash). Provides parsed JSON output and integrates with security policy for autonomy controls."
+glob_search = "Search for files matching a glob pattern within the workspace. Returns a sorted list of matching file paths relative to the workspace root. Examples: '**/*.rs' (all Rust files), 'src/**/mod.rs' (all mod.rs in src)."
+google_workspace = "Interact with Google Workspace services (Drive, Gmail, Calendar, Sheets, Docs, etc.) via the gws CLI. Requires gws to be installed and authenticated."
+hardware_board_info = "Return full board info (chip, architecture, memory map) for connected hardware. Use when: user asks for 'board info', 'what board do I have', 'connected hardware', 'chip info', 'what hardware', or 'memory map'."
+hardware_memory_map = "Return the memory map (flash and RAM address ranges) for connected hardware. Use when: user asks for 'upper and lower memory addresses', 'memory map', 'address space', or 'readable addresses'. Returns flash/RAM ranges from datasheets."
+hardware_memory_read = "Read actual memory/register values from Nucleo via USB. Use when: user asks to 'read register values', 'read memory at address', 'dump memory', 'lower memory 0-126', or 'give address and value'. Returns hex dump. Requires Nucleo connected via USB and probe feature. Params: address (hex, e.g. 0x20000000 for RAM start), length (bytes, default 128)."
+http_request = "Make HTTP requests to external APIs. Supports GET, POST, PUT, DELETE, PATCH, HEAD, OPTIONS methods. Security constraints: allowlist-only domains, no local/private hosts, configurable timeout and response size limits."
+image_info = "Read image file metadata (format, dimensions, size) and optionally return base64-encoded data."
+knowledge = "Manage a knowledge graph of architecture decisions, solution patterns, lessons learned, and experts. Actions: capture, search, relate, suggest, expert_find, lessons_extract, graph_stats."
+linkedin = "Manage LinkedIn: create posts, list your posts, comment, react, delete posts, view engagement, get profile info, and read the configured content strategy. Requires LINKEDIN_* credentials in .env file."
+memory_forget = "Remove a memory by key. Use to delete outdated facts or sensitive data. Returns whether the memory was found and removed."
+memory_recall = "Search long-term memory for relevant facts, preferences, or context. Returns scored results ranked by relevance."
+memory_store = "Store a fact, preference, or note in long-term memory. Use category 'core' for permanent facts, 'daily' for session notes, 'conversation' for chat context, or a custom category name."
+microsoft365 = "Microsoft 365 integration: manage Outlook mail, Teams messages, Calendar events, OneDrive files, and SharePoint search via Microsoft Graph API"
+model_routing_config = "Manage default model settings, scenario-based provider/model routes, classification rules, and delegate sub-agent profiles"
+notion = "Interact with Notion: query databases, read/create/update pages, and search the workspace."
+pdf_read = "Extract plain text from a PDF file in the workspace. Returns all readable text. Image-only or encrypted PDFs return an empty result. Requires the 'rag-pdf' build feature."
+project_intel = "Project delivery intelligence: generate status reports, detect risks, draft client updates, summarize sprints, and estimate effort. Read-only analysis tool."
+proxy_config = "Manage ZeroClaw proxy settings (scope: environment | zeroclaw | services), including runtime and process env application"
+pushover = "Send a Pushover notification to your device. Requires PUSHOVER_TOKEN and PUSHOVER_USER_KEY in .env file."
+schedule = """Manage scheduled shell-only tasks. Actions: create/add/once/list/get/cancel/remove/pause/resume. WARNING: This tool creates shell jobs whose output is only logged, NOT delivered to any channel. To send a scheduled message to Discord/Telegram/Slack/Matrix, use the cron_add tool with job_type='agent' and a delivery config like {"mode":"announce","channel":"discord","to":"<channel_id>"}."""
+screenshot = "Capture a screenshot of the current screen. Returns the file path and base64-encoded PNG data."
