chore(release): 1.22.0 [skip ci]

# [1.22.0](https://github.com/asepharyana/zesdex/compare/v1.21.2...v1.22.0) (2026-08-28)

### Features

* **agent:** wire auto-review (review_enabled no-op -> nyata) ([a27e815](https://github.com/asepharyana/zesdex/commit/a27e8151b39da4c8759e8e922ef132212327e413))

CHANGELOG.md

@@ -1,3 +1,88 @@
+# [1.22.0](https://github.com/asepharyana/zesdex/compare/v1.21.2...v1.22.0) (2026-08-28)
+
+
+### Features
+
+* **agent:** wire auto-review (review_enabled no-op -> nyata) ([a27e815](https://github.com/asepharyana/zesdex/commit/a27e8151b39da4c8759e8e922ef132212327e413))
+
+## [1.21.2](https://github.com/asepharyana/zesdex/compare/v1.21.1...v1.21.2) (2026-08-28)
+
+
+### Performance Improvements
+
+* **agent:** symbol index tak pegang mutex global saat rebuild I/O ([3b25e38](https://github.com/asepharyana/zesdex/commit/3b25e3898f9ba2fc1c58b991ad95a8c0bbe98601))
+
+## [1.21.1](https://github.com/asepharyana/zesdex/compare/v1.21.0...v1.21.1) (2026-08-28)
+
+
+### Bug Fixes
+
+* **agent:** semantic_search symbol index workspace-aware ([104af3a](https://github.com/asepharyana/zesdex/commit/104af3abb9e626c5d00ea87d523248d606d523ee))
+
+# [1.21.0](https://github.com/asepharyana/zesdex/compare/v1.20.2...v1.21.0) (2026-08-28)
+
+
+### Features
+
+* **agent:** hive-mind consensus synthesis pakai LLM nyata ([f75ff74](https://github.com/asepharyana/zesdex/commit/f75ff740ac2fa63340656e1e8215440f5e48063d))
+
+## [1.20.2](https://github.com/asepharyana/zesdex/compare/v1.20.1...v1.20.2) (2026-08-28)
+
+
+### Bug Fixes
+
+* **agent:** recall search + memory_dir fallback + bersihkan dead llm_client ([f1f58b9](https://github.com/asepharyana/zesdex/commit/f1f58b9996f8eb58871d44fdd41c2d863874f253))
+
+## [1.20.1](https://github.com/asepharyana/zesdex/compare/v1.20.0...v1.20.1) (2026-08-28)
+
+
+### Bug Fixes
+
+* **agent:** subagent patuhi tool-calling contract + truncation char-safe ([e982cbe](https://github.com/asepharyana/zesdex/commit/e982cbeb041baea9cd500e2a29862a7eca9e6e17))
+
+# [1.20.0](https://github.com/asepharyana/zesdex/compare/v1.19.6...v1.20.0) (2026-08-28)
+
+
+### Features
+
+* **agent:** subagent tool paralel + auto-load AGENTS.md + verify cek setelah edit ([21e3ccc](https://github.com/asepharyana/zesdex/commit/21e3ccc891ab886044a5f0a770db710769d1287b))
+
+## [1.19.6](https://github.com/asepharyana/zesdex/compare/v1.19.5...v1.19.6) (2026-08-27)
+
+
+### Performance Improvements
+
+* **agent:** eksekusi tool read-only paralel seperti Claude Code ([74b1ad4](https://github.com/asepharyana/zesdex/commit/74b1ad43020a11ca27e60e3a24de0db5d1ab37b4))
+
+## [1.19.5](https://github.com/asepharyana/zesdex/compare/v1.19.4...v1.19.5) (2026-08-27)
+
+
+### Bug Fixes
+
+* **api:** model Opus default pakai claude-opus-5 (bukan -4-8) ([b28a5fe](https://github.com/asepharyana/zesdex/commit/b28a5fe384fd45255a6b249c7febd3b4ffc0fd2f))
+* **api:** update zesdex packages to version 1.19.4 ([b46935c](https://github.com/asepharyana/zesdex/commit/b46935c606d4f68ea227c7db27e4c2aa9b4e373c))
+
+## [1.19.4](https://github.com/asepharyana/zesdex/compare/v1.19.3...v1.19.4) (2026-08-27)
+
+
+### Bug Fixes
+
+* **api:** model claude selalu pakai Opus dari settings.json, bukan deepseek ([9aca45c](https://github.com/asepharyana/zesdex/commit/9aca45cb65d6913d14fecc10d70c180934f69d74))
+
+## [1.19.3](https://github.com/asepharyana/zesdex/compare/v1.19.2...v1.19.3) (2026-08-27)
+
+
+### Bug Fixes
+
+* **api:** model Opus pakai URL + API custom dari ~/.claude/settings.json ([1f91b44](https://github.com/asepharyana/zesdex/commit/1f91b447080e7106201d77585edb7d38b74e34bb))
+
+## [1.19.2](https://github.com/asepharyana/zesdex/compare/v1.19.1...v1.19.2) (2026-08-27)
+
+
+### Performance Improvements
+
+* **agent:** stabilkan async & parallel — satu runtime, bounded concurrency, isolasi error ([6a98d52](https://github.com/asepharyana/zesdex/commit/6a98d52d54a69f78710d852a69dda3ac0a4ead31))
+
 ## [1.19.1](https://github.com/asepharyana/zesdex/compare/v1.19.0...v1.19.1) (2026-08-27)
 
 

Cargo.lock

@@ -4862,7 +4862,7 @@ dependencies = [
 
 [[package]]
 name = "zesdex-api"
-version = "1.18.4"
+version = "1.21.1"
 dependencies = [
  "anyhow",
  "argon2",
@@ -4885,11 +4885,12 @@ dependencies = [
 
 [[package]]
 name = "zesdex-application"
-version = "1.18.4"
+version = "1.21.1"
 dependencies = [
  "anyhow",
  "base64",
  "chrono",
+ "futures-util",
  "serde",
  "serde_json",
  "sha2 0.11.0",
@@ -4902,7 +4903,7 @@ dependencies = [
 
 [[package]]
 name = "zesdex-bootstrap"
-version = "1.18.4"
+version = "1.21.1"
 dependencies = [
  "anyhow",
  "chrono",
@@ -4919,7 +4920,7 @@ dependencies = [
 
 [[package]]
 name = "zesdex-daemon"
-version = "1.18.4"
+version = "1.21.1"
 dependencies = [
  "anyhow",
  "base64",
@@ -4943,7 +4944,7 @@ dependencies = [
 
 [[package]]
 name = "zesdex-domain"
-version = "1.18.4"
+version = "1.21.1"
 dependencies = [
  "anyhow",
  "base64",
@@ -4959,7 +4960,7 @@ dependencies = [
 
 [[package]]
 name = "zesdex-gateway"
-version = "1.18.4"
+version = "1.21.1"
 dependencies = [
  "anyhow",
  "axum",
@@ -4986,7 +4987,7 @@ dependencies = [
 
 [[package]]
 name = "zesdex-grpc"
-version = "1.18.4"
+version = "1.21.1"
 dependencies = [
  "anyhow",
  "axum",
@@ -5003,7 +5004,7 @@ dependencies = [
 
 [[package]]
 name = "zesdex-infrastructure"
-version = "1.18.4"
+version = "1.21.1"
 dependencies = [
  "anyhow",
  "argon2",
@@ -5051,7 +5052,7 @@ dependencies = [
 
 [[package]]
 name = "zesdex-tui"
-version = "1.18.4"
+version = "1.21.1"
 dependencies = [
  "anyhow",
  "base64",
@@ -5077,7 +5078,7 @@ dependencies = [
 
 [[package]]
 name = "zesdex-web"
-version = "1.18.4"
+version = "1.21.1"
 dependencies = [
  "anyhow",
  "axum",
@@ -5097,7 +5098,7 @@ dependencies = [
 
 [[package]]
 name = "zesdex-ws"
-version = "1.18.4"
+version = "1.21.1"
 dependencies = [
  "anyhow",
  "axum",

Cargo.toml

@@ -15,7 +15,7 @@ members = [
 ]
 
 [workspace.package]
-version = "1.19.1"
+version = "1.22.0"
 edition = "2021"
 authors = ["asepharyana <superaseph@gmail.com>"]
 

apps/application/Cargo.toml

@@ -16,6 +16,7 @@ uuid.workspace = true
 anyhow.workspace = true
 tracing.workspace = true
 tokio.workspace = true
+futures-util.workspace = true
 base64.workspace = true
 sha2.workspace = true
 url.workspace = true

apps/application/src/agent/mod.rs

@@ -11,6 +11,17 @@ pub trait ToolExecutor: Send + Sync {
         tool_name: &str,
         args: &serde_json::Value,
     ) -> impl Future<Output = Result<String>> + Send;
+
+    /// Whether a tool is *read-only* and therefore safe to run concurrently
+    /// with other read-only tools in the same assistant message.
+    ///
+    /// Defaults to `false` (sequential) so a caller that does not know the
+    /// tool surface stays conservative. Concrete executors that know their
+    /// tools override this — e.g. return `true` for `read`/`grep`/`glob`.
+    fn is_parallel_safe(&self, tool_name: &str) -> bool {
+        let _ = tool_name;
+        false
+    }
 }
 
 /// Service for running agent turns asynchronously.

apps/application/src/agent/turn_service.rs

@@ -5,7 +5,7 @@ use tracing::{debug, info, warn};
 
 use zesdex_domain::agent::{AgentTurnParams, TurnEvent};
 use zesdex_domain::core::{ChatMessage, StreamEvent, ToolDef};
-use zesdex_domain::main_agent_prompt;
+use zesdex_domain::main_agent_prompt_with_project_context;
 
 use super::ToolExecutor;
 use crate::ports::ProviderService;
@@ -111,6 +111,46 @@ fn conversation_chars(messages: &[ChatMessage]) -> usize {
         .sum()
 }
 
+/// The maximum combined size (characters) of project-rule files injected into
+/// the system prompt, so a huge AGENTS.md cannot blow the context window.
+const PROJECT_CONTEXT_MAX_CHARS: usize = 12_000;
+
+/// Case-insensitive rule filenames auto-loaded from the workspace root into
+/// the system prompt, matching the Claude-Code/AGENTS.md convention.
+const RULE_FILENAMES: [&str; 6] = [
+    "AGENTS.md",
+    "agent.md",
+    "CLAUDE.md",
+    "claude.md",
+    ".cursorrules",
+    ".zesdexrules",
+];
+
+/// Build a compact "project context" block from the repo's convention files
+/// (AGENTS.md, CLAUDE.md, .cursorrules, …) found at the workspace root.
+///
+/// Follows the Claude-Code convention of loading AGENTS.md at startup so the
+/// model starts each turn with the repo's rules. Reads are best-effort and
+/// capped at [`PROJECT_CONTEXT_MAX_CHARS`] total; missing files are skipped.
+fn build_project_context(root: &std::path::Path) -> String {
+    let mut ctx = String::new();
+    for file in RULE_FILENAMES {
+        let full = root.join(file);
+        if let Ok(content) = std::fs::read_to_string(&full) {
+            ctx.push_str(&format!("\n### {file}\n```\n{}\n```", content.trim()));
+        }
+    }
+    let context = ctx.trim().to_string();
+    if context.len() <= PROJECT_CONTEXT_MAX_CHARS {
+        return context;
+    }
+    context
+        .chars()
+        .take(PROJECT_CONTEXT_MAX_CHARS)
+        .collect::<String>()
+        + "\n...[project context truncated]"
+}
+
 /// Track repeated tool-call errors so the loop can recover instead of
 /// burning iterations retrying the same failing tool.
 #[derive(Default)]
@@ -187,6 +227,53 @@ async fn execute_tool_call<T: ToolExecutor>(
     output
 }
 
