fix gemini-schema

Fix Claude Code provider setup flow: wizard now shows Detect button instead of API key input for keyless providers, TUI wizards include claude-code in provider list, credentials detection checks both .credentials.json paths, subprocess env_clear prevents API key leaks. Fixes #376 #303.

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -0,0 +1,62 @@
+name: Bug Report
+description: Report a bug or unexpected behavior
+labels: ["bug"]
+body:
+  - type: textarea
+    id: description
+    attributes:
+      label: Description
+      description: What happened?
+      placeholder: Describe the bug clearly and concisely.
+    validations:
+      required: true
+
+  - type: textarea
+    id: expected
+    attributes:
+      label: Expected Behavior
+      description: What did you expect to happen?
+    validations:
+      required: true
+
+  - type: textarea
+    id: steps
+    attributes:
+      label: Steps to Reproduce
+      description: How can we reproduce this?
+      placeholder: |
+        1. Run `openfang start`
+        2. Open dashboard at http://localhost:4200
+        3. Click ...
+    validations:
+      required: true
+
+  - type: input
+    id: version
+    attributes:
+      label: OpenFang Version
+      description: Output of `openfang -V`
+      placeholder: "0.3.23"
+    validations:
+      required: true
+
+  - type: dropdown
+    id: os
+    attributes:
+      label: Operating System
+      options:
+        - Linux (x86_64)
+        - Linux (aarch64/ARM64)
+        - macOS (Apple Silicon)
+        - macOS (Intel)
+        - Windows
+        - Android (Termux)
+        - Other
+    validations:
+      required: true
+
+  - type: textarea
+    id: logs
+    attributes:
+      label: Logs / Screenshots
+      description: Paste relevant logs or attach screenshots.

.github/ISSUE_TEMPLATE/feature_request.yml

@@ -0,0 +1,24 @@
+name: Feature Request
+description: Suggest a new feature or improvement
+labels: ["enhancement"]
+body:
+  - type: textarea
+    id: description
+    attributes:
+      label: Description
+      description: What feature would you like?
+      placeholder: Describe the feature and why it would be useful.
+    validations:
+      required: true
+
+  - type: textarea
+    id: alternatives
+    attributes:
+      label: Alternatives Considered
+      description: Have you tried any workarounds?
+
+  - type: textarea
+    id: context
+    attributes:
+      label: Additional Context
+      description: Any other context, screenshots, or references.

.github/dependabot.yml

@@ -0,0 +1,17 @@
+version: 2
+updates:
+  - package-ecosystem: "cargo"
+    directory: "/"
+    schedule:
+      interval: "weekly"
+    open-pull-requests-limit: 5
+    labels:
+      - "dependencies"
+
+  - package-ecosystem: "github-actions"
+    directory: "/"
+    schedule:
+      interval: "weekly"
+    open-pull-requests-limit: 3
+    labels:
+      - "ci"

.github/pull_request_template.md

@@ -0,0 +1,19 @@
+## Summary
+
+<!-- What does this PR do? Link related issues with "Fixes #123". -->
+
+## Changes
+
+<!-- Brief list of what changed. -->
+
+## Testing
+
+- [ ] `cargo clippy --workspace --all-targets -- -D warnings` passes
+- [ ] `cargo test --workspace` passes
+- [ ] Live integration tested (if applicable)
+
+## Security
+
+- [ ] No new unsafe code
+- [ ] No secrets or API keys in diff
+- [ ] User input validated at boundaries

.github/workflows/release.yml

@@ -181,6 +181,11 @@ jobs:
       - name: Build CLI
         if: matrix.target != 'aarch64-unknown-linux-gnu'
         run: cargo build --release --target ${{ matrix.target }} --bin openfang
+      - name: Ad-hoc codesign CLI binary (macOS)
+        if: runner.os == 'macOS'
+        run: |
+          xattr -cr target/${{ matrix.target }}/release/openfang || true
+          codesign --force --sign - target/${{ matrix.target }}/release/openfang
       - name: Package (Unix)
         if: matrix.archive == 'tar.gz'
         run: |

Cargo.toml

@@ -18,7 +18,7 @@ members = [
 ]
 
 [workspace.package]
-version = "0.3.10"
+version = "0.3.42"
 edition = "2021"
 license = "Apache-2.0 OR MIT"
 repository = "https://github.com/RightNow-AI/openfang"
@@ -49,6 +49,7 @@ tracing-subscriber = { version = "0.3", features = ["env-filter", "json"] }
 
 # Time
 chrono = { version = "0.4", features = ["serde"] }
+chrono-tz = "0.10"
 
 # IDs
 uuid = { version = "1", features = ["v4", "serde"] }
@@ -119,9 +120,15 @@ colored = "3"
 aes-gcm = "0.10"
 argon2 = "0.5"
 
+# HTML entity decoding
+html-escape = "0.2"
+
 # Lightweight regex
 regex-lite = "0.1"
 
+# Socket options (SO_REUSEADDR)
+socket2 = "0.5"
+
 # Zip archive extraction
 zip = { version = "2", default-features = false, features = ["deflate"] }
 

README.md

@@ -19,7 +19,7 @@
 <p align="center">
   <img src="https://img.shields.io/badge/language-Rust-orange?style=flat-square" alt="Rust" />
   <img src="https://img.shields.io/badge/license-MIT-blue?style=flat-square" alt="MIT" />
-  <img src="https://img.shields.io/badge/version-0.1.0-green?style=flat-square" alt="v0.1.0" />
+  <img src="https://img.shields.io/badge/version-0.3.30-green?style=flat-square" alt="v0.3.30" />
   <img src="https://img.shields.io/badge/tests-1,767%2B%20passing-brightgreen?style=flat-square" alt="Tests" />
   <img src="https://img.shields.io/badge/clippy-0%20warnings-brightgreen?style=flat-square" alt="Clippy" />
   <a href="https://www.buymeacoffee.com/openfang" target="_blank"><img src="https://img.shields.io/badge/Buy%20Me%20a%20Coffee-FFDD00?style=flat-square&logo=buy-me-a-coffee&logoColor=black" alt="Buy Me A Coffee" /></a>
@@ -27,9 +27,9 @@
 
 ---
 
-> **v0.1.0 — First Release (February 2026)**
+> **v0.3.30 — Security Hardening Release (March 2026)**
 >
-> OpenFang is feature-complete but this is the first public release. You may encounter instability, rough edges, or breaking changes between minor versions. We ship fast and fix fast. Pin to a specific commit for production use until v1.0. [Report issues here.](https://github.com/RightNow-AI/openfang/issues)
+> OpenFang is feature-complete but still pre-1.0. You may encounter rough edges or breaking changes between minor versions. We ship fast and fix fast. Pin to a specific commit for production use until v1.0. [Report issues here.](https://github.com/RightNow-AI/openfang/issues)
 
 ---
 
@@ -371,7 +371,7 @@ cargo fmt --all -- --check
 
 ## Stability Notice
 
-OpenFang v0.1.0 is the first public release. The architecture is solid, the test suite is comprehensive, and the security model is comprehensive. That said:
+OpenFang v0.3.30 is pre-1.0. The architecture is solid, the test suite is comprehensive, and the security model is comprehensive. That said:
 
 - **Breaking changes** may occur between minor versions until v1.0
 - **Some Hands** are more mature than others (Browser and Researcher are the most battle-tested)
@@ -382,6 +382,12 @@ We ship fast and fix fast. The goal is a rock-solid v1.0 by mid-2026.
 
 ---
 
+## Security
+
+To report a security vulnerability, email **jaber@rightnowai.co**. We take all reports seriously and will respond within 48 hours.
+
+---
+
 ## License
 
 MIT — use it however you want.

SECURITY.md

@@ -4,7 +4,7 @@
 
 | Version | Supported          |
 |---------|--------------------|
-| 0.1.x   | :white_check_mark: |
+| 0.3.x   | :white_check_mark: |
 
 ## Reporting a Vulnerability
 
@@ -14,7 +14,7 @@ If you discover a security vulnerability in OpenFang, please report it responsib
 
 ### How to Report
 
-1. Email: **security@openfang.ai**
+1. Email: **jaber@rightnowai.co**
 2. Include:
    - Description of the vulnerability
    - Steps to reproduce

crates/openfang-api/Cargo.toml

@@ -33,9 +33,10 @@ governor = { workspace = true }
 tokio-stream = { workspace = true }
 subtle = { workspace = true }
 base64 = { workspace = true }
+socket2 = { workspace = true }
+reqwest = { workspace = true }
 
 [dev-dependencies]
 tokio-test = { workspace = true }
-reqwest = { workspace = true }
 tempfile = { workspace = true }
 uuid = { workspace = true }

crates/openfang-api/src/channel_bridge.rs

@@ -73,6 +73,33 @@ impl ChannelBridgeHandle for KernelBridgeAdapter {
         Ok(result.response)
     }
 
+    async fn send_message_with_blocks(
+        &self,
+        agent_id: AgentId,
+        blocks: Vec<openfang_types::message::ContentBlock>,
+    ) -> Result<String, String> {
+        // Extract text for the message parameter (used for memory recall / logging)
+        let text: String = blocks
+            .iter()
+            .filter_map(|b| match b {
+                openfang_types::message::ContentBlock::Text { text } => Some(text.as_str()),
+                _ => None,
+            })
+            .collect::<Vec<_>>()
+            .join("\n");
+        let text = if text.is_empty() {
+            "[Image]".to_string()
+        } else {
+            text
+        };
+        let result = self
+            .kernel
+            .send_message_with_blocks(agent_id, &text, blocks)
+            .await
+            .map_err(|e| format!("{e}"))?;
+        Ok(result.response)
+    }
+
     async fn find_agent_by_name(&self, name: &str) -> Result<Option<AgentId>, String> {
         Ok(self.kernel.registry.find_by_name(name).map(|e| e.id))
     }
@@ -648,7 +675,16 @@ impl ChannelBridgeHandle for KernelBridgeAdapter {
         self.kernel
             .set_agent_model(agent_id, model)
             .map_err(|e| format!("{e}"))?;
-        Ok(format!("Model switched to: {model}"))
+        // Read back resolved model+provider from registry
+        let entry = self
+            .kernel
+            .registry
+            .get(agent_id)
+            .ok_or_else(|| "Agent not found after model switch".to_string())?;
+        Ok(format!(
+            "Model switched to: {} (provider: {})",
+            entry.manifest.model.model, entry.manifest.model.provider
+        ))
     }
 
     async fn stop_run(&self, agent_id: AgentId) -> Result<String, String> {
@@ -1030,6 +1066,7 @@ pub async fn start_channel_bridge_with_config(
                 token,
                 tg_config.allowed_users.clone(),
                 poll_interval,
+                tg_config.api_url.clone(),
             ));
             adapters.push((adapter, tg_config.default_agent.clone()));
         }
@@ -1041,6 +1078,8 @@ pub async fn start_channel_bridge_with_config(
             let adapter = Arc::new(DiscordAdapter::new(
                 token,
                 dc_config.allowed_guilds.clone(),
+                dc_config.allowed_users.clone(),
+                dc_config.ignore_bots,
                 dc_config.intents,
             ));
             adapters.push((adapter, dc_config.default_agent.clone()));
@@ -1620,6 +1659,35 @@ pub async fn reload_channels_from_disk(
         *guard = None;
     }
 
+    // Re-read secrets.env so new API tokens are available in std::env
+    let secrets_path = state.kernel.config.home_dir.join("secrets.env");
+    if secrets_path.exists() {
+        if let Ok(content) = std::fs::read_to_string(&secrets_path) {
+            for line in content.lines() {
+                let trimmed = line.trim();
+                if trimmed.is_empty() || trimmed.starts_with('#') {
+                    continue;
+                }
+                if let Some(eq_pos) = trimmed.find('=') {
+                    let key = trimmed[..eq_pos].trim();
+                    let mut value = trimmed[eq_pos + 1..].trim().to_string();
+                    if !key.is_empty() {
+                        // Strip matching quotes
+                        if ((value.starts_with('"') && value.ends_with('"'))
+                            || (value.starts_with('\'') && value.ends_with('\'')))
+                            && value.len() >= 2
+                        {
+                            value = value[1..value.len() - 1].to_string();
+                        }
+                        // Always overwrite — the file is the source of truth after dashboard edits
+                        std::env::set_var(key, &value);
+                    }
+                }
+            }
+            info!("Reloaded secrets.env for channel hot-reload");
+        }
+    }
+
     // Re-read config from disk
     let config_path = state.kernel.config.home_dir.join("config.toml");
     let fresh_config = openfang_kernel::config::load_config(Some(&config_path));

crates/openfang-api/src/middleware.rs

@@ -45,39 +45,17 @@ pub async fn request_logging(request: Request<Body>, next: Next) -> Response<Bod
 
 /// Bearer token authentication middleware.
 ///
-/// When `api_key` is non-empty, all requests must include
-/// `Authorization: Bearer <api_key>`. If the key is empty, auth is bypassed.
+/// When `api_key` is non-empty (after trimming), requests to non-public
+/// endpoints must include `Authorization: Bearer <api_key>`.
+/// If the key is empty or whitespace-only, auth is disabled entirely
+/// (public/local development mode).
 pub async fn auth(
     axum::extract::State(api_key): axum::extract::State<String>,
     request: Request<Body>,
     next: Next,
 ) -> Response<Body> {
-    // If no API key configured, restrict to loopback addresses only.
-    if api_key.is_empty() {
-        let is_loopback = request
-            .extensions()
-            .get::<axum::extract::ConnectInfo<std::net::SocketAddr>>()
-            .map(|ci| ci.0.ip().is_loopback())
-            .unwrap_or(false);
-
-        if !is_loopback {
-            tracing::warn!(
-                "Rejected non-localhost request: no API key configured. \
-                 Set api_key in config.toml for remote access."
-            );
-            return Response::builder()
-                .status(StatusCode::FORBIDDEN)
-                .header("content-type", "application/json")
-                .body(Body::from(
-                    serde_json::json!({
-                        "error": "No API key configured. Remote access denied. Configure api_key in ~/.openfang/config.toml"
-                    })
-                    .to_string(),
-                ))
-                .unwrap_or_default();
-        }
-        return next.run(request).await;
-    }
+    // SECURITY: Capture method early for method-aware public endpoint checks.
+    let method = request.method().clone();
 
     // Shutdown is loopback-only (CLI on same machine) — skip token auth
     let path = request.uri().path();
@@ -86,52 +64,67 @@ pub async fn auth(
             .extensions()
             .get::<axum::extract::ConnectInfo<std::net::SocketAddr>>()
             .map(|ci| ci.0.ip().is_loopback())
-            .unwrap_or(true); // default true for unix sockets / tests
+            .unwrap_or(false); // SECURITY: default-deny — unknown origin is NOT loopback
         if is_loopback {
             return next.run(request).await;
         }
     }
 
-    // Public endpoints that don't require auth (dashboard needs these)
-    if path == "/"
+    // Public endpoints that don't require auth (dashboard needs these).
+    // SECURITY: /api/agents is GET-only (listing). POST (spawn) requires auth.
+    // SECURITY: Public endpoints are GET-only unless explicitly noted.
+    // POST/PUT/DELETE to any endpoint ALWAYS requires auth to prevent
+    // unauthenticated writes (cron job creation, skill install, etc.).
+    let is_get = method == axum::http::Method::GET;
+    let is_public = path == "/"
         || path == "/logo.png"
         || path == "/favicon.ico"
-        || path == "/.well-known/agent.json"
-        || path.starts_with("/a2a/")
+        || (path == "/.well-known/agent.json" && is_get)
+        || (path.starts_with("/a2a/") && is_get)
         || path == "/api/health"
         || path == "/api/health/detail"
         || path == "/api/status"
         || path == "/api/version"
-        || path == "/api/agents"
-        || path == "/api/profiles"
-        || path == "/api/config"
-        || path.starts_with("/api/uploads/")
+        || (path == "/api/agents" && is_get)
+        || (path == "/api/profiles" && is_get)
+        || (path == "/api/config" && is_get)
+        || (path == "/api/config/schema" && is_get)
+        || (path.starts_with("/api/uploads/") && is_get)
         // Dashboard read endpoints — allow unauthenticated so the SPA can
         // render before the user enters their API key.
-        || path == "/api/models"
-        || path == "/api/models/aliases"
-        || path == "/api/providers"
-        || path == "/api/budget"
-        || path == "/api/budget/agents"
-        || path.starts_with("/api/budget/agents/")
-        || path == "/api/network/status"
-        || path == "/api/a2a/agents"
-        || path == "/api/approvals"
-        || path.starts_with("/api/approvals/")
-        || path == "/api/channels"
-        || path == "/api/hands"
-        || path == "/api/hands/active"
-        || path.starts_with("/api/hands/")
-        || path == "/api/skills"
-        || path == "/api/sessions"
-        || path == "/api/integrations"
-        || path == "/api/integrations/available"
-        || path == "/api/integrations/health"
-        || path == "/api/workflows"
-        || path == "/api/logs/stream"
-        || path.starts_with("/api/cron/")
-        || path.starts_with("/api/providers/github-copilot/oauth/")
-    {
+        || (path == "/api/models" && is_get)
+        || (path == "/api/models/aliases" && is_get)
+        || (path == "/api/providers" && is_get)
+        || (path == "/api/budget" && is_get)
+        || (path == "/api/budget/agents" && is_get)
+        || (path.starts_with("/api/budget/agents/") && is_get)
+        || (path == "/api/network/status" && is_get)
+        || (path == "/api/a2a/agents" && is_get)
+        || (path == "/api/approvals" && is_get)
+        || (path.starts_with("/api/approvals/") && is_get)
+        || (path == "/api/channels" && is_get)
+        || (path == "/api/hands" && is_get)
+        || (path == "/api/hands/active" && is_get)
+        || (path.starts_with("/api/hands/") && is_get)
+        || (path == "/api/skills" && is_get)
+        || (path == "/api/sessions" && is_get)
+        || (path == "/api/integrations" && is_get)
+        || (path == "/api/integrations/available" && is_get)
+        || (path == "/api/integrations/health" && is_get)
+        || (path == "/api/workflows" && is_get)
+        || path == "/api/logs/stream"  // SSE stream, read-only
+        || (path.starts_with("/api/cron/") && is_get)
+        || path.starts_with("/api/providers/github-copilot/oauth/");
+
+    if is_public {
+        return next.run(request).await;
+    }
+
+    // If no API key configured (empty, whitespace-only, or missing), skip auth
+    // entirely. Users who don't set api_key accept that all endpoints are open.
+    // To secure the dashboard, set a non-empty api_key in config.toml.
+    let api_key = api_key.trim();
+    if api_key.is_empty() {
         return next.run(request).await;
     }
 
@@ -218,6 +211,10 @@ pub async fn security_headers(request: Request<Body>, next: Next) -> Response<Bo
         "cache-control",
         "no-store, no-cache, must-revalidate".parse().unwrap(),
     );
+    headers.insert(
+        "strict-transport-security",
+        "max-age=63072000; includeSubDomains".parse().unwrap(),
+    );
     response
 }
 

crates/openfang-api/src/openai_compat.rs

@@ -179,9 +179,8 @@ fn resolve_agent(state: &AppState, model: &str) -> Option<(AgentId, String)> {
         return Some((entry.id, entry.name.clone()));
     }
 
-    // 4. Fallback → first registered agent
-    let agents = state.kernel.registry.list();
-    agents.first().map(|e| (e.id, e.name.clone()))
+    // No match — return None so the caller returns a proper 404
+    None
 }
 
 // ── Message conversion ──────────────────────────────────────────────────────
@@ -336,7 +335,7 @@ pub async fn chat_completions(
                     index: 0,
                     message: ChoiceMessage {
                         role: "assistant",
-                        content: Some(result.response),
+                        content: Some(crate::ws::strip_think_tags(&result.response)),
                         tool_calls: None,
                     },
                     finish_reason: "stop",

crates/openfang-api/src/server.rs

@@ -50,11 +50,12 @@ pub async fn build_router(
         channels_config: tokio::sync::RwLock::new(channels_config),
         shutdown_notify: Arc::new(tokio::sync::Notify::new()),
         clawhub_cache: dashmap::DashMap::new(),
+        provider_probe_cache: openfang_runtime::provider_health::ProbeCache::new(),
     });
 
     // CORS: allow localhost origins by default. If API key is set, the API
     // is protected anyway. For development, permissive CORS is convenient.
-    let cors = if state.kernel.config.api_key.is_empty() {
+    let cors = if state.kernel.config.api_key.trim().is_empty() {
         // No auth → restrict CORS to localhost origins (include both 127.0.0.1 and localhost)
         let port = listen_addr.port();
         let mut origins: Vec<axum::http::HeaderValue> = vec![
@@ -102,7 +103,8 @@ pub async fn build_router(
             .allow_headers(tower_http::cors::Any)
     };
 
-    let api_key = state.kernel.config.api_key.clone();
+    // Trim whitespace so `api_key = ""` or `api_key = "  "` both disable auth.
+    let api_key = state.kernel.config.api_key.trim().to_string();
     let gcra_limiter = rate_limiter::create_rate_limiter();
 
     let app = Router::new()
@@ -126,7 +128,7 @@ pub async fn build_router(
         )
         .route(
             "/api/agents/{id}",
-            axum::routing::get(routes::get_agent).delete(routes::kill_agent),
+            axum::routing::get(routes::get_agent).delete(routes::kill_agent).patch(routes::patch_agent),
         )
         .route(
             "/api/agents/{id}/mode",
@@ -740,7 +742,8 @@ pub async fn run_daemon(
         if info_path.exists() {
             if let Ok(existing) = std::fs::read_to_string(info_path) {
                 if let Ok(info) = serde_json::from_str::<DaemonInfo>(&existing) {
-                    if is_process_alive(info.pid) {
+                    // PID alive AND the health endpoint responds → truly running
+                    if is_process_alive(info.pid) && is_daemon_responding(&info.listen_addr) {
                         return Err(format!(
                             "Another daemon (PID {}) is already running at {}",
                             info.pid, info.listen_addr
@@ -749,7 +752,8 @@ pub async fn run_daemon(
                     }
                 }
             }
-            // Stale PID file, remove it
+            // Stale PID file (process dead or different process reused PID), remove it
+            info!("Removing stale daemon info file");
             let _ = std::fs::remove_file(info_path);
         }
 
@@ -771,7 +775,22 @@ pub async fn run_daemon(
     info!("WebChat UI available at http://{addr}/",);
     info!("WebSocket endpoint: ws://{addr}/api/agents/{{id}}/ws",);
 
-    let listener = tokio::net::TcpListener::bind(addr).await?;
+    // Use SO_REUSEADDR to allow binding immediately after reboot (avoids TIME_WAIT).
+    let socket = socket2::Socket::new(
+        if addr.is_ipv4() {
+            socket2::Domain::IPV4
+        } else {
+            socket2::Domain::IPV6
+        },
+        socket2::Type::STREAM,
+        None,
+    )?;
+    socket.set_reuse_address(true)?;
+    socket.set_nonblocking(true)?;
+    socket.bind(&addr.into())?;
+    socket.listen(1024)?;
+    let listener =
+        tokio::net::TcpListener::from_std(std::net::TcpListener::from(socket))?;
 
     // Run server with graceful shutdown.
     // SECURITY: `into_make_service_with_connect_info` injects the peer
@@ -891,3 +910,26 @@ fn is_process_alive(pid: u32) -> bool {
         false
     }
 }
+
+/// Check if an OpenFang daemon is actually responding at the given address.
+/// This avoids false positives where a different process reused the same PID
+/// after a system reboot.
+fn is_daemon_responding(addr: &str) -> bool {
+    // Quick TCP connect check — don't make a full HTTP request to avoid delays
+    let addr_only = addr
+        .strip_prefix("http://")
+        .or_else(|| addr.strip_prefix("https://"))
+        .unwrap_or(addr);
+    if let Ok(sock_addr) = addr_only.parse::<std::net::SocketAddr>() {
+        std::net::TcpStream::connect_timeout(
+            &sock_addr,
+            std::time::Duration::from_millis(500),
+        )
+        .is_ok()
+    } else {
+        // Fallback: try connecting to hostname
+        std::net::TcpStream::connect(addr_only)
+            .map(|_| true)
+            .unwrap_or(false)
+    }
+}

crates/openfang-api/src/stream_chunker.rs

@@ -140,9 +140,23 @@ impl StreamChunker {
 }
 
 /// Find the last occurrence of a pattern within a byte range.
+///
+/// Both `range.start` and `range.end` are clamped to the nearest valid UTF-8
+/// char boundary so that slicing never panics on multi-byte content.
 fn find_last_in_range(text: &str, pattern: &str, range: &std::ops::Range<usize>) -> Option<usize> {
-    let search_text = &text[range.start..range.end.min(text.len())];
-    search_text.rfind(pattern).map(|pos| range.start + pos)
+    let len = text.len();
+    // Clamp end to text length and walk back to a char boundary
+    let mut end = range.end.min(len);
+    while end > 0 && !text.is_char_boundary(end) {
+        end -= 1;
+    }
+    // Walk start forward to the nearest char boundary (never past end)
+    let mut start = range.start.min(end);
+    while start < end && !text.is_char_boundary(start) {
+        start += 1;
+    }
+    let search_text = &text[start..end];
+    search_text.rfind(pattern).map(|pos| start + pos)
 }
 
 #[cfg(test)]

crates/openfang-api/src/types.rs

@@ -2,11 +2,16 @@
 
 use serde::{Deserialize, Serialize};
 
-/// Request to spawn an agent from a TOML manifest string.
+/// Request to spawn an agent from a TOML manifest string or a template name.
 #[derive(Debug, Deserialize)]
 pub struct SpawnRequest {
-    /// Agent manifest as TOML string.
+    /// Agent manifest as TOML string (optional if `template` is provided).
+    #[serde(default)]
     pub manifest_toml: String,
+    /// Template name from `~/.openfang/agents/{template}/agent.toml`.
+    /// When provided and `manifest_toml` is empty, the template is loaded automatically.
+    #[serde(default)]
+    pub template: Option<String>,
     /// Optional Ed25519 signed manifest envelope (JSON).
     /// When present, the signature is verified before spawning.
     #[serde(default)]

crates/openfang-api/src/ws.rs

@@ -146,19 +146,30 @@ pub async fn agent_ws(
     uri: axum::http::Uri,
 ) -> impl IntoResponse {
     // SECURITY: Authenticate WebSocket upgrades (bypasses middleware).
-    let api_key = &state.kernel.config.api_key;
+    // Trim whitespace so empty/whitespace-only api_key disables auth.
+    let api_key_raw = &state.kernel.config.api_key;
+    let api_key = api_key_raw.trim();
     if !api_key.is_empty() {
+        // SECURITY: Use constant-time comparison to prevent timing attacks on API key
+        let ct_eq = |token: &str, key: &str| -> bool {
+            use subtle::ConstantTimeEq;
+            if token.len() != key.len() {
+                return false;
+            }
+            token.as_bytes().ct_eq(key.as_bytes()).into()
+        };
+
         let header_auth = headers
             .get("authorization")
             .and_then(|v| v.to_str().ok())
             .and_then(|v| v.strip_prefix("Bearer "))
-            .map(|token| token == api_key)
+            .map(|token| ct_eq(token, api_key))
             .unwrap_or(false);
 
         let query_auth = uri
             .query()
             .and_then(|q| q.split('&').find_map(|pair| pair.strip_prefix("token=")))
-            .map(|token| token == api_key)
+            .map(|token| ct_eq(token, api_key))
             .unwrap_or(false);
 
         if !header_auth && !query_auth {
@@ -621,8 +632,12 @@ async fn handle_text_message(
                                 return;
                             }
 
+                            // Strip <think>...</think> blocks from model output
+                            // (e.g. MiniMax, DeepSeek reasoning tokens)
+                            let cleaned_response = strip_think_tags(&result.response);
+
                             // Guard: ensure we never send an empty response
-                            let content = if result.response.trim().is_empty() {
+                            let content = if cleaned_response.trim().is_empty() {
                                 format!(
                                     "[The agent completed processing but returned no text response. ({} in / {} out | {} iter)]",
                                     result.total_usage.input_tokens,
@@ -630,7 +645,7 @@ async fn handle_text_message(
                                     result.iterations,
                                 )
                             } else {
-                                result.response
+                                cleaned_response
                             };
 
                             // Estimate context pressure from last call
@@ -796,12 +811,19 @@ async fn handle_command(
             } else {
                 match state.kernel.set_agent_model(agent_id, args) {
                     Ok(()) => {
-                        let msg = if let Some(entry) = state.kernel.registry.get(agent_id) {
-                            format!("Model switched to: {} (provider: {})", entry.manifest.model.model, entry.manifest.model.provider)
+                        if let Some(entry) = state.kernel.registry.get(agent_id) {
+                            let model = &entry.manifest.model.model;
+                            let provider = &entry.manifest.model.provider;
+                            serde_json::json!({
+                                "type": "command_result",
+                                "command": cmd,
+                                "message": format!("Model switched to: {model} (provider: {provider})"),
+                                "model": model,
+                                "provider": provider
+                            })
                         } else {
-                            format!("Model switched to: {args}")
-                        };
-                        serde_json::json!({"type": "command_result", "command": cmd, "message": msg})
+                            serde_json::json!({"type": "command_result", "command": cmd, "message": format!("Model switched to: {args}")})
+                        }
                     }
                     Err(e) => {
                         serde_json::json!({"type": "error", "content": format!("Model switch failed: {e}")})
@@ -1114,10 +1136,19 @@ fn classify_streaming_error(err: &openfang_kernel::error::KernelError) -> String
         }
         llm_errors::LlmErrorCategory::Auth => "Verify your API key in config.".to_string(),
         llm_errors::LlmErrorCategory::ModelNotFound => {
-            "Model unavailable. Use /model to see options.".to_string()
+            if inner.contains("localhost:11434") || inner.contains("ollama") {
+                "Model not found on Ollama. Run `ollama pull <model>` to download it, then try again. Use /model to see options.".to_string()
+            } else {
+                "Model unavailable. Use /model to see options or check your provider configuration.".to_string()
+            }
         }
         llm_errors::LlmErrorCategory::Format => {
-            "LLM request failed. Check your API key and model configuration in Settings.".to_string()
+            // Claude Code CLI errors have actionable messages — pass them through
+            if inner.contains("Claude Code CLI") || inner.contains("claude auth") {
+                classified.raw_message.clone()
+            } else {
+                "LLM request failed. Check your API key and model configuration in Settings.".to_string()
+            }
         }
         _ => classified.sanitized_message,
     }
@@ -1152,6 +1183,27 @@ fn extract_status_code(s: &str) -> Option<u16> {
     None
 }
 
+/// Strip `<think>...</think>` blocks from model output.
+///
+/// Some models (MiniMax, DeepSeek, etc.) wrap their reasoning in `<think>` tags.
+/// These are internal chain-of-thought and shouldn't be shown to the user.
+pub fn strip_think_tags(text: &str) -> String {
+    let mut result = String::with_capacity(text.len());
+    let mut remaining = text;
+    while let Some(start) = remaining.find("<think>") {
+        result.push_str(&remaining[..start]);
+        if let Some(end) = remaining[start..].find("</think>") {
+            remaining = &remaining[(start + end + 8)..]; // 8 = "</think>".len()
+        } else {
+            // Unclosed <think> tag — strip to end
+            remaining = "";
+            break;
+        }
+    }
+    result.push_str(remaining);
+    result
+}
+
 // ---------------------------------------------------------------------------
 // Tests
 // ---------------------------------------------------------------------------
@@ -1228,4 +1280,18 @@ mod tests {
     fn test_sanitize_trims_whitespace() {
         assert_eq!(sanitize_user_input("  hello  "), "hello");
     }
+
+    #[test]
+    fn test_strip_think_tags() {
+        assert_eq!(
+            strip_think_tags("<think>reasoning here</think>The answer is 42."),
+            "The answer is 42."
+        );
+        assert_eq!(
+            strip_think_tags("Hello <think>\nsome thinking\n</think> world"),
+            "Hello  world"
+        );
+        assert_eq!(strip_think_tags("No thinking here"), "No thinking here");
+        assert_eq!(strip_think_tags("<think>all thinking</think>"), "");
+    }
 }

crates/openfang-api/static/css/components.css

@@ -69,6 +69,32 @@
   gap: 16px;
 }
 
+/* Card-based flex containers for agent chips and similar inline layouts */
+.card-flex {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 10px;
+}
+
+/* Nested list indentation inside cards, detail panels, and modals */
+.card ul, .card ol,
+.detail-grid ul, .detail-grid ol,
+.modal ul, .modal ol,
+.info-card ul, .info-card ol {
+  padding-left: 18px;
+  margin: 4px 0;
+}
+.card ul ul, .card ol ol,
+.modal ul ul, .modal ol ol {
+  padding-left: 16px;
+  margin: 2px 0;
+}
+.card li, .modal li, .info-card li {
+  margin-bottom: 2px;
+  font-size: 12px;
+  line-height: 1.5;
+}
+
 /* Glow effect on card hover */
 .card-glow {
   overflow: hidden;
@@ -90,13 +116,17 @@
   display: inline-flex;
   align-items: center;
   gap: 4px;
-  padding: 2px 8px;
+  padding: 3px 8px;
   border-radius: 20px;
   font-size: 10px;
   font-weight: 600;
   letter-spacing: 0.5px;
   text-transform: uppercase;
+  white-space: nowrap;
+  line-height: 1.2;
+  vertical-align: middle;
 }
+.badge + .badge { margin-left: 4px; }
 
 .badge-running { background: rgba(74,222,128,0.12); color: var(--success); }
 .badge-suspended { background: rgba(245,158,11,0.12); color: var(--warning); }
@@ -110,7 +140,7 @@
 .badge-error { background: rgba(239,68,68,0.12); color: var(--error); }
 .badge-muted { background: rgba(148,163,184,0.12); color: var(--text-dim); }
 .badge-info { background: rgba(59,130,246,0.12); color: var(--info); }
-.badge-dim { background: rgba(148,163,184,0.08); color: var(--text-dim); font-size: 0.65rem; }
+.badge-dim { background: rgba(148,163,184,0.08); color: var(--text-dim); font-size: 0.65rem; padding: 2px 6px; }
 .text-danger { color: var(--error); }
 
 /* Tables */
@@ -949,6 +979,14 @@ mark.search-highlight {
   padding: 8px 12px;
   border-bottom: 1px solid var(--border);
 }
+.model-switcher-search select {
+  max-width: 100px;
+  flex-shrink: 0;
+}
+.model-switcher-search select:focus {
+  outline: none;
+  border-color: var(--accent);
+}
 .model-switcher-search input {
   flex: 1;
   background: none;

crates/openfang-api/static/css/layout.css

@@ -55,6 +55,11 @@
   transform: scale(1.05);
 }
 
+[data-theme="light"] .sidebar-logo img,
+[data-theme="light"] .message-avatar img {
+  filter: invert(1);
+}
+
 .sidebar-header h1 {
   font-size: 14px;
   font-weight: 700;

crates/openfang-api/static/index_body.html

@@ -4,7 +4,8 @@
 <div x-show="$store.app.showAuthPrompt" style="position:fixed;inset:0;z-index:9999;display:flex;align-items:center;justify-content:center;background:rgba(0,0,0,0.6);backdrop-filter:blur(4px)" x-data="{ apiKeyInput: '' }">
   <div style="background:var(--bg-card,#1e1e2e);border:1px solid var(--border,#333);border-radius:12px;padding:2rem;max-width:400px;width:90%">
     <h3 style="margin:0 0 0.5rem;font-size:1.1rem">API Key Required</h3>
-    <p style="color:var(--text-dim,#888);font-size:0.85rem;margin:0 0 1rem">This instance requires an API key. Enter the key from your <code>config.toml</code>.</p>
+    <p style="color:var(--text-dim,#888);font-size:0.85rem;margin:0 0 0.5rem">This instance requires an API key. Enter the key from your <code>config.toml</code>.</p>
+    <p style="color:var(--text-dim,#666);font-size:0.75rem;margin:0 0 1rem">Add <code style="color:var(--accent-light,#a78bfa);background:var(--bg,#111);padding:1px 4px;border-radius:2px">api_key = "your-key"</code> at the <strong>top</strong> of <code>~/.openfang/config.toml</code> (not under any [section]).</p>
     <input type="password" x-model="apiKeyInput" placeholder="Enter API key..." @keydown.enter="$store.app.submitApiKey(apiKeyInput)" style="width:100%;padding:0.6rem;border-radius:6px;border:1px solid var(--border,#333);background:var(--bg-input,#151520);color:var(--text,#e0e0e0);font-size:0.9rem;box-sizing:border-box;margin-bottom:0.75rem">
     <button @click="$store.app.submitApiKey(apiKeyInput)" style="width:100%;padding:0.6rem;border-radius:6px;border:none;background:var(--accent,#7c3aed);color:#fff;font-weight:600;cursor:pointer;font-size:0.9rem">Unlock Dashboard</button>
   </div>
@@ -735,7 +736,7 @@
                   <span class="text-xs" style="color:var(--danger)" x-text="formatRecordingTime()"></span>
                 </div>
                 <textarea id="msg-input" rows="1" :placeholder="recording ? 'Recording... release to send' : 'Message OpenFang... (/ for commands)'"
-                          @keydown.enter.prevent="if(!$event.shiftKey){if(showModelPicker && filteredModelPicker.length){pickModel(filteredModelPicker[modelPickerIdx].id)}else if(showSlashMenu && filteredSlashCommands.length){executeSlashCommand(filteredSlashCommands[slashIdx].cmd)}else{sendMessage()}}"
+                          @keydown.enter.prevent="if(!$event.isComposing && !$event.shiftKey){if(showModelPicker && filteredModelPicker.length){pickModel(filteredModelPicker[modelPickerIdx].id)}else if(showSlashMenu && filteredSlashCommands.length){executeSlashCommand(filteredSlashCommands[slashIdx].cmd)}else{sendMessage()}}"
                           @keydown.escape="showSlashMenu = false; showModelPicker = false"
                           @keydown.arrow-up.prevent="if(showModelPicker){modelPickerIdx = Math.max(0, modelPickerIdx - 1)}else if(showSlashMenu){slashIdx = Math.max(0, slashIdx - 1)}"
                           @keydown.arrow-down.prevent="if(showModelPicker){modelPickerIdx = Math.min(filteredModelPicker.length - 1, modelPickerIdx + 1)}else if(showSlashMenu){slashIdx = Math.min(filteredSlashCommands.length - 1, slashIdx + 1)}"
@@ -769,6 +770,12 @@
                       <div class="model-switcher-search">
                         <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" style="flex-shrink:0;opacity:0.5"><circle cx="11" cy="11" r="8"/><line x1="21" y1="21" x2="16.65" y2="16.65"/></svg>
                         <input id="model-switcher-search" type="text" x-model="modelSwitcherFilter" placeholder="Search models..." @keydown.escape.stop="showModelSwitcher = false" @keydown.arrow-down.prevent="modelSwitcherIdx = Math.min(modelSwitcherIdx + 1, filteredSwitcherModels.length - 1)" @keydown.arrow-up.prevent="modelSwitcherIdx = Math.max(modelSwitcherIdx - 1, 0)" @keydown.enter.prevent="filteredSwitcherModels[modelSwitcherIdx] && switchModel(filteredSwitcherModels[modelSwitcherIdx])">
+                        <select x-model="modelSwitcherProviderFilter" style="background:var(--surface2);border:1px solid var(--border);border-radius:6px;color:var(--text-dim);font-size:11px;padding:2px 6px;cursor:pointer;font-family:var(--font-mono);flex-shrink:0">
+                          <option value="">All</option>
+                          <template x-for="pn in switcherProviders" :key="pn">
+                            <option :value="pn" x-text="pn"></option>
+                          </template>
+                        </select>
                       </div>
                       <div x-show="modelSwitching" style="display:flex;align-items:center;justify-content:center;padding:12px;gap:8px">
                         <div class="tool-card-spinner"></div>
@@ -841,6 +848,9 @@
                     <div class="text-xs text-dim font-mono" style="font-size:11px" x-text="agent.model_name"></div>
                   </div>
                   <span class="badge" :class="'badge-' + agent.state.toLowerCase()" x-text="agent.state" style="font-size:10px"></span>
+                  <button class="agent-chip-config-btn" @click.stop="showDetail(agent)" title="Agent settings" style="display:flex;align-items:center;justify-content:center;width:28px;height:28px;border-radius:50%;border:1px solid var(--border);background:transparent;cursor:pointer;color:var(--text-dim);transition:all 0.15s;flex-shrink:0" @mouseenter="$el.style.borderColor='var(--accent)';$el.style.color='var(--accent)';$el.style.background='var(--surface2)'" @mouseleave="$el.style.borderColor='var(--border)';$el.style.color='var(--text-dim)';$el.style.background='transparent'">
+                    <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><circle cx="12" cy="12" r="3"/><path d="M19.4 15a1.65 1.65 0 0 0 .33 1.82l.06.06a2 2 0 0 1 0 2.83 2 2 0 0 1-2.83 0l-.06-.06a1.65 1.65 0 0 0-1.82-.33 1.65 1.65 0 0 0-1 1.51V21a2 2 0 0 1-2 2 2 2 0 0 1-2-2v-.09A1.65 1.65 0 0 0 9 19.4a1.65 1.65 0 0 0-1.82.33l-.06.06a2 2 0 0 1-2.83 0 2 2 0 0 1 0-2.83l.06-.06A1.65 1.65 0 0 0 4.68 15a1.65 1.65 0 0 0-1.51-1H3a2 2 0 0 1-2-2 2 2 0 0 1 2-2h.09A1.65 1.65 0 0 0 4.6 9a1.65 1.65 0 0 0-.33-1.82l-.06-.06a2 2 0 0 1 0-2.83 2 2 0 0 1 2.83 0l.06.06A1.65 1.65 0 0 0 9 4.68a1.65 1.65 0 0 0 1-1.51V3a2 2 0 0 1 2-2 2 2 0 0 1 2 2v.09a1.65 1.65 0 0 0 1 1.51 1.65 1.65 0 0 0 1.82-.33l.06-.06a2 2 0 0 1 2.83 0 2 2 0 0 1 0 2.83l-.06.06A1.65 1.65 0 0 0 19.4 9a1.65 1.65 0 0 0 1.51 1H21a2 2 0 0 1 2 2 2 2 0 0 1-2 2h-.09a1.65 1.65 0 0 0-1.51 1z"/></svg>
+                  </button>
                 </div>
               </template>
             </div>
@@ -912,6 +922,36 @@
                       </template>
                     </div>
                     <div class="detail-row"><span class="detail-label">Created</span><span class="detail-value" x-text="detailAgent.created_at ? new Date(detailAgent.created_at).toLocaleString() : '-'"></span></div>
+
+                    <!-- Fallback Model Chain -->
+                    <div class="detail-row" style="align-items:flex-start">
+                      <span class="detail-label">Fallbacks</span>
+                      <div style="flex:1">
+                        <template x-if="detailAgent._fallbacks && detailAgent._fallbacks.length > 0">
+                          <div>
+                            <template x-for="(fb, idx) in detailAgent._fallbacks" :key="idx">
+                              <div class="flex gap-1 items-center" style="margin-bottom:4px">
+                                <span class="badge" style="font-size:11px;font-family:var(--font-mono)" x-text="(idx+1) + '. ' + fb.provider + '/' + fb.model"></span>
+                                <button class="btn btn-ghost btn-sm" style="padding:1px 4px;font-size:10px;color:var(--danger)" @click="removeFallback(idx)">&times;</button>
+                              </div>
+                            </template>
+                          </div>
+                        </template>
+                        <template x-if="!detailAgent._fallbacks || detailAgent._fallbacks.length === 0">
+                          <span class="text-dim" style="font-size:12px">None — add a fallback chain</span>
+                        </template>
+                        <template x-if="!editingFallback">
+                          <button class="btn btn-ghost btn-sm" style="padding:2px 8px;font-size:11px;margin-top:4px" @click="editingFallback = true; newFallbackValue = ''">+ Add</button>
+                        </template>
+                        <template x-if="editingFallback">
+                          <div class="flex gap-1 mt-1" style="align-items:center">
+                            <input class="form-input" style="width:220px;font-size:12px" x-model="newFallbackValue" placeholder="provider/model" @keydown.enter="addFallback()" @keydown.escape="editingFallback = false">
+                            <button class="btn btn-primary btn-sm" @click="addFallback()" style="padding:2px 10px;font-size:11px">Add</button>
+                            <button class="btn btn-ghost btn-sm" @click="editingFallback = false" style="padding:2px 8px;font-size:11px">Cancel</button>
+                          </div>
+                        </template>
+                      </div>
+                    </div>
                   </div>
                   <div class="flex gap-2 mt-4">
                     <button class="btn btn-primary" @click="chatWithAgent(detailAgent); showDetailModal = false">Chat</button>
@@ -1428,7 +1468,7 @@
                 <div>
                   <div class="form-group">
                     <label class="text-xs">Label</label>
-                    <input class="form-input" x-model="selectedNode.label" style="font-size:11px">
+                    <input class="form-input" x-model="selectedNode.label" @input="applyNodeEdit()" style="font-size:11px">
                   </div>
 
