fix(desktop): restart agents after setup fixes

Co-authored-by: npub137662rjc45tlwzl38m4f828yc3ymzugkudyp8w0jdpvxj7ysen3sm3uh7m <8fb5a50e58ad17f70bf13eea93a8e4c449b17116e34813b9f26858697890cce3@sprout-oss.stage.blox.sqprod.co>
Signed-off-by: npub137662rjc45tlwzl38m4f828yc3ymzugkudyp8w0jdpvxj7ysen3sm3uh7m <8fb5a50e58ad17f70bf13eea93a8e4c449b17116e34813b9f26858697890cce3@sprout-oss.stage.blox.sqprod.co>

desktop/src-tauri/src/commands/agent_models.rs

@@ -12,10 +12,11 @@ use crate::{
     managed_agents::{
         build_managed_agent_summary, current_instance_id, default_agent_workdir,
         discovery_env_with_baked_floor, find_managed_agent_mut, known_acp_runtime,
-        load_managed_agents, load_personas, managed_agent_avatar_url, missing_command_message,
-        normalize_agent_args, resolve_command, save_managed_agents, sync_managed_agent_processes,
-        try_regenerate_nest, AgentModelInfo, AgentModelsResponse, UpdateManagedAgentRequest,
-        UpdateManagedAgentResponse, DEFAULT_ACP_COMMAND,
+        load_managed_agents, load_personas, managed_agent_avatar_url,
+        maybe_restart_setup_mode_agent, missing_command_message, normalize_agent_args,
+        resolve_command, save_managed_agents, sync_managed_agent_processes, try_regenerate_nest,
+        AgentModelInfo, AgentModelsResponse, UpdateManagedAgentRequest, UpdateManagedAgentResponse,
+        DEFAULT_ACP_COMMAND,
     },
     relay::{relay_ws_url_with_override, sync_managed_agent_profile},
     util::now_iso,
@@ -786,9 +787,8 @@ async fn run_agent_models_command(
 
 /// Update mutable fields on an existing managed agent record.
 ///
-/// Does NOT auto-restart the agent. Runtime config changes (system prompt,
-/// parallelism, commands, toolsets) take effect on the next agent spawn.
-/// Name changes are synced to the relay immediately via a kind:0 re-publish.
+/// Does NOT auto-restart healthy agents; only setup-mode listeners are respawned.
+/// Other runtime config changes take effect on next spawn; renames sync now.
 #[tauri::command]
 pub async fn update_managed_agent(
     input: UpdateManagedAgentRequest,
@@ -910,6 +910,8 @@ pub async fn update_managed_agent(
 
         record.updated_at = now_iso();
 
+        maybe_restart_setup_mode_agent(&app, &state, record, &mut runtimes);
+
         save_managed_agents(&app, &records)?;
 
         let record = records

desktop/src-tauri/src/managed_agents/process_lifecycle.rs

@@ -134,6 +134,7 @@ pub fn finish_spawn(
     child: std::process::Child,
     log_path: std::path::PathBuf,
     spawn_config_hash: u64,
+    in_setup_mode: bool,
     agent_name: &str,
 ) -> super::ManagedAgentProcess {
     let job = create_job_for_child(child.id());
@@ -147,6 +148,7 @@ pub fn finish_spawn(
         child,
         log_path,
         spawn_config_hash,
+        in_setup_mode,
         job,
     }
 }

desktop/src-tauri/src/managed_agents/runtime.rs

@@ -5,6 +5,7 @@ use tauri::AppHandle;
 use super::agent_env::build_buzz_agent_provider_defaults;
 
 use crate::{
+    app_state::AppState,
     managed_agents::{
         append_log_marker, known_acp_runtime, login_shell_path, managed_agent_log_path,
         missing_command_message, normalize_agent_args, open_log_file, resolve_command,
@@ -1620,7 +1621,7 @@ pub fn spawn_agent_child(
     //
     // The JSON format mirrors `setup_mode::SetupPayload` in buzz-acp:
     //   { "agent_name": "...", "agent_pubkey": "...", "requirements": [{ "surface": "...", ... }] }
-    {
+    let in_setup_mode = {
         use crate::managed_agents::{
             agent_readiness, resolve_effective_agent_env, AgentReadiness, Requirement,
         };
@@ -1682,6 +1683,7 @@ pub fn spawn_agent_child(
         command.env_remove("BUZZ_ACP_SETUP_PAYLOAD");
 
         // Set the payload only when desktop computed NotReady.
+        let in_setup_mode = setup_payload_json.is_some();
         if let Some(json) = setup_payload_json {
             command.env("BUZZ_ACP_SETUP_PAYLOAD", json);
             eprintln!(
@@ -1689,7 +1691,8 @@ pub fn spawn_agent_child(
                 record.name
             );
         }
-    }
+        in_setup_mode
+    };
     // Only emit BUZZ_ACP_IDLE_TIMEOUT when the user has explicitly set an
     // override. When unset, the buzz-acp harness applies its own default
     // (see `DEFAULT_IDLE_TIMEOUT_SECS` in crates/buzz-acp/src/config.rs),
@@ -1890,6 +1893,7 @@ pub fn spawn_agent_child(
         child,
         log_path,
         spawn_config_hash,
+        in_setup_mode,
         &record.name,
     ));
     #[cfg(not(windows))]
@@ -1897,9 +1901,85 @@ pub fn spawn_agent_child(
         child,
         log_path,
         spawn_config_hash,
+        in_setup_mode,
     })
 }
 
+/// Returns `true` when a process that was spawned in setup mode would pass
+/// readiness with the record's current config. Setup-mode buzz-acp is a
+/// nudge-only listener and does not re-read desktop config after spawn, so this
+/// is the predicate for auto-restarting after an Edit Agent save.
+pub fn setup_mode_now_ready(
+    record: &ManagedAgentRecord,
+    personas: &[crate::managed_agents::types::PersonaRecord],
+    was_spawned_in_setup_mode: bool,
+) -> bool {
+    use crate::managed_agents::{agent_readiness, resolve_effective_agent_env, AgentReadiness};
+
+    if !was_spawned_in_setup_mode {
+        return false;
+    }
+
+    let effective_command = crate::managed_agents::effective_agent_command(
+        record.persona_id.as_deref(),
+        personas,
+        record.agent_command_override.as_deref(),
+    );
+    let runtime_meta = known_acp_runtime(&effective_command);
+    let effective = resolve_effective_agent_env(record, personas, runtime_meta);
+    matches!(agent_readiness(&effective), AgentReadiness::Ready)
+}
+
+/// Stop and respawn a setup-mode process if the current record now passes
+/// readiness. Called after Edit Agent saves mutate the record but before the
+/// store is persisted.
+pub fn maybe_restart_setup_mode_agent(
+    app: &AppHandle,
+    state: &AppState,
+    record: &mut ManagedAgentRecord,
+    runtimes: &mut HashMap<String, ManagedAgentProcess>,
+) {
+    let was_setup_mode = runtimes
+        .get(&record.pubkey)
+        .is_some_and(|runtime| runtime.in_setup_mode);
+    let personas = super::load_personas(app).unwrap_or_default();
+    if !setup_mode_now_ready(record, &personas, was_setup_mode) {
+        return;
+    }
+
+    let restart = (|| -> Result<(), String> {
+        let owner_hex = {
+            let keys = state.keys.lock().map_err(|e| e.to_string())?;
+            keys.public_key().to_hex()
+        };
+        stop_managed_agent_process(app, record, runtimes)?;
+        let process = spawn_agent_child(app, record, Some(&owner_hex))?;
+        let now = now_iso();
+        record.updated_at = now.clone();
+        record.runtime_pid = Some(process.child.id());
+        record.last_started_at = Some(now);
+        record.last_stopped_at = None;
+        record.last_exit_code = None;
+        record.last_error = None;
+        runtimes.insert(record.pubkey.clone(), process);
+        Ok(())
+    })();
+
+    match restart {
+        Ok(()) => eprintln!(
+            "buzz-desktop: auto-restarted agent {} out of setup mode",
+            record.name
+        ),
+        Err(error) => {
+            eprintln!(
+                "buzz-desktop: failed to auto-restart agent {} out of setup mode: {error}",
+                record.name
+            );
+            record.last_error = Some(format!("auto-restart failed: {error}"));
+        }
+    }
+}
+
 fn child_rust_log_filter() -> String {
     match std::env::var("RUST_LOG") {
         Ok(existing) if existing.contains("buzz_acp") => existing,

desktop/src-tauri/src/managed_agents/runtime/tests.rs

@@ -434,7 +434,47 @@ fn non_persona_agent_never_drifts() {
     assert!(!orphaned);
 }
 
-use super::runtime_metadata_env_vars;
+use super::{runtime_metadata_env_vars, setup_mode_now_ready};
+
+#[test]
+fn setup_mode_now_ready_requires_setup_mode_stamp() {
+    let mut record = fixture(RespondTo::Anyone, vec![], Some("tag".into()));
+    record.agent_command_override = Some("buzz-agent".into());
+    record.provider = Some("databricks_v2".into());
+    record.model = Some("goose-claude-4-6-opus".into());
+    record.env_vars.insert(
+        "DATABRICKS_HOST".into(),
+        "https://databricks.example.com".into(),
+    );
+
+    assert!(setup_mode_now_ready(&record, &[], true));
+    assert!(
+        !setup_mode_now_ready(&record, &[], false),
+        "healthy running agents must not be auto-restarted just because they are ready"
+    );
+}
+
+#[test]
+fn setup_mode_now_ready_stays_false_until_missing_config_is_fixed() {
+    let mut record = fixture(RespondTo::Anyone, vec![], Some("tag".into()));
+    record.agent_command_override = Some("buzz-agent".into());
+    record.provider = Some("databricks_v2".into());
+    record.env_vars.insert(
+        "DATABRICKS_HOST".into(),
+        "https://databricks.example.com".into(),
+    );
+
+    assert!(
+        !setup_mode_now_ready(&record, &[], true),
+        "missing model keeps the setup listener in place"
+    );
+
+    record.model = Some("goose-claude-4-6-opus".into());
+    assert!(
+        setup_mode_now_ready(&record, &[], true),
+        "filling the missing model makes the setup-mode process eligible for restart"
+    );
+}
 
 #[test]
 fn runtime_metadata_env_vars_injects_model_and_provider() {

desktop/src-tauri/src/managed_agents/types.rs

@@ -247,6 +247,13 @@ pub struct ManagedAgentProcess {
     /// `runtime_pid` have no `ManagedAgentProcess` entry, so their spawn
     /// config is unknown and the badge stays off.
     pub spawn_config_hash: u64,
+    /// `true` when this process was spawned with `BUZZ_ACP_SETUP_PAYLOAD` set —
+    /// readiness failed at spawn and buzz-acp is running as a nudge-only setup
+    /// listener instead of the real agent pool. Runtime-only — never persisted.
+    /// `update_managed_agent` uses this stamp to restart the harness once an
+    /// edit satisfies readiness, because setup-mode processes do not re-read
+    /// desktop config after launch.
+    pub in_setup_mode: bool,
     /// Win32 Job Object owning the harness + its entire process tree. Closing
     /// the handle (via `JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE`) kills the whole
     /// tree — the Windows mirror of the Unix process-group teardown. `None`