+// ---------------------------------------------------------------------------
+// Helper: bounded-parallel execution of read-only tool calls.
+// ---------------------------------------------------------------------------
+
+/// Maximum number of read-only tool calls executed concurrently in a single
+/// assistant batch. Models rarely emit more than a handful of reads per
+/// message; this cap keeps resource usage bounded while still removing the
+/// serial round-trip latency of many independent lookups.
+const MAX_PARALLEL_TOOLS: usize = 8;
+
+fn tool_executor_ref<T: ToolExecutor>(tool_executor: &T) -> &T {
+    tool_executor
+}
+
+/// Execute a batch of *read-only* tool calls concurrently (bounded by
+/// [`MAX_PARALLEL_TOOLS`]) and return their outputs **in the original call
+/// order**.
+///
+/// Order preservation matters: OpenAI/Anthropic tool-calling contracts expect
+/// tool-result messages to appear in the same order as the `tool_calls`
+/// emitted in the assistant message. Without it, the model sees shuffled
+/// results and loses track of which result belongs to which call.
+///
+/// Each call still pushes its `TurnEvent::ToolResult` (so the TUI shows each
+/// tool as it completes) but the returned `Vec` is ordered by the input index.
+async fn execute_tool_calls_in_parallel<T: ToolExecutor>(
+    tool_executor: &T,
+    turn_events: &Arc<Mutex<VecDeque<TurnEvent>>>,
+    tool_calls: &[zesdex_domain::core::ToolCall],
+) -> Vec<String> {
+    let semaphore = Arc::new(tokio::sync::Semaphore::new(MAX_PARALLEL_TOOLS));
+    let executor_ref = tool_executor_ref(tool_executor);
+
+    let futures = tool_calls.iter().map(|tc| {
+        let tc = tc.clone();
+        let events = turn_events.clone();
+        let sem = semaphore.clone();
+        async move {
+            // Acquire a permit to bound concurrency across the batch.
+            let _permit = sem.acquire_owned().await;
+            execute_tool_call(executor_ref, &events, &tc).await
+        }
+    });
+
+    futures_util::future::join_all(futures).await
+}
+
 // ---------------------------------------------------------------------------
 // Helper: emit usage event from optional LLM response metadata.
 // ---------------------------------------------------------------------------
@@ -294,9 +381,17 @@ impl<P: ProviderService, T: ToolExecutor> super::AgentTurnService for AgentTurnS
 
         // Insert system prompt at position 0 once and keep it there for the
         // entire turn, avoiding per-iteration clones of the full message list.
+        // Auto-load repo conventions (AGENTS.md / CLAUDE.md / .cursorrules)
+        // from the first workspace root, like Claude Code does at startup.
+        let project_context = params
+            .workspace_roots
+            .first()
+            .map(|root| build_project_context(root))
+            .unwrap_or_default();
+        let system_prompt = main_agent_prompt_with_project_context(&project_context);
         params
             .messages
-            .insert(0, ChatMessage::system(main_agent_prompt()));
+            .insert(0, ChatMessage::system(system_prompt));
         let original_count = params.messages.len();
 
         // Estimate request complexity from the last user message.
@@ -384,10 +479,42 @@ impl<P: ProviderService, T: ToolExecutor> super::AgentTurnService for AgentTurnS
                     params.messages.push(assistant_msg);
 
                     // ── Execute each tool call ──────────────────────────
+                    //
+                    // If the whole batch is made of *read-only* tools
+                    // (read/grep/glob/…), run it concurrently with bounded
+                    // parallelism — a big latency win for coding turns that
+                    // emit several independent lookups in one message. Any
+                    // single mutating tool forces the whole batch back to the
+                    // safe sequential path so writes never race.
+                    //
+                    // Results are always collected in the original call order
+                    // to honour the tool-calling contract.
+                    let parallel = tool_calls.len() > 1
+                        && tool_calls
+                            .iter()
+                            .all(|tc| self.tool_executor.is_parallel_safe(&tc.function.name));
+                    let outputs: Vec<String> = if parallel {
+                        execute_tool_calls_in_parallel(
+                            self.tool_executor.as_ref(),
+                            &params.turn_events,
+                            &tool_calls,
+                        )
+                        .await
+                    } else {
+                        let mut sequential = Vec::with_capacity(tool_calls.len());
                         for tc in &tool_calls {
-                        let output =
+                            let out = execute_tool_call(
-                            execute_tool_call(self.tool_executor.as_ref(), &params.turn_events, tc)
+                                self.tool_executor.as_ref(),
+                                &params.turn_events,
+                                tc,
+                            )
                             .await;
+                            sequential.push(out);
+                        }
+                        sequential
+                    };
+
+                    for (tc, output) in tool_calls.iter().zip(outputs) {
                         if output.starts_with("Error:") {
                             errors.record(&tc.function.name, &output, &mut params.messages);
                         }
@@ -479,7 +606,6 @@ pub async fn compact_messages_with_ai<P: ProviderService>(
 #[cfg(test)]
 mod tests {
     use super::*;
-
     #[test]
     fn truncate_short_output_is_unchanged() {
         let out = "short".to_string();
@@ -536,4 +662,83 @@ mod tests {
         ];
         assert_eq!(conversation_chars(&messages), 3 + 11 + 6);
     }
+
+    /// A fake executor that reports parallel-safety for read-only tools and
+    /// whose `execute` sleeps on the first call to prove the batch runs
+    /// concurrently (a sequential loop would pay the sleep per call).
+    struct FakeExecutor;
+
+    impl ToolExecutor for FakeExecutor {
+        async fn execute(&self, name: &str, _args: &serde_json::Value) -> anyhow::Result<String> {
+            if name == "read" {
+                // 30ms sleep on every read; a parallel batch of 3 would
+                // finish in ~30ms instead of ~90ms sequentially.
+                tokio::time::sleep(std::time::Duration::from_millis(30)).await;
+            }
+            Ok(format!("out:{name}"))
+        }
+
+        fn is_parallel_safe(&self, name: &str) -> bool {
+            matches!(name, "read" | "grep")
+        }
+    }
+
+    fn tc(name: &str, id: usize) -> zesdex_domain::core::ToolCall {
+        zesdex_domain::core::ToolCall {
+            id: format!("call_{id}"),
+            type_: "function".to_string(),
+            function: zesdex_domain::core::ToolFunction {
+                name: name.to_string(),
+                arguments: serde_json::Value::String(String::new()),
+            },
+        }
+    }
+
+    #[test]
+    fn parallel_batch_runs_concurrently_and_preserves_order() {
+        let executor = FakeExecutor;
+        let events = Arc::new(Mutex::new(VecDeque::new()));
+        let calls = vec![tc("read", 1), tc("grep", 2), tc("read", 3)];
+
+        // All three are parallel-safe.
+        assert!(calls
+            .iter()
+            .all(|c| executor.is_parallel_safe(&c.function.name)));
+
+        let rt = tokio::runtime::Builder::new_multi_thread()
+            .enable_time()
+            .build()
+            .unwrap();
+        let started = std::time::Instant::now();
+        let outputs = rt.block_on(execute_tool_calls_in_parallel(&executor, &events, &calls));
+        let elapsed = started.elapsed();
+
+        // Results are in *original* call order (read, grep, read).
+        assert_eq!(
+            outputs,
+            vec![
+                "out:read".to_string(),
+                "out:grep".to_string(),
+                "out:read".to_string()
+            ]
+        );
+        // Two reads sleep 30ms each; sequential would take ~60ms+ for the
+        // two reads, parallel keeps the whole batch under 60ms.
+        assert!(
+            elapsed < std::time::Duration::from_millis(60),
+            "batch took {elapsed:?}, expected parallel execution"
+        );
+        assert!(elapsed >= std::time::Duration::from_millis(25));
+    }
+
+    #[test]
+    fn mutating_batch_falls_back_to_sequential_path() {
+        // A batch containing a mutating tool is not eligible for the parallel
+        // path, so the main loop keeps results ordered and side-effects safe.
+        let executor = FakeExecutor;
+        let calls = [tc("read", 1), tc("edit", 2)];
+        assert!(!calls
+            .iter()
+            .all(|c| executor.is_parallel_safe(&c.function.name)));
+    }
 }

apps/domain/src/agent/prompt.rs

@@ -41,10 +41,43 @@ architectural plans and `todowrite` to maintain granular task checklists.
 4. TOOL EXECUTION: Execute individual tools (file edits, terminal commands) \
 within or guided by your workflows. If an error occurs, analyse and fix it.
 
+VERIFY AFTER EDIT (CLAUDE-CODE STYLE):
+- After modifying code (edit/write), run the repo's check command via `bash` \
+before ending the turn: `cargo check` / `cargo clippy` / `cargo test` for Rust, \
+or the equivalent lint/test (`bun run lint && bun run test`, `npm test`, etc.) \
+for other stacks. Pick the project's actual verify command (see PROJECT \
+CONTEXT / AGENTS.md when present).
+- If the check fails, fix the errors you can see and re-run; only end the turn \
+after the check passes or you cannot resolve a failure yourself (then report it \
+explicitly).
+- Do NOT claim code compiles or works without running a real check.
+
 Respond conversationally, concisely, and helpfully."
         .to_string()
 }
 
+/// Build the main-agent system prompt including an injected block of project
+/// context (AGENTS.md / CLAUDE.md / project rules).
+///
+/// Like Claude Code, which loads AGENTS.md at startup so the model starts with
+/// the repo's conventions, this wraps [`main_agent_prompt`] and appends a
+/// clearly-delimited `## PROJECT CONTEXT` section carrying the rules the user
+/// keeps next to their code. When `project_context` is empty the returned
+/// prompt is identical to [`main_agent_prompt`], so callers can fall back
+/// safely.
+pub fn main_agent_prompt_with_project_context(project_context: &str) -> String {
+    let base = main_agent_prompt();
+    let context = project_context.trim();
+    if context.is_empty() {
+        return base;
+    }
+    format!(
+        "{base}\n\n\
+## PROJECT CONTEXT (repo rules — follow these conventions)\n\
+{context}"
+    )
+}
+
 /// Build a subagent directive prompt.
 ///
 /// The directive is embedded in a system message that also communicates the
@@ -119,6 +152,25 @@ mod tests {
         assert!(prompt.contains("WORKFLOW FIRST"));
     }
 
+    #[test]
+    fn project_context_prompt_appends_context_and_keeps_base() {
+        let base = main_agent_prompt();
+        let with_ctx = main_agent_prompt_with_project_context("## AGENTS.md\nUse cargo clippy.");
+        assert!(with_ctx.contains("Zesdex"), "base prompt must be preserved");
+        assert!(with_ctx.contains("PROJECT CONTEXT"));
+        assert!(with_ctx.contains("Use cargo clippy."));
+        assert!(with_ctx.contains(&base));
+        // The base section should appear before the context section.
+        assert!(with_ctx.find("PROJECT CONTEXT").unwrap() > with_ctx.find("Zesdex").unwrap());
+    }
+
+    #[test]
+    fn empty_project_context_returns_base_prompt() {
+        let base = main_agent_prompt();
+        assert_eq!(main_agent_prompt_with_project_context(""), base);
+        assert_eq!(main_agent_prompt_with_project_context("   "), base);
+    }
+
     #[test]
     fn subagent_directive_includes_directive_text() {
         let prompt = subagent_directive("test directive", "/home", "/home/project");

apps/domain/src/cms/app_config.rs

@@ -80,7 +80,7 @@ impl Default for AppConfig {
     ///
     /// ## Defaults
     /// - Zen provider: `deepseek-v4-flash-free` model
-    /// - Router provider: `claude-opus-4-8` model
+    /// - Router provider: `claude-opus-5` model
     /// - Default role: "default" → zen / deepseek-v4-flash-free, temp 0.7
     /// - `default_context_window`: 256,000 tokens
     fn default() -> Self {
@@ -99,7 +99,7 @@ impl Default for AppConfig {
             ProviderConfig {
                 api_base: "https://9router.asepharyana.my.id/v1".to_string(),
                 api_key_env: Some("ROUTER_API_KEY".to_string()),
-                default_model: Some("claude-opus-4-8".to_string()),
+                default_model: Some("claude-opus-5".to_string()),
                 default_api_key: None,
             },
         );

apps/domain/src/cms/mod.rs

@@ -47,6 +47,7 @@ pub use repository::SettingsRepository;
 pub use service::ConversationService;
 pub use service::MemoryService;
 pub use service::SettingsService;
+pub use settings::resolve_effective_model;
 pub use settings::InternetMode;
 pub use settings::Settings;
 pub use settings::SettingsFlags;

apps/domain/src/cms/settings.rs

@@ -23,6 +23,8 @@ use std::collections::HashMap;
 
 use serde::{Deserialize, Serialize};
 
