diff --git a/docs/guides/guard.md b/docs/guides/guard.md
index 25d51cd..d24df9d 100644
--- a/docs/guides/guard.md
+++ b/docs/guides/guard.md
@@ -9,7 +9,9 @@ Without an explicit `guard` field, eval-magic detects packaged profiles from the
 
 ## Understand the two policy layers
 
-Containment checks run before command allowances. A command policy cannot override these checks:
+Containment checks run before command allowances. A command policy cannot override these checks.
+A command that violates both layers is denied once, with a verdict that names each blocking reason,
+so the denial never points at a fix that cannot unblock the command:
 
 - Direct write and patch tools must target the task environment.
 - Shell redirects and `tee` targets must resolve inside the task environment.
@@ -113,8 +115,8 @@ The packaged profiles are:
 - `language/rust` is detected from `Cargo.toml`. It allows `cargo build`, `check`, `test`, `run`,
   `fmt`, and `clippy`.
 - `language/javascript` is detected from `package.json`. It allows npm install, CI, test, build,
-  lint, and typecheck commands, plus corresponding pnpm, Yarn, and Bun install, add, test, build,
-  lint, and typecheck commands.
+  dev, start, lint, and typecheck commands, plus corresponding pnpm, Yarn, and Bun install, add,
+  test, build, dev, start, lint, and typecheck commands.
 - `framework/nextjs` is detected when `package.json` declares a `next` dependency. It allows npm,
   pnpm, Yarn, and Bun dev, build, and start scripts, plus direct `next` invocations through `npx`,
   `pnpm exec`, `yarn`, and `bunx`.
diff --git a/guard-profiles/language-javascript.toml b/guard-profiles/language-javascript.toml
index a05f013..5e6f285 100644
--- a/guard-profiles/language-javascript.toml
+++ b/guard-profiles/language-javascript.toml
@@ -5,24 +5,32 @@ allow_commands = [
   "npm ci",
   "npm test",
   "npm run build",
+  "npm run dev",
   "npm run lint",
+  "npm run start",
   "npm run typecheck",
   "pnpm install",
   "pnpm add",
   "pnpm test",
   "pnpm run build",
+  "pnpm run dev",
   "pnpm run lint",
+  "pnpm run start",
   "pnpm run typecheck",
   "yarn install",
   "yarn add",
   "yarn test",
   "yarn run build",
+  "yarn run dev",
   "yarn run lint",
+  "yarn run start",
   "yarn run typecheck",
   "bun install",
   "bun add",
   "bun test",
   "bun run build",
+  "bun run dev",
   "bun run lint",
+  "bun run start",
   "bun run typecheck",
 ]
diff --git a/src/adapters/guard.rs b/src/adapters/guard.rs
index 6d8f9f2..92b51aa 100644
--- a/src/adapters/guard.rs
+++ b/src/adapters/guard.rs
@@ -650,7 +650,7 @@ mod tests {
         assert_eq!(
             verdict("codex", payload, Some(marker())).expect("should block"),
             "{\"decision\":\"block\",\"reason\":\"eval guard: blocked Bash \
-             (package install/add) — runs outside the eval sandbox\"}"
+             (package install/add — runs outside the eval sandbox; command not allowed by eval guard policy)\"}"
         );
     }
 
diff --git a/src/sandbox/decide.rs b/src/sandbox/decide.rs
index 59b9f52..016dced 100644
--- a/src/sandbox/decide.rs
+++ b/src/sandbox/decide.rs
@@ -17,7 +17,7 @@ use crate::core::fs::artifact_path;
 
 use super::command_policy::COMMAND_POLICY_REASON;
 use super::policy::{
-    OUTPUT_REDIRECTION_REASON, apply_patch_paths, classify_bash_with_policy, is_patch_tool,
+    OUTPUT_REDIRECTION_REASON, apply_patch_paths, classify_bash_denials, is_patch_tool,
     is_shell_tool, is_under_any, is_write_tool, path_arg, resolve_path,
 };
 
