diff --git a/apps/infrastructure/src/agent/runner.rs b/apps/infrastructure/src/agent/runner.rs
index 49d17b4..8a8f0b1 100644
--- a/apps/infrastructure/src/agent/runner.rs
+++ b/apps/infrastructure/src/agent/runner.rs
@@ -11,6 +11,7 @@ use tracing::{debug, info, warn};
 use zesdex_domain::core::tool_call::sanitize_tool_arguments;
 use zesdex_domain::core::ChatMessage;
 use crate::llm::provider::LlmClient;
+use crate::subagent::auto::engine::trigger_auto_review;
 use crate::tools::{all_tools, tool_defs, ToolCtx};
 use crate::TurnEvent;
 
@@ -25,6 +26,8 @@ pub struct AgentTurnParams {
     pub api_key: String,
     pub model: String,
     pub api_base: Option<String>,
+    pub edit_count: u32,
+    pub consecutive_empty_reviews: u32,
 }
 
 /// Spawns an agent turn on a background OS thread.
@@ -198,6 +201,23 @@ fn run_turn(params: &mut AgentTurnParams) {
                     params
                         .messages
                         .push(ChatMessage::tool(tc.id.clone(), output.clone()));
+
+                    // Trigger auto-review after write/edit tool execution
+                    if name == "write" || name == "edit" {
+                        params.edit_count = params.edit_count.saturating_add(1);
+
+                        // If we have a workspace root, trigger review
+                        if let Some(root) = params.workspace_roots.first() {
+                            let _ = trigger_auto_review(
+                                root,
+                                params.edit_count,
+                                &mut params.consecutive_empty_reviews,
+                                3, // max_skip: skip after 3 consecutive empty reviews
+                                &params.turn_events,
+                                Some(&client),
+                            );
+                        }
+                    }
                 }
             }
             Err(e) => {
diff --git a/apps/infrastructure/src/subagent/auto/engine.rs b/apps/infrastructure/src/subagent/auto/engine.rs
new file mode 100644
index 0000000..195aa30
--- /dev/null
+++ b/apps/infrastructure/src/subagent/auto/engine.rs
@@ -0,0 +1,280 @@
+//! Auto-review engine — automatically checks git diff after file edits
+//! using an LLM subagent.
+//!
+//! Flow: after each write/edit tool execution in the agent turn, the runner
+//! calls `trigger_auto_review` which:
+//!   1. Runs `git diff --cached` and `git diff` to get working-tree changes
+//!   2. Sends the diff to a lightweight LLM call for quick review
+//!   3. Emits findings as `TurnEvent::SystemNote` on the event queue
+
+use std::collections::VecDeque;
+use std::path::Path;
+use std::process::Command;
+use std::sync::{Arc, Mutex};
+
+use anyhow::Result;
+use tracing::{debug, info, instrument, warn};
+
+use crate::llm::provider::LlmClient;
+use crate::subagent::gating::should_review;
+use crate::TurnEvent;
+
+/// Trigger an auto-review of recent git changes.
+///
+/// Flow:
+///   1. Check gating conditions (edit count, consecutive empty reviews)
+///   2. Run `git diff --cached` to get staged changes
+///   3. Run `git diff` to get unstaged changes
+///   4. If there are changes, call LLM for a quick review
+///   5. Emit findings as TurnEvent::SystemNote
+///
+/// Returns `(had_findings, total_findings)` tuple.
+#[instrument(skip(turn_events, llm_client))]
+pub fn trigger_auto_review(
+    workspace_root: &Path,
+    edit_count: u32,
+    consecutive_empty_reviews: &mut u32,
+    max_skip: u32,
+    turn_events: &Arc<Mutex<VecDeque<TurnEvent>>>,
+    llm_client: Option<&LlmClient>,
+) -> Result<(bool, usize)> {
+    // Gating check
+    if !should_review(edit_count, *consecutive_empty_reviews, max_skip) {
+        debug!("auto-review skipped by gating");
+        return Ok((false, 0));
+    }
+
+    info!("triggering auto-review");
+
+    // Run git diff to get changes
+    let diff = match get_git_diff(workspace_root) {
+        Ok(d) => d,
+        Err(e) => {
+            debug!(error = %e, "auto-review: git diff failed (not a git repo?)");
+            return Ok((false, 0));
+        }
+    };
+
+    if diff.is_empty() {
+        debug!("auto-review: no changes to review");
+        *consecutive_empty_reviews = consecutive_empty_reviews.saturating_add(1);
+        return Ok((false, 0));
+    }
+
+    // If we have an LLM client, do a real review
+    let review_result = if let Some(client) = llm_client {
+        perform_llm_review(client, &diff)?
+    } else {
+        // Fallback: simple heuristic review without LLM
+        perform_heuristic_review(&diff)
+    };
+
+    let had_findings = !review_result.is_empty();
+    let finding_count = review_result.len();
+
+    if had_findings {
+        *consecutive_empty_reviews = 0;
+        info!(finding_count, "auto-review produced findings");
+
+        // Emit findings as SystemNote events
+        for finding in &review_result {
+            let note = TurnEvent::SystemNote {
+                kind: "info".to_string(),
+                message: format!("🔍 Auto-Review: {finding}"),
+            };
+            if let Ok(mut q) = turn_events.lock() {
+                q.push_back(note);
+            }
+        }
+    } else {
+        *consecutive_empty_reviews = consecutive_empty_reviews.saturating_add(1);
+        info!("auto-review: no issues found");
+    }
+
+    Ok((had_findings, finding_count))
+}
+
+/// Run `git diff` to get workspace changes (both staged and unstaged).
+fn get_git_diff(workspace_root: &Path) -> Result<String> {
+    // Check if this is a git repo
+    let git_dir = workspace_root.join(".git");
+    if !git_dir.exists() {
+        return Ok(String::new());
+    }
+
+    // Get unstaged diff
+    let unstaged = Command::new("git")
+        .arg("diff")
+        .current_dir(workspace_root)
+        .output()
+        .map_err(|e| anyhow::anyhow!("git diff failed: {e}"))?;
+
+    // Get staged diff
+    let staged = Command::new("git")
+        .arg("diff")
+        .arg("--cached")
+        .current_dir(workspace_root)
+        .output()
+        .map_err(|e| anyhow::anyhow!("git diff --cached failed: {e}"))?;
+
+    let mut combined = String::new();
+
+    let staged_out = String::from_utf8_lossy(&staged.stdout).trim().to_string();
+    if !staged_out.is_empty() {
+        combined.push_str("=== Staged Changes ===\n");
+        combined.push_str(&staged_out);
+        combined.push('\n');
+    }
+
+    let unstaged_out = String::from_utf8_lossy(&unstaged.stdout).trim().to_string();
+    if !unstaged_out.is_empty() {
+        combined.push_str("=== Unstaged Changes ===\n");
+        combined.push_str(&unstaged_out);
+        combined.push('\n');
+    }
+
+    // Run `git diff --stat` for summary
+    let stat = Command::new("git")
+        .arg("diff")
+        .arg("--stat")
+        .current_dir(workspace_root)
+        .output()
+        .map_err(|e| anyhow::anyhow!("git diff --stat failed: {e}"))?;
+    let stat_out = String::from_utf8_lossy(&stat.stdout).trim().to_string();
+    if !stat_out.is_empty() {
+        combined.push_str("=== Summary ===\n");
+        combined.push_str(&stat_out);
+        combined.push('\n');
+    }
+
+    Ok(combined)
+}
+
+/// Perform an LLM-based review of the git diff.
+///
+/// Sends the diff to the model with a focused prompt asking it to find
+/// typos, missing imports, syntax errors, and other issues.
+fn perform_llm_review(client: &LlmClient, diff: &str) -> Result<Vec<String>> {
+    // Truncate very large diffs to avoid token limits
+    const MAX_DIFF_CHARS: usize = 8000;
+    let truncated_diff = if diff.len() > MAX_DIFF_CHARS {
+        format!("{}...\n[diff truncated at {} characters]",
+            &diff[..MAX_DIFF_CHARS], MAX_DIFF_CHARS)
+    } else {
+        diff.to_string()
+    };
+
+    let messages = vec![
+        zesdex_domain::core::ChatMessage::system(
+            "You are a focused code reviewer. Review the following git diff for:\n\
+             1. Typos and spelling errors\n\
+             2. Missing imports or undefined references\n\
+             3. Syntax errors or type mismatches\n\
+             4. Logic bugs or off-by-one errors\n\
+             5. Missing error handling\n\
+             6. Security issues\n\n\
+             Be concise. List each issue on a new line with severity [HIGH], [MEDIUM], or [LOW].\n\
+             If no issues are found, reply with exactly: NO_ISSUES_FOUND"
+                .to_string(),
+        ),
+        zesdex_domain::core::ChatMessage::user(format!(
+            "Review this git diff:\n\n```diff\n{truncated_diff}\n```"
+        )),
+    ];
+
+    match client.chat_with_tools_non_streaming(&messages, None, Some(1024), Some(0.3), None) {
+        Ok((response, _usage)) => {
+            let text = response.content.unwrap_or_default().trim().to_string();
+
+            if text.contains("NO_ISSUES_FOUND") {
+                return Ok(Vec::new());
+            }
+
+            // Parse findings line by line
+            let findings: Vec<String> = text
+                .lines()
+                .map(|l| l.trim().to_string())
+                .filter(|l| {
+                    !l.is_empty()
+                        && !l.starts_with("Here")
+                        && !l.starts_with("Let me")
+                        && !l.starts_with("I've")
+                        && !l.starts_with("The diff")
+                })
+                .collect();
+
+            Ok(findings)
+        }
+        Err(e) => {
+            warn!(error = %e, "auto-review LLM call failed");
+            Ok(Vec::new())
+        }
+    }
+}
+
+/// Perform a simple heuristic-based review without an LLM call.
+///
+/// This is a fallback when no LLM client is available. It checks for:
+///   - Missing semicolons
+///   - Unclosed brackets
+///   - `todo!()` or `unimplemented!()` macros left in code
+///   - Debug print statements
+///   - Extremely long functions
+fn perform_heuristic_review(diff: &str) -> Vec<String> {
+    let mut findings = Vec::new();
+
+    // Check for added lines (lines starting with +)
+    let added_lines: Vec<&str> = diff
+        .lines()
+        .filter(|l| l.starts_with('+') && !l.starts_with("+++"))
+        .collect();
+
+    let added_content: String = added_lines
+        .iter()
+        .map(|l| &l[1..]) // Strip leading +
+        .collect::<Vec<&str>>()
+        .join("\n");
+
+    // Check for todo! and unimplemented!
+    if added_content.contains("todo!()") {
+        findings.push("[MEDIUM] `todo!()` found in new code — replace with implementation".to_string());
+    }
+    if added_content.contains("unimplemented!()") {
+        findings.push("[MEDIUM] `unimplemented!()` found in new code — replace with implementation".to_string());
+    }
+
+    // Check for debug print statements
+    if added_content.contains("println!") || added_content.contains("dbg!") {
+        findings.push("[LOW] Debug print statements (println!/dbg!) found — consider removing before finalizing".to_string());
+    }
+    if added_content.contains("eprintln!") {
+        findings.push("[LOW] Debug eprintln! found — consider removing before finalizing".to_string());
+    }
+
+    // Check for unreachable or panic statements
+    if added_content.contains("panic!(\"reached") || added_content.contains("panic!(\"not implemented") {
+        findings.push("[HIGH] Unreachable code / panic found — implement the missing logic".to_string());
+    }
+
+    // Check for very long lines (>120 chars)
+    for (i, line) in added_lines.iter().enumerate() {
+        let content = &line[1..]; // Strip leading +
+        if content.len() > 120 && !content.trim_start().starts_with("//") {
+            let preview: String = content.chars().take(80).collect();
+            findings.push(format!(
+                "[LOW] Very long line ({} chars, line {} in diff) — consider breaking up:\n  `{}…`",
+                content.len(),
+                i + 1,
+                preview
+            ));
+        }
+    }
+
+    // Count new functions to detect very long additions
+    let fn_count = added_content.matches("fn ").count();
+    if fn_count > 5 {
+        findings.push("[INFO] Large number of new functions ({fn_count}) — consider whether this should be split into separate modules".to_string());
+    }
+
+    findings
+}
diff --git a/apps/infrastructure/src/subagent/auto/mod.rs b/apps/infrastructure/src/subagent/auto/mod.rs
index 38cc455..c107ab5 100644
--- a/apps/infrastructure/src/subagent/auto/mod.rs
+++ b/apps/infrastructure/src/subagent/auto/mod.rs
@@ -1,4 +1,5 @@
 //! Auto-subagents — automatically run review/test subagents at the end of
 //! each turn.
 
