diff --git a/crates/headroom-core/src/transforms/recommendations.rs b/crates/headroom-core/src/transforms/recommendations.rs
index a2efee865..6c5c842ab 100644
--- a/crates/headroom-core/src/transforms/recommendations.rs
+++ b/crates/headroom-core/src/transforms/recommendations.rs
@@ -28,6 +28,7 @@
 //! auth_mode = "payg"
 //! model_family = "claude-3-5"
 //! structure_hash = "deadbeef..."
+//! skip_compression_recommended = false
 //! strategy_hint = "smart_crusher"
 //! confidence = 0.87
 //! observations = 142
@@ -71,6 +72,8 @@ pub struct Recommendation {
     pub auth_mode: String,
     pub model_family: String,
     pub structure_hash: String,
+    #[serde(default)]
+    pub skip_compression_recommended: bool,
     pub strategy_hint: String,
     pub confidence: f64,
     pub observations: u64,
@@ -270,6 +273,7 @@ mod tests {
 auth_mode = "payg"
 model_family = "claude-3-5"
 structure_hash = "deadbeef"
+skip_compression_recommended = true
 strategy_hint = "smart_crusher"
 confidence = 0.87
 observations = 142
@@ -292,11 +296,21 @@ observations = 60
         let r = store
             .lookup(AuthMode::Payg, "claude-3-5", "deadbeef")
             .expect("hit");
+        assert!(r.skip_compression_recommended);
         assert_eq!(r.strategy_hint, "smart_crusher");
         assert!((r.confidence - 0.87).abs() < 1e-9);
         assert_eq!(r.observations, 142);
     }
 
+    #[test]
+    fn from_toml_str_defaults_missing_skip_field_to_false() {
+        let store = RecommendationStore::from_toml_str(sample_toml()).expect("parses");
+        let r = store
+            .lookup(AuthMode::OAuth, "gpt-4o", "cafebabe")
+            .expect("hit");
+        assert!(!r.skip_compression_recommended);
+    }
+
     #[test]
     fn lookup_returns_none_for_missing_slice() {
         let store = RecommendationStore::from_toml_str(sample_toml()).expect("parses");
diff --git a/headroom/cli/toin_publish.py b/headroom/cli/toin_publish.py
index c067b0828..89a408bab 100644
--- a/headroom/cli/toin_publish.py
+++ b/headroom/cli/toin_publish.py
@@ -15,6 +15,7 @@ result as TOML the Rust proxy loads at startup
     auth_mode = "payg"
     model_family = "claude-3-5"
     structure_hash = "deadbeef..."
+    skip_compression_recommended = false
     strategy_hint = "smart_crusher"
     confidence = 0.87
     observations = 142
@@ -30,7 +31,9 @@ alongside the Rust binary, and the proxy reads it once at startup.
 # Output stability
 