                   <!-- Agent config -->
@@ -1436,7 +1476,7 @@
                     <div>
                       <div class="form-group">
                         <label class="text-xs">Agent</label>
-                        <select class="form-select" x-model="selectedNode.config.agent_name" style="font-size:11px">
+                        <select class="form-select" x-model="selectedNode.config.agent_name" @change="applyNodeEdit()" style="font-size:11px">
                           <option value="">Select agent...</option>
                           <template x-for="a in agents" :key="a.id || a.name">
                             <option :value="a.name" x-text="a.name"></option>
@@ -1445,11 +1485,11 @@
                       </div>
                       <div class="form-group">
                         <label class="text-xs">Prompt Template</label>
-                        <textarea class="form-textarea" x-model="selectedNode.config.prompt" style="font-size:11px;min-height:60px" placeholder="{{input}}"></textarea>
+                        <textarea class="form-textarea" x-model="selectedNode.config.prompt" @input="applyNodeEdit()" style="font-size:11px;min-height:60px" placeholder="{{input}}"></textarea>
                       </div>
                       <div class="form-group">
                         <label class="text-xs">Model (optional)</label>
-                        <input class="form-input" x-model="selectedNode.config.model" style="font-size:11px" placeholder="Default model">
+                        <input class="form-input" x-model="selectedNode.config.model" @input="applyNodeEdit()" style="font-size:11px" placeholder="Default model">
                       </div>
                     </div>
                   </template>
@@ -1459,7 +1499,7 @@
                     <div>
                       <div class="form-group">
                         <label class="text-xs">Expression</label>
-                        <input class="form-input" x-model="selectedNode.config.expression" style="font-size:11px" placeholder="output.contains('yes')">
+                        <input class="form-input" x-model="selectedNode.config.expression" @input="applyNodeEdit()" style="font-size:11px" placeholder="output.contains('yes')">
                       </div>
                       <div class="text-xs text-dim">Top port = true, bottom port = false</div>
                     </div>
@@ -1470,11 +1510,11 @@
                     <div>
                       <div class="form-group">
                         <label class="text-xs">Max Iterations</label>
-                        <input type="number" class="form-input" x-model.number="selectedNode.config.max_iterations" style="font-size:11px" min="1" max="100">
+                        <input type="number" class="form-input" x-model.number="selectedNode.config.max_iterations" @input="applyNodeEdit()" style="font-size:11px" min="1" max="100">
                       </div>
                       <div class="form-group">
                         <label class="text-xs">Until (stop condition)</label>
-                        <input class="form-input" x-model="selectedNode.config.until" style="font-size:11px" placeholder="output === 'done'">
+                        <input class="form-input" x-model="selectedNode.config.until" @input="applyNodeEdit()" style="font-size:11px" placeholder="output === 'done'">
                       </div>
                     </div>
                   </template>
@@ -1484,7 +1524,7 @@
                     <div>
                       <div class="form-group">
                         <label class="text-xs">Fan-out Count</label>
-                        <input type="number" class="form-input" x-model.number="selectedNode.config.fan_count" style="font-size:11px" min="2" max="10">
+                        <input type="number" class="form-input" x-model.number="selectedNode.config.fan_count" @input="applyNodeEdit()" style="font-size:11px" min="2" max="10">
                       </div>
                     </div>
                   </template>
@@ -1494,7 +1534,7 @@
                     <div>
                       <div class="form-group">
                         <label class="text-xs">Strategy</label>
-                        <select class="form-select" x-model="selectedNode.config.strategy" style="font-size:11px">
+                        <select class="form-select" x-model="selectedNode.config.strategy" @change="applyNodeEdit()" style="font-size:11px">
                           <option value="all">Wait for all</option>
                           <option value="first">First to finish</option>
                           <option value="majority">Majority vote</option>
@@ -2787,7 +2827,7 @@ args = ["-y", "@modelcontextprotocol/server-filesystem", "/path"]</pre>
 
                         <!-- Text type -->
                         <template x-if="setting.setting_type === 'text'">
-                          <input type="text" class="input" x-model="settingsValues[setting.key]" :placeholder="setting.label" style="width:100%">
+                          <input type="text" class="form-input" x-model="settingsValues[setting.key]" :placeholder="setting.label" style="width:100%">
                         </template>
                       </div>
                     </template>
@@ -3003,6 +3043,30 @@ args = ["-y", "@modelcontextprotocol/server-filesystem", "/path"]</pre>
                 </div>
               </template>
             </div>
+            <!-- Add Custom Provider -->
+            <div class="info-card mt-4" style="border:1px solid var(--border)">
+              <h4 style="margin-top:0">Add Custom Provider</h4>
+              <p class="text-xs text-dim mb-2">Connect any OpenAI-compatible API (vLLM, LiteLLM, LocalAI, etc.)</p>
+              <div style="display:grid;grid-template-columns:1fr 1fr;gap:0.5rem">
+                <div>
+                  <label class="text-xs text-dim">Provider Name</label>
+                  <input class="form-input" x-model="customProviderName" placeholder="e.g. my-local-llm">
+                </div>
+                <div>
+                  <label class="text-xs text-dim">Base URL (required)</label>
+                  <input class="form-input" x-model="customProviderUrl" placeholder="http://localhost:8080/v1">
+                </div>
+              </div>
+              <div class="mt-2">
+                <label class="text-xs text-dim">API Key (optional)</label>
+                <input class="form-input" type="password" x-model="customProviderKey" placeholder="sk-... (leave blank if not needed)">
+              </div>
+              <button class="btn btn-primary btn-sm mt-2" @click="addCustomProvider()" :disabled="!customProviderName.trim() || !customProviderUrl.trim() || addingCustomProvider">
+                <span x-show="!addingCustomProvider">Add Provider</span>
+                <span x-show="addingCustomProvider" class="spinner" style="width:10px;height:10px;border-width:2px"></span>
+              </button>
+              <span class="text-xs text-dim ml-2" x-text="customProviderStatus"></span>
+            </div>
             <div class="empty-state" x-show="!providers.length">
               <h4>No providers found</h4>
               <p class="hint">Provider information could not be loaded. Check that the API is running.</p>
@@ -3067,7 +3131,7 @@ args = ["-y", "@modelcontextprotocol/server-filesystem", "/path"]</pre>
             </div>
             <div class="table-wrap" x-show="filteredModels.length">
               <table>
-                <thead><tr><th>Model</th><th>Provider</th><th>Tier</th><th>Context</th><th>Input Cost</th><th>Output Cost</th><th>Status</th></tr></thead>
+                <thead><tr><th>Model</th><th>Provider</th><th>Tier</th><th>Context</th><th>Input Cost</th><th>Output Cost</th><th>Status</th><th></th></tr></thead>
                 <tbody>
                   <template x-for="m in filteredModels" :key="m.id">
                     <tr>
@@ -3078,6 +3142,7 @@ args = ["-y", "@modelcontextprotocol/server-filesystem", "/path"]</pre>
                       <td class="text-xs" x-text="formatCost(m.input_cost_per_m)"></td>
                       <td class="text-xs" x-text="formatCost(m.output_cost_per_m)"></td>
                       <td><span class="badge" :class="m.available ? 'badge-success' : 'badge-muted'" x-text="m.available ? 'Available' : 'Needs Key'"></span></td>
+                      <td><button x-show="m.tier === 'custom'" class="btn btn-ghost btn-sm" @click="deleteCustomModel(m.id)" title="Delete custom model" style="padding:2px 6px;font-size:11px;color:var(--text-muted)">&#x2715;</button></td>
                     </tr>
                   </template>
                 </tbody>
@@ -3307,11 +3372,11 @@ args = ["-y", "@modelcontextprotocol/server-filesystem", "/path"]</pre>
           <!-- Network tab -->
           <div x-show="tab === 'network'" x-data="{
             netStatus: null, a2aAgents: [], a2aDiscoverUrl: '', a2aDiscovering: false,
-            async loadNetStatus() { try { this.netStatus = await (await fetch('/api/network/status')).json(); } catch(e) {} },
-            async loadA2aAgents() { try { let r = await (await fetch('/api/a2a/agents')).json(); this.a2aAgents = r.agents || []; } catch(e) {} },
+            async loadNetStatus() { try { this.netStatus = await OpenFangAPI.get('/api/network/status'); } catch(e) {} },
+            async loadA2aAgents() { try { let r = await OpenFangAPI.get('/api/a2a/agents'); this.a2aAgents = r.agents || []; } catch(e) {} },
             async discoverA2a() {
               if (!this.a2aDiscoverUrl) return; this.a2aDiscovering = true;
-              try { await fetch('/api/a2a/discover', {method:'POST',headers:{'Content-Type':'application/json'},body:JSON.stringify({url:this.a2aDiscoverUrl})}); this.a2aDiscoverUrl=''; await this.loadA2aAgents(); } catch(e) {}
+              try { await OpenFangAPI.post('/api/a2a/discover', {url:this.a2aDiscoverUrl}); this.a2aDiscoverUrl=''; await this.loadA2aAgents(); } catch(e) {}
               this.a2aDiscovering = false;
             }
           }" x-init="loadNetStatus(); loadA2aAgents()">
@@ -3399,8 +3464,8 @@ args = ["-y", "@modelcontextprotocol/server-filesystem", "/path"]</pre>
               this.budgetLoading = true;
               try {
                 let [b, a] = await Promise.all([
-                  fetch('/api/budget').then(r => r.json()),
-                  fetch('/api/budget/agents').then(r => r.json())
+                  OpenFangAPI.get('/api/budget'),
+                  OpenFangAPI.get('/api/budget/agents')
                 ]);
                 this.budgetData = b;
                 this.agentRanking = a.agents || [];
@@ -3423,10 +3488,10 @@ args = ["-y", "@modelcontextprotocol/server-filesystem", "/path"]</pre>
                 if (+this.editMonthly !== this.budgetData.monthly_limit) body.max_monthly_usd = +this.editMonthly;
                 let alertVal = (+this.editAlert) / 100;
                 if (Math.abs(alertVal - this.budgetData.alert_threshold) > 0.001) body.alert_threshold = alertVal;
-                await fetch('/api/budget', { method: 'PUT', headers: {'Content-Type':'application/json'}, body: JSON.stringify(body) });
+                await OpenFangAPI.put('/api/budget', body);
                 this.editMode = false;
                 await this.loadBudget();
-              } catch(e) { alert('Failed to save: ' + e); }
+              } catch(e) { OpenFangToast.error('Failed to save: ' + (e.message || e)); }
               this.saving = false;
             },
             pctColor(pct) { return pct >= 0.8 ? '#ef4444' : pct >= 0.5 ? '#eab308' : '#22c55e'; },
@@ -4388,7 +4453,29 @@ args = ["-y", "@modelcontextprotocol/server-filesystem", "/path"]</pre>
                 </div>
               </div>
 
-              <template x-if="selectedProviderObj && !providerIsConfigured(selectedProviderObj)">
+              <template x-if="selectedProviderObj && !providerIsConfigured(selectedProviderObj) && selectedProvider === 'claude-code'">
+                <div class="card" style="border-left:3px solid var(--accent);margin-top:16px">
+                  <div class="card-header">Configure Claude Code</div>
+                  <div class="text-xs text-dim mb-2" style="line-height:1.8">
+                    Claude Code uses its own CLI authentication &mdash; no API key needed.
+                  </div>
+                  <div style="background:var(--bg);border-radius:4px;padding:10px 12px;margin-bottom:12px;font-size:12px;line-height:1.8">
+                    <div><span style="color:var(--accent)">1.</span> Install: <code style="color:var(--accent-light);background:var(--bg-secondary);padding:1px 4px;border-radius:2px">npm install -g @anthropic-ai/claude-code</code></div>
+                    <div><span style="color:var(--accent)">2.</span> Authenticate: <code style="color:var(--accent-light);background:var(--bg-secondary);padding:1px 4px;border-radius:2px">claude auth</code></div>
+                    <div><span style="color:var(--accent)">3.</span> Click <strong>Detect</strong> below to verify</div>
+                  </div>
+                  <button class="btn btn-primary btn-sm" @click="detectClaudeCode()" :disabled="testingProvider">
+                    <span x-show="!testingProvider">Detect Claude Code</span>
+                    <span x-show="testingProvider" class="spinner" style="width:10px;height:10px;border-width:2px"></span>
+                  </button>
+                  <div x-show="testResult" class="mt-2">
+                    <div x-show="testResult && testResult.status === 'ok'" class="badge badge-success" style="padding:6px 12px">Claude Code detected<span x-show="testResult && testResult.latency_ms" x-text="' (' + (testResult ? testResult.latency_ms : '') + 'ms)'"></span></div>
+                    <div x-show="testResult && testResult.status !== 'ok'" class="badge badge-error" style="padding:6px 12px">Claude Code CLI not detected. Make sure you&rsquo;ve run: <code>npm install -g @anthropic-ai/claude-code &amp;&amp; claude auth</code></div>
+                  </div>
+                </div>
+              </template>
+
+              <template x-if="selectedProviderObj && !providerIsConfigured(selectedProviderObj) && selectedProvider !== 'claude-code'">
                 <div class="card" style="border-left:3px solid var(--accent);margin-top:16px">
                   <div class="card-header" x-text="'Configure ' + selectedProviderObj.display_name"></div>
                   <div class="text-xs text-dim mb-2" x-show="selectedProviderObj.api_key_env">

crates/openfang-api/static/js/api.js

@@ -161,6 +161,17 @@ var OpenFangAPI = (function() {
     return fetch(BASE + path, opts).then(function(r) {
       if (_connectionState !== 'connected') setConnectionState('connected');
       if (!r.ok) {
+        // On 401, auto-show auth prompt so the user can re-enter their key
+        if (r.status === 401 && typeof Alpine !== 'undefined') {
+          try {
+            var store = Alpine.store('app');
+            if (store && !store.showAuthPrompt) {
+              _authToken = '';
+              localStorage.removeItem('openfang-api-key');
+              store.showAuthPrompt = true;
+            }
+          } catch(e2) { /* ignore Alpine errors */ }
+        }
         return r.text().then(function(text) {
           var msg = '';
           try {

crates/openfang-api/static/js/pages/agents.js

@@ -63,6 +63,9 @@ function agentsPage() {
     editingModel: false,
     newModelValue: '',
     modelSaving: false,
+    // -- Fallback chain --
+    editingFallback: false,
+    newFallbackValue: '',
 
     // -- Templates state --
     tplTemplates: [],
@@ -316,12 +319,15 @@ function agentsPage() {
       OpenFangAPI.wsDisconnect();
     },
 
-    showDetail(agent) {
+    async showDetail(agent) {
       this.detailAgent = agent;
+      this.detailAgent._fallbacks = [];
       this.detailTab = 'info';
       this.agentFiles = [];
       this.editingFile = null;
       this.fileContent = '';
+      this.editingFallback = false;
+      this.newFallbackValue = '';
       this.configForm = {
         name: agent.name || '',
         system_prompt: agent.system_prompt || '',
@@ -331,6 +337,11 @@ function agentsPage() {
         vibe: (agent.identity && agent.identity.vibe) || ''
       };
       this.showDetailModal = true;
+      // Fetch full agent detail to get fallback_models
+      try {
+        var full = await OpenFangAPI.get('/api/agents/' + agent.id);
+        this.detailAgent._fallbacks = full.fallback_models || [];
+      } catch(e) { /* ignore */ }
     },
 
     killAgent(agent) {
@@ -400,7 +411,7 @@ function agentsPage() {
       var f = this.spawnForm;
       var si = this.spawnIdentity;
       var lines = [
-        'name = "' + f.name + '"',
+        'name = "' + f.name.replace(/\\/g, '\\\\').replace(/"/g, '\\"') + '"',
         'module = "builtin:chat"'
       ];
       if (f.profile && f.profile !== 'custom') {
@@ -409,7 +420,7 @@ function agentsPage() {
       lines.push('', '[model]');
       lines.push('provider = "' + f.provider + '"');
       lines.push('model = "' + f.model + '"');
-      lines.push('system_prompt = "' + f.systemPrompt.replace(/"/g, '\\"') + '"');
+      lines.push('system_prompt = """\n' + f.systemPrompt.replace(/\\/g, '\\\\').replace(/"""/g, '""\\"') + '\n"""');
       if (f.profile === 'custom') {
         lines.push('', '[capabilities]');
         if (f.caps.memory_read) lines.push('memory_read = ["*"]');
@@ -586,8 +597,9 @@ function agentsPage() {
       if (!this.detailAgent || !this.newModelValue.trim()) return;
       this.modelSaving = true;
       try {
-        await OpenFangAPI.put('/api/agents/' + this.detailAgent.id + '/model', { model: this.newModelValue.trim() });
-        OpenFangToast.success('Model changed (memory reset)');
+        var resp = await OpenFangAPI.put('/api/agents/' + this.detailAgent.id + '/model', { model: this.newModelValue.trim() });
+        var providerInfo = (resp && resp.provider) ? ' (provider: ' + resp.provider + ')' : '';
+        OpenFangToast.success('Model changed' + providerInfo + ' (memory reset)');
         this.editingModel = false;
         await Alpine.store('app').refreshAgents();
         // Refresh detailAgent
@@ -601,6 +613,41 @@ function agentsPage() {
       this.modelSaving = false;
     },
 
+    // ── Fallback model chain ──
+    async addFallback() {
+      if (!this.detailAgent || !this.newFallbackValue.trim()) return;
+      var parts = this.newFallbackValue.trim().split('/');
+      var provider = parts.length > 1 ? parts[0] : this.detailAgent.model_provider;
+      var model = parts.length > 1 ? parts.slice(1).join('/') : parts[0];
+      if (!this.detailAgent._fallbacks) this.detailAgent._fallbacks = [];
+      this.detailAgent._fallbacks.push({ provider: provider, model: model });
+      try {
+        await OpenFangAPI.patch('/api/agents/' + this.detailAgent.id + '/config', {
+          fallback_models: this.detailAgent._fallbacks
+        });
+        OpenFangToast.success('Fallback added: ' + provider + '/' + model);
+      } catch(e) {
+        OpenFangToast.error('Failed to save fallbacks: ' + e.message);
+        this.detailAgent._fallbacks.pop();
+      }
+      this.editingFallback = false;
+      this.newFallbackValue = '';
+    },
+
+    async removeFallback(idx) {
+      if (!this.detailAgent || !this.detailAgent._fallbacks) return;
+      var removed = this.detailAgent._fallbacks.splice(idx, 1);
+      try {
+        await OpenFangAPI.patch('/api/agents/' + this.detailAgent.id + '/config', {
+          fallback_models: this.detailAgent._fallbacks
+        });
+        OpenFangToast.success('Fallback removed');
+      } catch(e) {
+        OpenFangToast.error('Failed to save fallbacks: ' + e.message);
+        this.detailAgent._fallbacks.splice(idx, 0, removed[0]);
+      }
+    },
+
     // ── Tool filters ──
     async loadToolFilters() {
       if (!this.detailAgent) return;

crates/openfang-api/static/js/pages/chat.js

@@ -37,6 +37,7 @@ function chatPage() {
     // Model switcher dropdown
     showModelSwitcher: false,
     modelSwitcherFilter: '',
+    modelSwitcherProviderFilter: '',
     modelSwitcherIdx: 0,
     modelSwitching: false,
     _modelCache: null,
@@ -99,14 +100,25 @@ function chatPage() {
       return short.length > 24 ? short.substring(0, 22) + '\u2026' : short;
     },
 
+    get switcherProviders() {
+      var seen = {};
+      (this._modelCache || []).forEach(function(m) { seen[m.provider] = true; });
+      return Object.keys(seen).sort();
+    },
+
     get filteredSwitcherModels() {
       var models = this._modelCache || [];
-      if (!this.modelSwitcherFilter) return models;
-      var f = this.modelSwitcherFilter.toLowerCase();
+      var provFilter = this.modelSwitcherProviderFilter;
+      var textFilter = this.modelSwitcherFilter ? this.modelSwitcherFilter.toLowerCase() : '';
+      if (!provFilter && !textFilter) return models;
       return models.filter(function(m) {
-        return m.id.toLowerCase().indexOf(f) !== -1 ||
-               (m.display_name || '').toLowerCase().indexOf(f) !== -1 ||
-               m.provider.toLowerCase().indexOf(f) !== -1;
+        if (provFilter && m.provider !== provFilter) return false;
+        if (textFilter) {
+          return m.id.toLowerCase().indexOf(textFilter) !== -1 ||
+                 (m.display_name || '').toLowerCase().indexOf(textFilter) !== -1 ||
+                 m.provider.toLowerCase().indexOf(textFilter) !== -1;
+        }
+        return true;
       });
     },
 
@@ -220,6 +232,7 @@ function chatPage() {
       var now = Date.now();
       if (this._modelCache && (now - this._modelCacheTime) < 300000) {
         this.modelSwitcherFilter = '';
+        this.modelSwitcherProviderFilter = '';
         this.modelSwitcherIdx = 0;
         this.showModelSwitcher = true;
         this.$nextTick(function() {
@@ -234,6 +247,7 @@ function chatPage() {
         self._modelCacheTime = Date.now();
         self.modelPickerList = models;
         self.modelSwitcherFilter = '';
+        self.modelSwitcherProviderFilter = '';
         self.modelSwitcherIdx = 0;
         self.showModelSwitcher = true;
         self.$nextTick(function() {
@@ -250,9 +264,10 @@ function chatPage() {
       if (model.id === this.currentAgent.model_name) { this.showModelSwitcher = false; return; }
       var self = this;
       this.modelSwitching = true;
-      OpenFangAPI.put('/api/agents/' + this.currentAgent.id + '/model', { model: model.id }).then(function() {
-        self.currentAgent.model_name = model.id;
-        self.currentAgent.model_provider = model.provider;
+      OpenFangAPI.put('/api/agents/' + this.currentAgent.id + '/model', { model: model.id }).then(function(resp) {
+        // Use server-resolved model/provider to stay in sync (fixes #387/#466)
+        self.currentAgent.model_name = (resp && resp.model) || model.id;
+        self.currentAgent.model_provider = (resp && resp.provider) || model.provider;
         OpenFangToast.success('Switched to ' + (model.display_name || model.id));
         self.showModelSwitcher = false;
         self.modelSwitching = false;
@@ -406,9 +421,13 @@ function chatPage() {
         case '/model':
           if (self.currentAgent) {
             if (cmdArgs) {
-              OpenFangAPI.put('/api/agents/' + self.currentAgent.id + '/model', { model: cmdArgs }).then(function() {
-                self.currentAgent.model_name = cmdArgs;
-                self.messages.push({ id: ++msgId, role: 'system', text: 'Model switched to: `' + cmdArgs + '`', meta: '', tools: [] });
+              OpenFangAPI.put('/api/agents/' + self.currentAgent.id + '/model', { model: cmdArgs }).then(function(resp) {
+                // Use server-resolved model/provider (fixes #387/#466)
+                var resolvedModel = (resp && resp.model) || cmdArgs;
+                var resolvedProvider = (resp && resp.provider) || '';
+                self.currentAgent.model_name = resolvedModel;
+                if (resolvedProvider) { self.currentAgent.model_provider = resolvedProvider; }
+                self.messages.push({ id: ++msgId, role: 'system', text: 'Model switched to: `' + resolvedModel + '`' + (resolvedProvider ? ' (provider: `' + resolvedProvider + '`)' : ''), meta: '', tools: [] });
                 self.scrollToBottom();
               }).catch(function(e) { OpenFangToast.error('Model switch failed: ' + e.message); });
             } else {
@@ -517,7 +536,10 @@ function chatPage() {
                 is_error: !!t.is_error
               };
             });
-            return { id: ++msgId, role: role, text: text, meta: '', tools: tools };
+            var images = (m.images || []).map(function(img) {
+              return { file_id: img.file_id, filename: img.filename || 'image' };
+            });
+            return { id: ++msgId, role: role, text: text, meta: '', tools: tools, images: images };
           });
           self.$nextTick(function() { self.scrollToBottom(); });
         }

crates/openfang-api/static/js/pages/settings.js

@@ -26,6 +26,11 @@ function settingsPage() {
     providerTesting: {},
     providerTestResults: {},
     copilotOAuth: { polling: false, userCode: '', verificationUri: '', pollId: '', interval: 5 },
+    customProviderName: '',
+    customProviderUrl: '',
+    customProviderKey: '',
+    customProviderStatus: '',
+    addingCustomProvider: false,
     loading: true,
     loadError: '',
 
@@ -258,6 +263,17 @@ function settingsPage() {
       }
     },
 
+    async deleteCustomModel(modelId) {
+      if (!confirm('Delete custom model "' + modelId + '"?')) return;
+      try {
+        await OpenFangAPI.del('/api/models/custom/' + encodeURIComponent(modelId));
+        OpenFangToast.success('Model deleted');
+        await this.loadModels();
+      } catch(e) {
+        OpenFangToast.error('Failed to delete: ' + (e.message || 'Unknown error'));
+      }
+    },
+
     async loadConfigSchema() {
       try {
         var results = await Promise.all([
@@ -279,11 +295,14 @@ function settingsPage() {
 
     async saveConfigField(section, field, value) {
       var key = section + '.' + field;
+      // Root-level fields (api_key, api_listen, log_level) use just the field name
+      var sectionMeta = this.configSchema && this.configSchema[section];
+      var path = (sectionMeta && sectionMeta.root_level) ? field : key;
       this.configSaving[key] = true;
       try {
-        await OpenFangAPI.post('/api/config/set', { path: key, value: value });
+        await OpenFangAPI.post('/api/config/set', { path: path, value: value });
         this.configDirty[key] = false;
-        OpenFangToast.success('Saved ' + key);
+        OpenFangToast.success('Saved ' + field);
       } catch(e) {
         OpenFangToast.error('Failed to save: ' + e.message);
       }
@@ -333,7 +352,10 @@ function settingsPage() {
 
     providerAuthText(p) {
       if (p.auth_status === 'configured') return 'Configured';
-      if (p.auth_status === 'not_set' || p.auth_status === 'missing') return 'Not Set';
+      if (p.auth_status === 'not_set' || p.auth_status === 'missing') {
+        if (p.id === 'claude-code') return 'Not Installed';
+        return 'Not Set';
+      }
       return 'No Key Needed';
     },
 
@@ -488,6 +510,34 @@ function settingsPage() {
       this.providerUrlSaving[provider.id] = false;
     },
 
+    async addCustomProvider() {
+      var name = this.customProviderName.trim().toLowerCase().replace(/[^a-z0-9-]/g, '-').replace(/-+/g, '-');
+      if (!name) { OpenFangToast.error('Please enter a provider name'); return; }
+      var url = this.customProviderUrl.trim();
+      if (!url) { OpenFangToast.error('Please enter a base URL'); return; }
+      if (url.indexOf('http://') !== 0 && url.indexOf('https://') !== 0) {
+        OpenFangToast.error('URL must start with http:// or https://'); return;
+      }
+      this.addingCustomProvider = true;
+      this.customProviderStatus = '';
+      try {
+        var result = await OpenFangAPI.put('/api/providers/' + encodeURIComponent(name) + '/url', { base_url: url });
+        if (this.customProviderKey.trim()) {
+          await OpenFangAPI.post('/api/providers/' + encodeURIComponent(name) + '/key', { key: this.customProviderKey.trim() });
+        }
+        this.customProviderName = '';
+        this.customProviderUrl = '';
+        this.customProviderKey = '';
+        this.customProviderStatus = '';
+        OpenFangToast.success('Provider "' + name + '" added' + (result.reachable ? ' (reachable)' : ' (not reachable yet)'));
+        await this.loadProviders();
+      } catch(e) {
+        this.customProviderStatus = 'Error: ' + (e.message || 'Failed');
+        OpenFangToast.error('Failed to add provider: ' + e.message);
+      }
+      this.addingCustomProvider = false;
+    },
+
     // -- Security methods --
     async loadSecurity() {
       this.secLoading = true;

crates/openfang-api/static/js/pages/wizard.js

@@ -283,11 +283,13 @@ function wizardPage() {
     },
 
     get canGoNext() {
-      if (this.step === 2) return this.keySaved || this.hasConfiguredProvider;
+      if (this.step === 2) return this.keySaved || this.hasConfiguredProvider || this.claudeCodeDetected;
       if (this.step === 3) return this.agentName.trim().length > 0;
       return true;
     },
 
+    claudeCodeDetected: false,
+
     get hasConfiguredProvider() {
       var self = this;
       return this.providers.some(function(p) {
@@ -320,7 +322,7 @@ function wizardPage() {
     },
 
     get popularProviders() {
-      var popular = ['anthropic', 'openai', 'gemini', 'groq', 'deepseek', 'openrouter'];
+      var popular = ['anthropic', 'openai', 'gemini', 'groq', 'deepseek', 'openrouter', 'claude-code'];
       return this.providers.filter(function(p) {
         return popular.indexOf(p.id) >= 0;
       }).sort(function(a, b) {
@@ -329,7 +331,7 @@ function wizardPage() {
     },
 
     get otherProviders() {
-      var popular = ['anthropic', 'openai', 'gemini', 'groq', 'deepseek', 'openrouter'];
+      var popular = ['anthropic', 'openai', 'gemini', 'groq', 'deepseek', 'openrouter', 'claude-code'];
       return this.providers.filter(function(p) {
         return popular.indexOf(p.id) < 0;
       });
@@ -355,7 +357,8 @@ function wizardPage() {
         fireworks: { url: 'https://fireworks.ai/account/api-keys', text: 'Get your key from Fireworks AI' },
         perplexity: { url: 'https://www.perplexity.ai/settings/api', text: 'Get your key from Perplexity Settings' },
         cohere: { url: 'https://dashboard.cohere.com/api-keys', text: 'Get your key from the Cohere Dashboard' },
-        xai: { url: 'https://console.x.ai/', text: 'Get your key from the xAI Console' }
+        xai: { url: 'https://console.x.ai/', text: 'Get your key from the xAI Console' },
+        'claude-code': { url: 'https://docs.anthropic.com/en/docs/claude-code', text: 'Install: npm install -g @anthropic-ai/claude-code && claude auth (no API key needed)' }
       };
       return help[id] || null;
     },
@@ -408,6 +411,28 @@ function wizardPage() {
       this.testingProvider = false;
     },
 
+    async detectClaudeCode() {
+      this.testingProvider = true;
+      this.testResult = null;
+      try {
+        var result = await OpenFangAPI.post('/api/providers/claude-code/test', {});
+        this.testResult = result;
+        if (result.status === 'ok') {
+          this.claudeCodeDetected = true;
+          this.keySaved = true;
+          this.setupSummary.provider = 'Claude Code';
+          OpenFangToast.success('Claude Code detected (' + (result.latency_ms || '?') + 'ms)');
+        } else {
+          this.testResult = { status: 'error', error: 'Claude Code CLI not detected' };
+          OpenFangToast.error('Claude Code CLI not detected. Make sure you\'ve run: npm install -g @anthropic-ai/claude-code && claude auth');
+        }
+      } catch(e) {
+        this.testResult = { status: 'error', error: e.message };
+        OpenFangToast.error('Claude Code CLI not detected. Make sure you\'ve run: npm install -g @anthropic-ai/claude-code && claude auth');
+      }
+      this.testingProvider = false;
+    },
+
     // ── Step 3: Agent creation ──
 
     selectTemplate(index) {
@@ -468,13 +493,14 @@ function wizardPage() {
         gemini: 'gemini-2.5-flash',
         groq: 'llama-3.3-70b-versatile',
         deepseek: 'deepseek-chat',
-        openrouter: 'openrouter/auto',
+        openrouter: 'openrouter/google/gemini-2.5-flash',
         mistral: 'mistral-large-latest',
         together: 'meta-llama/Llama-3-70b-chat-hf',
         fireworks: 'accounts/fireworks/models/llama-v3p1-70b-instruct',
         perplexity: 'llama-3.1-sonar-large-128k-online',
         cohere: 'command-r-plus',
-        xai: 'grok-2'
+        xai: 'grok-2',
+        'claude-code': 'claude-code/sonnet'
       };
       return defaults[providerId] || '';
     },

crates/openfang-api/static/js/pages/workflow-builder.js

@@ -38,6 +38,11 @@ function workflowBuilder() {
     ],
 
     _renderScheduled: false,
+    _lastClickNodeId: null,
+    _lastClickTime: 0,
+    _didDrag: false,
+    _didConnect: false,
+    _didPan: false,
 
     async init() {
       var self = this;
@@ -334,8 +339,23 @@ function workflowBuilder() {
 
     onNodeMouseDown: function(node, e) {
       e.stopPropagation();
+      // Detect double-click manually — the native dblclick event never fires
+      // because scheduleRender() destroys and recreates all SVG elements between
+      // the first and second click, so the browser loses the DOM target for dblclick.
+      var now = Date.now();
+      if (this._lastClickNodeId === node.id && (now - this._lastClickTime) < 350) {
+        // Double-click detected — open editor instead of starting drag
+        this._lastClickNodeId = null;
+        this._lastClickTime = 0;
+        this.editNode(node);
+        return;
+      }
+      this._lastClickNodeId = node.id;
+      this._lastClickTime = now;
+
       this.selectedNode = node;
       this.selectedConnection = null;
+      this._didDrag = false;
       this.dragging = node.id;
       var rect = this._getCanvasRect();
       this.dragOffset = {
@@ -350,6 +370,7 @@ function workflowBuilder() {
       this.selectedConnection = null;
       this.showNodeEditor = false;
       // Start canvas pan
+      this._didPan = false;
       this.canvasDragging = true;
       this.canvasDragStart = { x: e.clientX - this.canvasOffset.x * this.zoom, y: e.clientY - this.canvasOffset.y * this.zoom };
     },
@@ -357,6 +378,7 @@ function workflowBuilder() {
     onCanvasMouseMove: function(e) {
       var rect = this._getCanvasRect();
       if (this.dragging) {
+        this._didDrag = true;
         var node = this.getNode(this.dragging);
         if (node) {
           node.x = Math.max(0, (e.clientX - rect.left) / this.zoom - this.canvasOffset.x - this.dragOffset.x);
@@ -364,12 +386,14 @@ function workflowBuilder() {
         }
         this.scheduleRender();
       } else if (this.connecting) {
+        this._didConnect = true;
         this.connectPreview = {
           x: (e.clientX - rect.left) / this.zoom - this.canvasOffset.x,
           y: (e.clientY - rect.top) / this.zoom - this.canvasOffset.y
         };
         this.scheduleRender();
       } else if (this.canvasDragging) {
+        this._didPan = true;
         this.canvasOffset = {
           x: (e.clientX - this.canvasDragStart.x) / this.zoom,
           y: (e.clientY - this.canvasDragStart.y) / this.zoom
@@ -378,11 +402,19 @@ function workflowBuilder() {
     },
 
     onCanvasMouseUp: function() {
+      // Only re-render if something actually moved. Rendering on every mouseup
+      // destroys SVG elements between clicks, which prevents dblclick detection.
+      var needsRender = this._didDrag || this._didConnect || this._didPan;
       this.dragging = null;
       this.connecting = null;
       this.connectPreview = null;
       this.canvasDragging = false;
-      this.scheduleRender();
+      this._didDrag = false;
+      this._didConnect = false;
+      this._didPan = false;
+      if (needsRender) {
+        this.scheduleRender();
+      }
     },
 
     onCanvasWheel: function(e) {
@@ -427,6 +459,12 @@ function workflowBuilder() {
     editNode: function(node) {
       this.selectedNode = node;
       this.showNodeEditor = true;
+      this.scheduleRender();
+    },
+
+    // Called from editor panel inputs to reflect changes on the canvas SVG
+    applyNodeEdit: function() {
+      this.scheduleRender();
     },
 