@@ -220,25 +220,49 @@ pub(crate) fn decide_with_cwd(
             .get("command")
             .and_then(Value::as_str)
             .unwrap_or("");
-        if let Some(classification) =
-            classify_bash_with_policy(command, &roots, invocation_cwd, guard_policy)
-        {
-            let hint = if classification.reason == OUTPUT_REDIRECTION_REASON {
-                scratch_hint(&roots)
+        let denials = classify_bash_denials(command, &roots, invocation_cwd, guard_policy);
+        if !denials.is_empty() {
+            // A containment denial stays a self-contained clause when the
+            // command policy also denies, so one verdict can name every
+            // blocking layer instead of sending the agent to fix a problem
+            // that cannot unblock the command.
+            let clause = |reason: &str| {
+                if reason == COMMAND_POLICY_REASON {
+                    reason.to_string()
+                } else {
+                    format!("{reason} — runs outside the eval sandbox")
+                }
+            };
+            let verdict = if let [only] = denials.as_slice() {
+                let boundary = if only.reason == COMMAND_POLICY_REASON {
+                    ""
+                } else {
+                    " — runs outside the eval sandbox"
+                };
+                format!("({}){boundary}", only.reason)
             } else {
-                String::new()
+                let clauses = denials
+                    .iter()
+                    .map(|denial| clause(denial.reason))
+                    .collect::<Vec<_>>()
+                    .join("; ");
+                format!("({clauses})")
             };
-            let boundary = if classification.reason == COMMAND_POLICY_REASON {
-                ""
+            let hint = if denials
+                .iter()
+                .any(|denial| denial.reason == OUTPUT_REDIRECTION_REASON)
+            {
+                scratch_hint(&roots)
             } else {
-                " — runs outside the eval sandbox"
+                String::new()
             };
+            let resolved_targets = denials
+                .into_iter()
+                .flat_map(|denial| denial.resolved_targets)
+                .collect();
             return GuardEvaluation::deny(
-                format!(
-                    "{GUARD_REASON_PREFIX}blocked {tool_name} ({}){boundary}{hint}",
-                    classification.reason,
-                ),
-                classification.resolved_targets,
+                format!("{GUARD_REASON_PREFIX}blocked {tool_name} {verdict}{hint}"),
+                resolved_targets,
             );
         }
     }
@@ -407,6 +431,62 @@ mod tests {
         );
     }
 
+    /// Issue #297: a command both layers would deny must get one verdict that
+    /// names both reasons. Naming only the redirect (with its actionable
+    /// scratch hint) sends the agent to fix a problem that cannot unblock the
+    /// command, because the command policy was already denying it.
+    #[test]
+    fn a_bash_denial_names_every_blocking_layer_in_one_verdict() {
+        let marker: GuardMarker = serde_json::from_value(json!({
+            "active": true,
+            "allowedRoots": ["/work/.eval-magic/task"],
+            "guardPolicy": { "allow_tools": ["cargo"] }
+        }))
+        .unwrap();
+
+        let denied = decide_with_cwd(
+            "Bash",
+            &json!({ "command": "npm run dev > /tmp/dev-server.log 2>&1 &" }),
+            Some(&marker),
+            now_ms(),
+            Path::new("/work/.eval-magic/task"),
+        );
+
+        assert!(!denied.decision.allow);
+        let reason = denied.decision.reason.unwrap();
+        assert!(reason.contains("output redirection to a file"), "{reason}");
+        assert!(
+            reason.contains("command not allowed by eval guard policy"),
+            "{reason}"
+        );
+        assert!(
+            reason.ends_with("For temporary or scratch files, use /work/.eval-magic/task/tmp."),
+            "{reason}"
+        );
+        assert_eq!(
+            denied.resolved_targets,
+            vec!["/tmp/dev-server.log".to_string()]
+        );
+    }
+
+    #[test]
+    fn a_bash_denial_from_containment_alone_keeps_the_single_reason_verdict() {
+        let d = decide_now(
+            "Bash",
+            json!({ "command": "echo hi > /tmp/out.log" }),
+            Some(&marker()),
+        );
+
+        assert_eq!(
+            d.reason.as_deref(),
+            Some(
+                "eval guard: blocked Bash (output redirection to a file) \
+                 — runs outside the eval sandbox. For temporary or scratch files, use \
+                 /work/.eval-magic/tmp."
+            )
+        );
+    }
+
     #[test]
     fn allows_bash_with_an_in_bounds_redirect() {
         let d = decide_now(
diff --git a/src/sandbox/guard_profiles.rs b/src/sandbox/guard_profiles.rs
index e805c36..e6dae2d 100644
--- a/src/sandbox/guard_profiles.rs
+++ b/src/sandbox/guard_profiles.rs
@@ -132,6 +132,36 @@ mod tests {
         assert!(!expanded.allow_commands.contains(&"npm test".to_string()));
     }
 