 Rows are sorted by ``(auth_mode, model_family, structure_hash)`` so the
-file diffs cleanly across publishes. Strategies ship as the
+file diffs cleanly across publishes. Rows with
+``skip_compression_recommended = true`` publish
+``strategy_hint = "skip_compression"``. Other rows ship the
 ToolPattern's learned ``optimal_strategy`` (or the dominant entry of
 ``strategy_success_rates``). Confidence is the pattern's existing
 confidence score — bounded ``[0.0, 0.95]`` by the confidence calculator.
@@ -98,16 +101,19 @@ def _format_row(
     auth_mode: str,
     model_family: str,
     structure_hash: str,
+    skip_compression_recommended: bool,
     strategy_hint: str,
     confidence: float,
     observations: int,
 ) -> str:
     """Render one ``[[recommendation]]`` block."""
+    skip_flag = "true" if skip_compression_recommended else "false"
     return (
         "[[recommendation]]\n"
         f'auth_mode = "{_toml_escape(auth_mode)}"\n'
         f'model_family = "{_toml_escape(model_family)}"\n'
         f'structure_hash = "{_toml_escape(structure_hash)}"\n'
+        f"skip_compression_recommended = {skip_flag}\n"
         f'strategy_hint = "{_toml_escape(strategy_hint)}"\n'
         f"confidence = {confidence:.4f}\n"
         f"observations = {observations}\n"
@@ -132,12 +138,18 @@ def _eligible_rows(
         if observations < min_observations:
             continue
         auth_mode, model_family, sig_hash = key
+        strategy_hint = (
+            "skip_compression"
+            if pattern.skip_compression_recommended
+            else _select_strategy(pattern)
+        )
         rows.append(
             {
                 "auth_mode": auth_mode,
                 "model_family": model_family,
                 "structure_hash": sig_hash,
-                "strategy_hint": _select_strategy(pattern),
+                "skip_compression_recommended": pattern.skip_compression_recommended,
+                "strategy_hint": strategy_hint,
                 "confidence": float(pattern.confidence),
                 "observations": observations,
             }
diff --git a/tests/test_toin_publish.py b/tests/test_toin_publish.py
index 1104acf89..a6c2bbe11 100644
--- a/tests/test_toin_publish.py
+++ b/tests/test_toin_publish.py
@@ -5,7 +5,8 @@ Pins:
 1. ``publish()`` writes a TOML file the stdlib ``tomllib`` can parse.
 2. Slices below ``--min-observations`` are filtered out.
 3. Rows include ``auth_mode``, ``model_family``, ``structure_hash``,
-   ``strategy_hint``, ``confidence``, ``observations`` — the schema
+   ``skip_compression_recommended``, ``strategy_hint``, ``confidence``,
+   ``observations`` — the schema
    ``crates/headroom-core/src/transforms/recommendations.rs`` consumes.
 4. The CLI entry point honors ``--output`` / ``--min-observations``.
 """
@@ -99,6 +100,7 @@ def test_publish_command_writes_toml(fresh_toin: ToolIntelligenceNetwork, tmp_pa
         "auth_mode",
         "model_family",
         "structure_hash",
+        "skip_compression_recommended",
         "strategy_hint",
         "confidence",
         "observations",
@@ -106,12 +108,49 @@ def test_publish_command_writes_toml(fresh_toin: ToolIntelligenceNetwork, tmp_pa
     assert row["auth_mode"] == "payg"
     assert row["model_family"] == "claude-3-5"
     assert row["structure_hash"] == sig.structure_hash
+    assert row["skip_compression_recommended"] is False
     assert row["strategy_hint"] == "smart_crusher"
     assert isinstance(row["confidence"], float)
     assert 0.0 <= row["confidence"] <= 1.0
     assert row["observations"] == 60
 
 
+def test_publish_preserves_skip_recommendation(
+    fresh_toin: ToolIntelligenceNetwork,
+    tmp_path: Path,
+) -> None:
+    """Skip-eligible rows publish the skip flag and skip strategy hint."""
+    items = [{"id": i, "status": "ok"} for i in range(20)]
+    sig = _record(
+        fresh_toin,
+        items=items,
+        n=60,
+        auth_mode="payg",
+        model_family="claude-3-5",
+    )
+    for _ in range(49):
+        fresh_toin.record_retrieval(
+            tool_signature_hash=sig.structure_hash,
+            retrieval_type="full",
+            strategy="smart_crusher",
+            auth_mode="payg",
+            model_family="claude-3-5",
+        )
+
+    output = tmp_path / "recommendations.toml"
+    rows_written = publish(
+        output_path=output,
+        min_observations=50,
+        toin=fresh_toin,
+    )
+    assert rows_written == 1
+
+    parsed = tomllib.loads(output.read_text(encoding="utf-8"))
+    row = parsed["recommendation"][0]
+    assert row["skip_compression_recommended"] is True
+    assert row["strategy_hint"] == "skip_compression"
+
+
 def test_publish_filters_below_min_observations(
     fresh_toin: ToolIntelligenceNetwork,
     tmp_path: Path,