+use super::app_config::AppConfig;
+
 /// Controls how much network access the agent is permitted during a session.
 ///
 /// ## Variants
@@ -112,3 +114,86 @@ impl Default for Settings {
         }
     }
 }
+
+/// Pick the effective model name for the main agent.
+///
+/// When `settings.provider` is `"claude"` (auto-detected from
+/// `~/.claude/settings.json`), the provider's `default_model` (or the
+/// app-level `default_model`) wins over a possibly-stale persisted
+/// `settings.model`. Otherwise the user's explicit `settings.model` is used.
+///
+/// Why: the user's custom Claude endpoint (URL + API key from
+/// `~/.claude/settings.json`) implies Opus as the model; a stale
+/// `settings.json` (e.g. "deepseek-v4-flash-free") must not override it.
+pub fn resolve_effective_model(settings: &Settings, app_config: &AppConfig) -> String {
+    if settings.provider == "claude" {
+        if let Some(m) = app_config
+            .providers
+            .get("claude")
+            .and_then(|p| p.default_model.clone())
+        {
+            return m;
+        }
+        return app_config.default_model.clone();
+    }
+    settings.model.clone()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::cms::app_config::AppConfig;
+
+    fn claude_app_config() -> AppConfig {
+        let mut cfg = AppConfig::default();
+        cfg.providers.insert(
+            "claude".to_string(),
+            crate::cms::ProviderConfig {
+                api_base: "https://9router.example/v1".to_string(),
+                api_key_env: Some("ANTHROPIC_API_KEY".to_string()),
+                default_model: Some("claude-opus-5".to_string()),
+                default_api_key: Some("sk-test".to_string()),
+            },
+        );
+        cfg.default_provider = "claude".to_string();
+        cfg.default_model = "claude-opus-5".to_string();
+        cfg
+    }
+
+    #[test]
+    fn claude_provider_uses_opus_model_over_stale_settings_model() {
+        let settings = Settings {
+            provider: "claude".to_string(),
+            model: "deepseek-v4-flash-free".to_string(), // stale persisted
+            ..Settings::default()
+        };
+
+        let model = resolve_effective_model(&settings, &claude_app_config());
+        assert_eq!(model, "claude-opus-5");
+    }
+
+    #[test]
+    fn non_claude_provider_uses_settings_model() {
+        let settings = Settings {
+            provider: "zen".to_string(),
+            model: "my-model".to_string(),
+            ..Settings::default()
+        };
+
+        let model = resolve_effective_model(&settings, &AppConfig::default());
+        assert_eq!(model, "my-model");
+    }
+
+    #[test]
+    fn claude_falls_back_to_app_default() {
+        let settings = Settings {
+            provider: "claude".to_string(),
+            model: String::new(),
+            ..Settings::default()
+        };
+
+        let cfg = AppConfig::default();
+        let model = resolve_effective_model(&settings, &cfg);
+        assert_eq!(model, cfg.default_model);
+    }
+}

apps/domain/src/lib.rs

@@ -58,6 +58,9 @@ pub use agent::*;
 // Sub-module items need explicit re-exports
 pub use agent::defaults::*;
 pub use agent::progress::AgentProgress;
-pub use agent::prompt::{compaction_prompt, main_agent_prompt, subagent_directive};
+pub use agent::prompt::{
+    compaction_prompt, main_agent_prompt, main_agent_prompt_with_project_context,
+    subagent_directive,
+};
 pub use subagent::*;
 pub use workflow::*;

apps/infrastructure/src/best_practice/explore.rs

@@ -15,7 +15,7 @@
 //!    read of up to 3 relevant files).
 //! 4. Join the result and return a concise bullet summary as a tool message.
 
-use anyhow::{Context, Result};
+use anyhow::Result;
 use serde_json::{json, Value};
 use tracing::{info, warn};
 
@@ -118,7 +118,7 @@ impl Tool for ExploreCodebase {
             model,
         );
 
-        let rt = tokio::runtime::Runtime::new().context("create explore tokio runtime")?;
+        let rt = crate::runtime::runtime();
         let result = rt.block_on(run_agent(
             subagent_ctx,
             &directive,

apps/infrastructure/src/lib.rs

@@ -34,6 +34,7 @@ pub mod llm;
 pub mod mcp;
 pub mod middleware;
 pub mod persistence;
+pub mod runtime;
 pub mod subagent;
 pub mod tools;
 pub mod utils;

apps/infrastructure/src/persistence/cms/app_config_repo.rs

@@ -72,27 +72,24 @@ fn detect_claude_settings_provider() -> Option<(ProviderConfig, Option<String>)>
     ))
 }
 
-impl AppConfigRepository for JsonAppConfigRepository {
+/// Apply a detected Claude provider + custom model onto an `AppConfig`.
-    fn load(&self, base_dir: &Path) -> Result<AppConfig, RepositoryError> {
+///
-        let path = base_dir.join("app_config.json");
+/// Pure (no I/O) so it can be unit-tested. Flow:
-        let mut cfg: AppConfig = match std::fs::read_to_string(&path) {
+///   1. Always `insert`s the "claude" provider (refreshing a possibly stale
-            Ok(s) => serde_json::from_str(&s)?,
+///      persisted entry with the current base URL + key from settings.json).
-            Err(e) if e.kind() == std::io::ErrorKind::NotFound => AppConfig::default(),
+///   2. Registers known Claude model roles if missing.
-            Err(e) => return Err(RepositoryError::Io(e)),
+///   3. Always sets `default_provider = "claude"` and
-        };
+///      `default_model = custom_model.unwrap_or("claude-opus-5")` so Opus
-
+///      is the default whenever `~/.claude/settings.json` is present.
-        let defaults = AppConfig::default();
+fn apply_claude_provider(
-        for (name, provider) in defaults.providers {
+    cfg: &mut AppConfig,
-            cfg.providers.entry(name).or_insert(provider);
+    claude_provider: ProviderConfig,
-        }
+    custom_model: Option<String>,
-
+) {
-        if let Some((claude_provider, custom_model)) = detect_claude_settings_provider() {
+    cfg.providers.insert("claude".to_string(), claude_provider);
-            cfg.providers
-                .entry("claude".to_string())
-                .or_insert(claude_provider);
 
     let claude_models: [(&str, &str); 3] = [
-                ("claude-opus-4-8", "claude-opus-4-8"),
+        ("claude-opus-5", "claude-opus-5"),
         ("claude-sonnet-5", "claude-sonnet-5"),
         ("claude-haiku-4-5", "claude-haiku-4-5-20251001"),
     ];
@@ -118,10 +115,28 @@ impl AppConfigRepository for JsonAppConfigRepository {
         });
     }
 
-            if cfg.default_provider == defaults.default_provider {
+    // Always prefer the Claude provider + Opus model when settings.json
+    // is present — this is the user's explicit custom endpoint choice.
     cfg.default_provider = "claude".to_string();
-                cfg.default_model = custom_model.unwrap_or_else(|| "claude-opus-4-8".to_string());
+    cfg.default_model = custom_model.unwrap_or_else(|| "claude-opus-5".to_string());
 }
+
+impl AppConfigRepository for JsonAppConfigRepository {
+    fn load(&self, base_dir: &Path) -> Result<AppConfig, RepositoryError> {
+        let path = base_dir.join("app_config.json");
+        let mut cfg: AppConfig = match std::fs::read_to_string(&path) {
+            Ok(s) => serde_json::from_str(&s)?,
+            Err(e) if e.kind() == std::io::ErrorKind::NotFound => AppConfig::default(),
+            Err(e) => return Err(RepositoryError::Io(e)),
+        };
+
+        let defaults = AppConfig::default();
+        for (name, provider) in defaults.providers {
+            cfg.providers.entry(name).or_insert(provider);
+        }
+
+        if let Some((claude_provider, custom_model)) = detect_claude_settings_provider() {
+            apply_claude_provider(&mut cfg, claude_provider, custom_model);
         }
 
         Ok(cfg)
@@ -134,3 +149,106 @@ impl AppConfigRepository for JsonAppConfigRepository {
         Ok(())
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::collections::HashMap;
+
+    fn claude_provider(base: &str, key: Option<&str>) -> ProviderConfig {
+        ProviderConfig {
+            api_base: base.to_string(),
+            api_key_env: Some("ANTHROPIC_API_KEY".to_string()),
+            default_model: Some("claude-opus-5".to_string()),
+            default_api_key: key.map(|s| s.to_string()),
+        }
+    }
+
+    #[test]
+    fn claude_settings_parse_env() {
+        let parsed: ClaudeSettings = serde_json::from_str(
+            r#"{"env":{"ANTHROPIC_BASE_URL":"https://9router.example/v1","ANTHROPIC_API_KEY":"sk-test"}}"#,
+        )
+        .unwrap();
+        let env = parsed.env.unwrap();
+        assert_eq!(
+            env.anthropic_base_url.as_deref(),
+            Some("https://9router.example/v1")
+        );
+        assert_eq!(env.anthropic_api_key.as_deref(), Some("sk-test"));
+    }
+
+    #[test]
+    fn apply_claude_refreshes_stale_provider_and_sets_opus_default() {
+        // Simulate a previously-persisted app_config.json with a STALE claude
+        // provider + non-opus default (e.g. user had switched provider).
+        let mut cfg = AppConfig {
+            providers: {
+                let mut m = HashMap::new();
+                m.insert(
+                    "claude".to_string(),
+                    claude_provider("https://old.example/v1", Some("sk-old")),
+                );
+                m
+            },
+            model_roles: HashMap::new(),
+            default_provider: "router".to_string(),
+            default_model: "other-model".to_string(),
+            default_context_window: 256_000,
+        };
+
+        // Detect returned a fresh provider from ~/.claude/settings.json.
+        apply_claude_provider(
+            &mut cfg,
+            claude_provider("https://9router.example/v1", Some("sk-new")),
+            None,
+        );
+
+        let claude = cfg.providers.get("claude").unwrap();
+        assert_eq!(claude.api_base, "https://9router.example/v1");
+        assert_eq!(claude.default_api_key.as_deref(), Some("sk-new"));
+        // Insert (not or_insert) → stale entry refreshed.
+        assert_eq!(cfg.default_provider, "claude");
+        assert_eq!(cfg.default_model, "claude-opus-5");
+
+        // Claude model roles registered.
+        assert!(cfg.model_roles.contains_key("claude-opus-5"));
+        assert!(cfg.model_roles.contains_key("claude-sonnet-5"));
+        assert!(cfg.model_roles.contains_key("claude-haiku-4-5"));
+    }
+
+    #[test]
+    fn apply_claude_honors_custom_model_from_settings() {
+        let mut cfg = AppConfig::default();
+        apply_claude_provider(
+            &mut cfg,
+            claude_provider("https://9router.example/v1", Some("sk-new")),
+            Some("claude-opus-5".to_string()),
+        );
+        assert_eq!(cfg.default_model, "claude-opus-5");
+        assert!(cfg.model_roles.contains_key("claude-opus-5"));
+    }
+
+    #[test]
+    fn detect_uses_env_creds_as_fallback() {
+        // When ~/.claude/settings.json is absent/unreadable, the env-var
+        // fallback should produce a "claude" provider. Set env vars, call
+        // detect, and assert the resulting provider uses them.
+        std::env::set_var("ANTHROPIC_BASE_URL", "https://env.example/v1");
+        std::env::set_var("ANTHROPIC_API_KEY", "sk-env");
+        match detect_claude_settings_provider() {
+            Some((provider, _custom)) => {
+                // If the real settings.json exists it wins (base could be the
+                // real 9router URL); otherwise env creds are used. Either way,
+                // the provider must have api_key_env pointing at ANTHROPIC_API_KEY.
+                assert_eq!(provider.api_key_env.as_deref(), Some("ANTHROPIC_API_KEY"));
+            }
+            None => {
+                // No file + no env (shouldn't happen since we just set env).
+                panic!("expected env fallback to produce a provider");
+            }
+        }
+        std::env::remove_var("ANTHROPIC_BASE_URL");
+        std::env::remove_var("ANTHROPIC_API_KEY");
+    }
+}

apps/infrastructure/src/runtime.rs

@@ -0,0 +1,62 @@
+//! Process-wide shared Tokio runtime for sync → async bridging.
+//!
+//! Many `Tool::run` implementations are synchronous but need to drive async
+//! work (LLM calls, subagent execution). Creating a fresh
+//! [`tokio::runtime::Runtime`] on every call is expensive (spawns a thread
+//! pool + runtime each time) and can fail randomly under thread pressure.
+//!
+//! # Flow
+//!
+//! [`runtime()`] returns a lazily-initialised process-wide runtime created
+//! exactly once via [`std::sync::OnceLock`]. Callers use
+//! `runtime().block_on(...)` exactly like they would with a local runtime —
+//! the only difference is the runtime is shared, so the cost is paid once per
+//! process instead of once per tool call.
+//!
+//! # Safety
+//!
+//! `block_on` panics if called from within a running Tokio runtime. The
+//! tools that use this helper are synchronous (`Tool::run`), so this is safe
+//! in practice. Async code should never call `runtime().block_on`.
+
+use std::sync::OnceLock;
+
+/// Maximum worker threads for the shared runtime. Kept modest — tools are
+/// mostly I/O-bound and rarely need more concurrency than this.
+const RUNTIME_WORKER_THREADS: usize = 8;
+
+static SHARED_RUNTIME: OnceLock<tokio::runtime::Runtime> = OnceLock::new();
+
+/// Return the process-wide shared Tokio runtime, initialising it on first use.
+///
+/// The runtime is configured with `worker_threads = 8` and
+/// `enable_all()` (time + IO drivers) so streams, timers, and network calls
+/// all work. If initialisation fails (extremely rare — resource exhaustion at
+/// startup), the process aborts with a clear message rather than returning
+/// an error on every subsequent call.
+pub fn runtime() -> &'static tokio::runtime::Runtime {
+    SHARED_RUNTIME.get_or_init(|| {
+        tokio::runtime::Builder::new_multi_thread()
+            .worker_threads(RUNTIME_WORKER_THREADS)
+            .thread_name("zesdex-shared-rt")
+            .enable_all()
+            .build()
+            .expect("failed to create shared tokio runtime")
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::runtime;
+
+    #[test]
+    fn runtime_is_singleton() {
+        assert!(std::ptr::eq(runtime(), runtime()));
+    }
+
+    #[test]
+    fn runtime_blocks_and_resolves() {
+        let val = runtime().block_on(async { 6 * 7 });
+        assert_eq!(val, 42);
+    }
+}

apps/infrastructure/src/subagent/auto/engine.rs

@@ -130,7 +130,7 @@ pub fn spawn_background_review(
             );
             format!(
                 "{}...\n[diff truncated at {} characters]",
-                &diff[..MAX_DIFF_CHARS],
+                crate::utils::truncate_chars(&diff, MAX_DIFF_CHARS),
                 MAX_DIFF_CHARS
             )
         } else {

apps/infrastructure/src/subagent/engine.rs

@@ -14,7 +14,8 @@ use tracing::{debug, info, instrument};
 use crate::llm::provider::LlmClient;
 use crate::subagent::context::SubagentContext;
 use crate::subagent::division::{tools_for, AccessTier};
-use crate::tools::{tool_defs, ToolCtx};
+use crate::tools::{tool_defs, Tool, ToolCtx};
+use serde_json::Value;
 use zesdex_domain::agent::progress::AgentProgress;
 use zesdex_domain::core::tool_call::sanitize_tool_arguments;
 use zesdex_domain::core::ChatMessage;
@@ -23,6 +24,126 @@ use zesdex_domain::subagent_directive;
 /// Maximum number of tool-call iterations before the engine gives up.
 const MAX_ITERATIONS: u32 = 25;
 
