fix(opencode): add GPT routing alongside OSS discovery

src/ucode/agents/__init__.py

@@ -363,6 +363,7 @@ def check_gateway_endpoint(state: dict, tool: str) -> bool:
             bool(state.get("claude_models"))
             or bool(state.get("codex_models"))
             or bool(state.get("gemini_models"))
+            or bool(state.get("oss_models"))
         )
     return False
 
@@ -373,7 +374,7 @@ def check_gateway_endpoint(state: dict, tool: str) -> bool:
     "codex": ("codex",),
     "gemini": ("gemini",),
     "copilot": ("claude", "codex"),
-    "pi": ("claude", "codex", "gemini"),
+    "pi": ("claude", "codex", "gemini", "oss"),
 }
 
 

src/ucode/agents/claude.py

@@ -27,6 +27,7 @@
 from ucode.databricks import (
     build_auth_shell_command,
     build_tool_base_url,
+    claude_model_supports_1m,
     get_databricks_token,
 )
 from ucode.launcher import exec_or_spawn
@@ -69,11 +70,6 @@ def _resolve_web_search_model(state: dict) -> str | None:
 
 
 WEB_SEARCH_MCP_NAME = "web_search"
-# Matches both the AI Gateway form (`databricks-claude-opus-4-8`) and the UC
-# model-services form (`system.ai.claude-opus-4-8`).
-_CLAUDE_MODEL_RE = re.compile(
-    r"^(?:system\.ai\.)?(?:databricks-)?claude-(opus|sonnet)-(\d+)-(\d+)(.*)$"
-)
 
 # Env keys the MLflow Stop hook reads to route traces. Written into the
 # settings `env` block alongside the hook itself.
@@ -333,19 +329,9 @@ def render_overlay(
 
 
 def _maybe_add_1m_suffix(model: str) -> str:
-    if model.endswith("[1m]"):
-        return model
-    match = _CLAUDE_MODEL_RE.match(model)
-    if not match:
+    if model.endswith("[1m]") or not claude_model_supports_1m(model):
         return model
-
-    family, major_raw, minor_raw, _ = match.groups()
-    major = int(major_raw)
-    minor = int(minor_raw)
-    should_suffix = (family == "opus" and (major, minor) >= (4, 6)) or (
-        family == "sonnet" and (major, minor) >= (4, 6)
-    )
-    return f"{model}[1m]" if should_suffix else model
+    return f"{model}[1m]"
 
 
 def _register_web_search_mcp(workspace: str, search_model: str, profile: str | None = None) -> bool:

src/ucode/agents/opencode.py

@@ -20,7 +20,9 @@
     TOKEN_REFRESH_INTERVAL_SECONDS,
     build_opencode_base_urls,
     get_databricks_token,
+    gpt_model_token_limits,
     model_token_limits,
+    preferred_gpt_model,
 )
 from ucode.state import mark_tool_managed, save_state
 from ucode.telemetry import agent_version, ucode_version
@@ -41,6 +43,7 @@
 PROVIDER_KEYS: list[list[str]] = [
     ["provider", "databricks-anthropic"],
     ["provider", "databricks-google"],
+    ["provider", "databricks-openai"],
     ["provider", "databricks-oss"],
 ]
 
@@ -51,7 +54,14 @@ def is_update_available() -> tuple[str, str] | None:
 
 def _resolve_model_selector(model: str, opencode_models: dict[str, list[str]]) -> str:
     """Return an OpenCode model selector in provider/model form when possible."""
-    if model.startswith(("databricks-anthropic/", "databricks-google/", "databricks-oss/")):
+    if model.startswith(
+        (
+            "databricks-anthropic/",
+            "databricks-google/",
+            "databricks-openai/",
+            "databricks-oss/",
+        )
+    ):
         return model
 
     anthropic_models = opencode_models.get("anthropic") or []
@@ -62,6 +72,10 @@ def _resolve_model_selector(model: str, opencode_models: dict[str, list[str]]) -
     if model in gemini_models:
         return f"databricks-google/{model}"
 