     // ── TOML Generation ──────────────────────────────────

crates/openfang-api/tests/api_integration_test.rs

@@ -77,6 +77,7 @@ async fn start_test_server_with_provider(
         channels_config: tokio::sync::RwLock::new(Default::default()),
         shutdown_notify: Arc::new(tokio::sync::Notify::new()),
         clawhub_cache: dashmap::DashMap::new(),
+        provider_probe_cache: openfang_runtime::provider_health::ProbeCache::new(),
     });
 
     let app = Router::new()
@@ -705,6 +706,7 @@ async fn start_test_server_with_auth(api_key: &str) -> TestServer {
         channels_config: tokio::sync::RwLock::new(Default::default()),
         shutdown_notify: Arc::new(tokio::sync::Notify::new()),
         clawhub_cache: dashmap::DashMap::new(),
+        provider_probe_cache: openfang_runtime::provider_health::ProbeCache::new(),
     });
 
     let api_key_state = state.kernel.config.api_key.clone();

crates/openfang-api/tests/daemon_lifecycle_test.rs

@@ -114,6 +114,7 @@ async fn test_full_daemon_lifecycle() {
         channels_config: tokio::sync::RwLock::new(Default::default()),
         shutdown_notify: Arc::new(tokio::sync::Notify::new()),
         clawhub_cache: dashmap::DashMap::new(),
+        provider_probe_cache: openfang_runtime::provider_health::ProbeCache::new(),
     });
 
     let app = Router::new()
@@ -238,6 +239,7 @@ async fn test_server_immediate_responsiveness() {
         channels_config: tokio::sync::RwLock::new(Default::default()),
         shutdown_notify: Arc::new(tokio::sync::Notify::new()),
         clawhub_cache: dashmap::DashMap::new(),
+        provider_probe_cache: openfang_runtime::provider_health::ProbeCache::new(),
     });
 
     let app = Router::new()

crates/openfang-api/tests/load_test.rs

@@ -58,6 +58,7 @@ async fn start_test_server() -> TestServer {
         channels_config: tokio::sync::RwLock::new(Default::default()),
         shutdown_notify: Arc::new(tokio::sync::Notify::new()),
         clawhub_cache: dashmap::DashMap::new(),
+        provider_probe_cache: openfang_runtime::provider_health::ProbeCache::new(),
     });
 
     let app = Router::new()

crates/openfang-channels/Cargo.toml

@@ -26,6 +26,7 @@ hmac = { workspace = true }
 sha2 = { workspace = true }
 base64 = { workspace = true }
 hex = { workspace = true }
+html-escape = { workspace = true }
 
 lettre = { workspace = true }
 imap = { workspace = true }

crates/openfang-channels/src/bluesky.rs

@@ -215,7 +215,7 @@ impl BlueskyAdapter {
         let chunks = split_message(text, MAX_MESSAGE_LEN);
 
         for chunk in chunks {
-            let now = Utc::now().format("%Y-%m-%dT%H:%M:%S%.3fZ").to_string();
+            let now = Utc::now().to_rfc3339_opts(chrono::SecondsFormat::Millis, true);
 
             let mut record = serde_json::json!({
                 "$type": "app.bsky.feed.post",
@@ -435,7 +435,11 @@ impl ChannelAdapter for BlueskyAdapter {
                     service_url
                 );
                 if let Some(ref seen) = last_seen_at {
-                    url.push_str(&format!("&seenAt={}", seen));
+                    let encoded: String = url::form_urlencoded::Serializer::new(String::new())
+                        .append_pair("seenAt", seen)
+                        .finish();
+                    url.push('&');
+                    url.push_str(&encoded);
                 }
 
                 let resp = match client.get(&url).bearer_auth(&token).send().await {
@@ -492,7 +496,7 @@ impl ChannelAdapter for BlueskyAdapter {
                 if last_seen_at.is_some() {
                     let mark_url = format!("{}/xrpc/app.bsky.notification.updateSeen", service_url);
                     let mark_body = serde_json::json!({
-                        "seenAt": Utc::now().format("%Y-%m-%dT%H:%M:%S%.3fZ").to_string(),
+                        "seenAt": Utc::now().to_rfc3339_opts(chrono::SecondsFormat::Millis, true),
                     });
                     let _ = client
                         .post(&mark_url)

crates/openfang-channels/src/bridge.rs

@@ -5,9 +5,13 @@
 
 use crate::formatter;
 use crate::router::AgentRouter;
-use crate::types::{ChannelAdapter, ChannelContent, ChannelMessage, ChannelUser};
+use crate::types::{
+    default_phase_emoji, AgentPhase, ChannelAdapter, ChannelContent, ChannelMessage, ChannelUser,
+    LifecycleReaction,
+};
 use async_trait::async_trait;
 use dashmap::DashMap;
+use openfang_types::message::ContentBlock;
 use futures::StreamExt;
 use openfang_types::agent::AgentId;
 use openfang_types::config::{ChannelOverrides, DmPolicy, GroupPolicy, OutputFormat};
@@ -25,6 +29,26 @@ pub trait ChannelBridgeHandle: Send + Sync {
     /// Send a message to an agent and get the text response.
     async fn send_message(&self, agent_id: AgentId, message: &str) -> Result<String, String>;
 
+    /// Send a message with structured content blocks (text + images) to an agent.
+    ///
+    /// Default implementation extracts text from blocks and falls back to `send_message()`.
+    async fn send_message_with_blocks(
+        &self,
+        agent_id: AgentId,
+        blocks: Vec<ContentBlock>,
+    ) -> Result<String, String> {
+        // Default: extract text from blocks and send as plain text
+        let text: String = blocks
+            .iter()
+            .filter_map(|b| match b {
+                ContentBlock::Text { text } => Some(text.as_str()),
+                _ => None,
+            })
+            .collect::<Vec<_>>()
+            .join("\n");
+        self.send_message(agent_id, &text).await
+    }
+
     /// Find an agent by name, returning its ID.
     async fn find_agent_by_name(&self, name: &str) -> Result<Option<AgentId>, String>;
 
@@ -359,6 +383,25 @@ async fn send_response(
     }
 }
 
+/// Send a lifecycle reaction (best-effort, non-blocking for supported adapters).
+///
+/// Silently ignores errors — reactions are non-critical UX polish.
+/// For Telegram, the underlying HTTP call is already fire-and-forget (spawned internally),
+/// so this await returns almost immediately.
+async fn send_lifecycle_reaction(
+    adapter: &dyn ChannelAdapter,
+    user: &ChannelUser,
+    message_id: &str,
+    phase: AgentPhase,
+) {
+    let reaction = LifecycleReaction {
+        emoji: default_phase_emoji(&phase).to_string(),
+        phase,
+        remove_previous: true,
+    };
+    let _ = adapter.send_reaction(user, message_id, &reaction).await;
+}
+
 /// Dispatch a single incoming message — handles bot commands or routes to an agent.
 ///
 /// Applies per-channel policies (DM/group filtering, rate limiting, formatting, threading).
@@ -407,8 +450,15 @@ async fn dispatch_message(
                     }
                 }
                 GroupPolicy::MentionOnly => {
-                    // Pass through — adapters should only forward mentioned messages.
-                    // This is a hint for adapters, not enforced here.
+                    // Only allow messages where the bot was @mentioned or commands.
+                    let was_mentioned = message.metadata.get("was_mentioned")
+                        .and_then(|v| v.as_bool())
+                        .unwrap_or(false);
+                    let is_command = matches!(&message.content, ChannelContent::Command { .. });
+                    if !was_mentioned && !is_command {
+                        debug!("Ignoring group message on {ct_str} (group_policy=mention_only, not mentioned)");
+                        return;
+                    }
                 }
                 GroupPolicy::All => {}
             }
@@ -439,24 +489,53 @@ async fn dispatch_message(
         }
     }
 
-    let text = match &message.content {
-        ChannelContent::Text(t) => t.clone(),
-        ChannelContent::Command { name, args } => {
-            let result = handle_command(name, args, handle, router, &message.sender).await;
-            send_response(adapter, &message.sender, result, thread_id, output_format).await;
-            return;
-        }
-        _ => {
-            send_response(
+    // Handle commands first (early return)
+    if let ChannelContent::Command { ref name, ref args } = message.content {
+        let result = handle_command(name, args, handle, router, &message.sender).await;
+        send_response(adapter, &message.sender, result, thread_id, output_format).await;
+        return;
+    }
+
+    // For images: download, base64 encode, and send as multimodal content blocks
+    if let ChannelContent::Image { ref url, ref caption } = message.content {
+        let blocks = download_image_to_blocks(url, caption.as_deref()).await;
+        if blocks.iter().any(|b| matches!(b, ContentBlock::Image { .. })) {
+            // We have actual image data — send as structured blocks for vision
+            dispatch_with_blocks(
+                blocks,
+                message,
+                handle,
+                router,
                 adapter,
-                &message.sender,
-                "I can only handle text messages for now.".to_string(),
+                ct_str,
                 thread_id,
                 output_format,
             )
             .await;
             return;
         }
+        // Image download failed — fall through to text description below
+    }
+
+    let text = match &message.content {
+        ChannelContent::Text(t) => t.clone(),
+        ChannelContent::Command { .. } => unreachable!(), // handled above
+        ChannelContent::Image { ref url, ref caption } => {
+            // Fallback when image download failed
+            match caption {
+                Some(c) => format!("[User sent a photo: {url}]\nCaption: {c}"),
+                None => format!("[User sent a photo: {url}]"),
+            }
+        }
+        ChannelContent::File { ref url, ref filename } => {
+            format!("[User sent a file ({filename}): {url}]")
+        }
+        ChannelContent::Voice { ref url, duration_seconds } => {
+            format!("[User sent a voice message ({duration_seconds}s): {url}]")
+        }
+        ChannelContent::Location { lat, lon } => {
+            format!("[User shared location: {lat}, {lon}]")
+        }
     };
 
     // Check if it's a slash command embedded in text (e.g. "/agents")
@@ -641,15 +720,210 @@ async fn dispatch_message(
     // Send typing indicator (best-effort)
     let _ = adapter.send_typing(&message.sender).await;
 
+    // Lifecycle reaction: ⏳ Queued → 🤔 Thinking → ✅ Done / ❌ Error
+    let msg_id = &message.platform_message_id;
+    send_lifecycle_reaction(adapter, &message.sender, msg_id, AgentPhase::Queued).await;
+    send_lifecycle_reaction(adapter, &message.sender, msg_id, AgentPhase::Thinking).await;
+
     // Send to agent and relay response
     match handle.send_message(agent_id, &text).await {
         Ok(response) => {
+            send_lifecycle_reaction(adapter, &message.sender, msg_id, AgentPhase::Done).await;
             send_response(adapter, &message.sender, response, thread_id, output_format).await;
             handle
                 .record_delivery(agent_id, ct_str, &message.sender.platform_id, true, None)
                 .await;
         }
         Err(e) => {
+            send_lifecycle_reaction(adapter, &message.sender, msg_id, AgentPhase::Error).await;
+            warn!("Agent error for {agent_id}: {e}");
+            let err_msg = format!("Agent error: {e}");
+            send_response(
+                adapter,
+                &message.sender,
+                err_msg.clone(),
+                thread_id,
+                output_format,
+            )
+            .await;
+            handle
+                .record_delivery(
+                    agent_id,
+                    ct_str,
+                    &message.sender.platform_id,
+                    false,
+                    Some(&err_msg),
+                )
+                .await;
+        }
+    }
+}
+
+/// Download an image from a URL and build content blocks for multimodal LLM input.
+///
+/// Returns a `Vec<ContentBlock>` containing an image block (base64-encoded) and
+/// optionally a text block for the caption. If the download fails, returns a
+/// text-only block describing the failure.
+async fn download_image_to_blocks(url: &str, caption: Option<&str>) -> Vec<ContentBlock> {
+    use base64::Engine;
+
+    // 5 MB limit to prevent memory abuse from oversized images
+    const MAX_IMAGE_BYTES: usize = 5 * 1024 * 1024;
+
+    let client = reqwest::Client::new();
+    let resp = match client.get(url).send().await {
+        Ok(r) => r,
+        Err(e) => {
+            warn!("Failed to download image from channel: {e}");
+            return vec![ContentBlock::Text {
+                text: format!("[Image download failed: {e}]"),
+            }];
+        }
+    };
+
+    // Detect media type from Content-Type header, fall back to URL extension
+    let content_type = resp
+        .headers()
+        .get("content-type")
+        .and_then(|v| v.to_str().ok())
+        .map(|ct| ct.split(';').next().unwrap_or(ct).trim().to_string());
+
+    let media_type = content_type.unwrap_or_else(|| {
+        if url.contains(".png") {
+            "image/png".to_string()
+        } else if url.contains(".gif") {
+            "image/gif".to_string()
+        } else if url.contains(".webp") {
+            "image/webp".to_string()
+        } else {
+            "image/jpeg".to_string()
+        }
+    });
+
+    let bytes = match resp.bytes().await {
+        Ok(b) => b,
+        Err(e) => {
+            warn!("Failed to read image bytes: {e}");
+            return vec![ContentBlock::Text {
+                text: format!("[Image read failed: {e}]"),
+            }];
+        }
+    };
+
+    if bytes.len() > MAX_IMAGE_BYTES {
+        warn!(
+            "Image too large ({} bytes), skipping vision — sending as text",
+            bytes.len()
+        );
+        let desc = match caption {
+            Some(c) => format!("[Image too large for vision ({} KB)]\nCaption: {c}", bytes.len() / 1024),
+            None => format!("[Image too large for vision ({} KB)]", bytes.len() / 1024),
+        };
+        return vec![ContentBlock::Text { text: desc }];
+    }
+
+    let data = base64::engine::general_purpose::STANDARD.encode(&bytes);
+
+    let mut blocks = Vec::new();
+
+    // Caption as text block first (gives the LLM context about the image)
+    if let Some(cap) = caption {
+        if !cap.is_empty() {
+            blocks.push(ContentBlock::Text {
+                text: cap.to_string(),
+            });
+        }
+    }
+
+    blocks.push(ContentBlock::Image { media_type, data });
+
+    blocks
+}
+
+/// Dispatch a multimodal message (content blocks) to an agent, handling routing
+/// and RBAC the same way as the text path.
+#[allow(clippy::too_many_arguments)]
+async fn dispatch_with_blocks(
+    blocks: Vec<ContentBlock>,
+    message: &ChannelMessage,
+    handle: &Arc<dyn ChannelBridgeHandle>,
+    router: &Arc<AgentRouter>,
+    adapter: &dyn ChannelAdapter,
+    ct_str: &str,
+    thread_id: Option<&str>,
+    output_format: OutputFormat,
+) {
+    // Route to agent (same logic as text path)
+    let agent_id = router.resolve(
+        &message.channel,
+        &message.sender.platform_id,
+        message.sender.openfang_user.as_deref(),
+    );
+
+    let agent_id = match agent_id {
+        Some(id) => id,
+        None => {
+            let fallback = handle.find_agent_by_name("assistant").await.ok().flatten();
+            let fallback = match fallback {
+                Some(id) => Some(id),
+                None => handle
+                    .list_agents()
+                    .await
+                    .ok()
+                    .and_then(|agents| agents.first().map(|(id, _)| *id)),
+            };
+            match fallback {
+                Some(id) => {
+                    router.set_user_default(message.sender.platform_id.clone(), id);
+                    id
+                }
+                None => {
+                    send_response(
+                        adapter,
+                        &message.sender,
+                        "No agents available. Start the dashboard at http://127.0.0.1:4200 to create one.".to_string(),
+                        thread_id,
+                        output_format,
+                    ).await;
+                    return;
+                }
+            }
+        }
+    };
+
+    // RBAC check
+    if let Err(denied) = handle
+        .authorize_channel_user(ct_str, &message.sender.platform_id, "chat")
+        .await
+    {
+        send_response(
+            adapter,
+            &message.sender,
+            format!("Access denied: {denied}"),
+            thread_id,
+            output_format,
+        )
+        .await;
+        return;
+    }
+
+    let _ = adapter.send_typing(&message.sender).await;
+
+    // Lifecycle reaction: ⏳ Queued → 🤔 Thinking → ✅ Done / ❌ Error
+    let msg_id = &message.platform_message_id;
+    send_lifecycle_reaction(adapter, &message.sender, msg_id, AgentPhase::Queued).await;
+    send_lifecycle_reaction(adapter, &message.sender, msg_id, AgentPhase::Thinking).await;
+
+    match handle.send_message_with_blocks(agent_id, blocks).await {
+        Ok(response) => {
+            send_lifecycle_reaction(adapter, &message.sender, msg_id, AgentPhase::Done).await;
+            send_response(adapter, &message.sender, response, thread_id, output_format).await;
+            handle
+                .record_delivery(agent_id, ct_str, &message.sender.platform_id, true, None)
+                .await;
+        }
+        Err(e) => {
+            send_lifecycle_reaction(adapter, &message.sender, msg_id, AgentPhase::Error).await;
             warn!("Agent error for {agent_id}: {e}");
             let err_msg = format!("Agent error: {e}");
             send_response(
@@ -1112,4 +1386,52 @@ mod tests {
             "irc"
         );
     }
+
+    #[tokio::test]
+    async fn test_send_message_with_blocks_default_fallback() {
+        // The default implementation of send_message_with_blocks extracts text
+        // from blocks and calls send_message
+        let agent_id = AgentId::new();
+        let handle: Arc<dyn ChannelBridgeHandle> = Arc::new(MockHandle {
+            agents: Mutex::new(vec![(agent_id, "vision-agent".to_string())]),
+        });
+
+        let blocks = vec![
+            ContentBlock::Text {
+                text: "What is in this photo?".to_string(),
+            },
+            ContentBlock::Image {
+                media_type: "image/jpeg".to_string(),
+                data: "base64data".to_string(),
+            },
+        ];
+
+        // Default impl should extract text and call send_message
+        let result = handle
+            .send_message_with_blocks(agent_id, blocks)
+            .await
+            .unwrap();
+        assert_eq!(result, "Echo: What is in this photo?");
+    }
+
+    #[tokio::test]
+    async fn test_send_message_with_blocks_image_only() {
+        // When there's no text block, the default should still work
+        let agent_id = AgentId::new();
+        let handle: Arc<dyn ChannelBridgeHandle> = Arc::new(MockHandle {
+            agents: Mutex::new(vec![(agent_id, "vision-agent".to_string())]),
+        });
+
+        let blocks = vec![ContentBlock::Image {
+            media_type: "image/png".to_string(),
+            data: "base64data".to_string(),
+        }];
+
+        // Default impl sends empty text when no text blocks
+        let result = handle
+            .send_message_with_blocks(agent_id, blocks)
+            .await
+            .unwrap();
+        assert_eq!(result, "Echo: ");
+    }
 }

crates/openfang-channels/src/discord.rs

@@ -39,6 +39,8 @@ pub struct DiscordAdapter {
     token: Zeroizing<String>,
     client: reqwest::Client,
     allowed_guilds: Vec<String>,
+    allowed_users: Vec<String>,
+    ignore_bots: bool,
     intents: u64,
     shutdown_tx: Arc<watch::Sender<bool>>,
     shutdown_rx: watch::Receiver<bool>,
@@ -51,12 +53,20 @@ pub struct DiscordAdapter {
 }
 
 impl DiscordAdapter {
-    pub fn new(token: String, allowed_guilds: Vec<String>, intents: u64) -> Self {
+    pub fn new(
+        token: String,
+        allowed_guilds: Vec<String>,
+        allowed_users: Vec<String>,
+        ignore_bots: bool,
+        intents: u64,
+    ) -> Self {
         let (shutdown_tx, shutdown_rx) = watch::channel(false);
         Self {
             token: Zeroizing::new(token),
             client: reqwest::Client::new(),
             allowed_guilds,
+            allowed_users,
+            ignore_bots,
             intents,
             shutdown_tx: Arc::new(shutdown_tx),
             shutdown_rx,
@@ -147,6 +157,8 @@ impl ChannelAdapter for DiscordAdapter {
         let token = self.token.clone();
         let intents = self.intents;
         let allowed_guilds = self.allowed_guilds.clone();
+        let allowed_users = self.allowed_users.clone();
+        let ignore_bots = self.ignore_bots;
         let bot_user_id = self.bot_user_id.clone();
         let session_id_store = self.session_id.clone();
         let resume_url_store = self.resume_gateway_url.clone();
@@ -307,7 +319,7 @@ impl ChannelAdapter for DiscordAdapter {
 
                                 "MESSAGE_CREATE" | "MESSAGE_UPDATE" => {
                                     if let Some(msg) =
-                                        parse_discord_message(d, &bot_user_id, &allowed_guilds)
+                                        parse_discord_message(d, &bot_user_id, &allowed_guilds, &allowed_users, ignore_bots)
                                             .await
                                     {
                                         debug!(
@@ -423,6 +435,8 @@ async fn parse_discord_message(
     d: &serde_json::Value,
     bot_user_id: &Arc<RwLock<Option<String>>>,
     allowed_guilds: &[String],
+    allowed_users: &[String],
+    ignore_bots: bool,
 ) -> Option<ChannelMessage> {
     let author = d.get("author")?;
     let author_id = author["id"].as_str()?;
@@ -434,8 +448,14 @@ async fn parse_discord_message(
         }
     }
 
-    // Filter out other bots
-    if author["bot"].as_bool() == Some(true) {
+    // Filter out other bots (configurable via ignore_bots)
+    if ignore_bots && author["bot"].as_bool() == Some(true) {
+        return None;
+    }
+
+    // Filter by allowed users
+    if !allowed_users.is_empty() && !allowed_users.iter().any(|u| u == author_id) {
+        debug!("Discord: ignoring message from unlisted user {author_id}");
         return None;
     }
 
@@ -486,6 +506,29 @@ async fn parse_discord_message(
         ChannelContent::Text(content_text.to_string())
     };
 
+    // Determine if this is a group message (guild_id present = server channel)
+    let is_group = d["guild_id"].as_str().is_some();
+
+    // Check if bot was @mentioned (for MentionOnly policy enforcement)
+    let was_mentioned = if let Some(ref bid) = *bot_user_id.read().await {
+        // Check Discord mentions array
+        let mentioned_in_array = d["mentions"]
+            .as_array()
+            .map(|arr| arr.iter().any(|m| m["id"].as_str() == Some(bid.as_str())))
+            .unwrap_or(false);
+        // Also check content for <@bot_id> or <@!bot_id> patterns
+        let mentioned_in_content =
+            content_text.contains(&format!("<@{bid}>")) || content_text.contains(&format!("<@!{bid}>"));
+        mentioned_in_array || mentioned_in_content
+    } else {
+        false
+    };
+
+    let mut metadata = HashMap::new();
+    if was_mentioned {
+        metadata.insert("was_mentioned".to_string(), serde_json::json!(true));
+    }
+
     Some(ChannelMessage {
         channel: ChannelType::Discord,
         platform_message_id: message_id.to_string(),
@@ -497,9 +540,9 @@ async fn parse_discord_message(
         content,
         target_agent: None,
         timestamp,
-        is_group: true,
+        is_group,
         thread_id: None,
-        metadata: HashMap::new(),
+        metadata,
     })
 }
 
@@ -523,7 +566,7 @@ mod tests {
             "timestamp": "2024-01-01T00:00:00+00:00"
         });
 
-        let msg = parse_discord_message(&d, &bot_id, &[]).await.unwrap();
+        let msg = parse_discord_message(&d, &bot_id, &[], &[], true).await.unwrap();
         assert_eq!(msg.channel, ChannelType::Discord);
         assert_eq!(msg.sender.display_name, "alice");
         assert_eq!(msg.sender.platform_id, "ch1");
@@ -545,7 +588,7 @@ mod tests {
             "timestamp": "2024-01-01T00:00:00+00:00"
         });
 
-        let msg = parse_discord_message(&d, &bot_id, &[]).await;
+        let msg = parse_discord_message(&d, &bot_id, &[], &[], true).await;
         assert!(msg.is_none());
     }
 
@@ -565,7 +608,52 @@ mod tests {
             "timestamp": "2024-01-01T00:00:00+00:00"
         });
 
-        let msg = parse_discord_message(&d, &bot_id, &[]).await;
+        let msg = parse_discord_message(&d, &bot_id, &[], &[], true).await;
+        assert!(msg.is_none());
+    }
+
+    #[tokio::test]
+    async fn test_parse_discord_ignore_bots_false_allows_other_bots() {
+        let bot_id = Arc::new(RwLock::new(Some("bot123".to_string())));
+        let d = serde_json::json!({
+            "id": "msg1",
+            "channel_id": "ch1",
+            "content": "Bot message",
+            "author": {
+                "id": "other_bot",
+                "username": "somebot",
+                "discriminator": "0",
+                "bot": true
+            },
+            "timestamp": "2024-01-01T00:00:00+00:00"
+        });
+
+        // With ignore_bots=false, other bots' messages should be allowed
+        let msg = parse_discord_message(&d, &bot_id, &[], &[], false).await;
+        assert!(msg.is_some());
+        let msg = msg.unwrap();
+        assert_eq!(msg.sender.display_name, "somebot");
+        assert!(matches!(msg.content, ChannelContent::Text(ref t) if t == "Bot message"));
+    }
+
+    #[tokio::test]
+    async fn test_parse_discord_ignore_bots_false_still_filters_self() {
+        let bot_id = Arc::new(RwLock::new(Some("bot123".to_string())));
+        let d = serde_json::json!({
+            "id": "msg1",
+            "channel_id": "ch1",
+            "content": "My own message",
+            "author": {
+                "id": "bot123",
+                "username": "openfang",
+                "discriminator": "0",
+                "bot": true
+            },
+            "timestamp": "2024-01-01T00:00:00+00:00"
+        });
+
+        // Even with ignore_bots=false, the bot's own messages must still be filtered
+        let msg = parse_discord_message(&d, &bot_id, &[], &[], false).await;
         assert!(msg.is_none());
     }
 
@@ -586,11 +674,11 @@ mod tests {
         });
 
         // Not in allowed guilds
-        let msg = parse_discord_message(&d, &bot_id, &["111".into(), "222".into()]).await;
+        let msg = parse_discord_message(&d, &bot_id, &["111".into(), "222".into()], &[], true).await;
         assert!(msg.is_none());
 
         // In allowed guilds
-        let msg = parse_discord_message(&d, &bot_id, &["999".into()]).await;
+        let msg = parse_discord_message(&d, &bot_id, &["999".into()], &[], true).await;
         assert!(msg.is_some());
     }
 
@@ -609,7 +697,7 @@ mod tests {
             "timestamp": "2024-01-01T00:00:00+00:00"
         });
 
-        let msg = parse_discord_message(&d, &bot_id, &[]).await.unwrap();
+        let msg = parse_discord_message(&d, &bot_id, &[], &[], true).await.unwrap();
         match &msg.content {
             ChannelContent::Command { name, args } => {
                 assert_eq!(name, "agent");
@@ -634,7 +722,7 @@ mod tests {
             "timestamp": "2024-01-01T00:00:00+00:00"
         });
 
-        let msg = parse_discord_message(&d, &bot_id, &[]).await;
+        let msg = parse_discord_message(&d, &bot_id, &[], &[], true).await;
         assert!(msg.is_none());
     }
 
@@ -653,7 +741,7 @@ mod tests {
             "timestamp": "2024-01-01T00:00:00+00:00"
         });
 
-        let msg = parse_discord_message(&d, &bot_id, &[]).await.unwrap();
+        let msg = parse_discord_message(&d, &bot_id, &[], &[], true).await.unwrap();
         assert_eq!(msg.sender.display_name, "alice#1234");
     }
 
@@ -675,16 +763,105 @@ mod tests {
         });
 
         // MESSAGE_UPDATE uses the same parse function as MESSAGE_CREATE
-        let msg = parse_discord_message(&d, &bot_id, &[]).await.unwrap();
+        let msg = parse_discord_message(&d, &bot_id, &[], &[], true).await.unwrap();
         assert_eq!(msg.channel, ChannelType::Discord);
         assert!(
             matches!(msg.content, ChannelContent::Text(ref t) if t == "Edited message content")
         );
     }
 
+    #[tokio::test]
+    async fn test_parse_discord_allowed_users_filter() {
+        let bot_id = Arc::new(RwLock::new(Some("bot123".to_string())));
+        let d = serde_json::json!({
+            "id": "msg1",
+            "channel_id": "ch1",
+            "content": "Hello",
+            "author": {
+                "id": "user999",
+                "username": "bob",
+                "discriminator": "0"
+            },
+            "timestamp": "2024-01-01T00:00:00+00:00"
+        });
+
+        // Not in allowed users
+        let msg = parse_discord_message(&d, &bot_id, &[], &["user111".into(), "user222".into()], true).await;
+        assert!(msg.is_none());
+
+        // In allowed users
+        let msg = parse_discord_message(&d, &bot_id, &[], &["user999".into()], true).await;
+        assert!(msg.is_some());
+
+        // Empty allowed_users = allow all
+        let msg = parse_discord_message(&d, &bot_id, &[], &[], true).await;
+        assert!(msg.is_some());
+    }
+
+    #[tokio::test]
+    async fn test_parse_discord_mention_detection() {
+        let bot_id = Arc::new(RwLock::new(Some("bot123".to_string())));
+
+        // Message with bot mentioned in mentions array
+        let d = serde_json::json!({
+            "id": "msg1",
+            "channel_id": "ch1",
+            "guild_id": "guild1",
+            "content": "Hey <@bot123> help me",
+            "mentions": [{"id": "bot123", "username": "openfang"}],
+            "author": {
+                "id": "user1",
+                "username": "alice",
+                "discriminator": "0"
+            },
+            "timestamp": "2024-01-01T00:00:00+00:00"
+        });
+
+        let msg = parse_discord_message(&d, &bot_id, &[], &[], true).await.unwrap();
+        assert!(msg.is_group);
+        assert_eq!(msg.metadata.get("was_mentioned").and_then(|v| v.as_bool()), Some(true));
+
+        // Message without mention in group
+        let d2 = serde_json::json!({
+            "id": "msg2",
+            "channel_id": "ch1",
+            "guild_id": "guild1",
+            "content": "Just chatting",
+            "author": {
+                "id": "user1",
+                "username": "alice",
+                "discriminator": "0"
+            },
+            "timestamp": "2024-01-01T00:00:00+00:00"
+        });
+
+        let msg2 = parse_discord_message(&d2, &bot_id, &[], &[], true).await.unwrap();
+        assert!(msg2.is_group);
+        assert!(!msg2.metadata.contains_key("was_mentioned"));
+    }
+
+    #[tokio::test]
+    async fn test_parse_discord_dm_not_group() {
+        let bot_id = Arc::new(RwLock::new(None));
+        let d = serde_json::json!({
+            "id": "msg1",
+            "channel_id": "dm-ch1",
+            "content": "Hello",
+            "author": {
+                "id": "user1",
+                "username": "alice",
+                "discriminator": "0"
+            },
+            "timestamp": "2024-01-01T00:00:00+00:00"
+        });
+
+        let msg = parse_discord_message(&d, &bot_id, &[], &[], true).await.unwrap();
+        assert!(!msg.is_group);
+    }
+
     #[test]
     fn test_discord_adapter_creation() {
-        let adapter = DiscordAdapter::new("test-token".to_string(), vec!["123".to_string(), "456".to_string()], 37376);
+        let adapter = DiscordAdapter::new("test-token".to_string(), vec!["123".to_string(), "456".to_string()], vec![], true, 37376);
         assert_eq!(adapter.name(), "discord");
         assert_eq!(adapter.channel_type(), ChannelType::Discord);
     }

crates/openfang-channels/src/email.rs

@@ -20,6 +20,21 @@ use tokio::sync::{mpsc, watch};
 use tracing::{debug, error, info, warn};
 use zeroize::Zeroizing;
 
+/// SASL PLAIN authenticator for IMAP servers that reject LOGIN
+/// (e.g., Lark/Larksuite which only advertise AUTH=PLAIN).
+struct PlainAuthenticator {
+    username: String,
+    password: String,
+}
+
+impl imap::Authenticator for PlainAuthenticator {
+    type Response = String;
+    fn process(&self, _data: &[u8]) -> Self::Response {
+        // SASL PLAIN: \0<username>\0<password>
+        format!("\x00{}\x00{}", self.username, self.password)
+    }
+}
+
 /// Reply context for email threading (In-Reply-To / Subject continuity).
 #[derive(Debug, Clone)]
 struct ReplyCtx {
@@ -203,9 +218,22 @@ fn fetch_unseen_emails(
     let client = imap::connect((host, port), host, &tls)
         .map_err(|e| format!("IMAP connect failed: {e}"))?;
 
-    let mut session = client
-        .login(username, password)
-        .map_err(|(e, _)| format!("IMAP login failed: {e}"))?;
+    // Try LOGIN first; fall back to AUTHENTICATE PLAIN for servers like Lark
+    // that reject LOGIN and only support AUTH=PLAIN (SASL).
+    let mut session = match client.login(username, password) {
+        Ok(s) => s,
+        Err((login_err, client)) => {
+            let authenticator = PlainAuthenticator {
+                username: username.to_string(),
+                password: password.to_string(),
+            };
+            client
+                .authenticate("PLAIN", &authenticator)
+                .map_err(|(e, _)| {
+                    format!("IMAP login failed: {login_err}; AUTH=PLAIN also failed: {e}")
+                })?
+        }
+    };
 
     let mut results = Vec::new();
 

crates/openfang-channels/src/mastodon.rs

@@ -298,17 +298,8 @@ fn strip_html_tags(html: &str) -> String {
         }
     }
 
-    // Decode HTML entities
-    let decoded = result
-        .replace("&", "&")
-        .replace("&lt;", "<")
-        .replace("&gt;", ">")
-        .replace("&quot;", "\"")
-        .replace("&#39;", "'")
-        .replace("&apos;", "'")
-        .replace("&#x27;", "'")
-        .replace("&nbsp;", " ");
-
+    // Decode HTML entities (handles named, decimal, and hex entities)
+    let decoded = html_escape::decode_html_entities(&result);
     decoded.trim().to_string()
 }
 

crates/openfang-channels/src/nostr.rs

@@ -165,7 +165,7 @@ impl ChannelAdapter for NostrAdapter {
     ) -> Result<Pin<Box<dyn Stream<Item = ChannelMessage> + Send>>, Box<dyn std::error::Error>>
     {
         let pubkey = self.derive_pubkey();
-        info!("Nostr adapter starting (pubkey: {}...)", &pubkey[..16]);
+        info!("Nostr adapter starting (pubkey: {}...)", openfang_types::truncate_str(&pubkey, 16));
 
         if self.relays.is_empty() {
             return Err("Nostr: no relay URLs configured".into());

crates/openfang-channels/src/telegram.rs

@@ -5,6 +5,7 @@
 
 use crate::types::{
     split_message, ChannelAdapter, ChannelContent, ChannelMessage, ChannelType, ChannelUser,
+    LifecycleReaction,
 };
 use async_trait::async_trait;
 use futures::Stream;
@@ -13,7 +14,7 @@ use std::pin::Pin;
 use std::sync::Arc;
 use std::time::Duration;
 use tokio::sync::{mpsc, watch};
-use tracing::{debug, error, info, warn};
+use tracing::{debug, info, warn};
 use zeroize::Zeroizing;
 
 /// Maximum backoff duration on API failures.
@@ -23,13 +24,18 @@ const INITIAL_BACKOFF: Duration = Duration::from_secs(1);
 /// Telegram long-polling timeout (seconds) — sent as the `timeout` parameter to getUpdates.
 const LONG_POLL_TIMEOUT: u64 = 30;
 
+/// Default Telegram Bot API base URL.
+const DEFAULT_API_URL: &str = "https://api.telegram.org";
+
 /// Telegram Bot API adapter using long-polling.
 pub struct TelegramAdapter {
     /// SECURITY: Bot token is zeroized on drop to prevent memory disclosure.
     token: Zeroizing<String>,
     client: reqwest::Client,
-    allowed_users: Vec<i64>,
+    allowed_users: Vec<String>,
     poll_interval: Duration,
+    /// Base URL for Telegram Bot API (supports proxies/mirrors).
+    api_base_url: String,
     shutdown_tx: Arc<watch::Sender<bool>>,
     shutdown_rx: watch::Receiver<bool>,
 }
@@ -39,13 +45,24 @@ impl TelegramAdapter {
     ///
     /// `token` is the raw bot token (read from env by the caller).
     /// `allowed_users` is the list of Telegram user IDs allowed to interact (empty = allow all).
-    pub fn new(token: String, allowed_users: Vec<i64>, poll_interval: Duration) -> Self {
+    /// `api_url` overrides the Telegram Bot API base URL (for proxies/mirrors).
+    pub fn new(
+        token: String,
+        allowed_users: Vec<String>,
+        poll_interval: Duration,
+        api_url: Option<String>,
+    ) -> Self {
         let (shutdown_tx, shutdown_rx) = watch::channel(false);
+        let api_base_url = api_url
+            .unwrap_or_else(|| DEFAULT_API_URL.to_string())
+            .trim_end_matches('/')
+            .to_string();
         Self {
             token: Zeroizing::new(token),
             client: reqwest::Client::new(),
             allowed_users,
             poll_interval,
+            api_base_url,
             shutdown_tx: Arc::new(shutdown_tx),
             shutdown_rx,
         }
@@ -53,7 +70,7 @@ impl TelegramAdapter {
 
     /// Validate the bot token by calling `getMe`.
     pub async fn validate_token(&self) -> Result<String, Box<dyn std::error::Error>> {
-        let url = format!("https://api.telegram.org/bot{}/getMe", self.token.as_str());
+        let url = format!("{}/bot{}/getMe", self.api_base_url, self.token.as_str());
         let resp: serde_json::Value = self.client.get(&url).send().await?.json().await?;
 
         if resp["ok"].as_bool() != Some(true) {
@@ -75,7 +92,8 @@ impl TelegramAdapter {
         text: &str,
     ) -> Result<(), Box<dyn std::error::Error>> {
         let url = format!(
-            "https://api.telegram.org/bot{}/sendMessage",
+            "{}/bot{}/sendMessage",
+            self.api_base_url,
             self.token.as_str()
         );
 
@@ -111,7 +129,8 @@ impl TelegramAdapter {
         caption: Option<&str>,
     ) -> Result<(), Box<dyn std::error::Error>> {
         let url = format!(
-            "https://api.telegram.org/bot{}/sendPhoto",
+            "{}/bot{}/sendPhoto",
+            self.api_base_url,
             self.token.as_str()
         );
         let mut body = serde_json::json!({
@@ -138,7 +157,8 @@ impl TelegramAdapter {
         filename: &str,
     ) -> Result<(), Box<dyn std::error::Error>> {
         let url = format!(
-            "https://api.telegram.org/bot{}/sendDocument",
+            "{}/bot{}/sendDocument",
+            self.api_base_url,
             self.token.as_str()
         );
         let body = serde_json::json!({
@@ -161,7 +181,8 @@ impl TelegramAdapter {
         voice_url: &str,
     ) -> Result<(), Box<dyn std::error::Error>> {
         let url = format!(
-            "https://api.telegram.org/bot{}/sendVoice",
+            "{}/bot{}/sendVoice",
+            self.api_base_url,
             self.token.as_str()
         );
         let body = serde_json::json!({
@@ -184,7 +205,8 @@ impl TelegramAdapter {
         lon: f64,
     ) -> Result<(), Box<dyn std::error::Error>> {
         let url = format!(
-            "https://api.telegram.org/bot{}/sendLocation",
+            "{}/bot{}/sendLocation",
+            self.api_base_url,
             self.token.as_str()
         );
         let body = serde_json::json!({
@@ -203,7 +225,8 @@ impl TelegramAdapter {
     /// Call `sendChatAction` to show "typing..." indicator.
     async fn api_send_typing(&self, chat_id: i64) -> Result<(), Box<dyn std::error::Error>> {
         let url = format!(
-            "https://api.telegram.org/bot{}/sendChatAction",
+            "{}/bot{}/sendChatAction",
+            self.api_base_url,
             self.token.as_str()
         );
         let body = serde_json::json!({
@@ -213,6 +236,37 @@ impl TelegramAdapter {
         let _ = self.client.post(&url).json(&body).send().await?;
         Ok(())
     }
+
+    /// Call `setMessageReaction` on the Telegram API (fire-and-forget).
+    ///
+    /// Sets or replaces the bot's emoji reaction on a message. Each new call
+    /// automatically replaces the previous reaction, so there is no need to
+    /// explicitly remove old ones.
+    fn fire_reaction(&self, chat_id: i64, message_id: i64, emoji: &str) {
+        let url = format!(
+            "{}/bot{}/setMessageReaction",
+            self.api_base_url,
+            self.token.as_str()
+        );
+        let body = serde_json::json!({
+            "chat_id": chat_id,
+            "message_id": message_id,
+            "reaction": [{"type": "emoji", "emoji": emoji}],
+        });
+        let client = self.client.clone();
+        tokio::spawn(async move {
+            match client.post(&url).json(&body).send().await {
+                Ok(resp) if !resp.status().is_success() => {
+                    let body_text = resp.text().await.unwrap_or_default();
+                    debug!("Telegram setMessageReaction failed: {body_text}");
+                }
+                Err(e) => {
+                    debug!("Telegram setMessageReaction error: {e}");
+                }
+                _ => {}
+            }
+        });
+    }
 }
 
 #[async_trait]
@@ -238,7 +292,8 @@ impl ChannelAdapter for TelegramAdapter {
         // still be active on Telegram's side for ~30s, causing 409 errors.
         {
             let delete_url = format!(
-                "https://api.telegram.org/bot{}/deleteWebhook",
+                "{}/bot{}/deleteWebhook",
+                self.api_base_url,
                 self.token.as_str()
             );
             match self
@@ -259,6 +314,7 @@ impl ChannelAdapter for TelegramAdapter {
         let client = self.client.clone();
         let allowed_users = self.allowed_users.clone();
         let poll_interval = self.poll_interval;
+        let api_base_url = self.api_base_url.clone();
         let mut shutdown = self.shutdown_rx.clone();
 
         tokio::spawn(async move {
@@ -272,7 +328,7 @@ impl ChannelAdapter for TelegramAdapter {
                 }
 
                 // Build getUpdates request
-                let url = format!("https://api.telegram.org/bot{}/getUpdates", token.as_str());
+                let url = format!("{}/bot{}/getUpdates", api_base_url, token.as_str());
                 let mut params = serde_json::json!({
                     "timeout": LONG_POLL_TIMEOUT,
                     "allowed_updates": ["message", "edited_message"],
@@ -318,10 +374,14 @@ impl ChannelAdapter for TelegramAdapter {
                     continue;
                 }
 