+    /// Issue #297: a plain package.json project (the pinned Weeknight fixture
+    /// is Vite + React) must be able to start its own dev server. `dev` and
+    /// `start` are generic lifecycle script names, not Next.js-specific, so
+    /// they belong to the language profile.
+    #[test]
+    fn language_javascript_allows_dev_and_start_scripts() {
+        let policy = GuardPolicyConfig {
+            profiles: vec!["language/javascript".to_string()],
+            ..GuardPolicyConfig::default()
+        };
+
+        let expanded = expand_policy(&policy).unwrap();
+
+        for command in [
+            "npm run dev",
+            "npm run start",
+            "pnpm run dev",
+            "pnpm run start",
+            "yarn run dev",
+            "yarn run start",
+            "bun run dev",
+            "bun run start",
+        ] {
+            assert!(
+                expanded.allow_commands.contains(&command.to_string()),
+                "{command}"
+            );
+        }
+    }
+
     #[test]
     fn detection_joins_language_and_framework_profiles_recursively() {
         let root = tempdir().unwrap();
diff --git a/src/sandbox/policy.rs b/src/sandbox/policy.rs
index 4799c67..3978f22 100644
--- a/src/sandbox/policy.rs
+++ b/src/sandbox/policy.rs
@@ -198,18 +198,43 @@ pub(crate) fn classify_bash_with_cwd(
     )
 }
 
-/// Classify one shell tool call under its resolved eval command policy.
+/// Classify one shell tool call under its resolved eval command policy,
+/// returning the first applicable denial.
 pub(crate) fn classify_bash_with_policy(
     command: &str,
     allowed_roots: &[String],
     invocation_cwd: &Path,
     policy: &crate::core::GuardPolicyConfig,
 ) -> Option<BashClassification> {
+    classify_bash_denials(command, allowed_roots, invocation_cwd, policy)
+        .into_iter()
+        .next()
+}
+
+/// Every layer that denies one shell tool call, containment checks first and
+/// the eval command policy second (at most one denial each).
+///
+/// The stray-write audit wants only the first denial, but the live guard owes
+/// the agent all of them in one verdict: a verdict that names just the
+/// redirect it found sends the agent to fix a problem that cannot unblock the
+/// command when the command policy was already denying it.
+pub(crate) fn classify_bash_denials(
+    command: &str,
+    allowed_roots: &[String],
+    invocation_cwd: &Path,
+    policy: &crate::core::GuardPolicyConfig,
+) -> Vec<BashClassification> {
     if command.is_empty() {
-        return None;
+        return Vec::new();
+    }
+    let mut denials = Vec::new();
+    if let Some(denial) = classify_fixed_containment(command, allowed_roots, invocation_cwd) {
+        denials.push(denial);
+    }
+    if let Some(denial) = super::command_policy::classify_command_policy(command, policy) {
+        denials.push(denial);
     }
-    classify_fixed_containment(command, allowed_roots, invocation_cwd)
-        .or_else(|| super::command_policy::classify_command_policy(command, policy))
+    denials
 }
 
