diff --git a/pyproject.toml b/pyproject.toml
index 86bda87..b7db686 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -4,7 +4,7 @@ build-backend = "uv_build"
 
 [project]
 name = "ucode"
-version = "0.1.0"
+version = "0.2.0"
 description = "Bootstrap Codex against a Databricks workspace with minimal setup."
 readme = "README.md"
 requires-python = ">=3.12"
diff --git a/src/ucode/agents/_mlflow_proxy.py b/src/ucode/agents/_mlflow_proxy.py
new file mode 100644
index 0000000..80e763a
--- /dev/null
+++ b/src/ucode/agents/_mlflow_proxy.py
@@ -0,0 +1,201 @@
+"""Local SSE-repair proxy for the MLflow chat-completions gateway.
+
+Some MLflow-served models (notably ``databricks-inkling``) stream a
+spec-incomplete response: on natural completion the terminal chunk omits the
+OpenAI-required ``finish_reason`` and the stream ends with only ``data: [DONE]``.
+Strict clients — Pi's ``openai-completions`` parser among them — reject this
+with "Stream ended without finish_reason", so the model is unusable.
+
+This proxy sits between Pi and the gateway. It forwards every request verbatim
+and, for streaming (``text/event-stream``) responses, repairs the terminator:
+when the stream ends — cleanly at ``[DONE]`` OR abnormally (an upstream drop /
+mid-stream 429) — after some data but without a ``finish_reason``, it
+synthesizes a ``finish_reason: "stop"`` chunk (and a ``[DONE]`` if the upstream
+never sent one). Non-streaming and error responses are relayed byte-for-byte
+with their original ``Content-Type``. Well-behaved streams never trigger the
+injection, so the proxy is a no-op for them.
+
+Tracked upstream as the gateway bug that should make this unnecessary; once the
+gateway emits ``finish_reason`` this module can be deleted.
+"""
+
+from __future__ import annotations
+
+import json
+import threading
+import urllib.error
+import urllib.request
+from http.server import BaseHTTPRequestHandler, HTTPServer
+from socketserver import ThreadingMixIn
+
+# Request headers we must not forward: hop-by-hop or ones urllib recomputes.
+_SKIP_REQUEST_HEADERS = frozenset({"host", "content-length", "accept-encoding", "connection"})
+
+# The proxy only exists to front the mlflow chat-completions route. Refuse any
+# other path so a co-located process can't turn the localhost relay into an
+# arbitrary authenticated workspace client (SSRF-to-workspace) using the token
+# we forward.
+_ALLOWED_PATH_PREFIX = "/ai-gateway/mlflow/"
+
+
+def _make_handler(upstream: str) -> type[BaseHTTPRequestHandler]:
+    class _Handler(BaseHTTPRequestHandler):
+        protocol_version = "HTTP/1.1"
+
+        def do_POST(self) -> None:  # noqa: N802 (BaseHTTPRequestHandler API)
+            # Read the body regardless so the socket is drained before we close.
+            length = int(self.headers.get("Content-Length", 0))
+            body = self.rfile.read(length)
+
+            if not self.path.startswith(_ALLOWED_PATH_PREFIX):
+                self._send_bytes(404, b'{"error":"path not allowed"}', "application/json")
+                return
+
+            req = urllib.request.Request(upstream + self.path, data=body, method="POST")
+            for key, value in self.headers.items():
+                if key.lower() not in _SKIP_REQUEST_HEADERS:
+                    req.add_header(key, value)
+            try:
+                upstream_resp = urllib.request.urlopen(req)  # noqa: S310 (trusted upstream)
+            except urllib.error.HTTPError as exc:
+                # Relay the gateway's own error verbatim (status + body).
+                payload = exc.read() if exc.fp else b""
+                content_type = exc.headers.get("Content-Type", "application/json")
+                self._send_bytes(exc.code, payload, content_type)
+                return
+            except (urllib.error.URLError, OSError) as exc:
+                # Connection-level failure (DNS/TLS/timeout/reset): give the
+                # client a clean 502 instead of a dead socket.
+                msg = json.dumps({"error": f"upstream unreachable: {exc}"}).encode()
+                self._send_bytes(502, msg, "application/json")
+                return
+
+            content_type = upstream_resp.headers.get("Content-Type", "")
+            if "text/event-stream" not in content_type:
+                # Non-streaming (or error-envelope) 200: relay unchanged.
+                self._send_bytes(
+                    upstream_resp.status, upstream_resp.read(), content_type or "application/json"
+                )
+                return
+
+            # Streaming: close the connection at end-of-stream so an HTTP/1.1
+            # client can detect the message boundary (we send no length/chunking).
+            self.close_connection = True
+            self.send_response(upstream_resp.status)
+            self.send_header("Content-Type", "text/event-stream")
+            self.send_header("Cache-Control", "no-cache")
+            self.send_header("Connection", "close")
+            self.end_headers()
+            self._relay_stream(upstream_resp)
+
+        def _relay_stream(self, upstream_resp) -> None:
+            saw_finish = False
+            saw_done = False
+            saw_data = False
+            last_id: str | None = None
+            try:
+                for raw in upstream_resp:  # yields line-by-line as bytes arrive
+                    payload = _sse_data_payload(raw)
+                    if payload is not None and payload != b"[DONE]":
+                        saw_data = True
+                        try:
+                            obj = json.loads(payload)
+                            last_id = obj.get("id", last_id)
+                            choice = (obj.get("choices") or [{}])[0]
+                            if choice.get("finish_reason"):
+                                saw_finish = True
+                        except (ValueError, AttributeError, IndexError):
+                            pass
+                        self._write(raw)
+                    elif payload == b"[DONE]":
+                        if not saw_finish:
+                            self._write(b"data: " + _finish_chunk(last_id) + b"\n\n")
+                            saw_finish = True
+                        self._write(raw)
+                        saw_done = True
+                    else:
+                        self._write(raw)
+                # Clean or abnormal end after data but without a terminator:
+                # synthesize one so strict clients don't error on a truncated
+                # stream (a mid-stream 429 is the common trigger).
+                if saw_data and not saw_finish:
+                    self._write(b"data: " + _finish_chunk(last_id) + b"\n\n")
+                if saw_data and not saw_done:
+                    self._write(b"data: [DONE]\n\n")
+            except OSError:
+                # Upstream dropped, or the client disconnected mid-write. Either
+                # way the socket is gone; swallow so the daemon thread exits
+                # quietly rather than tracebacking.
+                pass
+
+        def _send_bytes(self, status: int, data: bytes, content_type: str) -> None:
+            self.close_connection = True
+            try:
+                self.send_response(status)
+                self.send_header("Content-Type", content_type)
+                self.send_header("Content-Length", str(len(data)))
+                self.send_header("Connection", "close")
+                self.end_headers()
+                self.wfile.write(data)
+                self.wfile.flush()
+            except OSError:
+                pass
+
+        def _write(self, data: bytes) -> None:
+            self.wfile.write(data)
+            self.wfile.flush()
+
+        def log_message(self, format: str, *args) -> None:  # silence default stderr logging
+            pass
+
+    return _Handler
+
+
+def _sse_data_payload(raw: bytes) -> bytes | None:
+    """Return the payload of an SSE ``data:`` line (the spec allows an optional
+    single space after the colon), or None if the line isn't a data line."""
+    stripped = raw.strip()
+    if not stripped.startswith(b"data:"):
+        return None
+    payload = stripped[len(b"data:") :]
+    if payload.startswith(b" "):
+        payload = payload[1:]
+    return payload
+
+
+def _finish_chunk(chunk_id: str | None) -> bytes:
+    return json.dumps(
+        {
+            "id": chunk_id,
+            "object": "chat.completion.chunk",
+            "choices": [{"delta": {}, "index": 0, "finish_reason": "stop"}],
+        }
+    ).encode()
+
+
+class _Server(ThreadingMixIn, HTTPServer):
+    daemon_threads = True
+    allow_reuse_address = True
+
+
+def start(upstream: str) -> tuple[str, threading.Event]:
+    """Start the repair proxy on a free localhost port.
+
+    :param upstream: gateway origin to forward to, e.g.
+        ``https://workspace.databricks.com``.
+    :returns: ``(base_url, stop_event)`` where ``base_url`` is the local origin
+        Pi should target (``http://127.0.0.1:<port>``) and setting
+        ``stop_event`` shuts the server down.
+    """
+    server = _Server(("127.0.0.1", 0), _make_handler(upstream.rstrip("/")))
+    port = server.server_address[1]
+    stop_event = threading.Event()
+
+    threading.Thread(target=server.serve_forever, daemon=True).start()
+
+    def _watch_stop() -> None:
+        stop_event.wait()
+        server.shutdown()
+
+    threading.Thread(target=_watch_stop, daemon=True).start()
+    return f"http://127.0.0.1:{port}", stop_event
diff --git a/src/ucode/agents/claude.py b/src/ucode/agents/claude.py
index 422e763..82821a0 100644
--- a/src/ucode/agents/claude.py
+++ b/src/ucode/agents/claude.py
@@ -232,8 +232,12 @@ def _maybe_add_1m_suffix(model: str) -> str:
     family, major_raw, minor_raw, _ = match.groups()
     major = int(major_raw)
     minor = int(minor_raw)
+    # 1M-token context window support: Opus from 4.6, Sonnet from 4.5 (Sonnet
+    # 4.5 gained the 1M beta window before Opus did). Without the `[1m]`
+    # suffix the gateway serves these models with the default 200k context, so
+    # gate each family at its own floor rather than a shared version.
     should_suffix = (family == "opus" and (major, minor) >= (4, 6)) or (
-        family == "sonnet" and (major, minor) >= (4, 6)
+        family == "sonnet" and (major, minor) >= (4, 5)
     )
     return f"{model}[1m]" if should_suffix else model
 