+security_ops = "Security operations tool for managed cybersecurity services. Actions: triage_alert (classify/prioritize alerts), run_playbook (execute incident response steps), parse_vulnerability (parse scan results), generate_report (create security posture reports), list_playbooks (list available playbooks), alert_stats (summarize alert metrics)."
+shell = "Execute a shell command in the workspace directory"
+sop_advance = "Report the result of the current SOP step and advance to the next step. Provide the run_id, whether the step succeeded or failed, and a brief output summary."
+sop_approve = "Approve a pending SOP step that is waiting for operator approval. Returns the step instruction to execute. Use sop_status to see which runs are waiting."
+sop_execute = "Manually trigger a Standard Operating Procedure (SOP) by name. Returns the run ID and first step instruction. Use sop_list to see available SOPs."
+sop_list = "List all loaded Standard Operating Procedures (SOPs) with their triggers, priority, step count, and active run count. Optionally filter by name or priority."
+sop_status = "Query SOP execution status. Provide run_id for a specific run, or sop_name to list runs for that SOP. With no arguments, shows all active runs."
+swarm = "Orchestrate a swarm of agents to collaboratively handle a task. Supports sequential (pipeline), parallel (fan-out/fan-in), and router (LLM-selected) strategies."
+tool_search = """Fetch full schema definitions for deferred MCP tools so they can be called. Use "select:name1,name2" for exact match or keywords to search."""
+web_fetch = "Fetch a web page and return its content as clean plain text. HTML pages are automatically converted to readable text. JSON and plain text responses are returned as-is. Only GET requests; follows redirects. Security: allowlist-only domains, no local/private hosts."
+web_search_tool = "Search the web for information. Returns relevant search results with titles, URLs, and descriptions. Use this to find current information, news, or research topics."
+workspace = "Manage multi-client workspaces. Subcommands: list, switch, create, info, export. Each workspace provides isolated memory, audit, secrets, and tool restrictions."

tool_descriptions/zh-CN.toml

@@ -0,0 +1,60 @@
+# 中文工具描述 (简体中文)
+#
+# [tools] 下的每个键对应工具的 name() 返回值。
+# 值是显示在系统提示中的人类可读描述。
+# 缺少的键将回退到英文 (en.toml) 描述。
+
+[tools]
