diff --git a/headroom/perf/analyzer.py b/headroom/perf/analyzer.py
index 388e3a189..ab9b9c2c6 100644
--- a/headroom/perf/analyzer.py
+++ b/headroom/perf/analyzer.py
@@ -22,6 +22,7 @@
 log = logging.getLogger(__name__)
 
 LOG_DIR = _paths.log_dir()
+DEFAULT_SLOW_OPTIMIZATION_MS = 500.0
 
 # Matches: 2026-03-07 13:38:31,009 - headroom.proxy - INFO - [hr_...] PERF model=... ...
 _PERF_RE = re.compile(
@@ -536,15 +537,31 @@ def format_report(report: PerfReport) -> str:
         # Optimization latency
         opt_times = [r.optimization_ms for r in records if r.optimization_ms > 0]
         if opt_times:
-            avg_opt = sum(opt_times) / len(opt_times)
-            max_opt = max(opt_times)
+            overhead = build_overhead_summary(report)
+            opt_stats = overhead["optimization_ms"]
             lines.append("Optimization Overhead")
             lines.append("-" * 40)
-            lines.append(f"  Average:  {avg_opt:.0f}ms")
-            lines.append(f"  Max:      {max_opt:.0f}ms")
-            slow = [t for t in opt_times if t > 500]
-            if slow:
-                lines.append(f"  >500ms:   {len(slow)} requests")
+            lines.append(f"  Average:     {opt_stats['average_ms']:.0f}ms")
+            lines.append(
+                "  p50/p95/p99: "
+                f"{opt_stats['p50_ms']:.0f} / "
+                f"{opt_stats['p95_ms']:.0f} / "
+                f"{opt_stats['p99_ms']:.0f}ms"
+            )
+            lines.append(f"  Max:         {opt_stats['max_ms']:.0f}ms")
+            if opt_stats["slow_request_count"]:
+                lines.append(
+                    f"  >{overhead['slow_threshold_ms']:.0f}ms:      "
+                    f"{opt_stats['slow_request_count']} requests "
+                    f"({opt_stats['slow_request_pct']:.1f}%)"
+                )
+            if overhead["stage_breakdown"]:
+                lines.append("  Slowest stages:")
+                for stage in overhead["stage_breakdown"][:3]:
+                    lines.append(
+                        f"    {stage['stage']}: total {stage['total_ms']:.0f}ms, "
+                        f"p95 {stage['p95_ms']:.0f}ms, max {stage['max_ms']:.0f}ms"
+                    )
             lines.append("")
 
         # Throughput
@@ -723,6 +740,102 @@ def _percentile(data: list[float], pct: float) -> float:
     return sorted_data[lower]
 
 
+def _latency_stats(values: list[float]) -> dict[str, float | int]:
+    if not values:
+        return {
+            "count": 0,
+            "average_ms": 0.0,
+            "p50_ms": 0.0,
+            "p95_ms": 0.0,
+            "p99_ms": 0.0,
+            "max_ms": 0.0,
+        }
+    return {
+        "count": len(values),
+        "average_ms": round(sum(values) / len(values), 2),
+        "p50_ms": round(_percentile(values, 0.50), 2),
+        "p95_ms": round(_percentile(values, 0.95), 2),
+        "p99_ms": round(_percentile(values, 0.99), 2),
+        "max_ms": round(max(values), 2),
+    }
+
+
+def _slowest_stage(record: PerfRecord) -> tuple[str | None, float]:
+    stages = [(name, value) for name, value in record.stages.items() if value > 0]
+    if not stages:
+        return None, 0.0
+    return max(stages, key=lambda item: item[1])
+
+
+def build_overhead_summary(
+    report: PerfReport,
+    *,
+    slow_threshold_ms: float = DEFAULT_SLOW_OPTIMIZATION_MS,
+    limit: int = 10,
+) -> dict:
+    """Aggregate optimization latency and stage-level bottlenecks."""
+    records = report.perf_records
+    opt_times = [r.optimization_ms for r in records if r.optimization_ms > 0]
+    total_times = [r.total_ms for r in records if r.total_ms > 0]
+    slow_records = [r for r in records if r.optimization_ms > slow_threshold_ms]
+
+    stage_values: dict[str, list[float]] = {}
+    for record in records:
+        for stage, value in record.stages.items():
+            if value > 0:
+                stage_values.setdefault(stage, []).append(value)
+
+    stage_breakdown = []
+    for stage, values in stage_values.items():
+        stats = _latency_stats(values)
+        stage_breakdown.append(
+            {
+                "stage": stage,
+                **stats,
+                "total_ms": round(sum(values), 2),
+            }
+        )
+    stage_breakdown.sort(key=lambda row: row["total_ms"], reverse=True)
+
+    top_slow_requests = []
+    for record in sorted(
+        [r for r in records if r.optimization_ms > 0],
+        key=lambda r: r.optimization_ms,
+        reverse=True,
+    )[:limit]:
+        slowest_stage, slowest_stage_ms = _slowest_stage(record)
+        top_slow_requests.append(
+            {
+                "timestamp": record.timestamp,
+                "request_id": record.request_id,
+                "model": record.model,
+                "client": record.client,
+                "optimization_ms": record.optimization_ms,
+                "total_ms": record.total_ms,
+                "tokens_before": record.tokens_before,
+                "tokens_after": record.tokens_after,
+                "tokens_saved": record.tokens_saved,
+                "slowest_stage": slowest_stage,
+                "slowest_stage_ms": slowest_stage_ms,
+                "transforms": record.transforms,
+            }
+        )
+
+    return {
+        "slow_threshold_ms": slow_threshold_ms,
+        "optimization_ms": {
+            **_latency_stats(opt_times),
+            "slow_request_count": len(slow_records),
+            "slow_request_pct": round(len(slow_records) / len(records) * 100, 1)
+            if records
+            else 0.0,
+        },
+        "total_ms": _latency_stats(total_times),
+        "stage_breakdown": stage_breakdown,
+        "top_slow_requests": top_slow_requests,
+    }
+
+
 def calculate_throughput(report: PerfReport) -> dict:
     records = report.perf_records
     parsed_records = []
@@ -902,6 +1015,7 @@ def build_perf_summary(report: PerfReport) -> dict:
         "cache_hit_pct": cache_hit_pct,
         "by_model": by_model,
         "by_transform": by_transform,
+        "overhead": build_overhead_summary(report),
         "throughput": calculate_throughput(report),
         "log_files_read": report.log_files_read,
         "total_lines_parsed": report.total_lines_parsed,
@@ -1027,11 +1141,18 @@ def _generate_recommendations(report: PerfReport) -> list[str]:
                     )
 