diff --git a/src/ucode/agents/pi.py b/src/ucode/agents/pi.py
index e7c1760..3982be3 100644
--- a/src/ucode/agents/pi.py
+++ b/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 +16,20 @@
   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.
+- mlflow: `supportsStore: false` and `supportsStrictMode: false` — the MLflow
+  chat-completions gateway rejects OpenAI's `store` field and
+  `tools[].function.strict`.
 
-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.
+The `databricks-mlflow` provider carries the OSS chat-completions-only
+foundation models (Llama, Qwen, GLM, inkling, ...) 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).
+
+At launch the mlflow provider is routed through a local SSE-repair proxy (see
+`_mlflow_proxy`): some OSS models (inkling) omit the `finish_reason` on the
+streaming terminal chunk, which Pi's parser rejects. The proxy injects it and
+is a no-op for well-behaved models. Removable once the gateway is fixed.
 
 The bearer token is baked into the file and refreshed by a background thread
 while the session runs (same pattern as OpenCode/Copilot).
@@ -28,11 +38,13 @@
 from __future__ import annotations
 
 import os
+import re
 import signal
 import subprocess
 import threading
 
 from ucode.agent_updates import available_npm_package_update
+from ucode.agents import _mlflow_proxy
 from ucode.config_io import (
     APP_DIR,
     ToolSpec,
@@ -44,7 +56,11 @@
 from ucode.databricks import (
     TOKEN_REFRESH_INTERVAL_SECONDS,
     build_pi_base_urls,
+    claude_model_token_limits,
     get_databricks_token,
+    gpt_model_token_limits,
+    model_is_reasoning,
+    model_token_limits,
 )
 from ucode.state import mark_tool_managed, save_state
 from ucode.telemetry import agent_version, ucode_version
@@ -68,6 +84,7 @@
     "databricks-claude",
     "databricks-openai",
     "databricks-gemini",
+    "databricks-mlflow",
 )
 
 PROVIDER_KEYS: list[list[str]] = [["providers", name] for name in PROVIDER_NAMES]
@@ -86,6 +103,7 @@ 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:
@@ -97,9 +115,66 @@ 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.
+    """
+    limits = claude_model_token_limits(model_id)
+    entry: dict = {
+        "id": model_id,
+        "reasoning": True,
+        "input": ["text", "image"],
+        "contextWindow": limits["context"],
+        "maxTokens": limits["output"],
+    }
+    normalized = model_id.lower().replace(".", "-")
+    if re.search(r"claude-(?:opus|sonnet)-4-(?:[6-9]|[1-9]\d)", normalized):
+        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,
@@ -107,6 +182,7 @@ def render_overlay(
     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 +203,7 @@ 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 +213,7 @@ 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 +226,23 @@ 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 +269,7 @@ 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")
@@ -215,7 +307,10 @@ def default_model(state: dict) -> str | None:
     if codex_models:
         return codex_models[0]
     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:
@@ -241,7 +336,29 @@ def build_runtime_env(token: str) -> dict[str, str]:
     return env
 