-                // Handle conflict (another bot instance polling)
+                // Handle conflict (another bot instance or stale session polling).
+                // On daemon restart, the old long-poll may still be active on Telegram's
+                // side for up to 30s. Retry with backoff instead of stopping permanently.
                 if status.as_u16() == 409 {
-                    error!("Telegram 409 Conflict — another bot instance is running. Stopping.");
-                    break;
+                    warn!("Telegram 409 Conflict — stale polling session, retrying in {backoff:?}");
+                    tokio::time::sleep(backoff).await;
+                    backoff = (backoff * 2).min(MAX_BACKOFF);
+                    continue;
                 }
 
                 if !status.is_success() {
@@ -367,7 +427,7 @@ impl ChannelAdapter for TelegramAdapter {
                     }
 
                     // Parse the message
-                    let msg = match parse_telegram_update(update, &allowed_users) {
+                    let msg = match parse_telegram_update(update, &allowed_users, token.as_str(), &client, &api_base_url).await {
                         Some(m) => m,
                         None => continue, // filtered out or unparseable
                     };
@@ -437,6 +497,23 @@ impl ChannelAdapter for TelegramAdapter {
         self.api_send_typing(chat_id).await
     }
 
+    async fn send_reaction(
+        &self,
+        user: &ChannelUser,
+        message_id: &str,
+        reaction: &LifecycleReaction,
+    ) -> Result<(), Box<dyn std::error::Error>> {
+        let chat_id: i64 = user
+            .platform_id
+            .parse()
+            .map_err(|_| format!("Invalid Telegram chat_id: {}", user.platform_id))?;
+        let msg_id: i64 = message_id
+            .parse()
+            .map_err(|_| format!("Invalid Telegram message_id: {message_id}"))?;
+        self.fire_reaction(chat_id, msg_id, &reaction.emoji);
+        Ok(())
+    }
+
     async fn stop(&self) -> Result<(), Box<dyn std::error::Error>> {
         let _ = self.shutdown_tx.send(true);
         Ok(())
@@ -445,9 +522,36 @@ impl ChannelAdapter for TelegramAdapter {
 
 /// Parse a Telegram update JSON into a `ChannelMessage`, or `None` if filtered/unparseable.
 /// Handles both `message` and `edited_message` update types.
-fn parse_telegram_update(
+/// Resolve a Telegram file_id to a download URL via the Bot API.
+async fn telegram_get_file_url(
+    token: &str,
+    client: &reqwest::Client,
+    file_id: &str,
+    api_base_url: &str,
+) -> Option<String> {
+    let url = format!("{api_base_url}/bot{token}/getFile");
+    let resp = client
+        .post(&url)
+        .json(&serde_json::json!({"file_id": file_id}))
+        .send()
+        .await
+        .ok()?;
+    let body: serde_json::Value = resp.json().await.ok()?;
+    if body["ok"].as_bool() != Some(true) {
+        return None;
+    }
+    let file_path = body["result"]["file_path"].as_str()?;
+    Some(format!(
+        "{api_base_url}/file/bot{token}/{file_path}"
+    ))
+}
+
+async fn parse_telegram_update(
     update: &serde_json::Value,
-    allowed_users: &[i64],
+    allowed_users: &[String],
+    token: &str,
+    client: &reqwest::Client,
+    api_base_url: &str,
 ) -> Option<ChannelMessage> {
     let message = update
         .get("message")
@@ -455,8 +559,9 @@ fn parse_telegram_update(
     let from = message.get("from")?;
     let user_id = from["id"].as_i64()?;
 
-    // Security: check allowed_users
-    if !allowed_users.is_empty() && !allowed_users.contains(&user_id) {
+    // Security: check allowed_users (compare as strings for consistency)
+    let user_id_str = user_id.to_string();
+    if !allowed_users.is_empty() && !allowed_users.iter().any(|u| u == &user_id_str) {
         debug!("Telegram: ignoring message from unlisted user {user_id}");
         return None;
     }
@@ -472,41 +577,81 @@ fn parse_telegram_update(
 
     let chat_type = message["chat"]["type"].as_str().unwrap_or("private");
     let is_group = chat_type == "group" || chat_type == "supergroup";
-
-    let text = message["text"].as_str()?;
     let message_id = message["message_id"].as_i64().unwrap_or(0);
     let timestamp = message["date"]
         .as_i64()
         .and_then(|ts| chrono::DateTime::from_timestamp(ts, 0))
         .unwrap_or_else(chrono::Utc::now);
 
-    // Parse bot commands (Telegram sends entities for /commands)
-    let content = if let Some(entities) = message["entities"].as_array() {
-        let is_bot_command = entities
-            .iter()
-            .any(|e| e["type"].as_str() == Some("bot_command") && e["offset"].as_i64() == Some(0));
-        if is_bot_command {
-            let parts: Vec<&str> = text.splitn(2, ' ').collect();
-            let cmd_name = parts[0].trim_start_matches('/');
-            // Strip @botname from command (e.g. /agents@mybot -> agents)
-            let cmd_name = cmd_name.split('@').next().unwrap_or(cmd_name);
-            let args = if parts.len() > 1 {
-                parts[1].split_whitespace().map(String::from).collect()
+    // Determine content: text, photo, document, voice, or location
+    let content = if let Some(text) = message["text"].as_str() {
+        // Parse bot commands (Telegram sends entities for /commands)
+        if let Some(entities) = message["entities"].as_array() {
+            let is_bot_command = entities.iter().any(|e| {
+                e["type"].as_str() == Some("bot_command") && e["offset"].as_i64() == Some(0)
+            });
+            if is_bot_command {
+                let parts: Vec<&str> = text.splitn(2, ' ').collect();
+                let cmd_name = parts[0].trim_start_matches('/');
+                let cmd_name = cmd_name.split('@').next().unwrap_or(cmd_name);
+                let args = if parts.len() > 1 {
+                    parts[1].split_whitespace().map(String::from).collect()
+                } else {
+                    vec![]
+                };
+                ChannelContent::Command {
+                    name: cmd_name.to_string(),
+                    args,
+                }
             } else {
-                vec![]
-            };
-            ChannelContent::Command {
-                name: cmd_name.to_string(),
-                args,
+                ChannelContent::Text(text.to_string())
             }
         } else {
             ChannelContent::Text(text.to_string())
         }
+    } else if let Some(photos) = message["photo"].as_array() {
+        // Photos come as array of sizes; pick the largest (last)
+        let file_id = photos
+            .last()
+            .and_then(|p| p["file_id"].as_str())
+            .unwrap_or("");
+        let caption = message["caption"].as_str().map(String::from);
+        match telegram_get_file_url(token, client, file_id, api_base_url).await {
+            Some(url) => ChannelContent::Image { url, caption },
+            None => ChannelContent::Text(format!(
+                "[Photo received{}]",
+                caption.as_deref().map(|c| format!(": {c}")).unwrap_or_default()
+            )),
+        }
+    } else if message.get("document").is_some() {
+        let file_id = message["document"]["file_id"].as_str().unwrap_or("");
+        let filename = message["document"]["file_name"]
+            .as_str()
+            .unwrap_or("document")
+            .to_string();
+        match telegram_get_file_url(token, client, file_id, api_base_url).await {
+            Some(url) => ChannelContent::File { url, filename },
+            None => ChannelContent::Text(format!("[Document received: {filename}]")),
+        }
+    } else if message.get("voice").is_some() {
+        let file_id = message["voice"]["file_id"].as_str().unwrap_or("");
+        let duration = message["voice"]["duration"].as_u64().unwrap_or(0) as u32;
+        match telegram_get_file_url(token, client, file_id, api_base_url).await {
+            Some(url) => ChannelContent::Voice {
+                url,
+                duration_seconds: duration,
+            },
+            None => ChannelContent::Text(format!("[Voice message, {duration}s]")),
+        }
+    } else if message.get("location").is_some() {
+        let lat = message["location"]["latitude"].as_f64().unwrap_or(0.0);
+        let lon = message["location"]["longitude"].as_f64().unwrap_or(0.0);
+        ChannelContent::Location { lat, lon }
     } else {
-        ChannelContent::Text(text.to_string())
+        // Unsupported message type (stickers, polls, etc.)
+        return None;
     };
 
-    // Use chat_id as the platform_id (so responses go to the right chat)
     Some(ChannelMessage {
         channel: ChannelType::Telegram,
         platform_message_id: message_id.to_string(),
@@ -590,8 +735,12 @@ fn sanitize_telegram_html(text: &str) -> String {
 mod tests {
     use super::*;
 
-    #[test]
-    fn test_parse_telegram_update() {
+    fn test_client() -> reqwest::Client {
+        reqwest::Client::new()
+    }
+
+    #[tokio::test]
+    async fn test_parse_telegram_update() {
         let update = serde_json::json!({
             "update_id": 123456,
             "message": {
@@ -610,15 +759,16 @@ mod tests {
             }
         });
 
-        let msg = parse_telegram_update(&update, &[]).unwrap();
+        let client = test_client();
+        let msg = parse_telegram_update(&update, &[], "fake:token", &client, DEFAULT_API_URL).await.unwrap();
         assert_eq!(msg.channel, ChannelType::Telegram);
         assert_eq!(msg.sender.display_name, "Alice Smith");
         assert_eq!(msg.sender.platform_id, "111222333");
         assert!(matches!(msg.content, ChannelContent::Text(ref t) if t == "Hello, agent!"));
     }
 
-    #[test]
-    fn test_parse_telegram_command() {
+    #[tokio::test]
+    async fn test_parse_telegram_command() {
         let update = serde_json::json!({
             "update_id": 123457,
             "message": {
@@ -641,7 +791,8 @@ mod tests {
             }
         });
 
-        let msg = parse_telegram_update(&update, &[]).unwrap();
+        let client = test_client();
+        let msg = parse_telegram_update(&update, &[], "fake:token", &client, DEFAULT_API_URL).await.unwrap();
         match &msg.content {
             ChannelContent::Command { name, args } => {
                 assert_eq!(name, "agent");
@@ -651,8 +802,8 @@ mod tests {
         }
     }
 
-    #[test]
-    fn test_allowed_users_filter() {
+    #[tokio::test]
+    async fn test_allowed_users_filter() {
         let update = serde_json::json!({
             "update_id": 123458,
             "message": {
@@ -670,21 +821,25 @@ mod tests {
             }
         });
 
+        let client = test_client();
+
         // Empty allowed_users = allow all
-        let msg = parse_telegram_update(&update, &[]);
+        let msg = parse_telegram_update(&update, &[], "fake:token", &client, DEFAULT_API_URL).await;
         assert!(msg.is_some());
 
         // Non-matching allowed_users = filter out
-        let msg = parse_telegram_update(&update, &[111, 222]);
+        let blocked: Vec<String> = vec!["111".to_string(), "222".to_string()];
+        let msg = parse_telegram_update(&update, &blocked, "fake:token", &client, DEFAULT_API_URL).await;
         assert!(msg.is_none());
 
         // Matching allowed_users = allow
-        let msg = parse_telegram_update(&update, &[999]);
+        let allowed: Vec<String> = vec!["999".to_string()];
+        let msg = parse_telegram_update(&update, &allowed, "fake:token", &client, DEFAULT_API_URL).await;
         assert!(msg.is_some());
     }
 
-    #[test]
-    fn test_parse_telegram_edited_message() {
+    #[tokio::test]
+    async fn test_parse_telegram_edited_message() {
         let update = serde_json::json!({
             "update_id": 123459,
             "edited_message": {
@@ -704,7 +859,8 @@ mod tests {
             }
         });
 
-        let msg = parse_telegram_update(&update, &[]).unwrap();
+        let client = test_client();
+        let msg = parse_telegram_update(&update, &[], "fake:token", &client, DEFAULT_API_URL).await.unwrap();
         assert_eq!(msg.channel, ChannelType::Telegram);
         assert_eq!(msg.sender.display_name, "Alice Smith");
         assert!(matches!(msg.content, ChannelContent::Text(ref t) if t == "Edited message!"));
@@ -725,8 +881,8 @@ mod tests {
         assert_eq!(b4, Duration::from_secs(60)); // stays at cap
     }
 
-    #[test]
-    fn test_parse_command_with_botname() {
+    #[tokio::test]
+    async fn test_parse_command_with_botname() {
         let update = serde_json::json!({
             "update_id": 100,
             "message": {
@@ -739,7 +895,8 @@ mod tests {
             }
         });
 
-        let msg = parse_telegram_update(&update, &[]).unwrap();
+        let client = test_client();
+        let msg = parse_telegram_update(&update, &[], "fake:token", &client, DEFAULT_API_URL).await.unwrap();
         match &msg.content {
             ChannelContent::Command { name, args } => {
                 assert_eq!(name, "agents");
@@ -748,4 +905,121 @@ mod tests {
             other => panic!("Expected Command, got {other:?}"),
         }
     }
+
+    #[tokio::test]
+    async fn test_parse_telegram_location() {
+        let update = serde_json::json!({
+            "update_id": 200,
+            "message": {
+                "message_id": 50,
+                "from": { "id": 123, "first_name": "Alice" },
+                "chat": { "id": 123, "type": "private" },
+                "date": 1700000000,
+                "location": { "latitude": 51.5074, "longitude": -0.1278 }
+            }
+        });
+
+        let client = test_client();
+        let msg = parse_telegram_update(&update, &[], "fake:token", &client, DEFAULT_API_URL).await.unwrap();
+        assert!(matches!(msg.content, ChannelContent::Location { .. }));
+    }
+
+    #[tokio::test]
+    async fn test_parse_telegram_photo_fallback() {
+        // When getFile fails (fake token), photo messages should fall back to
+        // a text description rather than being silently dropped.
+        let update = serde_json::json!({
+            "update_id": 300,
+            "message": {
+                "message_id": 60,
+                "from": { "id": 123, "first_name": "Alice" },
+                "chat": { "id": 123, "type": "private" },
+                "date": 1700000000,
+                "photo": [
+                    { "file_id": "small_id", "file_unique_id": "a", "width": 90, "height": 90, "file_size": 1234 },
+                    { "file_id": "large_id", "file_unique_id": "b", "width": 800, "height": 600, "file_size": 45678 }
+                ],
+                "caption": "Check this out"
+            }
+        });
+
+        let client = test_client();
+        let msg = parse_telegram_update(&update, &[], "fake:token", &client, DEFAULT_API_URL).await.unwrap();
+        // With a fake token, getFile will fail, so we get a text fallback
+        match &msg.content {
+            ChannelContent::Text(t) => {
+                assert!(t.contains("Photo received"));
+                assert!(t.contains("Check this out"));
+            }
+            ChannelContent::Image { caption, .. } => {
+                // If somehow the HTTP call succeeded (unlikely with fake token),
+                // verify caption was extracted
+                assert_eq!(caption.as_deref(), Some("Check this out"));
+            }
+            other => panic!("Expected Text or Image fallback for photo, got {other:?}"),
+        }
+    }
+
+    #[tokio::test]
+    async fn test_parse_telegram_document_fallback() {
+        let update = serde_json::json!({
+            "update_id": 301,
+            "message": {
+                "message_id": 61,
+                "from": { "id": 123, "first_name": "Alice" },
+                "chat": { "id": 123, "type": "private" },
+                "date": 1700000000,
+                "document": {
+                    "file_id": "doc_id",
+                    "file_unique_id": "c",
+                    "file_name": "report.pdf",
+                    "file_size": 102400
+                }
+            }
+        });
+
+        let client = test_client();
+        let msg = parse_telegram_update(&update, &[], "fake:token", &client, DEFAULT_API_URL).await.unwrap();
+        match &msg.content {
+            ChannelContent::Text(t) => {
+                assert!(t.contains("Document received"));
+                assert!(t.contains("report.pdf"));
+            }
+            ChannelContent::File { filename, .. } => {
+                assert_eq!(filename, "report.pdf");
+            }
+            other => panic!("Expected Text or File for document, got {other:?}"),
+        }
+    }
+
+    #[tokio::test]
+    async fn test_parse_telegram_voice_fallback() {
+        let update = serde_json::json!({
+            "update_id": 302,
+            "message": {
+                "message_id": 62,
+                "from": { "id": 123, "first_name": "Alice" },
+                "chat": { "id": 123, "type": "private" },
+                "date": 1700000000,
+                "voice": {
+                    "file_id": "voice_id",
+                    "file_unique_id": "d",
+                    "duration": 15
+                }
+            }
+        });
+
+        let client = test_client();
+        let msg = parse_telegram_update(&update, &[], "fake:token", &client, DEFAULT_API_URL).await.unwrap();
+        match &msg.content {
+            ChannelContent::Text(t) => {
+                assert!(t.contains("Voice message"));
+                assert!(t.contains("15s"));
+            }
+            ChannelContent::Voice { duration_seconds, .. } => {
+                assert_eq!(*duration_seconds, 15);
+            }
+            other => panic!("Expected Text or Voice for voice message, got {other:?}"),
+        }
+    }
 }

crates/openfang-cli/src/main.rs

@@ -1073,11 +1073,23 @@ pub(crate) fn find_daemon() -> Option<String> {
 }
 
 /// Build an HTTP client for daemon calls.
+///
+/// When api_key is configured in config.toml, the client automatically
+/// includes a `Authorization: Bearer <key>` header on every request.
+/// When api_key is empty or missing, no auth header is sent.
 pub(crate) fn daemon_client() -> reqwest::blocking::Client {
-    reqwest::blocking::Client::builder()
-        .timeout(std::time::Duration::from_secs(120))
-        .build()
-        .expect("Failed to build HTTP client")
+    let mut builder = reqwest::blocking::Client::builder()
+        .timeout(std::time::Duration::from_secs(120));
+
+    if let Some(key) = read_api_key() {
+        let mut headers = reqwest::header::HeaderMap::new();
+        if let Ok(val) = reqwest::header::HeaderValue::from_str(&format!("Bearer {key}")) {
+            headers.insert(reqwest::header::AUTHORIZATION, val);
+        }
+        builder = builder.default_headers(headers);
+    }
+
+    builder.build().expect("Failed to build HTTP client")
 }
 
 /// Helper: send a request to the daemon and parse the JSON body.
@@ -1163,6 +1175,12 @@ fn cmd_init(quick: bool) {
 
     if quick {
         cmd_init_quick(&openfang_dir);
+    } else if !std::io::IsTerminal::is_terminal(&std::io::stdin())
+        || !std::io::IsTerminal::is_terminal(&std::io::stdout())
+    {
+        ui::hint("Non-interactive terminal detected — running in quick mode");
+        ui::hint("For the interactive wizard, run: openfang init (in a terminal)");
+        cmd_init_quick(&openfang_dir);
     } else {
         cmd_init_interactive(&openfang_dir);
     }
@@ -1284,8 +1302,17 @@ fn launch_desktop_app(_openfang_dir: &std::path::Path) {
             if let Some(base) = find_daemon() {
                 let url = format!("{base}/");
                 if !open_in_browser(&url) {
-                    ui::hint(&format!("Visit: {url}"));
+                    // Browser launch failed entirely (e.g., sandbox EPERM,
+                    // no display server, container environment).
+                    ui::hint("Could not open a browser automatically.");
                 }
+                // Always print the URL so the user can open it manually,
+                // even when open_in_browser reported success — the spawned
+                // opener may still fail asynchronously.
+                ui::hint(&format!("Dashboard: {url}"));
+            } else {
+                ui::hint("Daemon is not running. Start it with: openfang start");
+                ui::hint("Then open: http://127.0.0.1:4200");
             }
         }
     }
@@ -1306,8 +1333,14 @@ fn detect_best_provider() -> (&'static str, &'static str, &'static str) {
         ui::success("Detected Gemini (GOOGLE_API_KEY)");
         return ("gemini", "GOOGLE_API_KEY", "gemini-2.5-flash");
     }
+    // Check if Ollama is running locally (no API key needed)
+    if check_ollama_available() {
+        ui::success("Detected Ollama running locally (no API key needed)");
+        return ("ollama", "OLLAMA_API_KEY", "llama3.2");
+    }
     ui::hint("No LLM provider API keys found");
     ui::hint("Groq offers a free tier: https://console.groq.com");
+    ui::hint("Or install Ollama for local models: https://ollama.com");
     ("groq", "GROQ_API_KEY", "llama-3.3-70b-versatile")
 }
 
@@ -1327,12 +1360,21 @@ fn provider_list() -> Vec<(&'static str, &'static str, &'static str, &'static st
         (
             "openrouter",
             "OPENROUTER_API_KEY",
-            "openrouter/auto",
+            "openrouter/google/gemini-2.5-flash",
             "OpenRouter",
         ),
     ]
 }
 
+/// Quick probe to check if Ollama is running on localhost.
+fn check_ollama_available() -> bool {
+    std::net::TcpStream::connect_timeout(
+        &std::net::SocketAddr::from(([127, 0, 0, 1], 11434)),
+        std::time::Duration::from_millis(500),
+    )
+    .is_ok()
+}
+
 /// Write config.toml if it doesn't already exist.
 fn write_config_if_missing(
     openfang_dir: &std::path::Path,
@@ -1434,11 +1476,14 @@ fn cmd_start(config: Option<PathBuf>) {
 }
 
 /// Read the api_key from ~/.openfang/config.toml (if any).
+///
+/// Returns `None` when the key is missing, empty, or whitespace-only —
+/// meaning the daemon is running in public (unauthenticated) mode.
 fn read_api_key() -> Option<String> {
     let config_path = cli_openfang_home().join("config.toml");
     let text = std::fs::read_to_string(config_path).ok()?;
     let table: toml::Value = text.parse().ok()?;
-    let key = table.get("api_key")?.as_str()?;
+    let key = table.get("api_key")?.as_str()?.trim();
     if key.is_empty() {
         None
     } else {
@@ -1450,11 +1495,7 @@ fn cmd_stop() {
     match find_daemon() {
         Some(base) => {
             let client = daemon_client();
-            let mut req = client.post(format!("{base}/api/shutdown"));
-            if let Some(key) = read_api_key() {
-                req = req.bearer_auth(key);
-            }
-            match req.send() {
+            match client.post(format!("{base}/api/shutdown")).send() {
                 Ok(r) if r.status().is_success() => {
                     // Wait for daemon to actually stop (up to 5 seconds)
                     for _ in 0..10 {
@@ -2060,20 +2101,23 @@ fn cmd_doctor(json: bool, repair: bool) {
             }
             let answer = prompt_input("    Create default config? [Y/n] ");
             if answer.is_empty() || answer.starts_with('y') || answer.starts_with('Y') {
-                let default_config = r#"# OpenFang Agent OS configuration
+                let (provider, api_key_env, model) = detect_best_provider();
+                let default_config = format!(
+                    r#"# OpenFang Agent OS configuration
 # See https://github.com/RightNow-AI/openfang for documentation
 
 # For Docker, change to "0.0.0.0:4200" or set OPENFANG_LISTEN env var.
 api_listen = "127.0.0.1:4200"
 
 [default_model]
-provider = "groq"
-model = "llama-3.3-70b-versatile"
-api_key_env = "GROQ_API_KEY"
+provider = "{provider}"
+model = "{model}"
+api_key_env = "{api_key_env}"
 
 [memory]
 decay_rate = 0.05
-"#;
+"#
+                );
                 let _ = std::fs::create_dir_all(&openfang_dir);
                 if std::fs::write(&config_path, default_config).is_ok() {
                     restrict_file_permissions(&config_path);
@@ -2530,12 +2574,20 @@ decay_rate = 0.05
         }
 
         // Check for prompt injection issues in skill definitions
+        // Only flag Critical-severity warnings (Warning-level hits are expected
+        // in bundled skills that mention shell commands in educational context).
         let skills = skill_reg.list();
         let mut injection_warnings = 0;
         for skill in &skills {
             if let Some(ref prompt) = skill.manifest.prompt_context {
                 let warnings = openfang_skills::verify::SkillVerifier::scan_prompt_content(prompt);
-                if !warnings.is_empty() {
+                let has_critical = warnings.iter().any(|w| {
+                    matches!(
+                        w.severity,
+                        openfang_skills::verify::WarningSeverity::Critical
+                    )
+                });
+                if has_critical {
                     injection_warnings += 1;
                     if !json {
                         ui::check_warn(&format!(
@@ -2602,7 +2654,7 @@ decay_rate = 0.05
                         checks.push(serde_json::json!({"check": "daemon_uptime", "status": "ok", "secs": uptime}));
                     }
                     if let Some(db_status) = body.get("database").and_then(|v| v.as_str()) {
-                        if db_status == "ok" {
+                        if db_status == "connected" || db_status == "ok" {
                             if !json {
                                 ui::check_ok("Database connectivity: OK");
                             }
@@ -2676,12 +2728,18 @@ decay_rate = 0.05
         match client.get(format!("{base}/api/integrations/health")).send() {
             Ok(resp) if resp.status().is_success() => {
                 if let Ok(body) = resp.json::<serde_json::Value>() {
-                    if let Some(obj) = body.as_object() {
-                        let healthy = obj
-                            .values()
-                            .filter(|v| v.get("healthy").and_then(|h| h.as_bool()).unwrap_or(false))
+                    let entries = body.get("health").and_then(|h| h.as_array());
+                    if let Some(arr) = entries {
+                        let healthy = arr
+                            .iter()
+                            .filter(|v| {
+                                v.get("status")
+                                    .and_then(|s| s.as_str())
+                                    .map(|s| s.eq_ignore_ascii_case("ready"))
+                                    .unwrap_or(false)
+                            })
                             .count();
-                        let total = obj.len();
+                        let total = arr.len();
                         if healthy == total {
                             if !json {
                                 ui::check_ok(&format!(
@@ -2924,10 +2982,31 @@ pub(crate) fn open_in_browser(url: &str) -> bool {
     }
     #[cfg(target_os = "linux")]
     {
-        std::process::Command::new("xdg-open")
-            .arg(url)
-            .spawn()
-            .is_ok()
+        // Try multiple openers in order. xdg-open is the standard, but it
+        // (or the browser it launches) can fail with EPERM in sandboxed
+        // environments (containers, Snap, Flatpak, user-namespace
+        // restrictions). Fall through to alternatives if any opener fails.
+        let openers = [
+            "xdg-open",
+            "sensible-browser",
+            "x-www-browser",
+            "firefox",
+            "google-chrome",
+            "chromium",
+            "chromium-browser",
+        ];
+        for opener in &openers {
+            let result = std::process::Command::new(opener)
+                .arg(url)
+                .stdin(std::process::Stdio::null())
+                .stdout(std::process::Stdio::null())
+                .stderr(std::process::Stdio::null())
+                .spawn();
+            if result.is_ok() {
+                return true;
+            }
+        }
+        false
     }
     #[cfg(not(any(target_os = "windows", target_os = "macos", target_os = "linux")))]
     {

crates/openfang-cli/src/mcp.rs

@@ -216,8 +216,13 @@ fn read_message(reader: &mut impl BufRead) -> io::Result<Option<Value>> {
 /// Write a Content-Length framed JSON-RPC response to the writer.
 fn write_message(writer: &mut impl Write, msg: &Value) {
     let body = serde_json::to_string(msg).unwrap_or_default();
-    let _ = write!(writer, "Content-Length: {}\r\n\r\n{}", body.len(), body);
-    let _ = writer.flush();
+    if let Err(e) = write!(writer, "Content-Length: {}\r\n\r\n{}", body.len(), body) {
+        eprintln!("MCP write error: {e}");
+        return;
+    }
+    if let Err(e) = writer.flush() {
+        eprintln!("MCP flush error: {e}");
+    }
 }
 
 /// Handle a JSON-RPC message and return an optional response.
@@ -225,6 +230,10 @@ fn handle_message(backend: &McpBackend, msg: &Value) -> Option<Value> {
     let method = msg["method"].as_str().unwrap_or("");
     let id = msg.get("id").cloned();
 
+    // Per JSON-RPC 2.0 spec: requests MUST have an id field.
+    // Use null if missing so we always send a response.
+    let rid = id.unwrap_or(Value::Null);
+
     match method {
         "initialize" => {
             let result = json!({
@@ -234,10 +243,10 @@ fn handle_message(backend: &McpBackend, msg: &Value) -> Option<Value> {
                 },
                 "serverInfo": {
                     "name": "openfang",
-                    "version": "0.1.0"
+                    "version": env!("CARGO_PKG_VERSION")
                 }
             });
-            Some(jsonrpc_response(id?, result))
+            Some(jsonrpc_response(rid, result))
         }
 
         "notifications/initialized" => None, // Notification, no response
@@ -269,7 +278,7 @@ fn handle_message(backend: &McpBackend, msg: &Value) -> Option<Value> {
                     })
                 })
                 .collect();
-            Some(jsonrpc_response(id?, json!({ "tools": tools })))
+            Some(jsonrpc_response(rid, json!({ "tools": tools })))
         }
 
         "tools/call" => {
@@ -281,14 +290,14 @@ fn handle_message(backend: &McpBackend, msg: &Value) -> Option<Value> {
                 .to_string();
 
             if message.is_empty() {
-                return Some(jsonrpc_error(id?, -32602, "Missing 'message' argument"));
+                return Some(jsonrpc_error(rid, -32602, "Missing 'message' argument"));
             }
 
             let agent_id = match backend.resolve_tool_agent(tool_name) {
                 Some(id) => id,
                 None => {
                     return Some(jsonrpc_error(
-                        id?,
+                        rid,
                         -32602,
                         &format!("Unknown tool: {tool_name}"),
                     ));
@@ -297,7 +306,7 @@ fn handle_message(backend: &McpBackend, msg: &Value) -> Option<Value> {
 
             match backend.send_message(&agent_id, &message) {
                 Ok(response) => Some(jsonrpc_response(
-                    id?,
+                    rid,
                     json!({
                         "content": [{
                             "type": "text",
@@ -306,7 +315,7 @@ fn handle_message(backend: &McpBackend, msg: &Value) -> Option<Value> {
                     }),
                 )),
                 Err(e) => Some(jsonrpc_response(
-                    id?,
+                    rid,
                     json!({
                         "content": [{
                             "type": "text",
@@ -319,8 +328,8 @@ fn handle_message(backend: &McpBackend, msg: &Value) -> Option<Value> {
         }
 
         _ => {
-            // Unknown method
-            id.map(|id| jsonrpc_error(id, -32601, &format!("Method not found: {method}")))
+            // Unknown method — always respond with error
+            Some(jsonrpc_error(rid, -32601, &format!("Method not found: {method}")))
         }
     }
 }

crates/openfang-cli/src/tui/screens/init_wizard.rs

@@ -72,7 +72,7 @@ const PROVIDERS: &[ProviderInfo] = &[
         name: "openrouter",
         display: "OpenRouter",
         env_var: "OPENROUTER_API_KEY",
-        default_model: "openrouter/auto",
+        default_model: "openrouter/google/gemini-2.5-flash",
         needs_key: true,
         hint: "",
     },
@@ -100,6 +100,102 @@ const PROVIDERS: &[ProviderInfo] = &[
         needs_key: true,
         hint: "",
     },
+    ProviderInfo {
+        name: "xai",
+        display: "xAI (Grok)",
+        env_var: "XAI_API_KEY",
+        default_model: "grok-4-0709",
+        needs_key: true,
+        hint: "",
+    },
+    ProviderInfo {
+        name: "perplexity",
+        display: "Perplexity",
+        env_var: "PERPLEXITY_API_KEY",
+        default_model: "sonar-pro",
+        needs_key: true,
+        hint: "",
+    },
+    ProviderInfo {
+        name: "cohere",
+        display: "Cohere",
+        env_var: "COHERE_API_KEY",
+        default_model: "command-a-03-2025",
+        needs_key: true,
+        hint: "",
+    },
+    ProviderInfo {
+        name: "cerebras",
+        display: "Cerebras",
+        env_var: "CEREBRAS_API_KEY",
+        default_model: "llama-4-scout-17b-16e-instruct",
+        needs_key: true,
+        hint: "fast inference",
+    },
+    ProviderInfo {
+        name: "sambanova",
+        display: "SambaNova",
+        env_var: "SAMBANOVA_API_KEY",
+        default_model: "DeepSeek-R1",
+        needs_key: true,
+        hint: "fast inference",
+    },
+    ProviderInfo {
+        name: "qwen",
+        display: "Qwen (Alibaba)",
+        env_var: "QWEN_API_KEY",
+        default_model: "qwen-plus",
+        needs_key: true,
+        hint: "",
+    },
+    ProviderInfo {
+        name: "huggingface",
+        display: "Hugging Face",
+        env_var: "HUGGINGFACE_API_KEY",
+        default_model: "meta-llama/Llama-3.3-70B-Instruct",
+        needs_key: true,
+        hint: "",
+    },
+    ProviderInfo {
+        name: "github-copilot",
+        display: "GitHub Copilot",
+        env_var: "GITHUB_TOKEN",
+        default_model: "gpt-4o",
+        needs_key: true,
+        hint: "via PAT",
+    },
+    ProviderInfo {
+        name: "replicate",
+        display: "Replicate",
+        env_var: "REPLICATE_API_KEY",
+        default_model: "meta/meta-llama-3-70b-instruct",
+        needs_key: true,
+        hint: "",
+    },
+    ProviderInfo {
+        name: "venice",
+        display: "Venice.ai",
+        env_var: "VENICE_API_KEY",
+        default_model: "venice-uncensored",
+        needs_key: true,
+        hint: "uncensored",
+    },
+    ProviderInfo {
+        name: "ai21",
+        display: "AI21",
+        env_var: "AI21_API_KEY",
+        default_model: "jamba-1.5-large",
+        needs_key: true,
+        hint: "",
+    },
+    ProviderInfo {
+        name: "claude-code",
+        display: "Claude Code",
+        env_var: "",
+        default_model: "claude-code/sonnet",
+        needs_key: false,
+        hint: "no API key",
+    },
     ProviderInfo {
         name: "ollama",
         display: "Ollama",
@@ -116,6 +212,14 @@ const PROVIDERS: &[ProviderInfo] = &[
         needs_key: false,
         hint: "local",
     },
+    ProviderInfo {
+        name: "vllm",
+        display: "vLLM",
+        env_var: "VLLM_API_KEY",
+        default_model: "local-model",
+        needs_key: false,
+        hint: "local",
+    },
 ];
 
 // ── Public result type ─────────────────────────────────────────────────────
@@ -287,15 +391,23 @@ impl State {
         self.provider_order.clear();
         let gemini_via_google = std::env::var("GOOGLE_API_KEY").is_ok();
         for (i, p) in PROVIDERS.iter().enumerate() {
-            let detected =
-                std::env::var(p.env_var).is_ok() || (p.name == "gemini" && gemini_via_google);
+            let detected = if p.name == "claude-code" {
+                openfang_runtime::drivers::claude_code::claude_code_available()
+            } else {
+                (!p.env_var.is_empty() && std::env::var(p.env_var).is_ok())
+                    || (p.name == "gemini" && gemini_via_google)
+            };
             if detected {
                 self.provider_order.push(i);
             }
         }
         for (i, p) in PROVIDERS.iter().enumerate() {
-            let detected =
-                std::env::var(p.env_var).is_ok() || (p.name == "gemini" && gemini_via_google);
+            let detected = if p.name == "claude-code" {
+                openfang_runtime::drivers::claude_code::claude_code_available()
+            } else {
+                (!p.env_var.is_empty() && std::env::var(p.env_var).is_ok())
+                    || (p.name == "gemini" && gemini_via_google)
+            };
             if !detected {
                 self.provider_order.push(i);
             }
@@ -334,7 +446,10 @@ impl State {
 
     fn is_provider_detected(&self, prov_idx: usize) -> bool {
         let p = &PROVIDERS[prov_idx];
-        std::env::var(p.env_var).is_ok()
+        if p.name == "claude-code" {
+            return openfang_runtime::drivers::claude_code::claude_code_available();
+        }
+        (!p.env_var.is_empty() && std::env::var(p.env_var).is_ok())
             || (p.name == "gemini" && std::env::var("GOOGLE_API_KEY").is_ok())
     }
 
@@ -468,6 +583,13 @@ fn tier_label(tier: ModelTier) -> &'static str {
 // ── Entry point ────────────────────────────────────────────────────────────
 
 pub fn run() -> InitResult {
+    // Guard against non-TTY environments (Docker, piped, CI/CD)
+    if !std::io::IsTerminal::is_terminal(&std::io::stdin())
+        || !std::io::IsTerminal::is_terminal(&std::io::stdout())
+    {
+        return InitResult::Cancelled;
+    }
+
     let original_hook = std::panic::take_hook();
     std::panic::set_hook(Box::new(move |info| {
         ratatui::restore();
@@ -988,6 +1110,12 @@ complex_threshold = 500
     };
 
     let config_path = openfang_dir.join("config.toml");
+    let api_key_line = if p.env_var.is_empty() {
+        String::new()
+    } else {
+        format!("api_key_env = \"{}\"", p.env_var)
+    };
+
     let config = format!(
         r#"# OpenFang Agent OS configuration
 # See https://github.com/RightNow-AI/openfang for documentation
@@ -997,13 +1125,12 @@ api_listen = "127.0.0.1:4200"
 [default_model]
 provider = "{provider}"
 model = "{model}"
-api_key_env = "{env_var}"
+{api_key_line}
 
 [memory]
 decay_rate = 0.05
 {routing_section}"#,
         provider = p.name,
-        env_var = p.env_var,
     );
 
     match std::fs::write(&config_path, &config) {
@@ -1600,7 +1727,13 @@ fn draw_provider(f: &mut Frame, area: Rect, state: &mut State) {
                 Span::styled("  ", Style::default())
             };
             let name_span = Span::raw(format!("{:<14}", p.display));
-            let hint_text = if detected {
+            let hint_text = if p.name == "claude-code" {
+                if detected {
+                    "CLI detected".to_string()
+                } else {
+                    "no API key needed".to_string()
+                }
+            } else if detected {
                 format!("{} detected", p.env_var)
             } else if !p.needs_key {
                 "local, no key needed".to_string()
@@ -1918,11 +2051,7 @@ fn draw_routing_pick(f: &mut Frame, area: Rect, state: &mut State, tier: usize)
                 .split('/')
                 .next_back()
                 .unwrap_or(&state.routing_models[t]);
-            let display = if short.len() > 14 {
-                &short[..14]
-            } else {
-                short
-            };
+            let display = openfang_types::truncate_str(short, 14);
             summary_spans.push(Span::styled(
                 format!("{name}:{display}"),
                 Style::default().fg(*c),

crates/openfang-cli/src/tui/screens/memory.rs

@@ -317,7 +317,10 @@ fn draw_agent_select(f: &mut Frame, area: Rect, state: &mut MemoryState) {
             .iter()
             .map(|a| {
                 let id_short = if a.id.len() > 12 {
-                    format!("{}\u{2026}", &a.id[..12])
+                    format!(
+                        "{}\u{2026}",
+                        openfang_types::truncate_str(&a.id, 12)
+                    )
                 } else {
                     a.id.clone()
                 };
@@ -405,7 +408,10 @@ fn draw_kv_browser(f: &mut Frame, area: Rect, state: &mut MemoryState) {
             .iter()
             .map(|kv| {
                 let val_display = if kv.value.len() > 40 {
-                    format!("{}\u{2026}", &kv.value[..39])
+                    format!(
+                        "{}\u{2026}",
+                        openfang_types::truncate_str(&kv.value, 39)
+                    )
                 } else {
                     kv.value.clone()
                 };

crates/openfang-cli/src/tui/screens/peers.rs

@@ -149,7 +149,10 @@ pub fn draw(f: &mut Frame, area: Rect, state: &mut PeersState) {
             .iter()
             .map(|p| {
                 let id_short = if p.node_id.len() > 12 {
-                    format!("{}\u{2026}", &p.node_id[..12])
+                    format!(
+                        "{}\u{2026}",
+                        openfang_types::truncate_str(&p.node_id, 12)
+                    )
                 } else {
                     p.node_id.clone()
                 };

crates/openfang-cli/src/tui/screens/sessions.rs

@@ -251,7 +251,10 @@ pub fn draw(f: &mut Frame, area: Rect, state: &mut SessionsState) {
             .map(|&idx| {
                 let s = &state.sessions[idx];
                 let id_short = if s.id.len() > 12 {
-                    format!("{}\u{2026}", &s.id[..12])
+                    format!(
+                        "{}\u{2026}",
+                        openfang_types::truncate_str(&s.id, 12)
+                    )
                 } else {
                     s.id.clone()
                 };

crates/openfang-cli/src/tui/screens/wizard.rs

@@ -40,7 +40,7 @@ const PROVIDERS: &[ProviderInfo] = &[
     ProviderInfo {
         name: "openrouter",
         env_var: "OPENROUTER_API_KEY",
-        default_model: "anthropic/claude-sonnet-4-20250514",
+        default_model: "google/gemini-2.5-flash",
         needs_key: true,
     },
     ProviderInfo {
@@ -127,6 +127,12 @@ const PROVIDERS: &[ProviderInfo] = &[
         default_model: "codegeex-4",
         needs_key: true,
     },
+    ProviderInfo {
+        name: "claude-code",
+        env_var: "",
+        default_model: "claude-code/sonnet",
+        needs_key: false,
+    },
     ProviderInfo {
         name: "ollama",
         env_var: "OLLAMA_API_KEY",
@@ -215,13 +221,23 @@ impl WizardState {
         self.provider_order.clear();
         // Detected providers first
         for (i, p) in PROVIDERS.iter().enumerate() {
-            if std::env::var(p.env_var).is_ok() {
+            let detected = if p.name == "claude-code" {
+                openfang_runtime::drivers::claude_code::claude_code_available()
+            } else {
+                !p.env_var.is_empty() && std::env::var(p.env_var).is_ok()
+            };
+            if detected {
                 self.provider_order.push(i);
             }
         }
         // Then the rest
         for (i, p) in PROVIDERS.iter().enumerate() {
-            if std::env::var(p.env_var).is_err() {
+            let detected = if p.name == "claude-code" {
+                openfang_runtime::drivers::claude_code::claude_code_available()
+            } else {
+                !p.env_var.is_empty() && std::env::var(p.env_var).is_ok()
+            };
+            if !detected {
                 self.provider_order.push(i);
             }
         }
@@ -376,6 +392,8 @@ impl WizardState {
 
         let api_key_line = if !self.api_key_input.is_empty() {
             format!("api_key = \"{}\"", self.api_key_input)
+        } else if p.env_var.is_empty() {
+            String::new()
         } else {
             format!("api_key_env = \"{}\"", p.env_var)
         };
@@ -506,9 +524,15 @@ fn draw_provider(f: &mut Frame, area: Rect, state: &mut WizardState) {
         .iter()
         .map(|&idx| {
             let p = &PROVIDERS[idx];
-            let hint = if !p.needs_key {
+            let hint = if p.name == "claude-code" {
+                if openfang_runtime::drivers::claude_code::claude_code_available() {
+                    "CLI detected".to_string()
+                } else {
+                    "no API key needed".to_string()
+                }
+            } else if !p.needs_key {
                 "local, no key needed".to_string()
-            } else if std::env::var(p.env_var).is_ok() {
+            } else if !p.env_var.is_empty() && std::env::var(p.env_var).is_ok() {
                 format!("{} detected", p.env_var)
             } else {
                 format!("requires {}", p.env_var)

crates/openfang-extensions/integrations/aws.toml

@@ -8,7 +8,7 @@ tags = ["cloud", "amazon", "infrastructure", "s3", "ec2", "lambda", "devops"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-aws"]
+args = ["-y", "@aws-mcp/server-aws"]
 
