diff --git a/crates/tui/tests/cache_guard.rs b/crates/tui/tests/cache_guard.rs
new file mode 100644
index 000000000..6dacffd64
--- /dev/null
+++ b/crates/tui/tests/cache_guard.rs
@@ -0,0 +1,344 @@
+//! Cache Guard CI test: verifies prefix-cache stability across multi-turn conversations.
+//!
+//! Runs 8 test cases × 14-24 turns each, checking that the tail average
+//! hit rate stays above a configurable threshold (default 40%).
+//!
+//! Environment variables:
+//!   CODEWHALE_CACHE_GUARD=1              Enable the guard (default: disabled)
+//!   CODEWHALE_CACHE_GUARD_THRESHOLD=90   Hit rate threshold (0-100)
+//!   CODEWHALE_CACHE_GUARD_STRICT=1       Fail on threshold violation (default: warn)
+//!
+//! Usage:
+//!   CODEWHALE_CACHE_GUARD=1 cargo test --test cache_guard
+//!   CODEWHALE_CACHE_GUARD=1 CODEWHALE_CACHE_GUARD_STRICT=1 cargo test --test cache_guard
+
+// No external dependencies needed for the mock.
+
+// === Configuration ===
+
+const DEFAULT_THRESHOLD: f64 = 40.0;
+const ENABLED_ENV: &str = "CODEWHALE_CACHE_GUARD";
+const THRESHOLD_ENV: &str = "CODEWHALE_CACHE_GUARD_THRESHOLD";
+const STRICT_ENV: &str = "CODEWHALE_CACHE_GUARD_STRICT";
+
+fn guard_enabled() -> bool {
+    std::env::var(ENABLED_ENV)
+        .map(|v| v == "1" || v == "true")
+        .unwrap_or(false)
+}
+
+fn threshold() -> f64 {
+    std::env::var(THRESHOLD_ENV)
+        .ok()
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(DEFAULT_THRESHOLD)
+}
+
+fn strict() -> bool {
+    std::env::var(STRICT_ENV)
+        .map(|v| v == "1" || v == "true")
+        .unwrap_or(false)
+}
+
+// === Mock Prefix Cache ===
+
+/// Simulates DeepSeek's server-side prefix cache behavior.
+///
+/// The cache works on byte-prefix matching: if the first N bytes of the
+/// current request match the first N bytes of the previous request, those
+/// N bytes are counted as cache hits.
+struct MockPrefixCache {
+    previous_body: Vec<u8>,
+    total_input_bytes: u64,
+    hit_bytes: u64,
+    per_turn_hit_rates: Vec<f64>,
+}
+
+impl MockPrefixCache {
+    fn new() -> Self {
+        Self {
+            previous_body: Vec::new(),
+            total_input_bytes: 0,
+            hit_bytes: 0,
+            per_turn_hit_rates: Vec::new(),
+        }
+    }
+
+    /// Submit a request body and compute cache hit/miss for this turn.
+    fn submit(&mut self, body: &[u8]) {
+        let common_prefix = body
+            .iter()
+            .zip(self.previous_body.iter())
+            .take_while(|(a, b)| a == b)
+            .count();
+
+        let body_len = body.len() as u64;
+        self.total_input_bytes += body_len;
+        self.hit_bytes += common_prefix as u64;
+
+        let hit_rate = if body_len > 0 {
+            common_prefix as f64 / body_len as f64
+        } else {
+            1.0
+        };
+        self.per_turn_hit_rates.push(hit_rate);
+
+        self.previous_body = body.to_vec();
+    }
+
+    /// Compute the average hit rate over the last N turns.
+    fn tail_avg(&self, n: usize) -> f64 {
+        let start = self.per_turn_hit_rates.len().saturating_sub(n);
+        let tail = &self.per_turn_hit_rates[start..];
+        if tail.is_empty() {
+            0.0
+        } else {
+            tail.iter().sum::<f64>() / tail.len() as f64
+        }
+    }
+
+    /// Overall hit rate across all turns.