 
+def _start_oss_proxy(state: dict) -> threading.Event | None:
+    """Route the mlflow provider through the local SSE-repair proxy.
+
+    Rewrites ``state["base_urls"]["pi"]["oss"]`` in-memory to the proxy origin
+    so both the initial config write and every token-refresh rewrite point Pi
+    at the proxy. The gateway origin is always re-derived from the workspace
+    (never the persisted `oss` URL, which may hold a dead proxy port from a
+    prior session). Returns the proxy's stop event, or None when no OSS models
+    are configured. See `_mlflow_proxy`.
+    """
+    if not (state.get("oss_models") or []):
+        return None
+    pi_urls = state.setdefault("base_urls", {}).setdefault(
+        "pi", build_pi_base_urls(state["workspace"])
+    )
+    origin = build_pi_base_urls(state["workspace"])["oss"].split("/ai-gateway/", 1)[0]
+    proxy_base, stop_event = _mlflow_proxy.start(origin)
+    pi_urls["oss"] = f"{proxy_base}/ai-gateway/mlflow/v1"
+    return stop_event
+
+
 def launch(state: dict, tool_args: list[str]) -> None:
+    proxy_stop = _start_oss_proxy(state)
     token = _refresh_token_once(state)
     env = build_runtime_env(token)
 
@@ -262,6 +379,8 @@ def launch(state: dict, tool_args: list[str]) -> None:
     finally:
         stop_event.set()
         refresher.join(timeout=1)
+        if proxy_stop is not None:
+            proxy_stop.set()
 
     raise SystemExit(returncode)
 
diff --git a/src/ucode/cli.py b/src/ucode/cli.py
index 9ebdae9..9800d9d 100644
--- a/src/ucode/cli.py
+++ b/src/ucode/cli.py
@@ -324,7 +324,7 @@ def configure_shared_state(
     )
     want_gemini = fetch_all or "gemini" in tools or "opencode" in tools or "pi" in tools
     want_codex = fetch_all or "codex" in tools or "copilot" in tools or "pi" in tools
-    want_oss = fetch_all or "opencode" in tools
+    want_oss = fetch_all or "opencode" in tools or "pi" in tools
 
     claude_reason: str | None = None
     gemini_reason: str | None = None
diff --git a/src/ucode/databricks.py b/src/ucode/databricks.py
index 9f89c6b..8aa70e4 100644
--- a/src/ucode/databricks.py
+++ b/src/ucode/databricks.py
@@ -1097,8 +1097,31 @@ def build_auth_shell_command(
 _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-")
+# support a new family. Covers the full chat-completions-only cohort on the
+# `/ai-gateway/mlflow/v1` route (verified 2026-07-16): kimi, glm, inkling,
+# llama, qwen, gpt-oss, gemma.
+_OSS_MODEL_FAMILIES = (
+    "kimi-",
+    "glm-",
+    "inkling",
+    "llama-",
+    "qwen",
+    "gpt-oss-",
+    "gemma-",
+)
+
+# Non-chat services that can share a family substring with a chat model (e.g.
+# `qwen3-embedding-*` matches the `qwen` family) but can't back a chat agent.
+# The model-services API doesn't expose api_types, so exclude by name.
+_OSS_NON_CHAT_SUBSTRINGS = ("embedding", "embed", "rerank")
+
+
+def _is_oss_chat_model(model_id: str) -> bool:
+    """True if the id matches an OSS chat family and isn't a non-chat service."""
+    if any(bad in model_id for bad in _OSS_NON_CHAT_SUBSTRINGS):
+        return False
+    return any(family in model_id for family in _OSS_MODEL_FAMILIES)
+
 
 # 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
@@ -1107,24 +1130,130 @@ def build_auth_shell_command(
 # 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.
+#
+# Output caps probed from the gateway 2026-07-16 (it 400s with "max_tokens (N)
+# cannot exceed <cap>"); context windows from each model's docs/description
+# (conservative when unstated). If the gateway raises a cap or ships a new
+# model, update this table.
 _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},
+    "glm": {"context": 1_000_000, "output": 65_536},
+    "inkling": {"context": 128_000, "output": 65_536},
+    "kimi": {"context": 128_000, "output": 65_536},
+    "gpt-oss": {"context": 128_000, "output": 25_000},
+    "qwen35": {"context": 128_000, "output": 25_000},
+    "qwen3-next": {"context": 128_000, "output": 10_000},
+    "llama-4": {"context": 128_000, "output": 8_192},
+    "llama-3": {"context": 128_000, "output": 8_192},
+    "gemma": {"context": 128_000, "output": 8_192},
 }
 
+# Conservative fallback for a model that IS an OSS chat model (bucketed by
+# `_is_oss_chat_model`) but has no specific `_MODEL_TOKEN_LIMITS` entry — e.g. a
+# `qwen3-coder` or future `llama-5`. The discovery families are intentionally
+# broader than the limits table, so without a floor such a model would be
+# offered with no output cap and 400 on a default-size request. 8192 is the
+# lowest cap observed across the cohort, so every mlflow chat model accepts it;
+# pinning it only risks truncating a longer response (the safe failure
+# direction), never a hard 400.
+_OSS_FALLBACK_LIMITS = {"context": 128_000, "output": 8_192}
+
+# OSS families that emit reasoning (openai_reasoning capability, verified
+# 2026-07-16). Pi renders their streamed reasoning_content as thinking when the
+# model entry sets reasoning:true; agents whose SDK surfaces reasoning natively
+# (OpenCode's @ai-sdk/openai) need no flag.
+_OSS_REASONING_FAMILIES = ("glm", "inkling", "kimi", "gpt-oss", "qwen35")
+
+
+def model_is_reasoning(model_id: str) -> bool:
+    """True if the OSS model reports reasoning output (family-matched)."""
+    return any(family in model_id for family in _OSS_REASONING_FAMILIES)
+
 