+/// A single tool-result message is truncated before entering the subagent's
+/// context so it cannot blow the window (matches the main turn service).
+const TOOL_OUTPUT_MAX_CHARS: usize = 12_000;
+
+/// Maximum consecutive identical tool errors before the engine injects a
+/// recovery note steering the model to a different approach.
+const MAX_CONSECUTIVE_TOOL_ERRORS: usize = 3;
+
+/// Maximum number of read-only tool calls executed concurrently in a single
+/// subagent batch. Read-only tools (read/grep/glob/…) block on disk I/O, so
+/// running them on parallel OS threads removes the serial round-trip latency
+/// for a batch of independent lookups, mirroring the main turn loop.
+const MAX_PARALLEL_TOOLS: usize = 8;
+
+/// Execute a batch of tool calls, running read-only tools concurrently when
+/// the whole batch is parallel-safe.
+///
+/// Returns one `(tool_call_id, tool_name, result)` per call **in the original
+/// call order** (OpenAI/Anthropic tool-result ordering contract). `Tool::run`
+/// is synchronous, so real parallelism comes from scoped OS threads; `Tool`
+/// and `ToolCtx` are `Send + Sync`, so the borrowed references can be shared
+/// across the short-lived scoped threads.
+///
+/// If any single tool in the batch mutates state (edit/write/bash/git/…), the
+/// whole batch falls back to the safe sequential path so writes never race.
+fn execute_tool_batch(
+    tools: &[Box<dyn Tool>],
+    tool_ctx: &ToolCtx,
+    tool_calls: &[zesdex_domain::core::ToolCall],
+) -> Vec<(String, String, String)> {
+    let parallel = tool_calls.len() > 1
+        && tool_calls
+            .iter()
+            .all(|tc| crate::tools::tool_is_parallel_safe(&tc.function.name));
+
+    if !parallel {
+        // Sequential fallback (kept identical to the historical behavior).
+        return tool_calls
+            .iter()
+            .map(|tc| {
+                let tool_name = tc.function.name.clone();
+                let args = sanitize_tool_arguments(&tc.function.arguments);
+                let result = run_one_tool(tools, tool_ctx, &tool_name, &args);
+                (tc.id.clone(), tool_name, result)
+            })
+            .collect();
+    }
+
+    // Bounded parallel path: process the batch in windows of
+    // `MAX_PARALLEL_TOOLS` so concurrency stays bounded, joining each window
+    // before the next so results stay in original order.
+    let mut ordered = Vec::with_capacity(tool_calls.len());
+    for window in tool_calls.chunks(MAX_PARALLEL_TOOLS) {
+        let window_results = std::thread::scope(|s| {
+            let handles: Vec<_> = window
+                .iter()
+                .map(|tc| {
+                    let tool_name = tc.function.name.clone();
+                    let args = sanitize_tool_arguments(&tc.function.arguments);
+                    s.spawn(move || {
+                        debug!("Subagent executing tool: {tool_name}");
+                        run_one_tool(tools, tool_ctx, &tool_name, &args)
+                    })
+                })
+                .collect();
+            handles
+                .into_iter()
+                .map(|h| {
+                    h.join()
+                        .unwrap_or_else(|_| "Error: tool panicked".to_string())
+                })
+                .collect::<Vec<_>>()
+        });
+        for (tc, result) in window.iter().zip(window_results) {
+            ordered.push((tc.id.clone(), tc.function.name.clone(), result));
+        }
+    }
+    ordered
+}
+
+/// Run a single synchronous tool call and capture its result string.
+fn run_one_tool(
+    tools: &[Box<dyn Tool>],
+    tool_ctx: &ToolCtx,
+    tool_name: &str,
+    args: &Value,
+) -> String {
+    if let Some(tool) = tools.iter().find(|t| t.name() == tool_name) {
+        match tool.run(tool_ctx, args) {
+            Ok(output) => output,
+            Err(e) => format!("Error: {e}"),
+        }
+    } else {
+        format!("Unknown tool: {tool_name}")
+    }
+}
+
+/// Pick a `max_tokens` budget proportional to the directive's length.
+fn adaptive_max_tokens(directive_len: usize) -> u32 {
+    if directive_len <= 80 {
+        800
+    } else if directive_len <= 400 {
+        1600
+    } else {
+        4096
+    }
+}
+
+fn truncate_tool_output(output: String) -> String {
+    if output.len() <= TOOL_OUTPUT_MAX_CHARS {
+        return output;
+    }
+    let mut result: String = output.chars().take(TOOL_OUTPUT_MAX_CHARS).collect();
+    result.push_str(&format!(
+        "\n...[truncated {} chars]",
+        output.len() - TOOL_OUTPUT_MAX_CHARS
+    ));
+    result
+}
+
 /// Emit an `AgentProgress` event onto the turn-event queue, if one is
 /// configured in the `ToolCtx`.
 fn report_progress(tool_ctx: &ToolCtx, progress: AgentProgress) {
@@ -85,15 +206,21 @@ pub async fn run_agent(
         Some(ctx.base_url.clone()),
     );
 
+    let max_tokens = adaptive_max_tokens(directive.len());
+
+    // Track repeated tool errors so the agent can recover from a dead end.
+    let mut consecutive_errors = 0usize;
+    let mut last_tool = String::new();
+
     // Limited iteration loop so we don't run forever
     for iteration in 0..MAX_ITERATIONS {
         use zesdex_application::ports::ProviderService;
         let (response_msg, _usage) = client
-            .chat(&messages, Some(defs.clone()), Some(4096), None)
+            .chat(&messages, Some(defs.clone()), Some(max_tokens), Some(0.2))
             .await?;
 
         let content = response_msg.content.clone().unwrap_or_default();
-        let tool_calls = response_msg.tool_calls.unwrap_or_default();
+        let tool_calls = response_msg.tool_calls.clone().unwrap_or_default();
 
         // If no tool calls, we're done — return content
         if tool_calls.is_empty() {
@@ -102,37 +229,48 @@ pub async fn run_agent(
             return Ok(content);
         }
 
-        // Execute tool calls
+        // Push the assistant message (with its tool_calls) BEFORE executing
-        for tc in &tool_calls {
+        // so the tool-calling contract is honoured: tool results reference
-            let tool_name = &tc.function.name;
+        // the calls declared in the preceding assistant message. Without
-            let args = sanitize_tool_arguments(&tc.function.arguments);
+        // this, the history is malformed (`[...tool, tool, assistant]`).
+        messages.push(response_msg);
 
-            debug!("Subagent executing tool: {tool_name}");
+        // Execute tool calls — read-only batches run concurrently (bounded,
+        // order preserved); any mutating tool forces the safe sequential path.
+        let results = execute_tool_batch(&tools, &tool_ctx, &tool_calls);
+
+        for (id, tool_name, result) in results {
+            debug!("Subagent tool {tool_name} finished");
 
             report_progress(
                 &tool_ctx,
                 AgentProgress::running(
                     "subagent",
-                    format!("{}:{}", directive, tool_name),
+                    format!("{}:{tool_name}", directive),
                     Some(tool_name.clone()),
                 ),
             );
 
-            let result = if let Some(tool) = tools.iter().find(|t| t.name() == tool_name) {
+            // Error-recovery: if the same tool keeps failing, inject a
-                match tool.run(&tool_ctx, &args) {
+            // system note steering the model to a different approach.
-                    Ok(output) => output,
+            if result.starts_with("Error:") {
-                    Err(e) => format!("Error: {e}"),
+                if last_tool.as_str() == tool_name.as_str() {
+                    consecutive_errors += 1;
+                } else {
+                    consecutive_errors = 1;
+                    last_tool = tool_name.clone();
+                }
+                if consecutive_errors >= MAX_CONSECUTIVE_TOOL_ERRORS {
+                    messages.push(ChatMessage::system(
+                        zesdex_domain::agent::prompt::error_recovery_note(&tool_name, &result),
+                    ));
+                    consecutive_errors = 0;
                 }
             } else {
-                format!("Unknown tool: {tool_name}")
+                consecutive_errors = 0;
-            };
-
-            messages.push(ChatMessage::tool(tc.id.clone(), result));
             }
 
-        // Add assistant response if there was text content
+            messages.push(ChatMessage::tool(id, truncate_tool_output(result)));
-        if !content.is_empty() {
-            messages.push(ChatMessage::assistant(Some(content)));
         }
     }
 
@@ -149,3 +287,111 @@ pub async fn run_agent(
         "Subagent reached iteration limit ({MAX_ITERATIONS})"
     ))
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::tools::ToolCtxBuilder;
+    use serde_json::json;
+
+    /// A deterministic mock tool whose `run` returns its own name (opting into
+    /// an optional sleep to make parallel-vs-sequential observable).
+    struct MockTool {
+        name: &'static str,
+        sleep_ms: u64,
+    }
+
+    impl MockTool {
+        fn new(name: &'static str, sleep_ms: u64) -> Self {
+            Self { name, sleep_ms }
+        }
+    }
+
+    impl Tool for MockTool {
+        fn name(&self) -> &'static str {
+            self.name
+        }
+        fn description(&self) -> &'static str {
+            "mock tool for tests"
+        }
+        fn parameters(&self) -> Value {
+            json!({"type":"object","properties":{}})
+        }
+        fn run(&self, _ctx: &ToolCtx, _args: &Value) -> Result<String> {
+            if self.sleep_ms > 0 {
+                std::thread::sleep(std::time::Duration::from_millis(self.sleep_ms));
+            }
+            Ok(self.name.to_string())
+        }
+    }
+
+    fn tc(name: &str, id: usize) -> zesdex_domain::core::ToolCall {
+        zesdex_domain::core::ToolCall {
+            id: format!("call_{id}"),
+            type_: "function".to_string(),
+            function: zesdex_domain::core::ToolFunction {
+                name: name.to_string(),
+                arguments: serde_json::Value::String(String::new()),
+            },
+        }
+    }
+
+    fn ctx() -> ToolCtx {
+        ToolCtxBuilder::default().build()
+    }
+
+    #[test]
+    fn parallel_batch_preserves_original_order() {
+        let tools: Vec<Box<dyn Tool>> = vec![
+            Box::new(MockTool::new("read", 0)),
+            Box::new(MockTool::new("grep", 0)),
+        ];
+        let calls = vec![tc("read", 1), tc("grep", 2), tc("read", 3)];
+
+        let results = execute_tool_batch(&tools, &ctx(), &calls);
+
+        // Results keep the assistant's original call order.
+        let names: Vec<&str> = results.iter().map(|(_, n, _)| n.as_str()).collect();
+        assert_eq!(names, vec!["read", "grep", "read"]);
+        // IDs follow the same original order (ordering contract).
+        let ids: Vec<&str> = results.iter().map(|(id, _, _)| id.as_str()).collect();
+        assert_eq!(ids, vec!["call_1", "call_2", "call_3"]);
+    }
+
+    #[test]
+    fn parallel_read_batch_is_faster_than_sequential() {
+        // Both reads sleep 30ms each. Parallel should finish ~30ms (both run
+        // at once), sequential would take ~60ms.
+        let tools: Vec<Box<dyn Tool>> = vec![Box::new(MockTool::new("read", 30))];
+        let calls = vec![tc("read", 1), tc("read", 2)];
+
+        let started = std::time::Instant::now();
+        let results = execute_tool_batch(&tools, &ctx(), &calls);
+        let elapsed = started.elapsed();
+
+        assert_eq!(results.len(), 2);
+        assert!(
+            elapsed < std::time::Duration::from_millis(55),
+            "parallel read batch took {elapsed:?}, expected concurrent execution"
+        );
+        assert!(elapsed >= std::time::Duration::from_millis(25));
+    }
+
+    #[test]
+    fn mutating_tool_forces_sequential_batch() {
+        // A batch containing a mutating tool ("write") must NOT run in
+        // parallel — the single 30ms read runs alone, then the write runs.
+        let tools: Vec<Box<dyn Tool>> = vec![
+            Box::new(MockTool::new("read", 30)),
+            Box::new(MockTool::new("write", 0)),
+        ];
+        let calls = vec![tc("read", 1), tc("write", 2)];
+
+        let results = execute_tool_batch(&tools, &ctx(), &calls);
+
+        let names: Vec<&str> = results.iter().map(|(_, n, _)| n.as_str()).collect();
+        assert_eq!(names, vec!["read", "write"]);
+        let ids: Vec<&str> = results.iter().map(|(id, _, _)| id.as_str()).collect();
+        assert_eq!(ids, vec!["call_1", "call_2"]);
+    }
+}