+backup = "创建、列出、验证和恢复工作区备份"
+browser = "基于可插拔后端（agent-browser、rust-native、computer_use）的网页/浏览器自动化。支持 DOM 操作以及通过 computer-use 辅助工具进行的可选系统级操作（mouse_move、mouse_click、mouse_drag、key_type、key_press、screen_capture）。使用 'snapshot' 将交互元素映射到引用（@e1、@e2）。对 open 操作强制执行 browser.allowed_domains。"
+browser_delegate = "将基于浏览器的任务委派给具有浏览器功能的 CLI，用于与 Teams、Outlook、Jira、Confluence 等 Web 应用交互"
+browser_open = "在系统浏览器中打开经批准的 HTTPS URL。安全约束：仅允许列表域名，禁止本地/私有主机，禁止抓取。"
+cloud_ops = "云转型咨询工具。分析 IaC 计划、评估迁移路径、审查成本，并根据良好架构框架支柱检查架构。只读：不创建或修改云资源。"
+cloud_patterns = "云模式库。根据工作负载描述，建议适用的云原生架构模式（容器化、无服务器、数据库现代化等）。"
+composio = "通过 Composio 在 1000 多个应用上执行操作（Gmail、Notion、GitHub、Slack 等）。使用 action='list' 查看可用操作（包含参数名称）。使用 action='execute' 配合 action_name/tool_slug 和 params 运行操作。如果不确定具体参数，可传入 'text' 并用自然语言描述需求（Composio 将通过 NLP 解析正确参数）。使用 action='list_accounts' 或 action='connected_accounts' 列出 OAuth 已连接账户。使用 action='connect' 配合 app/auth_config_id 获取 OAuth URL。省略时自动解析 connected_account_id。"
+content_search = "在工作区内按正则表达式搜索文件内容。支持 ripgrep (rg)，可回退到 grep。输出模式：'content'（带上下文的匹配行）、'files_with_matches'（仅文件路径）、'count'（每个文件的匹配计数）。"
+cron_add = "创建带有 cron/at/every 计划的定时任务（shell 或 agent）。使用 job_type='agent' 配合 prompt 按计划运行 AI 代理。要将输出发送到频道（Discord、Telegram、Slack、Mattermost、Matrix），请设置 delivery 配置。这是通过频道向用户发送定时/延迟消息的首选工具。"
+cron_list = "列出所有已计划的 cron 任务"
+cron_remove = "按 ID 删除 cron 任务"
+cron_run = "立即强制运行 cron 任务并记录运行历史"
+cron_runs = "列出 cron 任务的最近运行历史"
+cron_update = "修改现有 cron 任务（计划、命令、提示、启用状态、投递配置、模型等）"
+data_management = "工作区数据保留、清理和存储统计"
+delegate = "将子任务委派给专用代理。适用场景：任务受益于不同模型（如快速摘要、深度推理、代码生成）。子代理默认运行单个提示；设置 agentic=true 后可通过过滤的工具调用循环进行迭代。"
+file_edit = "通过替换精确匹配的字符串来编辑文件"
+file_read = "读取带行号的文件内容。支持通过 offset 和 limit 进行部分读取。可从 PDF 提取文本；其他二进制文件使用有损 UTF-8 转换读取。"
+file_write = "将内容写入工作区中的文件"
+git_operations = "执行结构化的 Git 操作（status、diff、log、branch、commit、add、checkout、stash）。提供解析后的 JSON 输出，并与安全策略集成以实现自主控制。"
+glob_search = "在工作区内搜索匹配 glob 模式的文件。返回相对于工作区根目录的排序文件路径列表。示例：'**/*.rs'（所有 Rust 文件）、'src/**/mod.rs'（src 中所有 mod.rs）。"
+google_workspace = "与 Google Workspace 服务（Drive、Gmail、Calendar、Sheets、Docs 等）交互。通过 gws CLI 操作，需要 gws 已安装并认证。"
+hardware_board_info = "返回已连接硬件的完整板卡信息（芯片、架构、内存映射）。适用场景：用户询问板卡信息、连接的硬件、芯片信息等。"
+hardware_memory_map = "返回已连接硬件的内存映射（Flash 和 RAM 地址范围）。适用场景：用户询问内存地址、地址空间或可读地址。返回数据手册中的 Flash/RAM 范围。"
+hardware_memory_read = "通过 USB 从 Nucleo 读取实际内存/寄存器值。适用场景：用户要求读取寄存器值、读取内存地址、转储内存等。返回十六进制转储。需要 Nucleo 通过 USB 连接并启用 probe 功能。"