 fn classify_fixed_containment(
diff --git a/src/sandbox/policy/tests/command_policy.rs b/src/sandbox/policy/tests/command_policy.rs
index 9ece762..516218d 100644
--- a/src/sandbox/policy/tests/command_policy.rs
+++ b/src/sandbox/policy/tests/command_policy.rs
@@ -154,3 +154,78 @@ fn shell_wrapper_allowance_cannot_bypass_fixed_containment() {
         Some("cargo build/test output")
     );
 }
+
+#[test]
+fn denials_reports_every_denying_layer_containment_first() {
+    let roots = vec!["/work/env".to_string()];
+    let policy = crate::core::GuardPolicyConfig {
+        allow_commands: vec!["cargo build".to_string()],
+        ..crate::core::GuardPolicyConfig::default()
+    };
+
+    let denials = classify_bash_denials(
+        "npm run dev > /tmp/dev-server.log",
+        &roots,
+        Path::new("/work/env"),
+        &policy,
+    );
+
+    let reasons: Vec<&str> = denials.iter().map(|denial| denial.reason).collect();
+    assert_eq!(
+        reasons,
+        [
+            "output redirection to a file",
+            "command not allowed by eval guard policy"
+        ]
+    );
+    assert_eq!(
+        denials[0].resolved_targets,
+        vec!["/tmp/dev-server.log".to_string()]
+    );
+}
+
+#[test]
+fn denials_reports_a_single_layer_when_only_one_applies() {
+    let roots = vec!["/work/env".to_string()];
+
+    // Containment only: echo is unrecognized by the command policy.
+    let denials = classify_bash_denials(
+        "echo hi > /tmp/out.log",
+        &roots,
+        Path::new("/work/env"),
+        &crate::core::GuardPolicyConfig::default(),
+    );
+    assert_eq!(denials.len(), 1);
+    assert_eq!(denials[0].reason, "output redirection to a file");
+
+    // Policy only: the redirect target is in bounds, but npm is claimed and
+    // not allowed.
+    let policy = crate::core::GuardPolicyConfig {
+        allow_commands: vec!["cargo build".to_string()],
+        ..crate::core::GuardPolicyConfig::default()
+    };
+    let denials = classify_bash_denials(
+        "npm run dev > /work/env/dev-server.log",
+        &roots,
+        Path::new("/work/env"),
+        &policy,
+    );
+    assert_eq!(denials.len(), 1);
+    assert_eq!(
+        denials[0].reason,
+        "command not allowed by eval guard policy"
+    );
+}
+
+#[test]
+fn denials_is_empty_for_an_allowed_command() {
+    let roots = vec!["/work/env".to_string()];
+    let policy = crate::core::GuardPolicyConfig {
+        allow_commands: vec!["cargo build".to_string()],
+        ..crate::core::GuardPolicyConfig::default()
+    };
+
+    assert!(
+        classify_bash_denials("cargo build", &roots, Path::new("/work/env"), &policy).is_empty()
+    );
+}
diff --git a/tests/cli/guard.rs b/tests/cli/guard.rs
index b03fce0..79921f7 100644
--- a/tests/cli/guard.rs
+++ b/tests/cli/guard.rs
@@ -204,7 +204,7 @@ fn guard_codex_block_verdict_bytes_are_stable() {
         .success()
         .stdout(
             "{\"decision\":\"block\",\"reason\":\"eval guard: blocked Bash \
-             (package install/add) — runs outside the eval sandbox\"}",
+             (package install/add — runs outside the eval sandbox; command not allowed by eval guard policy)\"}",
         );
 }
 
diff --git a/tests/run/guard_policy.rs b/tests/run/guard_policy.rs
index e49e7b7..2a2b66b 100644
--- a/tests/run/guard_policy.rs
+++ b/tests/run/guard_policy.rs
@@ -105,3 +105,50 @@ fn per_eval_guard_replaces_the_default_and_disables_detection() {
         &serde_json::json!({ "allow_commands": ["npm run dev"] })
     );
 }
+
+/// Issue #297: the pinned default fixture (Vite + React, no `next` dependency)
+/// must detect `language/javascript` alone, and the frozen policy must allow
+/// the lifecycle scripts a browser check needs.
+#[test]
+fn a_plain_package_json_detects_language_javascript_with_dev_and_start() {
+    let tmp = tempfile::TempDir::new().unwrap();
+    let evals = r#"{
+      "skill_name": "mr-review",
+      "evals": [{
+        "id": "e1",
+        "prompt": "build the app",
+        "expected_output": "built",
+        "files": ["package.json"]
+      }]
+    }"#;
+    let (skill_dir, cwd) = setup(tmp.path(), evals);
+    fs::write(
+        skill_dir.join("mr-review/evals/package.json"),
+        r#"{"dependencies":{"vite":"5.0.0","react":"18.0.0"}}"#,
+    )
+    .unwrap();
+
+    skill_eval()
+        .current_dir(&cwd)
+        .args(["run", "--skill-dir"])
+        .arg(&skill_dir)
+        .args(["--skill", "mr-review", "--mode", "new-skill", "--guard"])
+        .assert()
+        .success();
+
+    let dispatch = read_json(&iteration_dir(&cwd).join("dispatch.json"));
+    let policy = &dispatch["tasks"][0]["guard_policy"];
+    assert_eq!(
+        policy["profiles"],
+        serde_json::json!(["language/javascript"])
+    );
+    let commands: Vec<&str> = policy["allow_commands"]
+        .as_array()
+        .unwrap()
+        .iter()
+        .filter_map(|command| command.as_str())
+        .collect();
+    assert!(commands.contains(&"npm run dev"), "{commands:?}");
+    assert!(commands.contains(&"npm run start"), "{commands:?}");
+    assert!(!commands.contains(&"npx next dev"), "{commands:?}");
+}