apps/infrastructure/src/subagent/provider.rs

@@ -63,15 +63,21 @@ impl SubagentProvider {
 
 /// Resolve subagent provider and model from settings.
 ///
-/// Flow: reads `settings.provider` and `settings.model` → if model is empty,
+/// Flow: reads `settings.provider` and `settings.model` → if provider is
+/// empty, falls back to `app_config.default_provider` → if model is empty,
 /// falls back to the provider config's `default_model` → if that is also
-/// empty, uses the domain default model constant.
+/// empty, uses `app_config.default_model` → finally the domain default model
+/// constant.
 #[instrument]
 pub fn resolve_subagent_provider(
     settings: &zesdex_domain::cms::Settings,
     app_config: &zesdex_domain::cms::AppConfig,
 ) -> (String, String) {
-    let provider = settings.provider.clone();
+    let provider = if settings.provider.is_empty() {
+        app_config.default_provider.clone()
+    } else {
+        settings.provider.clone()
+    };
     let model = settings.model.clone();
 
     // Use the default model from the provider config if available
@@ -80,6 +86,7 @@ pub fn resolve_subagent_provider(
             .providers
             .get(&provider)
             .and_then(|p| p.default_model.clone())
+            .or_else(|| Some(app_config.default_model.clone()))
             .unwrap_or_else(|| zesdex_domain::agent::defaults::DEFAULT_MODEL.to_string())
     } else {
         model

apps/infrastructure/src/subagent/spawn.rs

@@ -32,7 +32,6 @@ pub fn spawn_subagent(
 ) -> thread::JoinHandle<Result<String>> {
     info!("Spawning subagent: {directive}");
     thread::spawn(move || {
-        let rt = tokio::runtime::Runtime::new()?;
+        crate::runtime::runtime().block_on(run_agent(ctx, &directive, access, tool_ctx))
-        rt.block_on(run_agent(ctx, &directive, access, tool_ctx))
     })
 }

apps/infrastructure/src/tools/context.rs

@@ -66,6 +66,14 @@ impl ToolCtxBuilder {
         self.session_dir = v;
         self
     }
+    pub fn memory_dir(mut self, v: PathBuf) -> Self {
+        self.memory_dir = v;
+        self
+    }
+    pub fn worktrees_dir(mut self, v: PathBuf) -> Self {
+        self.worktrees_dir = v;
+        self
+    }
     pub fn workspaces(mut self, v: Vec<PathBuf>) -> Self {
         self.workspaces = v;
         self

apps/infrastructure/src/tools/executor.rs

@@ -15,6 +15,13 @@ impl InfrastructureToolExecutor {
             tools: all_tools(),
         }
     }
+
+    /// Whether a tool is read-only and safe to execute concurrently with
+    /// other parallel-safe tools. Delegates to the registry so the main
+    /// turn loop and subagent engine share one source of truth.
+    pub fn is_parallel_safe(tool_name: &str) -> bool {
+        crate::tools::tool_is_parallel_safe(tool_name)
+    }
 }
 
 impl ToolExecutor for InfrastructureToolExecutor {
@@ -37,4 +44,8 @@ impl ToolExecutor for InfrastructureToolExecutor {
             }
         }
     }
+
+    fn is_parallel_safe(&self, tool_name: &str) -> bool {
+        Self::is_parallel_safe(tool_name)
+    }
 }

apps/infrastructure/src/tools/memory/forget.rs

@@ -43,7 +43,8 @@ impl Tool for Forget {
         let name = crate::tools::arg_str(args, "name")?;
         info!(name, "forget invoked");
         let repo = crate::persistence::cms::memory_repo::MarkdownMemoryRepository::new();
-        repo.delete(&ctx.memory_dir, &name)?;
+        let memory_dir = crate::tools::memory::resolve_memory_dir(&ctx.memory_dir);
+        repo.delete(&memory_dir, &name)?;
         info!(name, "memory deleted");
         Ok(format!("Memory '{}' deleted", name))
     }

apps/infrastructure/src/tools/memory/mod.rs

@@ -1,5 +1,21 @@
 //! Memory management tools — remember, recall, forget.
 
+use std::path::PathBuf;
+
 pub mod forget;
 pub mod recall;
 pub mod remember;
+
+/// Resolve the directory the memory tools should read/write.
+///
+/// Prefer an explicitly-configured `ToolCtx.memory_dir`. If that is empty
+/// (a `ToolCtx` is often built without setting `memory_dir`), fall back to
+/// the canonical persistent memory location from `Store` so memories are not
+/// silently written into the current working directory.
+pub fn resolve_memory_dir(ctx_memory_dir: &std::path::Path) -> PathBuf {
+    if ctx_memory_dir.as_os_str().is_empty() {
+        zesdex_domain::core::Store::new().memory_dir
+    } else {
+        ctx_memory_dir.to_path_buf()
+    }
+}

apps/infrastructure/src/tools/memory/recall.rs

@@ -45,16 +45,47 @@ impl Tool for Recall {
     #[instrument(skip(self, ctx, args))]
     fn run(&self, ctx: &ToolCtx, args: &Value) -> Result<String> {
         let repo = crate::persistence::cms::memory_repo::MarkdownMemoryRepository::new();
+        let memory_dir = crate::tools::memory::resolve_memory_dir(&ctx.memory_dir);
 
         let specific_name = args.get("name").and_then(|v| v.as_str());
+        let search = args.get("search").and_then(|v| v.as_str());
 
         if let Some(name) = specific_name {
             info!(name, "recall loading specific memory");
-            let memory = repo.load(&ctx.memory_dir, name)?;
+            let memory = repo.load(&memory_dir, name)?;
-            Ok(serde_json::to_string_pretty(&memory)?)
+            return Ok(serde_json::to_string_pretty(&memory)?);
-        } else {
+        }
+
+        if let Some(query) = search {
+            let query = query.trim().to_lowercase();
+            info!(search = %query, "recall searching memories");
+            if query.is_empty() {
+                return Ok("Search query is empty".to_string());
+            }
+            let names = repo.list(&memory_dir)?;
+            let mut matches: Vec<String> = Vec::new();
+            for name in &names {
+                // Load each memory and match against name/description/content.
+                if let Ok(m) = repo.load(&memory_dir, name) {
+                    let haystack =
+                        format!("{} {} {}", m.name, m.description, m.content).to_lowercase();
+                    if haystack.contains(&query) {
+                        matches.push(m.name);
+                    }
+                }
+            }
+            if matches.is_empty() {
+                return Ok(format!("No memories match '{query}'"));
+            }
+            return Ok(format!(
+                "Memories matching '{query}' ({}):\n{}",
+                matches.len(),
+                matches.join("\n")
+            ));
+        }
+
         info!("recall listing all memories");
-            let names = repo.list(&ctx.memory_dir)?;
+        let names = repo.list(&memory_dir)?;
         if names.is_empty() {
             info!("no memories found");
             return Ok("No memories saved yet".to_string());
@@ -62,4 +93,84 @@ impl Tool for Recall {
         Ok(format!("Available memories:\n{}", names.join("\n")))
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::tools::ToolCtxBuilder;
+    use zesdex_domain::cms::MemoryRepository;
+
+    fn save_mem(name: &str, description: &str, content: &str, dir: &std::path::Path) {
+        let repo = crate::persistence::cms::memory_repo::MarkdownMemoryRepository::new();
+        let memory = zesdex_domain::cms::Memory {
+            name: name.to_string(),
+            description: description.to_string(),
+            content: content.to_string(),
+            kind: "reference".to_string(),
+            created_at: 0,
+            updated_at: 0,
+            outcome: None,
+            lifecycle: "active".to_string(),
+            scope: None,
+            before_snippet: None,
+            after_snippet: None,
+            provenances: Vec::new(),
+        };
+        repo.save(dir, &memory).unwrap();
+    }
+
+    #[test]
+    fn search_filters_memories_by_keyword() {
+        let dir = std::env::temp_dir().join(format!("zdx-mem-test-{}", uuid::Uuid::new_v4()));
+        std::fs::create_dir_all(&dir).unwrap();
+        save_mem(
+            "rust-concurrency",
+            "tokio spawn patterns",
+            "how to use async tasks",
+            &dir,
+        );
+        save_mem(
+            "docker-deploy",
+            "deploy via compose",
+            "container orchestration",
+            &dir,
+        );
+
+        let tool_ctx = ToolCtxBuilder::default().memory_dir(dir.clone()).build();
+        let args = serde_json::json!({ "search": "tokio" });
+
+        let out = Recall.run(&tool_ctx, &args).unwrap();
+        assert!(
+            out.contains("rust-concurrency"),
+            "should match rust-concurrency, got: {out}"
+        );
+        assert!(
+            !out.contains("docker-deploy"),
+            "docker-deploy should not match tokio"
+        );
+
+        // A query with no match reports so.
+        let no_match = Recall
+            .run(&tool_ctx, &serde_json::json!({ "search": "zzzznope" }))
+            .unwrap();
+        assert!(no_match.contains("No memories match"), "{no_match}");
+
+        std::fs::remove_dir_all(&dir).ok();
+    }
+
+    #[test]
+    fn resolve_memory_dir_falls_back_to_store_when_empty() {
+        // An empty ToolCtx.memory_dir is resolved to the canonical Store path.
+        let resolved = crate::tools::memory::resolve_memory_dir(std::path::Path::new(""));
+        assert!(!resolved.as_os_str().is_empty());
+        assert!(
+            resolved.ends_with("memory"),
+            "expected memory dir, got {resolved:?}"
+        );
+
+        // An explicit memory_dir is preserved.
+        let explicit =
+            crate::tools::memory::resolve_memory_dir(std::path::Path::new("/tmp/custom-memory"));
+        assert_eq!(explicit, std::path::Path::new("/tmp/custom-memory"));
+    }
 }

apps/infrastructure/src/tools/memory/remember.rs

@@ -81,7 +81,8 @@ impl Tool for Remember {
         };
 
         let repo = crate::persistence::cms::memory_repo::MarkdownMemoryRepository::new();
-        repo.save(&ctx.memory_dir, &memory)?;
+        let memory_dir = crate::tools::memory::resolve_memory_dir(&ctx.memory_dir);
+        repo.save(&memory_dir, &memory)?;
 
         info!(name, "memory saved");
         Ok(format!("Memory '{}' saved", name))

apps/infrastructure/src/tools/mod.rs

@@ -59,7 +59,7 @@ pub mod workflow;
 // like `crate::tools::{Tool, ToolCtx}` continue to work.
 pub use context::{ToolCtx, ToolCtxBuilder};
 pub use graduated::{check_graduated_checks, GraduatedCheck};
-pub use registry::{all_tools, tool_defs, tool_is_risky};
+pub use registry::{all_tools, tool_defs, tool_is_parallel_safe, tool_is_risky};
 pub use util::{arg_str, execute_cmd, log_write_edit_tool, resolve_path};
 
 /// Common interface every agent-invocable tool implements.

apps/infrastructure/src/tools/parallel_delegate.rs

@@ -123,7 +123,7 @@ impl Tool for ParallelDelegate {
                 .collect()
         } else {
             // Auto-split using LLM
-            let rt = tokio::runtime::Runtime::new()?;
+            let rt = crate::runtime::runtime();
             let directives = rt.block_on(auto_split_task(
                 &task,
                 max_parallel,
@@ -146,9 +146,13 @@ impl Tool for ParallelDelegate {
             "parallel delegation: starting subagents"
         );
 
-        // Spawn agents in parallel
+        // Spawn agents in parallel — bounded: never more than `max_parallel`
-        let mut handles = Vec::new();
+        // subagent threads in flight at once (Claude Code-style isolation).
-        for (i, (directive, access)) in directives.iter().enumerate() {
+        let mut results: Vec<(usize, String, String)> = Vec::new();
+        for batch in directives.chunks(max_parallel) {
+            let mut handles = Vec::with_capacity(batch.len());
+            for (i, (directive, access)) in batch.iter().enumerate() {
+                let global_idx = results.len() + i;
                 let subagent_ctx = SubagentContext::new(
                     directive.clone(),
                     ctx.clone(),
@@ -158,13 +162,12 @@ impl Tool for ParallelDelegate {
                     model.clone(),
                 );
 
