support glm-5-2 on opencode (#186)

* Broaden OSS model bucket to catch all non-claude/gpt/gemini families

discover_model_services bucketed OSS models with an allowlist ("kimi-"
only), so GLM/Llama/Qwen/DeepSeek and any future OSS family were
discovered from UC model-services but dropped on the floor — never
reaching opencode_models, so they couldn't be listed or selected in
OpenCode (e.g. system.ai.glm-5-2).

Bucket OSS by exclusion instead: anything not claude/gpt/gemini, minus
non-chat model services (embeddings, rerankers) that can't back a chat
agent. New OSS families now surface without a code change.

Co-authored-by: Isaac

* Bound GLM token limits (200k context / 25k output) for OSS-serving agents

The Databricks gateway caps GLM's max output (completion) tokens and
rejects requests above it. These limits are a property of the model +
its /ai-gateway/mlflow/v1 route, not of any one agent, so expose them as
a single model_token_limits() helper (family->limits table) in
databricks.py. Each agent translates the limits into its own config
dialect.

OpenCode's per-model schema is strict (`additionalProperties: false`)
and its `limit` object requires BOTH context and output — supplying only
output made the whole opencode.json invalid ("Configuration is
invalid"). The table now always provides both fields, so OpenCode pins
`limit: {context, output}` and clamps `max_tokens` to a gateway-accepted
value. Adding a model/limit is a one-line edit shared across
opencode/pi/copilot.

Co-authored-by: Isaac

* Restrict OSS model bucket to kimi and glm allowlist

The previous catch-all bucketed any non-claude/gpt/gemini model (llama,
qwen, deepseek, ...) into OSS, offering families we haven't validated.
Go back to an explicit allowlist, but include both kimi and glm.

_OSS_MODEL_FAMILIES = ("kimi-", "glm-") replaces the exclusion logic and
its non-chat marker filter (embeddings/rerankers never match the
allowlist, so no filtering is needed).

Co-authored-by: Isaac

* Revert test_no_matching_families_reports_sample to llama sample

Under the kimi/glm allowlist, llama-4-maverick again matches no family,
so the original test (asserting it's reported in the "no families" reason)
holds without the embedding-service substitute.

Co-authored-by: Isaac

* Trim verbose comments on OSS allowlist

Co-authored-by: Isaac

Author: lilly-luo

src/ucode/agents/opencode.py

@@ -20,6 +20,7 @@
     TOKEN_REFRESH_INTERVAL_SECONDS,
     build_opencode_base_urls,
     get_databricks_token,
+    model_token_limits,
 )
 from ucode.state import mark_tool_managed, save_state
 from ucode.telemetry import agent_version, ucode_version
@@ -69,6 +70,20 @@ def _resolve_model_selector(model: str, opencode_models: dict[str, list[str]]) -
     return model
 
 
+def _oss_model_overlay(model: str, ua_header: dict[str, str]) -> dict:
+    """Per-model overlay for an OSS model entry.
+
+    All OSS models carry the User-Agent header; models with known token limits
+    also pin `limit` (context + output) so OpenCode clamps `max_tokens` to a
+    value the gateway accepts. OpenCode's schema requires both fields together,
+    so the limits table always supplies both."""
+    overlay: dict = {"headers": ua_header}
+    limits = model_token_limits(model)
+    if limits is not None:
+        overlay["limit"] = limits
+    return overlay
+
+
 def render_overlay(
     model: str,
     token: str,
@@ -130,7 +145,7 @@ def render_overlay(
                 "apiKey": token,
                 "headers": auth_headers,
             },
-            "models": {m: {"headers": ua_header} for m in oss_models},
+            "models": {m: _oss_model_overlay(m, ua_header) for m in oss_models},
         }
         keys.append(["provider", "databricks-oss"])
 

src/ucode/databricks.py

@@ -1096,6 +1096,34 @@ def build_auth_shell_command(
 # Databricks-managed foundation models under `system.ai`.
 _MODEL_SERVICE_REQUIRED_PREFIX = "system.ai."
 
+# Supported OSS chat families, matched by name substring. Add an entry to
+# support a new family.
+_OSS_MODEL_FAMILIES = ("kimi-", "glm-")
+
+# Per-family token limits (context window + max output tokens). These are a
+# property of the model + its `/ai-gateway/mlflow/v1` route (the gateway rejects
+# requests whose output exceeds the cap), not of any one agent — so every agent
+# that serves OSS models reads this single table and translates it into its own
+# config dialect. Both fields are provided because agents like OpenCode require
+# 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.
+    "glm": {"context": 200_000, "output": 25_000},
+}
+
+
+def model_token_limits(model_id: str) -> dict[str, int] | None:
+    """Return ``{"context": ..., "output": ...}`` limits for ``model_id``, or None.
+
+    Matches by family substring (e.g. any ``*glm*`` id). None means the model
+    has no known limits and the agent should not pin any."""
+    for family, limits in _MODEL_TOKEN_LIMITS.items():
+        if family in model_id:
+            return dict(limits)
+    return None
+
 
 def _model_service_id(service: dict) -> str | None:
     """Extract the `system.ai.<model-name>` id from one model-service entry.
@@ -1223,7 +1251,8 @@ def discover_model_services(
 
     codex_models = [m for m in ids if "gpt-" in m]
     gemini_models = sorted([m for m in ids if "gemini-" in m], key=model_version_sort_key)
-    oss_models = [m for m in ids if "kimi-" in m]
+
+    oss_models = [m for m in ids if any(family in m for family in _OSS_MODEL_FAMILIES)]
 
     if not (claude_models or codex_models or gemini_models or oss_models):
         sample = ", ".join(ids[:5])

tests/test_agent_opencode.py

@@ -93,6 +93,21 @@ def test_oss_base_url(self):
         options = overlay["provider"]["databricks-oss"]["options"]
         assert options["baseURL"] == f"{WS}/ai-gateway/mlflow/v1"
 
+    def test_glm_gets_token_limits(self):
+        models = {"oss": ["system.ai.glm-5-2"]}
+        overlay, _ = opencode.render_overlay("system.ai.glm-5-2", "tok", _base_urls(), models)
+        glm = overlay["provider"]["databricks-oss"]["models"]["system.ai.glm-5-2"]
+        # OpenCode's schema requires both context and output on `limit`.
+        assert glm["limit"] == {"context": 200000, "output": 25000}
+
+    def test_non_glm_oss_model_has_no_output_cap(self):
+        models = {"oss": ["system.ai.kimi-k2-7-code"]}
+        overlay, _ = opencode.render_overlay(
+            "system.ai.kimi-k2-7-code", "tok", _base_urls(), models
+        )
+        kimi = overlay["provider"]["databricks-oss"]["models"]["system.ai.kimi-k2-7-code"]
+        assert "limit" not in kimi
+
     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

@@ -140,6 +140,23 @@ def _model_service(model_id: str) -> dict:
     return {"name": f"model-services/{model_id}"}
 
 
+class TestModelTokenLimits:
+    def test_glm_is_capped(self):
+        assert db_mod.model_token_limits("system.ai.glm-5-2") == {
+            "context": 200_000,
+            "output": 25_000,
+        }
+
+    def test_glm_matches_any_version(self):
+        assert db_mod.model_token_limits("system.ai.glm-4-6-flash") == {
+            "context": 200_000,
+            "output": 25_000,
+        }
+
+    def test_uncapped_model_returns_none(self):
+        assert db_mod.model_token_limits("system.ai.kimi-k2-7-code") is None
+
+
 class TestDiscoverModelServices:
     def test_buckets_families_by_name(self, monkeypatch):
         payload = {
@@ -151,6 +168,7 @@ def test_buckets_families_by_name(self, monkeypatch):
                 _model_service("system.ai.gemini-2-5-flash"),
                 _model_service("system.ai.gemini-3-5-flash"),
                 _model_service("system.ai.kimi-k2-7-code"),
+                _model_service("system.ai.glm-5-2"),
                 _model_service("system.ai.llama-4-maverick"),
             ]
         }
@@ -169,9 +187,30 @@ def test_buckets_families_by_name(self, monkeypatch):
         assert codex == ["system.ai.gpt-5"]
         # Gemini ordered newest-first via the shared sort key.
         assert gemini[0] == "system.ai.gemini-3-5-flash"
-        assert oss == ["system.ai.kimi-k2-7-code"]
-        # llama is not bucketed into any of the four families.
-        assert "system.ai.llama-4-maverick" not in codex + gemini + oss
+        # kimi and glm are the allowlisted OSS families; llama is not.
+        assert oss == ["system.ai.glm-5-2", "system.ai.kimi-k2-7-code"]
+
+    def test_oss_allowlist_drops_unsupported_families(self, monkeypatch):
+        # Only kimi/glm are allowlisted; other families are dropped.
+        payload = {
+            "model_services": [
+                _model_service("system.ai.glm-5-2"),
+                _model_service("system.ai.kimi-k2-7-code"),
+                _model_service("system.ai.qwen-3-coder"),
+                _model_service("system.ai.deepseek-v3"),
+                _model_service("system.ai.gte-large-embed"),
+                _model_service("system.ai.bge-reranker-v2"),
+            ]
+        }
+        monkeypatch.setattr(
+            db_mod, "_http_get_json", lambda url, token, timeout=10: (payload, None)
+        )
+
+        claude, codex, gemini, oss, reason = db_mod.discover_model_services(WS, "token")
+
+        assert reason is None
+        assert (claude, codex, gemini) == ({}, [], [])
+        assert oss == ["system.ai.glm-5-2", "system.ai.kimi-k2-7-code"]
 
     def test_paginates_via_next_page_token(self, monkeypatch):
         pages = {