+    openai_models = opencode_models.get("openai") or []
+    if model in openai_models:
+        return f"databricks-openai/{model}"
+
     oss_models = opencode_models.get("oss") or []
     if model in oss_models:
         return f"databricks-oss/{model}"
@@ -83,6 +97,15 @@ def _oss_model_overlay(model: str, ua_header: dict[str, str]) -> dict:
     return overlay
 
 
+def _openai_model_overlay(model: str, ua_header: dict[str, str]) -> dict:
+    """Per-model Responses API options and explicit GPT token limits."""
+    return {
+        "headers": ua_header,
+        "limit": gpt_model_token_limits(model),
+        "options": {"useResponsesApi": True},
+    }
+
+
 def render_overlay(
     model: str,
     token: str,
@@ -101,6 +124,7 @@ def render_overlay(
 
     anthropic_models = opencode_models.get("anthropic") or []
     gemini_models = opencode_models.get("gemini") or []
+    openai_models = opencode_models.get("openai") or []
     oss_models = opencode_models.get("oss") or []
 
     providers: dict = {}
@@ -136,6 +160,22 @@ def render_overlay(
             "models": {m: {"headers": ua_header} for m in gemini_models},
         }
         keys.append(["provider", "databricks-google"])
+    if openai_models:
+        # @ai-sdk/openai supports both the Responses API and the legacy
+        # chat-completions API. Databricks GPT-5 / GPT-5.6 / Codex models are
+        # Responses-only on /ai-gateway/codex/v1, so the per-model flag
+        # `useResponsesApi: true` lives in models.<m>.options where opencode
+        # reads it (provider-level options is read by the SDK only).
+        providers["databricks-openai"] = {
+            "npm": "@ai-sdk/openai",
+            "options": {
+                "baseURL": opencode_base_urls["openai"],
+                "apiKey": token,
+                "headers": auth_headers,
+            },
+            "models": {m: _openai_model_overlay(m, ua_header) for m in openai_models},
+        }
+        keys.append(["provider", "databricks-openai"])
     if oss_models:
         providers["databricks-oss"] = {
             "npm": "@ai-sdk/openai",
@@ -232,6 +272,9 @@ def default_model(state: dict) -> str | None:
     anthropic = opencode_models.get("anthropic") or []
     if anthropic:
         return anthropic[0]
+    openai = preferred_gpt_model(opencode_models.get("openai") or [])
+    if openai:
+        return openai
     gemini = opencode_models.get("gemini") or []
     if gemini:
         return gemini[0]

src/ucode/agents/pi.py

@@ -1,12 +1,13 @@
 """Pi coding agent: writes ~/.pi/agent/models.json with Databricks-backed providers.
 
-Pi (https://pi.dev) is a multi-provider coding agent. We register three
+Pi (https://pi.dev) is a multi-provider coding agent. We register four
 providers in its `models.json`, each speaking the API dialect best suited to
 that family's gateway path:
 
 - `databricks-claude`  (api: anthropic-messages)       → /ai-gateway/anthropic
 - `databricks-openai`  (api: openai-responses)         → /ai-gateway/codex/v1
 - `databricks-gemini`  (api: google-generative-ai)     → /ai-gateway/gemini/v1beta
+- `databricks-mlflow`  (api: openai-completions)       → /ai-gateway/mlflow/v1
 
 Per-provider `compat` flags work around fields the gateway translators reject:
 
