fix(TASK-035): close agent-bracket leaks and prove compression/role-hook seams

- Open the top-level agent bracket only at committed dispatch sites so
  inspection-flag early returns (--list-sessions, --info, --macro,
  --rebuild-rag, --init-*, --build-tools) never leak agent.started
- Close the agent bracket ahead of process::exit in shell_execute's
  execute branch, where the dispatch wrapper's close is unreachable
- Extract the ask-loop auto-compression seam and cover the
  session.compressed success path at both call sites with sink tests
- Prove Session.role_hooks never serializes and that injected
  role_hooks YAML is ignored on deserialization

src/config/session.rs

@@ -1266,6 +1266,37 @@ mod tests {
         assert!(!yaml.contains("token_usage"));
     }
 
+    #[test]
+    fn session_role_hooks_is_not_serialized() {
+        let content = "---\nhooks:\n  turn.completed:\n    - name: role-probe\n      command: role-cmd\n---\nPrompt";
+        let mut session = Session::default();
+        session.set_role(Role::new("hooked", content));
+        assert!(session.role_hooks().is_some());
+
+        let yaml = serde_yaml::to_string(&session).unwrap();
+        assert!(
+            !yaml.contains("role_hooks"),
+            "the role-hooks snapshot must never reach session YAML"
+        );
+        assert!(
+            !yaml.contains("role-cmd"),
+            "no hook command text may leak into session YAML under any key"
+        );
+    }
+
+    #[test]
+    fn session_deserialization_ignores_injected_role_hooks() {
+        // A session file is attacker-writable data: a `role_hooks:` key
+        // crafted into it must never deserialize into executable hook
+        // config — hooks are only ever re-read from the role file itself.
+        let yaml = "model: ''\nmessages: []\nrole_hooks:\n  session.started:\n    - name: injected\n      command: touch /tmp/pwned\n";
+        let session: Session = serde_yaml::from_str(yaml).unwrap();
+        assert!(
+            session.role_hooks().is_none(),
+            "injected role_hooks in session YAML must be ignored"
+        );
+    }
+
     #[test]
     fn session_cost_accumulates_and_survives_yaml_round_trip() {
         let mut session = Session::default();

src/main.rs

@@ -388,7 +388,12 @@ async fn run(
         let app = Arc::clone(&ctx.app.config);
         ctx.use_agent(app.as_ref(), agent, session, abort_signal.clone())
             .await?;
-        ctx.top_level_agent_started();
+        // The top-level agent bracket does NOT open here: the inspection
+        // flags below (--list-sessions, --info, --macro, --rebuild-rag,
+        // --init-*) return early without running the agent, and an open
+        // bracket on those paths would leak `agent.started` with no
+        // terminal event. It opens at the dispatch sites further down,
+        // where an agent run is actually committed.
     } else {
         let app: Arc<AppConfig> = Arc::clone(&ctx.app.config);
         if let Some(prompt) = &cli.temp_role {
@@ -525,6 +530,9 @@ async fn run(
         return Ok(());
     }
     if cli.execute && !is_repl {
+        // Dispatch committed: open the top-level agent bracket only now,
+        // past every inspection-flag early return above.
+        ctx.top_level_agent_started();
         let result = async {
             let input = create_input(&ctx, text, &cli.file, abort_signal.clone()).await?;
             shell_execute(&mut ctx, &SHELL, input, abort_signal.clone()).await
@@ -550,6 +558,12 @@ async fn run(
         return acp::run_acp_server(ctx, abort_signal).await;
     }
 
+    // Dispatch committed (headless directive or REPL): open the top-level
+    // agent bracket only now, past every inspection-flag early return
+    // above. The ACP server stays unbracketed deliberately. The headless
+    // arm closes the bracket below; the REPL closes it at teardown (or on
+    // an interactive agent switch).
+    ctx.top_level_agent_started();
     match is_repl {
         false => {
             let result = async {
@@ -692,6 +706,12 @@ async fn shell_execute(
                     if code == 0 && app.save_shell_history {
                         let _ = append_to_shell_history(&shell.name, &eval_str, code);
                     }
+                    // `process::exit` never returns, so the dispatch
+                    // wrapper's bracket close in `run()` is unreachable
+                    // from here: close the top-level agent bracket now.
+                    // The agent run itself succeeded — the exit code
+                    // belongs to the user's shell command, not to us.
+                    ctx.top_level_agent_finished(None);
                     hooks::drain_pending(EXIT_HOOK_DRAIN_TIMEOUT).await;
                     process::exit(code);
                 }

src/repl/mod.rs

@@ -1644,10 +1644,7 @@ async fn ask_inner(
             };
             eprintln!("\n📢 {}", color.italic().paint("Compressing the session."),);
 
-            match ctx.compress_session().await {
-                Ok(()) => hooks::fire(HookEvent::SessionCompressed, ctx, &[], None),
-                Err(err) => warn!("Failed to compress the session: {err}"),
-            }
+            auto_compress_session(ctx).await;
             if let Some(session) = ctx.session.as_mut() {
                 session.set_compressing(false);
             }
@@ -1698,6 +1695,18 @@ fn print_turn_divider() {
     println!();
 }
 