+http_request = "向外部 API 发送 HTTP 请求。支持 GET、POST、PUT、DELETE、PATCH、HEAD、OPTIONS 方法。安全约束：仅允许列表域名，禁止本地/私有主机，可配置超时和响应大小限制。"
+image_info = "读取图片文件元数据（格式、尺寸、大小），可选返回 base64 编码数据。"
+knowledge = "管理架构决策、解决方案模式、经验教训和专家的知识图谱。操作：capture、search、relate、suggest、expert_find、lessons_extract、graph_stats。"
+linkedin = "管理 LinkedIn：创建帖子、列出帖子、评论、点赞、删除帖子、查看互动数据、获取个人资料信息，以及阅读配置的内容策略。需要在 .env 文件中配置 LINKEDIN_* 凭据。"
+memory_forget = "按键删除记忆。用于删除过时事实或敏感数据。返回记忆是否被找到并删除。"
+memory_recall = "在长期记忆中搜索相关事实、偏好或上下文。返回按相关性排名的评分结果。"
+memory_store = "在长期记忆中存储事实、偏好或笔记。使用类别 'core' 存储永久事实，'daily' 存储会话笔记，'conversation' 存储聊天上下文，或使用自定义类别名称。"
+microsoft365 = "Microsoft 365 集成：通过 Microsoft Graph API 管理 Outlook 邮件、Teams 消息、日历事件、OneDrive 文件和 SharePoint 搜索"
+model_routing_config = "管理默认模型设置、基于场景的提供商/模型路由、分类规则和委派子代理配置"
+notion = "与 Notion 交互：查询数据库、读取/创建/更新页面、搜索工作区。"
+pdf_read = "从工作区中的 PDF 文件提取纯文本。返回所有可读文本。仅图片或加密的 PDF 返回空结果。需要 'rag-pdf' 构建功能。"
+project_intel = "项目交付智能：生成状态报告、检测风险、起草客户更新、总结冲刺、估算工作量。只读分析工具。"
+proxy_config = "管理 ZeroClaw 代理设置（范围：environment | zeroclaw | services），包括运行时和进程环境应用"
+pushover = "向设备发送 Pushover 通知。需要在 .env 文件中配置 PUSHOVER_TOKEN 和 PUSHOVER_USER_KEY。"
+schedule = "管理仅限 shell 的定时任务。操作：create/add/once/list/get/cancel/remove/pause/resume。警告：此工具创建的 shell 任务输出仅记录日志，不会发送到任何频道。要向 Discord/Telegram/Slack/Matrix 发送定时消息，请使用 cron_add 工具。"
+screenshot = "捕获当前屏幕截图。返回文件路径和 base64 编码的 PNG 数据。"
+security_ops = "托管网络安全服务的安全运营工具。操作：triage_alert（分类/优先级排序警报）、run_playbook（执行事件响应步骤）、parse_vulnerability（解析扫描结果）、generate_report（创建安全态势报告）、list_playbooks（列出可用剧本）、alert_stats（汇总警报指标）。"
+shell = "在工作区目录中执行 shell 命令"
+sop_advance = "报告当前 SOP 步骤的结果并前进到下一步。提供 run_id、步骤是否成功或失败，以及简短的输出摘要。"
+sop_approve = "批准等待操作员批准的待处理 SOP 步骤。返回要执行的步骤指令。使用 sop_status 查看哪些运行正在等待。"
+sop_execute = "按名称手动触发标准操作程序 (SOP)。返回运行 ID 和第一步指令。使用 sop_list 查看可用 SOP。"
+sop_list = "列出所有已加载的标准操作程序 (SOP)，包括触发器、优先级、步骤数和活跃运行数。可按名称或优先级筛选。"
+sop_status = "查询 SOP 执行状态。提供 run_id 查看特定运行，或提供 sop_name 列出该 SOP 的所有运行。无参数时显示所有活跃运行。"
+swarm = "编排代理群以协作处理任务。支持顺序（管道）、并行（扇出/扇入）和路由器（LLM 选择）策略。"
+tool_search = "获取延迟 MCP 工具的完整 schema 定义以便调用。使用 \"select:name1,name2\" 精确匹配或关键词搜索。"
+web_fetch = "获取网页并以纯文本形式返回内容。HTML 页面自动转换为可读文本。JSON 和纯文本响应按原样返回。仅 GET 请求；跟随重定向。安全：仅允许列表域名，禁止本地/私有主机。"
+web_search_tool = "搜索网络获取信息。返回包含标题、URL 和描述的相关搜索结果。用于查找当前信息、新闻或研究主题。"
+workspace = "管理多客户端工作区。子命令：list、switch、create、info、export。每个工作区提供隔离的记忆、审计、密钥和工具限制。"