fix(brain): attach the default safety audit log

README.md

@@ -314,7 +314,7 @@ milestone.
 | **Every Ubuntu LTS validated** — 22.04, 24.04 and 26.04 all at 79/79, each with a replay twin that reproduces it | ✅ |
 | Telegram approval interface | 📋 roadmap |
 
-**1,845 Rust tests and 72 frontend tests** form the current deterministic
+**1,846 Rust tests and 72 frontend tests** form the current deterministic
 release baseline.
 
 ## Configure your LLM

crates/sysknife-brain/src/planner.rs

@@ -754,6 +754,10 @@ impl LlmPlanner {
     /// HTTP client cannot be initialised (rare; only fails if the TLS
     /// subsystem is unavailable).
     ///
+    /// Safety-fence rejection logging is enabled by default and writes to
+    /// [`SafetyAuditLog::default_path`]. Use [`Self::new`] for a planner
+    /// without runtime defaults.
+    ///
     /// Rate limiting is **enabled by default** at [`DEFAULT_MAX_RPM`] requests
     /// per minute. Override with the `SYSKNIFE_MAX_RPM` environment variable.
     /// Call `with_rate_limiter` after this to replace the default limiter, or
@@ -859,9 +863,13 @@ impl LlmPlanner {
             None => provider,
         };
 
-        let mut planner = Self::new(provider, state_client, config.max_turns);
+        let mut planner = Self::new(provider, state_client, config.max_turns)
+            .with_audit_log(SafetyAuditLog::new(SafetyAuditLog::default_path()));
+        // `from_config` is the production construction path used by the CLI,
+        // MCP server, and shell. Attach the default safety log here so every
+        // runtime planner records fence rejections; direct `new` remains
+        // opt-in for callers such as tests and embedded consumers.
         planner.prefs_path = Some(sysknife_core::config::prefs_path());
-
         if replaying {
             // No rate limiter under replay. It exists to bound spend and load on a
             // provider, and a replay reaches neither: every answer comes off disk.
@@ -1516,6 +1524,18 @@ mod tests {
             .expect("ollama defaults need no credentials")
     }
 
+    /// The integration test exercises the rejection path and reads the JSONL;
+    /// this test pins that the production constructor actually attaches the
+    /// log without mutating process-global path configuration.
+    #[test]
+    fn configured_planner_attaches_audit_log() {
+        let planner = ollama_planner();
+        assert!(
+            planner.audit_log.is_some(),
+            "from_config must attach the safety audit log"
+        );
+    }
+
     #[test]
     fn into_authorized_replaces_every_step_risk() {
         let step = |name: &str, risk| {

docs/distro-support.md

@@ -82,7 +82,7 @@ family and the atomic story family are implemented and covered by the workspace
 suite. What is missing is a way to put the helpers somewhere the daemon's own
 grants already point.
 
-The deterministic workspace baseline is 1,845 Rust tests plus 72 frontend
+The deterministic workspace baseline is 1,846 Rust tests plus 72 frontend
 tests. Those tests verify action construction, policy, approval, storage, and
 UI behavior, but they do not replace a real distribution VM run.
 

docs/introduction.md

@@ -141,7 +141,7 @@ flow.
 
 ## Status
 
-190 typed actions · 1,845 Rust tests + 72 frontend tests · MIT
+190 typed actions · 1,846 Rust tests + 72 frontend tests · MIT
 
 SysKnife is the reference implementation of the
 [LACS specification](https://github.com/lacs-project/specification) — a

tests/evidence/workspace-tests.json

@@ -4,6 +4,6 @@
     "tests": "cargo nextest run --workspace --locked"
   },
   "frontend_tests": 72,
-  "tests": 1845,
+  "tests": 1846,
   "version": 2
 }