+/// Auto-compression seam for the ask loop: `session.compressed` fires only
+/// when compression actually succeeded. A failed auto-compression is
+/// survivable mid-session, so the error is logged and swallowed rather
+/// than aborting the turn — unlike `.compress session`, which surfaces
+/// the failure to the user.
+async fn auto_compress_session(ctx: &mut RequestContext) {
+    match ctx.compress_session().await {
+        Ok(()) => hooks::fire(HookEvent::SessionCompressed, ctx, &[], None),
+        Err(err) => warn!("Failed to compress the session: {err}"),
+    }
+}
+
 fn should_continue(ctx: &RequestContext) -> bool {
     let config = ctx.auto_continue_config();
     ctx.app.config.function_calling_support
@@ -2005,7 +2014,8 @@ pub fn split_args_text(line: &str, is_win: bool) -> (Vec<String>, &str) {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::config::{AppState, Session, WorkingMode};
+    use crate::client::{ClientConfig, Model};
+    use crate::config::{AppState, RoleLike, Session, WorkingMode};
     use crate::hooks::{HookDef, HooksMap, test_sink};
     use anyhow::anyhow;
     use serial_test::serial;
@@ -2180,6 +2190,113 @@ mod tests {
         );
     }
 
+    /// A ctx whose compression can actually SUCCEED offline: `dry_run`
+    /// makes the summarization call echo instead of hitting the network,
+    /// the default openai client config makes the session model
+    /// resolvable, and the session already holds a user exchange so
+    /// `compress_session` has something to summarize.
+    fn compressible_ctx(marker: &str) -> RequestContext {
+        let mut hooks_map = HooksMap::default();
+        hooks_map.insert(
+            "session.compressed".to_string(),
+            vec![HookDef {
+                name: format!("{marker}-session.compressed"),
+                command: "true".to_string(),
+            }],
+        );
+        let mut app = AppState::test_default();
+        app.config = Arc::new(AppConfig {
+            hooks: hooks_map,
+            dry_run: true,
+            clients: vec![ClientConfig::default()],
+            ..Default::default()
+        });
+        let mut ctx = RequestContext::new(Arc::new(app), WorkingMode::Repl);
+        let mut session = Session::default();
+        session.set_model(Model::new("openai", "test-compress-model"));
+        ctx.session = Some(session);
+        let input = Input::from_str(&ctx, "hello there", None).unwrap();
+        ctx.session
+            .as_mut()
+            .unwrap()
+            .add_message(&input, "hi")
+            .unwrap();
+        ctx
+    }
+
+    fn session_compressed_count(marker: &str) -> usize {
+        test_sink::snapshot()
+            .iter()
+            .filter(|capture| capture.hook_name == format!("{marker}-session.compressed"))
+            .count()
+    }
+
+    #[tokio::test]
+    #[serial]
+    async fn compress_command_success_fires_session_compressed_once() {
+        let _sink = test_sink::install();
+        let marker = "compress-ok-w2p";
+        let mut ctx = compressible_ctx(marker);
+        let abort_signal = create_abort_signal();
+
+        // Boxed: `run_repl_command`'s state machine is far larger than a test
+        // thread's stack.
+        let result = Box::pin(run_repl_command(
+            &mut ctx,
+            abort_signal,
+            ".compress session",
+        ))
+        .await;
+
+        assert!(
+            result.is_ok(),
+            "dry-run compression must succeed: {result:?}"
+        );
+        assert_eq!(
+            session_compressed_count(marker),
+            1,
+            "a successful `.compress session` must fire session.compressed exactly once"
+        );
+    }
+
+    #[tokio::test]
+    #[serial]
+    async fn auto_compress_success_fires_session_compressed_once() {
+        // Covers the ask-loop call site through its extracted seam;
+        // driving `needs_compression` itself requires a full LLM turn and
+        // is exercised end-to-end by integration coverage, not here.
+        let _sink = test_sink::install();
+        let marker = "auto-compress-ok-f8r";
+        let mut ctx = compressible_ctx(marker);
+
+        auto_compress_session(&mut ctx).await;
+
+        assert_eq!(
+            session_compressed_count(marker),
+            1,
+            "a successful auto-compression must fire session.compressed exactly once"
+        );
+    }
+
+    #[tokio::test]
+    #[serial]
+    async fn auto_compress_failure_fires_no_session_compressed_and_swallows_error() {
+        let _sink = test_sink::install();
+        let marker = "auto-compress-fail-d3k";
+        let mut ctx = compressible_ctx(marker);
+        // An empty session makes compress_session fail; the seam must
+        // swallow the error (no panic, no propagation) and fire nothing.
+        ctx.session = Some(Session::default());
+
+        auto_compress_session(&mut ctx).await;
+
+        assert_eq!(
+            session_compressed_count(marker),
+            0,
+            "a failed auto-compression must fire no session.compressed hook"
+        );
+    }
+
     #[test]
     fn test_process_command_line() {
         assert_eq!(parse_command(" ."), Some((".", None)));