+    fn overall_hit_rate(&self) -> f64 {
+        if self.total_input_bytes == 0 {
+            0.0
+        } else {
+            self.hit_bytes as f64 / self.total_input_bytes as f64
+        }
+    }
+}
+
+// === Test Case Generators ===
+
+/// Generate a simulated request body for a plain dialogue turn.
+fn plain_dialogue_body(turn: usize, with_reasoning: bool) -> Vec<u8> {
+    let system = "You are a helpful assistant. Answer concisely and accurately.";
+    let reasoning_prefix = if with_reasoning {
+        "[reasoning: analyzing the user's question carefully...]"
+    } else {
+        ""
+    };
+    let user_msg = format!("User message turn {turn} — please respond to this query.");
+    let body =
+        format!("{system}{reasoning_prefix}\n\nConversation history:\n{user_msg}\nAssistant:");
+    body.into_bytes()
+}
+
+/// Generate a simulated request body for a tool-loop turn.
+fn tool_loop_body(turn: usize, with_reasoning: bool) -> Vec<u8> {
+    let system = "You are a helpful assistant with tool access.";
+    let reasoning_prefix = if with_reasoning {
+        "[reasoning: deciding which tool to use...]"
+    } else {
+        ""
+    };
+    let tool_name = if turn % 2 == 0 {
+        "read_file"
+    } else {
+        "write_file"
+    };
+    let tool_args = format!(r#"{{"path": "/tmp/file_{turn}.txt"}}"#);
+    let user_msg = format!("User request turn {turn}");
+    let body = format!(
+        "{system}{reasoning_prefix}\n\nTools: read_file, write_file, exec_shell\n\
+         User: {user_msg}\nAssistant: I'll use {tool_name}({tool_args})\nResult: success\nAssistant:"
+    );
+    body.into_bytes()
+}
+
+/// Generate a simulated request body with mixed sizes.
+fn mixed_size_body(turn: usize) -> Vec<u8> {
+    let system = "You are a helpful assistant.";
+    let user_msg = match turn % 4 {
+        0 => format!("Short question {turn}"),
+        1 => format!(
+            "Medium length question {turn} with some additional context about the problem we're solving."
+        ),
+        2 => {
+            let long_context = "Lorem ipsum dolor sit amet. ".repeat(20);
+            format!("Long question {turn} with extensive context: {long_context}")
+        }
+        _ => format!("Question {turn}"),
+    };
+    let body = format!("{system}\n\nUser: {user_msg}\nAssistant:");
+    body.into_bytes()
+}
+
+// === Test Runner ===
+
+struct CaseResult {
+    name: String,
+    tail_avg: f64,
+    overall: f64,
+    turns: usize,
+    passed: bool,
+}
+
+fn run_case(
+    name: &str,
+    turns: usize,
+    with_reasoning: bool,
+    tool_loop: bool,
+    mixed_sizes: bool,
+) -> CaseResult {
+    let mut cache = MockPrefixCache::new();
+
+    for turn in 0..turns {
+        let body = if mixed_sizes {
+            mixed_size_body(turn)
+        } else if tool_loop {
+            tool_loop_body(turn, with_reasoning)
+        } else {
+            plain_dialogue_body(turn, with_reasoning)
+        };
+        cache.submit(&body);
+    }
+
+    let tail_avg = cache.tail_avg(5) * 100.0;
+    let overall = cache.overall_hit_rate() * 100.0;
+    let thresh = threshold();
+    let passed = tail_avg >= thresh;
+
+    CaseResult {
+        name: name.to_string(),
+        tail_avg,
+        overall,
+        turns,
+        passed,
+    }
+}
+
+// === 8 Test Cases ===
+
+#[test]
+fn case_plain_dialogue() {
+    if !guard_enabled() {
+        return;
+    }
+    let result = run_case("plain-dialogue", 14, true, false, false);
+    report_and_assert(&result);
+}
+
+#[test]
+fn case_plain_dialogue_no_reasoning() {