 [[required_env]]
 name = "AWS_ACCESS_KEY_ID"

crates/openfang-extensions/integrations/azure-mcp.toml

@@ -8,7 +8,7 @@ tags = ["cloud", "microsoft", "infrastructure", "azure", "devops", "enterprise"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-azure"]
+args = ["-y", "@azure/mcp@latest", "server", "start"]
 
 [[required_env]]
 name = "AZURE_SUBSCRIPTION_ID"

crates/openfang-extensions/integrations/bitbucket.toml

@@ -8,7 +8,7 @@ tags = ["git", "vcs", "code", "pull-requests", "ci", "atlassian"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-bitbucket"]
+args = ["-y", "@atlassian-mcp-server/bitbucket"]
 
 [[required_env]]
 name = "BITBUCKET_USERNAME"

crates/openfang-extensions/integrations/brave-search.toml

@@ -8,7 +8,7 @@ tags = ["search", "web", "brave", "api", "information-retrieval"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-brave-search"]
+args = ["-y", "@modelcontextprotocol/server-brave-search"]
 
 [[required_env]]
 name = "BRAVE_API_KEY"

crates/openfang-extensions/integrations/discord-mcp.toml

@@ -8,7 +8,7 @@ tags = ["chat", "messaging", "community", "gaming", "voice"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-discord"]
+args = ["-y", "mcp-discord"]
 
 [[required_env]]
 name = "DISCORD_BOT_TOKEN"

crates/openfang-extensions/integrations/dropbox.toml

@@ -8,7 +8,7 @@ tags = ["files", "storage", "cloud-storage", "sync", "sharing"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-dropbox"]
+args = ["-y", "@microagents/mcp-server-dropbox"]
 
 [[required_env]]
 name = "DROPBOX_ACCESS_TOKEN"

crates/openfang-extensions/integrations/elasticsearch.toml

@@ -8,7 +8,7 @@ tags = ["search", "database", "indexing", "analytics", "full-text"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-elasticsearch"]
+args = ["-y", "@elastic/mcp-server-elasticsearch"]
 
 [[required_env]]
 name = "ELASTICSEARCH_URL"

crates/openfang-extensions/integrations/exa-search.toml

@@ -8,7 +8,7 @@ tags = ["search", "web", "ai", "neural", "semantic", "information-retrieval"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-exa"]
+args = ["-y", "exa-mcp-server"]
 
 [[required_env]]
 name = "EXA_API_KEY"

crates/openfang-extensions/integrations/gcp-mcp.toml

@@ -8,7 +8,7 @@ tags = ["cloud", "google", "infrastructure", "gce", "gcs", "bigquery", "devops"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-gcp"]
+args = ["-y", "@google-cloud/gcloud-mcp"]
 
 [[required_env]]
 name = "GOOGLE_APPLICATION_CREDENTIALS"

crates/openfang-extensions/integrations/github.toml

@@ -8,7 +8,7 @@ tags = ["git", "vcs", "code", "issues", "pull-requests", "ci"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-github"]
+args = ["-y", "@modelcontextprotocol/server-github"]
 
 [[required_env]]
 name = "GITHUB_PERSONAL_ACCESS_TOKEN"

crates/openfang-extensions/integrations/gitlab.toml

@@ -8,7 +8,7 @@ tags = ["git", "vcs", "code", "merge-requests", "ci", "devops"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-gitlab"]
+args = ["-y", "@modelcontextprotocol/server-gitlab"]
 
 [[required_env]]
 name = "GITLAB_PERSONAL_ACCESS_TOKEN"

crates/openfang-extensions/integrations/gmail.toml

@@ -8,7 +8,7 @@ tags = ["email", "google", "messaging", "inbox", "communication"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@anthropic/server-gmail"]
+args = ["-y", "@gongrzhe/server-gmail-autoauth-mcp"]
 
 [oauth]
 provider = "google"

crates/openfang-extensions/integrations/google-calendar.toml

@@ -8,7 +8,7 @@ tags = ["calendar", "scheduling", "google", "events", "meetings"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@anthropic/server-google-calendar"]
+args = ["-y", "@cocal/google-calendar-mcp"]
 
 [oauth]
 provider = "google"

crates/openfang-extensions/integrations/google-drive.toml

@@ -8,7 +8,7 @@ tags = ["files", "storage", "google", "documents", "cloud-storage"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@anthropic/server-google-drive"]
+args = ["-y", "@modelcontextprotocol/server-gdrive"]
 
 [oauth]
 provider = "google"

crates/openfang-extensions/integrations/jira.toml

@@ -8,7 +8,7 @@ tags = ["project-management", "issues", "agile", "atlassian", "tracking"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-atlassian"]
+args = ["-y", "@aashari/mcp-server-atlassian-jira"]
 
 [[required_env]]
 name = "JIRA_API_TOKEN"

crates/openfang-extensions/integrations/linear.toml

@@ -8,7 +8,7 @@ tags = ["project-management", "issues", "agile", "tracking", "sprint"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-linear"]
+args = ["-y", "linear-mcp"]
 
 [[required_env]]
 name = "LINEAR_API_KEY"

crates/openfang-extensions/integrations/mongodb.toml

@@ -8,7 +8,7 @@ tags = ["database", "nosql", "document", "mongo", "queries"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-mongodb"]
+args = ["-y", "@mongodb-js/mongodb-mcp-server"]
 
 [[required_env]]
 name = "MONGODB_URI"

crates/openfang-extensions/integrations/notion.toml

@@ -8,7 +8,7 @@ tags = ["notes", "wiki", "knowledge-base", "documentation", "databases"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-notion"]
+args = ["-y", "@notionhq/notion-mcp-server"]
 
 [[required_env]]
 name = "NOTION_API_KEY"

crates/openfang-extensions/integrations/postgresql.toml

@@ -8,7 +8,7 @@ tags = ["database", "sql", "relational", "postgres", "queries"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-postgres"]
+args = ["-y", "@modelcontextprotocol/server-postgres"]
 
 [[required_env]]
 name = "POSTGRES_CONNECTION_STRING"

crates/openfang-extensions/integrations/redis.toml

@@ -8,7 +8,7 @@ tags = ["database", "cache", "key-value", "in-memory", "nosql"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-redis"]
+args = ["-y", "@modelcontextprotocol/server-redis"]
 
 [[required_env]]
 name = "REDIS_URL"

crates/openfang-extensions/integrations/sentry.toml

@@ -8,7 +8,7 @@ tags = ["monitoring", "errors", "debugging", "observability", "apm"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-sentry"]
+args = ["-y", "@sentry/mcp-server"]
 
 [[required_env]]
 name = "SENTRY_AUTH_TOKEN"

crates/openfang-extensions/integrations/slack.toml

@@ -8,7 +8,7 @@ tags = ["chat", "messaging", "team", "channels", "collaboration"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-slack"]
+args = ["-y", "@modelcontextprotocol/server-slack"]
 
 [[required_env]]
 name = "SLACK_BOT_TOKEN"

crates/openfang-extensions/integrations/sqlite-mcp.toml

@@ -8,7 +8,7 @@ tags = ["database", "sql", "relational", "sqlite", "local", "embedded"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-sqlite"]
+args = ["-y", "@modelcontextprotocol/server-sqlite"]
 
 [[required_env]]
 name = "SQLITE_DB_PATH"

crates/openfang-extensions/integrations/teams-mcp.toml

@@ -8,7 +8,7 @@ tags = ["chat", "messaging", "microsoft", "enterprise", "collaboration"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-teams"]
+args = ["-y", "teams-mcp"]
 
 [oauth]
 provider = "microsoft"

crates/openfang-extensions/integrations/todoist.toml

@@ -8,7 +8,7 @@ tags = ["tasks", "todo", "project-management", "productivity", "gtd"]
 [transport]
 type = "stdio"
 command = "npx"
-args = ["@modelcontextprotocol/server-todoist"]
+args = ["-y", "todoist-mcp"]
 
 [[required_env]]
 name = "TODOIST_API_KEY"

crates/openfang-hands/bundled/browser/HAND.toml

@@ -17,8 +17,8 @@ tools = [
 key = "python3"
 label = "Python 3 must be installed"
 requirement_type = "binary"
-check_value = "python"
-description = "Python 3 is required to run Playwright, the browser automation library that powers this hand."
+check_value = "python3"
+description = "Python 3 is required for installing and running the Playwright browser automation library. Python 3.8 or newer is recommended."
 
 [requires.install]
 macos = "brew install python3"
@@ -26,27 +26,25 @@ windows = "winget install Python.Python.3.12"
 linux_apt = "sudo apt install python3"
 linux_dnf = "sudo dnf install python3"
 linux_pacman = "sudo pacman -S python"
+pip = "python3 --version"
 manual_url = "https://www.python.org/downloads/"
-estimated_time = "2-5 min"
+estimated_time = "1-3 min"
 
 [[requires]]
-key = "playwright"
-label = "Playwright must be installed"
+key = "chromium"
+label = "Chromium or Google Chrome must be installed"
 requirement_type = "binary"
-check_value = "playwright"
-description = "Playwright is a browser automation framework. After installing via pip, you also need to install browser binaries."
+check_value = "chromium"
+description = "A Chromium-based browser is required. Google Chrome, Chromium, or any Chromium derivative will work. You can also set the CHROME_PATH environment variable to point to your browser binary."
 
 [requires.install]
-macos = "pip3 install playwright && playwright install chromium"
-windows = "pip install playwright && playwright install chromium"
-linux_apt = "pip3 install playwright && playwright install chromium"
-pip = "pip install playwright && playwright install chromium"
-manual_url = "https://playwright.dev/python/docs/intro"
-estimated_time = "3-5 min"
-steps = [
-    "Install Playwright: pip install playwright",
-    "Install browser binaries: playwright install chromium",
-]
+macos = "brew install --cask google-chrome"
+windows = "winget install Google.Chrome"
+linux_apt = "sudo apt install chromium-browser"
+linux_dnf = "sudo dnf install chromium"
+linux_pacman = "sudo pacman -S chromium"
+manual_url = "https://www.google.com/chrome/"
+estimated_time = "1-3 min"
 
 # ─── Configurable settings ───────────────────────────────────────────────────
 

crates/openfang-hands/bundled/clip/HAND.toml

@@ -118,6 +118,7 @@ key = "elevenlabs_api_key"
 label = "ElevenLabs API Key"
 description = "API key from elevenlabs.io for high-quality text-to-speech. Required when ElevenLabs TTS is selected."
 setting_type = "text"
+env_var = "ELEVENLABS_API_KEY"
 default = ""
 
 # ─── Publishing settings ────────────────────────────────────────────────────

crates/openfang-hands/src/bundled.rs

@@ -187,7 +187,7 @@ mod tests {
         assert_eq!(def.name, "Browser Hand");
         assert_eq!(def.category, crate::HandCategory::Productivity);
         assert!(def.skill_content.is_some());
-        assert!(!def.requires.is_empty()); // requires python3, playwright
+        assert!(!def.requires.is_empty()); // requires python3 + chromium
         assert_eq!(def.requires.len(), 2);
         assert!(def.tools.contains(&"browser_navigate".to_string()));
         assert!(def.tools.contains(&"browser_click".to_string()));

crates/openfang-hands/src/lib.rs

@@ -29,6 +29,8 @@ pub enum HandError {
     TomlParse(String),
     #[error("IO error: {0}")]
     Io(#[from] std::io::Error),
+    #[error("Config error: {0}")]
+    Config(String),
 }
 
 pub type HandResult<T> = Result<T, HandError>;
@@ -172,6 +174,10 @@ pub struct HandSetting {
     pub default: String,
     #[serde(default)]
     pub options: Vec<HandSettingOption>,
+    /// Env var name to expose when a text-type setting has a value
+    /// (e.g. `ELEVENLABS_API_KEY` for an API key text field).
+    #[serde(default)]
+    pub env_var: Option<String>,
 }
 
 /// Result of resolving user-chosen settings against the schema.
@@ -227,6 +233,9 @@ pub fn resolve_settings(
             HandSettingType::Text => {
                 if !chosen_value.is_empty() {
                     lines.push(format!("- {}: {}", setting.label, chosen_value));
+                    if let Some(ref env) = setting.env_var {
+                        env_vars.push(env.clone());
+                    }
                 }
             }
         }
@@ -550,6 +559,7 @@ metrics = []
                     binary: None,
                 },
             ],
+            env_var: None,
         }];
 
         // User picks groq
@@ -583,6 +593,7 @@ metrics = []
                     binary: None,
                 },
             ],
+            env_var: None,
         }];
 
         // Empty config → uses default "auto"
@@ -604,6 +615,7 @@ metrics = []
                 setting_type: HandSettingType::Toggle,
                 default: "false".to_string(),
                 options: vec![],
+                env_var: None,
             },
             HandSetting {
                 key: "custom_model".to_string(),
@@ -612,6 +624,7 @@ metrics = []
                 setting_type: HandSettingType::Text,
                 default: String::new(),
                 options: vec![],
+                env_var: None,
             },
         ];
 

crates/openfang-hands/src/registry.rs

@@ -52,6 +52,59 @@ impl HandRegistry {
         }
     }
 
+    /// Persist active hand state to disk so it survives restarts.
+    pub fn persist_state(&self, path: &std::path::Path) -> HandResult<()> {
+        let entries: Vec<serde_json::Value> = self
+            .instances
+            .iter()
+            .filter(|e| e.status == HandStatus::Active)
+            .map(|e| {
+                serde_json::json!({
+                    "hand_id": e.hand_id,
+                    "config": e.config,
+                    "agent_id": e.agent_id,
+                })
+            })
+            .collect();
+        let json = serde_json::to_string_pretty(&entries)
+            .map_err(|e| HandError::Config(format!("serialize hand state: {e}")))?;
+        std::fs::write(path, json)
+            .map_err(|e| HandError::Config(format!("write hand state: {e}")))?;
+        Ok(())
+    }
+
+    /// Load persisted hand state and re-activate hands.
+    /// Returns list of (hand_id, config, old_agent_id) that should be activated.
+    /// The `old_agent_id` is the agent UUID from before the restart, used to
+    /// reassign cron jobs to the newly spawned agent (issue #402).
+    pub fn load_state(
+        path: &std::path::Path,
+    ) -> Vec<(String, HashMap<String, serde_json::Value>, Option<AgentId>)> {
+        let data = match std::fs::read_to_string(path) {
+            Ok(d) => d,
+            Err(_) => return Vec::new(),
+        };
+        let entries: Vec<serde_json::Value> = match serde_json::from_str(&data) {
+            Ok(e) => e,
+            Err(e) => {
+                warn!("Failed to parse hand state file: {e}");
+                return Vec::new();
+            }
+        };
+        entries
+            .into_iter()
+            .filter_map(|e| {
+                let hand_id = e["hand_id"].as_str()?.to_string();
+                let config: HashMap<String, serde_json::Value> =
+                    serde_json::from_value(e["config"].clone()).unwrap_or_default();
+                let old_agent_id: Option<AgentId> = e
+                    .get("agent_id")
+                    .and_then(|v| serde_json::from_value(v.clone()).ok());
+                Some((hand_id, config, old_agent_id))
+            })
+            .collect()
+    }
+
     /// Load all bundled hand definitions. Returns count of definitions loaded.
     pub fn load_bundled(&self) -> usize {
         let bundled = bundled::bundled_hands();
@@ -311,8 +364,25 @@ impl Default for HandRegistry {
 fn check_requirement(req: &HandRequirement) -> bool {
     match req.requirement_type {
         RequirementType::Binary => {
-            // Check if binary exists on PATH
-            which_binary(&req.check_value)
+            // Special handling for python3: must actually run the command and verify
+            // the output contains "Python 3", because Windows ships a python3.exe
+            // Store shim that exists on PATH but doesn't actually work.
+            if req.check_value == "python3" {
+                return check_python3_available();
+            }
+            // Check if binary exists on PATH.
+            if which_binary(&req.check_value) {
+                return true;
+            }
+            if req.check_value == "chromium" {
+                // Try common Chromium/Chrome binary names across platforms
+                return which_binary("chromium-browser")
+                    || which_binary("google-chrome")
+                    || which_binary("google-chrome-stable")
+                    || which_binary("chrome")
+                    || std::env::var("CHROME_PATH").map(|v| !v.is_empty()).unwrap_or(false);
+            }
+            false
         }
         RequirementType::EnvVar | RequirementType::ApiKey => {
             // Check if env var is set and non-empty
@@ -323,6 +393,44 @@ fn check_requirement(req: &HandRequirement) -> bool {
     }
 }
 
+/// Check if Python 3 is actually available by running the command and checking
+/// the version output. This avoids false negatives from Windows Store shims
+/// (python3.exe that just opens the Microsoft Store) and false positives from
+/// Python 2 installations where `python` exists but is Python 2.
+fn check_python3_available() -> bool {
+    // Try "python3 --version" first (Linux/macOS, some Windows installs)
+    if run_returns_python3("python3") {
+        return true;
+    }
+    // Try "python --version" (Windows commonly uses this, Docker containers too)
+    if run_returns_python3("python") {
+        return true;
+    }
+    false
+}
+
+/// Run `{cmd} --version` and return true if the output contains "Python 3".
+fn run_returns_python3(cmd: &str) -> bool {
+    match std::process::Command::new(cmd)
+        .arg("--version")
+        .stdout(std::process::Stdio::piped())
+        .stderr(std::process::Stdio::piped())
+        .stdin(std::process::Stdio::null())
+        .output()
+    {
+        Ok(output) => {
+            if !output.status.success() {
+                return false;
+            }
+            // Python --version may print to stdout or stderr depending on version
+            let stdout = String::from_utf8_lossy(&output.stdout);
+            let stderr = String::from_utf8_lossy(&output.stderr);
+            stdout.contains("Python 3") || stderr.contains("Python 3")
+        }
+        Err(_) => false,
+    }
+}
+
 /// Check if a binary is on PATH (cross-platform).
 fn which_binary(name: &str) -> bool {
     let path_var = std::env::var("PATH").unwrap_or_default();

crates/openfang-kernel/Cargo.toml

@@ -23,6 +23,7 @@ crossbeam = { workspace = true }
 tracing = { workspace = true }
 tracing-subscriber = { workspace = true }
 chrono = { workspace = true }
+chrono-tz = { workspace = true }
 uuid = { workspace = true }
 thiserror = { workspace = true }
 async-trait = { workspace = true }

crates/openfang-kernel/src/config.rs

@@ -50,6 +50,25 @@ pub fn load_config(path: Option<&Path>) -> KernelConfig {
                         tbl.remove("include");
                     }
 
+                    // Migrate misplaced api_key/api_listen from [api] section to root level.
+                    // The old config schema incorrectly grouped these under [api], so many
+                    // users have them in the wrong place. Move them up if not already at root.
+                    if let toml::Value::Table(ref mut tbl) = root_value {
+                        if let Some(toml::Value::Table(api_section)) = tbl.get("api").cloned() {
+                            for key in &["api_key", "api_listen", "log_level"] {
+                                if !tbl.contains_key(*key) {
+                                    if let Some(val) = api_section.get(*key) {
+                                        tracing::info!(
+                                            key,
+                                            "Migrating misplaced config field from [api] to root level"
+                                        );
+                                        tbl.insert(key.to_string(), val.clone());
+                                    }
+                                }
+                            }
+                        }
+                    }
+
                     match root_value.try_into::<KernelConfig>() {
                         Ok(config) => {
                             info!(path = %config_path.display(), "Loaded configuration");

crates/openfang-kernel/src/config_reload.rs

@@ -241,6 +241,10 @@ pub fn build_reload_plan(old: &KernelConfig, new: &KernelConfig) -> ReloadPlan {
         plan.hot_actions.push(HotAction::ReloadProviderUrls);
     }
 
+    if field_changed(&old.provider_api_keys, &new.provider_api_keys) {
+        plan.noop_changes.push("provider_api_keys changed (takes effect on next driver init)".to_string());
+    }
+
     // ----- No-op fields -----
 
     if old.log_level != new.log_level {

crates/openfang-kernel/src/cron.rs

@@ -216,6 +216,46 @@ impl CronScheduler {
         self.jobs.iter().map(|r| r.value().job.clone()).collect()
     }
 
+    /// Reassign all cron jobs from `old_agent_id` to `new_agent_id`.
+    ///
+    /// Used when a hand agent is respawned (e.g. after daemon restart) and
+    /// gets a new UUID. Without this, persisted cron jobs would reference
+    /// the stale old agent ID and fail silently.
+    ///
+    /// Returns the number of jobs reassigned.
+    pub fn reassign_agent_jobs(&self, old_agent_id: AgentId, new_agent_id: AgentId) -> usize {
+        let mut count = 0;
+        for mut entry in self.jobs.iter_mut() {
+            if entry.value().job.agent_id == old_agent_id {
+                entry.value_mut().job.agent_id = new_agent_id;
+                // Reset consecutive errors so the job gets a fresh start
+                // with the new agent.
+                entry.value_mut().consecutive_errors = 0;
+                if !entry.value().job.enabled {
+                    // Re-enable jobs that were auto-disabled due to the stale
+                    // agent ID causing repeated failures.
+                    if entry.value().last_status.as_deref().is_some_and(|s| {
+                        s.contains("not found") || s.contains("No such agent")
+                    }) {
+                        entry.value_mut().job.enabled = true;
+                        entry.value_mut().job.next_run =
+                            Some(compute_next_run(&entry.value().job.schedule));
+                    }
+                }
+                count += 1;
+            }
+        }
+        if count > 0 {
+            info!(
+                old_agent = %old_agent_id,
+                new_agent = %new_agent_id,
+                count,
+                "Reassigned cron jobs to new agent"
+            );
+        }
+        count
+    }
+
     /// Total number of tracked jobs.
     pub fn total_jobs(&self) -> usize {
         self.jobs.len()
@@ -324,7 +364,7 @@ pub fn compute_next_run_after(
     match schedule {
         CronSchedule::At { at } => *at,
         CronSchedule::Every { every_secs } => after + Duration::seconds(*every_secs as i64),
-        CronSchedule::Cron { expr, tz: _ } => {
+        CronSchedule::Cron { expr, tz } => {
             // Convert standard 5/6-field cron to 7-field for the `cron` crate.
             // Standard 5-field: min hour dom month dow
             // 6-field:          sec min hour dom month dow
@@ -341,10 +381,33 @@ pub fn compute_next_run_after(
             let base = after + Duration::seconds(1);
 
             match seven_field.parse::<cron::Schedule>() {
-                Ok(sched) => sched
-                    .after(&base)
-                    .next()
-                    .unwrap_or_else(|| after + Duration::hours(1)),
+                Ok(sched) => {
+                    // If a timezone is specified, compute the next fire time in
+                    // that timezone so DST and local offsets are respected, then
+                    // convert back to UTC for storage.
+                    let next_utc = match tz.as_deref() {
+                        Some(tz_str) if !tz_str.is_empty() && tz_str != "UTC" => {
+                            match tz_str.parse::<chrono_tz::Tz>() {
+                                Ok(timezone) => {
+                                    let base_local = base.with_timezone(&timezone);
+                                    sched
+                                        .after(&base_local)
+                                        .next()
+                                        .map(|dt| dt.with_timezone(&Utc))
+                                }
+                                Err(_) => {
+                                    warn!(
+                                        "Invalid timezone '{}' in cron job, falling back to UTC",
+                                        tz_str
+                                    );
+                                    sched.after(&base).next()
+                                }
+                            }
+                        }
+                        _ => sched.after(&base).next(),
+                    };
+                    next_utc.unwrap_or_else(|| after + Duration::hours(1))
+                }
                 Err(e) => {
                     warn!("Failed to parse cron expression '{}': {}", expr, e);
                     after + Duration::hours(1)
@@ -361,7 +424,7 @@ pub fn compute_next_run_after(
 #[cfg(test)]
 mod tests {
     use super::*;
-    use chrono::Duration;
+    use chrono::{Duration, Timelike};
     use openfang_types::scheduler::{CronAction, CronDelivery};
 
     /// Build a minimal valid `CronJob` with an `Every` schedule.
@@ -787,4 +850,259 @@ mod tests {
             status.len()
         );
     }
+
+    // -- timezone-aware cron (#473) -----------------------------------------
+
+    #[test]
+    fn test_cron_tz_shifts_next_run() {
+        // "0 9 * * *" in America/New_York (UTC-5 or UTC-4 depending on DST).
+        // The next fire time in UTC should differ from a plain UTC "0 9 * * *".
+        let schedule_utc = CronSchedule::Cron {
+            expr: "0 9 * * *".into(),
+            tz: None,
+        };
+        let schedule_ny = CronSchedule::Cron {
+            expr: "0 9 * * *".into(),
+            tz: Some("America/New_York".into()),
+        };
+        let now = Utc::now();
+        let next_utc = compute_next_run_after(&schedule_utc, now);
+        let next_ny = compute_next_run_after(&schedule_ny, now);
+
+        // The New York schedule should fire at 09:00 Eastern, which is 13:00
+        // or 14:00 UTC (depending on DST). In either case, it should NOT
+        // equal the plain UTC 09:00 result.
+        assert_ne!(
+            next_utc, next_ny,
+            "Timezone-aware schedule should produce a different UTC time"
+        );
+
+        // Verify the New York result, when converted to ET, shows hour 09.
+        let ny_tz: chrono_tz::Tz = "America/New_York".parse().unwrap();
+        let next_ny_local = next_ny.with_timezone(&ny_tz);
+        assert_eq!(
+            next_ny_local.hour(),
+            9,
+            "Expected 09:00 in America/New_York, got {:02}:{:02}",
+            next_ny_local.hour(),
+            next_ny_local.minute()
+        );
+    }
+
+    #[test]
+    fn test_cron_tz_none_defaults_to_utc() {
+        // tz: None should behave identically to tz: Some("UTC").
+        let schedule_none = CronSchedule::Cron {
+            expr: "30 12 * * *".into(),
+            tz: None,
+        };
+        let schedule_utc = CronSchedule::Cron {
+            expr: "30 12 * * *".into(),
+            tz: Some("UTC".into()),
+        };
+        let now = Utc::now();
+        let next_none = compute_next_run_after(&schedule_none, now);
+        let next_utc = compute_next_run_after(&schedule_utc, now);
+        assert_eq!(next_none, next_utc);
+    }
+
+    #[test]
+    fn test_cron_tz_empty_string_defaults_to_utc() {
+        let schedule_empty = CronSchedule::Cron {
+            expr: "30 12 * * *".into(),
+            tz: Some(String::new()),
+        };
+        let schedule_none = CronSchedule::Cron {
+            expr: "30 12 * * *".into(),
+            tz: None,
+        };
+        let now = Utc::now();
+        assert_eq!(
+            compute_next_run_after(&schedule_empty, now),
+            compute_next_run_after(&schedule_none, now)
+        );
+    }
+
+    #[test]
+    fn test_cron_tz_invalid_falls_back_to_utc() {
+        // An invalid timezone string should fall back to UTC, not panic.
+        let schedule_bad = CronSchedule::Cron {
+            expr: "0 9 * * *".into(),
+            tz: Some("Not/A_Timezone".into()),
+        };
+        let schedule_utc = CronSchedule::Cron {
+            expr: "0 9 * * *".into(),
+            tz: None,
+        };
+        let now = Utc::now();
+        let next_bad = compute_next_run_after(&schedule_bad, now);
+        let next_utc = compute_next_run_after(&schedule_utc, now);
+        // Invalid tz falls back to UTC computation — same result.
+        assert_eq!(next_bad, next_utc);
+    }
+
+    #[test]
+    fn test_cron_tz_asia_shanghai() {
+        // "0 8 * * *" in Asia/Shanghai (UTC+8) should fire at 00:00 UTC.
+        let schedule = CronSchedule::Cron {
+            expr: "0 8 * * *".into(),
+            tz: Some("Asia/Shanghai".into()),
+        };
+        let now = Utc::now();
+        let next = compute_next_run_after(&schedule, now);
+
+        let shanghai_tz: chrono_tz::Tz = "Asia/Shanghai".parse().unwrap();
+        let local = next.with_timezone(&shanghai_tz);
+        assert_eq!(local.hour(), 8);
+        assert_eq!(local.minute(), 0);
+
+        // In UTC, 08:00 Shanghai = 00:00 UTC.
+        assert_eq!(next.hour(), 0, "08:00 CST should be 00:00 UTC");
+    }
+
+    // -- reassign_agent_jobs (#461) -----------------------------------------
+
+    #[test]
+    fn test_reassign_agent_jobs_basic() {
+        let (sched, _tmp) = make_scheduler(100);
+        let old_agent = AgentId::new();
+        let new_agent = AgentId::new();
+
+        let mut j1 = make_job(old_agent);
+        j1.name = "cron-a".into();
+        let mut j2 = make_job(old_agent);
+        j2.name = "cron-b".into();
+
+        let id1 = sched.add_job(j1, false).unwrap();
+        let id2 = sched.add_job(j2, false).unwrap();
+
+        let count = sched.reassign_agent_jobs(old_agent, new_agent);
+        assert_eq!(count, 2);
+
+        // Both jobs should now belong to the new agent
+        let job1 = sched.get_job(id1).unwrap();
+        assert_eq!(job1.agent_id, new_agent);
+        let job2 = sched.get_job(id2).unwrap();
+        assert_eq!(job2.agent_id, new_agent);
+
+        // Old agent should have zero jobs
+        assert!(sched.list_jobs(old_agent).is_empty());
+        // New agent should have both
+        assert_eq!(sched.list_jobs(new_agent).len(), 2);
+    }
+
+    #[test]
+    fn test_reassign_agent_jobs_does_not_touch_other_agents() {
+        let (sched, _tmp) = make_scheduler(100);
+        let agent_a = AgentId::new();
+        let agent_b = AgentId::new();
+        let agent_c = AgentId::new();
+
+        let mut ja = make_job(agent_a);
+        ja.name = "job-a".into();
+        let mut jb = make_job(agent_b);
+        jb.name = "job-b".into();
+
+        let _id_a = sched.add_job(ja, false).unwrap();
+        let id_b = sched.add_job(jb, false).unwrap();
+
+        // Reassign agent_a -> agent_c
+        let count = sched.reassign_agent_jobs(agent_a, agent_c);
+        assert_eq!(count, 1);
+
+        // agent_b's job should be untouched
+        let job_b = sched.get_job(id_b).unwrap();
+        assert_eq!(job_b.agent_id, agent_b);
+    }
+
+    #[test]
+    fn test_reassign_agent_jobs_no_match_returns_zero() {
+        let (sched, _tmp) = make_scheduler(100);
+        let agent = AgentId::new();
+        let other = AgentId::new();
+
+        let job = make_job(agent);
+        sched.add_job(job, false).unwrap();
+
+        // Reassign a non-existent agent
+        let count = sched.reassign_agent_jobs(AgentId::new(), other);
+        assert_eq!(count, 0);
+    }
+
+    #[test]
+    fn test_reassign_agent_jobs_resets_consecutive_errors() {
+        let (sched, _tmp) = make_scheduler(100);
+        let old_agent = AgentId::new();
+        let new_agent = AgentId::new();
+
+        let job = make_job(old_agent);
+        let id = sched.add_job(job, false).unwrap();
+
+        // Simulate some failures
+        sched.record_failure(id, "agent not found");
+        sched.record_failure(id, "agent not found");
+        let meta = sched.get_meta(id).unwrap();
+        assert_eq!(meta.consecutive_errors, 2);
+
+        // Reassign
+        sched.reassign_agent_jobs(old_agent, new_agent);
+
+        // Errors should be reset
+        let meta = sched.get_meta(id).unwrap();
+        assert_eq!(meta.consecutive_errors, 0);
+        assert_eq!(meta.job.agent_id, new_agent);
+    }
+
+    #[test]
+    fn test_reassign_agent_jobs_reenables_disabled_stale_jobs() {
+        let (sched, _tmp) = make_scheduler(100);
+        let old_agent = AgentId::new();
+        let new_agent = AgentId::new();
+
+        let job = make_job(old_agent);
+        let id = sched.add_job(job, false).unwrap();
+
+        // Simulate enough failures to auto-disable (with "not found" message)
+        for _ in 0..MAX_CONSECUTIVE_ERRORS {
+            sched.record_failure(id, "No such agent");
+        }
+        let meta = sched.get_meta(id).unwrap();
+        assert!(!meta.job.enabled, "Job should be auto-disabled");
+
+        // Reassign should re-enable it
+        sched.reassign_agent_jobs(old_agent, new_agent);
+
+        let meta = sched.get_meta(id).unwrap();
+        assert!(meta.job.enabled, "Job should be re-enabled after reassignment");
+        assert_eq!(meta.consecutive_errors, 0);
+        assert_eq!(meta.job.agent_id, new_agent);
+    }
+
+    #[test]
+    fn test_reassign_agent_jobs_persists_after_roundtrip() {
+        let tmp = tempfile::tempdir().unwrap();
+        let old_agent = AgentId::new();
+        let new_agent = AgentId::new();
+
+        // Create scheduler, add job, reassign, persist
+        let id = {
+            let sched = CronScheduler::new(tmp.path(), 100);
+            let job = make_job(old_agent);
+            let id = sched.add_job(job, false).unwrap();
+
+            sched.reassign_agent_jobs(old_agent, new_agent);
+            sched.persist().unwrap();
+            id
+        };
+
+        // Load from disk and verify the agent_id was persisted
+        {
+            let sched = CronScheduler::new(tmp.path(), 100);
+            sched.load().unwrap();
+
+            let job = sched.get_job(id).unwrap();
+            assert_eq!(job.agent_id, new_agent);
+            assert!(sched.list_jobs(old_agent).is_empty());
+        }
+    }
 }

crates/openfang-kernel/src/kernel.rs

@@ -50,7 +50,8 @@ impl LlmDriver for StubDriver {
     async fn complete(&self, _request: CompletionRequest) -> Result<CompletionResponse, LlmError> {
         Err(LlmError::MissingApiKey(
             "No LLM provider configured. Set an API key (e.g. GROQ_API_KEY) and restart, \
-             or configure a provider via the dashboard."
+             configure a provider via the dashboard, \
+             or use Ollama for local models (no API key needed)."
                 .to_string(),
         ))
     }
@@ -545,10 +546,20 @@ impl OpenFangKernel {
                 .map_err(|e| KernelError::BootFailed(format!("Memory init failed: {e}")))?,
         );
 
-        // Create LLM driver
+        // Create LLM driver.
+        // For the API key, try: 1) explicit api_key_env from config, 2) provider_api_keys
+        // mapping, 3) auth profiles, 4) convention {PROVIDER}_API_KEY. This ensures
+        // custom providers (e.g. nvidia, azure) work without hardcoded env var names.
+        let default_api_key = if !config.default_model.api_key_env.is_empty() {
+            std::env::var(&config.default_model.api_key_env).ok()
+        } else {
+            // api_key_env not set — resolve using provider_api_keys / convention
+            let env_var = config.resolve_api_key_env(&config.default_model.provider);
+            std::env::var(&env_var).ok()
+        };
         let driver_config = DriverConfig {
             provider: config.default_model.provider.clone(),
-            api_key: std::env::var(&config.default_model.api_key_env).ok(),
+            api_key: default_api_key,
             base_url: config
                 .default_model
                 .base_url
@@ -566,20 +577,54 @@ impl OpenFangKernel {
                 warn!(
                     provider = %config.default_model.provider,
                     error = %e,
-                    "Primary LLM driver init failed — dashboard will still be accessible"
+                    "Primary LLM driver init failed — trying auto-detect"
                 );
+                // Auto-detect: scan env for any configured provider key
+                if let Some((provider, model, env_var)) = drivers::detect_available_provider() {
+                    let auto_config = DriverConfig {
+                        provider: provider.to_string(),
+                        api_key: std::env::var(env_var).ok(),
+                        base_url: config.provider_urls.get(provider).cloned(),
+                    };
+                    match drivers::create_driver(&auto_config) {
+                        Ok(d) => {
+                            info!(
+                                provider = %provider,
+                                model = %model,
+                                "Auto-detected provider from {} — using as default",
+                                env_var
+                            );
+                            driver_chain.push(d);
+                            // Update the running config so agents get the right model
+                            config.default_model.provider = provider.to_string();
+                            config.default_model.model = model.to_string();
+                            config.default_model.api_key_env = env_var.to_string();
+                        }
+                        Err(e2) => {
+                            warn!(provider = %provider, error = %e2, "Auto-detected provider also failed");
+                        }
+                    }
+                }
             }
         }
 
-        // Add fallback providers to the chain
+        // Add fallback providers to the chain (with model names for cross-provider fallback)
+        let mut model_chain: Vec<(Arc<dyn LlmDriver>, String)> = Vec::new();
+        // Primary driver uses empty model name (uses the request's model field as-is)
+        for d in &driver_chain {
+            model_chain.push((d.clone(), String::new()));
+        }
         for fb in &config.fallback_providers {
+            let fb_api_key = if !fb.api_key_env.is_empty() {
+                std::env::var(&fb.api_key_env).ok()
+            } else {
+                // Resolve using provider_api_keys / convention for custom providers
+                let env_var = config.resolve_api_key_env(&fb.provider);
+                std::env::var(&env_var).ok()
+            };
             let fb_config = DriverConfig {
                 provider: fb.provider.clone(),
-                api_key: if fb.api_key_env.is_empty() {
-                    None
-                } else {
-                    std::env::var(&fb.api_key_env).ok()
-                },
+                api_key: fb_api_key,
                 base_url: fb
                     .base_url
                     .clone()
@@ -592,7 +637,8 @@ impl OpenFangKernel {
                         model = %fb.model,
                         "Fallback provider configured"
                     );
-                    driver_chain.push(d);
+                    driver_chain.push(d.clone());
+                    model_chain.push((d, fb.model.clone()));
                 }
                 Err(e) => {
                     warn!(
@@ -606,8 +652,8 @@ impl OpenFangKernel {
 
