diff --git a/packages/hub-client/src/catalog-seed-data.ts b/packages/hub-client/src/catalog-seed-data.ts
index 0ad37a43a..742bdfd3d 100644
--- a/packages/hub-client/src/catalog-seed-data.ts
+++ b/packages/hub-client/src/catalog-seed-data.ts
@@ -217,10 +217,13 @@ export const CATALOG_SEEDS: Readonly<
       baseURL: "http://localhost:11434/v1",
     },
     // Curated against a real instance's `/api/tags` (see this ticket's
-    // report): whichever of these two the person's own Ollama actually
-    // has loaded, `qwen3.8:27b` leads since it is the one confirmed to
-    // serve tool calls and thinking.
+    // report): whichever of these the person's own Ollama actually has
+    // loaded. `gpt-oss:20b` leads — CL-6477's overnight run confirmed it
+    // correct and tool-obedient, and 12-25x faster than `qwen3.8:27b` (8s
+    // vs 100-216s for the same task) — with `qwen3.8:27b` kept next as
+    // the prior confirmed-good default and `qwen3.5:9b-mlx` last.
     models: [
+      { canonicalName: "gpt-oss:20b", displayName: "GPT-OSS 20B (Ollama)" },
       { canonicalName: "qwen3.8:27b", displayName: "Qwen 3.8 27B (Ollama)" },
       {
         canonicalName: "qwen3.5:9b-mlx",
diff --git a/packages/hub-client/src/model-capability.test.ts b/packages/hub-client/src/model-capability.test.ts
index 56cfde9ec..e89ccf483 100644
--- a/packages/hub-client/src/model-capability.test.ts
+++ b/packages/hub-client/src/model-capability.test.ts
@@ -50,6 +50,51 @@ describe("preferCompletionCapable", () => {
       [],
     );
   });
+
+  // CL-6477: "embeddinggemma" has no delimiter between "embedding" and
+  // "gemma", unlike every other embedding model name this filter has seen
+  // (nomic-embed-text, all-minilm, bge-m3, qwen3-embedding all delimit
+  // with "-" or "_"). A regex that required a trailing delimiter after
+  // "embed(ding)?" let it through, where it then won the alphabetical
+  // default-model tiebreak and answered every chat turn with "does not
+  // support chat".
+  test("drops an uncataloged embeddinggemma offering even with no delimiter after 'embedding'", () => {
+    const offerings = [noData("embeddinggemma:300m"), completion("qwen3:8b")];
+    expect(
+      preferCompletionCapable(offerings, capabilitiesOf, nameOf).map(
+        (o) => o.name,
+      ),
+    ).toEqual(["qwen3:8b"]);
+  });
+
+  test("still filters the delimited embedding-model names (no regression)", () => {
+    const offerings = [
+      noData("nomic-embed-text"),
+      noData("all-minilm"),
+      noData("bge-m3"),
+      noData("qwen3-embedding"),
+      completion("qwen3:8b"),
+    ];
+    expect(
+      preferCompletionCapable(offerings, capabilitiesOf, nameOf).map(
+        (o) => o.name,
+      ),
+    ).toEqual(["qwen3:8b"]);
+  });
+
+  test("prefers real capability data over the name fallback: a probed completion offering named like an embedding model is kept", () => {
+    const probedCompletion: Offering = {
+      name: "embeddinggemma:300m",
+      capabilities: ["plain-text"],
+    };
+    expect(
+      preferCompletionCapable(
+        [probedCompletion, noData("all-minilm")],
+        capabilitiesOf,
+        nameOf,
+      ).map((o) => o.name),
+    ).toEqual(["embeddinggemma:300m"]);
+  });
 });
 