-            debug!(agent_index = i, access = ?access, "spawning parallel agent");
+                debug!(agent_index = global_idx, access = ?access, "spawning parallel agent");
                 let handle = spawn_subagent(subagent_ctx, directive.clone(), *access, ctx.clone());
-            handles.push((i, handle));
+                handles.push((global_idx, handle));
             }
 
-        // Join all results
+            // Join this batch before spawning the next.
-        let mut results: Vec<(usize, String, String)> = Vec::new();
             for (i, handle) in handles {
                 match handle.join() {
                     Ok(Ok(output)) => {
@@ -185,10 +188,11 @@ impl Tool for ParallelDelegate {
                     }
                 }
             }
+        }
 
         // Consolidate results
         if synthesize && results.len() > 1 {
-            let rt = tokio::runtime::Runtime::new()?;
+            let rt = crate::runtime::runtime();
             let consolidated =
                 rt.block_on(consolidate_results(&results, &base_url, &api_key, &model))?;
             Ok(format!(

apps/infrastructure/src/tools/registry.rs

@@ -52,6 +52,34 @@ pub fn tool_is_risky(name: &str) -> bool {
     matches!(name, "write" | "delete" | "edit" | "bash" | "git_operator")
 }
 
+/// Whether a tool by name is read-only and therefore safe to run *in
+/// parallel* with other tool calls within the same assistant message.
+///
+/// Read-only tools only inspect the workspace (read files, grep, glob,
+/// semantic search, list symbols, recall memory, web search, directory
+/// listing). They have no side effects, so concurrent execution cannot
+/// create data races or conflicting writes.
+///
+/// Everything else (edits, writes, deletes, shell, git, planning, memory
+/// writes, agent/spawn orchestration) stays sequential to preserve
+/// correctness.
+pub fn tool_is_parallel_safe(name: &str) -> bool {
+    matches!(
+        name,
+        "read"
+            | "grep"
+            | "glob"
+            | "semantic_search"
+            | "list_symbols"
+            | "web_search"
+            | "recall"
+            | "dir_list"
+            | "pong"
+            | "seq_think"
+            | "dir_cache_update"
+    )
+}
+
 /// Convert a list of tools into provider-facing `ToolDef` request schema.
 pub fn tool_defs(tools: &[Box<dyn super::Tool>]) -> Vec<zesdex_domain::core::ToolDef> {
     tools
@@ -66,3 +94,59 @@ pub fn tool_defs(tools: &[Box<dyn super::Tool>]) -> Vec<zesdex_domain::core::Too
         })
         .collect()
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn read_only_tools_are_parallel_safe() {
+        for name in [
+            "read",
+            "grep",
+            "glob",
+            "semantic_search",
+            "list_symbols",
+            "web_search",
+            "recall",
+            "dir_list",
+            "pong",
+            "seq_think",
+            "dir_cache_update",
+        ] {
+            assert!(
+                tool_is_parallel_safe(name),
+                "{name} should be parallel-safe"
+            );
+        }
+    }
+
+    #[test]
+    fn mutating_and_shell_tools_are_not_parallel_safe() {
+        for name in [
+            "edit",
+            "write",
+            "delete",
+            "bash",
+            "git_operator",
+            "git_worktree",
+            "remember",
+            "forget",
+            "todowrite",
+            "todofinish",
+            "plan_enter",
+            "plan_ready",
+            "workflow_run",
+            "hive_mind",
+            "spawn_agents",
+            "spawn_pipeline",
+            "parallel_delegate",
+            "explore_codebase",
+        ] {
+            assert!(
+                !tool_is_parallel_safe(name),
+                "{name} should NOT be parallel-safe"
+            );
+        }
+    }
+}

apps/infrastructure/src/tools/semantic_search.rs

@@ -121,7 +121,66 @@ pub struct CodeSymbol {
 }
 
 /// The in-memory symbol index, shared via a global static.
-static SYMBOL_INDEX: Mutex<Option<SymbolIndex>> = Mutex::new(None);
+///
+/// Keyed by workspace path: each workspace owns its own `SymbolIndex`, so
+/// searching one repo never leaks stale symbols from another, and a rebuild
+/// triggered for workspace A cannot clobber the index of B. The lock is only
+/// ever held briefly (to check/insert/look up) — never across the I/O-heavy
+/// walk in `rebuild`, which runs locally and is swapped in under a short lock.
+static SYMBOL_INDEX: std::sync::OnceLock<Mutex<HashMap<String, SymbolIndex>>> =
+    std::sync::OnceLock::new();
+
+/// Access the (lazily initialised) global per-workspace symbol index map.
+fn symbol_index_map() -> &'static Mutex<HashMap<String, SymbolIndex>> {
+    SYMBOL_INDEX.get_or_init(|| Mutex::new(HashMap::new()))
+}
+
+/// Ensure the per-workspace symbol index is built, returning the symbol count.
+///
+/// If `force` is true, or the workspace has no cached (non-empty) index yet,
+/// the index is rebuilt. The rebuild itself runs OUTSIDE the global lock
+/// (the walk can take seconds on a large repo), then the result is stored
+/// under a short lock so concurrent searches never block on the I/O. Returns
+/// the number of symbols now cached for the workspace.
+fn ensure_symbol_index(workspace: &str, force: bool) -> Result<usize> {
+    let ready = {
+        let map = symbol_index_map()
+            .lock()
+            .map_err(|e| anyhow::anyhow!("index lock failed: {e}"))?;
+        !force && map.get(workspace).is_some_and(|i| !i.is_empty())
+    };
+
+    if !ready {
+        // Rebuild locally, off the global lock (I/O heavy).
+        let mut fresh = SymbolIndex::new();
+        let count = fresh.rebuild(workspace)?;
+        // Swap in under a short lock; keep an existing non-empty index if a
+        // concurrent rebuild already populated this workspace.
+        let mut map = symbol_index_map()
+            .lock()
+            .map_err(|e| anyhow::anyhow!("index lock failed: {e}"))?;
+        if map.get(workspace).is_none_or(|i| i.is_empty()) {
+            map.insert(workspace.to_string(), fresh);
+        }
+        return Ok(count);
+    }
+
+    let map = symbol_index_map()
+        .lock()
+        .map_err(|e| anyhow::anyhow!("index lock failed: {e}"))?;
+    Ok(map.get(workspace).map_or(0, |i| i.len()))
+}
+
+/// Lock and return a borrow to the global per-workspace symbol index map.
+///
+/// The caller must have called [`ensure_symbol_index`] first, then looks up
+/// its workspace key; the lookup is short and in-memory, so holding the guard
+/// for the search is fine.
+fn symbol_index() -> Result<std::sync::MutexGuard<'static, HashMap<String, SymbolIndex>>> {
+    symbol_index_map()
+        .lock()
+        .map_err(|e| anyhow::anyhow!("index lock failed: {e}"))
+}
 
 // ---------------------------------------------------------------------------
 //  Language-specific regexes (lazily compiled)
@@ -270,6 +329,17 @@ impl SymbolIndex {
         self.symbols.is_empty()
     }
 
+    /// Returns true when the cached index must be rebuilt for the given
+    /// workspace — either because nothing has been indexed yet, or because the
+    /// requested workspace differs from the one the index was built for.
+    ///
+    /// Without this, searching a *different* workspace after the first one
+    /// silently returns stale symbols from the previously indexed repo
+    /// (a misleading result for a coding agent).
+    pub fn needs_rebuild(&self, workspace: &str) -> bool {
+        self.is_empty() || self.workspace_path.as_deref() != Some(workspace)
+    }
+
     pub fn len(&self) -> usize {
         self.symbols.len()
     }
@@ -1250,15 +1320,11 @@ impl Tool for SemanticSearch {
             "semantic search"
         );
 
-        let mut guard = SYMBOL_INDEX
+        // Build (or load) the per-workspace index without holding the global
-            .lock()
+        // lock across the I/O-heavy walk.
-            .map_err(|e| anyhow::anyhow!("index lock failed: {e}"))?;
+        let _count = ensure_symbol_index(&workspace, rebuild)?;
-        let index = guard.get_or_insert_with(SymbolIndex::new);
+        let map = symbol_index()?;
-
+        let index = map.get(&workspace).expect("index should be ensured");
-        if rebuild || index.is_empty() {
-            let count = index.rebuild(&workspace)?;
-            debug!(symbol_count = count, "symbol index rebuilt");
-        }
 
         // Map kind filter to enum
         let target_kind = match kind_filter {
@@ -1398,12 +1464,15 @@ impl Tool for RebuildIndex {
 
         info!("rebuilding multi-language symbol index");
 
-        let mut guard = SYMBOL_INDEX
+        // Force a rebuild of this workspace's index. The walk runs off the
-            .lock()
+        // global lock (via ensure_symbol_index) so it cannot stall concurrent
-            .map_err(|e| anyhow::anyhow!("index lock failed: {e}"))?;
+        // searches.
-        let index = guard.get_or_insert_with(SymbolIndex::new);
+        let count = ensure_symbol_index(&workspace, true)?;
-        let count = index.rebuild(&workspace)?;
+        let map = symbol_index()?;
-        let by_lang = index.count_by_language();
+        let by_lang = match map.get(&workspace) {
+            Some(i) => i.count_by_language(),
+            None => Vec::new(),
+        };
 
         let mut out = format!(
             "Symbol index rebuilt successfully. {} symbols indexed.\n\n",
@@ -1497,15 +1566,11 @@ impl Tool for ListSymbols {
             .map(|p| p.to_string_lossy().to_string())
             .unwrap_or_else(|| ".".to_string());
 
-        let mut guard = SYMBOL_INDEX
+        // Build (or load) the per-workspace index without holding the global
-            .lock()
+        // lock across the I/O-heavy walk.
-            .map_err(|e| anyhow::anyhow!("index lock failed: {e}"))?;
+        let _count = ensure_symbol_index(&workspace, rebuild)?;
-        let index = guard.get_or_insert_with(SymbolIndex::new);
+        let map = symbol_index()?;
-
+        let index = map.get(&workspace).expect("index should be ensured");
-        if rebuild || index.is_empty() {
-            let count = index.rebuild(&workspace)?;
-            info!(symbol_count = count, "symbol index rebuilt for list");
-        }
 
         let target_lang = match lang_filter {
             "rust" => Some(Language::Rust),
@@ -1746,4 +1811,80 @@ mod tests {
         let index = SymbolIndex::new();
         assert!(index.search("anything", 10).is_empty());
     }
+
+    #[test]
+    fn test_needs_rebuild_workspace_aware() {
+        let ws_a = std::env::temp_dir().join(format!("ws_a_{}", uuid::Uuid::new_v4()));
+        let ws_b = std::env::temp_dir().join(format!("ws_b_{}", uuid::Uuid::new_v4()));
+        std::fs::create_dir_all(&ws_a).unwrap();
+        std::fs::create_dir_all(&ws_b).unwrap();
+        std::fs::write(ws_a.join("a.rs"), "pub fn fn_in_a() {}\n").unwrap();
+        std::fs::write(ws_b.join("b.rs"), "pub fn fn_in_b() {}\n").unwrap();
+
+        let mut index = SymbolIndex::new();
+        let a = ws_a.to_string_lossy().to_string();
+        let b = ws_b.to_string_lossy().to_string();
+
+        // Fresh index: needs rebuild for any workspace.
+        assert!(index.needs_rebuild(&a));
+
+        // After rebuilding A, searching A needs no rebuild...
+        index.rebuild(&a).unwrap();
+        assert!(!index.needs_rebuild(&a));
+        // ...but searching B DOES (stale index otherwise).
+        assert!(
+            index.needs_rebuild(&b),
+            "workspace switch must trigger rebuild"
+        );
+
+        // Rebuilding B flips the cached workspace.
+        index.rebuild(&b).unwrap();
+        assert!(!index.needs_rebuild(&b));
+        assert!(index.needs_rebuild(&a));
+
+        std::fs::remove_dir_all(&ws_a).ok();
+        std::fs::remove_dir_all(&ws_b).ok();
+    }
+
+    #[test]
+    fn test_ensure_symbol_index_per_workspace_isolation() {
+        let ws_a = std::env::temp_dir().join(format!("iso_a_{}", uuid::Uuid::new_v4()));
+        let ws_b = std::env::temp_dir().join(format!("iso_b_{}", uuid::Uuid::new_v4()));
+        std::fs::create_dir_all(&ws_a).unwrap();
+        std::fs::create_dir_all(&ws_b).unwrap();
+        std::fs::write(ws_a.join("a.rs"), "pub fn only_in_a() {}\n").unwrap();
+        std::fs::write(ws_b.join("b.rs"), "pub fn only_in_b() {}\n").unwrap();
+
+        let a = ws_a.to_string_lossy().to_string();
+        let b = ws_b.to_string_lossy().to_string();
+
+        // Build A and B independently through the shared global helper.
+        let count_a = ensure_symbol_index(&a, false).unwrap();
+        assert!(
+            count_a >= 1,
+            "workspace A should index its fn, got {count_a}"
+        );
+        let count_b = ensure_symbol_index(&b, false).unwrap();
+        assert!(
+            count_b >= 1,
+            "workspace B should index its fn, got {count_b}"
+        );
+
+        // Rebuilding A must not have clobbered B and vice-versa.
+        let count_a_again = ensure_symbol_index(&a, true).unwrap();
+        assert!(count_a_again >= 1);
+
+        // Each workspace's cached index is independently correct.
+        {
+            let map = symbol_index().unwrap();
+            let idx_a = map.get(&a).unwrap();
+            assert!(!idx_a.search("only_in_a", 5).is_empty());
+            assert!(idx_a.search("only_in_b", 5).is_empty());
+            let idx_b = map.get(&b).unwrap();
+            assert!(!idx_b.search("only_in_b", 5).is_empty());
+        }
+
+        std::fs::remove_dir_all(&ws_a).ok();
+        std::fs::remove_dir_all(&ws_b).ok();
+    }
 }