+pub mod engine;
 pub mod paths;
diff --git a/apps/infrastructure/src/subagent/mod.rs b/apps/infrastructure/src/subagent/mod.rs
index 0bdabdb..07bc11b 100644
--- a/apps/infrastructure/src/subagent/mod.rs
+++ b/apps/infrastructure/src/subagent/mod.rs
@@ -1,6 +1,7 @@
 //! Subagent spawning and execution engine — spawn managed sub-processes
 //! for test generation, architecture review, security review, etc.
 
+pub mod auto;
 pub mod context;
 pub mod division;
 pub mod engine;
diff --git a/apps/infrastructure/src/tools/mod.rs b/apps/infrastructure/src/tools/mod.rs
index c50f786..3326b92 100644
--- a/apps/infrastructure/src/tools/mod.rs
+++ b/apps/infrastructure/src/tools/mod.rs
@@ -19,13 +19,16 @@ pub mod fs;
 pub mod git;
 pub mod lsp;
 pub mod memory;
+pub mod parallel_delegate;
 pub mod plan;
 pub mod search;
+pub mod semantic_search;
 pub mod sequential_think;
 pub mod shell;
 pub mod shell_filter;
 pub mod spawn;
 pub mod utility;
+pub mod web_search;
 pub mod workflow;
 
 pub use git::git_cred;
@@ -204,6 +207,10 @@ pub fn all_tools() -> Vec<Box<dyn Tool>> {
         Box::new(lsp::LspDefinition),
         Box::new(lsp::LspReferences),
         Box::new(lsp::LspDisconnect),
+        Box::new(web_search::WebSearch),
+        Box::new(semantic_search::SemanticSearch),
+        Box::new(semantic_search::RebuildIndex),
+        Box::new(parallel_delegate::ParallelDelegate),
     ]
 }
 
diff --git a/apps/infrastructure/src/tools/parallel_delegate.rs b/apps/infrastructure/src/tools/parallel_delegate.rs
new file mode 100644
index 0000000..707a52a
--- /dev/null
+++ b/apps/infrastructure/src/tools/parallel_delegate.rs
@@ -0,0 +1,377 @@
+//! Parallel delegation tool — splits a large task into sub-tasks that run
+//! concurrently across multiple subagents, then synthesises results.
+//!
+//! Flow: receive task description + optional breakdown → LLM analyses the
+//! task and splits it into parallel directives → spawn subagents for each →
+//! collect results → return consolidated output.
+
+use anyhow::Result;
+use serde_json::{json, Value};
+use tracing::{debug, info, instrument, warn};
+
+use zesdex_domain::cms::{AppConfigRepository, SettingsRepository};
+use zesdex_domain::core::Store;
+
+use crate::persistence::{JsonAppConfigRepository, JsonSettingsRepository};
+use crate::subagent::context::SubagentContext;
+use crate::subagent::division::AccessTier;
+use crate::subagent::spawn::spawn_subagent;
+use crate::tools::{Tool, ToolCtx};
+
+/// Delegate a large task to multiple subagents running in parallel.
+///
+/// Flow:
+///   1. If `directives` are provided explicitly, use them directly.
+///   2. Otherwise, use LLM to analyse the task and split it into directives.
+///   3. Spawn a subagent for each directive concurrently.
+///   4. Join all results and consolidate into a single response.
+pub struct ParallelDelegate;
+
+impl Tool for ParallelDelegate {
+    fn name(&self) -> &'static str {
+        "parallel_delegate"
+    }
+
+    fn description(&self) -> &'static str {
+        "Delegate a large task to multiple subagents running in parallel for 3x faster completion"
+    }
+
+    fn parameters(&self) -> Value {
+        json!({
+            "type": "object",
+            "properties": {
+                "task": {
+                    "type": "string",
+                    "description": "The task to be split and delegated to parallel agents"
+                },
+                "directives": {
+                    "type": "array",
+                    "items": {
+                        "type": "object",
+                        "properties": {
+                            "directive": {"type": "string", "description": "Directive for one parallel agent"},
+                            "access": {"type": "string", "enum": ["read", "write", "full"], "description": "Access tier for this agent (default: write)"}
+                        },
+                        "required": ["directive"]
+                    },
+                    "description": "Optional explicit list of parallel directives (if not provided, the LLM will auto-split the task)"
+                },
+                "synthesize": {
+                    "type": "boolean",
+                    "description": "Whether to synthesize results into a unified response (default: true)",
+                    "default": true
+                },
+                "max_parallel": {
+                    "type": "integer",
+                    "description": "Maximum number of parallel agents (default: 3, max: 8)",
+                    "default": 3
+                }
+            },
+            "required": ["task"]
+        })
+    }
+
+    #[instrument(skip(self, ctx, args))]
+    fn run(&self, ctx: &ToolCtx, args: &Value) -> Result<String> {
+        let task = crate::tools::arg_str(args, "task")?;
+        let synthesize = args
+            .get("synthesize")
+            .and_then(|v| v.as_bool())
+            .unwrap_or(true);
+        let max_parallel = args
+            .get("max_parallel")
+            .and_then(|v| v.as_u64())
+            .unwrap_or(3)
+            .min(8) as usize;
+
+        // Load LLM credentials
+        let store = Store::new();
+        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(|| "https://opencode.ai/zen/v1".to_string());
+
+        let api_key = crate::llm::provider::resolve_api_key(&settings, &app_config);
+
+        // Determine directives
+        let directives: Vec<(String, AccessTier)> = if let Some(dirs) =
+            args.get("directives").and_then(|v| v.as_array())
+        {
+            // Explicit directives provided
+            dirs.iter()
+                .filter_map(|d| {
+                    let directive = d.get("directive").and_then(|v| v.as_str())?;
+                    let access_str = d
+                        .get("access")
+                        .and_then(|v| v.as_str())
+                        .unwrap_or("write");
+                    let access = match access_str {
+                        "read" => AccessTier::Read,
+                        "full" => AccessTier::Full,
+                        _ => AccessTier::Write,
+                    };
+                    Some((directive.to_string(), access))
+                })
+                .collect()
+        } else {
+            // Auto-split using LLM
+            let rt = tokio::runtime::Runtime::new()?;
+            let directives = rt.block_on(auto_split_task(
+                &task,
+                max_parallel,
+                &base_url,
+                &api_key,
+                &model,
+            ))?;
+            directives
+        };
+
+        if directives.is_empty() {
+            anyhow::bail!("no directives could be derived for the task");
+        }
+
+        let directive_count = directives.len();
+        info!(
+            task = %task,
+            directive_count,
+            max_parallel,
+            "parallel delegation: starting subagents"
+        );
+
+        // Spawn agents in parallel
+        let mut handles = Vec::new();
+        for (i, (directive, access)) in directives.iter().enumerate() {
+            let subagent_ctx = SubagentContext::new(
+                directive.clone(),
+                ctx.clone(),
+                format!("{access:?}"),
+                base_url.clone(),
+                api_key.clone(),
+                model.clone(),
+            );
+
+            debug!(agent_index = i, access = ?access, "spawning parallel agent");
+            let handle = spawn_subagent(
+                subagent_ctx,
+                directive.clone(),
+                access.clone(),
+                ctx.clone(),
+            );
+            handles.push((i, handle));
+        }
+
+        // Join all results
+        let mut results: Vec<(usize, String, String)> = Vec::new();
+        for (i, handle) in handles {
+            match handle.join() {
+                Ok(Ok(output)) => {
+                    info!(agent_index = i, "parallel agent completed");
+                    results.push((i, directives[i].0.clone(), output));
+                }
+                Ok(Err(e)) => {
+                    warn!(agent_index = i, error = %e, "parallel agent failed");
+                    results.push((
+                        i,
+                        directives[i].0.clone(),
+                        format!("[ERROR] {e}"),
+                    ));
+                }
+                Err(e) => {
+                    warn!(agent_index = i, error = ?e, "parallel agent panicked");
+                    results.push((
+                        i,
+                        directives[i].0.clone(),
+                        "[ERROR] Agent panicked".to_string(),
+                    ));
+                }
+            }
+        }
+
+        // Consolidate results
+        if synthesize && results.len() > 1 {
+            let consolidated = consolidate_results(&results, &base_url, &api_key, &model)?;
+            Ok(format!(
+                "## Parallel Delegation Complete\n\n**Task:** {task}\n**Parallel agents:** {}\n\n{}",
+                results.len(),
+                consolidated
+            ))
+        } else {
+            let mut output = format!(
+                "## Parallel Delegation Complete\n\n**Task:** {task}\n**Parallel agents:** {}\n\n",
+                results.len()
+            );
+            for (i, directive, result) in &results {
+                output.push_str(&format!("---\n### Agent {}: {}\n\n{}\n", i, directive, result));
+            }
+            Ok(output)
+        }
+    }
+}
+
+/// Use LLM to analyse a task and split it into parallel directives.
+async fn auto_split_task(
+    task: &str,
+    max_parallel: usize,
+    base_url: &str,
+    api_key: &str,
+    model: &str,
+) -> Result<Vec<(String, AccessTier)>> {
+    let client = crate::llm::provider::LlmClient::new(
+        api_key.to_string(),
+        model.to_string(),
+        Some(base_url.to_string()),
+    );
+
+    let sys_msg = zesdex_domain::core::ChatMessage::system(
+        format!(
+            "You are a task decomposition expert. Split the following task into {max_parallel} \
+             independent sub-tasks that can run in parallel. Each sub-task must be self-contained \
+             and produce useful output independently.\n\n\
+             Output your response as a JSON array of objects, each with:\n\
+             - \"directive\": a clear, detailed instruction for a subagent\n\
+             - \"access\": one of \"read\", \"write\", or \"full\"\n\n\
+             IMPORTANT: Return ONLY valid JSON, no other text. Example:\n\
+             [{{\"directive\": \"Create the User model with fields...\", \"access\": \"write\"}}]"
+        )
+    );
+
+    let user_msg =
+        zesdex_domain::core::ChatMessage::user(format!("Task: {task}\n\nSplit into {max_parallel} parallel directives:"));
+
+    match client
+        .chat_with_tools_non_streaming(&[sys_msg, user_msg], None, Some(2048), Some(0.4), None)
+    {
+        Ok((response, _)) => {
+            let text = response.content.unwrap_or_default();
+            parse_directives_json(&text, max_parallel)
+        }
+        Err(e) => {
+            warn!(error = %e, "auto-split LLM call failed, using fallback splitting");
+            Ok(fallback_split(task, max_parallel))
+        }
+    }
+}
+
+/// Parse JSON directives from LLM response.
+fn parse_directives_json(text: &str, max_parallel: usize) -> Result<Vec<(String, AccessTier)>> {
+    // Find JSON array in the response
+    let json_start = text.find('[');
+    let json_end = text.rfind(']');
+
+    let json_str = match (json_start, json_end) {
+        (Some(start), Some(end)) if end > start => &text[start..=end],
+        _ => {
+            // Try parsing the whole text as JSON
+            text.trim()
+        }
+    };
+
+    let parsed: Vec<Value> = serde_json::from_str(json_str)
+        .map_err(|e| anyhow::anyhow!("failed to parse directives JSON: {e}"))?;
+
+    let directives: Vec<(String, AccessTier)> = parsed
+        .into_iter()
+        .take(max_parallel)
+        .filter_map(|v| {
+            let directive = v.get("directive")?.as_str()?.to_string();
+            let access_str = v.get("access").and_then(|a| a.as_str()).unwrap_or("write");
+            let access = match access_str {
+                "read" => AccessTier::Read,
+                "full" => AccessTier::Full,
+                _ => AccessTier::Write,
+            };
+            Some((directive, access))
+        })
+        .collect();
+
+    Ok(directives)
+}
+
+/// Fallback splitting when LLM call fails.
+fn fallback_split(task: &str, max_parallel: usize) -> Vec<(String, AccessTier)> {
+    // Simple heuristic split by common patterns
+    let mut directives = Vec::new();
+
+    if task.contains("backend") || task.contains("api") || task.contains("server") {
+        directives.push((
+            format!("Implement the backend/API components for: {task}"),
+            AccessTier::Write,
+        ));
+    }
+
+    if task.contains("frontend") || task.contains("ui") || task.contains("client") {
+        directives.push((
+            format!("Implement the frontend/UI components for: {task}"),
+            AccessTier::Write,
+        ));
+    }
+
+    if task.contains("test") || task.contains("unit") {