         // Use the chain, or create a stub driver if everything failed
         let driver: Arc<dyn LlmDriver> = if driver_chain.len() > 1 {
-            Arc::new(openfang_runtime::drivers::fallback::FallbackDriver::new(
-                driver_chain,
+            Arc::new(openfang_runtime::drivers::fallback::FallbackDriver::with_models(
+                model_chain,
             ))
         } else if let Some(single) = driver_chain.into_iter().next() {
             single
@@ -754,12 +800,14 @@ impl OpenFangKernel {
             Arc<dyn openfang_runtime::embedding::EmbeddingDriver + Send + Sync>,
         > = {
             use openfang_runtime::embedding::create_embedding_driver;
+            let configured_model = &config.memory.embedding_model;
             if let Some(ref provider) = config.memory.embedding_provider {
-                // Explicit config takes priority
+                // Explicit config takes priority — use the configured embedding model
                 let api_key_env = config.memory.embedding_api_key_env.as_deref().unwrap_or("");
-                match create_embedding_driver(provider, "text-embedding-3-small", api_key_env) {
+                let custom_url = config.provider_urls.get(provider.as_str()).map(|s| s.as_str());
+                match create_embedding_driver(provider, configured_model, api_key_env, custom_url) {
                     Ok(d) => {
-                        info!(provider = %provider, "Embedding driver configured from memory config");
+                        info!(provider = %provider, model = %configured_model, "Embedding driver configured from memory config");
                         Some(Arc::from(d))
                     }
                     Err(e) => {
@@ -768,8 +816,13 @@ impl OpenFangKernel {
                     }
                 }
             } else if std::env::var("OPENAI_API_KEY").is_ok() {
-                match create_embedding_driver("openai", "text-embedding-3-small", "OPENAI_API_KEY")
-                {
+                let model = if configured_model == "all-MiniLM-L6-v2" {
+                    "text-embedding-3-small"
+                } else {
+                    configured_model.as_str()
+                };
+                let openai_url = config.provider_urls.get("openai").map(|s| s.as_str());
+                match create_embedding_driver("openai", model, "OPENAI_API_KEY", openai_url) {
                     Ok(d) => {
                         info!("Embedding driver auto-detected: OpenAI");
                         Some(Arc::from(d))
@@ -781,7 +834,13 @@ impl OpenFangKernel {
                 }
             } else {
                 // Try Ollama (local, no key needed)
-                match create_embedding_driver("ollama", "nomic-embed-text", "") {
+                let model = if configured_model == "all-MiniLM-L6-v2" {
+                    "nomic-embed-text"
+                } else {
+                    configured_model.as_str()
+                };
+                let ollama_url = config.provider_urls.get("ollama").map(|s| s.as_str());
+                match create_embedding_driver("ollama", model, "", ollama_url) {
                     Ok(d) => {
                         info!("Embedding driver auto-detected: Ollama (local)");
                         Some(Arc::from(d))
@@ -889,7 +948,7 @@ impl OpenFangKernel {
             workflows: WorkflowEngine::new(),
             triggers: TriggerEngine::new(),
             background,
-            audit_log: Arc::new(AuditLog::new()),
+            audit_log: Arc::new(AuditLog::with_db(memory.usage_conn())),
             metering,
             default_driver: driver,
             wasm_sandbox,
@@ -1016,24 +1075,31 @@ impl OpenFangKernel {
                         &mut restored_entry.manifest.resources,
                     );
 
-                    // Apply default_model to restored agents (same logic as spawn)
+                    // Apply default_model to restored agents.
+                    //
+                    // Two cases:
+                    // 1. Agent has empty/default provider → always apply default_model
+                    // 2. Agent named "assistant" (auto-spawned) → update to match
+                    //    default_model so config.toml changes take effect on restart
                     {
+                        let dm = &kernel.config.default_model;
                         let is_default_provider = restored_entry.manifest.model.provider.is_empty()
                             || restored_entry.manifest.model.provider == "default";
                         let is_default_model = restored_entry.manifest.model.model.is_empty()
                             || restored_entry.manifest.model.model == "default";
-                        if is_default_provider && is_default_model {
-                            let dm = &kernel.config.default_model;
+                        let is_auto_spawned = restored_entry.name == "assistant"
+                            && restored_entry.manifest.description == "General-purpose assistant";
+                        if is_default_provider && is_default_model || is_auto_spawned {
                             if !dm.provider.is_empty() {
                                 restored_entry.manifest.model.provider = dm.provider.clone();
                             }
                             if !dm.model.is_empty() {
                                 restored_entry.manifest.model.model = dm.model.clone();
                             }
-                            if !dm.api_key_env.is_empty() && restored_entry.manifest.model.api_key_env.is_none() {
+                            if !dm.api_key_env.is_empty() {
                                 restored_entry.manifest.model.api_key_env = Some(dm.api_key_env.clone());
                             }
-                            if dm.base_url.is_some() && restored_entry.manifest.model.base_url.is_none() {
+                            if dm.base_url.is_some() {
                                 restored_entry.manifest.model.base_url.clone_from(&dm.base_url);
                             }
                         }
@@ -1299,12 +1365,47 @@ impl OpenFangKernel {
             .await
     }
 
+    /// Send a multimodal message (text + images) to an agent and get a response.
+    ///
+    /// Used by channel bridges when a user sends a photo — the image is downloaded,
+    /// base64 encoded, and passed as `ContentBlock::Image` alongside any caption text.
+    pub async fn send_message_with_blocks(
+        &self,
+        agent_id: AgentId,
+        message: &str,
+        blocks: Vec<openfang_types::message::ContentBlock>,
+    ) -> KernelResult<AgentLoopResult> {
+        let handle: Option<Arc<dyn KernelHandle>> = self
+            .self_handle
+            .get()
+            .and_then(|w| w.upgrade())
+            .map(|arc| arc as Arc<dyn KernelHandle>);
+        self.send_message_with_handle_and_blocks(agent_id, message, handle, Some(blocks))
+            .await
+    }
+
     /// Send a message with an optional kernel handle for inter-agent tools.
     pub async fn send_message_with_handle(
         &self,
         agent_id: AgentId,
         message: &str,
         kernel_handle: Option<Arc<dyn KernelHandle>>,
+    ) -> KernelResult<AgentLoopResult> {
+        self.send_message_with_handle_and_blocks(agent_id, message, kernel_handle, None)
+            .await
+    }
+
+    /// Send a message with optional content blocks and an optional kernel handle.
+    ///
+    /// When `content_blocks` is `Some`, the LLM agent loop receives structured
+    /// multimodal content (text + images) instead of just a text string. This
+    /// enables vision models to process images sent from channels like Telegram.
+    pub async fn send_message_with_handle_and_blocks(
+        &self,
+        agent_id: AgentId,
+        message: &str,
+        kernel_handle: Option<Arc<dyn KernelHandle>>,
+        content_blocks: Option<Vec<openfang_types::message::ContentBlock>>,
     ) -> KernelResult<AgentLoopResult> {
         // Enforce quota before running the agent loop
         self.scheduler
@@ -1323,7 +1424,7 @@ impl OpenFangKernel {
             self.execute_python_agent(&entry, agent_id, message).await
         } else {
             // Default: LLM agent loop (builtin:chat or any unrecognized module)
-            self.execute_llm_agent(&entry, agent_id, message, kernel_handle)
+            self.execute_llm_agent(&entry, agent_id, message, kernel_handle, content_blocks)
                 .await
         };
 
@@ -1711,6 +1812,7 @@ impl OpenFangKernel {
                 Some(&kernel_clone.hooks),
                 ctx_window,
                 Some(&kernel_clone.process_manager),
+                None, // content_blocks (streaming path uses text only for now)
             )
             .await;
 
@@ -1930,6 +2032,7 @@ impl OpenFangKernel {
         agent_id: AgentId,
         message: &str,
         kernel_handle: Option<Arc<dyn KernelHandle>>,
+        content_blocks: Option<Vec<openfang_types::message::ContentBlock>>,
     ) -> KernelResult<AgentLoopResult> {
         // Check metering quota before starting
         self.metering
@@ -2191,6 +2294,7 @@ impl OpenFangKernel {
             Some(&self.hooks),
             ctx_window,
             Some(&self.process_manager),
+            content_blocks,
         )
         .await
         .map_err(KernelError::OpenFang)?;
@@ -2292,6 +2396,9 @@ impl OpenFangKernel {
             .update_session_id(agent_id, new_session.id)
             .map_err(KernelError::OpenFang)?;
 
+        // Reset quota tracking so /new clears "token quota exceeded"
+        self.scheduler.reset_usage(agent_id);
+
         info!(agent_id = %agent_id, "Session reset (summary saved to memory)");
         Ok(())
     }
@@ -2461,7 +2568,7 @@ impl OpenFangKernel {
                 .take(5)
                 .enumerate()
                 .map(|(i, t)| {
-                    let truncated = if t.len() > 200 { &t[..200] } else { t };
+                    let truncated = openfang_types::truncate_str(t, 200);
                     format!("{}. {}", i + 1, truncated)
                 })
                 .collect::<Vec<_>>()
@@ -2961,6 +3068,7 @@ impl OpenFangKernel {
 
         // If an agent with this hand's name already exists, remove it first
         let existing = self.registry.list().into_iter().find(|e| e.name == def.agent.name);
+        let old_agent_id = existing.as_ref().map(|e| e.id);
         if let Some(old) = existing {
             info!(agent = %old.name, id = %old.id, "Removing existing hand agent for reactivation");
             let _ = self.kill_agent(old.id);
@@ -2969,6 +3077,18 @@ impl OpenFangKernel {
         // Spawn the agent
         let agent_id = self.spawn_agent(manifest)?;
 
+        // Migrate cron jobs from old agent to new agent so they survive restarts.
+        // Without this, persisted cron jobs would reference the stale old UUID
+        // and fail silently (issue #461).
+        if let Some(old_id) = old_agent_id {
+            let migrated = self.cron_scheduler.reassign_agent_jobs(old_id, agent_id);
+            if migrated > 0 {
+                if let Err(e) = self.cron_scheduler.persist() {
+                    warn!("Failed to persist cron jobs after agent migration: {e}");
+                }
+            }
+        }
+
         // Link agent to instance
         self.hand_registry
             .set_agent(instance.instance_id, agent_id)
@@ -2981,6 +3101,9 @@ impl OpenFangKernel {
             "Hand activated with agent"
         );
 
+        // Persist hand state so it survives restarts
+        self.persist_hand_state();
+
         // Return instance with agent set
         Ok(self
             .hand_registry
@@ -3012,9 +3135,19 @@ impl OpenFangKernel {
                 }
             }
         }
+        // Persist hand state so it survives restarts
+        self.persist_hand_state();
         Ok(())
     }
 
+    /// Persist active hand state to disk.
+    fn persist_hand_state(&self) {
+        let state_path = self.config.home_dir.join("hand_state.json");
+        if let Err(e) = self.hand_registry.persist_state(&state_path) {
+            warn!(error = %e, "Failed to persist hand state");
+        }
+    }
+
     /// Pause a hand (marks it paused; agent stays alive but won't receive new work).
     pub fn pause_hand(&self, instance_id: uuid::Uuid) -> KernelResult<()> {
         self.hand_registry
@@ -3291,6 +3424,44 @@ impl OpenFangKernel {
     /// Iterates the agent registry and starts background tasks for agents with
     /// `Continuous`, `Periodic`, or `Proactive` schedules.
     pub fn start_background_agents(self: &Arc<Self>) {
+        // Restore previously active hands from persisted state
+        let state_path = self.config.home_dir.join("hand_state.json");
+        let saved_hands = openfang_hands::registry::HandRegistry::load_state(&state_path);
+        if !saved_hands.is_empty() {
+            info!("Restoring {} persisted hand(s)", saved_hands.len());
+            for (hand_id, config, old_agent_id) in saved_hands {
+                match self.activate_hand(&hand_id, config) {
+                    Ok(inst) => {
+                        info!(hand = %hand_id, instance = %inst.instance_id, "Hand restored");
+                        // Reassign cron jobs from the pre-restart agent ID to the
+                        // newly spawned agent so scheduled tasks survive daemon
+                        // restarts (issue #402). activate_hand only handles
+                        // reassignment when an existing agent is found in the live
+                        // registry, which is empty on a fresh boot.
+                        if let (Some(old_id), Some(new_id)) = (old_agent_id, inst.agent_id) {
+                            if old_id != new_id {
+                                let migrated =
+                                    self.cron_scheduler.reassign_agent_jobs(old_id, new_id);
+                                if migrated > 0 {
+                                    info!(
+                                        hand = %hand_id,
+                                        old_agent = %old_id,
+                                        new_agent = %new_id,
+                                        migrated,
+                                        "Reassigned cron jobs after restart"
+                                    );
+                                    if let Err(e) = self.cron_scheduler.persist() {
+                                        warn!("Failed to persist cron jobs after hand restore: {e}");
+                                    }
+                                }
+                            }
+                        }
+                    }
+                    Err(e) => warn!(hand = %hand_id, error = %e, "Failed to restore hand"),
+                }
+            }
+        }
+
         let agents = self.registry.list();
         let mut bg_agents: Vec<(openfang_types::agent::AgentId, String, ScheduleMode)> =
             Vec::new();
@@ -3829,25 +4000,36 @@ impl OpenFangKernel {
 
     /// Resolve the LLM driver for an agent.
     ///
-    /// If the agent's manifest specifies a different provider than the kernel default,
-    /// a dedicated driver is created. Otherwise the kernel's default driver is reused.
+    /// Always creates a fresh driver using current environment variables so that
+    /// API keys saved via the dashboard (`set_provider_key`) take effect immediately
+    /// without requiring a daemon restart. Uses the hot-reloaded default model
+    /// override when available.
     /// If fallback models are configured, wraps the primary in a `FallbackDriver`.
     fn resolve_driver(&self, manifest: &AgentManifest) -> KernelResult<Arc<dyn LlmDriver>> {
         let agent_provider = &manifest.model.provider;
-        let default_provider = &self.config.default_model.provider;
 
-        // If agent uses same provider as kernel default and has no custom overrides, reuse
+        // Use the effective default model: hot-reloaded override takes priority
+        // over the boot-time config. This ensures that when a user saves a new
+        // API key via the dashboard and the default provider is switched,
+        // resolve_driver sees the updated provider/model/api_key_env.
+        let override_guard = self
+            .default_model_override
+            .read()
+            .unwrap_or_else(|e: std::sync::PoisonError<_>| e.into_inner());
+        let effective_default = override_guard
+            .as_ref()
+            .unwrap_or(&self.config.default_model);
+        let default_provider = &effective_default.provider;
+
         let has_custom_key = manifest.model.api_key_env.is_some();
         let has_custom_url = manifest.model.base_url.is_some();
 
-        let primary = if agent_provider == default_provider && !has_custom_key && !has_custom_url {
-            Arc::clone(&self.default_driver)
-        } else {
-            // Create a dedicated driver for this agent.
-            //
-            // IMPORTANT: When the agent's provider differs from the default,
-            // we must NOT pass the default provider's API key. Instead, pass None
-            // so create_driver() can look up the correct env var for the target provider.
+        // Always create a fresh driver by reading current env vars.
+        // This ensures API keys saved at runtime (via dashboard POST
+        // /api/providers/{name}/key which calls std::env::set_var) are
+        // picked up immediately — the boot-time default_driver cache is
+        // only used as a final fallback when driver creation fails.
+        let primary = {
             let api_key = if has_custom_key {
                 // Agent explicitly set an API key env var — use it
                 manifest
@@ -3856,29 +4038,26 @@ impl OpenFangKernel {
                     .as_ref()
                     .and_then(|env| std::env::var(env).ok())
             } else if agent_provider == default_provider {
-                // Same provider — use default key
-                std::env::var(&self.config.default_model.api_key_env).ok()
-            } else {
-                // Different provider — check auth profiles first, then let
-                // create_driver() look up the correct env var automatically.
-                if let Some(profiles) = self.config.auth_profiles.get(agent_provider.as_str()) {
-                    let mut sorted: Vec<_> = profiles.iter().collect();
-                    sorted.sort_by_key(|p| p.priority);
-                    sorted
-                        .first()
-                        .and_then(|best| std::env::var(&best.api_key_env).ok())
+                // Same provider as effective default — use its env var
+                if !effective_default.api_key_env.is_empty() {
+                    std::env::var(&effective_default.api_key_env).ok()
                 } else {
-                    // Pass None — create_driver() has per-provider env var lookups
-                    None
+                    let env_var = self.config.resolve_api_key_env(agent_provider);
+                    std::env::var(&env_var).ok()
                 }
+            } else {
+                // Different provider — check auth profiles, provider_api_keys,
+                // and convention-based env var. For custom providers (not in the
+                // hardcoded list), this is the primary path for API key resolution.
+                let env_var = self.config.resolve_api_key_env(agent_provider);
+                std::env::var(&env_var).ok()
             };
 
             // Don't inherit default provider's base_url when switching providers
             let base_url = if has_custom_url {
                 manifest.model.base_url.clone()
             } else if agent_provider == default_provider {
-                self.config
-                    .default_model
+                effective_default
                     .base_url
                     .clone()
                     .or_else(|| self.config.provider_urls.get(agent_provider.as_str()).cloned())
@@ -3893,9 +4072,27 @@ impl OpenFangKernel {
                 base_url,
             };
 
-            drivers::create_driver(&driver_config).map_err(|e| {
-                KernelError::BootFailed(format!("Agent LLM driver init failed: {e}"))
-            })?
+            match drivers::create_driver(&driver_config) {
+                Ok(d) => d,
+                Err(e) => {
+                    // If fresh driver creation fails (e.g. key not yet set for this
+                    // provider), fall back to the boot-time default driver. This
+                    // keeps existing agents working while the user is still
+                    // configuring providers via the dashboard.
+                    if agent_provider == default_provider && !has_custom_key && !has_custom_url {
+                        debug!(
+                            provider = %agent_provider,
+                            error = %e,
+                            "Fresh driver creation failed, falling back to boot-time default"
+                        );
+                        Arc::clone(&self.default_driver)
+                    } else {
+                        return Err(KernelError::BootFailed(format!(
+                            "Agent LLM driver init failed: {e}"
+                        )));
+                    }
+                }
+            }
         };
 
         // If fallback models are configured, wrap in FallbackDriver
@@ -3904,12 +4101,16 @@ impl OpenFangKernel {
             let mut chain: Vec<(std::sync::Arc<dyn openfang_runtime::llm_driver::LlmDriver>, String)> =
                 vec![(primary.clone(), String::new())];
             for fb in &manifest.fallback_models {
+                let fb_api_key = if let Some(env) = &fb.api_key_env {
+                    std::env::var(env).ok()
+                } else {
+                    // Resolve using provider_api_keys / convention for custom providers
+                    let env_var = self.config.resolve_api_key_env(&fb.provider);
+                    std::env::var(&env_var).ok()
+                };
                 let config = DriverConfig {
                     provider: fb.provider.clone(),
-                    api_key: fb
-                        .api_key_env
-                        .as_ref()
-                        .and_then(|env| std::env::var(env).ok()),
+                    api_key: fb_api_key,
                     base_url: fb
                         .base_url
                         .clone()
@@ -4281,6 +4482,10 @@ impl OpenFangKernel {
             _ => all_builtins,
         };
 
+        // Track tool names added by skills/MCP — these already passed their own
+        // allowlist filters and should bypass the per-agent capability check.
+        let mut extension_tool_names = std::collections::HashSet::new();
+
         // Add skill-provided tools (filtered by agent's skill allowlist)
         let skill_tools = {
             let registry = self
@@ -4294,6 +4499,7 @@ impl OpenFangKernel {
             }
         };
         for skill_tool in skill_tools {
+            extension_tool_names.insert(skill_tool.name.clone());
             all_tools.push(ToolDefinition {
                 name: skill_tool.name.clone(),
                 description: skill_tool.description.clone(),
@@ -4304,6 +4510,9 @@ impl OpenFangKernel {
         // Add MCP tools (filtered by agent's MCP server allowlist)
         if let Ok(mcp_tools) = self.mcp_tools.lock() {
             if mcp_allowlist.is_empty() {
+                for t in mcp_tools.iter() {
+                    extension_tool_names.insert(t.name.clone());
+                }
                 all_tools.extend(mcp_tools.iter().cloned());
             } else {
                 // Normalize allowlist names for matching
@@ -4311,16 +4520,19 @@ impl OpenFangKernel {
                     .iter()
                     .map(|s| openfang_runtime::mcp::normalize_name(s))
                     .collect();
-                all_tools.extend(
-                    mcp_tools
-                        .iter()
-                        .filter(|t| {
-                            openfang_runtime::mcp::extract_mcp_server(&t.name)
-                                .map(|s| normalized.iter().any(|n| n == s))
-                                .unwrap_or(false)
-                        })
-                        .cloned(),
-                );
+                let filtered: Vec<_> = mcp_tools
+                    .iter()
+                    .filter(|t| {
+                        openfang_runtime::mcp::extract_mcp_server(&t.name)
+                            .map(|s| normalized.iter().any(|n| n == s))
+                            .unwrap_or(false)
+                    })
+                    .cloned()
+                    .collect();
+                for t in &filtered {
+                    extension_tool_names.insert(t.name.clone());
+                }
+                all_tools.extend(filtered);
             }
         }
 
@@ -4356,14 +4568,17 @@ impl OpenFangKernel {
             return all_tools;
         }
 
-        // Filter to tools the agent has capability for
+        // Filter to tools the agent has capability for.
+        // MCP and skill tools bypass this check — they already passed their
+        // own allowlist filters above (mcp_servers / skills on the manifest).
         all_tools
             .into_iter()
             .filter(|tool| {
-                caps.iter().any(|c| match c {
-                    Capability::ToolInvoke(name) => name == &tool.name || name == "*",
-                    _ => false,
-                })
+                extension_tool_names.contains(&tool.name)
+                    || caps.iter().any(|c| match c {
+                        Capability::ToolInvoke(name) => name == &tool.name || name == "*",
+                        _ => false,
+                    })
             })
             .collect()
     }
@@ -4632,8 +4847,8 @@ fn apply_budget_defaults(
     budget: &openfang_types::config::BudgetConfig,
     resources: &mut ResourceQuota,
 ) {
-    // Only override hourly if agent has the built-in default (1.0) and global is set
-    if budget.max_hourly_usd > 0.0 && resources.max_cost_per_hour_usd == 1.0 {
+    // Only override hourly if agent has unlimited (0.0) and global is set
+    if budget.max_hourly_usd > 0.0 && resources.max_cost_per_hour_usd == 0.0 {
         resources.max_cost_per_hour_usd = budget.max_hourly_usd;
     }
     // Only override daily/monthly if agent has unlimited (0.0) and global is set
@@ -4651,17 +4866,27 @@ fn apply_budget_defaults(
 /// This is a defense-in-depth fallback — models should ideally be in the catalog.
 fn infer_provider_from_model(model: &str) -> Option<String> {
     let lower = model.to_lowercase();
-    // Check for explicit provider prefix (e.g., "minimax/MiniMax-M2.5")
-    if let Some(prefix) = lower.split('/').next() {
+    // Check for explicit provider prefix with / or : delimiter
+    // (e.g., "minimax/MiniMax-M2.5" or "qwen:qwen-plus")
+    let (prefix, has_delim) = if let Some(idx) = lower.find('/') {
+        (&lower[..idx], true)
+    } else if let Some(idx) = lower.find(':') {
+        (&lower[..idx], true)
+    } else {
+        (lower.as_str(), false)
+    };
+    if has_delim {
         match prefix {
             "minimax" | "gemini" | "anthropic" | "openai" | "groq" | "deepseek" | "mistral"
             | "cohere" | "xai" | "ollama" | "together" | "fireworks" | "perplexity"
             | "cerebras" | "sambanova" | "replicate" | "huggingface" | "ai21" | "codex"
-            | "claude-code" | "copilot" | "github-copilot" | "qwen" | "zhipu" | "moonshot"
-            | "openrouter" => {
-                if model.contains('/') {
-                    return Some(prefix.to_string());
-                }
+            | "claude-code" | "copilot" | "github-copilot" | "qwen" | "zhipu" | "zai" | "moonshot"
+            | "openrouter" | "volcengine" | "doubao" | "dashscope" => {
+                return Some(prefix.to_string());
+            }
+            // "kimi" is a brand alias for moonshot
+            "kimi" => {
+                return Some("moonshot".to_string());
             }
             _ => {}
         }
@@ -4696,6 +4921,8 @@ fn infer_provider_from_model(model: &str) -> Option<String> {
         Some("qianfan".to_string())
     } else if lower.starts_with("abab") {
         Some("minimax".to_string())
+    } else if lower.starts_with("moonshot") || lower.starts_with("kimi") {
+        Some("moonshot".to_string())
     } else {
         None
     }
@@ -5205,7 +5432,7 @@ impl KernelHandle for OpenFangKernel {
             .unwrap_or_else(|e| e.into_inner());
         agents
             .iter()
-            .map(|(url, card)| (card.name.clone(), url.clone()))
+            .map(|(_, card)| (card.name.clone(), card.url.clone()))
             .collect()
     }
 
@@ -5218,7 +5445,7 @@ impl KernelHandle for OpenFangKernel {
         agents
             .iter()
             .find(|(_, card)| card.name.to_lowercase() == name_lower)
-            .map(|(url, _)| url.clone())
+            .map(|(_, card)| card.url.clone())
     }
 
     async fn send_channel_message(
@@ -5257,6 +5484,59 @@ impl KernelHandle for OpenFangKernel {
         Ok(format!("Message sent to {} via {}", recipient, channel))
     }
 
+    async fn send_channel_media(
+        &self,
+        channel: &str,
+        recipient: &str,
+        media_type: &str,
+        media_url: &str,
+        caption: Option<&str>,
+        filename: Option<&str>,
+    ) -> Result<String, String> {
+        let adapter = self
+            .channel_adapters
+            .get(channel)
+            .ok_or_else(|| {
+                let available: Vec<String> = self
+                    .channel_adapters
+                    .iter()
+                    .map(|e| e.key().clone())
+                    .collect();
+                format!(
+                    "Channel '{}' not found. Available channels: {:?}",
+                    channel, available
+                )
+            })?
+            .clone();
+
+        let user = openfang_channels::types::ChannelUser {
+            platform_id: recipient.to_string(),
+            display_name: recipient.to_string(),
+            openfang_user: None,
+        };
+
+        let content = match media_type {
+            "image" => openfang_channels::types::ChannelContent::Image {
+                url: media_url.to_string(),
+                caption: caption.map(|s| s.to_string()),
+            },
+            "file" => openfang_channels::types::ChannelContent::File {
+                url: media_url.to_string(),
+                filename: filename.unwrap_or("file").to_string(),
+            },
+            _ => {
+                return Err(format!("Unsupported media type: '{media_type}'. Use 'image' or 'file'."));
+            }
+        };
+
+        adapter
+            .send(&user, content)
+            .await
+            .map_err(|e| format!("Channel media send failed: {e}"))?;
+
+        Ok(format!("{} sent to {} via {}", media_type, recipient, channel))
+    }
+
     async fn spawn_agent_checked(
         &self,
         manifest_toml: &str,

crates/openfang-kernel/src/metering.rs

@@ -343,6 +343,11 @@ fn estimate_cost_rates(model: &str) -> (f64, f64) {
         return (0.40, 0.40);
     }
 
+    // ── Venice.ai ──────────────────────────────────────────────
+    if model.contains("venice") {
+        return (0.20, 0.90);
+    }
+
     // ── Open-source (Groq, Together, etc.) ─────────────────────
     if model.contains("llama-4-maverick") {
         return (0.50, 0.77);
@@ -376,11 +381,20 @@ fn estimate_cost_rates(model: &str) -> (f64, f64) {
     }
 
     // ── Zhipu / GLM ─────────────────────────────────────────────
-    if model.contains("glm-4-flash") {
-        return (0.10, 0.10);
+    if model.contains("glm-5") {
+        return (1.00, 3.20);
+    }
+    if model.contains("glm-4.7") {
+        return (0.60, 2.20);
+    }
+    if model.contains("glm-4-flash") || model.contains("glm-4.5-flash") {
+        return (0.0, 0.0); // free tier
+    }
+    if model.contains("glm-4.5") {
+        return (0.60, 2.20);
     }
     if model.contains("glm") {
-        return (1.50, 5.00);
+        return (0.60, 2.20);
     }
     if model.contains("codegeex") {
         return (0.10, 0.10);
@@ -391,6 +405,20 @@ fn estimate_cost_rates(model: &str) -> (f64, f64) {
         return (0.80, 0.80);
     }
 
+    // ── Volcano Engine / Doubao ────────────────────────────────
+    if model.contains("doubao-seed-code") {
+        return (0.50, 1.00);
+    }
+    if model.contains("doubao") && model.contains("mini") {
+        return (0.10, 0.10);
+    }
+    if model.contains("doubao") && model.contains("lite") {
+        return (0.30, 0.60);
+    }
+    if model.contains("doubao") {
+        return (0.80, 2.00);
+    }
+
     // ── Baidu ERNIE ─────────────────────────────────────────────
     if model.contains("ernie") {
         return (2.00, 6.00);

crates/openfang-kernel/src/registry.rs

@@ -177,6 +177,21 @@ impl AgentRegistry {
         Ok(())
     }
 
+    /// Update an agent's fallback model chain.
+    pub fn update_fallback_models(
+        &self,
+        id: AgentId,
+        fallback_models: Vec<openfang_types::agent::FallbackModel>,
+    ) -> OpenFangResult<()> {
+        let mut entry = self
+            .agents
+            .get_mut(&id)
+            .ok_or_else(|| OpenFangError::AgentNotFound(id.to_string()))?;
+        entry.manifest.fallback_models = fallback_models;
+        entry.last_active = chrono::Utc::now();
+        Ok(())
+    }
+
     /// Update an agent's skill allowlist.
     pub fn update_skills(&self, id: AgentId, skills: Vec<String>) -> OpenFangResult<()> {
         let mut entry = self

crates/openfang-kernel/src/scheduler.rs

@@ -100,6 +100,15 @@ impl AgentScheduler {
         Ok(())
     }
 
+    /// Reset usage tracking for an agent (e.g. on session reset).
+    pub fn reset_usage(&self, agent_id: AgentId) {
+        if let Some(mut tracker) = self.usage.get_mut(&agent_id) {
+            tracker.total_tokens = 0;
+            tracker.tool_calls = 0;
+            tracker.window_start = Instant::now();
+        }
+    }
+
     /// Abort an agent's active task.
     pub fn abort_task(&self, agent_id: AgentId) {
         if let Some((_, handle)) = self.tasks.remove(&agent_id) {

crates/openfang-kernel/src/wizard.rs

@@ -54,10 +54,11 @@ impl SetupWizard {
     /// model configuration, capabilities, and schedule.
     pub fn build_plan(intent: AgentIntent) -> SetupPlan {
         // Map model tier to provider/model
+        // Use "default" so the kernel applies config.toml's [default_model].
+        // Only "complex" tier gets an explicit Anthropic override.
         let (provider, model) = match intent.model_tier.as_str() {
-            "simple" => ("groq", "llama-3.3-70b-versatile"),
             "complex" => ("anthropic", "claude-sonnet-4-20250514"),
-            _ => ("groq", "llama-3.3-70b-versatile"), // medium default
+            _ => ("default", "default"),
         };
 
         // Build capabilities from intent
@@ -285,7 +286,7 @@ mod tests {
         let plan = SetupWizard::build_plan(intent);
 
         assert_eq!(plan.manifest.name, "research-bot");
-        assert_eq!(plan.manifest.model.provider, "groq");
+        assert_eq!(plan.manifest.model.provider, "default");
         assert!(plan
             .manifest
             .capabilities

crates/openfang-memory/src/migration.rs

@@ -5,7 +5,7 @@
 use rusqlite::Connection;
 
 /// Current schema version.
-const SCHEMA_VERSION: u32 = 7;
+const SCHEMA_VERSION: u32 = 8;
 
 /// Run all migrations to bring the database up to date.
 pub fn run_migrations(conn: &Connection) -> Result<(), rusqlite::Error> {
@@ -39,6 +39,10 @@ pub fn run_migrations(conn: &Connection) -> Result<(), rusqlite::Error> {
         migrate_v7(conn)?;
     }
 
+    if current_version < 8 {
+        migrate_v8(conn)?;
+    }
+
     set_schema_version(conn, SCHEMA_VERSION)?;
     Ok(())
 }
@@ -299,6 +303,31 @@ fn migrate_v7(conn: &Connection) -> Result<(), rusqlite::Error> {
     Ok(())
 }
 
+/// Version 8: Add audit_entries table for persistent Merkle audit trail.
+fn migrate_v8(conn: &Connection) -> Result<(), rusqlite::Error> {
+    conn.execute_batch(
+        "
+        CREATE TABLE IF NOT EXISTS audit_entries (
+            seq INTEGER PRIMARY KEY,
+            timestamp TEXT NOT NULL,
+            agent_id TEXT NOT NULL,
+            action TEXT NOT NULL,
+            detail TEXT NOT NULL,
+            outcome TEXT NOT NULL,
+            prev_hash TEXT NOT NULL,
+            hash TEXT NOT NULL
+        );
+        CREATE INDEX IF NOT EXISTS idx_audit_agent ON audit_entries(agent_id);
+        CREATE INDEX IF NOT EXISTS idx_audit_timestamp ON audit_entries(timestamp);
+        CREATE INDEX IF NOT EXISTS idx_audit_action ON audit_entries(action);
+
+        INSERT OR IGNORE INTO migrations (version, applied_at, description)
+        VALUES (8, datetime('now'), 'Add audit_entries table for persistent Merkle audit trail');
+        ",
+    )?;
+    Ok(())
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

crates/openfang-memory/src/structured.rs

@@ -130,11 +130,19 @@ impl StructuredStore {
             "ALTER TABLE agents ADD COLUMN session_id TEXT DEFAULT ''",
             [],
         );
+        // Add identity column (migration compat)
+        let _ = conn.execute(
+            "ALTER TABLE agents ADD COLUMN identity TEXT DEFAULT '{}'",
+            [],
+        );
+
+        let identity_json = serde_json::to_string(&entry.identity)
+            .map_err(|e| OpenFangError::Serialization(e.to_string()))?;
 
         conn.execute(
-            "INSERT INTO agents (id, name, manifest, state, created_at, updated_at, session_id)
-             VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7)
-             ON CONFLICT(id) DO UPDATE SET name = ?2, manifest = ?3, state = ?4, updated_at = ?6, session_id = ?7",
+            "INSERT INTO agents (id, name, manifest, state, created_at, updated_at, session_id, identity)
+             VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7, ?8)
+             ON CONFLICT(id) DO UPDATE SET name = ?2, manifest = ?3, state = ?4, updated_at = ?6, session_id = ?7, identity = ?8",
             rusqlite::params![
                 entry.id.0.to_string(),
                 entry.name,
@@ -143,6 +151,7 @@ impl StructuredStore {
                 entry.created_at.to_rfc3339(),
                 now,
                 entry.session_id.0.to_string(),
+                identity_json,
             ],
         )
         .map_err(|e| OpenFangError::Memory(e.to_string()))?;
@@ -157,10 +166,13 @@ impl StructuredStore {
             .map_err(|e| OpenFangError::Internal(e.to_string()))?;
 
         let mut stmt = conn
-            .prepare("SELECT id, name, manifest, state, created_at, updated_at, session_id FROM agents WHERE id = ?1")
+            .prepare("SELECT id, name, manifest, state, created_at, updated_at, session_id, identity FROM agents WHERE id = ?1")
             .or_else(|_| {
-                // Fallback without session_id column for old DBs
-                conn.prepare("SELECT id, name, manifest, state, created_at, updated_at FROM agents WHERE id = ?1")
+                conn.prepare("SELECT id, name, manifest, state, created_at, updated_at, session_id FROM agents WHERE id = ?1")
+                    .or_else(|_| {
+                        // Fallback without session_id column for old DBs
+                        conn.prepare("SELECT id, name, manifest, state, created_at, updated_at FROM agents WHERE id = ?1")
+                    })
             })
             .map_err(|e| OpenFangError::Memory(e.to_string()))?;
 
@@ -175,11 +187,16 @@ impl StructuredStore {
             } else {
                 None
             };
-            Ok((name, manifest_blob, state_str, created_str, session_id_str))
+            let identity_str: Option<String> = if col_count >= 8 {
+                row.get(7).ok()
+            } else {
+                None
+            };
+            Ok((name, manifest_blob, state_str, created_str, session_id_str, identity_str))
         });
 
         match result {
-            Ok((name, manifest_blob, state_str, created_str, session_id_str)) => {
+            Ok((name, manifest_blob, state_str, created_str, session_id_str, identity_str)) => {
                 let manifest = rmp_serde::from_slice(&manifest_blob)
                     .map_err(|e| OpenFangError::Serialization(e.to_string()))?;
                 let state = serde_json::from_str(&state_str)
@@ -191,6 +208,9 @@ impl StructuredStore {
                     .and_then(|s| uuid::Uuid::parse_str(&s).ok())
                     .map(openfang_types::agent::SessionId)
                     .unwrap_or_else(openfang_types::agent::SessionId::new);
+                let identity = identity_str
+                    .and_then(|s| serde_json::from_str(&s).ok())
+                    .unwrap_or_default();
                 Ok(Some(AgentEntry {
                     id: agent_id,
                     name,
@@ -203,7 +223,7 @@ impl StructuredStore {
                     children: vec![],
                     session_id,
                     tags: vec![],
-                    identity: Default::default(),
+                    identity,
                     onboarding_completed: false,
                     onboarding_completed_at: None,
                 }))
@@ -239,11 +259,14 @@ impl StructuredStore {
             .lock()
             .map_err(|e| OpenFangError::Internal(e.to_string()))?;
 
-        // Try with session_id column first, fall back without
+        // Try with identity+session_id columns first, fall back gracefully
         let mut stmt = conn
             .prepare(
-                "SELECT id, name, manifest, state, created_at, updated_at, session_id FROM agents",
+                "SELECT id, name, manifest, state, created_at, updated_at, session_id, identity FROM agents",
             )
+            .or_else(|_| {
+                conn.prepare("SELECT id, name, manifest, state, created_at, updated_at, session_id FROM agents")
+            })
             .or_else(|_| {
                 conn.prepare("SELECT id, name, manifest, state, created_at, updated_at FROM agents")
             })
@@ -262,6 +285,11 @@ impl StructuredStore {
                 } else {
                     None
                 };
+                let identity_str: Option<String> = if col_count >= 8 {
+                    row.get(7).ok()
+                } else {
+                    None
+                };
                 Ok((
                     id_str,
                     name,
@@ -269,6 +297,7 @@ impl StructuredStore {
                     state_str,
                     created_str,
                     session_id_str,
+                    identity_str,
                 ))
             })
             .map_err(|e| OpenFangError::Memory(e.to_string()))?;
@@ -278,7 +307,7 @@ impl StructuredStore {
         let mut repair_queue: Vec<(String, Vec<u8>, String)> = Vec::new();
 
         for row in rows {
-            let (id_str, name, manifest_blob, state_str, created_str, session_id_str) = match row {
+            let (id_str, name, manifest_blob, state_str, created_str, session_id_str, identity_str) = match row {
                 Ok(r) => r,
                 Err(e) => {
                     tracing::warn!("Skipping agent row with read error: {e}");
@@ -342,6 +371,10 @@ impl StructuredStore {
                 .map(openfang_types::agent::SessionId)
                 .unwrap_or_else(openfang_types::agent::SessionId::new);
 
+            let identity = identity_str
+                .and_then(|s| serde_json::from_str(&s).ok())
+                .unwrap_or_default();
+
             agents.push(AgentEntry {
                 id: agent_id,
                 name,
@@ -354,7 +387,7 @@ impl StructuredStore {
                 children: vec![],
                 session_id,
                 tags: vec![],
-                identity: Default::default(),
+                identity,
                 onboarding_completed: false,
                 onboarding_completed_at: None,
             });

crates/openfang-migrate/src/openclaw.rs

@@ -64,11 +64,11 @@ struct OpenClawModels {
 #[serde(default, rename_all = "camelCase")]
 struct OpenClawRootTools {
     #[allow(dead_code)]
-    profile: Option<String>,
+    profile: Option<serde_json::Value>,
     #[allow(dead_code)]
-    allow: Option<Vec<String>>,
+    allow: Option<serde_json::Value>,
     #[allow(dead_code)]
-    deny: Option<Vec<String>>,
+    deny: Option<serde_json::Value>,
 }
 
 #[derive(Debug, Default, Deserialize)]
@@ -110,17 +110,36 @@ struct OpenClawAgentEntry {
     model: Option<OpenClawAgentModel>,
     tools: Option<OpenClawAgentTools>,
     workspace: Option<String>,
-    skills: Option<Vec<String>>,
+    skills: Option<serde_json::Value>,
     identity: Option<String>,
 }
 
 #[derive(Debug, Default, Clone, Deserialize)]
 #[serde(default, rename_all = "camelCase")]
 struct OpenClawAgentTools {
-    profile: Option<String>,
-    allow: Option<Vec<String>>,
-    deny: Option<Vec<String>>,
-    also_allow: Option<Vec<String>>,
+    profile: Option<serde_json::Value>,
+    allow: Option<serde_json::Value>,
+    deny: Option<serde_json::Value>,
+    also_allow: Option<serde_json::Value>,
+}
+
+/// Extract a profile name from a Value (string or {name: "..."}  object).
+fn extract_profile(val: &serde_json::Value) -> Option<String> {
+    val.as_str()
+        .map(|s| s.to_string())
+        .or_else(|| val.get("name").and_then(|v| v.as_str()).map(|s| s.to_string()))
+}
+
+/// Extract a list of strings from a Value (array of strings, single string, or object keys).
+fn extract_string_list(val: &serde_json::Value) -> Vec<String> {
+    match val {
+        serde_json::Value::Array(arr) => {
+            arr.iter().filter_map(|v| v.as_str()).map(|s| s.to_string()).collect()
+        }
+        serde_json::Value::String(s) => vec![s.clone()],
+        serde_json::Value::Object(map) => map.keys().cloned().collect(),
+        _ => vec![],
+    }
 }
 
 #[derive(Debug, Default, Deserialize)]
@@ -149,7 +168,7 @@ struct OpenClawChannels {
 #[serde(default, rename_all = "camelCase")]
 struct OpenClawTelegramConfig {
     bot_token: Option<String>,
-    allow_from: Option<Vec<String>>,
+    allow_from: Option<serde_json::Value>,
     group_policy: Option<String>,
     dm_policy: Option<String>,
     enabled: Option<bool>,
@@ -162,7 +181,7 @@ struct OpenClawDiscordConfig {
     guilds: Option<serde_json::Value>,
     dm_policy: Option<String>,
     group_policy: Option<String>,
-    allow_from: Option<Vec<String>>,
+    allow_from: Option<serde_json::Value>,
     enabled: Option<bool>,
 }
 