 describe("hasCompletionCapableModel", () => {
@@ -72,4 +117,24 @@ describe("hasCompletionCapableModel", () => {
       ),
     ).toBe(false);
   });
+
+  test("false when the only offering is an uncataloged embeddinggemma pull (CL-6477)", () => {
+    expect(
+      hasCompletionCapableModel(
+        [noData("embeddinggemma:300m")],
+        capabilitiesOf,
+        nameOf,
+      ),
+    ).toBe(false);
+  });
+
+  test("a chat-capable model is selected as the default when both an embeddinggemma pull and a chat model are present", () => {
+    const offerings = [noData("embeddinggemma:300m"), completion("qwen3:8b")];
+    const chatDefault = preferCompletionCapable(
+      offerings,
+      capabilitiesOf,
+      nameOf,
+    )[0];
+    expect(chatDefault?.name).toBe("qwen3:8b");
+  });
 });
diff --git a/packages/hub-client/src/model-capability.ts b/packages/hub-client/src/model-capability.ts
index 2d3b17c0b..5c7a59051 100644
--- a/packages/hub-client/src/model-capability.ts
+++ b/packages/hub-client/src/model-capability.ts
@@ -23,9 +23,24 @@
 // case this falls back to recognizing common embedding-model name
 // families by convention — the same signal CL-6351's first pass used
 // before the pinned catalog existed — rather than trusting an empty list
-// as "no data" and letting an embedding model win anyway.
+// as "no data" and letting an embedding model win anyway. Capability data
+// is always tried first; the name check below only ever runs as a
+// belt-and-braces fallback for the uncataloged case.
+//
+// CL-6477: `embed(ding)?` carries no required delimiter after it, unlike
+// the other families here. Google's `embeddinggemma:300m` — an
+// uncataloged Ollama pull with no capability data, same as any other
+// local embedding model — has no delimiter between "embedding" and the
+// "gemma" model family name, so a trailing-delimiter requirement lets it
+// slip past the filter entirely and then win the alphabetical
+// default-model tiebreak with a model that answers every chat turn with
+// "does not support chat" (CL-6351 reopened through a name its own fix
+// missed). The other short abbreviations here (`minilm`, `bge`, `gte`,
+// `e5`, `arctic-embed`) keep a required trailing delimiter: they are
+// short enough that matching them as a bare substring risks false
+// positives no real model name has exercised yet.
 const EMBEDDING_MODEL_NAME_PATTERN =
-  /(^|[-_/])(embed(ding)?|minilm|bge|gte|e5|arctic-embed)(-|_|:|$)/i;
+  /(^|[-_/])embed(ding)?|(^|[-_/])(minilm|bge|gte|e5|arctic-embed)(-|_|:|$)/i;
 
 function isEmbeddingModelName(canonicalName: string): boolean {
   return EMBEDDING_MODEL_NAME_PATTERN.test(canonicalName);
diff --git a/packages/onboarding/test/complete-credential.test.ts b/packages/onboarding/test/complete-credential.test.ts
index e264ab886..cf2acd4dd 100644
--- a/packages/onboarding/test/complete-credential.test.ts
+++ b/packages/onboarding/test/complete-credential.test.ts
@@ -149,9 +149,8 @@ describe("modelSourceFor", () => {
   });
 
   // CL-6366 red/green: a fresh instance whose live catalog carries only
-  // llama3.2 (the curated `qwen3.8:27b` name is absent entirely) still
-  // resolves to what the instance actually serves, never a pin it can't
-  // answer for.
+  // llama3.2 (the curated default name is absent entirely) still resolves
+  // to what the instance actually serves, never a pin it can't answer for.
   test("ollama resolves to the instance's own seeded model, never the curated pin it may lack", async () => {
     const api: ApiCall = async (method, path) => {
       if (method === "GET" && path === `/api/tenants/${TENANT_ID}/models`) {
@@ -218,7 +217,7 @@ describe("modelSourceFor", () => {
             capabilities: ["plain-text"],
           },
           {
-            canonicalName: "qwen3.8:27b",
+            canonicalName: "gpt-oss:20b",
             providerName: "ollama",
             capabilities: ["plain-text"],
           },
@@ -234,7 +233,7 @@ describe("modelSourceFor", () => {
       "ollama",
       "ollama",
     );
-    expect(result.model).toBe("qwen3.8:27b");
+    expect(result.model).toBe("gpt-oss:20b");
   });
 
   test("ollama's baseURLOverride is normalized to the /v1 form", async () => {