diff --git a/crates/buzz-acp/src/config.rs b/crates/buzz-acp/src/config.rs
index d9596858460..cd3385b40db 100644
--- a/crates/buzz-acp/src/config.rs
+++ b/crates/buzz-acp/src/config.rs
@@ -159,6 +159,36 @@ impl std::fmt::Display for PermissionMode {
     }
 }
 
+/// One-line operator warning for a permission mode that silently denies the
+/// agent's own reply path. Returns `None` for modes that don't.
+///
+/// `dontAsk` is the default. It overrides the agent's configured default mode,
+/// so every tool call not matched by a preauthorization rule is denied without
+/// prompting — including the bundled `buzz` CLI the agent is prompted to reply
+/// with. The resulting failure is invisible from the outside: the denial
+/// happens *inside* the agent, so no `session/request_permission` ever reaches
+/// this harness and there is nothing for it to log. An agent stays connected
+/// and subscribed, consumes a full turn, and publishes nothing.
+///
+/// Stating it once at startup costs one line and saves operators from
+/// diagnosing an empty channel as a relay or delivery problem.
+///
+/// Kept as a free function over the mode so the text is unit-testable without
+/// constructing a full [`Config`].
+pub fn unattended_denial_warning(mode: PermissionMode) -> Option<&'static str> {
+    matches!(mode, PermissionMode::DontAsk).then_some(
+        "permission_mode=dontAsk — tool calls not matched by a preauthorization rule are denied \
+         inside the agent without prompting, including the bundled `buzz` CLI agents use to \
+         publish replies, so an agent may consume a full turn and post nothing. Matching \
+         permissions.allow rules do still apply, but Claude Code ignores a project's \
+         .claude/settings.json in a workspace that was never trusted interactively — use \
+         .claude/settings.local.json to preauthorize commands for a headless agent directory. \
+         To leave the agent's own default mode in charge instead, set --permission-mode \
+         (BUZZ_ACP_PERMISSION_MODE) to `default`; unattended permission requests that reach the \
+         harness are still rejected either way.",
+    )
+}
+
 /// CLI args for `buzz-acp models` — query available models from an agent.
 ///
 /// This is a standalone `Parser` (not a subcommand variant) because the
@@ -427,8 +457,31 @@ pub struct CliArgs {
     /// Permission mode for agents that support `session/set_config_option`
     /// with `configId: "mode"` (e.g. `claude-agent-acp`).
     ///
-    /// Defaults to `dontAsk`, which rejects operations that need interactive
-    /// approval because Buzz does not expose a human permission prompt.
+    /// Defaults to `dontAsk`, which denies operations that need interactive
+    /// approval, without prompting, because Buzz does not expose a human
+    /// permission prompt. It is applied with `session/set_config_option`
+    /// *after* the session is created, so it overrides the agent's configured
+    /// **default mode** — but not its preauthorization: matching
+    /// `permissions.allow` rules are still evaluated per tool call, so a
+    /// pre-authorized command runs and only unmatched ones are denied.
+    ///
+    /// One caveat bites headless deployments. Claude Code ignores
+    /// `permissions.allow` from a project's `.claude/settings.json` in a
+    /// workspace that was never trusted interactively — which is every
+    /// desktop-managed agent directory — and says so only on its own stderr.
+    /// `.claude/settings.local.json` is honoured in that case and is the
+    /// reliable target for a seeded rule.
+    ///
+    /// An unmatched call under `dontAsk` includes the bundled `buzz` CLI,
+    /// which is how agents publish their replies — so an agent can receive a
+    /// mention, do the work, and then be denied the send.
+    ///
+    /// Select `default` to leave the agent's own configuration in charge: the
+    /// harness sends no `set_config_option` at all in that mode. This is the
+    /// supported way to run an agent that pre-authorizes its own tools, and it
+    /// does not weaken the harness guarantee — any `session/request_permission`
+    /// that reaches buzz-acp is still rejected, because there is no human to
+    /// ask.
     #[arg(
         long,
         env = "BUZZ_ACP_PERMISSION_MODE",
@@ -2276,6 +2329,57 @@ channels = "ALL"
         assert!(!PermissionMode::Plan.is_default());
     }
 
+    #[test]
+    /// Only `dontAsk` denies the agent's reply path, so only `dontAsk` warns.
+    /// A warning on every mode would be noise operators learn to skip.
+    fn test_unattended_denial_warning_only_for_dont_ask() {
+        assert!(unattended_denial_warning(PermissionMode::DontAsk).is_some());
+        assert!(unattended_denial_warning(PermissionMode::Default).is_none());
+        assert!(unattended_denial_warning(PermissionMode::AcceptEdits).is_none());
+        assert!(unattended_denial_warning(PermissionMode::Plan).is_none());
+    }
+
+    #[test]
+    /// The warning is only useful if it names the way out. Pin both the escape
+    /// hatch and the guarantee it does not weaken, so a future edit that drops
+    /// either one fails here instead of shipping a dead-end warning.
+    fn test_unattended_denial_warning_names_the_escape_hatch() {
+        let warning = unattended_denial_warning(PermissionMode::DontAsk)
+            .expect("dontAsk must produce a warning");
+        assert!(
+            warning.contains("BUZZ_ACP_PERMISSION_MODE"),
+            "warning must name the env var operators can set, got: {warning}"
+        );
+        assert!(
+            warning.contains("`default`"),
+            "warning must name the mode that defers to the agent's config, got: {warning}"
+        );
+        assert!(
+            warning.contains("still rejected"),
+            "warning must state the harness guarantee is unchanged, got: {warning}"
+        );
+    }
+
+    #[test]
+    /// The warning must not tell operators that agent-side preauthorization is
+    /// futile — matching allow rules are evaluated per tool call even under
+    /// `dontAsk`. The trap is the workspace-trust gate, which silently drops
+    /// `permissions.allow` from a project `settings.json` in a nest that was
+    /// never opened interactively, so name the file that does work.
+    fn test_unattended_denial_warning_points_at_settings_local() {
+        let warning = unattended_denial_warning(PermissionMode::DontAsk)
+            .expect("dontAsk must produce a warning");
+        assert!(
+            warning.contains("settings.local.json"),
+            "warning must name the settings file honoured in an untrusted workspace, \
+             got: {warning}"
+        );
+        assert!(
+            warning.contains("do still apply"),
+            "warning must not imply allow rules are ineffective under dontAsk, got: {warning}"
+        );
+    }
+
     #[test]
     fn test_permission_mode_display() {
         assert_eq!(format!("{}", PermissionMode::DontAsk), "dontAsk");
diff --git a/crates/buzz-acp/src/lib.rs b/crates/buzz-acp/src/lib.rs
index 65c9dd6203c..c5029905804 100644
--- a/crates/buzz-acp/src/lib.rs
+++ b/crates/buzz-acp/src/lib.rs
@@ -1578,6 +1578,13 @@ async fn tokio_main() -> Result<()> {
 
     tracing::info!("buzz-acp starting: {}", config.summary());
 
+    // The dontAsk denial happens inside the agent, so it never surfaces as a
+    // harness log line. Say it once here rather than leave an empty channel as
+    // the only symptom.
+    if let Some(warning) = crate::config::unattended_denial_warning(config.permission_mode) {
+        tracing::warn!("buzz-acp: {warning}");
+    }
+
     let observer = config
         .relay_observer
         .then(observer::ObserverHandle::in_process);