+        directives.push((
+            format!("Write unit tests for: {task}"),
+            AccessTier::Read,
+        ));
+    }
+
+    if directives.is_empty() {
+        // Just split the task into generic chunks
+        for i in 0..max_parallel {
+            directives.push((
+                format!("Part {} of parallel task: {task}", i + 1),
+                AccessTier::Write,
+            ));
+        }
+    }
+
+    directives.truncate(max_parallel);
+    directives
+}
+
+/// Use LLM to consolidate multiple agent results into a single response.
+fn consolidate_results(
+    results: &[(usize, String, String)],
+    base_url: &str,
+    api_key: &str,
+    model: &str,
+) -> Result<String> {
+    let client = crate::llm::provider::LlmClient::new(
+        api_key.to_string(),
+        model.to_string(),
+        Some(base_url.to_string()),
+    );
+
+    let mut summary = String::new();
+    for (i, directive, result) in results {
+        summary.push_str(&format!("### Agent {}: {}\n{}\n\n", i, directive, result));
+    }
+
+    let sys_msg = zesdex_domain::core::ChatMessage::system(
+        "You are a synthesis expert. Consolidate the following parallel agent outputs \
+         into a single coherent response. Remove duplication, reconcile conflicts, and \
+         present the unified result in a well-structured format."
+            .to_string(),
+    );
+
+    let user_msg = zesdex_domain::core::ChatMessage::user(format!(
+        "Consolidate the following parallel agent outputs:\n\n{summary}"
+    ));
+
+    match client.chat_with_tools_non_streaming(&[sys_msg, user_msg], None, Some(2048), Some(0.3), None) {
+        Ok((response, _)) => Ok(response.content.unwrap_or_else(|| summary.clone())),
+        Err(e) => {
+            warn!(error = %e, "consolidation LLM call failed, using raw concatenation");
+            Ok(summary)
+        }
+    }
+}
diff --git a/apps/infrastructure/src/tools/semantic_search.rs b/apps/infrastructure/src/tools/semantic_search.rs
new file mode 100644
index 0000000..626ea9b
--- /dev/null
+++ b/apps/infrastructure/src/tools/semantic_search.rs
@@ -0,0 +1,690 @@
+//! Semantic code search tool — indexes all code symbols (functions, structs,
+//! enums, traits, modules) in a project and allows searching by name,
+//! concept, or meaning.
+//!
+//! Flow: walk workspace files → parse Rust source for symbol declarations →
+//! build in-memory index → search by fuzzy/prefix match on symbol names and
+//! doc comments.
+
+use crate::tools::{Tool, ToolCtx};
+use anyhow::Result;
+use regex::Regex;
+use serde::{Deserialize, Serialize};
+use serde_json::{json, Value};
+use std::collections::HashMap;
+use std::sync::Mutex;
+use tracing::{debug, info, instrument, warn};
+
+// ---------------------------------------------------------------------------
+//  Symbol index types
+// ---------------------------------------------------------------------------
+
+/// The kind of a code symbol.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
+pub enum SymbolKind {
+    Function,
+    Struct,
+    Enum,
+    Trait,
+    Module,
+    Impl,
+    Type,
+    Constant,
+    Macro,
+    Other,
+}
+
+impl std::fmt::Display for SymbolKind {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            SymbolKind::Function => write!(f, "fn"),
+            SymbolKind::Struct => write!(f, "struct"),
+            SymbolKind::Enum => write!(f, "enum"),
+            SymbolKind::Trait => write!(f, "trait"),
+            SymbolKind::Module => write!(f, "mod"),
+            SymbolKind::Impl => write!(f, "impl"),
+            SymbolKind::Type => write!(f, "type"),
+            SymbolKind::Constant => write!(f, "const"),
+            SymbolKind::Macro => write!(f, "macro"),
+            SymbolKind::Other => write!(f, "symbol"),
+        }
+    }
+}
+
+/// A single code symbol entry in the index.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct CodeSymbol {
+    /// Symbol name (e.g. "run_agent", "AppStateRest").
+    pub name: String,
+    /// Kind of symbol.
+    pub kind: SymbolKind,
+    /// File path relative to workspace root.
+    pub file: String,
+    /// Line number (1-indexed).
+    pub line: usize,
+    /// Parent symbol (e.g. struct name for impl methods).
+    pub parent: Option<String>,
+    /// Doc comment text, if any.
+    pub doc_comment: Option<String>,
+    /// Short context (the declaration line).
+    pub context: String,
+}
+
+/// The in-memory symbol index, shared via a global static.
+static SYMBOL_INDEX: Mutex<Option<SymbolIndex>> = Mutex::new(None);
+
+/// Lazily compile regexes for symbol extraction.
+fn compiled_regexes() -> (
+    Regex,
+    Regex,
+    Regex,
+    Regex,
+    Regex,
+    Regex,
+    Regex,
+    Regex,
+    Regex,
+) {
+    (
+        Regex::new(r"(?m)^\s*(?:pub\s+)?(?:(?:unsafe\s+)?async\s+)?fn\s+(\w+)").unwrap(),
+        Regex::new(r"(?m)^\s*(?:pub\s+)?struct\s+(\w+)").unwrap(),
+        Regex::new(r"(?m)^\s*(?:pub\s+)?enum\s+(\w+)").unwrap(),
+        Regex::new(r"(?m)^\s*(?:pub\s+)?(?:(?:unsafe\s+)?)?trait\s+(\w+)").unwrap(),
+        Regex::new(r"(?m)^\s*(?:pub\s+)?mod\s+(\w+)").unwrap(),
+        Regex::new(r"(?m)^\s*(?:pub\s+)?(?:unsafe\s+)?impl(?:\s*<[^>]*>)?\s+(?:for\s+)?(\w+)").unwrap(),
+        Regex::new(r"(?m)^\s*(?:pub\s+)?type\s+(\w+)").unwrap(),
+        Regex::new(r"(?m)^\s*(?:pub\s+)?const\s+(\w+)").unwrap(),
+        Regex::new(r"(?m)^\s*(?:pub\s+)?macro_rules!\s*\(\s*(\w+)").unwrap(),
+    )
+}
+
+/// A symbol index cache.
+#[derive(Debug, Clone)]
+pub struct SymbolIndex {
+    symbols: Vec<CodeSymbol>,
+    workspace_path: Option<String>,
+}
+
+impl SymbolIndex {
+    pub fn new() -> Self {
+        SymbolIndex {
+            symbols: Vec::new(),
+            workspace_path: None,
+        }
+    }
+
+    pub fn is_empty(&self) -> bool {
+        self.symbols.is_empty()
+    }
+
+    pub fn len(&self) -> usize {
+        self.symbols.len()
+    }
+
+    pub fn rebuild(&mut self, workspace: &str) -> Result<usize> {
+        let path = std::path::Path::new(workspace);
+        if !path.exists() {
+            anyhow::bail!("workspace path does not exist: {workspace}");
+        }
+
+        let mut symbols = Vec::new();
+        let walker = ignore::Walk::new(path);
+
+        for entry in walker.flatten() {
+            let file_path = entry.path();
+            if !file_path.is_file() {
+                continue;
+            }
+
+            // Only index Rust files
+            let ext = file_path.extension().and_then(|e| e.to_str()).unwrap_or("");
+            if ext != "rs" {
+                continue;
+            }
+
+            let rel_path = file_path
+                .strip_prefix(path)
+                .unwrap_or(file_path)
+                .display()
+                .to_string();
+
+            match std::fs::read_to_string(file_path) {
+                Ok(content) => {
+                    let file_symbols = extract_symbols(&content, &rel_path);
+                    symbols.extend(file_symbols);
+                }
+                Err(e) => {
+                    debug!(file = %rel_path, error = %e, "failed to read file for indexing");
+                }
+            }
+        }
+
+        symbols.sort_by(|a, b| a.name.cmp(&b.name));
+        self.symbols = symbols;
+        self.workspace_path = Some(workspace.to_string());
+
+        let count = self.symbols.len();
+        info!(symbol_count = count, workspace = %workspace, "symbol index rebuilt");
+        Ok(count)
+    }
+
+    pub fn search(&self, query: &str, max_results: usize) -> Vec<&CodeSymbol> {
+        if self.symbols.is_empty() {
+            return Vec::new();
+        }
+
+        let query_lower = query.to_lowercase();
+        let query_words: Vec<String> = query_lower.split_whitespace().map(|s| s.to_string()).collect();
+
+        let mut scored: Vec<(i32, &CodeSymbol)> = self
+            .symbols
+            .iter()
+            .filter_map(|sym| {
+                let score = score_symbol(sym, &query_lower, &query_words);
+                if score > 0 {
+                    Some((score, sym))
+                } else {
+                    None
+                }
+            })
+            .collect();
+
+        // Sort by score descending, then by name ascending
+        scored.sort_by(|a, b| b.0.cmp(&a.0).then_with(|| a.1.name.cmp(&b.1.name)));
+
+        scored
+            .into_iter()
+            .take(max_results)
+            .map(|(_, sym)| sym)
+            .collect()
+    }
+}
+
+impl Default for SymbolIndex {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+/// Score a symbol against a search query.
+fn score_symbol(sym: &CodeSymbol, query_lower: &str, query_words: &[String]) -> i32 {
+    let name_lower = sym.name.to_lowercase();
+    let mut score: i32 = 0;
+
+    // Exact match = highest score
+    if name_lower == *query_lower {
+        score += 1000;
+    }
+
+    // Prefix match
+    if name_lower.starts_with(query_lower) {
+        score += 500;
+    }
+
+    // Contains
+    if name_lower.contains(query_lower) {
+        score += 200;
+    }
+
+    // Word-by-word matching
+    for word in query_words {
+        if name_lower.contains(word) {
+            score += 50;
+        }
+    }
+
+    // Doc comment match
+    if let Some(ref doc) = sym.doc_comment {
+        let doc_lower = doc.to_lowercase();
+        if doc_lower.contains(query_lower) {
+            score += 30;
+        }
+        for word in query_words {
+            if doc_lower.contains(word) {
+                score += 10;
+            }
+        }
+    }
+
+    // Context match
+    let context_lower = sym.context.to_lowercase();
+    if context_lower.contains(query_lower) {
+        score += 20;
+    }
+
+    score
+}
+
+/// Extract code symbols from Rust source content.
+fn extract_symbols(content: &str, rel_path: &str) -> Vec<CodeSymbol> {
+    let (fn_re, struct_re, enum_re, trait_re, mod_re, impl_re, type_re, const_re, macro_re) =
+        compiled_regexes();
+
+    let mut symbols = Vec::new();
+    let lines: Vec<&str> = content.lines().collect();
+
+    // Extract doc comments that precede declarations
+    let doc_comments = extract_doc_comments(&lines);
+
+    for (i, line) in lines.iter().enumerate() {
+        let line_num = i + 1;
+        let trimmed = line.trim();
+
+        // Check for function declarations
+        if let Some(caps) = fn_re.captures(trimmed) {
+            let name = caps.get(1).unwrap().as_str().to_string();
+            let doc = doc_comments.get(&line_num).cloned();
+            symbols.push(CodeSymbol {
+                name,
+                kind: SymbolKind::Function,
+                file: rel_path.to_string(),
+                line: line_num,
+                parent: None,
+                doc_comment: doc,
+                context: trimmed.to_string(),
+            });
+        }
+
+        // Check for struct declarations
+        if let Some(caps) = struct_re.captures(trimmed) {
+            let name = caps.get(1).unwrap().as_str().to_string();
+            let doc = doc_comments.get(&line_num).cloned();
+            symbols.push(CodeSymbol {
+                name,
+                kind: SymbolKind::Struct,
+                file: rel_path.to_string(),
+                line: line_num,
+                parent: None,
+                doc_comment: doc,
+                context: trimmed.to_string(),
+            });
+        }
+
+        // Check for enum declarations
+        if let Some(caps) = enum_re.captures(trimmed) {
+            let name = caps.get(1).unwrap().as_str().to_string();
+            let doc = doc_comments.get(&line_num).cloned();
+            symbols.push(CodeSymbol {
+                name,
+                kind: SymbolKind::Enum,
+                file: rel_path.to_string(),
+                line: line_num,
+                parent: None,
+                doc_comment: doc,
