diff --git a/.claude/commands/github-pr.md b/.claude/commands/github-pr.md
index ccc26ce3..e15e9c6d 100644
--- a/.claude/commands/github-pr.md
+++ b/.claude/commands/github-pr.md
@@ -1,6 +1,6 @@
 ---
 description: Commit, create PR, monitor CI, fix failures, merge
-allowed-tools: Bash(git status:*), Bash(git diff:*), Bash(git add:*), Bash(git commit:*), Bash(git push:*), Bash(git branch:*), Bash(git checkout:*), Bash(git log:*), Bash(gh pr:*), Bash(gh run:*), Bash(cargo check:*), Bash(cargo test:*), Bash(cargo clippy:*), Read, Edit, Write, Grep, Glob
+allowed-tools: Bash(git status:*), Bash(git diff:*), Bash(git add:*), Bash(git commit:*), Bash(git push:*), Bash(git branch:*), Bash(git checkout:*), Bash(git log:*), Bash(git fetch:*), Bash(git merge:*), Bash(git rebase:*), Bash(git cherry-pick:*), Bash(git reset:*), Bash(gh pr:*), Bash(gh run:*), Bash(gh api:*), Bash(cargo check:*), Bash(cargo test:*), Bash(cargo clippy:*), Bash(bazel:*), Bash(rustfmt:*), Bash(sleep:*), Read, Edit, Write, Grep, Glob
 ---
 
 # GitHub PR 全流程
@@ -54,6 +54,43 @@ allowed-tools: Bash(git status:*), Bash(git diff:*), Bash(git add:*), Bash(git c
 
 4. 记录 PR 编号，后续阶段会用到。
 
+## Phase 2.5: Conflict Detection
+
+CI 不触发的常见原因是 PR 有 conflict。创建 PR 后立即检查：
+
+```
+gh pr view <PR#> --json mergeable,mergeStateStatus
+```
+
+- **mergeStateStatus: "CLEAN"** → 无冲突，继续 Phase 3
+- **mergeStateStatus: "DIRTY" / mergeable: "CONFLICTING"** → 需要解决冲突：
+
+**解决流程：**
+1. `git fetch origin main` 获取最新 main
+2. `git log --oneline origin/main..HEAD` 查看当前分支比 main 多了哪些 commit
+3. 判断多余 commit 是否已经在 main 上（通过其他 PR squash 合并过）：
+   - `git log --oneline <merge-base>..origin/main` 对比 main 上的新 commit
+   - 如果当前分支的旧 commit 已在 main 上 squash 合并，只需保留本次 PR 的 commit
+4. 解决方案（按优先级）：
+   - **Cherry-pick 法**（推荐，当旧 commit 已在 main 上）：
+     ```
+     git reset --hard origin/main
+     git cherry-pick <本次PR的commit>
+     git push --force-with-lease
+     ```
+   - **Rebase 法**（通用）：
+     ```
+     git rebase origin/main
+     # 逐个解决冲突
+     git push --force-with-lease
+     ```
+5. 推送后重新检查 `gh pr view <PR#> --json mergeable,mergeStateStatus`
+
+**注意**：worktree 中 `gh pr merge` 可能因为无法 checkout main 而失败，此时用 API 合并：
+```
+gh api repos/{owner}/{repo}/pulls/{PR#}/merge -f merge_method=squash
+```
+
 ## Phase 3: Monitor CI
 
 1. 等待 10 秒让 CI 启动，然后开始轮询：
@@ -115,6 +152,7 @@ allowed-tools: Bash(git status:*), Bash(git diff:*), Bash(git add:*), Bash(git c
 
 - 每个阶段完成后简要汇报进度
 - 遇到不确定的决策（如 CI 失败原因不明）时询问用户
-- 绝不 force push
 - 绝不直接 push 到 main/master
+- Force push 仅限 `--force-with-lease` 且仅在 rebase/cherry-pick 解决冲突后使用
 - 如果 PR 需要 review approval 才能 merge，报告并等待用户指示
+- 在 worktree 中合并 PR 时，优先使用 `gh api` 而非 `gh pr merge`（避免 checkout main 失败）
diff --git a/crates/loopal-agent-client/src/client.rs b/crates/loopal-agent-client/src/client.rs
index 896931ed..8e366fba 100644
--- a/crates/loopal-agent-client/src/client.rs
+++ b/crates/loopal-agent-client/src/client.rs
@@ -55,6 +55,7 @@ impl AgentClient {
         no_sandbox: bool,
         resume: Option<&str>,
         lifecycle: Option<&str>,
+        agent_type: Option<&str>,
     ) -> anyhow::Result<String> {
         let params = serde_json::json!({
             "cwd": cwd.to_string_lossy(),
@@ -65,6 +66,7 @@ impl AgentClient {
             "no_sandbox": no_sandbox,
             "resume": resume,
             "lifecycle": lifecycle,
+            "agent_type": agent_type,
         });
         let result = self
             .connection
diff --git a/crates/loopal-agent-hub/src/dispatch/dispatch_handlers.rs b/crates/loopal-agent-hub/src/dispatch/dispatch_handlers.rs
index 8b13bedb..4e97c542 100644
--- a/crates/loopal-agent-hub/src/dispatch/dispatch_handlers.rs
+++ b/crates/loopal-agent-hub/src/dispatch/dispatch_handlers.rs
@@ -160,6 +160,7 @@ pub async fn handle_spawn_agent(
     let model = params["model"].as_str().map(String::from);
     let prompt = params["prompt"].as_str().map(String::from);
     let permission_mode = params["permission_mode"].as_str().map(String::from);
+    let agent_type = params["agent_type"].as_str().map(String::from);
 
     // Parent: use explicit "parent" field from params if present (cross-hub),
     // otherwise use from_agent (local spawn).
@@ -180,6 +181,7 @@ pub async fn handle_spawn_agent(
             prompt,
             parent,
             permission_mode,
+            agent_type,
         )
         .await
     });
diff --git a/crates/loopal-agent-hub/src/spawn_manager.rs b/crates/loopal-agent-hub/src/spawn_manager.rs
index 06d1df5d..a573e6ca 100644
--- a/crates/loopal-agent-hub/src/spawn_manager.rs
+++ b/crates/loopal-agent-hub/src/spawn_manager.rs
@@ -12,6 +12,7 @@ use loopal_protocol::{AgentEvent, AgentEventPayload, Envelope};
 use crate::hub::Hub;
 