         # Optimization latency
-        slow = [r for r in report.perf_records if r.optimization_ms > 500]
-        if len(slow) > len(report.perf_records) * 0.2:
+        overhead = build_overhead_summary(report)
+        opt_stats = overhead["optimization_ms"]
+        slow_count = int(opt_stats["slow_request_count"])
+        if slow_count > len(report.perf_records) * 0.2:
+            stage_hint = ""
+            if overhead["stage_breakdown"]:
+                stage_hint = f"; top stage: {overhead['stage_breakdown'][0]['stage']}"
             recs.append(
-                f"{len(slow)} requests took >500ms for optimization — "
-                "consider reducing transform pipeline"
+                f"{slow_count} requests took >{overhead['slow_threshold_ms']:.0f}ms "
+                "for optimization"
+                f"{stage_hint} — consider reducing heavy transforms or lowering "
+                "HEADROOM_COMPRESSION_TIMEOUT_SECONDS"
             )
 
     if report.router_records:
diff --git a/tests/test_cli_perf_format.py b/tests/test_cli_perf_format.py
index 7e1941181..38f9e0eae 100644
--- a/tests/test_cli_perf_format.py
+++ b/tests/test_cli_perf_format.py
@@ -15,6 +15,7 @@
     PerfRecord,
     PerfReport,
     TransformRecord,
+    build_overhead_summary,
     build_perf_summary,
     perf_records_as_dicts,
 )
@@ -116,6 +117,49 @@ def test_build_perf_summary_empty_report_no_zero_division():
     assert summary["savings_pct"] == 0.0
     assert summary["cache_hit_pct"] == 0.0
     assert summary["by_model"] == []
+    assert summary["overhead"]["optimization_ms"]["count"] == 0
+
+
+def test_build_overhead_summary_attributes_slow_stages():
+    report = PerfReport(
+        perf_records=[
+            PerfRecord(
+                timestamp="2026-06-05 10:00:00,000",
+                request_id="fast",
+                model="gpt-5",
+                tokens_before=1000,
+                tokens_after=500,
+                tokens_saved=500,
+                optimization_ms=100.0,
+                total_ms=300.0,
+                stages={"cache_align": 10.0, "content_router": 90.0},
+            ),
+            PerfRecord(
+                timestamp="2026-06-05 10:01:00,000",
+                request_id="slow",
+                model="gpt-5",
+                tokens_before=1000,
+                tokens_after=500,
+                tokens_saved=500,
+                optimization_ms=700.0,
+                total_ms=900.0,
+                stages={"kompress": 650.0, "content_router": 40.0},
+            ),
+        ]
+    )
+
+    overhead = build_overhead_summary(report, slow_threshold_ms=500.0)
+
+    assert overhead["optimization_ms"]["count"] == 2
+    assert overhead["optimization_ms"]["average_ms"] == 400.0
+    assert overhead["optimization_ms"]["p50_ms"] == 400.0
+    assert overhead["optimization_ms"]["p95_ms"] == 670.0
+    assert overhead["optimization_ms"]["p99_ms"] == 694.0
+    assert overhead["optimization_ms"]["slow_request_count"] == 1
+    assert overhead["stage_breakdown"][0]["stage"] == "kompress"
+    assert overhead["stage_breakdown"][0]["total_ms"] == 650.0
+    assert overhead["top_slow_requests"][0]["request_id"] == "slow"
+    assert overhead["top_slow_requests"][0]["slowest_stage"] == "kompress"
 
 
 def test_perf_records_as_dicts_roundtrips_fields():
@@ -144,6 +188,7 @@ def test_perf_json_format(runner, monkeypatch):
     assert data["savings_pct"] == 50.0
     assert "by_model" in data
     assert data["total_requests"] == 2
+    assert data["overhead"]["optimization_ms"]["p95_ms"] == 11.8
 
 
 def test_perf_json_raw_is_array(runner, monkeypatch):
@@ -214,6 +259,7 @@ def test_perf_text_default_unchanged(runner, monkeypatch):
     result = runner.invoke(main, ["perf"])
     assert result.exit_code == 0, result.output
     assert "Headroom Performance Report" in result.output
+    assert "p50/p95/p99" in result.output
 
 
 def test_perf_rejects_unknown_format(runner, monkeypatch):