+                context: trimmed.to_string(),
+            });
+        }
+
+        // Check for trait declarations
+        if let Some(caps) = trait_re.captures(trimmed) {
+            let name = caps.get(1).unwrap().as_str().to_string();
+            let doc = doc_comments.get(&line_num).cloned();
+            symbols.push(CodeSymbol {
+                name,
+                kind: SymbolKind::Trait,
+                file: rel_path.to_string(),
+                line: line_num,
+                parent: None,
+                doc_comment: doc,
+                context: trimmed.to_string(),
+            });
+        }
+
+        // Check for module declarations
+        if let Some(caps) = mod_re.captures(trimmed) {
+            let name = caps.get(1).unwrap().as_str().to_string();
+            symbols.push(CodeSymbol {
+                name,
+                kind: SymbolKind::Module,
+                file: rel_path.to_string(),
+                line: line_num,
+                parent: None,
+                doc_comment: None,
+                context: trimmed.to_string(),
+            });
+        }
+
+        // Check for type alias declarations
+        if let Some(caps) = type_re.captures(trimmed) {
+            let name = caps.get(1).unwrap().as_str().to_string();
+            let doc = doc_comments.get(&line_num).cloned();
+            symbols.push(CodeSymbol {
+                name,
+                kind: SymbolKind::Type,
+                file: rel_path.to_string(),
+                line: line_num,
+                parent: None,
+                doc_comment: doc,
+                context: trimmed.to_string(),
+            });
+        }
+
+        // Check for const declarations
+        if let Some(caps) = const_re.captures(trimmed) {
+            let name = caps.get(1).unwrap().as_str().to_string();
+            let doc = doc_comments.get(&line_num).cloned();
+            symbols.push(CodeSymbol {
+                name,
+                kind: SymbolKind::Constant,
+                file: rel_path.to_string(),
+                line: line_num,
+                parent: None,
+                doc_comment: doc,
+                context: trimmed.to_string(),
+            });
+        }
+
+        // Check for macro declarations
+        if let Some(caps) = macro_re.captures(trimmed) {
+            let name = caps.get(1).unwrap().as_str().to_string();
+            symbols.push(CodeSymbol {
+                name,
+                kind: SymbolKind::Macro,
+                file: rel_path.to_string(),
+                line: line_num,
+                parent: None,
+                doc_comment: None,
+                context: trimmed.to_string(),
+            });
+        }
+
+        // Parse impl blocks for method-level indexing
+        if let Some(caps) = impl_re.captures(trimmed) {
+            let impl_for = caps.get(1).unwrap().as_str().to_string();
+            // Look for methods inside this impl block
+            let mut brace_depth: i32 = 0;
+            let mut started = false;
+            for (j, l) in lines[i..].iter().enumerate() {
+                for ch in l.chars() {
+                    match ch {
+                        '{' => {
+                            brace_depth += 1;
+                            started = true;
+                        }
+                        '}' => {
+                            brace_depth -= 1;
+                        }
+                        _ => {}
+                    }
+                }
+                if started && brace_depth <= 0 && j > 1 {
+                    break; // End of impl block
+                }
+                if j > 0 {
+                    let inner_line = l.trim();
+                    if let Some(mcaps) = fn_re.captures(inner_line) {
+                        let method_name = mcaps.get(1).unwrap().as_str().to_string();
+                        let abs_line = i + j + 1;
+                        let doc = doc_comments.get(&abs_line).cloned();
+                        symbols.push(CodeSymbol {
+                            name: format!("{impl_for}::{method_name}"),
+                            kind: SymbolKind::Function,
+                            file: rel_path.to_string(),
+                            line: abs_line,
+                            parent: Some(impl_for.clone()),
+                            doc_comment: doc,
+                            context: inner_line.to_string(),
+                        });
+                    }
+                }
+            }
+        }
+    }
+
+    symbols
+}
+
+/// Extract doc comments (/// or //!) that precede each line.
+fn extract_doc_comments(lines: &[&str]) -> HashMap<usize, String> {
+    let mut map = HashMap::new();
+    let mut i = 0;
+    while i < lines.len() {
+        let line = lines[i].trim();
+        if line.starts_with("///") {
+            let mut doc = String::new();
+            while i < lines.len() {
+                let l = lines[i].trim();
+                if l.starts_with("///") {
+                    if !doc.is_empty() {
+                        doc.push(' ');
+                    }
+                    doc.push_str(l.trim_start_matches("///").trim());
+                    i += 1;
+                } else {
+                    break;
+                }
+            }
+            // Associate doc with the next non-empty, non-doc, non-attribute line
+            let target_line = find_next_declaration_line(lines, i);
+            if let Some(tl) = target_line {
+                map.insert(tl + 1, doc);
+            }
+        } else {
+            i += 1;
+        }
+    }
+    map
+}
+
+/// Find the next line that looks like a declaration (not doc, not attr).
+fn find_next_declaration_line(lines: &[&str], start: usize) -> Option<usize> {
+    for i in start..lines.len() {
+        let trimmed = lines[i].trim();
+        if trimmed.is_empty()
+            || trimmed.starts_with("///")
+            || trimmed.starts_with("//!")
+            || trimmed.starts_with('#')
+        {
+            continue;
+        }
+        return Some(i);
+    }
+    None
+}
+
+// ---------------------------------------------------------------------------
+//  Tool: SemanticSearch — search the symbol index
+// ---------------------------------------------------------------------------
+
+/// Search for code symbols by name, concept, or semantic meaning.
+///
+/// Flow: ensure index is built → search by query → return formatted results.
+pub struct SemanticSearch;
+
+impl Tool for SemanticSearch {
+    fn name(&self) -> &'static str {
+        "semantic_search"
+    }
+
+    fn description(&self) -> &'static str {
+        "Search for code symbols (functions, structs, enums, traits) by name, concept, or meaning"
+    }
+
+    fn parameters(&self) -> Value {
+        json!({
+            "type": "object",
+            "properties": {
+                "query": {
+                    "type": "string",
+                    "description": "Search query — function name, struct name, or concept (e.g. 'payment handler', 'auth middleware', 'user repository')"
+                },
+                "kind": {
+                    "type": "string",
+                    "enum": ["fn", "struct", "enum", "trait", "mod", "all"],
+                    "description": "Filter by symbol kind (default: all)",
+                    "default": "all"
+                },
+                "max_results": {
+                    "type": "integer",
+                    "description": "Maximum results (default 10, max 30)",
+                    "default": 10
+                },
+                "rebuild_index": {
+                    "type": "boolean",
+                    "description": "Force rebuild the symbol index before searching (default false)",
+                    "default": false
+                }
+            },
+            "required": ["query"]
+        })
+    }
+
+    #[instrument(skip(self, ctx, args))]
+    fn run(&self, ctx: &ToolCtx, args: &Value) -> Result<String> {
+        let query = crate::tools::arg_str(args, "query")?;
+        let kind_filter = args
+            .get("kind")
+            .and_then(|v| v.as_str())
+            .unwrap_or("all");
+        let max_results = args
+            .get("max_results")
+            .and_then(|v| v.as_u64())
+            .unwrap_or(10)
+            .min(30) as usize;
+        let rebuild = args
+            .get("rebuild_index")
+            .and_then(|v| v.as_bool())
+            .unwrap_or(false);
+
+        let workspace = ctx
+            .workspaces
+            .first()
+            .map(|p| p.to_string_lossy().to_string())
+            .unwrap_or_else(|| ".".to_string());
+
+        info!(query = %query, kind = %kind_filter, max_results, rebuild, "semantic search");
+
+        // Get or rebuild the index
+        let mut guard = SYMBOL_INDEX.lock().map_err(|e| anyhow::anyhow!("index lock failed: {e}"))?;
+        let index = guard.get_or_insert_with(SymbolIndex::new);
+
+        if rebuild || index.is_empty() {
+            let count = index.rebuild(&workspace)?;
+            debug!(symbol_count = count, "symbol index rebuilt");
+        }
+
+        let results = index.search(&query, max_results * 2); // Get extra for filtering
+
+        // Apply kind filter
+        let filtered: Vec<&&CodeSymbol> = if kind_filter != "all" {
+            let target_kind = match kind_filter {
+                "fn" => SymbolKind::Function,
+                "struct" => SymbolKind::Struct,
+                "enum" => SymbolKind::Enum,
+                "trait" => SymbolKind::Trait,
+                "mod" => SymbolKind::Module,
+                _ => SymbolKind::Other,
+            };
+            results
+                .iter()
+                .filter(|s| s.kind == target_kind)
+                .take(max_results)
+                .collect()
+        } else {
+            results.iter().take(max_results).collect()
+        };
+
+        if filtered.is_empty() {
+            return Ok(format!(
+                "No symbols found matching '{query}'.\n\
+                 Try a different query, or use `rebuild_index: true` to rebuild the index first."
+            ));
+        }
+
+        let total = index.len();
+        info!(matched = filtered.len(), total_indexed = total, "semantic search completed");
+
+        // Group results by file for cleaner output
+        let mut by_file: std::collections::BTreeMap<String, Vec<&&CodeSymbol>> =
+            std::collections::BTreeMap::new();
+        for sym in &filtered {
+            by_file.entry(sym.file.clone()).or_default().push(*sym);
+        }
+
+        let mut output = format!(
+            "## Semantic Search Results\n\n**Query:** {query}\n**Index size:** {total} symbols\n**Matches:** {}\n\n",
+            filtered.len()
+        );
+
+        for (file, symbols) in &by_file {
+            output.push_str(&format!("### `{file}`\n\n"));
+            for sym in symbols {
+                let kind_str = sym.kind.to_string();
+                let parent_str = sym
+                    .parent
+                    .as_ref()
+                    .map(|p| format!(" [{p}]"))
+                    .unwrap_or_default();
+                let doc_str = sym
+                    .doc_comment
+                    .as_ref()
+                    .map(|d| {
+                        let truncated: String = d.chars().take(100).collect();
+                        format!(" — {truncated}")
+                    })
+                    .unwrap_or_default();
+
+                let context_trimmed = sym.context.trim();
+                let context_ellipsis = if context_trimmed.len() > 80 { "…" } else { "" };
+
+                output.push_str(&format!(
+                    "- `{kind_str}` **{}**{} at line {} `{}`{}{}\n",
+                    sym.name,
+                    parent_str,
+                    sym.line,
+                    context_trimmed,