@@ -15,11 +16,17 @@
   pi uses for every request. With this flag pi omits the per-tool field and
   sends the legacy `anthropic-beta: fine-grained-tool-streaming-...` header
   instead, which the gateway accepts.
-
-OSS / Databricks-foundation models (Llama, Qwen, etc.) are not exposed via
-pi today — they live behind /ai-gateway/mlflow/v1 with per-model
-`max_tokens` caps that pi has no global way to honor without per-model
-config we don't currently maintain.
+- mlflow: `supportsStore: false` and `supportsStrictMode: false` — the MLflow
+  chat-completions gateway rejects OpenAI's `store` field and
+  `tools[].function.strict`.
+
+The `databricks-mlflow` provider carries the validated OSS coding models
+(GLM and Kimi) discovered upstream. Per model it sets
+`contextWindow`/`maxTokens` from `databricks.model_token_limits` and
+`reasoning` from `databricks.model_is_reasoning` (so Pi renders the gateway's
+streamed reasoning_content as thinking). Inkling is intentionally not offered
+until the gateway emits a terminal `finish_reason` on natural completion
+(issue #215).
 
 The bearer token is baked into the file and refreshed by a background thread
 while the session runs (same pattern as OpenCode/Copilot).
@@ -44,7 +51,12 @@
 from ucode.databricks import (
     TOKEN_REFRESH_INTERVAL_SECONDS,
     build_pi_base_urls,
+    claude_model_capabilities,
     get_databricks_token,
+    gpt_model_token_limits,
+    model_is_reasoning,
+    model_token_limits,
+    preferred_gpt_model,
 )
 from ucode.state import mark_tool_managed, save_state
 from ucode.telemetry import agent_version, ucode_version
@@ -68,6 +80,7 @@
     "databricks-claude",
     "databricks-openai",
     "databricks-gemini",
+    "databricks-mlflow",
 )
 
 PROVIDER_KEYS: list[list[str]] = [["providers", name] for name in PROVIDER_NAMES]
@@ -86,6 +99,7 @@ def _resolve_model_selector(
     claude_models: dict[str, str],
     codex_models: list[str],
     gemini_models: list[str],
+    oss_models: list[str],
 ) -> str:
     """Return a Pi model selector in `<provider>/<model>` form when possible."""
     for name in PROVIDER_NAMES:
@@ -97,16 +111,73 @@ def _resolve_model_selector(
         return f"databricks-openai/{model}"
     if model in gemini_models:
         return f"databricks-gemini/{model}"
+    if model in oss_models:
+        return f"databricks-mlflow/{model}"
     return model
 