apps/infrastructure/src/tools/utility/dir_cache_update.rs

@@ -63,7 +63,7 @@ impl crate::tools::Tool for DirCacheUpdate {
         // Persist the resolved paths into the shared DirCache so the TUI
         // and other tools can read the cached listing without re-scanning.
         let dc = ctx.dir_cache.clone();
-        let rt = tokio::runtime::Runtime::new()?;
+        let rt = crate::runtime::runtime();
         rt.block_on(async { dc.write().await.set(resolved).await });
 
         info!(count, "directory cache updated");

apps/infrastructure/src/tools/workflow.rs

@@ -9,7 +9,6 @@ use anyhow::Result;
 use serde_json::{json, Value};
 use tracing::{debug, info, instrument, warn};
 
-use crate::llm::provider::LlmClient;
 use crate::tools::{arg_str, Tool, ToolCtx};
 use crate::workflow::engine::execution::execute_workflow;
 use crate::workflow::hive_mind::cycle::execute_cycle;
@@ -60,14 +59,8 @@ impl Tool for WorkflowRun {
             phase_names.join(", ")
         );
 
-        let llm_client = LlmClient::new(
+        let rt = crate::runtime::runtime();
-            crate::llm::provider::DEFAULT_API_KEY.to_string(),
+        let result: Vec<String> = rt.block_on(async { execute_workflow(&script, ctx).await })?;
-            zesdex_domain::agent::defaults::DEFAULT_MODEL.to_string(),
-            None,
-        );
-        let rt = tokio::runtime::Runtime::new()?;
-        let result: Vec<String> =
-            rt.block_on(async { execute_workflow(&script, ctx, &llm_client).await })?;
 
         info!(phase_count = result.len(), "Workflow completed");
         Ok(format!(
@@ -229,7 +222,7 @@ impl Tool for HiveMind {
             .ok_or_else(|| anyhow::anyhow!("missing 'cycles' array"))?;
 
         info!("Hive mind starting with {} cycles", cycles_val.len());
-        let rt = tokio::runtime::Runtime::new()?;
+        let rt = crate::runtime::runtime();
         let mut all_node_outputs = Vec::new();
 
         for (cycle_idx, cycle_val) in cycles_val.iter().enumerate() {

apps/infrastructure/src/utils.rs

@@ -221,7 +221,14 @@ pub fn build_rich_context(root: &Path) -> String {
     ));
 
     // 2. Custom Rules
-    let rule_files = [".cursorrules", ".zesdexrules", "claude.md", "agent.md"];
+    let rule_files = [
+        "AGENTS.md",
+        "CLAUDE.md",
+        ".cursorrules",
+        ".zesdexrules",
+        "claude.md",
+        "agent.md",
+    ];
     for file in rule_files {
         let p = root.join(file);
         if let Ok(content) = std::fs::read_to_string(&p) {
@@ -268,7 +275,7 @@ pub fn build_rich_context(root: &Path) -> String {
         let p = root.join(file);
         if let Ok(content) = std::fs::read_to_string(&p) {
             let snippet = if content.len() > 1500 {
-                format!("{}\n... (truncated)", &content[..1500])
+                format!("{}\n... (truncated)", truncate_chars(&content, 1500))
             } else {
                 content
             };
@@ -294,7 +301,7 @@ pub fn build_rich_context(root: &Path) -> String {
     let readme_path = root.join("README.md");
     if let Ok(content) = std::fs::read_to_string(&readme_path) {
         let snippet = if content.len() > 1000 {
-            format!("{}\n... (truncated)", &content[..1000])
+            format!("{}\n... (truncated)", truncate_chars(&content, 1000))
         } else {
             content
         };
@@ -341,3 +348,53 @@ pub fn build_rich_context(root: &Path) -> String {
 
     ctx.trim_end().to_string()
 }
+
+/// Truncate a string to at most `max_chars` **characters**, never cutting a
+/// multi-byte UTF-8 code point in half.
+///
+/// `&s[..n]` with `n` a raw byte index panics when `n` lands inside a
+/// multi-byte character (e.g. an emoji, `→`, or CJK in a README/diff). This
+/// helper slices on character boundaries so content is safely capped at a
+/// byte budget while remaining valid UTF-8.
+pub fn truncate_chars(s: &str, max_chars: usize) -> String {
+    if s.chars().count() <= max_chars {
+        return s.to_string();
+    }
+    s.chars().take(max_chars).collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn truncate_chars_leaves_short_strings_unchanged() {
+        assert_eq!(truncate_chars("short", 100), "short");
+        assert_eq!(truncate_chars("", 5), "");
+    }
+
+    #[test]
+    fn truncate_chars_cuts_to_max_chars() {
+        assert_eq!(truncate_chars("hello world", 5), "hello");
+    }
+
+    #[test]
+    fn truncate_chars_never_splits_multibyte_utf8() {
+        // 4 chars each: '→' is 3 bytes. Byte-slicing at 5 would panic; char
+        // slicing must not.
+        let s = "a→b→c→d";
+        let t = truncate_chars(s, 5);
+        assert_eq!(t, "a→b→c");
+        assert!(t.chars().count() <= 5);
+        // No replacement char must appear (valid UTF-8 preserved).
+        assert!(!t.contains('\u{FFFD}'));
+    }
+
+    #[test]
+    fn truncate_chars_handles_emoji() {
+        let s = "🚀🚀🚀🚀";
+        let t = truncate_chars(s, 2);
+        assert_eq!(t, "🚀🚀");
+        assert!(t.chars().count() == 2);
+    }
+}

apps/infrastructure/src/workflow/engine/execution.rs

@@ -3,7 +3,6 @@
 use anyhow::Result;
 use tracing::{info, instrument};
 
-use crate::llm::provider::LlmClient;
 use crate::tools::ToolCtx;
 use crate::workflow::engine::primitives::execute_primitive;
 use zesdex_domain::workflow::WorkflowScript;
@@ -11,12 +10,8 @@ use zesdex_domain::workflow::WorkflowScript;
 /// Execute each phase of a workflow script sequentially.
 ///
 /// Flow: for each phase → execute_primitive → collect result.
-#[instrument(skip(tool_ctx, _llm_client))]
+#[instrument(skip(tool_ctx))]
-pub async fn execute_workflow(
+pub async fn execute_workflow(script: &WorkflowScript, tool_ctx: &ToolCtx) -> Result<Vec<String>> {
-    script: &WorkflowScript,
-    tool_ctx: &ToolCtx,
-    _llm_client: &LlmClient,
-) -> Result<Vec<String>> {
     info!(
         "Executing workflow: {} ({} phases)",
         script.name,

apps/infrastructure/src/workflow/hive_mind/cycle.rs

@@ -1,11 +1,15 @@
 //! Hive-mind cycle execution — run one cycle of parallel nodes.
 //!
 //! Flow: load settings → resolve LLM credentials → run all directives in the
-//! cycle concurrently via try_join_all → collect Vec<NodeOutput>.
+//! cycle concurrently via a BOUNDED buffer (`buffer_unordered(MAX)`) → collect
+//! `Vec<NodeOutput>`. Unlike `try_join_all`, a single failing node does NOT
+//! fail the whole cycle — failed nodes are logged and replaced with an
+//! `[ERROR]` output so the remaining results are preserved (like Claude
+//! Code's isolated subagents).
 
 use anyhow::Result;
-use futures_util::future::try_join_all;
+use futures_util::stream::StreamExt;
-use tracing::info;
+use tracing::{info, warn};
 
 use zesdex_domain::cms::{AppConfigRepository, SettingsRepository};
 use zesdex_domain::core::Store;
@@ -15,16 +19,21 @@ use crate::subagent::context::SubagentContext;
 use crate::subagent::division::AccessTier;
 use crate::subagent::engine::run_agent;
 use crate::tools::ToolCtx;
-use zesdex_domain::workflow::{CognitiveCycle, NodeOutput};
+use zesdex_domain::workflow::{CognitiveCycle, NodeDirective, NodeOutput};
+
+/// Maximum number of hive-mind nodes running concurrently per cycle.
+/// Keeps thread/runtime pressure bounded (Claude Code-style).
+const MAX_CONCURRENT_NODES: usize = 8;
 
 /// Execute one cycle: run each node directive and collect outputs.
 ///
 /// Flow:
 ///   1. Load `Settings` and `AppConfig` from the store directory.
 ///   2. Resolve provider, model, base_url, and api_key.
-///   3. Spawn all directives concurrently — each builds a `SubagentContext`
+///   3. Spawn directives with bounded concurrency — each builds a
-///      and calls `run_agent` (Full access).
+///      `SubagentContext` and calls `run_agent`.
-///   4. `try_join_all` waits for all to complete, then collect `NodeOutput`s.
+///   4. Collect `NodeOutput`s; failed nodes are logged and replaced with an
+///      `[ERROR]` placeholder so the cycle still completes.
 pub async fn execute_cycle(cycle: &CognitiveCycle, tool_ctx: &ToolCtx) -> Result<Vec<NodeOutput>> {
     info!(
         "Executing cycle {} with {} directives",
@@ -53,10 +62,9 @@ pub async fn execute_cycle(cycle: &CognitiveCycle, tool_ctx: &ToolCtx) -> Result
 
     let cycle_index = cycle.index;
 
-    use zesdex_domain::workflow::NodeDirective;
+    // Run all directives with bounded concurrency. Each node is its own
-
+    // future; failures are collected, not propagated (isolated errors).
-    // Run all directives in this cycle concurrently.
+    let tasks: Vec<_> = cycle
-    let handles: Vec<_> = cycle
         .directives
         .iter()
         .enumerate()
@@ -79,18 +87,33 @@ pub async fn execute_cycle(cycle: &CognitiveCycle, tool_ctx: &ToolCtx) -> Result
                 _ => AccessTier::Read,
             };
 
+            let node_id = format!("Node-{}-{}", cycle_index, i);
             async move {
-                let result = run_agent(ctx, &dir, access, tc).await?;
+                match run_agent(ctx, &dir, access, tc).await {
-                Ok::<NodeOutput, anyhow::Error>(NodeOutput {
+                    Ok(output) => Ok::<NodeOutput, anyhow::Error>(NodeOutput {
-                    id: format!("Node-{}-{}", cycle_index, i),
+                        id: node_id.clone(),
                         directive: dir,
-                    output: result,
+                        output,
+                    }),
+                    Err(e) => {
+                        warn!(node = %node_id, error = %e, "hive-mind node failed (isolated)");
+                        Ok::<NodeOutput, anyhow::Error>(NodeOutput {
+                            id: node_id,
+                            directive: dir,
+                            output: format!("[ERROR] {e}"),
                         })
                     }
+                }
+            }
         })
         .collect();
 
-    let results = try_join_all(handles).await?;
+    // Bounded concurrency: run at most MAX_CONCURRENT_NODES futures at once.
+    let mut stream = futures_util::stream::iter(tasks).buffer_unordered(MAX_CONCURRENT_NODES);
+    let mut results = Vec::with_capacity(cycle.directives.len());
+    while let Some(node) = stream.next().await {
+        results.push(node?);
+    }
 
     Ok(results)
 }

apps/infrastructure/src/workflow/hive_mind/synthesis.rs

@@ -1,28 +1,128 @@
 //! Consensus synthesis — reconciles multiple node outputs into one assessment.
+//!
+//! Flow: load settings → resolve LLM credentials → ask the model to distill the
+//! node outputs into a single consensus (conflicts, agreements, key findings)
+//! → return the synthesized text. If the LLM call fails for any reason, we
+//! degrade gracefully to a concatenation-based summary so consensus synthesis
+//! never breaks the surrounding hive-mind cycle (mirrors the isolated-errors
+//! philosophy used for the nodes themselves).
 