+                    doc_str,
+                    context_ellipsis
+                ));
+            }
+            output.push('\n');
+        }
+
+        output.push_str(&format!(
+            "---\n*{} symbols indexed. Use `rebuild_index: true` to refresh.*\n",
+            total
+        ));
+
+        Ok(output)
+    }
+}
+
+/// Tool: Rebuild the symbol index explicitly.
+pub struct RebuildIndex;
+
+impl Tool for RebuildIndex {
+    fn name(&self) -> &'static str {
+        "rebuild_index"
+    }
+
+    fn description(&self) -> &'static str {
+        "Rebuild the code symbol index for semantic search"
+    }
+
+    fn parameters(&self) -> Value {
+        json!({
+            "type": "object",
+            "properties": {}
+        })
+    }
+
+    #[instrument(skip(self, ctx, _args))]
+    fn run(&self, ctx: &ToolCtx, _args: &Value) -> Result<String> {
+        let workspace = ctx
+            .workspaces
+            .first()
+            .map(|p| p.to_string_lossy().to_string())
+            .unwrap_or_else(|| ".".to_string());
+
+        info!("rebuilding symbol index");
+
+        let mut guard = SYMBOL_INDEX.lock().map_err(|e| anyhow::anyhow!("index lock failed: {e}"))?;
+        let index = guard.get_or_insert_with(SymbolIndex::new);
+        let count = index.rebuild(&workspace)?;
+
+        Ok(format!(
+            "Symbol index rebuilt successfully. {} symbols indexed.",
+            count
+        ))
+    }
+}
diff --git a/apps/infrastructure/src/tools/web_search.rs b/apps/infrastructure/src/tools/web_search.rs
new file mode 100644
index 0000000..b643cef
--- /dev/null
+++ b/apps/infrastructure/src/tools/web_search.rs
@@ -0,0 +1,283 @@
+//! Web search & documentation scraper tool — search the web for docs, APIs,
+//! and troubleshooting information using a SearXNG instance.
+//!
+//! Flow: take a search query → call SearXNG JSON API → parse results →
+//! optionally fetch full page content → return formatted markdown.
+
+use anyhow::Result;
+use serde_json::{json, Value};
+use tracing::{debug, info, instrument, warn};
+
+use crate::tools::{arg_str, Tool, ToolCtx};
+
+/// Default SearXNG instance URL.
+const DEFAULT_SEARXNG_URL: &str = "https://searxng.imrnes.team";
+
+/// Search the web for documentation, APIs, and technical information.
+///
+/// Flow: build query → call SearXNG JSON endpoint → parse results →
+/// optionally scrape full page content → return formatted result.
+pub struct WebSearch;
+
+impl Tool for WebSearch {
+    fn name(&self) -> &'static str {
+        "web_search"
+    }
+
+    fn description(&self) -> &'static str {
+        "Search the web for documentation, APIs, and technical information. Uses SearXNG instance."
+    }
+
+    fn parameters(&self) -> Value {
+        json!({
+            "type": "object",
+            "properties": {
+                "query": {
+                    "type": "string",
+                    "description": "Search query (e.g. 'Next.js 15 app router documentation')"
+                },
+                "max_results": {
+                    "type": "integer",
+                    "description": "Maximum number of results to return (default 5, max 10)",
+                    "default": 5
+                },
+                "fetch_content": {
+                    "type": "boolean",
+                    "description": "Whether to fetch full page content from each result (default false)",
+                    "default": false
+                },
+                "categories": {
+                    "type": "string",
+                    "description": "Search categories (e.g. 'general', 'science', 'it', 'news'). Default: general",
+                    "default": "general"
+                }
+            },
+            "required": ["query"]
+        })
+    }
+
+    #[instrument(skip(self, _ctx, args))]
+    fn run(&self, _ctx: &ToolCtx, args: &Value) -> Result<String> {
+        let query = arg_str(args, "query")?;
+        let max_results = args
+            .get("max_results")
+            .and_then(|v| v.as_u64())
+            .unwrap_or(5)
+            .min(10) as usize;
+        let fetch_content = args
+            .get("fetch_content")
+            .and_then(|v| v.as_bool())
+            .unwrap_or(false);
+        let categories = args
+            .get("categories")
+            .and_then(|v| v.as_str())
+            .unwrap_or("general");
+
+        info!(query = %query, max_results, fetch_content, categories = %categories, "web search starting");
+
+        let searxng_url = std::env::var("SEARXNG_URL")
+            .unwrap_or_else(|_| DEFAULT_SEARXNG_URL.to_string());
+
+        // Build the SearXNG JSON search URL
+        let search_url = format!(
+            "{}/search?format=json&q={}&categories={}&language=en-US",
+            searxng_url,
+            urlencoding(&query),
+            categories
+        );
+
+        debug!(search_url = %search_url, "calling SearXNG API");
+
+        // Make the HTTP request
+        let client = reqwest::blocking::Client::builder()
+            .timeout(std::time::Duration::from_secs(30))
+            .user_agent("Zesdex/1.0 (AI Coding Agent)")
+            .build()
+            .map_err(|e| anyhow::anyhow!("failed to create HTTP client: {e}"))?;
+
+        let response = client
+            .get(&search_url)
+            .send()
+            .map_err(|e| anyhow::anyhow!("search request failed: {e}"))?;
+
+        if !response.status().is_success() {
+            warn!(status = %response.status(), "SearXNG returned non-success status");
+            anyhow::bail!("SearXNG returned HTTP {}", response.status());
+        }
+
+        let body: Value = response
+            .json()
+            .map_err(|e| anyhow::anyhow!("failed to parse SearXNG JSON response: {e}"))?;
+
+        let results = body
+            .get("results")
+            .and_then(|v| v.as_array())
+            .map(|arr| {
+                arr.iter()
+                    .filter(|r| {
+                        r.get("url").and_then(|u| u.as_str()).is_some()
+                            && r.get("title").and_then(|t| t.as_str()).is_some()
+                    })
+                    .take(max_results)
+                    .collect::<Vec<_>>()
+            })
+            .unwrap_or_default();
+
+        let result_count = results.len();
+        if results.is_empty() {
+            info!("no search results found");
+            return Ok(format!("No search results found for '{query}'."));
+        }
+
+        info!(result_count, "web search completed");
+
+        let mut output = format!("## Web Search Results for: {query}\n\n");
+
+        for (i, result) in results.iter().enumerate() {
+            let title = result
+                .get("title")
+                .and_then(|v| v.as_str())
+                .unwrap_or("Untitled");
+            let url = result
+                .get("url")
+                .and_then(|v| v.as_str())
+                .unwrap_or("");
+            let snippet = result
+                .get("content")
+                .and_then(|v| v.as_str())
+                .unwrap_or("")
+                .trim();
+
+            output.push_str(&format!("### {}. [{title}]({url})\n\n", i + 1));
+
+            if !snippet.is_empty() {
+                output.push_str(&format!("> {snippet}\n\n"));
+            }
+
+            // Optionally fetch full page content
+            if fetch_content && !url.is_empty() {
+                match fetch_page_content(url, &client) {
+                    Ok(content) => {
+                        let truncated = limit_lines(&content, 80);
+                        output.push_str(&format!("**Content:**\n```\n{truncated}\n```\n\n"));
+                    }
+                    Err(e) => {
+                        debug!(url = %url, error = %e, "failed to fetch page content");
+                        output.push_str(&format!("*(Content fetch failed: {e})*\n\n"));
+                    }
+                }
+            }
+
+            // Show engine info if available
+            if let Some(engine) = result.get("engine").and_then(|v| v.as_str()) {
+                output.push_str(&format!("*Source: {engine}*\n\n"));
+            }
+        }
+
+        // Add suggestion for more specific queries if very few results
+        if result_count < 3 {
+            output.push_str("---\n*Few results. Try a more specific query or different categories.*\n");
+        }
+
+        Ok(output)
+    }
+}
+
+/// URL-encode a string for use in query parameters.
+fn urlencoding(input: &str) -> String {
+    input
+        .chars()
+        .map(|c| match c {
+            'A'..='Z' | 'a'..='z' | '0'..='9' | '-' | '_' | '.' | '~' => c.to_string(),
+            ' ' => '+'.to_string(),
+            _ => {
+                let bytes = c.to_string().into_bytes();
+                bytes
+                    .iter()
+                    .map(|&b| format!("%{:02X}", b))
+                    .collect::<String>()
+            }
+        })
+        .collect()
+}
+
+/// Fetch and extract readable text content from a URL.
+fn fetch_page_content(url: &str, client: &reqwest::blocking::Client) -> Result<String> {
+    let resp = client
+        .get(url)
+        .timeout(std::time::Duration::from_secs(15))
+        .send()
+        .map_err(|e| anyhow::anyhow!("HTTP request failed: {e}"))?;
+
+    let html = resp
+        .text()
+        .map_err(|e| anyhow::anyhow!("failed to read response body: {e}"))?;
+
+    // Parse HTML and extract text
+    let document = scraper::Html::parse_document(&html);
+
+    // Remove script and style elements
+    let selector = scraper::Selector::parse("script, style, nav, footer, header")
+        .map_err(|e| anyhow::anyhow!("invalid selector: {e}"))?;
+    let cleaned = document.clone();
+    for element in cleaned.select(&selector) {
+        let inner = element.inner_html();
+        let _ = inner;
+    }
+
+    // Extract text from main content areas
+    let content_selectors = [
+        "article",
+        "main",
+        ".content",
+        ".documentation",
+        ".doc-content",
+        ".post-content",
+        ".entry-content",
+        "#content",
+        "body",
+    ];
+
+    let mut text = String::new();
+    for sel_str in &content_selectors {
+        if let Ok(sel) = scraper::Selector::parse(sel_str) {
+            if let Some(element) = cleaned.select(&sel).next() {
+                text = element.text().collect::<Vec<_>>().join(" ");
+                if text.len() > 100 {
+                    break;
+                }
+            }
+        }
+    }
+
+    if text.is_empty() {
+        // Fallback: just get all body text
+        if let Ok(body_sel) = scraper::Selector::parse("body") {
+            text = cleaned
+                .select(&body_sel)
+                .next()
+                .map(|e| e.text().collect::<Vec<_>>().join(" "))
+                .unwrap_or_default();
+        }
+    }
+
+    // Clean up whitespace
+    let cleaned_text = text
+        .split_whitespace()
+        .collect::<Vec<_>>()
+        .join(" ");
+
+    if cleaned_text.is_empty() {
+        anyhow::bail!("no readable content found on page");
+    }
+
+    Ok(cleaned_text)
+}
+
+/// Limit a string to at most `max_lines` lines.
+fn limit_lines(s: &str, max_lines: usize) -> String {
+    s.lines()
+        .take(max_lines)
+        .collect::<Vec<_>>()
+        .join("\n")
+}
diff --git a/apps/interfaces/daemon/src/handler.rs b/apps/interfaces/daemon/src/handler.rs
index d65044f..ff28cf2 100644
--- a/apps/interfaces/daemon/src/handler.rs
+++ b/apps/interfaces/daemon/src/handler.rs
@@ -330,6 +330,8 @@ fn handle_submit_input(state: &mut AppStateRest, text: String) {
         api_key,
         model: state.settings.model.clone(),
         api_base: provider_cfg.map(|cfg| cfg.api_base.clone()),
+        edit_count: 0,
+        consecutive_empty_reviews: 0,
     };
 