 
+def _pi_claude_model_entry(model_id: str) -> dict:
+    """Build a Claude entry with explicit limits.
+
+    Databricks model ids do not match Pi's built-in Anthropic ids, so a bare
+    custom entry silently gets Pi's 128k context / 4k output defaults.
+    """
+    capabilities = claude_model_capabilities(model_id)
+    entry: dict = {
+        "id": model_id,
+        "reasoning": True,
+        "input": ["text", "image"],
+        "contextWindow": capabilities.context,
+        "maxTokens": capabilities.output,
+    }
+    if capabilities.force_adaptive_thinking:
+        entry["compat"] = {"forceAdaptiveThinking": True}
+    return entry
+
+
+def _pi_oss_model_entry(model_id: str) -> dict:
+    """Build a Pi mlflow model entry enriched from the shared limits/reasoning
+    tables: `reasoning:true` for reasoning models (Pi renders their streamed
+    reasoning_content as thinking), and `contextWindow`/`maxTokens` from
+    `model_token_limits`. Fields are omitted when unknown so Pi keeps its
+    default."""
+    entry: dict = {"id": model_id}
+    if model_is_reasoning(model_id):
+        entry["reasoning"] = True
+    limits = model_token_limits(model_id)
+    if limits:
+        if limits.get("context"):
+            entry["contextWindow"] = limits["context"]
+        if limits.get("output"):
+            entry["maxTokens"] = limits["output"]
+    return entry
+
+
+def _pi_gpt_model_entry(model_id: str) -> dict:
+    """Build a Pi openai (codex) model entry with `contextWindow`/`maxTokens`
+    from `databricks.gpt_model_token_limits`. GPT ids aren't in Pi's built-in
+    catalog, so without an explicit window Pi falls back to a small default and
+    truncates long sessions."""
+    limits = gpt_model_token_limits(model_id)
+    entry: dict = {
+        "id": model_id,
+        "contextWindow": limits["context"],
+        "maxTokens": limits["output"],
+    }
+    if "gpt-5" in model_id.lower().replace(".", "-"):
+        entry["reasoning"] = True
+        entry["input"] = ["text", "image"]
+    return entry
+
+
 def render_overlay(
     model: str,
     token: str,
     pi_base_urls: dict[str, str],
     claude_models: dict[str, str],
     codex_models: list[str],
     gemini_models: list[str],
+    oss_models: list[str],
 ) -> tuple[dict, list[list[str]]]:
     """Return (overlay, managed_key_paths) for ~/.pi/agent/models.json."""
     providers: dict = {}
@@ -127,7 +198,7 @@ def render_overlay(
             # the legacy beta header instead when this is false.
             "compat": {"supportsEagerToolInputStreaming": False},
             "headers": ua_headers,
-            "models": [{"id": m} for m in claude_ids],
+            "models": [_pi_claude_model_entry(m) for m in claude_ids],
         }
         keys.append(["providers", "databricks-claude"])
     if codex_models:
@@ -137,7 +208,7 @@ def render_overlay(
             "apiKey": token,
             "authHeader": True,
             "headers": ua_headers,
-            "models": [{"id": m} for m in codex_models],
+            "models": [_pi_gpt_model_entry(m) for m in codex_models],
         }
         keys.append(["providers", "databricks-openai"])
     if gemini_models:
@@ -150,8 +221,23 @@ def render_overlay(
             "models": [{"id": m} for m in gemini_models],
         }
         keys.append(["providers", "databricks-gemini"])
+    if oss_models:
+        providers["databricks-mlflow"] = {
+            "baseUrl": pi_base_urls["oss"],
+            "api": "openai-completions",
+            "apiKey": token,
+            "authHeader": True,
+            # MLflow chat-completions gateway rejects OpenAI's `store` field
+            # and per-tool `strict`. Pi omits both when these are false.
+            "compat": {"supportsStore": False, "supportsStrictMode": False},
+            "headers": ua_headers,
+            "models": [_pi_oss_model_entry(m) for m in oss_models],
+        }
+        keys.append(["providers", "databricks-mlflow"])
     overlay: dict = {
-        "model": _resolve_model_selector(model, claude_models, codex_models, gemini_models),
+        "model": _resolve_model_selector(
+            model, claude_models, codex_models, gemini_models, oss_models
+        ),
     }
     if providers:
         overlay["providers"] = providers
@@ -178,6 +264,7 @@ def write_tool_config(
         state.get("claude_models") or {},
         state.get("codex_models") or [],
         state.get("gemini_models") or [],
+        state.get("oss_models") or [],
     )
     existing = read_json_safe(PI_CONFIG_PATH)
     providers = existing.get("providers")
@@ -206,16 +293,19 @@ def _write_settings(model_selector: str) -> None:
 
 
 def default_model(state: dict) -> str | None:
-    """Prefer Claude opus → sonnet → haiku; fall back to codex, gemini."""
+    """Prefer Claude opus → sonnet → haiku; fall back to codex, Gemini, then OSS."""
     claude_models = state.get("claude_models") or {}
     for family in ("opus", "sonnet", "haiku"):
         if claude_models.get(family):
             return claude_models[family]
-    codex_models = state.get("codex_models") or []
-    if codex_models:
-        return codex_models[0]
+    codex_model = preferred_gpt_model(state.get("codex_models") or [])
+    if codex_model:
+        return codex_model
     gemini_models = state.get("gemini_models") or []
-    return gemini_models[0] if gemini_models else None
+    if gemini_models:
+        return gemini_models[0]
+    oss_models = state.get("oss_models") or []
+    return oss_models[0] if oss_models else None
 
 
 def _refresh_token_once(state: dict, *, force_refresh: bool = False) -> str: