feat(goal): /config skeptic-local — auto-managed local model for the skeptic

crates/hi-agent/src/agent/lifecycle.rs

@@ -108,18 +108,13 @@ impl crate::Agent {
             model: config.model.clone(),
         };
         // Opt-in: route the skeptic review to a separate OpenAI-compatible
-        // endpoint (e.g. a local hi-local server) when configured.
-        let skeptic_provider: Option<Arc<dyn Provider>> =
-            config.skeptic_endpoint.as_deref().map(|url| {
-                let key = config
-                    .skeptic_endpoint_key
-                    .clone()
-                    .unwrap_or_else(|| "local".to_string());
-                Arc::new(hi_ai::OpenAiProvider::new(url.to_string(), key)) as Arc<dyn Provider>
-            });
+        // endpoint (e.g. a local hi-local server) when configured. Shared with
+        // the runtime `/config skeptic-local` toggle so their wiring can't drift.
+        let skeptic_provider = crate::local_skeptic::build_skeptic_provider(&config);
         Ok(Self {
             provider,
             skeptic_provider,
+            local_skeptic: None,
             config,
             runtime,
             task_context: None,
@@ -733,9 +728,11 @@ impl crate::Agent {
         self.runtime.background().kill_started_after(before)
     }
 
-    /// Stop every background process owned by this agent runtime.
+    /// Stop every background process owned by this agent runtime, plus any
+    /// auto-managed local skeptic server, on session shutdown.
     pub fn kill_background_processes(&self) {
         self.runtime.background().kill_all();
+        self.stop_local_skeptic_server();
     }
 
     /// Finalize a turn whose future was cancelled by its frontend. Reconcile

crates/hi-agent/src/command.rs

@@ -605,6 +605,11 @@ pub enum ConfigArg {
     /// `On` forces streaming, `Off` forces resident, `Auto` (the default) lets
     /// the loader auto-enable when the model exceeds the memory budget.
     MoeStreaming(MoeStreamingMode),
+    /// `/config skeptic-local <on|off>` — turn the auto-managed local model for
+    /// the `/goal` skeptic review on or off. `on` detects the machine's backend,
+    /// downloads a small default review model if needed, and spawns a local
+    /// server; `off` stops it and restores the prior skeptic settings.
+    SkepticLocal(bool),
     /// Unrecognized option or bad value; carries a usage/error hint.
     Invalid(String),
 }
@@ -720,12 +725,27 @@ pub fn parse_config_arg(arg: &str) -> ConfigArg {
                 }
             }
         }
+        "skeptic-local" | "local-skeptic" => match val.to_ascii_lowercase().as_str() {
+            "" => ConfigArg::Invalid("usage: /config skeptic-local <on|off>".into()),
+            "on" | "enable" | "enabled" | "1" | "true" | "yes" => ConfigArg::SkepticLocal(true),
+            "off" | "disable" | "disabled" | "0" | "false" | "no" => ConfigArg::SkepticLocal(false),
+            _ => ConfigArg::Invalid(format!(
+                "unknown skeptic-local mode '{val}' — use on or off"
+            )),
+        },
         other => ConfigArg::Invalid(format!(
-            "unknown /config option '{other}' — try: show, reasoning <level>, temp <value>, steps <n|auto|off>, moe-streaming <on|off|auto>"
+            "unknown /config option '{other}' — try: show, reasoning <level>, temp <value>, steps <n|auto|off>, moe-streaming <on|off|auto>, skeptic-local <on|off>"
         )),
     }
 }
 
+/// Whether a `/config` argument is the async `skeptic-local` toggle. The CLI
+/// routes this through its async handler (it may download a model and spawn a
+/// server) rather than the synchronous `/config` path.
+pub fn config_is_skeptic_local(arg: &str) -> bool {
+    matches!(parse_config_arg(arg), ConfigArg::SkepticLocal(_))
+}
+
 fn read_only_macro_prompt(kind: &str, topic: &str) -> String {
     let topic = topic.trim();
     let topic = if topic.is_empty() {

crates/hi-agent/src/lib.rs

@@ -10,6 +10,7 @@ mod context_index;
 mod decision;
 mod goal;
 mod heuristics;
+pub mod local_skeptic;
 mod memory;
 mod outcome;
 mod prompt;
@@ -37,6 +38,7 @@ pub use config::{
 };
 pub use heuristics::humanize_count;
 pub use hi_tools::{PlanStatus, PlanStep};
+pub use local_skeptic::LocalSkepticOutcome;
 pub use memory::{
     AnnotatedBullet, global_memory_file, memory_file, read_global_memory, read_memory,
     read_project_annotated, should_distill_memory,
@@ -557,6 +559,11 @@ pub struct Agent {
     /// as it always has. Lets the frequent, fail-open review loop run on a local
     /// model while the driver stays on the session model.
     pub(crate) skeptic_provider: Option<Arc<dyn Provider>>,
+    /// Session state for an auto-managed local skeptic server started by
+    /// `/config skeptic-local on` (`None` when off). Held so the server can be
+    /// stopped, the prior skeptic settings restored, and the process killed on
+    /// session shutdown.
+    pub(crate) local_skeptic: Option<crate::local_skeptic::LocalSkepticState>,
     pub(crate) config: AgentConfig,
     pub(crate) runtime: WorkspaceRuntime,
     /// Per-turn ranked repository data and scoped instructions.