     zesdex_infrastructure::agent::spawn_agent_turn(params);
diff --git a/apps/interfaces/tui/src/action.rs b/apps/interfaces/tui/src/action.rs
index 5d7d2a2..5c898b4 100644
--- a/apps/interfaces/tui/src/action.rs
+++ b/apps/interfaces/tui/src/action.rs
@@ -12,6 +12,7 @@
 //! *how* state is updated — only *what* action to produce.
 
 use crate::state::Overlay;
+use std::process::Command;
 use tracing::debug;
 
 /// A single well-typed event in the TUI that mutates `AppStateRest`.
@@ -85,6 +86,10 @@ pub enum Action {
     AbortTurn,
     /// Request AI-summary compaction of the conversation history.
     Compact,
+    /// Show the git diff preview overlay.
+    ShowDiff,
+    /// Scroll the diff overlay.
+    DiffScroll(i32),
 }
 
 /// Apply an `Action` to `AppStateRest`.
@@ -309,5 +314,107 @@ pub fn apply_action(state: &mut crate::state::AppStateRest, action: Action) {
         Action::Compact => {
             state.toast_info("Compacting conversation...".to_string());
         }
+        Action::ShowDiff => {
+            // Run git diff to get current changes
+            let diff_output = git_diff_output(state);
+            state.misc.diff_content = diff_output;
+            state.misc.diff_scroll = 0;
+            state.misc.overlay = crate::state::Overlay::Diff;
+            state.mark_dirty();
+        }
+        Action::DiffScroll(amount) => {
+            let max_scroll = state
+                .misc
+                .diff_content
+                .lines()
+                .count()
+                .saturating_sub(1);
+            let new_scroll = (state.misc.diff_scroll as i32 + amount).max(0) as usize;
+            state.misc.diff_scroll = new_scroll.min(max_scroll);
+            state.mark_dirty();
+        }
     }
 }
+
+/// Run `git diff` and return the output for display in the diff overlay.
+///
+/// Flow: runs `git diff HEAD` (staged + unstaged changes), and falls back
+/// to `git diff` if HEAD has no commits yet. Returns a summary of what
+/// changed with colored +/- markers.
+fn git_diff_output(state: &crate::state::AppStateRest) -> String {
+    let root = state
+        .workspace_roots
+        .first()
+        .cloned()
+        .unwrap_or_else(|| std::path::PathBuf::from("."));
+
+    let mut output = String::new();
+
+    // Try diff against HEAD
+    let head_result = Command::new("git")
+        .args(["diff", "HEAD"])
+        .current_dir(&root)
+        .output();
+
+    match head_result {
+        Ok(out) => {
+            let stdout = String::from_utf8_lossy(&out.stdout).trim().to_string();
+            if !stdout.is_empty() {
+                output.push_str("=== Changes (against HEAD) ===\n");
+                output.push_str(&stdout);
+                output.push('\n');
+            }
+        }
+        Err(_) => {
+            // Fallback: no HEAD yet (new repo)
+            let diff_result = Command::new("git")
+                .args(["diff"])
+                .current_dir(&root)
+                .output();
+            if let Ok(out) = diff_result {
+                let stdout = String::from_utf8_lossy(&out.stdout).trim().to_string();
+                if !stdout.is_empty() {
+                    output.push_str("=== Unstaged Changes ===\n");
+                    output.push_str(&stdout);
+                    output.push('\n');
+                }
+            }
+        }
+    }
+
+    // Get staged changes too
+    let staged_result = Command::new("git")
+        .args(["diff", "--cached"])
+        .current_dir(&root)
+        .output();
+
+    if let Ok(out) = staged_result {
+        let stdout = String::from_utf8_lossy(&out.stdout).trim().to_string();
+        if !stdout.is_empty() {
+            output.push_str("=== Staged Changes ===\n");
+            output.push_str(&stdout);
+            output.push('\n');
+        }
+    }
+
+    // Also get status summary
+    let status_result = Command::new("git")
+        .args(["status", "--short"])
+        .current_dir(&root)
+        .output();
+
+    if let Ok(out) = status_result {
+        let stdout = String::from_utf8_lossy(&out.stdout).trim().to_string();
+        if !stdout.is_empty() {
+            output.push_str("=== Summary ===\n");
+            output.push_str(&stdout);
+            output.push('\n');
+        }
+    }
+
+    if output.is_empty() {
+        output = "No changes detected in the working tree.".to_string();
+    }
+
+    output
+}
diff --git a/apps/interfaces/tui/src/controller/command.rs b/apps/interfaces/tui/src/controller/command.rs
index 511d7b4..07bfa06 100644
--- a/apps/interfaces/tui/src/controller/command.rs
+++ b/apps/interfaces/tui/src/controller/command.rs
@@ -36,6 +36,8 @@ pub enum Command {
     PlanOpen,
     /// `/usage` — open the usage-stats overlay.
     UsageOpen,
+    /// `/diff` — show git diff preview.
+    Diff,
     /// Catch-all: unrecognised or non-slash input.
     Unknown(String),
 }
@@ -89,6 +91,7 @@ pub fn parse_command(text: &str) -> Command {
         "/todo" => Command::TodoOpen,
         "/plan" => Command::PlanOpen,
         "/usage" => Command::UsageOpen,
+        "/diff" => Command::Diff,
         _ => Command::Unknown(cmd.to_string()),
     };
 
@@ -142,6 +145,9 @@ pub fn apply_command(cmd: Command) -> Vec<crate::action::Action> {
         Command::UsageOpen => {
             vec![crate::action::Action::OpenOverlay(crate::state::Overlay::Usage)]
         }
+        Command::Diff => {
+            vec![crate::action::Action::ShowDiff]
+        }
         Command::Unknown(text) => {
             if text.starts_with('/') {
                 vec![crate::action::Action::SystemNote {
diff --git a/apps/interfaces/tui/src/controller/input.rs b/apps/interfaces/tui/src/controller/input.rs
index d1b328d..88c83cc 100644
--- a/apps/interfaces/tui/src/controller/input.rs
+++ b/apps/interfaces/tui/src/controller/input.rs
@@ -82,6 +82,31 @@ pub fn handle_key(key: KeyEvent, state: &mut AppStateRest) -> Vec<Action> {
         }
     }
 
+    // ── Diff overlay ──────────────────────────────────────────────────────
+    if state.misc.overlay == Overlay::Diff {
+        match key.code {
+            KeyCode::Char('c') if key.modifiers.contains(KeyModifiers::CONTROL) => {
+                return vec![Action::QuitConfirm];
+            }
+            KeyCode::Up | KeyCode::Char('k') => {
+                return vec![Action::DiffScroll(-1)];
+            }
+            KeyCode::Down | KeyCode::Char('j') => {
+                return vec![Action::DiffScroll(1)];
+            }
+            KeyCode::PageUp => {
+                return vec![Action::DiffScroll(-20)];
+            }
+            KeyCode::PageDown => {
+                return vec![Action::DiffScroll(20)];
+            }
+            KeyCode::Esc => {
+                return vec![Action::CloseOverlay];
+            }
+            _ => {}
+        }
+    }
+
     // ── Learning overlay ──────────────────────────────────────────────────
     if state.misc.overlay == Overlay::Learning {
         match key.code {
@@ -164,6 +189,9 @@ pub fn handle_key(key: KeyEvent, state: &mut AppStateRest) -> Vec<Action> {
             }
             Vec::new()
         }
+        KeyCode::Char('d') if key.modifiers.contains(KeyModifiers::ALT) => {
+            vec![crate::action::Action::ShowDiff]
+        }
         KeyCode::Enter => {
             if state.input.autocomplete_visible {
                 state.input.select_autocomplete();
diff --git a/apps/interfaces/tui/src/state.rs b/apps/interfaces/tui/src/state.rs
index aae2677..50028ad 100644
--- a/apps/interfaces/tui/src/state.rs
+++ b/apps/interfaces/tui/src/state.rs
@@ -439,6 +439,8 @@ pub enum Overlay {
     ModelSelector,
     /// "Clear conversation?" confirmation.
     ClearConfirm,
+    /// Git diff preview overlay.
+    Diff,
 }
 
 impl Overlay {
@@ -462,6 +464,7 @@ impl Overlay {
             Overlay::Loading => "loading",
             Overlay::ModelSelector => "model_selector",
             Overlay::ClearConfirm => "clear_confirm",
+            Overlay::Diff => "diff",
         }
     }
 
@@ -510,6 +513,10 @@ pub struct MiscState {
     pub lesson_running: bool,
     /// Text waiting to be written to the system clipboard.
     pub pending_clipboard_copy: Option<String>,
+    /// Cached git diff content for the preview overlay.
+    pub diff_content: String,
+    /// Scroll offset for the diff overlay.
+    pub diff_scroll: usize,
 }
 
 impl MiscState {
@@ -529,6 +536,8 @@ impl MiscState {
             plan_content: String::new(),
             lesson_running: false,
             pending_clipboard_copy: None,
+            diff_content: String::new(),
+            diff_scroll: 0,
         }
     }
 
diff --git a/apps/interfaces/tui/src/turn.rs b/apps/interfaces/tui/src/turn.rs
index c17593b..d75395b 100644
--- a/apps/interfaces/tui/src/turn.rs
+++ b/apps/interfaces/tui/src/turn.rs
@@ -61,6 +61,17 @@ pub fn spawn_agent_turn(state: &mut AppStateRest, text: String) {
 
     info!("delegating agent turn to infrastructure engine (model: {})", model);
 
+    let edit_count = state
+        .session_runtime
+        .as_ref()
+        .map(|rt| rt.edit_count)
+        .unwrap_or(0);
+    let consecutive_empty_reviews = state
+        .session_runtime
+        .as_ref()
+        .map(|rt| rt.consecutive_empty_reviews)
+        .unwrap_or(0);
+
     let params = AgentTurnParams {
         messages,
         session_dir,
@@ -71,6 +82,8 @@ pub fn spawn_agent_turn(state: &mut AppStateRest, text: String) {
         api_key,
         model,
         api_base,
+        edit_count,
+        consecutive_empty_reviews,
     };
 
     backend_spawn_agent_turn(params);
diff --git a/apps/interfaces/tui/src/view/overlays/diff.rs b/apps/interfaces/tui/src/view/overlays/diff.rs
new file mode 100644
index 0000000..b43c1a2
--- /dev/null
+++ b/apps/interfaces/tui/src/view/overlays/diff.rs
@@ -0,0 +1,138 @@
+//! Diff preview overlay — renders git diff output with color-coded
+//! additions (green) and deletions (red) in a scrollable panel.
+//!
+//! Flow: read `state.misc.diff_content` → parse diff lines → apply syntax
+//! coloring (green `+`, red `-`, blue header, dim context) → render in a
+//! scrollable paragraph widget.
+
+use crate::state::AppStateRest;
+use ratatui::layout::Rect;
+use ratatui::style::{Modifier, Style};
+use ratatui::text::{Line, Span};
+use ratatui::widgets::{Block, Paragraph, Wrap};
+use ratatui::Frame;
+use super::super::theme::Theme;
+use super::overlay_block;
+
+/// Render the diff preview overlay with color-coded +/- lines and scrolling.
+///
+/// The overlay fills most of the screen and shows:
+/// - Green lines for additions (`+`)
+/// - Red lines for deletions (`-`)
+/// - Blue/dim headers (diff --git, @@ hunk headers)
+/// - Dim context lines
+/// - Scroll hint at the bottom
+pub fn render(frame: &mut Frame, area: Rect, block: Block<'static>, state: &AppStateRest) {
+    let block = overlay_block(block, " Git Diff Preview ", Theme::PRIMARY);
+    let inner = block.inner(area);
+
+    // Render the block background
+    frame.render_widget(block.clone(), area);
+
+    let diff_content = &state.misc.diff_content;
+    let diff_scroll = state.misc.diff_scroll;
+
+    // Parse lines with ANSI-style coloring
+    let lines: Vec<Line> = diff_content
+        .lines()
+        .skip(diff_scroll)
+        .map(|line| {
+            let trimmed = line.trim_end();
+            if trimmed.starts_with('+') && !trimmed.starts_with("+++") {
+                // Addition — green
+                Line::from(Span::styled(
+                    format!("{trimmed}\n"),
+                    Style::default().fg(Theme::SUCCESS),
+                ))
+            } else if trimmed.starts_with('-') && !trimmed.starts_with("---") {
+                // Deletion — red
+                Line::from(Span::styled(
+                    format!("{trimmed}\n"),
+                    Style::default().fg(Theme::ERROR),
+                ))
+            } else if trimmed.starts_with("@@") {
+                // Hunk header — cyan
+                Line::from(Span::styled(
+                    format!("{trimmed}\n"),
+                    Style::default()
+                        .fg(Theme::INFO)
+                        .add_modifier(Modifier::DIM),
+                ))
+            } else if trimmed.starts_with("diff --git")
+                || trimmed.starts_with("index ")
+                || trimmed.starts_with("--- ")
+                || trimmed.starts_with("+++ ")
+                || trimmed.starts_with("===")
+            {
+                // File header — blue/bold
+                Line::from(Span::styled(
+                    format!("{trimmed}\n"),
+                    Style::default()
+                        .fg(Theme::PRIMARY)
+                        .add_modifier(Modifier::BOLD),
+                ))
+            } else {
+                // Context — dim
+                Line::from(Span::styled(
+                    format!("{trimmed}\n"),
+                    Style::default().fg(Theme::TEXT_DIM),
+                ))
+            }
+        })
+        .collect();
+
+    let total_lines = diff_content.lines().count();
+    let visible_lines = inner.height as usize;
+    let has_more = diff_scroll + visible_lines < total_lines;
+
+    // Create the content paragraph
+    let paragraph = Paragraph::new(lines).wrap(Wrap { trim: false });
+    frame.render_widget(paragraph, inner);
+
+    // Render scroll indicator at the bottom
+    if has_more {
+        let scroll_pct = if total_lines > 0 {
+            ((diff_scroll as f64 / total_lines as f64) * 100.0) as u32
+        } else {
+            0
+        };
+        let scroll_text = format!(
+            "  Lines {}-{} of {} ({}%)  ",
+            diff_scroll + 1,
+            (diff_scroll + visible_lines).min(total_lines),
+            total_lines,
+            scroll_pct
+        );
+
+        let status_line = Line::from(Span::styled(
+            scroll_text,
+            Style::default()
+                .fg(Theme::TEXT_DIM)
+                .add_modifier(Modifier::ITALIC),
+        ));
+
+        let status_area = Rect {
+            x: area.x,
+            y: area.y + area.height - 1,
+            width: area.width,
+            height: 1,
+        };
+        frame.render_widget(Paragraph::new(status_line), status_area);
+    }
+
+    // Render keybinding hint at the top
+    let hint = Line::from(Span::styled(
+        "  ↑/↓: scroll  Esc: close  /diff: refresh",
+        Style::default()
+            .fg(Theme::TEXT_DIM)
+            .add_modifier(Modifier::ITALIC),
+    ));
+
+    let hint_area = Rect {
+        x: area.x,
+        y: area.y,
+        width: area.width,
+        height: 1,
+    };
+    frame.render_widget(Paragraph::new(hint), hint_area);
+}
diff --git a/apps/interfaces/tui/src/view/overlays/mod.rs b/apps/interfaces/tui/src/view/overlays/mod.rs
index c10a824..d8e2496 100644
--- a/apps/interfaces/tui/src/view/overlays/mod.rs
+++ b/apps/interfaces/tui/src/view/overlays/mod.rs
@@ -4,6 +4,7 @@
 
 pub mod bash;
 pub mod clear_confirm;
+pub mod diff;
 pub mod editor;
 pub mod effort;
 pub mod help;
@@ -123,5 +124,8 @@ pub fn render_overlay(
         Overlay::ClearConfirm => {
             clear_confirm::render(frame, overlay_area, block, state);
         }
+        Overlay::Diff => {
+            diff::render(frame, overlay_area, block, state);
+        }
     }
 }
diff --git a/apps/interfaces/ws/src/lib.rs b/apps/interfaces/ws/src/lib.rs
index a477e93..c2c2a56 100644
--- a/apps/interfaces/ws/src/lib.rs
+++ b/apps/interfaces/ws/src/lib.rs
@@ -92,6 +92,8 @@ async fn handle_socket(mut socket: WebSocket, state: Arc<WsState>) {
                                 api_key,
                                 model,
                                 api_base: None,
+                                edit_count: 0,
+                                consecutive_empty_reviews: 0,
                             };
 
                             zesdex_infrastructure::agent::spawn_agent_turn(params);