fix(opencode): make the OSS provider usable at all

src/ucode/agents/opencode.py

@@ -143,6 +143,12 @@ def render_overlay(
                 "baseURL": opencode_base_urls["oss"],
                 "apiKey": token,
                 "headers": auth_headers,
+                # OpenCode stamps `prompt_cache_key: <sessionID>` on every
+                # `@ai-sdk/openai` request; the Databricks gateway's strict
+                # validator rejects the unknown field with HTTP 400. OpenCode
+                # reads this opt-out from the provider `options` only — a
+                # per-model `options` entry does not suppress the field.
+                "setCacheKey": False,
             },
             "models": {m: _oss_model_overlay(m, ua_header) for m in oss_models},
         }

src/ucode/databricks.py

@@ -1495,9 +1495,16 @@ def classify_model_family(model_id: str) -> str | None:
 # context and output together. Keyed by family substring; add an entry to bound
 # a new model.
 _MODEL_TOKEN_LIMITS: dict[str, dict[str, int]] = {
-    # GLM-4.6: 200k context, but the gateway caps output well below the model's
-    # native 128k — pin 25k so requests aren't rejected.
+    # Every `output` here is the gateway's cap, which sits below the model's
+    # native output length. Neither the model-services listing nor the
+    # serving-endpoints API reports the cap, so it cannot be discovered.
     "glm": {"context": 200_000, "output": 25_000},
+    "qwen": {"context": 262_144, "output": 25_000},
+    "gpt-oss": {"context": 131_072, "output": 25_000},
+    # Keyed on the full name, not `llama`: the Llama 3 endpoints have a 128k
+    # context, so a bare `llama` key would pin 1M on them.
+    "llama-4-maverick": {"context": 1_000_000, "output": 8_192},
+    "gemma": {"context": 131_072, "output": 8_192},
 }
 
 

tests/test_agent_opencode.py

@@ -116,6 +116,21 @@ def test_non_glm_oss_model_has_no_output_cap(self):
         kimi = overlay["provider"]["databricks-oss"]["models"]["system.ai.kimi-k2-7-code"]
         assert "limit" not in kimi
 
+    def test_qwen_gets_token_limits(self):
+        model = "system.ai.qwen35-122b-a10b"
+        overlay, _ = opencode.render_overlay(model, "tok", _base_urls(), {"oss": [model]})
+        qwen = overlay["provider"]["databricks-oss"]["models"][model]
+        assert qwen["limit"] == {"context": 262144, "output": 25000}
+
+    def test_oss_provider_opts_out_of_the_prompt_cache_key(self):
+        # OpenCode stamps `prompt_cache_key` on every `@ai-sdk/openai` request and
+        # the gateway rejects the unknown field. It reads the opt-out from the
+        # provider options only, so a per-model entry would not suppress it.
+        model = "system.ai.qwen35-122b-a10b"
+        overlay, _ = opencode.render_overlay(model, "tok", _base_urls(), {"oss": [model]})
+        assert overlay["provider"]["databricks-oss"]["options"]["setCacheKey"] is False
+        assert "setCacheKey" not in overlay["provider"]["databricks-oss"]["models"][model]
+
     def test_token_in_api_key(self):
         models = {"anthropic": ["claude-sonnet"]}
         overlay, _ = opencode.render_overlay("claude-sonnet", "mytoken", _base_urls(), models)

tests/test_databricks.py

@@ -244,6 +244,20 @@ def test_glm_matches_any_version(self):
     def test_uncapped_model_returns_none(self):
         assert db_mod.model_token_limits("system.ai.kimi-k2-7-code") is None
 
+    def test_gateway_caps_measured_against_the_workspace(self):
+        # Each `output` is the cap the gateway enforces; a request above it fails
+        # with HTTP 400.
+        assert db_mod.model_token_limits("system.ai.qwen35-122b-a10b")["output"] == 25_000
+        assert db_mod.model_token_limits("system.ai.gpt-oss-120b")["output"] == 25_000
+        assert db_mod.model_token_limits("system.ai.gpt-oss-20b")["output"] == 25_000
+        assert db_mod.model_token_limits("system.ai.llama-4-maverick")["output"] == 8_192
+        assert db_mod.model_token_limits("system.ai.gemma-3-12b")["output"] == 8_192
+
+    def test_llama_3_endpoints_do_not_take_the_maverick_context(self):
+        # A bare `llama` key would pin Maverick's 1M context on the 128k Llama 3
+        # endpoints.
+        assert db_mod.model_token_limits("system.ai.meta-llama-3-3-70b-instruct") is None
+
 
 class TestDiscoverModelServices:
     def test_buckets_families_by_name(self, monkeypatch):