 /// Spawn a real agent process, initialize, start, and register in Hub.
+#[allow(clippy::too_many_arguments)]
 pub async fn spawn_and_register(
     hub: Arc<Mutex<Hub>>,
     name: String,
@@ -20,6 +21,7 @@ pub async fn spawn_and_register(
     prompt: Option<String>,
     parent: Option<String>,
     permission_mode: Option<String>,
+    agent_type: Option<String>,
 ) -> Result<String, String> {
     info!(agent = %name, parent = ?parent, "spawn: forking process");
     let agent_proc = loopal_agent_client::AgentProcess::spawn(None)
@@ -44,6 +46,7 @@ pub async fn spawn_and_register(
             false,
             None,
             Some("ephemeral"), // sub-agents always exit on idle
+            agent_type.as_deref(),
         )
         .await
     {
diff --git a/crates/loopal-agent-hub/tests/suite/e2e_bootstrap_test.rs b/crates/loopal-agent-hub/tests/suite/e2e_bootstrap_test.rs
index 203795ca..4788d610 100644
--- a/crates/loopal-agent-hub/tests/suite/e2e_bootstrap_test.rs
+++ b/crates/loopal-agent-hub/tests/suite/e2e_bootstrap_test.rs
@@ -59,6 +59,7 @@ async fn full_bootstrap_hub_to_agent_roundtrip() {
             true, // no sandbox
             None,
             None, // lifecycle: default
+            None, // agent_type
         )
         .await
         .expect("start_agent should work");
diff --git a/crates/loopal-agent-server/src/agent_setup.rs b/crates/loopal-agent-server/src/agent_setup.rs
index 2eb7e80d..92914960 100644
--- a/crates/loopal-agent-server/src/agent_setup.rs
+++ b/crates/loopal-agent-server/src/agent_setup.rs
@@ -123,6 +123,7 @@ pub fn build_with_frontend(
         &cwd.to_string_lossy(),
         &skills_summary,
         &config.memory,
+        start.agent_type.as_deref(),
     );
 
     // Append MCP server instructions (from initialize handshake).
diff --git a/crates/loopal-agent-server/src/memory_adapter.rs b/crates/loopal-agent-server/src/memory_adapter.rs
index b630fa6a..a71e4d4f 100644
--- a/crates/loopal-agent-server/src/memory_adapter.rs
+++ b/crates/loopal-agent-server/src/memory_adapter.rs
@@ -49,6 +49,7 @@ impl MemoryProcessor for ServerMemoryProcessor {
             cwd_override: None,
             permission_mode: None,
             target_hub: None,
+            agent_type: None,
         };
         spawn_agent(&self.shared, params).await?;
         info!("memory-maintainer agent spawned via Hub");
diff --git a/crates/loopal-agent-server/src/params.rs b/crates/loopal-agent-server/src/params.rs
index a40744f6..b69ef24d 100644
--- a/crates/loopal-agent-server/src/params.rs
+++ b/crates/loopal-agent-server/src/params.rs
@@ -16,6 +16,8 @@ pub struct StartParams {
     pub resume: Option<String>,
     /// Explicit lifecycle mode. Ephemeral exits on idle, Persistent waits.
     pub lifecycle: loopal_runtime::LifecycleMode,
+    /// Agent type for fragment selection (e.g. "explore", "plan").
+    pub agent_type: Option<String>,
 }
 
 /// Build a Kernel from config (production path: MCP, tools).
diff --git a/crates/loopal-agent-server/src/session_start.rs b/crates/loopal-agent-server/src/session_start.rs
index 3dac01e9..8ff6e915 100644
--- a/crates/loopal-agent-server/src/session_start.rs
+++ b/crates/loopal-agent-server/src/session_start.rs
@@ -62,6 +62,7 @@ pub(crate) async fn start_session(
         no_sandbox: params["no_sandbox"].as_bool().unwrap_or(false),
         resume: params["resume"].as_str().map(String::from),
         lifecycle,
+        agent_type: params["agent_type"].as_str().map(String::from),
     };
 
     let mut config = load_config(&cwd)?;
diff --git a/crates/loopal-agent-server/tests/suite/hub_harness.rs b/crates/loopal-agent-server/tests/suite/hub_harness.rs
index a7ea2761..70aae7c1 100644
--- a/crates/loopal-agent-server/tests/suite/hub_harness.rs
+++ b/crates/loopal-agent-server/tests/suite/hub_harness.rs
@@ -130,6 +130,7 @@ pub async fn build_hub_harness_with(
         no_sandbox: true,
         resume: None,
         lifecycle: loopal_runtime::LifecycleMode::Persistent,
+        agent_type: None,
     };
     let (hub_conn, _hub_peer) = loopal_ipc::duplex_pair();
     let hub_connection = std::sync::Arc::new(loopal_ipc::Connection::new(hub_conn));
diff --git a/crates/loopal-agent/src/spawn.rs b/crates/loopal-agent/src/spawn.rs
index a7a191fb..4e840f64 100644
--- a/crates/loopal-agent/src/spawn.rs
+++ b/crates/loopal-agent/src/spawn.rs
@@ -19,6 +19,8 @@ pub struct SpawnParams {
     pub permission_mode: Option<String>,
     /// Target hub for cross-hub spawn (e.g. "hub-b"). None = local hub.
     pub target_hub: Option<String>,
+    /// Agent type for fragment selection (e.g. "explore", "plan").
+    pub agent_type: Option<String>,
 }
 
 /// Result returned from Hub after spawning.
@@ -45,6 +47,7 @@ pub async fn spawn_agent(
         "model": params.model,
         "prompt": params.prompt,
         "permission_mode": params.permission_mode,
+        "agent_type": params.agent_type,
     });
     if let Some(ref hub) = params.target_hub {
         request["target_hub"] = json!(hub);
diff --git a/crates/loopal-agent/src/tools/collaboration/agent.rs b/crates/loopal-agent/src/tools/collaboration/agent.rs
index d96da663..e1831e2b 100644
--- a/crates/loopal-agent/src/tools/collaboration/agent.rs
+++ b/crates/loopal-agent/src/tools/collaboration/agent.rs
@@ -130,6 +130,7 @@ async fn action_spawn(
             cwd_override,
             permission_mode: Some(perm_mode.to_string()),
             target_hub,
+            agent_type: subagent_type.map(String::from),
         },
     )
     .await;
diff --git a/crates/loopal-agent/tests/suite/bridge_chain_test.rs b/crates/loopal-agent/tests/suite/bridge_chain_test.rs
index 36560f1d..9e7df193 100644
--- a/crates/loopal-agent/tests/suite/bridge_chain_test.rs
+++ b/crates/loopal-agent/tests/suite/bridge_chain_test.rs
@@ -49,6 +49,7 @@ async fn full_chain_sub_agent_result_delivered_to_parent() {
             false,
             None,
             None,
+            None,
         )
         .await
         .expect("start_agent");
diff --git a/crates/loopal-agent/tests/suite/bridge_child_test.rs b/crates/loopal-agent/tests/suite/bridge_child_test.rs
index 9bd602b2..52350afd 100644
--- a/crates/loopal-agent/tests/suite/bridge_child_test.rs
+++ b/crates/loopal-agent/tests/suite/bridge_child_test.rs
@@ -66,6 +66,7 @@ pub(crate) async fn start_bridge_client(
             false,
             None,
             None,
+            None,
         )
         .await
         .expect("start_agent");
@@ -156,6 +157,7 @@ async fn bridge_cancel_sends_shutdown() {
             false,
             None,
             None,
+            None,
         )
         .await
         .unwrap();
diff --git a/crates/loopal-context/src/system_prompt.rs b/crates/loopal-context/src/system_prompt.rs
index 0ddf0c25..faa463fc 100644
--- a/crates/loopal-context/src/system_prompt.rs
+++ b/crates/loopal-context/src/system_prompt.rs
@@ -12,6 +12,7 @@ pub fn build_system_prompt(
     cwd: &str,
     skills_summary: &str,
     memory: &str,
+    agent_type: Option<&str>,
 ) -> String {
     let mut registry = FragmentRegistry::new(system_fragments());
 
@@ -23,26 +24,15 @@ pub fn build_system_prompt(
 
     let builder = PromptBuilder::new(registry);
 