 use anyhow::Result;
-use tracing::info;
+use tracing::{info, warn};
 
-use crate::tools::ToolCtx;
+use zesdex_domain::cms::{AppConfigRepository, SettingsRepository};
+use zesdex_domain::core::message::ChatMessage;
+use zesdex_domain::core::Store;
 use zesdex_domain::workflow::NodeOutput;
 
-/// Synthesize a consensus from all node outputs.
+use crate::persistence::{JsonAppConfigRepository, JsonSettingsRepository};
+use crate::tools::ToolCtx;
+
+/// Maximum characters of node output to feed into the synthesis prompt per node.
+/// Keeps the prompt bounded so a huge/talkative node cannot blow up the request.
+const MAX_NODE_OUTPUT_CHARS: usize = 4000;
+
+/// Synthesize a consensus from all node outputs using the LLM.
 ///
-/// Flow: combine node outputs → return consensus text.
+/// Flow: combine node outputs → ask the model to reconcile them into a single
-/// Uses simple concatenation-based synthesis (avoids LLM call dependency).
+/// consensus → return the synthesized text. Falls back to a plain
+/// concatenation summary if the LLM is unreachable or the call fails.
 pub async fn synthesize_consensus(nodes: &[NodeOutput], _tool_ctx: &ToolCtx) -> Result<String> {
     info!("Synthesizing consensus from {} nodes", nodes.len());
 
-    let mut combined = String::new();
+    let combined = build_combined_body(nodes);
-    for node in nodes {
+
-        combined.push_str(&format!(
+    // 1. Resolve LLM credentials (same source of truth as execute_cycle).
-            "\n## {} — {}\n\n{}\n",
+    let store = Store::new();
-            node.id, node.directive, node.output
+    let settings = JsonSettingsRepository::new()
+        .load(&store.base_dir)
+        .unwrap_or_default();
+    let app_config = JsonAppConfigRepository::new()
+        .load(&store.base_dir)
+        .unwrap_or_default();
+
+    let (provider, model) =
+        crate::subagent::provider::resolve_subagent_provider(&settings, &app_config);
+    let base_url = app_config
+        .providers
+        .get(&provider)
+        .map(|p| p.api_base.clone())
+        .unwrap_or_else(|| zesdex_domain::agent::defaults::DEFAULT_API_BASE.to_string());
+    let api_key = crate::llm::provider::resolve_api_key(&settings, &app_config);
+
+    let client = crate::llm::provider::LlmClient::new(api_key, model, Some(base_url));
+
+    // 2. Build the synthesis prompt.
+    let system_msg = ChatMessage::system(
+        "You are a consensus synthesizer for a multi-agent hive mind. \
+         Several independent nodes analysed a problem and produced the outputs \
+         below. Distill them into ONE coherent consensus report with these \
+         sections:\n\
+         - AGREEMENTS: points multiple nodes converge on.\n\
+         - CONFLICTS: contradictory conclusions, with which node(s) support each side.\n\
+         - KEY FINDINGS: the most important, actionable takeaways.\n\
+         - RECOMMENDATION: a single recommended next action, or 'no clear consensus' \
+         if the outputs are too divergent.\n\
+         Be concise and factual. If a node errored, note it and ignore its content.\n\
+         Do not invent facts not present in the node outputs.",
+    );
+    let user_msg = ChatMessage::user(format!(
+        "Consolidate these {} node outputs into a single consensus:\n\n{}",
+        nodes.len(),
+        combined
     ));
+
+    // 3. Call the model and gracefully degrade on failure.
+    match call_consensus(&client, &[system_msg, user_msg]).await {
+        Ok(text) => {
+            let trimmed = text.trim();
+            if trimmed.is_empty() {
+                warn!("consensus LLM returned empty output; falling back to concat summary");
+                Ok(concat_summary(nodes))
+            } else {
+                Ok(format!(
+                    "# Consensus Synthesis\n\n\
+                     Nodes synthesized: {}\n\n{}",
+                    nodes.len(),
+                    trimmed
+                ))
+            }
+        }
+        Err(e) => {
+            warn!(error = %e, "consensus LLM call failed; falling back to concat summary");
+            Ok(concat_summary(nodes))
+        }
+    }
 }
 
-    Ok(format!(
+/// Run the LLM consensus call and return the assistant text.
-        "# Consensus Synthesis\n\
+async fn call_consensus(
+    client: &crate::llm::provider::LlmClient,
+    messages: &[ChatMessage],
+) -> Result<String> {
+    use zesdex_application::ports::ProviderService;
+    let (msg, _) = client.chat(messages, None, Some(1024), Some(0.3)).await?;
+    Ok(msg.content.unwrap_or_default())
+}
+
+/// Build the concatenated node-output body for the prompt.
+fn build_combined_body(nodes: &[NodeOutput]) -> String {
+    let mut combined = String::new();
+    for node in nodes {
+        let output = crate::utils::truncate_chars(&node.output, MAX_NODE_OUTPUT_CHARS);
+        combined.push_str(&format!(
+            "\n## {} — {}\n{}\n",
+            node.id, node.directive, output
+        ));
+    }
+    combined
+}
+
+/// Fallback: a plain concatenation summary (the behaviour of the original stub).
+fn concat_summary(nodes: &[NodeOutput]) -> String {
+    let combined = build_combined_body(nodes);
+    format!(
+        "# Consensus Synthesis\n\n\
          Nodes synthesized: {}\n\n\
          ## Summary\n\
          The following node outputs were collected:\n\
@@ -31,5 +131,47 @@ pub async fn synthesize_consensus(nodes: &[NodeOutput], _tool_ctx: &ToolCtx) ->
          Review the individual node outputs above for detailed findings.",
         nodes.len(),
         combined
-    ))
+    )
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use zesdex_domain::workflow::NodeOutput;
+
+    fn node(id: &str, output: &str) -> NodeOutput {
+        NodeOutput {
+            id: id.to_string(),
+            directive: "directive".to_string(),
+            output: output.to_string(),
+        }
+    }
+
+    #[test]
+    fn build_combined_body_truncates_oversized_output() {
+        let long = "é".repeat(MAX_NODE_OUTPUT_CHARS + 500);
+        let body = build_combined_body(&[node("n1", &long)]);
+        // Must contain the header and a char-truncated (<= cap) payload without
+        // panicking on a multi-byte boundary.
+        assert!(body.contains("## n1"));
+        // Header "## n1 — directive\n" ~= 20 chars, so char count stays just above cap.
+        let char_count = body.chars().count();
+        assert!(
+            char_count <= MAX_NODE_OUTPUT_CHARS + 50,
+            "expected body near {MAX_NODE_OUTPUT_CHARS} chars, got {char_count}"
+        );
+        assert!(
+            char_count > 1000,
+            "expected a many-node output, got small: {char_count}"
+        );
+    }
+
+    #[test]
+    fn concat_summary_includes_all_node_ids() {
+        let nodes = vec![node("node-a", "a out"), node("node-b", "b out")];
+        let summary = concat_summary(&nodes);
+        assert!(summary.contains("node-a"));
+        assert!(summary.contains("node-b"));
+        assert!(summary.contains("Nodes synthesized: 2"));
+    }
 }

apps/interfaces/daemon/src/handler.rs

@@ -316,13 +316,25 @@ fn handle_submit_input(state: &mut AppStateRest, text: String) {
         in_flight: std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false)),
         abort: state.abort_flag.clone(),
         api_key: api_key.clone(),
-        model: state.settings.model.clone(),
+        model: zesdex_domain::cms::resolve_effective_model(&state.settings, &state.app_config),
         api_base: provider_cfg.as_ref().map(|cfg| cfg.api_base.clone()),
     };
 
+    // Capture state for the optional background auto-review so it can run
+    // with the same resolved provider. Done here, before `api_key` /
+    // `provider_cfg` are moved into the client below. The reviewer is
+    // fire-and-forget and skips itself when the workspace has no diff.
+    let review_enabled = state.settings.flags.review_enabled;
+    let review_key = api_key.clone();
+    let review_model =
+        zesdex_domain::cms::resolve_effective_model(&state.settings, &state.app_config);
+    let review_base = provider_cfg.as_ref().map(|c| c.api_base.clone());
+    let review_ws = params.workspace_roots.clone();
+    let review_events = params.turn_events.clone();
+
     let client = std::sync::Arc::new(zesdex_infrastructure::llm::provider::LlmClient::new(
         api_key,
-        state.settings.model.clone(),
+        zesdex_domain::cms::resolve_effective_model(&state.settings, &state.app_config),
         provider_cfg.map(|cfg| cfg.api_base.clone()),
     ));
 
@@ -348,6 +360,15 @@ fn handle_submit_input(state: &mut AppStateRest, text: String) {
     use zesdex_application::agent::AgentTurnService;
     tokio::spawn(async move {
         let _ = turn_service.run_turn(params).await;
+        if review_enabled {
+            zesdex_infrastructure::subagent::auto::engine::spawn_background_review(
+                review_ws,
+                review_events,
+                review_key,
+                review_model,
+                review_base,
+            );
+        }
     });
 }
 
@@ -465,14 +486,13 @@ fn handle_compact(state: &mut AppStateRest) {
         .get(provider_name)
         .cloned()
         .unwrap_or_default();
-    let model = state.settings.model.clone();
+    let model = zesdex_domain::cms::resolve_effective_model(&state.settings, &state.app_config);
     let api_base = provider_cfg.map(|cfg| cfg.api_base.clone());
 
     let client = zesdex_infrastructure::llm::provider::LlmClient::new(api_key, model, api_base);
 
     if let Some(ref mut rt) = state.session_runtime {
-        let tokio_rt =
+        let tokio_rt = zesdex_infrastructure::runtime::runtime();
-            tokio::runtime::Runtime::new().expect("create tokio runtime for AI compaction");
         if let Ok(()) = tokio_rt.block_on(
             zesdex_application::agent::turn_service::compact_messages_with_ai(
                 &mut rt.messages,

apps/interfaces/tui/src/turn.rs

@@ -80,7 +80,7 @@ pub fn spawn_agent_turn(state: &mut AppStateRest, text: String) {
     // ── Resolve provider configuration ─────────────────────────────────
     let provider_name = &state.settings.provider;
     let api_key = resolve_api_key(state, provider_name);
-    let model = state.settings.model.clone();
+    let model = zesdex_domain::cms::resolve_effective_model(&state.settings, &state.app_config);
     let api_base = resolve_api_base(state, provider_name);
 
     // ── Build message list ─────────────────────────────────────────────
@@ -110,6 +110,16 @@ pub fn spawn_agent_turn(state: &mut AppStateRest, text: String) {
         api_base: api_base.clone(),
     };
 
+    // Capture for the optional background auto-review before `api_key` /
+    // `model` / `api_base` move into the client below. The reviewer is
+    // fire-and-forget and skips itself when the workspace has no diff.
+    let review_enabled = state.settings.flags.review_enabled;
+    let review_key = api_key.clone();
+    let review_model = model.clone();
+    let review_base = api_base.clone();
+    let review_ws = workspace_roots.clone();
+    let review_events = turn_events.clone();
+
     let client = std::sync::Arc::new(LlmClient::new(api_key, model, api_base));
 
     let tool_ctx = ToolCtx::builder()
@@ -127,6 +137,15 @@ pub fn spawn_agent_turn(state: &mut AppStateRest, text: String) {
 
     tokio::spawn(async move {
         let _ = turn_service.run_turn(params).await;
+        if review_enabled {
+            zesdex_infrastructure::subagent::auto::engine::spawn_background_review(
+                review_ws,
+                review_events,
+                review_key,
+                review_model,
+                review_base,
+            );
+        }
     });
 }
 

apps/interfaces/ws/src/lib.rs

@@ -130,6 +130,17 @@ async fn handle_socket(mut socket: WebSocket, state: Arc<WsState>) {
                                 api_base: None,
                             };
 
+                            // Capture for the optional background auto-review
+                            // before `api_key`/`model` move into the client.
+                            // This minimal WS channel has no settings toggle, so
+                            // review fires whenever a prompt runs (consistent
+                            // with the default review_enabled=true).
+                            let review_key = api_key.clone();
+                            let review_model = model.clone();
+                            let review_base = None;
+                            let review_ws = workspace_roots.clone();
+                            let review_events = turn_events.clone();
+
                             let client = std::sync::Arc::new(
                                 zesdex_infrastructure::llm::provider::LlmClient::new(
                                     api_key, model, None,
@@ -159,6 +170,13 @@ async fn handle_socket(mut socket: WebSocket, state: Arc<WsState>) {
                             use zesdex_application::agent::AgentTurnService;
                             tokio::spawn(async move {
                                 let _ = turn_service.run_turn(params).await;
+                                zesdex_infrastructure::subagent::auto::engine::spawn_background_review(
+                                    review_ws,
+                                    review_events,
+                                    review_key,
+                                    review_model,
+                                    review_base,
+                                );
                             });
 
                             let tx_clone = tx.clone();