diff --git a/cli.py b/cli.py
index c308b5608e..f2502538e4 100644
--- a/cli.py
+++ b/cli.py
@@ -6546,16 +6546,17 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
                 base_read = cur_read
             delta_prompt = cur_prompt - base_prompt
             delta_read = cur_read - base_read
-            if delta_prompt > 0 and delta_read >= 0:
-                pct = int(round((delta_read / delta_prompt) * 100))
-                pct = max(0, min(100, pct))
+            # A zero-read regime hides the segment entirely (no cache data
+            # is not the same as a 0% hit worth alarming about), and the pct
+            # stays a float so renderers control their own precision.
+            if delta_prompt > 0 and delta_read > 0:
+                pct = max(0.0, min(100.0, (delta_read / delta_prompt) * 100))
                 snapshot["cache_hit_pct"] = pct
-                snapshot["cache_hit_label"] = f"{pct}%"
-            elif cur_prompt > 0 and cur_read >= 0 and base_prompt == 0 and base_read == 0:
-                pct = int(round((cur_read / cur_prompt) * 100)) if cur_prompt else 0
-                pct = max(0, min(100, pct))
+                snapshot["cache_hit_label"] = f"{pct:.0f}%"
+            elif cur_prompt > 0 and cur_read > 0 and base_prompt == 0 and base_read == 0:
+                pct = max(0.0, min(100.0, (cur_read / cur_prompt) * 100))
                 snapshot["cache_hit_pct"] = pct
-                snapshot["cache_hit_label"] = f"{pct}%"
+                snapshot["cache_hit_label"] = f"{pct:.0f}%"
             else:
                 snapshot["cache_hit_pct"] = None
                 snapshot["cache_hit_label"] = ""
diff --git a/tests/cli/test_cli_status_bar.py b/tests/cli/test_cli_status_bar.py
index 5b07bc15e2..76d25a8578 100644
--- a/tests/cli/test_cli_status_bar.py
+++ b/tests/cli/test_cli_status_bar.py
@@ -612,3 +612,106 @@ class TestCacheHitRate:
 
         # cache_read / prompt_tokens = 5000 / 10000 = 50%
         assert "◎ 50.0%" in text
+
+
+class TestRollingLatencyVelocity:
+    def _with_history(self, cli_obj, latencies, outputs):
+        from collections import deque
+        cli_obj.agent._api_latency_history = deque(latencies, maxlen=10)
+        cli_obj.agent._api_output_history = deque(outputs, maxlen=10)
+        return cli_obj
+
+    def test_latency_and_tps_shown_in_wide_terminal(self):
+        cli_obj = _attach_agent(
+            _make_cli(),
+            prompt_tokens=10_000, completion_tokens=2_000, total_tokens=12_000,
+            api_calls=5, context_tokens=12_000, context_length=200_000,
+        )
+        self._with_history(cli_obj, [2.0, 4.0], [120, 180])
+
+        text = cli_obj._build_status_bar_text(width=140)
+
+        assert "\u25f7 3.0s" in text           # mean latency (2+4)/2
+        assert "\u2191 50 t/s" in text          # true throughput 300/6.0
+
+    def test_latency_hidden_without_history(self):
+        cli_obj = _attach_agent(
+            _make_cli(),
+            prompt_tokens=10_000, completion_tokens=2_000, total_tokens=12_000,
+            api_calls=5, context_tokens=12_000, context_length=200_000,
+        )
+        text = cli_obj._build_status_bar_text(width=140)
+        assert "\u25f7" not in text
+        assert "t/s" not in text
+
+    def test_latency_and_tps_respect_field_filter(self):
+        cli_obj = _attach_agent(
+            _make_cli(),
+            prompt_tokens=10_000, completion_tokens=2_000, total_tokens=12_000,
+            api_calls=5, context_tokens=12_000, context_length=200_000,
+        )
+        self._with_history(cli_obj, [2.0], [100])
+        with patch.object(cli_mod, "CLI_CONFIG", {"display": {"status_bar": {"fields": ["model", "duration"]}}}):
+            text = cli_obj._build_status_bar_text(width=140)
+        assert "\u25f7" not in text
+        assert "t/s" not in text
+
+    def test_negative_latency_guard(self):
+        cli_obj = _attach_agent(
+            _make_cli(),
+            prompt_tokens=10_000, completion_tokens=2_000, total_tokens=12_000,
+            api_calls=5, context_tokens=12_000, context_length=200_000,
+        )
+        self._with_history(cli_obj, [-0.8], [100])
+        snapshot = cli_obj._get_status_bar_snapshot()
+        assert snapshot["avg_latency"] is None
+        assert snapshot["avg_velocity"] is None
+
+
+class TestCacheHitBaselineReset:
+    def test_baseline_resets_on_model_switch(self):
+        cli_obj = _attach_agent(
+            _make_cli(),
+            prompt_tokens=10_000, completion_tokens=2_000, total_tokens=12_000,
+            api_calls=5, context_tokens=12_000, context_length=200_000,
+            cache_read_tokens=9_000,
+        )
+        first = cli_obj._get_status_bar_snapshot()
+        assert first["cache_hit_pct"] == 90.0
+
+        # Switch model. The bar repaints every frame, so the switch is
+        # observed (and the baseline reset) before new tokens accrue.