+    // Tool names/descriptions feed Minijinja conditionals in fragments
+    // (e.g. `{% if "Bash" in tool_names %}`). Full JSON schemas are sent
+    // separately via ChatParams.tools — no need to duplicate them here.
     let tool_names: Vec<String> = tools.iter().map(|t| t.name.clone()).collect();
     let tool_descriptions: HashMap<String, String> = tools
         .iter()
         .map(|t| (t.name.clone(), t.description.clone()))
         .collect();
 
-    // Build tool schema section (kept as-is for LLM function calling)
-    let tools_section = if tools.is_empty() {
-        String::new()
-    } else {
-        let mut s = String::from("# Available Tools\n");
-        for tool in tools {
-            s.push_str(&format!(
-                "\n## {}\n{}\nParameters: {}\n",
-                tool.name, tool.description, tool.input_schema
-            ));
-        }
-        s
-    };
-
     let ctx = PromptContext {
         cwd: cwd.to_string(),
         platform: std::env::consts::OS.to_string(),
@@ -61,18 +51,10 @@ pub fn build_system_prompt(
         skills_summary: skills_summary.to_string(),
         features: Vec::new(),
         agent_name: None,
-        agent_type: None,
+        agent_type: agent_type.map(String::from),
     };
 
-    let mut prompt = builder.build(&ctx);
-
-    // Append tool schemas after fragments (LLM needs the JSON schemas)
-    if !tools_section.is_empty() {
-        prompt.push_str("\n\n");
-        prompt.push_str(&tools_section);
-    }
-
-    prompt
+    builder.build(&ctx)
 }
 
 fn today() -> String {
diff --git a/crates/loopal-context/tests/suite.rs b/crates/loopal-context/tests/suite.rs
index ea44a364..9f73f891 100644
--- a/crates/loopal-context/tests/suite.rs
+++ b/crates/loopal-context/tests/suite.rs
@@ -13,6 +13,8 @@ mod pipeline_test;
 mod smart_compact_test;
 #[path = "suite/store_test.rs"]
 mod store_test;
+#[path = "suite/system_prompt_agent_test.rs"]
+mod system_prompt_agent_test;
 #[path = "suite/system_prompt_test.rs"]
 mod system_prompt_test;
 #[path = "suite/token_counter_test.rs"]
diff --git a/crates/loopal-context/tests/suite/system_prompt_agent_test.rs b/crates/loopal-context/tests/suite/system_prompt_agent_test.rs
new file mode 100644
index 00000000..66fcc623
--- /dev/null
+++ b/crates/loopal-context/tests/suite/system_prompt_agent_test.rs
@@ -0,0 +1,108 @@
+use loopal_context::build_system_prompt;
+use loopal_tool_api::ToolDefinition;
+
+#[test]
+fn explore_subagent_full_prompt() {
+    let tools = vec![
+        ToolDefinition {
+            name: "Read".into(),
+            description: "Read a file".into(),
+            input_schema: serde_json::json!({"type": "object"}),
+        },
+        ToolDefinition {
+            name: "Grep".into(),
+            description: "Search file contents".into(),
+            input_schema: serde_json::json!({"type": "object"}),
+        },
+        ToolDefinition {
+            name: "Bash".into(),
+            description: "Execute commands".into(),
+            input_schema: serde_json::json!({"type": "object"}),
+        },
+    ];
+    let result = build_system_prompt(
+        "Project rules",
+        &tools,
+        "act",
+        "/project",
+        "",
+        "",
+        Some("explore"),
+    );
+
+    // Explore-specific content present
+    assert!(
+        result.contains("READ-ONLY MODE"),
+        "explore fragment should be included"
+    );
+    // Default sub-agent fragment excluded (explore is specific)
+    assert!(
+        !result.contains("sub-agent named"),
+        "default-subagent should be excluded when explore matches"
+    );
+    // Core fragments still present
+    assert!(
+        result.contains("Output Efficiency"),
+        "core fragments should be included for sub-agents"
+    );
+    // Tool usage policy still present
+    assert!(
+        result.contains("Tool Usage Policy"),
+        "tool usage policy should be included for sub-agents"
+    );
+    // User instructions still present
+    assert!(result.contains("Project rules"), "instructions missing");
+    // Tool schemas NOT in prompt
+    assert!(
+        !result.contains("# Available Tools"),
+        "tool schemas should not be in system prompt"
+    );
+}
+
+#[test]
+fn root_agent_excludes_agent_fragments() {
+    let result = build_system_prompt("Base", &[], "act", "/workspace", "", "", None);
+    // No agent fragments in root prompt
+    assert!(
+        !result.contains("sub-agent named"),
+        "default-subagent should not appear for root agent"
+    );
+    assert!(
+        !result.contains("READ-ONLY MODE"),
+        "explore fragment should not appear for root agent"
+    );
+    // Core fragments present
+    assert!(result.contains("Output Efficiency"));
+}
+
+#[test]
+fn plan_subagent_gets_plan_fragment() {
+    let result = build_system_prompt("", &[], "act", "/work", "", "", Some("plan"));
+    assert!(
+        result.contains("software architect"),
+        "plan fragment should be included"
+    );
+    assert!(
+        !result.contains("sub-agent named"),
+        "default-subagent excluded when plan matches"
+    );
+    assert!(
+        !result.contains("READ-ONLY MODE"),
+        "explore fragment should not be included for plan agent"
+    );
+}
+
+#[test]
+fn general_subagent_gets_default_fragment() {
+    let result = build_system_prompt("", &[], "act", "/work", "", "", Some("general"));
+    // Default sub-agent fragment (fallback for unknown types)
+    assert!(
+        result.contains("sub-agent named"),
+        "default-subagent should be included for unknown agent types"
+    );
+    // Specialized fragments excluded
+    assert!(
+        !result.contains("READ-ONLY MODE"),
+        "explore fragment should not appear for general"
+    );
+}
diff --git a/crates/loopal-context/tests/suite/system_prompt_test.rs b/crates/loopal-context/tests/suite/system_prompt_test.rs
index 433455b8..620f9f0d 100644
--- a/crates/loopal-context/tests/suite/system_prompt_test.rs
+++ b/crates/loopal-context/tests/suite/system_prompt_test.rs
@@ -3,26 +3,28 @@ use loopal_tool_api::ToolDefinition;
 
 #[test]
 fn includes_instructions() {
-    let result = build_system_prompt("You are helpful.", &[], "act", "/tmp", "", "");
+    let result = build_system_prompt("You are helpful.", &[], "act", "/tmp", "", "", None);
     assert!(result.contains("You are helpful."));
 }
 
 #[test]
-fn includes_tool_schemas() {
+fn tool_schemas_not_in_system_prompt() {
     let tools = vec![ToolDefinition {
         name: "read".into(),
         description: "Read a file".into(),
         input_schema: serde_json::json!({"type": "object"}),
     }];
-    let result = build_system_prompt("Base", &tools, "act", "/workspace", "", "");
-    assert!(result.contains("# Available Tools"));
-    assert!(result.contains("## read"));
-    assert!(result.contains("Read a file"));
+    let result = build_system_prompt("Base", &tools, "act", "/workspace", "", "", None);
+    // Tool schemas should NOT appear in system prompt — they go via ChatParams.tools
+    assert!(!result.contains("# Available Tools"));
+    assert!(!result.contains("## read"));
+    // But tool names should still feed fragment conditionals
+    assert!(result.contains("Base")); // instructions still present
 }
 
 #[test]
 fn includes_fragments() {
-    let result = build_system_prompt("Base", &[], "act", "/workspace", "", "");
+    let result = build_system_prompt("Base", &[], "act", "/workspace", "", "", None);
     // Core fragments should be present
     assert!(
         result.contains("Output Efficiency"),
@@ -35,15 +37,28 @@ fn includes_fragments() {
 }
 
 #[test]
-fn includes_environment() {
-    let result = build_system_prompt("Base", &[], "act", "/Users/dev/project", "", "");
-    assert!(result.contains("/Users/dev/project"), "cwd not rendered");
+fn cwd_available_in_subagent_prompt() {
+    // Root agent: cwd is injected per-turn via env_context, not in static prompt.
+    // Sub-agent: cwd appears in default-subagent fragment template.
+    let result = build_system_prompt(
+        "Base",
+        &[],
+        "act",
+        "/Users/dev/project",
+        "",
+        "",
+        Some("general"), // any agent_type makes is_subagent() true
+    );
+    assert!(
+        result.contains("/Users/dev/project"),
+        "cwd not rendered in sub-agent prompt"
+    );
 }
 
 #[test]
 fn includes_skills() {
     let skills = "# Available Skills\n- /commit: Generate a git commit message";
-    let result = build_system_prompt("Base", &[], "act", "/workspace", skills, "");
+    let result = build_system_prompt("Base", &[], "act", "/workspace", skills, "", None);
     assert!(result.contains("Available Skills"));
     assert!(result.contains("/commit"));
 }
@@ -57,6 +72,7 @@ fn includes_memory() {
         "/workspace",
         "",
         "## Key Patterns\n- Use DI",
+        None,
     );
     assert!(result.contains("# Project Memory"));
     assert!(result.contains("Key Patterns"));
@@ -64,7 +80,7 @@ fn includes_memory() {
 
 #[test]
 fn empty_memory_no_section() {
-    let result = build_system_prompt("Base", &[], "act", "/workspace", "", "");
+    let result = build_system_prompt("Base", &[], "act", "/workspace", "", "", None);
     assert!(!result.contains("Project Memory"));
 }
 
@@ -76,14 +92,14 @@ fn tool_conditional_fragments() {
         description: "Execute commands".into(),
         input_schema: serde_json::json!({"type": "object"}),
     }];
-    let result = build_system_prompt("Base", &tools, "act", "/workspace", "", "");
+    let result = build_system_prompt("Base", &tools, "act", "/workspace", "", "", None);
     assert!(
         result.contains("Bash Tool Guidelines"),
         "bash guidelines missing when Bash tool present"
     );
 
     // Without Bash tool → no bash guidelines
-    let result_no_bash = build_system_prompt("Base", &[], "act", "/workspace", "", "");
+    let result_no_bash = build_system_prompt("Base", &[], "act", "/workspace", "", "", None);
     assert!(
         !result_no_bash.contains("Bash Tool Guidelines"),
         "bash guidelines should not appear without Bash"
@@ -136,23 +152,22 @@ fn report_token_usage() {
     let mem = "## Architecture\n- 17 Rust crates\n- 200-line limit";
     let skills = "# Available Skills\n- /commit: Git commit\n- /review-pr: Review PR";
 
-    let bare = build_system_prompt("", &[], "act", "/project", "", "");
-    let with_tools = build_system_prompt("", &tools, "act", "/project", "", "");
-    let full_act = build_system_prompt(instr, &tools, "act", "/project", skills, mem);
-    let full_plan = build_system_prompt(instr, &tools, "plan", "/project", skills, mem);
+    let bare = build_system_prompt("", &[], "act", "/project", "", "", None);
+    let with_tools = build_system_prompt("", &tools, "act", "/project", "", "", None);
+    let full_act = build_system_prompt(instr, &tools, "act", "/project", skills, mem, None);
+    let full_plan = build_system_prompt(instr, &tools, "plan", "/project", skills, mem, None);
 
     let t_bare = estimate_tokens(&bare);
     let t_tools = estimate_tokens(&with_tools);
     let t_act = estimate_tokens(&full_act);
     let t_plan = estimate_tokens(&full_plan);
 
-    eprintln!("\n=== System Prompt Token Analysis ===");
     eprintln!(
         "Fragments only:              {} tokens ({} chars)",
         t_bare,
         bare.len()
     );
-    eprintln!("Fragments + 21 tool schemas: {t_tools} tokens");
+    eprintln!("Fragments + 21 tools (cond): {t_tools} tokens");
     eprintln!(
         "Full (act, 21 tools):        {} tokens ({} chars)",
         t_act,
@@ -166,7 +181,7 @@ fn report_token_usage() {
     eprintln!("Plan overhead:               +{} tokens", t_plan - t_act);
     eprintln!("--- Breakdown ---");
     eprintln!("  Behavior fragments: {t_bare} tokens");
-    eprintln!("  Tool schemas:       {} tokens", t_tools - t_bare);
+    eprintln!("  Tool-conditional:   {} tokens", t_tools - t_bare);
     eprintln!("  Instructions:       {} tokens", estimate_tokens(instr));
     eprintln!("  Skills:             {} tokens", estimate_tokens(skills));
     eprintln!(
diff --git a/crates/loopal-meta-hub/tests/e2e/cluster_harness.rs b/crates/loopal-meta-hub/tests/e2e/cluster_harness.rs
index 10790cc4..475b7abd 100644
--- a/crates/loopal-meta-hub/tests/e2e/cluster_harness.rs
+++ b/crates/loopal-meta-hub/tests/e2e/cluster_harness.rs
@@ -106,7 +106,7 @@ impl HubHandle {
         client.initialize().await.expect("initialize");
         let cwd = std::env::temp_dir();
         client
-            .start_agent(&cwd, None, Some("act"), None, None, true, None, None)
+            .start_agent(&cwd, None, Some("act"), None, None, true, None, None, None)
             .await
             .expect("start_agent");
 
diff --git a/crates/loopal-prompt-system/prompts/agents/default-subagent.md b/crates/loopal-prompt-system/prompts/agents/default-subagent.md
index 5dc4b7a8..e0562179 100644
--- a/crates/loopal-prompt-system/prompts/agents/default-subagent.md
+++ b/crates/loopal-prompt-system/prompts/agents/default-subagent.md
@@ -5,10 +5,19 @@ priority: 100
 ---
 You are a sub-agent named '{{ agent_name | default("sub-agent") }}'. Your working directory is: {{ cwd }}.
 
-Complete the task given to you. When done, output your findings or results directly as a clear summary.
+Complete the task assigned to you and report your results.
 
-Guidelines:
-- Be thorough but efficient.
-- Return file paths as absolute paths.
-- If you encounter issues, report them clearly rather than guessing.
-- Do not make changes outside the scope of your assigned task.
+## Rules
+
+1. **Stay in scope**: Do not make changes outside your assigned task.
+2. **Read before modifying**: Always read a file's current contents before editing it.
+3. **Verify your work**: If you modify code, confirm it compiles or passes basic checks before reporting success.
+4. **Report results, not process**: Focus your output on what you found or accomplished. Skip narrating each step.
+5. **Use absolute paths**: Always reference files with their full absolute path.
+
+## Output
+
+When done, provide a clear summary of:
+- What was accomplished (or what was found, for research tasks)
+- Any issues encountered or decisions made
+- File paths of modified or relevant files
diff --git a/crates/loopal-prompt-system/prompts/agents/explore.md b/crates/loopal-prompt-system/prompts/agents/explore.md
index 3920d3fa..357dd3bb 100644
--- a/crates/loopal-prompt-system/prompts/agents/explore.md
+++ b/crates/loopal-prompt-system/prompts/agents/explore.md
@@ -1,35 +1,46 @@
 ---
 name: Explore Agent
 category: agents
+condition: agent
+condition_value: explore
 priority: 100
 ---
-You are a file search specialist. You excel at thoroughly navigating and exploring codebases.
+You are a codebase exploration specialist. Your sole purpose is to search, read, and analyze existing code — nothing else.
 
 === CRITICAL: READ-ONLY MODE — NO FILE MODIFICATIONS ===
-This is a READ-ONLY exploration task. You are STRICTLY PROHIBITED from:
-- Creating new files (no Write, touch, or file creation of any kind)
-- Modifying existing files (no Edit operations)
-- Deleting files (no rm or deletion)
-- Moving or copying files (no mv or cp)
+You are STRICTLY PROHIBITED from:
+- Creating, modifying, deleting, moving, or copying files
 - Using redirect operators (>, >>, |) or heredocs to write to files
-- Running ANY commands that change system state
+- Running commands that change system or repository state (git add, git commit, npm install, etc.)
 
-Your role is EXCLUSIVELY to search and analyze existing code.
+## Search Strategy
 
-Your strengths:
-- Rapidly finding files using glob patterns
-- Searching code with powerful regex patterns
-- Reading and analyzing file contents
+1. **Start broad**: Use Glob to understand directory structure and find files by name patterns.
+2. **Narrow down**: Use Grep with regex to locate specific code patterns, function definitions, or string literals.
+3. **Read specifics**: Use Read when you know the exact file path and need full context.
+4. **Maximize parallelism**: Launch up to 5 independent tool calls in a single turn. Do not serialize searches that can run concurrently.
 
-Guidelines:
-- Use Glob for file pattern matching
-- Use Grep for content search with regex
-- Use Read when you know the specific file path
-- Use Bash ONLY for read-only operations (ls, git status, git log, git diff, find, cat, head, tail)
-- NEVER use Bash for: mkdir, touch, rm, cp, mv, git add, git commit, npm install, or any modification
-- Return file paths as absolute paths
-- Do not use emojis
+## Tool Preferences
 
-NOTE: You are meant to be a fast agent. Make efficient use of tools and spawn multiple parallel tool calls wherever possible.
+- **Glob** for file pattern matching (`**/*.rs`, `src/**/mod.rs`)
+- **Grep** for content search with regex (function signatures, imports, error patterns)
+- **Read** for reading known files (always use absolute paths)
+- **Bash** ONLY for: `ls`, `git log`, `git diff`, `git blame`, `wc`, `find` (read-only operations)
+- NEVER use Bash for: `mkdir`, `touch`, `rm`, `cp`, `mv`, `git add`, `git commit`
 
-Complete the search request efficiently and report findings clearly.
+## Output Requirements
+
+Structure your findings clearly:
+
+1. **Files found**: List relevant file paths with one-line descriptions of their role.
+2. **Key code excerpts**: Include the most relevant code snippets with `file_path:line_number` format.
+3. **Patterns observed**: Architectural decisions, naming conventions, or recurring patterns you noticed.
+4. **Not found / Uncertain**: Explicitly state what you searched for but could not find. Never fabricate results.
+
+## Guidelines
+
+- Return all file paths as absolute paths.
+- When a search term has multiple possible spellings or conventions (snake_case, camelCase, kebab-case), try all variations.
+- If the scope is unclear, ask for clarification rather than guessing.
+- Keep your response focused and avoid unnecessary commentary.
+- Do not use emojis.
diff --git a/crates/loopal-prompt-system/prompts/agents/plan.md b/crates/loopal-prompt-system/prompts/agents/plan.md
index b02cf190..49447b86 100644
--- a/crates/loopal-prompt-system/prompts/agents/plan.md
+++ b/crates/loopal-prompt-system/prompts/agents/plan.md
@@ -1,22 +1,41 @@
 ---
 name: Plan Agent
 category: agents
+condition: agent
+condition_value: plan
 priority: 100
 ---
-You are a software architect agent. Your job is to design implementation approaches based on exploration results.
-
-Guidelines:
-- Analyze the provided context thoroughly before proposing solutions.
-- Consider multiple approaches and evaluate trade-offs (simplicity, performance, maintainability).
-- Identify critical files that will need modification.
-- Note existing patterns and utilities that should be reused.
-- Flag potential risks, edge cases, and breaking changes.
-- Provide a concrete, step-by-step implementation plan.
-
-Your output should include:
-1. **Approach summary** — one paragraph describing the chosen strategy and why.
-2. **Files to modify** — list with file paths and what changes are needed.
-3. **Implementation steps** — ordered list of specific changes.
-4. **Risks and considerations** — anything the implementer should watch out for.
-
-Keep your plan actionable and specific. Avoid vague suggestions like "refactor as needed."
+You are a software architect agent. Your job is to explore the codebase, understand existing patterns, and design a concrete implementation plan.
+
+## Critical Rules
+
+1. **Read before designing**: You MUST read the critical files yourself. Never design based on assumptions or summaries — verify by reading the actual code.
+2. **Reference existing code**: Identify reusable functions, utilities, types, and patterns. Always cite them with `file_path:line_number`.
+3. **Be specific**: Your plan must be detailed enough for an implementation agent that has never seen this codebase to execute it directly.
+4. **One recommended approach**: Present your best approach with clear reasoning. Do not list multiple alternatives without a recommendation.
+
+## Tools Available
+
+You have read-only access. Use:
+- **Glob** to find files by pattern
+- **Grep** to search for code patterns and references
+- **Read** to examine file contents
+- **Bash** for read-only commands only (ls, git log, git diff, wc)
+
+## Output Format
+
+### Approach Summary
+One paragraph: what you propose, why this approach over alternatives, and the key design decision.
+
+### Critical Files
+List every file that will be modified or created, with a one-line description of the change:
+```
+- `path/to/file.rs` — add new_function() for X
+- `path/to/test.rs` — add test coverage for Y
+```
+
+### Implementation Steps
+Numbered, ordered list of specific changes. Each step should reference exact file paths and function names.
+
+### Risks and Considerations
+Potential issues: breaking changes, edge cases, performance concerns, or dependencies that need attention.
diff --git a/crates/loopal-prompt-system/tests/suite.rs b/crates/loopal-prompt-system/tests/suite.rs
index f8f5223f..85948d82 100644
--- a/crates/loopal-prompt-system/tests/suite.rs
+++ b/crates/loopal-prompt-system/tests/suite.rs
@@ -1,2 +1,4 @@
+#[path = "suite/fragments_agent_test.rs"]
+mod fragments_agent_test;
 #[path = "suite/fragments_test.rs"]
 mod fragments_test;
diff --git a/crates/loopal-prompt-system/tests/suite/fragments_agent_test.rs b/crates/loopal-prompt-system/tests/suite/fragments_agent_test.rs
new file mode 100644
index 00000000..3364a0f1
--- /dev/null
+++ b/crates/loopal-prompt-system/tests/suite/fragments_agent_test.rs
@@ -0,0 +1,90 @@
+use loopal_prompt::{Condition, FragmentRegistry, PromptBuilder, PromptContext};
+use loopal_prompt_system::system_fragments;
+
+#[test]
+fn agent_fragments_have_correct_conditions() {
+    let frags = system_fragments();
+    let explore = frags.iter().find(|f| f.id == "agents/explore").unwrap();
+    assert_eq!(
+        explore.condition,
+        Condition::Agent("explore".into()),
+        "explore.md should have Agent(\"explore\") condition"
+    );
+
+    let plan = frags.iter().find(|f| f.id == "agents/plan").unwrap();
+    assert_eq!(
+        plan.condition,
+        Condition::Agent("plan".into()),
+        "plan.md should have Agent(\"plan\") condition"
+    );
+
+    let default = frags
+        .iter()
+        .find(|f| f.id == "agents/default-subagent")
+        .unwrap();
+    assert_eq!(
+        default.condition,
+        Condition::Always,
+        "default-subagent.md should have Always condition (fallback)"
+    );
+}
+
+#[test]
+fn explore_subagent_gets_explore_fragment_and_core() {
+    let frags = system_fragments();
+    let registry = FragmentRegistry::new(frags);
+
+    let ctx = PromptContext {
+        agent_type: Some("explore".into()),
+        cwd: "/project".into(),
+        tool_names: vec!["Read".into(), "Grep".into(), "Glob".into(), "Bash".into()],
+        ..Default::default()
+    };
+    let selected = registry.select(&ctx);
+    let ids: Vec<&str> = selected.iter().map(|f| f.id.as_str()).collect();
+
+    // Gets explore-specific fragment
+    assert!(ids.contains(&"agents/explore"), "explore fragment missing");
+    // Does NOT get default fallback
+    assert!(
+        !ids.contains(&"agents/default-subagent"),
+        "default should be excluded when explore matches"
+    );
+    // Still gets core/tasks/tools fragments
+    assert!(
+        ids.contains(&"core/identity"),
+        "core identity should be included for sub-agents"
+    );
+    assert!(
+        ids.contains(&"tools/usage-policy"),
+        "usage-policy should be included for sub-agents"
+    );
+}
+
+#[test]
+fn subagent_prompt_includes_core_plus_agent_fragment() {
+    let frags = system_fragments();
+    let registry = FragmentRegistry::new(frags);
+    let builder = PromptBuilder::new(registry);
+
+    let ctx = PromptContext {
+        agent_type: Some("general".into()),
+        cwd: "/work".into(),
+        tool_names: vec!["Read".into(), "Bash".into()],
+        ..Default::default()
+    };
+    let prompt = builder.build(&ctx);
+
+    // Core behavioral fragments present
+    assert!(
+        prompt.contains("Output Efficiency"),
+        "core fragment missing"
+    );
+    // Default sub-agent fragment present (fallback for "general")
+    assert!(
+        prompt.contains("sub-agent"),
+        "default-subagent fragment missing"
+    );
+    // cwd rendered in default-subagent template
+    assert!(prompt.contains("/work"), "cwd not rendered");
+}
diff --git a/crates/loopal-prompt-system/tests/suite/fragments_test.rs b/crates/loopal-prompt-system/tests/suite/fragments_test.rs
index 5c1c07ad..1500c3a0 100644
--- a/crates/loopal-prompt-system/tests/suite/fragments_test.rs
+++ b/crates/loopal-prompt-system/tests/suite/fragments_test.rs
@@ -93,7 +93,8 @@ fn full_prompt_build() {
         prompt.contains("Executing Actions with Care"),
         "safety fragment missing"
     );
-    assert!(prompt.contains("/home/user/project"), "cwd not rendered");
+    // cwd is injected per-turn via env_context for root agent, not in static prompt.
+    // Sub-agent fragments (which do use cwd) are excluded when is_subagent=false.
 }
 