@@ -173,7 +192,7 @@ struct OpenClawSlackConfig {
     app_token: Option<String>,
     dm_policy: Option<String>,
     group_policy: Option<String>,
-    allow_from: Option<Vec<String>>,
+    allow_from: Option<serde_json::Value>,
     enabled: Option<bool>,
 }
 
@@ -182,7 +201,7 @@ struct OpenClawSlackConfig {
 struct OpenClawWhatsAppConfig {
     auth_dir: Option<String>,
     dm_policy: Option<String>,
-    allow_from: Option<Vec<String>>,
+    allow_from: Option<serde_json::Value>,
     group_policy: Option<String>,
     enabled: Option<bool>,
 }
@@ -195,7 +214,7 @@ struct OpenClawSignalConfig {
     http_port: Option<u16>,
     account: Option<String>,
     dm_policy: Option<String>,
-    allow_from: Option<Vec<String>>,
+    allow_from: Option<serde_json::Value>,
     enabled: Option<bool>,
 }
 
@@ -205,9 +224,9 @@ struct OpenClawMatrixConfig {
     homeserver: Option<String>,
     user_id: Option<String>,
     access_token: Option<String>,
-    rooms: Option<Vec<String>>,
+    rooms: Option<serde_json::Value>,
     dm_policy: Option<String>,
-    allow_from: Option<Vec<String>>,
+    allow_from: Option<serde_json::Value>,
     enabled: Option<bool>,
 }
 
@@ -228,7 +247,7 @@ struct OpenClawTeamsConfig {
     app_password: Option<String>,
     tenant_id: Option<String>,
     dm_policy: Option<String>,
-    allow_from: Option<Vec<String>>,
+    allow_from: Option<serde_json::Value>,
     enabled: Option<bool>,
 }
 
@@ -240,9 +259,9 @@ struct OpenClawIrcConfig {
     tls: Option<bool>,
     nick: Option<String>,
     password: Option<String>,
-    channels: Option<Vec<String>>,
+    channels: Option<serde_json::Value>,
     dm_policy: Option<String>,
-    allow_from: Option<Vec<String>>,
+    allow_from: Option<serde_json::Value>,
     enabled: Option<bool>,
 }
 
@@ -252,7 +271,7 @@ struct OpenClawMattermostConfig {
     bot_token: Option<String>,
     base_url: Option<String>,
     dm_policy: Option<String>,
-    allow_from: Option<Vec<String>>,
+    allow_from: Option<serde_json::Value>,
     enabled: Option<bool>,
 }
 
@@ -272,7 +291,7 @@ struct OpenClawIMessageConfig {
     cli_path: Option<String>,
     db_path: Option<String>,
     dm_policy: Option<String>,
-    allow_from: Option<Vec<String>>,
+    allow_from: Option<serde_json::Value>,
     enabled: Option<bool>,
 }
 
@@ -282,7 +301,7 @@ struct OpenClawBlueBubblesConfig {
     server_url: Option<String>,
     password: Option<String>,
     dm_policy: Option<String>,
-    allow_from: Option<Vec<String>>,
+    allow_from: Option<serde_json::Value>,
     enabled: Option<bool>,
 }
 
@@ -488,16 +507,18 @@ fn build_channel_table(
     fields: Vec<(&str, toml::Value)>,
     dm_policy: Option<&str>,
     group_policy: Option<&str>,
-    allow_from: Option<&[String]>,
+    allow_from: Option<&serde_json::Value>,
 ) -> toml::Value {
     let mut table = toml::map::Map::new();
     for (key, val) in fields {
         table.insert(key.to_string(), val);
     }
 
+    let allow_list = allow_from.map(extract_string_list).unwrap_or_default();
+
     // Add overrides sub-table if any policy is set
     let has_overrides =
-        dm_policy.is_some() || group_policy.is_some() || allow_from.is_some_and(|a| !a.is_empty());
+        dm_policy.is_some() || group_policy.is_some() || !allow_list.is_empty();
 
     if has_overrides {
         let mut overrides = toml::map::Map::new();
@@ -515,14 +536,12 @@ fn build_channel_table(
                 toml::Value::String(mapped.to_string()),
             );
         }
-        if let Some(users) = allow_from {
-            if !users.is_empty() {
-                let arr: Vec<toml::Value> = users
-                    .iter()
-                    .map(|u| toml::Value::String(u.clone()))
-                    .collect();
-                overrides.insert("allowed_users".to_string(), toml::Value::Array(arr));
-            }
+        if !allow_list.is_empty() {
+            let arr: Vec<toml::Value> = allow_list
+                .iter()
+                .map(|u| toml::Value::String(u.clone()))
+                .collect();
+            overrides.insert("allowed_users".to_string(), toml::Value::Array(arr));
         }
         table.insert("overrides".to_string(), toml::Value::Table(overrides));
     }
@@ -811,13 +830,14 @@ fn scan_from_json5(base: &Path, config_path: &Path, result: &mut ScanResult) {
                 .tools
                 .as_ref()
                 .and_then(|t| t.allow.as_ref())
-                .map(|a| a.len())
+                .map(|a| extract_string_list(a).len())
                 .or_else(|| {
                     entry
                         .tools
                         .as_ref()
                         .and_then(|t| t.profile.as_ref())
-                        .map(|p| tools_for_profile(p).len())
+                        .and_then(extract_profile)
+                        .map(|p| tools_for_profile(&p).len())
                 })
                 .unwrap_or(3);
 
@@ -1253,7 +1273,8 @@ fn migrate_channels_from_json(
                 "bot_token_env",
                 toml::Value::String("TELEGRAM_BOT_TOKEN".into()),
             )];
-            if let Some(ref users) = tg.allow_from {
+            if let Some(ref users_val) = tg.allow_from {
+                let users = extract_string_list(users_val);
                 if !users.is_empty() {
                     let arr: Vec<toml::Value> = users
                         .iter()
@@ -1268,7 +1289,7 @@ fn migrate_channels_from_json(
                     fields,
                     tg.dm_policy.as_deref(),
                     tg.group_policy.as_deref(),
-                    tg.allow_from.as_deref(),
+                    tg.allow_from.as_ref(),
                 ),
             );
             report.imported.push(MigrateItem {
@@ -1295,7 +1316,7 @@ fn migrate_channels_from_json(
                     fields,
                     dc.dm_policy.as_deref(),
                     dc.group_policy.as_deref(),
-                    dc.allow_from.as_deref(),
+                    dc.allow_from.as_ref(),
                 ),
             );
             report.imported.push(MigrateItem {
@@ -1331,7 +1352,7 @@ fn migrate_channels_from_json(
                     fields,
                     sl.dm_policy.as_deref(),
                     sl.group_policy.as_deref(),
-                    sl.allow_from.as_deref(),
+                    sl.allow_from.as_ref(),
                 ),
             );
             report.imported.push(MigrateItem {
@@ -1372,7 +1393,8 @@ fn migrate_channels_from_json(
                 "access_token_env",
                 toml::Value::String("WHATSAPP_ACCESS_TOKEN".into()),
             )];
-            if let Some(ref users) = wa.allow_from {
+            if let Some(ref users_val) = wa.allow_from {
+                let users = extract_string_list(users_val);
                 if !users.is_empty() {
                     let arr: Vec<toml::Value> = users
                         .iter()
@@ -1387,7 +1409,7 @@ fn migrate_channels_from_json(
                     fields,
                     wa.dm_policy.as_deref(),
                     wa.group_policy.as_deref(),
-                    wa.allow_from.as_deref(),
+                    wa.allow_from.as_ref(),
                 ),
             );
             report.imported.push(MigrateItem {
@@ -1418,7 +1440,7 @@ fn migrate_channels_from_json(
                     fields,
                     sig.dm_policy.as_deref(),
                     None,
-                    sig.allow_from.as_deref(),
+                    sig.allow_from.as_ref(),
                 ),
             );
             report.imported.push(MigrateItem {
@@ -1445,7 +1467,8 @@ fn migrate_channels_from_json(
             if let Some(ref uid) = mx.user_id {
                 fields.push(("user_id", toml::Value::String(uid.clone())));
             }
-            if let Some(ref rooms) = mx.rooms {
+            if let Some(ref rooms_val) = mx.rooms {
+                let rooms = extract_string_list(rooms_val);
                 if !rooms.is_empty() {
                     let arr: Vec<toml::Value> = rooms
                         .iter()
@@ -1460,7 +1483,7 @@ fn migrate_channels_from_json(
                     fields,
                     mx.dm_policy.as_deref(),
                     None,
-                    mx.allow_from.as_deref(),
+                    mx.allow_from.as_ref(),
                 ),
             );
             report.imported.push(MigrateItem {
@@ -1534,7 +1557,7 @@ fn migrate_channels_from_json(
                     fields,
                     tm.dm_policy.as_deref(),
                     None,
-                    tm.allow_from.as_deref(),
+                    tm.allow_from.as_ref(),
                 ),
             );
             report.imported.push(MigrateItem {
@@ -1567,7 +1590,8 @@ fn migrate_channels_from_json(
             if irc.password.is_some() {
                 fields.push(("password_env", toml::Value::String("IRC_PASSWORD".into())));
             }
-            if let Some(ref chans) = irc.channels {
+            if let Some(ref chans_val) = irc.channels {
+                let chans = extract_string_list(chans_val);
                 if !chans.is_empty() {
                     let arr: Vec<toml::Value> = chans
                         .iter()
@@ -1582,7 +1606,7 @@ fn migrate_channels_from_json(
                     fields,
                     irc.dm_policy.as_deref(),
                     None,
-                    irc.allow_from.as_deref(),
+                    irc.allow_from.as_ref(),
                 ),
             );
             report.imported.push(MigrateItem {
@@ -1612,7 +1636,7 @@ fn migrate_channels_from_json(
                     fields,
                     mm.dm_policy.as_deref(),
                     None,
-                    mm.allow_from.as_deref(),
+                    mm.allow_from.as_ref(),
                 ),
             );
             report.imported.push(MigrateItem {
@@ -1770,9 +1794,10 @@ fn convert_agent_from_json(
     // Resolve tools
     let mut unmapped_tools = Vec::new();
     let tools: Vec<String> = if let Some(ref agent_tools) = entry.tools {
-        if let Some(ref allow) = agent_tools.allow {
+        if let Some(ref allow_val) = agent_tools.allow {
+            let allow = extract_string_list(allow_val);
             let mut mapped = Vec::new();
-            for t in allow {
+            for t in &allow {
                 if is_known_openfang_tool(t) {
                     mapped.push(t.clone());
                 } else if let Some(of_name) = map_tool_name(t) {
@@ -1782,8 +1807,9 @@ fn convert_agent_from_json(
                 }
             }
             // also_allow
-            if let Some(ref also) = agent_tools.also_allow {
-                for t in also {
+            if let Some(ref also_val) = agent_tools.also_allow {
+                let also = extract_string_list(also_val);
+                for t in &also {
                     if is_known_openfang_tool(t) {
                         mapped.push(t.clone());
                     } else if let Some(of_name) = map_tool_name(t) {
@@ -1794,8 +1820,9 @@ fn convert_agent_from_json(
                 }
             }
             mapped
-        } else if let Some(ref profile) = agent_tools.profile {
-            tools_for_profile(profile)
+        } else if let Some(ref profile_val) = agent_tools.profile {
+            let profile_name = extract_profile(profile_val).unwrap_or_default();
+            tools_for_profile(&profile_name)
         } else {
             resolve_default_tools(defaults)
         }
@@ -1894,8 +1921,10 @@ fn convert_agent_from_json(
 
     // Tool profile hint
     if let Some(ref agent_tools) = entry.tools {
-        if let Some(ref profile) = agent_tools.profile {
-            toml_str.push_str(&format!("\nprofile = \"{profile}\"\n"));
+        if let Some(ref profile_val) = agent_tools.profile {
+            if let Some(profile) = extract_profile(profile_val) {
+                toml_str.push_str(&format!("\nprofile = \"{profile}\"\n"));
+            }
         }
     }
 
@@ -1905,12 +1934,15 @@ fn convert_agent_from_json(
 fn resolve_default_tools(defaults: Option<&OpenClawAgentDefaults>) -> Vec<String> {
     if let Some(defs) = defaults {
         if let Some(ref tools) = defs.tools {
-            if let Some(ref profile) = tools.profile {
-                return tools_for_profile(profile);
+            if let Some(ref profile_val) = tools.profile {
+                if let Some(profile) = extract_profile(profile_val) {
+                    return tools_for_profile(&profile);
+                }
             }
-            if let Some(ref allow) = tools.allow {
+            if let Some(ref allow_val) = tools.allow {
+                let allow = extract_string_list(allow_val);
                 let mut mapped = Vec::new();
-                for t in allow {
+                for t in &allow {
                     if is_known_openfang_tool(t) {
                         mapped.push(t.clone());
                     } else if let Some(of_name) = map_tool_name(t) {

crates/openfang-runtime/Cargo.toml

@@ -29,7 +29,9 @@ hex = { workspace = true }
 zeroize = { workspace = true }
 dashmap = { workspace = true }
 regex-lite = { workspace = true }
+rusqlite = { workspace = true }
 tokio-tungstenite = "0.24"
+shlex = "1"
 
 [dev-dependencies]
 tokio-test = { workspace = true }

crates/openfang-runtime/src/a2a.rs

@@ -87,8 +87,8 @@ pub struct A2aTask {
     /// Optional session identifier for conversation continuity.
     #[serde(default)]
     pub session_id: Option<String>,
-    /// Current task status.
-    pub status: A2aTaskStatus,
+    /// Current task status (accepts both string and object forms).
+    pub status: A2aTaskStatusWrapper,
     /// Messages exchanged during the task.
     #[serde(default)]
     pub messages: Vec<A2aMessage>,
@@ -115,6 +115,44 @@ pub enum A2aTaskStatus {
     Failed,
 }
 
+/// Wrapper that accepts either a bare status string (`"completed"`)
+/// or the object form (`{"state": "completed", "message": null}`)
+/// used by some A2A implementations.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(untagged)]
+pub enum A2aTaskStatusWrapper {
+    /// Object form: `{"state": "completed", "message": ...}`.
+    Object {
+        state: A2aTaskStatus,
+        #[serde(default)]
+        message: Option<serde_json::Value>,
+    },
+    /// Bare enum form: `"completed"`.
+    Enum(A2aTaskStatus),
+}
+
+impl A2aTaskStatusWrapper {
+    /// Extract the underlying `A2aTaskStatus` regardless of encoding form.
+    pub fn state(&self) -> &A2aTaskStatus {
+        match self {
+            Self::Object { state, .. } => state,
+            Self::Enum(s) => s,
+        }
+    }
+}
+
+impl From<A2aTaskStatus> for A2aTaskStatusWrapper {
+    fn from(status: A2aTaskStatus) -> Self {
+        Self::Enum(status)
+    }
+}
+
+impl PartialEq<A2aTaskStatus> for A2aTaskStatusWrapper {
+    fn eq(&self, other: &A2aTaskStatus) -> bool {
+        self.state() == other
+    }
+}
+
 /// A2A message in a task conversation.
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct A2aMessage {
@@ -145,9 +183,23 @@ pub enum A2aPart {
 
 /// A2A artifact produced by a task.
 #[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
 pub struct A2aArtifact {
-    /// Artifact name.
-    pub name: String,
+    /// Artifact name (optional per spec).
+    #[serde(default)]
+    pub name: Option<String>,
+    /// Human-readable description.
+    #[serde(default)]
+    pub description: Option<String>,
+    /// Arbitrary metadata.
+    #[serde(default)]
+    pub metadata: Option<serde_json::Value>,
+    /// Artifact index in the sequence.
+    #[serde(default)]
+    pub index: Option<u32>,
+    /// Whether this is the last chunk of a streamed artifact.
+    #[serde(default)]
+    pub last_chunk: Option<bool>,
     /// Artifact content parts.
     pub parts: Vec<A2aPart>,
 }
@@ -185,7 +237,7 @@ impl A2aTaskStore {
                 .iter()
                 .filter(|(_, t)| {
                     matches!(
-                        t.status,
+                        t.status.state(),
                         A2aTaskStatus::Completed | A2aTaskStatus::Failed | A2aTaskStatus::Cancelled
                     )
                 })
@@ -211,7 +263,7 @@ impl A2aTaskStore {
     pub fn update_status(&self, task_id: &str, status: A2aTaskStatus) -> bool {
         let mut tasks = self.tasks.lock().unwrap_or_else(|e| e.into_inner());
         if let Some(task) = tasks.get_mut(task_id) {
-            task.status = status;
+            task.status = status.into();
             true
         } else {
             false
@@ -224,7 +276,7 @@ impl A2aTaskStore {
         if let Some(task) = tasks.get_mut(task_id) {
             task.messages.push(response);
             task.artifacts.extend(artifacts);
-            task.status = A2aTaskStatus::Completed;
+            task.status = A2aTaskStatus::Completed.into();
         }
     }
 
@@ -233,7 +285,7 @@ impl A2aTaskStore {
         let mut tasks = self.tasks.lock().unwrap_or_else(|e| e.into_inner());
         if let Some(task) = tasks.get_mut(task_id) {
             task.messages.push(error_message);
-            task.status = A2aTaskStatus::Failed;
+            task.status = A2aTaskStatus::Failed.into();
         }
     }
 
@@ -490,7 +542,7 @@ mod tests {
         let task = A2aTask {
             id: "task-1".to_string(),
             session_id: None,
-            status: A2aTaskStatus::Submitted,
+            status: A2aTaskStatus::Submitted.into(),
             messages: vec![],
             artifacts: vec![],
         };
@@ -498,30 +550,71 @@ mod tests {
 
         // Simulate progression
         let working = A2aTask {
-            status: A2aTaskStatus::Working,
+            status: A2aTaskStatus::Working.into(),
             ..task.clone()
         };
         assert_eq!(working.status, A2aTaskStatus::Working);
 
         let completed = A2aTask {
-            status: A2aTaskStatus::Completed,
+            status: A2aTaskStatus::Completed.into(),
             ..task.clone()
         };
         assert_eq!(completed.status, A2aTaskStatus::Completed);
 
         let cancelled = A2aTask {
-            status: A2aTaskStatus::Cancelled,
+            status: A2aTaskStatus::Cancelled.into(),
             ..task.clone()
         };
         assert_eq!(cancelled.status, A2aTaskStatus::Cancelled);
 
         let failed = A2aTask {
-            status: A2aTaskStatus::Failed,
+            status: A2aTaskStatus::Failed.into(),
             ..task
         };
         assert_eq!(failed.status, A2aTaskStatus::Failed);
     }
 
+    #[test]
+    fn test_a2a_task_status_wrapper_object_form() {
+        // Test deserialization of the object form: {"state": "completed", "message": null}
+        let json = r#"{"state":"completed","message":null}"#;
+        let wrapper: A2aTaskStatusWrapper = serde_json::from_str(json).unwrap();
+        assert_eq!(wrapper, A2aTaskStatus::Completed);
+        assert_eq!(wrapper.state(), &A2aTaskStatus::Completed);
+
+        // Test with a message payload
+        let json_with_msg =
+            r#"{"state":"working","message":{"text":"Processing..."}}"#;
+        let wrapper2: A2aTaskStatusWrapper = serde_json::from_str(json_with_msg).unwrap();
+        assert_eq!(wrapper2, A2aTaskStatus::Working);
+
+        // Test bare string form
+        let json_bare = r#""completed""#;
+        let wrapper3: A2aTaskStatusWrapper = serde_json::from_str(json_bare).unwrap();
+        assert_eq!(wrapper3, A2aTaskStatus::Completed);
+    }
+
+    #[test]
+    fn test_a2a_artifact_optional_fields() {
+        // name is now optional — artifact with no name should deserialize
+        let json = r#"{"parts":[{"type":"text","text":"hello"}]}"#;
+        let artifact: A2aArtifact = serde_json::from_str(json).unwrap();
+        assert!(artifact.name.is_none());
+        assert!(artifact.description.is_none());
+        assert!(artifact.metadata.is_none());
+        assert!(artifact.index.is_none());
+        assert!(artifact.last_chunk.is_none());
+        assert_eq!(artifact.parts.len(), 1);
+
+        // Full artifact with all optional fields
+        let json_full = r#"{"name":"output.txt","description":"The result","metadata":{"key":"val"},"index":0,"lastChunk":true,"parts":[]}"#;
+        let full: A2aArtifact = serde_json::from_str(json_full).unwrap();
+        assert_eq!(full.name.as_deref(), Some("output.txt"));
+        assert_eq!(full.description.as_deref(), Some("The result"));
+        assert_eq!(full.index, Some(0));
+        assert_eq!(full.last_chunk, Some(true));
+    }
+
     #[test]
     fn test_a2a_message_serde() {
         let msg = A2aMessage {
@@ -554,7 +647,7 @@ mod tests {
         let task = A2aTask {
             id: "t-1".to_string(),
             session_id: None,
-            status: A2aTaskStatus::Working,
+            status: A2aTaskStatus::Working.into(),
             messages: vec![],
             artifacts: vec![],
         };
@@ -571,7 +664,7 @@ mod tests {
         let task = A2aTask {
             id: "t-2".to_string(),
             session_id: None,
-            status: A2aTaskStatus::Working,
+            status: A2aTaskStatus::Working.into(),
             messages: vec![],
             artifacts: vec![],
         };
@@ -599,7 +692,7 @@ mod tests {
         let task = A2aTask {
             id: "t-3".to_string(),
             session_id: None,
-            status: A2aTaskStatus::Working,
+            status: A2aTaskStatus::Working.into(),
             messages: vec![],
             artifacts: vec![],
         };
@@ -618,7 +711,7 @@ mod tests {
             let task = A2aTask {
                 id: format!("t-{i}"),
                 session_id: None,
-                status: A2aTaskStatus::Completed,
+                status: A2aTaskStatus::Completed.into(),
                 messages: vec![],
                 artifacts: vec![],
             };
@@ -630,7 +723,7 @@ mod tests {
         let task = A2aTask {
             id: "t-2".to_string(),
             session_id: None,
-            status: A2aTaskStatus::Working,
+            status: A2aTaskStatus::Working.into(),
             messages: vec![],
             artifacts: vec![],
         };

crates/openfang-runtime/src/audit.rs

@@ -3,11 +3,15 @@
 //! Every auditable event is appended to an append-only log where each entry
 //! contains the SHA-256 hash of its own contents concatenated with the hash of
 //! the previous entry, forming a tamper-evident chain (similar to a blockchain).
+//!
+//! When a database connection is provided (`with_db`), entries are persisted to
+//! the `audit_entries` table (schema V8) so the trail survives daemon restarts.
 
 use chrono::Utc;
+use rusqlite::Connection;
 use serde::{Deserialize, Serialize};
 use sha2::{Digest, Sha256};
-use std::sync::Mutex;
+use std::sync::{Arc, Mutex};
 
 /// Categories of auditable actions within the agent runtime.
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -77,26 +81,102 @@ fn compute_entry_hash(
 /// An append-only, tamper-evident audit log using a Merkle hash chain.
 ///
 /// Thread-safe — all access is serialised through internal mutexes.
+/// Optionally backed by SQLite for persistence across daemon restarts.
 pub struct AuditLog {
     entries: Mutex<Vec<AuditEntry>>,
     tip: Mutex<String>,
+    /// Optional database connection for persistent storage.
+    db: Option<Arc<Mutex<Connection>>>,
 }
 
 impl AuditLog {
-    /// Creates a new empty audit log.
+    /// Creates a new empty audit log (in-memory only, no persistence).
     ///
     /// The initial tip hash is 64 zero characters (the "genesis" sentinel).
     pub fn new() -> Self {
         Self {
             entries: Mutex::new(Vec::new()),
             tip: Mutex::new("0".repeat(64)),
+            db: None,
         }
     }
 
+    /// Creates an audit log backed by a database connection.
+    ///
+    /// On construction, loads all existing entries from the `audit_entries`
+    /// table and verifies the Merkle chain integrity. New entries are written
+    /// to both the in-memory chain and the database.
+    pub fn with_db(conn: Arc<Mutex<Connection>>) -> Self {
+        let mut entries = Vec::new();
+        let mut tip = "0".repeat(64);
+
+        // Load existing entries from database
+        if let Ok(db) = conn.lock() {
+            let result = db.prepare(
+                "SELECT seq, timestamp, agent_id, action, detail, outcome, prev_hash, hash FROM audit_entries ORDER BY seq ASC",
+            );
+            if let Ok(mut stmt) = result {
+                let rows = stmt.query_map([], |row| {
+                    let action_str: String = row.get(3)?;
+                    let action = match action_str.as_str() {
+                        "ToolInvoke" => AuditAction::ToolInvoke,
+                        "CapabilityCheck" => AuditAction::CapabilityCheck,
+                        "AgentSpawn" => AuditAction::AgentSpawn,
+                        "AgentKill" => AuditAction::AgentKill,
+                        "AgentMessage" => AuditAction::AgentMessage,
+                        "MemoryAccess" => AuditAction::MemoryAccess,
+                        "FileAccess" => AuditAction::FileAccess,
+                        "NetworkAccess" => AuditAction::NetworkAccess,
+                        "ShellExec" => AuditAction::ShellExec,
+                        "AuthAttempt" => AuditAction::AuthAttempt,
+                        "WireConnect" => AuditAction::WireConnect,
+                        "ConfigChange" => AuditAction::ConfigChange,
+                        _ => AuditAction::ToolInvoke, // fallback
+                    };
+                    Ok(AuditEntry {
+                        seq: row.get(0)?,
+                        timestamp: row.get(1)?,
+                        agent_id: row.get(2)?,
+                        action,
+                        detail: row.get(4)?,
+                        outcome: row.get(5)?,
+                        prev_hash: row.get(6)?,
+                        hash: row.get(7)?,
+                    })
+                });
+                if let Ok(rows) = rows {
+                    for entry in rows.flatten() {
+                        tip = entry.hash.clone();
+                        entries.push(entry);
+                    }
+                }
+            }
+        }
+
+        let count = entries.len();
+        let log = Self {
+            entries: Mutex::new(entries),
+            tip: Mutex::new(tip),
+            db: Some(conn),
+        };
+
+        // Verify chain integrity on load
+        if count > 0 {
+            if let Err(e) = log.verify_integrity() {
+                tracing::error!("Audit trail integrity check FAILED on boot: {e}");
+            } else {
+                tracing::info!("Audit trail loaded: {count} entries, chain integrity OK");
+            }
+        }
+
+        log
+    }
+
     /// Records a new auditable event and returns the SHA-256 hash of the entry.
     ///
     /// The entry is atomically appended to the chain with the current tip as
     /// its `prev_hash`, and the tip is advanced to the new hash.
+    /// If a database connection is available, the entry is also persisted.
     pub fn record(
         &self,
         agent_id: impl Into<String>,
@@ -119,7 +199,7 @@ impl AuditLog {
             seq, &timestamp, &agent_id, &action, &detail, &outcome, &prev_hash,
         );
 
-        entries.push(AuditEntry {
+        let entry = AuditEntry {
             seq,
             timestamp,
             agent_id,
@@ -128,8 +208,28 @@ impl AuditLog {
             outcome,
             prev_hash,
             hash: hash.clone(),
-        });
+        };
 
+        // Persist to database if available
+        if let Some(ref db) = self.db {
+            if let Ok(conn) = db.lock() {
+                let _ = conn.execute(
+                    "INSERT INTO audit_entries (seq, timestamp, agent_id, action, detail, outcome, prev_hash, hash) VALUES (?1, ?2, ?3, ?4, ?5, ?6, ?7, ?8)",
+                    rusqlite::params![
+                        entry.seq as i64,
+                        &entry.timestamp,
+                        &entry.agent_id,
+                        entry.action.to_string(),
+                        &entry.detail,
+                        &entry.outcome,
+                        &entry.prev_hash,
+                        &entry.hash,
+                    ],
+                );
+            }
+        }
+
+        entries.push(entry);
         *tip = hash.clone();
         hash
     }
@@ -271,4 +371,52 @@ mod tests {
         assert_eq!(log.tip_hash(), h2);
         assert_ne!(h2, h1);
     }
+
+    #[test]
+    fn test_audit_persists_to_db() {
+        let conn = Connection::open_in_memory().unwrap();
+        conn.execute_batch(
+            "CREATE TABLE audit_entries (
+                seq INTEGER PRIMARY KEY,
+                timestamp TEXT NOT NULL,
+                agent_id TEXT NOT NULL,
+                action TEXT NOT NULL,
+                detail TEXT NOT NULL,
+                outcome TEXT NOT NULL,
+                prev_hash TEXT NOT NULL,
+                hash TEXT NOT NULL
+            )",
+        )
+        .unwrap();
+
+        let db = Arc::new(Mutex::new(conn));
+
+        // Record entries with DB
+        let log = AuditLog::with_db(Arc::clone(&db));
+        log.record("agent-1", AuditAction::AgentSpawn, "spawn test", "ok");
+        log.record("agent-1", AuditAction::ShellExec, "ls", "ok");
+        assert_eq!(log.len(), 2);
+
+        // Verify entries in database
+        let db_conn = db.lock().unwrap();
+        let count: i64 = db_conn
+            .query_row("SELECT COUNT(*) FROM audit_entries", [], |row| row.get(0))
+            .unwrap();
+        assert_eq!(count, 2);
+        drop(db_conn);
+
+        // Simulate restart: create new AuditLog from same DB
+        let log2 = AuditLog::with_db(Arc::clone(&db));
+        assert_eq!(log2.len(), 2);
+        assert!(log2.verify_integrity().is_ok());
+
+        // Chain continues correctly after restart
+        log2.record("agent-2", AuditAction::ToolInvoke, "file_read", "ok");
+        assert_eq!(log2.len(), 3);
+        assert!(log2.verify_integrity().is_ok());
+
+        // Verify tip is correct
+        let entries = log2.recent(3);
+        assert_eq!(entries[2].prev_hash, entries[1].hash);
+    }
 }

crates/openfang-runtime/src/context_budget.rs

@@ -65,22 +65,23 @@ pub fn truncate_tool_result_dynamic(content: &str, budget: &ContextBudget) -> St
     }
 
     // Find last newline before the cap to break cleanly (char-boundary safe)
-    let safe_cap = if content.is_char_boundary(cap) {
-        cap
-    } else {
-        content[..cap].char_indices().next_back().map(|(i, _)| i).unwrap_or(0)
-    };
-    let search_start = safe_cap.saturating_sub(200);
-    let break_point = content[search_start..safe_cap]
+    let mut safe_cap = cap.min(content.len());
+    while safe_cap > 0 && !content.is_char_boundary(safe_cap) {
+        safe_cap -= 1;
+    }
+    let mut search_start = safe_cap.saturating_sub(200);
+    // Ensure search_start is a valid char boundary
+    while search_start > 0 && !content.is_char_boundary(search_start) {
+        search_start -= 1;
+    }
+    let mut break_point = content[search_start..safe_cap]
         .rfind('\n')
         .map(|pos| search_start + pos)
         .unwrap_or(safe_cap.saturating_sub(100));
     // Ensure break_point is also a char boundary
-    let break_point = if content.is_char_boundary(break_point) {
-        break_point
-    } else {
-        content[..break_point].char_indices().next_back().map(|(i, _)| i).unwrap_or(0)
-    };
+    while break_point > 0 && !content.is_char_boundary(break_point) {
+        break_point -= 1;
+    }
 
     format!(
         "{}\n\n[TRUNCATED: result was {} chars, showing first {} (budget: {}% of {}K context window)]",
@@ -201,28 +202,16 @@ fn truncate_to(content: &str, max_chars: usize) -> String {
     if content.len() <= max_chars {
         return content.to_string();
     }
-    let keep = max_chars.saturating_sub(80).min(content.len());
-    // Ensure keep is a valid char boundary
-    let keep = if content.is_char_boundary(keep) {
-        keep
-    } else {
-        content[..keep]
-            .char_indices()
-            .next_back()
-            .map(|(i, _)| i)
-            .unwrap_or(0)
-    };
-    let search_start = keep.saturating_sub(100);
-    // Ensure search_start is a valid char boundary
-    let search_start = if content.is_char_boundary(search_start) {
-        search_start
-    } else {
-        content[..search_start]
-            .char_indices()
-            .next_back()
-            .map(|(i, _)| i)
-            .unwrap_or(0)
-    };
+    let mut keep = max_chars.saturating_sub(80).min(content.len());
+    // Walk back to a valid char boundary
+    while keep > 0 && !content.is_char_boundary(keep) {
+        keep -= 1;
+    }
+    let mut search_start = keep.saturating_sub(100);
+    // Walk back to a valid char boundary
+    while search_start > 0 && !content.is_char_boundary(search_start) {
+        search_start -= 1;
+    }
     // Try to break at newline
     let break_point = content[search_start..keep]
         .rfind('\n')
@@ -319,4 +308,47 @@ mod tests {
             }
         }
     }
+
+    #[test]
+    fn test_truncate_tool_result_multibyte_chinese() {
+        // Tiny budget: cap = 30% of 100 * 2.0 = 60 bytes
+        let budget = ContextBudget::new(100);
+        // Each Chinese char is 3 bytes in UTF-8; 100 chars = 300 bytes
+        let content: String = "\u{4f60}\u{597d}\u{4e16}\u{754c}".repeat(25);
+        assert_eq!(content.len(), 300);
+        // Must not panic on multi-byte content
+        let result = truncate_tool_result_dynamic(&content, &budget);
+        assert!(result.contains("[TRUNCATED:"));
+        // The visible portion must be valid UTF-8 (implicit: no panic)
+        assert!(result.is_char_boundary(0));
+    }
+
+    #[test]
+    fn test_truncate_to_multibyte_emoji() {
+        // Each emoji is 4 bytes; 200 emojis = 800 bytes
+        let content: String = "\u{1f600}".repeat(200);
+        let result = truncate_to(&content, 100);
+        assert!(result.contains("[COMPACTED:"));
+        // Must not panic and must produce valid UTF-8
+        assert!(result.is_char_boundary(0));
+    }
+
+    #[test]
+    fn test_context_guard_multibyte_tool_results() {
+        let budget = ContextBudget::new(100);
+        // Chinese text: 500 chars * 3 bytes = 1500 bytes
+        let big_chinese: String = "\u{4e2d}\u{6587}\u{6d4b}\u{8bd5}\u{6570}\u{636e}".repeat(83);
+        let mut messages = vec![Message {
+            role: openfang_types::message::Role::User,
+            content: MessageContent::Blocks(vec![ContentBlock::ToolResult {
+                tool_use_id: "t1".to_string(),
+                tool_name: String::new(),
+                content: big_chinese,
+                is_error: false,
+            }]),
+        }];
+        // Must not panic on multi-byte content
+        let compacted = apply_context_guard(&mut messages, &budget, &[]);
+        assert!(compacted > 0);
+    }
 }

crates/openfang-runtime/src/context_overflow.rs

@@ -102,13 +102,11 @@ pub fn recover_from_overflow(
             for block in blocks.iter_mut() {
                 if let ContentBlock::ToolResult { content, .. } = block {
                     if content.len() > tool_truncation_limit {
-                        let keep = tool_truncation_limit.saturating_sub(80);
-                        // Find a valid char boundary at or before `keep`
-                        let safe_keep = if content.is_char_boundary(keep) {
-                            keep
-                        } else {
-                            content[..keep].char_indices().next_back().map(|(i, _)| i).unwrap_or(0)
-                        };
+                        let mut safe_keep = tool_truncation_limit.saturating_sub(80);
+                        // Walk back to a valid char boundary
+                        while safe_keep > 0 && !content.is_char_boundary(safe_keep) {
+                            safe_keep -= 1;
+                        }
                         *content = format!(
                             "{}\n\n[OVERFLOW RECOVERY: truncated from {} to {} chars]",
                             &content[..safe_keep],
@@ -244,4 +242,26 @@ mod tests {
         // we should cascade through stages
         assert_ne!(stage, RecoveryStage::None);
     }
+
+    #[test]
+    fn test_stage3_multibyte_tool_truncation() {
+        // Chinese text (3 bytes per char) in tool results must not panic
+        let chinese_result: String = "\u{4f60}\u{597d}\u{4e16}\u{754c}".repeat(1250); // 5000 chars, 15000 bytes
+        let mut msgs = vec![
+            Message::user("hi"),
+            Message {
+                role: Role::User,
+                content: MessageContent::Blocks(vec![ContentBlock::ToolResult {
+                    tool_use_id: "t1".to_string(),
+                    tool_name: String::new(),
+                    content: chinese_result,
+                    is_error: false,
+                }]),
+            },
+        ];
+        // Tiny context window to force stage 3 tool truncation
+        let stage = recover_from_overflow(&mut msgs, "system", &[], 500);
+        // Must not panic — the truncation at byte boundaries could split a 3-byte char
+        assert_ne!(stage, RecoveryStage::None);
+    }
 }

crates/openfang-runtime/src/docker_sandbox.rs

@@ -61,19 +61,15 @@ fn validate_image_name(image: &str) -> Result<(), String> {
 }
 
 /// SECURITY: Sanitize command — reject dangerous shell metacharacters.
+/// Delegates to the comprehensive subprocess_sandbox check.
 fn validate_command(command: &str) -> Result<(), String> {
     if command.is_empty() {
         return Err("Command cannot be empty".into());
     }
-    // Reject backticks and $() which could enable command injection
-    let dangerous = ["`", "$(", "${"];
-    for pattern in &dangerous {
-        if command.contains(pattern) {
-            return Err(format!(
-                "Command contains disallowed pattern '{}' — potential injection",
-                pattern
-            ));
-        }
+    if let Some(reason) = crate::subprocess_sandbox::contains_shell_metacharacters(command) {
+        return Err(format!(
+            "Command blocked: contains {reason} — potential injection"
+        ));
     }
     Ok(())
 }
@@ -205,21 +201,23 @@ pub async fn exec_in_sandbox(
     let stderr = String::from_utf8_lossy(&output.stderr).to_string();
     let exit_code = output.status.code().unwrap_or(-1);
 
-    // Truncate large outputs
+    // Truncate large outputs (char-boundary safe to avoid UTF-8 panics)
     let max_output = 50_000;
     let stdout = if stdout.len() > max_output {
+        let safe_end = crate::str_utils::safe_truncate_str(&stdout, max_output);
         format!(
             "{}... [truncated, {} total bytes]",
-            &stdout[..max_output],
+            safe_end,
             stdout.len()
         )
     } else {
         stdout
     };
     let stderr = if stderr.len() > max_output {
+        let safe_end = crate::str_utils::safe_truncate_str(&stderr, max_output);
         format!(
             "{}... [truncated, {} total bytes]",
-            &stderr[..max_output],
+            safe_end,
             stderr.len()
         )
     } else {
@@ -489,7 +487,12 @@ mod tests {
     fn test_validate_command_valid() {
         assert!(validate_command("python script.py").is_ok());
         assert!(validate_command("ls -la /workspace").is_ok());
-        assert!(validate_command("echo hello | grep h").is_ok());
+    }
+
+    #[test]
+    fn test_validate_command_pipe_blocked() {
+        // SECURITY: Pipes now blocked by comprehensive metacharacter check
+        assert!(validate_command("echo hello | grep h").is_err());
     }
 
     #[test]

crates/openfang-runtime/src/drivers/anthropic.rs

@@ -415,12 +415,13 @@ impl LlmDriver for AnthropicDriver {
                             }
                         }
                         "content_block_delta" => {
+                            let block_idx = json["index"].as_u64().unwrap_or(0) as usize;
                             let delta = &json["delta"];
                             match delta["type"].as_str().unwrap_or("") {
                                 "text_delta" => {
                                     if let Some(text) = delta["text"].as_str() {
                                         if let Some(ContentBlockAccum::Text(ref mut t)) =
-                                            blocks.last_mut()
+                                            blocks.get_mut(block_idx)
                                         {
                                             t.push_str(text);
                                         }
@@ -436,7 +437,7 @@ impl LlmDriver for AnthropicDriver {
                                         if let Some(ContentBlockAccum::ToolUse {
                                             ref mut input_json,
                                             ..
-                                        }) = blocks.last_mut()
+                                        }) = blocks.get_mut(block_idx)
                                         {
                                             input_json.push_str(partial);
                                         }
@@ -450,7 +451,7 @@ impl LlmDriver for AnthropicDriver {
                                 "thinking_delta" => {
                                     if let Some(thinking) = delta["thinking"].as_str() {
                                         if let Some(ContentBlockAccum::Thinking(ref mut t)) =
-                                            blocks.last_mut()
+                                            blocks.get_mut(block_idx)
                                         {
                                             t.push_str(thinking);
                                         }
@@ -460,11 +461,12 @@ impl LlmDriver for AnthropicDriver {
                             }
                         }
                         "content_block_stop" => {
+                            let block_idx = json["index"].as_u64().unwrap_or(0) as usize;
                             if let Some(ContentBlockAccum::ToolUse {
                                 id,
                                 name,
                                 input_json,
-                            }) = blocks.last()
+                            }) = blocks.get(block_idx)
                             {
                                 let input: serde_json::Value =
                                     serde_json::from_str(input_json).unwrap_or_default();

crates/openfang-runtime/src/drivers/claude_code.rs

@@ -12,6 +12,38 @@ use serde::Deserialize;
 use tokio::io::AsyncBufReadExt;
 use tracing::{debug, warn};
 
+/// Environment variable names (and suffixes) to strip from the subprocess
+/// to prevent leaking API keys from other providers. We keep the full env
+/// intact (so Node.js, NVM, SSL, proxies, etc. all work) and only remove
+/// secrets that belong to other LLM providers.
+const SENSITIVE_ENV_EXACT: &[&str] = &[
+    "OPENAI_API_KEY",
+    "ANTHROPIC_API_KEY",
+    "GEMINI_API_KEY",
+    "GOOGLE_API_KEY",
+    "GROQ_API_KEY",
+    "DEEPSEEK_API_KEY",
+    "MISTRAL_API_KEY",
+    "TOGETHER_API_KEY",
+    "FIREWORKS_API_KEY",
+    "OPENROUTER_API_KEY",
+    "PERPLEXITY_API_KEY",
+    "COHERE_API_KEY",
+    "AI21_API_KEY",
+    "CEREBRAS_API_KEY",
+    "SAMBANOVA_API_KEY",
+    "HUGGINGFACE_API_KEY",
+    "XAI_API_KEY",
+    "REPLICATE_API_TOKEN",
+    "BRAVE_API_KEY",
+    "TAVILY_API_KEY",
+    "ELEVENLABS_API_KEY",
+];
+
+/// Suffixes that indicate a secret — remove any env var ending with these
+/// unless it starts with `CLAUDE_`.
+const SENSITIVE_SUFFIXES: &[&str] = &["_SECRET", "_TOKEN", "_PASSWORD"];
+
 /// LLM driver that delegates to the Claude Code CLI.
 pub struct ClaudeCodeDriver {
     cli_path: String,
@@ -80,13 +112,44 @@ impl ClaudeCodeDriver {
             _ => Some(stripped.to_string()),
         }
     }
+
+    /// Apply security env filtering to a command.
+    ///
+    /// Instead of `env_clear()` (which breaks Node.js, NVM, SSL, proxies),
+    /// we keep the full environment and only remove known sensitive API keys
+    /// from other LLM providers.
+    fn apply_env_filter(cmd: &mut tokio::process::Command) {
+        for key in SENSITIVE_ENV_EXACT {
+            cmd.env_remove(key);
+        }
+        // Remove any env var with a sensitive suffix, unless it's CLAUDE_*
+        for (key, _) in std::env::vars() {
+            if key.starts_with("CLAUDE_") {
+                continue;
+            }
+            let upper = key.to_uppercase();
+            for suffix in SENSITIVE_SUFFIXES {
+                if upper.ends_with(suffix) {
+                    cmd.env_remove(&key);
+                    break;
+                }
+            }
+        }
+    }
 }
 
 /// JSON output from `claude -p --output-format json`.
+///
+/// The CLI may return the response text in different fields depending on
+/// version: `result`, `content`, or `text`. We try all three.
 #[derive(Debug, Deserialize)]
 struct ClaudeJsonOutput {
     result: Option<String>,
     #[serde(default)]
+    content: Option<String>,
+    #[serde(default)]
+    text: Option<String>,
+    #[serde(default)]
     usage: Option<ClaudeUsage>,
     #[serde(default)]
     #[allow(dead_code)]
@@ -134,7 +197,8 @@ impl LlmDriver for ClaudeCodeDriver {
             cmd.arg("--model").arg(model);
         }
 
-        // SECURITY: Don't inherit all env vars — only safe ones
+        Self::apply_env_filter(&mut cmd);
+
         cmd.stdout(std::process::Stdio::piped());
         cmd.stderr(std::process::Stdio::piped());
 
@@ -143,13 +207,41 @@ impl LlmDriver for ClaudeCodeDriver {
         let output = cmd
             .output()
             .await
-            .map_err(|e| LlmError::Http(format!("Failed to spawn claude CLI: {e}")))?;
+            .map_err(|e| LlmError::Http(format!(
+                "Claude Code CLI not found or failed to start ({}). \
+                 Install: npm install -g @anthropic-ai/claude-code && claude auth",
+                e
+            )))?;
 
         if !output.status.success() {
-            let stderr = String::from_utf8_lossy(&output.stderr);
+            let stderr = String::from_utf8_lossy(&output.stderr).trim().to_string();
+            let stdout = String::from_utf8_lossy(&output.stdout).trim().to_string();
+            let detail = if !stderr.is_empty() { &stderr } else { &stdout };
+            let code = output.status.code().unwrap_or(1);
+
+            // Provide actionable error messages
+            let message = if detail.contains("not authenticated")
+                || detail.contains("auth")
+                || detail.contains("login")
+                || detail.contains("credentials")
+            {
+                format!(
+                    "Claude Code CLI is not authenticated. Run: claude auth\nDetail: {detail}"
+                )
+            } else if detail.contains("permission")
+                || detail.contains("--dangerously-skip-permissions")
+            {
+                format!(
+                    "Claude Code CLI requires permissions acceptance. \
+                     Run: claude --dangerously-skip-permissions (once to accept)\nDetail: {detail}"
+                )
+            } else {
+                format!("Claude Code CLI exited with code {code}: {detail}")
+            };
+
             return Err(LlmError::Api {
-                status: output.status.code().unwrap_or(1) as u16,
-                message: format!("Claude CLI failed: {stderr}"),
+                status: code as u16,
+                message,
             });
         }
 
@@ -157,7 +249,10 @@ impl LlmDriver for ClaudeCodeDriver {
 
         // Try JSON parse first
         if let Ok(parsed) = serde_json::from_str::<ClaudeJsonOutput>(&stdout) {
-            let text = parsed.result.unwrap_or_default();
+            let text = parsed.result
+                .or(parsed.content)
+                .or(parsed.text)
+                .unwrap_or_default();
             let usage = parsed.usage.unwrap_or_default();
             return Ok(CompletionResponse {
                 content: vec![ContentBlock::Text { text: text.clone() }],
@@ -195,12 +290,15 @@ impl LlmDriver for ClaudeCodeDriver {
         cmd.arg("-p")
             .arg(&prompt)
             .arg("--output-format")
-            .arg("stream-json");
+            .arg("stream-json")
+            .arg("--verbose");
 
         if let Some(ref model) = model_flag {
             cmd.arg("--model").arg(model);
         }
 
+        Self::apply_env_filter(&mut cmd);
+
         cmd.stdout(std::process::Stdio::piped());
         cmd.stderr(std::process::Stdio::piped());
 
@@ -208,7 +306,11 @@ impl LlmDriver for ClaudeCodeDriver {
 
         let mut child = cmd
             .spawn()
-            .map_err(|e| LlmError::Http(format!("Failed to spawn claude CLI: {e}")))?;
+            .map_err(|e| LlmError::Http(format!(
+                "Claude Code CLI not found or failed to start ({}). \
+                 Install: npm install -g @anthropic-ai/claude-code && claude auth",
+                e
+            )))?;
 
         let stdout = child
             .stdout
@@ -232,7 +334,7 @@ impl LlmDriver for ClaudeCodeDriver {
             match serde_json::from_str::<ClaudeStreamEvent>(&line) {
                 Ok(event) => {
                     match event.r#type.as_str() {
-                        "content" | "text" => {
+                        "content" | "text" | "assistant" | "content_block_delta" => {
                             if let Some(ref content) = event.content {
                                 full_text.push_str(content);
                                 let _ = tx
@@ -316,10 +418,16 @@ pub fn claude_code_available() -> bool {
         || claude_credentials_exist()
 }
 
-/// Check if Claude credentials file exists (~/.claude/.credentials.json).
+/// Check if Claude credentials file exists.
+///
+/// Different Claude CLI versions store credentials at different paths:
+/// - `~/.claude/.credentials.json` (older versions)
+/// - `~/.claude/credentials.json` (newer versions)
 fn claude_credentials_exist() -> bool {
     if let Some(home) = home_dir() {
-        home.join(".claude").join(".credentials.json").exists()
+        let claude_dir = home.join(".claude");
+        claude_dir.join(".credentials.json").exists()
+            || claude_dir.join("credentials.json").exists()
     } else {
         false
     }
@@ -402,4 +510,14 @@ mod tests {
         let driver = ClaudeCodeDriver::new(Some(String::new()));
         assert_eq!(driver.cli_path, "claude");
     }
+
+    #[test]
+    fn test_sensitive_env_list_coverage() {
+        // Ensure all major provider keys are in the strip list
+        assert!(SENSITIVE_ENV_EXACT.contains(&"OPENAI_API_KEY"));
+        assert!(SENSITIVE_ENV_EXACT.contains(&"ANTHROPIC_API_KEY"));
+        assert!(SENSITIVE_ENV_EXACT.contains(&"GEMINI_API_KEY"));
+        assert!(SENSITIVE_ENV_EXACT.contains(&"GROQ_API_KEY"));
+        assert!(SENSITIVE_ENV_EXACT.contains(&"DEEPSEEK_API_KEY"));
+    }
 }

crates/openfang-runtime/src/drivers/copilot.rs

@@ -205,6 +205,11 @@ impl CopilotDriver {
             token.base_url.clone()
         };
         super::openai::OpenAIDriver::new(token.token.to_string(), base_url)
+            .with_extra_headers(vec![
+                ("Editor-Version".to_string(), "vscode/1.96.0".to_string()),
+                ("Editor-Plugin-Version".to_string(), "copilot/1.250.0".to_string()),
+                ("Copilot-Integration-Id".to_string(), "vscode-chat".to_string()),
+            ])
     }
 }
 

crates/openfang-runtime/src/drivers/gemini.rs

@@ -92,6 +92,9 @@ struct GeminiInlineData {
 struct GeminiFunctionCallData {
     name: String,
     args: serde_json::Value,
+    /// Gemini 2.5+ thinking models return this on functionCall parts.
+    #[serde(rename = "thoughtSignature", default, skip_serializing_if = "Option::is_none")]
+    thought_signature: Option<String>,
 }
 
 #[derive(Debug, Serialize, Deserialize, Clone)]
@@ -161,8 +164,29 @@ struct GeminiErrorResponse {
 }
 
 #[derive(Debug, Deserialize)]
+#[allow(dead_code)]
 struct GeminiErrorDetail {
     message: String,
+    #[serde(default)]
+    code: Option<u16>,
+    #[serde(default)]
+    status: Option<String>,
+}
+
+/// Parse a Gemini error response body, handling multiple Google API error formats.
+fn parse_gemini_error(body: &str) -> String {
+    if let Ok(e) = serde_json::from_str::<GeminiErrorResponse>(body) {
+        let mut msg = e.error.message;
+        if let Some(status) = e.error.status {
+            msg = format!("{status}: {msg}");
+        }
+        return msg;
+    }
+    // Google sometimes returns bare JSON arrays or HTML error pages
+    if body.starts_with('<') {
+        return "Google API returned an HTML error page — check your API key and model name".to_string();
+    }
+    body.to_string()
 }
 
 // ── Message conversion ─────────────────────────────────────────────────
@@ -202,6 +226,7 @@ fn convert_messages(
                                 function_call: GeminiFunctionCallData {
                                     name: name.clone(),
                                     args: input.clone(),
+                                    thought_signature: None,
                                 },
                             });
                         }
@@ -393,8 +418,8 @@ impl LlmDriver for GeminiDriver {
         let max_retries = 3;
         for attempt in 0..=max_retries {
             let url = format!(
-                "{}/v1beta/models/{}:generateContent",
-                self.base_url, request.model
+                "{}/v1beta/models/{}:generateContent?key={}",
+                self.base_url, request.model, self.api_key.as_str()
             );
             debug!(url = %url, attempt, "Sending Gemini API request");
 
@@ -430,9 +455,13 @@ impl LlmDriver for GeminiDriver {
 
             if !resp.status().is_success() {
                 let body = resp.text().await.unwrap_or_default();
-                let message = serde_json::from_str::<GeminiErrorResponse>(&body)
-                    .map(|e| e.error.message)
-                    .unwrap_or(body);
+                let message = parse_gemini_error(&body);
+                if status == 401 || status == 403 {
+                    return Err(LlmError::AuthenticationFailed(message));
+                }
+                if status == 404 {
+                    return Err(LlmError::ModelNotFound(message));
+                }
                 return Err(LlmError::Api { status, message });
             }
 
@@ -473,8 +502,8 @@ impl LlmDriver for GeminiDriver {
         let max_retries = 3;
         for attempt in 0..=max_retries {
             let url = format!(
-                "{}/v1beta/models/{}:streamGenerateContent?alt=sse",
-                self.base_url, request.model
+                "{}/v1beta/models/{}:streamGenerateContent?alt=sse&key={}",
+                self.base_url, request.model, self.api_key.as_str()
             );
             debug!(url = %url, attempt, "Sending Gemini streaming request");
 
@@ -513,9 +542,13 @@ impl LlmDriver for GeminiDriver {
 
             if !resp.status().is_success() {
                 let body = resp.text().await.unwrap_or_default();
-                let message = serde_json::from_str::<GeminiErrorResponse>(&body)
-                    .map(|e| e.error.message)
-                    .unwrap_or(body);
+                let message = parse_gemini_error(&body);
+                if status == 401 || status == 403 {
+                    return Err(LlmError::AuthenticationFailed(message));
+                }
+                if status == 404 {
+                    return Err(LlmError::ModelNotFound(message));
+                }
                 return Err(LlmError::Api { status, message });
             }
 

crates/openfang-runtime/src/drivers/mod.rs

@@ -14,11 +14,13 @@ pub mod openai;
 use crate::llm_driver::{DriverConfig, LlmDriver, LlmError};
 use openfang_types::model_catalog::{
     AI21_BASE_URL, ANTHROPIC_BASE_URL, CEREBRAS_BASE_URL, COHERE_BASE_URL, DEEPSEEK_BASE_URL,
-    FIREWORKS_BASE_URL, GEMINI_BASE_URL, GROQ_BASE_URL, HUGGINGFACE_BASE_URL, LMSTUDIO_BASE_URL,
+    FIREWORKS_BASE_URL, GEMINI_BASE_URL, GROQ_BASE_URL, HUGGINGFACE_BASE_URL, LEMONADE_BASE_URL,
+    LMSTUDIO_BASE_URL,
     MINIMAX_BASE_URL, MISTRAL_BASE_URL, MOONSHOT_BASE_URL, OLLAMA_BASE_URL, OPENAI_BASE_URL,
     OPENROUTER_BASE_URL, PERPLEXITY_BASE_URL, QIANFAN_BASE_URL, QWEN_BASE_URL,
-    REPLICATE_BASE_URL, SAMBANOVA_BASE_URL, TOGETHER_BASE_URL, VLLM_BASE_URL, XAI_BASE_URL,
-    ZHIPU_BASE_URL, ZHIPU_CODING_BASE_URL,
+    REPLICATE_BASE_URL, SAMBANOVA_BASE_URL, TOGETHER_BASE_URL, VENICE_BASE_URL, VLLM_BASE_URL,
+    VOLCENGINE_BASE_URL, VOLCENGINE_CODING_BASE_URL, XAI_BASE_URL, ZAI_BASE_URL,
+    ZAI_CODING_BASE_URL, ZHIPU_BASE_URL, ZHIPU_CODING_BASE_URL,
 };
 use std::sync::Arc;
 
@@ -88,6 +90,11 @@ fn provider_defaults(provider: &str) -> Option<ProviderDefaults> {
             api_key_env: "LMSTUDIO_API_KEY",
             key_required: false,
         }),
+        "lemonade" => Some(ProviderDefaults {
+            base_url: LEMONADE_BASE_URL,
+            api_key_env: "LEMONADE_API_KEY",
+            key_required: false,
+        }),
         "perplexity" => Some(ProviderDefaults {
             base_url: PERPLEXITY_BASE_URL,
             api_key_env: "PERPLEXITY_API_KEY",
@@ -148,7 +155,7 @@ fn provider_defaults(provider: &str) -> Option<ProviderDefaults> {
             api_key_env: "MOONSHOT_API_KEY",
             key_required: true,
         }),
-        "qwen" | "dashscope" => Some(ProviderDefaults {
+        "qwen" | "dashscope" | "model_studio" => Some(ProviderDefaults {
             base_url: QWEN_BASE_URL,
             api_key_env: "DASHSCOPE_API_KEY",
             key_required: true,
@@ -168,11 +175,36 @@ fn provider_defaults(provider: &str) -> Option<ProviderDefaults> {
             api_key_env: "ZHIPU_API_KEY",
             key_required: true,
         }),
+        "zai" => Some(ProviderDefaults {
+            base_url: ZAI_BASE_URL,
+            api_key_env: "ZHIPU_API_KEY",
+            key_required: true,
+        }),
+        "zai_coding" => Some(ProviderDefaults {
+            base_url: ZAI_CODING_BASE_URL,
+            api_key_env: "ZHIPU_API_KEY",
+            key_required: true,
+        }),
         "qianfan" | "baidu" => Some(ProviderDefaults {
             base_url: QIANFAN_BASE_URL,
             api_key_env: "QIANFAN_API_KEY",
             key_required: true,
         }),
+        "volcengine" | "doubao" => Some(ProviderDefaults {
+            base_url: VOLCENGINE_BASE_URL,
+            api_key_env: "VOLCENGINE_API_KEY",
+            key_required: true,
+        }),
+        "volcengine_coding" => Some(ProviderDefaults {
+            base_url: VOLCENGINE_CODING_BASE_URL,
+            api_key_env: "VOLCENGINE_API_KEY",
+            key_required: true,
+        }),
+        "venice" => Some(ProviderDefaults {
+            base_url: VENICE_BASE_URL,
+            api_key_env: "VENICE_API_KEY",
+            key_required: true,
+        }),
         _ => None,
     }
 }
@@ -309,27 +341,84 @@ pub fn create_driver(config: &DriverConfig) -> Result<Arc<dyn LlmDriver>, LlmErr
         return Ok(Arc::new(openai::OpenAIDriver::new(api_key, base_url)));
     }
 
-    // Unknown provider — if base_url is set, treat as custom OpenAI-compatible
+    // Unknown provider — if base_url is set, treat as custom OpenAI-compatible.
+    // For custom providers, try the convention {PROVIDER_UPPER}_API_KEY as env var
+    // when no explicit api_key was passed. This lets users just set e.g. NVIDIA_API_KEY
+    // in their environment and use provider = "nvidia" without extra config.
     if let Some(ref base_url) = config.base_url {
-        let api_key = config.api_key.clone().unwrap_or_default();
+        let api_key = config.api_key.clone().unwrap_or_else(|| {
+            let env_var = format!("{}_API_KEY", provider.to_uppercase().replace('-', "_"));
+            std::env::var(&env_var).unwrap_or_default()
+        });
         return Ok(Arc::new(openai::OpenAIDriver::new(
             api_key,
             base_url.clone(),
         )));
     }
 
+    // No base_url either — last resort: check if the user set an API key env var
+    // using the convention {PROVIDER_UPPER}_API_KEY. If found, use OpenAI-compatible
+    // driver with a default base URL derived from common patterns.
+    {
+        let env_var = format!("{}_API_KEY", provider.to_uppercase().replace('-', "_"));
+        if let Ok(api_key) = std::env::var(&env_var) {
+            if !api_key.is_empty() {
+                return Err(LlmError::Api {
+                    status: 0,
+                    message: format!(
+                        "Provider '{}' has API key ({} is set) but no base_url configured. \
+                         Add base_url to your [default_model] config or set it in [provider_urls].",
+                        provider, env_var
+                    ),
+                });
+            }
+        }
+    }
+
     Err(LlmError::Api {
         status: 0,
         message: format!(
             "Unknown provider '{}'. Supported: anthropic, gemini, openai, groq, openrouter, \
              deepseek, together, mistral, fireworks, ollama, vllm, lmstudio, perplexity, \
              cohere, ai21, cerebras, sambanova, huggingface, xai, replicate, github-copilot, \
-             codex, claude-code. Or set base_url for a custom OpenAI-compatible endpoint.",
+             venice, codex, claude-code. Or set base_url for a custom OpenAI-compatible endpoint.",
             provider
         ),
     })
 }
 
+/// Detect the first available provider by scanning environment variables.
+///
+/// Returns `(provider, model, api_key_env)` for the first provider that has a
+/// configured API key, checked in a user-friendly priority order.
+pub fn detect_available_provider() -> Option<(&'static str, &'static str, &'static str)> {
+    // Priority: popular cloud providers first, then niche, then local
+    const PROBE_ORDER: &[(&str, &str, &str)] = &[
+        ("openai", "gpt-4o", "OPENAI_API_KEY"),
+        ("anthropic", "claude-sonnet-4-20250514", "ANTHROPIC_API_KEY"),
+        ("gemini", "gemini-2.5-flash", "GEMINI_API_KEY"),
+        ("groq", "llama-3.3-70b-versatile", "GROQ_API_KEY"),
+        ("deepseek", "deepseek-chat", "DEEPSEEK_API_KEY"),
+        ("openrouter", "openrouter/google/gemini-2.5-flash", "OPENROUTER_API_KEY"),
+        ("mistral", "mistral-large-latest", "MISTRAL_API_KEY"),
+        ("together", "meta-llama/Llama-3-70b-chat-hf", "TOGETHER_API_KEY"),
+        ("fireworks", "accounts/fireworks/models/llama-v3p1-70b-instruct", "FIREWORKS_API_KEY"),
+        ("xai", "grok-2", "XAI_API_KEY"),
+        ("perplexity", "llama-3.1-sonar-large-128k-online", "PERPLEXITY_API_KEY"),
+        ("cohere", "command-r-plus", "COHERE_API_KEY"),
+    ];
+    for &(provider, model, env_var) in PROBE_ORDER {
+        if std::env::var(env_var).ok().filter(|v| !v.is_empty()).is_some() {
+            return Some((provider, model, env_var));
+        }
+    }
+    // Also check GOOGLE_API_KEY as alias for Gemini
+    if std::env::var("GOOGLE_API_KEY").ok().filter(|v| !v.is_empty()).is_some() {
+        return Some(("gemini", "gemini-2.5-flash", "GOOGLE_API_KEY"));
+    }
+    None
+}
+
 /// List all known provider names.
 pub fn known_providers() -> &'static [&'static str] {
     &[
@@ -360,6 +449,8 @@ pub fn known_providers() -> &'static [&'static str] {
         "zhipu",
         "zhipu_coding",
         "qianfan",
+        "volcengine",
+        "venice",
         "codex",
         "claude-code",
     ]
@@ -455,9 +546,10 @@ mod tests {
         assert!(providers.contains(&"zhipu"));
         assert!(providers.contains(&"zhipu_coding"));
         assert!(providers.contains(&"qianfan"));
+        assert!(providers.contains(&"volcengine"));
         assert!(providers.contains(&"codex"));
         assert!(providers.contains(&"claude-code"));
-        assert_eq!(providers.len(), 29);
+        assert_eq!(providers.len(), 31);
     }
 
     #[test]
@@ -497,4 +589,61 @@ mod tests {
         assert_eq!(d.api_key_env, "HF_API_KEY");
         assert!(d.key_required);
     }
+
+    #[test]
+    fn test_custom_provider_convention_env_var() {
+        // Set NVIDIA_API_KEY env var, then create a custom "nvidia" provider with base_url.
+        // The driver should pick up the key automatically via convention.
+        let unique_key = "test-nvidia-key-12345";
+        std::env::set_var("NVIDIA_API_KEY", unique_key);
+        let config = DriverConfig {
+            provider: "nvidia".to_string(),
+            api_key: None, // not explicitly passed
+            base_url: Some("https://integrate.api.nvidia.com/v1".to_string()),
+        };
+        let driver = create_driver(&config);
+        assert!(driver.is_ok(), "Custom provider with env var convention should succeed");
+        std::env::remove_var("NVIDIA_API_KEY");
+    }
+
+    #[test]
+    fn test_custom_provider_no_key_no_url_errors() {
+        // Custom provider with neither API key nor base_url should error.
+        let config = DriverConfig {
+            provider: "nvidia".to_string(),
+            api_key: None,
+            base_url: None,
+        };
+        let driver = create_driver(&config);
+        assert!(driver.is_err());
+    }
+
+    #[test]
+    fn test_custom_provider_key_no_url_helpful_error() {
+        // Custom provider with key set (via env) but no base_url should give helpful error.
+        let unique_key = "test-nvidia-key-67890";
+        std::env::set_var("NVIDIA_API_KEY", unique_key);
+        let config = DriverConfig {
+            provider: "nvidia".to_string(),
+            api_key: None,
+            base_url: None,
+        };
+        let result = create_driver(&config);
+        assert!(result.is_err());
+        let err = result.err().unwrap().to_string();
+        assert!(err.contains("base_url"), "Error should mention base_url: {}", err);
+        std::env::remove_var("NVIDIA_API_KEY");
+    }
+
+    #[test]
+    fn test_custom_provider_explicit_key_with_url() {
+        // When api_key is explicitly passed, it should be used regardless of env var.
+        let config = DriverConfig {
+            provider: "my-custom-provider".to_string(),
+            api_key: Some("explicit-key".to_string()),
+            base_url: Some("https://api.example.com/v1".to_string()),
+        };
+        let driver = create_driver(&config);
+        assert!(driver.is_ok());
+    }
 }

crates/openfang-runtime/src/drivers/openai.rs

@@ -16,6 +16,7 @@ pub struct OpenAIDriver {
     api_key: Zeroizing<String>,
     base_url: String,
     client: reqwest::Client,
+    extra_headers: Vec<(String, String)>,
 }
 
 impl OpenAIDriver {
@@ -25,8 +26,15 @@ impl OpenAIDriver {
             api_key: Zeroizing::new(api_key),
             base_url,
             client: reqwest::Client::new(),
+            extra_headers: Vec::new(),
         }
     }
+
+    /// Create a driver with additional HTTP headers (e.g. for Copilot IDE auth).
+    pub fn with_extra_headers(mut self, headers: Vec<(String, String)>) -> Self {
+        self.extra_headers = headers;
+        self
+    }
 }
 
 #[derive(Debug, Serialize)]
@@ -39,13 +47,17 @@ struct OaiRequest {
     /// New token limit field required by GPT-5 and o-series reasoning models.
     #[serde(skip_serializing_if = "Option::is_none")]
     max_completion_tokens: Option<u32>,
-    temperature: f32,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    temperature: Option<f32>,
     #[serde(skip_serializing_if = "Vec::is_empty")]
     tools: Vec<OaiTool>,
     #[serde(skip_serializing_if = "Option::is_none")]
     tool_choice: Option<serde_json::Value>,
     #[serde(skip_serializing_if = "std::ops::Not::not")]
     stream: bool,
+    /// Request usage stats in streaming responses (OpenAI extension, supported by Groq et al).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    stream_options: Option<serde_json::Value>,
 }
 
 /// Returns true if a model uses `max_completion_tokens` instead of `max_tokens`.
@@ -58,6 +70,24 @@ fn uses_completion_tokens(model: &str) -> bool {
         || m.starts_with("o4")
 }
 
+/// Returns true if a model rejects the `temperature` parameter.
+///
+/// OpenAI's o-series reasoning models and GPT-5-mini variants only accept
+/// `temperature=1` (the default). Sending any other value causes a 400 error.
+/// We proactively omit `temperature` for these models to avoid wasting a retry.
+fn rejects_temperature(model: &str) -> bool {
+    let m = model.to_lowercase();
+    // o-series reasoning models: o1, o1-mini, o1-preview, o3, o3-mini, o3-pro, o4-mini, etc.
+    m.starts_with("o1")
+        || m.starts_with("o3")
+        || m.starts_with("o4")
+        // GPT-5-mini is a reasoning model that rejects temperature
+        || m.starts_with("gpt-5-mini")
+        || m.starts_with("gpt5-mini")
+        // Catch any model explicitly tagged as "reasoning"
+        || m.contains("-reasoning")
+}
+
 #[derive(Debug, Serialize)]
 struct OaiMessage {
     role: String,
@@ -202,7 +232,9 @@ impl LlmDriver for OpenAIDriver {
                                 has_tool_results = true;
                                 oai_messages.push(OaiMessage {
                                     role: "tool".to_string(),
-                                    content: Some(OaiMessageContent::Text(content.clone())),
+                                    content: Some(OaiMessageContent::Text(
+                                        if content.is_empty() { "(empty)".to_string() } else { content.clone() }
+                                    )),
                                     tool_calls: None,
                                     tool_call_id: Some(tool_use_id.clone()),
                                 });
@@ -250,10 +282,19 @@ impl LlmDriver for OpenAIDriver {
                             _ => {}
                         }
                     }
+                    let has_tool_calls = !tool_calls.is_empty();
                     oai_messages.push(OaiMessage {
                         role: "assistant".to_string(),
+                        // ZHIPU (GLM) rejects assistant messages where content is
+                        // null or omitted when tool_calls are present (error 1214).
+                        // Always send an empty string so every OpenAI-compat
+                        // provider gets a valid payload.
                         content: if text_parts.is_empty() {
-                            None
+                            if has_tool_calls {
+                                Some(OaiMessageContent::Text(String::new()))
+                            } else {
+                                None
+                            }
                         } else {
                             Some(OaiMessageContent::Text(text_parts.join("")))
                         },
@@ -301,10 +342,11 @@ impl LlmDriver for OpenAIDriver {
             messages: oai_messages,
             max_tokens: mt,
             max_completion_tokens: mct,
-            temperature: request.temperature,
+            temperature: if rejects_temperature(&request.model) { None } else { Some(request.temperature) },
             tools: oai_tools,
             tool_choice,
             stream: false,
+            stream_options: None,
         };
 
         let max_retries = 3;
@@ -322,6 +364,9 @@ impl LlmDriver for OpenAIDriver {
                 req_builder = req_builder
                     .header("authorization", format!("Bearer {}", self.api_key.as_str()));
             }
+            for (k, v) in &self.extra_headers {
+                req_builder = req_builder.header(k, v);
+            }
 
             let resp = req_builder
                 .send()
@@ -360,6 +405,18 @@ impl LlmDriver for OpenAIDriver {
                     }
                 }
 
+                // o-series / reasoning models: strip temperature if rejected
+                if status == 400
+                    && body.contains("temperature")
+                    && body.contains("unsupported_parameter")
+                    && oai_request.temperature.is_some()
+                    && attempt < max_retries
+                {
+                    warn!(model = %oai_request.model, "Stripping temperature for this model");
+                    oai_request.temperature = None;
+                    continue;
+                }
+
                 // GPT-5 / o-series: switch from max_tokens to max_completion_tokens
                 if status == 400
                     && body.contains("max_tokens")
@@ -388,6 +445,28 @@ impl LlmDriver for OpenAIDriver {
                     continue;
                 }
 
+                // Model doesn't support function calling — retry without tools
+                // (e.g. GLM-5 on DashScope returns 500 "internal error" when tools are sent)
+                let body_lower = body.to_lowercase();
+                if !oai_request.tools.is_empty()
+                    && attempt < max_retries
+                    && (status == 500
+                        || body_lower.contains("internal error")
+                        || (status == 400
+                            && (body_lower.contains("does not support tools")
+                                || body_lower.contains("tool")
+                                    && body_lower.contains("not supported"))))
+                {
+                    warn!(
+                        model = %oai_request.model,
+                        status,
+                        "Model may not support tools, retrying without tools"
+                    );
+                    oai_request.tools.clear();
+                    oai_request.tool_choice = None;
+                    continue;
+                }
+
                 return Err(LlmError::Api {
                     status,
                     message: body,
@@ -523,7 +602,9 @@ impl LlmDriver for OpenAIDriver {
                         {
                             oai_messages.push(OaiMessage {
                                 role: "tool".to_string(),
-                                content: Some(OaiMessageContent::Text(content.clone())),
+                                content: Some(OaiMessageContent::Text(
+                                    if content.is_empty() { "(empty)".to_string() } else { content.clone() }
+                                )),
                                 tool_calls: None,
                                 tool_call_id: Some(tool_use_id.clone()),
                             });
@@ -550,10 +631,15 @@ impl LlmDriver for OpenAIDriver {
                             _ => {}
                         }
                     }
+                    let has_tool_calls = !tool_calls_out.is_empty();
                     oai_messages.push(OaiMessage {
                         role: "assistant".to_string(),
                         content: if text_parts.is_empty() {
-                            None
+                            if has_tool_calls {
+                                Some(OaiMessageContent::Text(String::new()))
+                            } else {
+                                None
+                            }
                         } else {
                             Some(OaiMessageContent::Text(text_parts.join("")))
                         },
@@ -601,10 +687,11 @@ impl LlmDriver for OpenAIDriver {
             messages: oai_messages,
             max_tokens: mt,
             max_completion_tokens: mct,
-            temperature: request.temperature,
+            temperature: if rejects_temperature(&request.model) { None } else { Some(request.temperature) },
             tools: oai_tools,
             tool_choice,
             stream: true,
+            stream_options: Some(serde_json::json!({"include_usage": true})),
         };
 
         // Retry loop for the initial HTTP request
@@ -623,6 +710,9 @@ impl LlmDriver for OpenAIDriver {
                 req_builder = req_builder
                     .header("authorization", format!("Bearer {}", self.api_key.as_str()));
             }
+            for (k, v) in &self.extra_headers {
+                req_builder = req_builder.header(k, v);
+            }
 
             let resp = req_builder
                 .send()
@@ -662,6 +752,18 @@ impl LlmDriver for OpenAIDriver {
                     }
                 }
 
+                // o-series / reasoning models: strip temperature if rejected
+                if status == 400
+                    && body.contains("temperature")
+                    && body.contains("unsupported_parameter")
+                    && oai_request.temperature.is_some()
+                    && attempt < max_retries
+                {
+                    warn!(model = %oai_request.model, "Stripping temperature for this model (stream)");
+                    oai_request.temperature = None;
+                    continue;
+                }
+
                 // GPT-5 / o-series: switch from max_tokens to max_completion_tokens
                 if status == 400
                     && body.contains("max_tokens")
@@ -690,6 +792,40 @@ impl LlmDriver for OpenAIDriver {
                     continue;
                 }
 
+                // Provider doesn't support stream_options — retry without it
+                if status == 400
+                    && oai_request.stream_options.is_some()
+                    && attempt < max_retries
+                    && (body.contains("stream_options")
+                        || body.contains("stream_option")
+                        || body.contains("Unrecognized request argument"))
+                {
+                    warn!(model = %oai_request.model, "Stripping stream_options (unsupported by provider)");
+                    oai_request.stream_options = None;
+                    continue;
+                }
+
+                // Model doesn't support function calling — retry without tools
+                let body_lower = body.to_lowercase();
+                if !oai_request.tools.is_empty()
+                    && attempt < max_retries
+                    && (status == 500
+                        || body_lower.contains("internal error")
+                        || (status == 400
+                            && (body_lower.contains("does not support tools")
+                                || body_lower.contains("tool")
+                                    && body_lower.contains("not supported"))))
+                {
+                    warn!(
+                        model = %oai_request.model,
+                        status,
+                        "Model may not support tools (stream), retrying without tools"
+                    );
+                    oai_request.tools.clear();
+                    oai_request.tool_choice = None;
+                    continue;
+                }
+
                 return Err(LlmError::Api {
                     status,
                     message: body,
@@ -703,10 +839,13 @@ impl LlmDriver for OpenAIDriver {
             let mut tool_accum: Vec<(String, String, String)> = Vec::new();
             let mut finish_reason: Option<String> = None;
             let mut usage = TokenUsage::default();
+            let mut chunk_count: u32 = 0;
+            let mut sse_line_count: u32 = 0;
 
             let mut byte_stream = resp.bytes_stream();
             while let Some(chunk_result) = byte_stream.next().await {
                 let chunk = chunk_result.map_err(|e| LlmError::Http(e.to_string()))?;
+                chunk_count += 1;
                 buffer.push_str(&String::from_utf8_lossy(&chunk));
 
                 // Process complete lines
@@ -718,6 +857,7 @@ impl LlmDriver for OpenAIDriver {
                         continue;
                     }
 
+                    sse_line_count += 1;
                     let data = match line.strip_prefix("data:") {
                         Some(d) => d.trim_start(),
                         None => continue,
@@ -812,6 +952,33 @@ impl LlmDriver for OpenAIDriver {
                 }
             }
 
+            // Log stream summary for diagnostics
+            let is_empty_stream = text_content.is_empty()
+                && tool_accum.is_empty()
+                && usage.input_tokens == 0
+                && usage.output_tokens == 0;
+            if is_empty_stream {
+                warn!(
+                    chunks = chunk_count,
+                    sse_lines = sse_line_count,
+                    finish = ?finish_reason,
+                    buffer_remaining = buffer.len(),
+                    "SSE stream returned empty: 0 content, 0 tokens — likely a silently failed request"
+                );
+            } else {
+                debug!(
+                    chunks = chunk_count,
+                    sse_lines = sse_line_count,
+                    text_len = text_content.len(),
+                    tool_count = tool_accum.len(),
+                    finish = ?finish_reason,
+                    input_tokens = usage.input_tokens,
+                    output_tokens = usage.output_tokens,
+                    buffer_remaining = buffer.len(),
+                    "SSE stream completed"
+                );
+            }
+
             // Build the final response
             let mut content = Vec::new();
             let mut tool_calls = Vec::new();
@@ -1008,4 +1175,94 @@ mod tests {
         let resp = result.unwrap();
         assert_eq!(resp.tool_calls[0].name, "shell_exec");
     }
+
+    // ----- rejects_temperature tests -----
+
+    #[test]
+    fn test_rejects_temperature_o1_models() {
+        assert!(rejects_temperature("o1"));
+        assert!(rejects_temperature("o1-mini"));
+        assert!(rejects_temperature("o1-mini-2024-09-12"));
+        assert!(rejects_temperature("o1-preview"));
+        assert!(rejects_temperature("o1-preview-2024-09-12"));
+    }
+
+    #[test]
+    fn test_rejects_temperature_o3_models() {
+        assert!(rejects_temperature("o3"));
+        assert!(rejects_temperature("o3-mini"));
+        assert!(rejects_temperature("o3-mini-2025-01-31"));
+        assert!(rejects_temperature("o3-pro"));
+    }
+
+    #[test]
+    fn test_rejects_temperature_o4_models() {
+        assert!(rejects_temperature("o4-mini"));
+        assert!(rejects_temperature("o4-mini-2025-04-16"));
+    }
+
+    #[test]
+    fn test_rejects_temperature_gpt5_mini() {
+        assert!(rejects_temperature("gpt-5-mini"));
+        assert!(rejects_temperature("gpt-5-mini-2025-08-07"));
+        assert!(rejects_temperature("gpt5-mini"));
+        assert!(rejects_temperature("GPT-5-MINI-2025-08-07"));
+    }
+
+    #[test]
+    fn test_rejects_temperature_reasoning_suffix() {
+        assert!(rejects_temperature("some-model-reasoning"));
+        assert!(rejects_temperature("deepseek-r1-reasoning"));
+    }
+
+    #[test]
+    fn test_does_not_reject_temperature_normal_models() {
+        assert!(!rejects_temperature("gpt-4o"));
+        assert!(!rejects_temperature("gpt-4o-mini"));
+        assert!(!rejects_temperature("gpt-5"));
+        assert!(!rejects_temperature("gpt-5-2025-06-01"));
+        assert!(!rejects_temperature("claude-sonnet-4-20250514"));
+        assert!(!rejects_temperature("llama-3.3-70b-versatile"));
+        assert!(!rejects_temperature("deepseek-chat"));
+    }
+
+    // ----- uses_completion_tokens tests -----
+
+    #[test]
+    fn test_uses_completion_tokens_gpt5() {
+        assert!(uses_completion_tokens("gpt-5"));
+        assert!(uses_completion_tokens("gpt-5-mini"));
+        assert!(uses_completion_tokens("gpt-5-mini-2025-08-07"));
+        assert!(uses_completion_tokens("gpt5-mini"));
+    }
+
+    #[test]
+    fn test_uses_completion_tokens_o_series() {
+        assert!(uses_completion_tokens("o1"));
+        assert!(uses_completion_tokens("o1-mini"));
+        assert!(uses_completion_tokens("o3"));
+        assert!(uses_completion_tokens("o3-mini"));
+        assert!(uses_completion_tokens("o3-pro"));
+        assert!(uses_completion_tokens("o4-mini"));
+    }
+
+    #[test]
+    fn test_does_not_use_completion_tokens_normal_models() {
+        assert!(!uses_completion_tokens("gpt-4o"));
+        assert!(!uses_completion_tokens("gpt-4o-mini"));
+        assert!(!uses_completion_tokens("llama-3.3-70b"));
+    }
+
+    // ----- extract_max_tokens_limit tests -----
+
+    #[test]
+    fn test_extract_max_tokens_limit() {
+        let body = r#"max_tokens must be less than or equal to `8192`"#;
+        assert_eq!(extract_max_tokens_limit(body), Some(8192));
+    }
+
+    #[test]
+    fn test_extract_max_tokens_limit_no_match() {
+        assert_eq!(extract_max_tokens_limit("some random error"), None);
+    }
 }