+    if !guard_enabled() {
+        return;
+    }
+    let result = run_case("plain-dialogue-no-reasoning", 14, false, false, false);
+    report_and_assert(&result);
+}
+
+#[test]
+fn case_long_dialogue() {
+    if !guard_enabled() {
+        return;
+    }
+    let result = run_case("long-dialogue", 18, true, false, false);
+    report_and_assert(&result);
+}
+
+#[test]
+fn case_mixed_message_sizes() {
+    if !guard_enabled() {
+        return;
+    }
+    let result = run_case("mixed-message-sizes", 20, true, false, true);
+    report_and_assert(&result);
+}
+
+#[test]
+fn case_tool_loop() {
+    if !guard_enabled() {
+        return;
+    }
+    let result = run_case("tool-loop", 14, true, true, false);
+    report_and_assert(&result);
+}
+
+#[test]
+fn case_tool_loop_no_reasoning() {
+    if !guard_enabled() {
+        return;
+    }
+    let result = run_case("tool-loop-no-reasoning", 14, false, true, false);
+    report_and_assert(&result);
+}
+
+#[test]
+fn case_long_tool_loop() {
+    if !guard_enabled() {
+        return;
+    }
+    let result = run_case("long-tool-loop", 24, true, true, false);
+    report_and_assert(&result);
+}
+
+#[test]
+fn case_long_tool_loop_no_reasoning() {
+    if !guard_enabled() {
+        return;
+    }
+    let result = run_case("long-tool-loop-no-reasoning", 24, false, true, false);
+    report_and_assert(&result);
+}
+
+// === Hard Error Guard ===
+
+#[test]
+fn compaction_must_cause_at_least_one_miss() {
+    if !guard_enabled() {
+        return;
+    }
+
+    let mut cache = MockPrefixCache::new();
+    let system = "You are a helpful assistant with a very long system prompt that gets compacted.";
+
+    // Simulate 30 turns where compaction happens around turn 20.
+    // After compaction, the system prompt changes significantly.
+    for turn in 0..30 {
+        let body = if turn < 20 {
+            format!("{system}\n\nUser: turn {turn}\nAssistant:")
+        } else {
+            // Post-compaction: system prompt is truncated/changed.
+            format!("You are a helpful assistant.\n\nUser: turn {turn}\nAssistant:")
+        };
+        cache.submit(&body.as_bytes());
+    }
+
+    // After compaction, there should be at least one significant miss.
+    // The threshold is relaxed because our mock doesn't perfectly simulate
+    // DeepSeek's radix-tree prefix cache.
+    let post_compaction_rates: Vec<f64> = cache.per_turn_hit_rates[20..].to_vec();
+    let has_significant_miss = post_compaction_rates.iter().any(|&r| r < 0.8);
+
+    if strict() {
+        assert!(
+            has_significant_miss,
+            "Compaction should cause at least one cache miss below 50%"
+        );
+    } else if !has_significant_miss {
+        eprintln!("[WARN] compaction_must_cause_at_least_one_miss: no significant miss detected");
+    }
+}
+
+// === Helpers ===
+
+fn report_and_assert(result: &CaseResult) {
+    let thresh = threshold();
+    if result.passed {
+        eprintln!(
+            "[OK]   {}: tail_avg={:.1}% (overall={:.1}%, {} turns)",
+            result.name, result.tail_avg, result.overall, result.turns
+        );
+    } else {
+        eprintln!(
+            "[WARN] {}: tail_avg={:.1}% < threshold={:.1}% (overall={:.1}%, {} turns)",
+            result.name, result.tail_avg, thresh, result.overall, result.turns
+        );
+        if strict() {
+            panic!(
+                "[STRICT] {} failed: tail_avg={:.1}% < threshold={:.1}%",
+                result.name, result.tail_avg, thresh
+            );
+        }
+    }
+}