 #[test]
diff --git a/crates/loopal-prompt/src/builder.rs b/crates/loopal-prompt/src/builder.rs
index d7525116..82d58a2d 100644
--- a/crates/loopal-prompt/src/builder.rs
+++ b/crates/loopal-prompt/src/builder.rs
@@ -13,20 +13,15 @@ impl PromptBuilder {
 
     /// Build the full system prompt for the given context.
     ///
-    /// Assembly order:
-    /// 1. User instructions (raw, highest priority)
-    /// 2. Matched & rendered fragments (sorted by priority)
+    /// Assembly order (identity-first for stronger LLM attention):
+    /// 1. Matched & rendered fragments (sorted by priority: core → tasks → tools → modes)
+    /// 2. User instructions (project-specific, after core behavioral rules)
     /// 3. Skills summary
     /// 4. Project memory (tail)
     pub fn build(&self, ctx: &PromptContext) -> String {
         let mut parts = Vec::new();
 
-        // 1. User instructions (injected raw)
-        if !ctx.instructions.is_empty() {
-            parts.push(ctx.instructions.clone());
-        }
-
-        // 2. Fragments: select → sort → render → collect
+        // 1. Fragments: select → sort → render → collect
         for frag in self.registry.select(ctx) {
             let rendered = self.registry.render(frag, ctx);
             let trimmed = rendered.trim();
@@ -35,6 +30,11 @@ impl PromptBuilder {
             }
         }
 
+        // 2. User instructions (after fragments so identity/core rules come first)
+        if !ctx.instructions.is_empty() {
+            parts.push(ctx.instructions.clone());
+        }
+
         // 3. Skills summary
         if !ctx.skills_summary.is_empty() {
             parts.push(ctx.skills_summary.clone());
@@ -48,30 +48,8 @@ impl PromptBuilder {
         parts.join("\n\n")
     }
 
-    /// Build a prompt for a specific sub-agent type.
-    ///
-    /// Looks for a fragment with id "agents/{agent_type}" and renders it.
-    /// Falls back to the default sub-agent prompt if not found.
-    pub fn build_agent_prompt(&self, agent_type: &str, ctx: &PromptContext) -> String {
-        let agent_id = format!("agents/{agent_type}");
-        if let Some(frag) = self.registry.fragments().iter().find(|f| f.id == agent_id) {
-            self.registry.render(frag, ctx)
-        } else {
-            default_agent_prompt(ctx)
-        }
-    }
-
     /// Access the underlying registry.
     pub fn registry(&self) -> &FragmentRegistry {
         &self.registry
     }
 }
-
-fn default_agent_prompt(ctx: &PromptContext) -> String {
-    let name = ctx.agent_name.as_deref().unwrap_or("sub-agent");
-    format!(
-        "You are a sub-agent named '{name}'. Your working directory is: {}. \
-         Complete the task given to you and report your findings.",
-        ctx.cwd
-    )
-}
diff --git a/crates/loopal-prompt/src/context.rs b/crates/loopal-prompt/src/context.rs
index e9021345..c3b049af 100644
--- a/crates/loopal-prompt/src/context.rs
+++ b/crates/loopal-prompt/src/context.rs
@@ -37,10 +37,18 @@ pub struct PromptContext {
     // -- Sub-agent context --
     #[serde(skip_serializing_if = "Option::is_none")]
     pub agent_name: Option<String>,
+    /// Agent type for fragment selection. Some(_) implies this is a sub-agent.
     #[serde(skip_serializing_if = "Option::is_none")]
     pub agent_type: Option<String>,
 }
 
+impl PromptContext {
+    /// Whether this context is for a sub-agent (determined by agent_type presence).
+    pub fn is_subagent(&self) -> bool {
+        self.agent_type.is_some()
+    }
+}
+
 impl Default for PromptContext {
     fn default() -> Self {
         Self {
diff --git a/crates/loopal-prompt/src/fragment.rs b/crates/loopal-prompt/src/fragment.rs
index 98ab2e50..995043c4 100644
--- a/crates/loopal-prompt/src/fragment.rs
+++ b/crates/loopal-prompt/src/fragment.rs
@@ -37,6 +37,8 @@ pub enum Condition {
     Feature(String),
     /// Only when the specified tool is available.
     Tool(String),
+    /// Only when spawned as the specified agent type ("explore", "plan").
+    Agent(String),
 }
 
 // -- Frontmatter deserialization --
@@ -128,6 +130,7 @@ fn parse_condition(kind: Option<&str>, value: Option<&str>) -> Condition {
         Some("mode") => Condition::Mode(value.unwrap_or("plan").to_string()),
         Some("feature") => Condition::Feature(value.unwrap_or("").to_string()),
         Some("tool") => Condition::Tool(value.unwrap_or("").to_string()),
+        Some("agent") => Condition::Agent(value.unwrap_or("").to_string()),
         Some(other) => {
             tracing::warn!(
                 condition = other,
diff --git a/crates/loopal-prompt/src/registry.rs b/crates/loopal-prompt/src/registry.rs
index 7ddeee8b..fe0c9458 100644
--- a/crates/loopal-prompt/src/registry.rs
+++ b/crates/loopal-prompt/src/registry.rs
@@ -1,7 +1,7 @@
 use std::path::Path;
 
 use crate::context::PromptContext;
-use crate::fragment::{Condition, Fragment, parse_fragment, parse_fragments_from_dir};
+use crate::fragment::{Category, Condition, Fragment, parse_fragment, parse_fragments_from_dir};
 use crate::render::PromptRenderer;
 
 /// Manages a collection of prompt fragments with selection, rendering,
@@ -41,11 +41,37 @@ impl FragmentRegistry {
     }
 
     /// Select fragments that match the given context, sorted by priority.
+    ///
+    /// Agents-category fragments are excluded for the root agent (non-subagent).
+    /// When a specific Agent condition matches (e.g. "explore"), the default
+    /// fallback agent fragment (Always + Agents) is excluded to avoid overlap.
     pub fn select<'a>(&'a self, ctx: &PromptContext) -> Vec<&'a Fragment> {
+        let is_subagent = ctx.is_subagent();
+
+        // Check if any Agent(type) condition matches for this context.
+        let has_specific_agent = is_subagent
+            && self.fragments.iter().any(|f| {
+                f.category == Category::Agents
+                    && matches!(&f.condition, Condition::Agent(t) if ctx.agent_type.as_deref() == Some(t.as_str()))
+            });
+
         let mut matched: Vec<&Fragment> = self
             .fragments
             .iter()
-            .filter(|f| condition_matches(&f.condition, ctx))
+            .filter(|f| {
+                // Agents category only relevant for sub-agents
+                if f.category == Category::Agents && !is_subagent {
+                    return false;
+                }
+                // If a specific agent fragment matches, skip the Always fallback
+                if has_specific_agent
+                    && f.category == Category::Agents
+                    && f.condition == Condition::Always
+                {
+                    return false;
+                }
+                condition_matches(&f.condition, ctx)
+            })
             .collect();
         matched.sort_by_key(|f| f.priority);
         matched
@@ -68,6 +94,7 @@ fn condition_matches(cond: &Condition, ctx: &PromptContext) -> bool {
         Condition::Mode(m) => ctx.mode == *m,
         Condition::Feature(f) => ctx.features.contains(f),
         Condition::Tool(t) => ctx.tool_names.contains(t),
+        Condition::Agent(t) => ctx.agent_type.as_deref() == Some(t.as_str()),
     }
 }
 
diff --git a/crates/loopal-prompt/tests/suite.rs b/crates/loopal-prompt/tests/suite.rs
index 0e4d555f..7dee5b07 100644
--- a/crates/loopal-prompt/tests/suite.rs
+++ b/crates/loopal-prompt/tests/suite.rs
@@ -2,6 +2,8 @@
 mod builder_test;
 #[path = "suite/fragment_test.rs"]
 mod fragment_test;
+#[path = "suite/registry_agent_test.rs"]
+mod registry_agent_test;
 #[path = "suite/registry_test.rs"]
 mod registry_test;
 #[path = "suite/render_test.rs"]
diff --git a/crates/loopal-prompt/tests/suite/builder_test.rs b/crates/loopal-prompt/tests/suite/builder_test.rs
index 08eb1f31..2fc4fd7a 100644
--- a/crates/loopal-prompt/tests/suite/builder_test.rs
+++ b/crates/loopal-prompt/tests/suite/builder_test.rs
@@ -1,13 +1,11 @@
-use loopal_prompt::{
-    Category, Condition, Fragment, FragmentRegistry, PromptBuilder, PromptContext,
-};
+use loopal_prompt::{FragmentRegistry, PromptBuilder, PromptContext};
 
-fn frag(id: &str, priority: u16, content: &str) -> Fragment {
-    Fragment {
+fn frag(id: &str, priority: u16, content: &str) -> loopal_prompt::Fragment {
+    loopal_prompt::Fragment {
         id: id.to_string(),
         name: id.to_string(),
-        category: Category::Core,
-        condition: Condition::Always,
+        category: loopal_prompt::Category::Core,
+        condition: loopal_prompt::Condition::Always,
         priority,
         content: content.to_string(),
     }
@@ -36,7 +34,7 @@ fn build_includes_instructions_and_memory() {
         ..Default::default()
     };
     let prompt = builder.build(&ctx);
-    assert!(prompt.starts_with("Be helpful."));
+    assert!(prompt.contains("Be helpful."));
     assert!(prompt.contains("# Project Memory"));
     assert!(prompt.contains("User prefers Rust."));
 }
@@ -54,33 +52,18 @@ fn build_skips_empty_renders() {
 }
 
 #[test]
-fn build_agent_prompt_fallback() {
-    let builder = PromptBuilder::new(FragmentRegistry::new(vec![]));
-    let ctx = PromptContext {
-        agent_name: Some("explorer".into()),
-        cwd: "/work".into(),
-        ..Default::default()
-    };
-    let prompt = builder.build_agent_prompt("nonexistent", &ctx);
-    assert!(prompt.contains("explorer"));
-    assert!(prompt.contains("/work"));
-}
-
-#[test]
-fn build_agent_prompt_uses_fragment() {
-    let frags = vec![Fragment {
-        id: "agents/explore".into(),
-        name: "Explore".into(),
-        category: Category::Agents,
-        condition: Condition::Always,
-        priority: 100,
-        content: "You are an explorer in {{ cwd }}.".into(),
-    }];
+fn fragments_come_before_instructions() {
+    let frags = vec![frag("core/id", 100, "## Identity")];
     let builder = PromptBuilder::new(FragmentRegistry::new(frags));
     let ctx = PromptContext {
-        cwd: "/project".into(),
+        instructions: "## User Instructions".into(),
         ..Default::default()
     };
-    let prompt = builder.build_agent_prompt("explore", &ctx);
-    assert_eq!(prompt, "You are an explorer in /project.");
+    let prompt = builder.build(&ctx);
+    let frag_pos = prompt.find("## Identity").unwrap();
+    let instr_pos = prompt.find("## User Instructions").unwrap();
+    assert!(
+        frag_pos < instr_pos,
+        "fragments ({frag_pos}) should come before instructions ({instr_pos})"
+    );
 }
diff --git a/crates/loopal-prompt/tests/suite/fragment_test.rs b/crates/loopal-prompt/tests/suite/fragment_test.rs
index 93348004..a2b823b1 100644
--- a/crates/loopal-prompt/tests/suite/fragment_test.rs
+++ b/crates/loopal-prompt/tests/suite/fragment_test.rs
@@ -67,3 +67,21 @@ fn returns_none_without_frontmatter() {
     let raw = "Just plain text, no frontmatter.";
     assert!(parse_fragment("bad", raw).is_none());
 }
+
+#[test]
+fn parse_agent_condition() {
+    let raw = "\
+---
+name: Explore Agent
+category: agents
+condition: agent
+condition_value: explore
+priority: 100
+---
+Explore instructions.
+";
+    let frag = parse_fragment("agents/explore", raw).unwrap();
+    assert_eq!(frag.category, Category::Agents);
+    assert_eq!(frag.condition, Condition::Agent("explore".to_string()));
+    assert_eq!(frag.priority, 100);
+}
diff --git a/crates/loopal-prompt/tests/suite/registry_agent_test.rs b/crates/loopal-prompt/tests/suite/registry_agent_test.rs
new file mode 100644
index 00000000..1e3f222a
--- /dev/null
+++ b/crates/loopal-prompt/tests/suite/registry_agent_test.rs
@@ -0,0 +1,141 @@
+use loopal_prompt::{Category, Condition, Fragment, FragmentRegistry, PromptContext};
+
+fn agent_frag(
+    id: &str,
+    priority: u16,
+    condition: Condition,
+    category: Category,
+    content: &str,
+) -> Fragment {
+    Fragment {
+        id: id.to_string(),
+        name: id.to_string(),
+        category,
+        condition,
+        priority,
+        content: content.to_string(),
+    }
+}
+
+#[test]
+fn agent_condition_matches_only_when_type_set() {
+    let frags = vec![agent_frag(
+        "agents/explore",
+        100,
+        Condition::Agent("explore".into()),
+        Category::Agents,
+        "explore prompt",
+    )];
+    let registry = FragmentRegistry::new(frags);
+
+    // Root agent (no agent_type) → agents excluded
+    let root_ctx = PromptContext::default();
+    assert!(registry.select(&root_ctx).is_empty());
+
+    // Sub-agent with wrong type → Agent condition doesn't match
+    let wrong_type = PromptContext {
+        agent_type: Some("plan".into()),
+        ..Default::default()
+    };
+    assert!(registry.select(&wrong_type).is_empty());
+
+    // Sub-agent with correct type → matches
+    let correct_type = PromptContext {
+        agent_type: Some("explore".into()),
+        ..Default::default()
+    };
+    let selected: Vec<&str> = registry
+        .select(&correct_type)
+        .iter()
+        .map(|f| f.id.as_str())
+        .collect();
+    assert_eq!(selected, vec!["agents/explore"]);
+}
+
+#[test]
+fn default_agent_excluded_when_specific_agent_matches() {
+    let frags = vec![
+        agent_frag(
+            "agents/default",
+            100,
+            Condition::Always,
+            Category::Agents,
+            "default sub-agent",
+        ),
+        agent_frag(
+            "agents/explore",
+            100,
+            Condition::Agent("explore".into()),
+            Category::Agents,
+            "explore prompt",
+        ),
+    ];
+    let registry = FragmentRegistry::new(frags);
+
+    // Sub-agent "general" (no specific match) → gets default only
+    let general = PromptContext {
+        agent_type: Some("general".into()),
+        ..Default::default()
+    };
+    let ids: Vec<&str> = registry
+        .select(&general)
+        .iter()
+        .map(|f| f.id.as_str())
+        .collect();
+    assert_eq!(ids, vec!["agents/default"]);
+
+    // Sub-agent "explore" → gets explore, NOT default
+    let explore = PromptContext {
+        agent_type: Some("explore".into()),
+        ..Default::default()
+    };
+    let ids: Vec<&str> = registry
+        .select(&explore)
+        .iter()
+        .map(|f| f.id.as_str())
+        .collect();
+    assert_eq!(ids, vec!["agents/explore"]);
+}
+
+#[test]
+fn agents_category_excluded_for_root() {
+    let frags = vec![
+        Fragment {
+            id: "core/identity".into(),
+            name: "core/identity".into(),
+            category: Category::Core,
+            condition: Condition::Always,
+            priority: 100,
+            content: "identity".into(),
+        },
+        agent_frag(
+            "agents/default",
+            200,
+            Condition::Always,
+            Category::Agents,
+            "default agent",
+        ),
+    ];
+    let registry = FragmentRegistry::new(frags);
+
+    // Root agent → only core, no agents
+    let root_ctx = PromptContext::default();
+    let ids: Vec<&str> = registry
+        .select(&root_ctx)
+        .iter()
+        .map(|f| f.id.as_str())
+        .collect();
+    assert_eq!(ids, vec!["core/identity"]);
+
+    // Sub-agent → both
+    let sub_ctx = PromptContext {
+        agent_type: Some("general".into()),
+        ..Default::default()
+    };
+    let ids: Vec<&str> = registry
+        .select(&sub_ctx)
+        .iter()
+        .map(|f| f.id.as_str())
+        .collect();
+    assert_eq!(ids, vec!["core/identity", "agents/default"]);
+}
diff --git a/crates/loopal-runtime/tests/agent_loop/llm_test.rs b/crates/loopal-runtime/tests/agent_loop/llm_test.rs
index 39f59c70..9f98ce22 100644
--- a/crates/loopal-runtime/tests/agent_loop/llm_test.rs
+++ b/crates/loopal-runtime/tests/agent_loop/llm_test.rs
@@ -205,6 +205,7 @@ fn report_real_system_prompt_tokens() {
         "/Users/dev/project",
         "",
         "",
+        None,
     );
     runner.params.config.system_prompt = real_prompt.clone();
     let params = runner
@@ -221,6 +222,7 @@ fn report_real_system_prompt_tokens() {
         "/Users/dev/project",
         "",
         "",
+        None,
     );
     let fragment_tokens = loopal_context::estimate_tokens(&prompt_no_tools);
 
diff --git a/crates/loopal-test-support/src/ipc_harness.rs b/crates/loopal-test-support/src/ipc_harness.rs
index f3c383fe..d945b2e1 100644
--- a/crates/loopal-test-support/src/ipc_harness.rs
+++ b/crates/loopal-test-support/src/ipc_harness.rs
@@ -77,6 +77,7 @@ pub async fn build_ipc_harness(
             false,
             None,
             None,
+            None,
         )
         .await
         .expect("agent/start failed");
diff --git a/src/bootstrap/hub_bootstrap.rs b/src/bootstrap/hub_bootstrap.rs
index c5eb93ad..8df4a4b9 100644
--- a/src/bootstrap/hub_bootstrap.rs
+++ b/src/bootstrap/hub_bootstrap.rs
@@ -67,6 +67,7 @@ pub async fn bootstrap_hub_and_agent(
             cli.no_sandbox,
             cli.resume.as_deref(),
             lifecycle_str,
+            None, // root agent has no agent_type
         )
         .await?;