+        cli_obj.model = "openai/gpt-5"
+        cli_obj.agent.model = "openai/gpt-5"
+        reset_snap = cli_obj._get_status_bar_snapshot()
+        assert reset_snap["cache_hit_pct"] is None  # new regime, no data yet
+
+        cli_obj.agent.session_prompt_tokens = 12_000
+        cli_obj.agent.session_cache_read_tokens = 9_500
+        second = cli_obj._get_status_bar_snapshot()
+        # Delta since switch: 500/2000 = 25%, not the lifetime 79%.
+        assert second["cache_hit_pct"] == 25.0
+
+    def test_baseline_resets_on_compression(self):
+        cli_obj = _attach_agent(
+            _make_cli(),
+            prompt_tokens=10_000, completion_tokens=2_000, total_tokens=12_000,
+            api_calls=5, context_tokens=12_000, context_length=200_000,
+            cache_read_tokens=8_000,
+        )
+        cli_obj._get_status_bar_snapshot()
+
+        cli_obj.agent.context_compressor.compression_count = 1
+        cli_obj._get_status_bar_snapshot()  # repaint observes the compression
+
+        cli_obj.agent.session_prompt_tokens = 14_000
+        cli_obj.agent.session_cache_read_tokens = 8_400
+        snap = cli_obj._get_status_bar_snapshot()
+        assert snap["cache_hit_pct"] == 10.0  # 400/4000 post-compression
+
+    def test_title_field_filter_hides_session_badge(self):
+        cli_obj = _make_cli()
+        cli_obj._pending_title = "weekly-digest"
+        with patch.object(cli_mod, "CLI_CONFIG", {"display": {"status_bar": {"fields": ["model", "duration"]}}}):
+            text = cli_obj._build_status_bar_text(width=80)
+        assert "weekly-digest" not in text
diff --git a/website/docs/user-guide/configuration.md b/website/docs/user-guide/configuration.md
index 3263158ba5..27ff51bd54 100644
--- a/website/docs/user-guide/configuration.md
+++ b/website/docs/user-guide/configuration.md
@@ -1980,14 +1980,15 @@ display:
     fields: ["model", "duration", "total_tokens"]   # visibility only; built-in order is preserved
 ```
 
-Supported fields: `model`, `context_detail` (used/total tokens), `context_pct` (percent + meter), `compressions`, `bg_tasks`, `bg_processes`, `bg_subagents`, `goal`, `duration`, `prompt_elapsed`, `idle_since`, `focus`, `yolo`, and `total_tokens` (session Σ — opt-in only, never shown by default).
+Supported fields: `model`, `context_detail` (used/total tokens), `context_pct` (percent + meter), `cache_hit` (prompt cache hit ratio — resets on model switch and compression), `latency` (rolling mean API latency, last 10 calls), `tps` (rolling output tokens/sec, last 10 calls), `compressions`, `bg_tasks`, `bg_processes`, `bg_subagents`, `goal`, `duration`, `prompt_elapsed`, `idle_since`, `focus`, `yolo`, `stash`, `battery`, `title` (right-aligned session badge), and `total_tokens` (session Σ — opt-in only, never shown by default).
 
 Notes:
 
 - An empty list (the default) keeps the standard set — everything except `total_tokens`.
 - The config controls **visibility, not order**; fields render in their built-in positions.
-- Narrow terminals still drop wide-mode-only fields (`context_detail`, `prompt_elapsed`, `idle_since`) regardless of config.
-- The battery indicator, session title, and prompt-stash indicator have their own toggles (`/battery`, `/title`) and are not governed by this list.
+- Narrow terminals still drop wide-mode-only fields (`context_detail`, `cache_hit`, `latency`, `tps`, `prompt_elapsed`, `idle_since`) regardless of config (`cache_hit` also shows in the medium ≥52-col tier).
+- `latency`/`tps` stay hidden until API calls have been recorded (e.g. the Codex app-server backend reports no latency).
+- `battery` and `title` visibility here compose with their own toggles (`/battery`, `/title`) — both must be on for the segment to show.
 - Display-only: no effect on prompt caching or request payloads. Changes take effect on the next session start.
 
 ### Runtime-metadata footer (gateway only)