 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."""
+    Prefers a specific `_MODEL_TOKEN_LIMITS` family entry (e.g. any ``*glm*``
+    id). Any other OSS chat model falls back to a conservative floor so it is
+    never offered uncapped (which would 400). None only for non-OSS ids, where
+    the agent should not pin any limit."""
     for family, limits in _MODEL_TOKEN_LIMITS.items():
         if family in model_id:
             return dict(limits)
+    if _is_oss_chat_model(model_id):
+        return dict(_OSS_FALLBACK_LIMITS)
     return None
 
 
+# Pi treats every custom model without explicit metadata as 128k context / 4k
+# output. Gateway ids are custom ids (not Pi's built-ins), so preserve the
+# upstream windows explicitly. Entries are ordered most-specific first after
+# normalizing dotted OpenAI ids and hyphenated Databricks ids to one form.
+# GPT-5.6 Sol/Terra/Luna support the opt-in 1.05M window; 272k is merely Pi's
+# built-in short-context pricing default, not the model's hard context limit.
+_GPT_TOKEN_LIMITS: tuple[tuple[str, dict[str, int]], ...] = (
+    ("gpt-5-6-sol", {"context": 1_050_000, "output": 128_000}),
+    ("gpt-5-6-terra", {"context": 1_050_000, "output": 128_000}),
+    ("gpt-5-6-luna", {"context": 1_050_000, "output": 128_000}),
+    ("gpt-5-5-pro", {"context": 1_050_000, "output": 128_000}),
+    ("gpt-5-4-pro", {"context": 1_050_000, "output": 128_000}),
+    ("gpt-5-5", {"context": 272_000, "output": 128_000}),
+    ("gpt-5-4-mini", {"context": 400_000, "output": 128_000}),
+    ("gpt-5-4-nano", {"context": 400_000, "output": 128_000}),
+    ("gpt-5-4", {"context": 272_000, "output": 128_000}),
+    ("gpt-5", {"context": 400_000, "output": 128_000}),
+    ("gpt-4-1", {"context": 1_047_576, "output": 32_768}),
+    ("gpt-4o", {"context": 128_000, "output": 16_384}),
+    ("gpt-4-turbo", {"context": 128_000, "output": 4_096}),
+    ("gpt-4", {"context": 8_192, "output": 8_192}),
+)
+_GPT_FALLBACK_LIMITS = {"context": 128_000, "output": 16_384}
+
+
+def _normalized_foundation_model_id(model_id: str) -> str:
+    """Strip route prefixes and normalize dotted versions to hyphens."""
+    tail = model_id.split("/")[-1]
+    if tail.startswith("system.ai."):
+        tail = tail[len("system.ai.") :]
+    if tail.startswith("databricks-"):
+        tail = tail[len("databricks-") :]
+    return tail.lower().replace(".", "-")
+
+
+def gpt_model_token_limits(model_id: str) -> dict[str, int]:
+    """Return Pi metadata limits for a GPT (codex/openai) gateway model."""
+    tail = _normalized_foundation_model_id(model_id)
+    for family, limits in _GPT_TOKEN_LIMITS:
+        if tail.startswith(family):
+            return dict(limits)
+    return dict(_GPT_FALLBACK_LIMITS)
+
+
+_CLAUDE_FALLBACK_LIMITS = {"context": 200_000, "output": 64_000}
+
+
+def claude_model_token_limits(model_id: str) -> dict[str, int]:
+    """Return Pi metadata limits for a Claude gateway model.
+
+    Claude gateway ids are custom to Pi, so omitting these fields silently
+    applies Pi's 128k/4k custom-model defaults. Current 1M families mirror
+    Anthropic's model metadata: Opus >=4.6 and Sonnet >=4.5. Sonnet 4.5 has a
+    64k output cap; later 1M models have 128k output.
+    """
+    tail = _normalized_foundation_model_id(model_id)
+    match = re.match(r"claude-(opus|sonnet|haiku)-(\d+)-(\d+)", tail)
+    if not match:
+        return dict(_CLAUDE_FALLBACK_LIMITS)
+    family, major_raw, minor_raw = match.groups()
+    version = (int(major_raw), int(minor_raw))
+    if family == "opus" and version >= (4, 6):
+        return {"context": 1_000_000, "output": 128_000}
+    if family == "sonnet" and version >= (4, 6):
+        return {"context": 1_000_000, "output": 128_000}
+    if family == "sonnet" and version >= (4, 5):
+        return {"context": 1_000_000, "output": 64_000}
+    return dict(_CLAUDE_FALLBACK_LIMITS)
+
+
 def _model_service_id(service: dict) -> str | None:
     """Extract the `system.ai.<model-name>` id from one model-service entry.
 
@@ -1249,10 +1378,13 @@ def discover_model_services(
         if candidates:
             claude_models[family] = candidates[0]
 
-    codex_models = [m for m in ids if "gpt-" in m]
+    # `gpt-oss-*` also contains "gpt-" but is a chat-completions-only OSS model
+    # (served via /ai-gateway/mlflow/v1), NOT an openai-responses codex model —
+    # exclude it here so it isn't offered under the codex provider (which 400s).
+    codex_models = [m for m in ids if "gpt-" in m and "gpt-oss" not 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 any(family in m for family in _OSS_MODEL_FAMILIES)]
+    oss_models = [m for m in ids if _is_oss_chat_model(m)]
 
     if not (claude_models or codex_models or gemini_models or oss_models):
         sample = ", ".join(ids[:5])
@@ -2133,6 +2265,7 @@ def build_pi_base_urls(workspace: str) -> dict[str, str]:
         "claude": build_tool_base_url("claude", workspace),
         "openai": build_tool_base_url("codex", workspace),
         "gemini": build_tool_base_url("gemini", workspace) + "/v1beta",
+        "oss": f"{workspace}/ai-gateway/mlflow/v1",
     }
 
 
diff --git a/tests/test_agent_claude.py b/tests/test_agent_claude.py
index b2d3a27..a69fb9d 100644
--- a/tests/test_agent_claude.py
+++ b/tests/test_agent_claude.py
@@ -48,6 +48,30 @@ def test_adds_1m_suffix_for_sonnet_4_6_and_later(self):
             overlay["env"]["ANTHROPIC_DEFAULT_SONNET_MODEL"] == "databricks-claude-sonnet-4-7[1m]"
         )
 
+    def test_adds_1m_suffix_for_sonnet_4_5(self):
+        # Sonnet 4.5 supports the 1M context window (its 1M beta shipped before
+        # Opus's), so it must get the [1m] suffix even though it predates the
+        # Opus 4.6 floor.
+        overlay, _ = claude.render_overlay(
+            WS, "s4", claude_models={"sonnet": "databricks-claude-sonnet-4-5"}
+        )
+        assert (
+            overlay["env"]["ANTHROPIC_DEFAULT_SONNET_MODEL"] == "databricks-claude-sonnet-4-5[1m]"
+        )
+
+    def test_does_not_add_1m_suffix_for_sonnet_4_4(self):
+        overlay, _ = claude.render_overlay(
+            WS, "s4", claude_models={"sonnet": "databricks-claude-sonnet-4-4"}
+        )
+        assert overlay["env"]["ANTHROPIC_DEFAULT_SONNET_MODEL"] == "databricks-claude-sonnet-4-4"
+
+    def test_does_not_add_1m_suffix_for_opus_4_5(self):
+        # Opus's 1M window starts at 4.6, so 4.5 stays on the default context.
+        overlay, _ = claude.render_overlay(
+            WS, "s4", claude_models={"opus": "databricks-claude-opus-4-5"}
+        )
+        assert overlay["env"]["ANTHROPIC_DEFAULT_OPUS_MODEL"] == "databricks-claude-opus-4-5"
+
     def test_does_not_add_1m_suffix_for_haiku(self):
         overlay, _ = claude.render_overlay(
             WS, "s4", claude_models={"haiku": "databricks-claude-haiku-4-6"}
diff --git a/tests/test_agent_opencode.py b/tests/test_agent_opencode.py
index 0960c64..3cc7348 100644
--- a/tests/test_agent_opencode.py
+++ b/tests/test_agent_opencode.py
@@ -98,15 +98,24 @@ def test_glm_gets_token_limits(self):
         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}
+        # Probed 2026-07-16: glm-5-2 is 1M context / 65536 output.
+        assert glm["limit"] == {"context": 1_000_000, "output": 65_536}
 
-    def test_non_glm_oss_model_has_no_output_cap(self):
+    def test_kimi_gets_token_limits(self):
+        # kimi is now a capped OSS family (128k context / 65536 output).
         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
+        assert kimi["limit"] == {"context": 128_000, "output": 65_536}
+
+    def test_uncapped_oss_model_has_no_limit(self):
+        # A model outside the limits table gets no `limit` (client default).
+        models = {"oss": ["system.ai.mystery-7b"]}
+        overlay, _ = opencode.render_overlay("system.ai.mystery-7b", "tok", _base_urls(), models)
+        entry = overlay["provider"]["databricks-oss"]["models"]["system.ai.mystery-7b"]
+        assert "limit" not in entry
 
     def test_token_in_api_key(self):
         models = {"anthropic": ["claude-sonnet"]}
diff --git a/tests/test_agent_pi.py b/tests/test_agent_pi.py
index 0afc5fb..c7242cb 100644
--- a/tests/test_agent_pi.py
+++ b/tests/test_agent_pi.py
@@ -17,6 +17,7 @@ def _base_urls() -> dict[str, str]:
         "claude": f"{WS}/ai-gateway/anthropic",
         "openai": f"{WS}/ai-gateway/codex/v1",
         "gemini": f"{WS}/ai-gateway/gemini/v1beta",
+        "oss": f"{WS}/ai-gateway/mlflow/v1",
     }
 
 
@@ -26,6 +27,7 @@ def _empty() -> dict:
         "claude_models": {},
         "codex_models": [],
         "gemini_models": [],
+        "oss_models": [],
     }
 
 
@@ -39,6 +41,7 @@ def _overlay(model: str, token: str = "tok", **kwargs):
         bundle["claude_models"],
         bundle["codex_models"],
         bundle["gemini_models"],
+        bundle["oss_models"],
     )
 
 
@@ -75,23 +78,114 @@ def test_openai_provider_uses_openai_responses(self):
         assert provider["api"] == "openai-responses"
         assert provider["baseUrl"] == f"{WS}/ai-gateway/codex/v1"
 
+    def test_gpt56_sol_model_entry_pins_1m_context(self):
+        # Gateway ids are custom to Pi, so explicit metadata is required to
+        # avoid its 128k custom-model default.
+        overlay, _ = _overlay("gpt-5-6-sol", codex_models=["gpt-5-6-sol"])
+        entry = overlay["providers"]["databricks-openai"]["models"][0]
+        assert entry["id"] == "gpt-5-6-sol"
+        assert entry["contextWindow"] == 1_050_000
+        assert entry["maxTokens"] == 128_000
+        assert entry["reasoning"] is True
+        assert entry["input"] == ["text", "image"]
+
+    def test_gpt_model_entries_use_model_specific_windows(self):
+        overlay, _ = _overlay(
+            "system.ai.gpt-5-2",
+            codex_models=[
+                "system.ai.gpt-5-2",
+                "databricks-gpt-5-4-nano",
+                "databricks-gpt-5-6-sol",
+            ],
+        )
+        windows = {
+            m["id"]: m["contextWindow"] for m in overlay["providers"]["databricks-openai"]["models"]
+        }
+        assert windows == {
+            "system.ai.gpt-5-2": 400_000,
+            "databricks-gpt-5-4-nano": 400_000,
+            "databricks-gpt-5-6-sol": 1_050_000,
+        }
+
+    def test_claude_entries_pin_limits_and_capabilities(self):
+        overlay, _ = _overlay(
+            "databricks-claude-opus-4-8",
+            claude_models={
+                "opus": "databricks-claude-opus-4-8",
+                "sonnet": "system.ai.claude-sonnet-4-5",
+                "haiku": "databricks-claude-haiku-4-5",
+            },
+        )
+        entries = {m["id"]: m for m in overlay["providers"]["databricks-claude"]["models"]}
+        opus = entries["databricks-claude-opus-4-8"]
+        assert opus["contextWindow"] == 1_000_000
+        assert opus["maxTokens"] == 128_000
+        assert opus["reasoning"] is True
+        assert opus["input"] == ["text", "image"]
+        assert opus["compat"] == {"forceAdaptiveThinking": True}
+        assert entries["system.ai.claude-sonnet-4-5"]["contextWindow"] == 1_000_000
+        assert entries["system.ai.claude-sonnet-4-5"]["maxTokens"] == 64_000
+        assert entries["databricks-claude-haiku-4-5"]["contextWindow"] == 200_000
+
     def test_gemini_provider_uses_google_generative_ai(self):
         overlay, _ = _overlay("gemini-2", gemini_models=["gemini-2"])
         provider = overlay["providers"]["databricks-gemini"]
         assert provider["api"] == "google-generative-ai"
         assert provider["baseUrl"] == f"{WS}/ai-gateway/gemini/v1beta"
 
-    def test_all_three_providers_when_all_present(self):
+    def test_mlflow_provider_uses_openai_completions(self):
+        overlay, _ = _overlay("system.ai.glm-5-2", oss_models=["system.ai.glm-5-2"])
+        provider = overlay["providers"]["databricks-mlflow"]
+        assert provider["api"] == "openai-completions"
+        assert provider["baseUrl"] == f"{WS}/ai-gateway/mlflow/v1"
+        assert provider["compat"] == {"supportsStore": False, "supportsStrictMode": False}
+
+    def test_no_mlflow_provider_when_no_oss_models(self):
+        overlay, _ = _overlay("gpt-5", codex_models=["gpt-5"])
+        assert "databricks-mlflow" not in overlay.get("providers", {})
+
+    def test_all_four_providers_when_all_present(self):
         overlay, _ = _overlay(
             "claude-sonnet",
             claude_models={"sonnet": "claude-sonnet"},
             codex_models=["gpt-5"],
             gemini_models=["gemini-2"],
+            oss_models=["system.ai.glm-5-2"],
         )
         assert set(overlay["providers"].keys()) == {
             "databricks-claude",
             "databricks-openai",
             "databricks-gemini",
+            "databricks-mlflow",
+        }
+
+
+class TestRenderOverlayOssEnrichment:
+    """OSS mlflow model entries carry reasoning + contextWindow + maxTokens
+    from the shared databricks.model_token_limits / model_is_reasoning tables."""
+
+    def test_reasoning_model_enriched(self):
+        overlay, _ = _overlay("system.ai.glm-5-2", oss_models=["system.ai.glm-5-2"])
+        entry = overlay["providers"]["databricks-mlflow"]["models"][0]
+        assert entry["id"] == "system.ai.glm-5-2"
+        assert entry["reasoning"] is True
+        assert entry["contextWindow"] == 1_000_000
+        assert entry["maxTokens"] == 65_536
+
+    def test_non_reasoning_model_omits_reasoning(self):
+        overlay, _ = _overlay(
+            "system.ai.llama-4-maverick", oss_models=["system.ai.llama-4-maverick"]
+        )
+        entry = overlay["providers"]["databricks-mlflow"]["models"][0]
+        assert "reasoning" not in entry
+        assert entry["contextWindow"] == 128_000
+        assert entry["maxTokens"] == 8_192
+
+    def test_unknown_oss_model_bare(self):
+        # No limits/reasoning table entry -> only id, client keeps defaults.
+        overlay, _ = _overlay("system.ai.mystery-7b", oss_models=["system.ai.mystery-7b"])
+        assert overlay["providers"]["databricks-mlflow"]["models"][0] == {
+            "id": "system.ai.mystery-7b"
         }
 
 
@@ -193,6 +287,10 @@ def test_prefixes_gemini_model(self):
         overlay, _ = _overlay("gemini-2", gemini_models=["gemini-2"])
         assert overlay["model"] == "databricks-gemini/gemini-2"
 
+    def test_prefixes_oss_model(self):
+        overlay, _ = _overlay("system.ai.glm-5-2", oss_models=["system.ai.glm-5-2"])
+        assert overlay["model"] == "databricks-mlflow/system.ai.glm-5-2"
+
     def test_preserves_already_prefixed_model(self):
         overlay, _ = _overlay(
             "databricks-claude/claude-sonnet",
@@ -228,6 +326,15 @@ def test_falls_back_to_gemini(self):
         state = {"claude_models": {}, "codex_models": [], "gemini_models": ["gemini-2"]}
         assert pi.default_model(state) == "gemini-2"
 
+    def test_falls_back_to_oss_last(self):
+        state = {
+            "claude_models": {},
+            "codex_models": [],
+            "gemini_models": [],
+            "oss_models": ["system.ai.glm-5-2"],
+        }
+        assert pi.default_model(state) == "system.ai.glm-5-2"
+
     def test_returns_none_when_empty(self):
         assert pi.default_model({}) is None
         assert (
diff --git a/tests/test_databricks.py b/tests/test_databricks.py
index f59aa1d..2775c0a 100644
--- a/tests/test_databricks.py
+++ b/tests/test_databricks.py
@@ -142,19 +142,115 @@ def _model_service(model_id: str) -> dict:
 
 class TestModelTokenLimits:
     def test_glm_is_capped(self):
+        # Probed 2026-07-16: glm-5-2 accepts 1M context / 65536 output.
         assert db_mod.model_token_limits("system.ai.glm-5-2") == {
-            "context": 200_000,
-            "output": 25_000,
+            "context": 1_000_000,
+            "output": 65_536,
         }
 
     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,
+            "context": 1_000_000,
+            "output": 65_536,
         }
 
-    def test_uncapped_model_returns_none(self):
-        assert db_mod.model_token_limits("system.ai.kimi-k2-7-code") is None
+    def test_kimi_is_capped(self):
+        assert db_mod.model_token_limits("system.ai.kimi-k2-7-code") == {
+            "context": 128_000,
+            "output": 65_536,
+        }
+
+    def test_llama_lower_output_cap(self):
+        assert db_mod.model_token_limits("system.ai.llama-4-maverick")["output"] == 8_192
+
+    def test_qwen_versions_distinct_caps(self):
+        # qwen35 and qwen3-next have different output ceilings; keyed precisely.
+        assert db_mod.model_token_limits("system.ai.qwen35-122b-a10b")["output"] == 25_000
+        assert (
+            db_mod.model_token_limits("system.ai.qwen3-next-80b-a3b-instruct")["output"] == 10_000
+        )
+
+    def test_oss_chat_model_without_specific_entry_uses_fallback(self):
+        # qwen3-coder is an OSS chat model (matches the `qwen` family) but has
+        # no specific limits entry — it must get the conservative floor, never
+        # None (which would leave it uncapped and 400 on a default request).
+        limits = db_mod.model_token_limits("system.ai.qwen3-coder-480b")
+        assert limits == {"context": 128_000, "output": 8_192}
+
+    def test_non_oss_model_returns_none(self):
+        # deepseek is outside every family (not an OSS chat model) -> no cap.
+        assert db_mod.model_token_limits("system.ai.deepseek-v3") is None
+
+    def test_embedding_model_returns_none_not_fallback(self):
+        # An embeddings model is excluded from OSS chat, so it must NOT get the
+        # chat fallback limits.
+        assert db_mod.model_token_limits("system.ai.qwen3-embedding-0-6b") is None
+
+
+class TestGptModelTokenLimits:
+    def test_gpt56_sol_serves_1m_context_across_id_forms(self):
+        for model_id in (
+            "gpt-5.6-sol",
+            "system.ai.gpt-5-6-sol",
+            "databricks-gpt-5-6-sol",
+            "databricks-openai/gpt-5.6-sol",
+        ):
+            assert db_mod.gpt_model_token_limits(model_id) == {
+                "context": 1_050_000,
+                "output": 128_000,
+            }
+
+    def test_gpt5_windows_are_model_specific(self):
+        assert db_mod.gpt_model_token_limits("system.ai.gpt-5-2")["context"] == 400_000
+        assert db_mod.gpt_model_token_limits("databricks-gpt-5-4")["context"] == 272_000
+        assert db_mod.gpt_model_token_limits("databricks-gpt-5-4-nano")["context"] == 400_000
+        assert db_mod.gpt_model_token_limits("gpt-5.5-pro")["context"] == 1_050_000
+
+    def test_gpt41_window(self):
+        assert db_mod.gpt_model_token_limits("databricks-gpt-4-1") == {
+            "context": 1_047_576,
+            "output": 32_768,
+        }
+
+    def test_unknown_gpt_uses_conservative_floor(self):
+        assert db_mod.gpt_model_token_limits("gpt-6-turbo") == {
+            "context": 128_000,
+            "output": 16_384,
+        }
+
+
+class TestClaudeModelTokenLimits:
+    def test_1m_models_and_output_caps(self):
+        assert db_mod.claude_model_token_limits("databricks-claude-opus-4-8") == {
+            "context": 1_000_000,
+            "output": 128_000,
+        }
+        assert db_mod.claude_model_token_limits("system.ai.claude-sonnet-4-5") == {
+            "context": 1_000_000,
+            "output": 64_000,
+        }
+        assert db_mod.claude_model_token_limits("claude-sonnet-4-6[1m]") == {
+            "context": 1_000_000,
+            "output": 128_000,
+        }
+
+    def test_older_and_unknown_claude_use_200k_floor(self):
+        expected = {"context": 200_000, "output": 64_000}
+        assert db_mod.claude_model_token_limits("claude-opus-4-5") == expected
+        assert db_mod.claude_model_token_limits("claude-haiku-4-5") == expected
+        assert db_mod.claude_model_token_limits("claude-future") == expected
+
+
+class TestModelIsReasoning:
+    def test_reasoning_families(self):
+        assert db_mod.model_is_reasoning("system.ai.glm-5-2") is True
+        assert db_mod.model_is_reasoning("system.ai.inkling") is True
+        assert db_mod.model_is_reasoning("system.ai.qwen35-122b-a10b") is True
+
+    def test_non_reasoning_families(self):
+        assert db_mod.model_is_reasoning("system.ai.llama-4-maverick") is False
+        assert db_mod.model_is_reasoning("system.ai.qwen3-next-80b-a3b-instruct") is False
+        assert db_mod.model_is_reasoning("system.ai.gemma-3-12b") is False
 
 
 class TestDiscoverModelServices:
@@ -187,16 +283,21 @@ 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"
-        # kimi and glm are the allowlisted OSS families; llama is not.
-        assert oss == ["system.ai.glm-5-2", "system.ai.kimi-k2-7-code"]
+        # Full OSS cohort now allowlisted: kimi, glm, AND llama.
+        assert oss == [
+            "system.ai.glm-5-2",
+            "system.ai.kimi-k2-7-code",
+            "system.ai.llama-4-maverick",
+        ]
 
     def test_oss_allowlist_drops_unsupported_families(self, monkeypatch):
-        # Only kimi/glm are allowlisted; other families are dropped.
+        # The cohort covers kimi/glm/qwen/llama/gpt-oss/gemma/inkling; families
+        # outside it (deepseek) and non-chat services (embeddings/rerankers) drop.
         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.qwen35-122b-a10b"),
+                _model_service("system.ai.inkling"),
                 _model_service("system.ai.deepseek-v3"),
                 _model_service("system.ai.gte-large-embed"),
                 _model_service("system.ai.bge-reranker-v2"),
@@ -210,7 +311,50 @@ def test_oss_allowlist_drops_unsupported_families(self, monkeypatch):
 
         assert reason is None
         assert (claude, codex, gemini) == ({}, [], [])
-        assert oss == ["system.ai.glm-5-2", "system.ai.kimi-k2-7-code"]
+        # deepseek/embeddings/rerankers dropped; cohort members kept.
+        assert "system.ai.glm-5-2" in oss
+        assert "system.ai.qwen35-122b-a10b" in oss
+        assert "system.ai.inkling" in oss
+        assert "system.ai.deepseek-v3" not in oss
+        assert "system.ai.gte-large-embed" not in oss
+
+    def test_embedding_model_sharing_family_substring_excluded(self, monkeypatch):
+        # qwen3-embedding matches the `qwen` family but is an embeddings model,
+        # not a chat model — it must not be bucketed as OSS chat.
+        payload = {
+            "model_services": [
+                _model_service("system.ai.qwen35-122b-a10b"),
+                _model_service("system.ai.qwen3-embedding-0-6b"),
+            ]
+        }
+        monkeypatch.setattr(
+            db_mod, "_http_get_json", lambda url, token, timeout=10: (payload, None)
+        )
+
+        _, _, _, oss, _ = db_mod.discover_model_services(WS, "token")
+
+        assert "system.ai.qwen35-122b-a10b" in oss
+        assert "system.ai.qwen3-embedding-0-6b" not in oss
+
+    def test_gpt_oss_is_oss_not_codex(self, monkeypatch):
+        # gpt-oss-* contains "gpt-" but is chat-completions-only, not an
+        # openai-responses codex model. It must bucket into oss ONLY, never
+        # codex (whose openai-responses route would 400 for it).
+        payload = {
+            "model_services": [
+                _model_service("system.ai.gpt-5"),
+                _model_service("system.ai.gpt-oss-120b"),
+            ]
+        }
+        monkeypatch.setattr(
+            db_mod, "_http_get_json", lambda url, token, timeout=10: (payload, None)
+        )
+
+        _, codex, _, oss, _ = db_mod.discover_model_services(WS, "token")
+
+        assert codex == ["system.ai.gpt-5"]
+        assert "system.ai.gpt-oss-120b" in oss
+        assert "system.ai.gpt-oss-120b" not in codex
 
     def test_paginates_via_next_page_token(self, monkeypatch):
         pages = {
@@ -248,7 +392,8 @@ def test_http_failure_returns_reason(self, monkeypatch):
         assert reason == "HTTP 500 Server Error"
 
     def test_no_matching_families_reports_sample(self, monkeypatch):
-        payload = {"model_services": [_model_service("system.ai.llama-4-maverick")]}
+        # deepseek is outside every claude/gpt/gemini/oss family bucket.
+        payload = {"model_services": [_model_service("system.ai.deepseek-v3")]}
         monkeypatch.setattr(
             db_mod, "_http_get_json", lambda url, token, timeout=10: (payload, None)
         )
@@ -256,7 +401,7 @@ def test_no_matching_families_reports_sample(self, monkeypatch):
         claude, codex, gemini, oss, reason = db_mod.discover_model_services(WS, "token")
 
         assert (claude, codex, gemini, oss) == ({}, [], [], [])
-        assert reason is not None and "llama-4-maverick" in reason
+        assert reason is not None and "deepseek-v3" in reason
 
     def test_ignores_non_system_ai_schemas(self, monkeypatch):
         # The metastore listing returns services from every schema; only
diff --git a/tests/test_mlflow_proxy.py b/tests/test_mlflow_proxy.py
new file mode 100644
index 0000000..5aeb16a
--- /dev/null
+++ b/tests/test_mlflow_proxy.py
@@ -0,0 +1,194 @@
+"""Tests for agents/_mlflow_proxy.py — the SSE finish_reason repair proxy."""
+
+from __future__ import annotations
+
+import json
+import threading
+import urllib.error
+import urllib.request
+from http.server import BaseHTTPRequestHandler, HTTPServer
+
+from ucode.agents import _mlflow_proxy
+
+# An inkling-style stream: content deltas then [DONE], NO finish_reason anywhere.
+_STREAM_NO_FINISH = (
+    b'data: {"id":"c1","choices":[{"delta":{"content":"ok"},"index":0}]}\n\ndata: [DONE]\n\n'
+)
+# A well-behaved stream: terminal chunk carries finish_reason.
+_STREAM_WITH_FINISH = (
+    b'data: {"id":"c2","choices":[{"delta":{"content":"ok"},"index":0}]}\n\n'
+    b'data: {"id":"c2","choices":[{"delta":{},"finish_reason":"stop","index":0}]}\n\n'
+    b"data: [DONE]\n\n"
+)
+# An abnormally-terminated stream: content delta, then the upstream just stops —
+# no finish_reason AND no [DONE] (what a mid-stream 429/drop looks like).
+_STREAM_ABRUPT_END = b'data: {"id":"c3","choices":[{"delta":{"content":"ok"},"index":0}]}\n\n'
+# SSE data line with NO space after the colon (spec-legal) and no finish_reason.
+_STREAM_NO_SPACE = (
+    b'data:{"id":"c4","choices":[{"delta":{"content":"ok"},"index":0}]}\n\ndata:[DONE]\n\n'
+)
+# A non-streaming JSON response (some models / non-stream requests).
+_JSON_BODY = b'{"id":"c5","choices":[{"message":{"content":"ok"},"finish_reason":"stop"}]}'
+
+
+def _make_fake_gateway(
+    payload: bytes, content_type: str = "text/event-stream", status: int = 200
+) -> tuple[str, HTTPServer]:
+    class _Fake(BaseHTTPRequestHandler):
+        protocol_version = "HTTP/1.1"
+
+        def do_POST(self):  # noqa: N802
+            self.send_response(status)
+            self.send_header("Content-Type", content_type)
+            self.send_header("Content-Length", str(len(payload)))
+            self.end_headers()
+            self.wfile.write(payload)
+
+        def log_message(self, format: str, *args):
+            pass
+
+    server = HTTPServer(("127.0.0.1", 0), _Fake)
+    threading.Thread(target=server.serve_forever, daemon=True).start()
+    return f"http://127.0.0.1:{server.server_address[1]}", server
+
+
+def _post_through_proxy(upstream: str, path: str = "/ai-gateway/mlflow/v1/chat/completions"):
+    """POST through the proxy; returns (status, body, content_type)."""
+    base, stop = _mlflow_proxy.start(upstream)
+    try:
+        req = urllib.request.Request(
+            f"{base}{path}",
+            data=b'{"model":"databricks-inkling","stream":true}',
+            headers={"Content-Type": "application/json"},
+            method="POST",
+        )
+        try:
+            with urllib.request.urlopen(req, timeout=5) as resp:
+                return resp.status, resp.read().decode("utf-8"), resp.headers.get("Content-Type")
+        except urllib.error.HTTPError as exc:
+            return exc.code, exc.read().decode("utf-8"), exc.headers.get("Content-Type")
+    finally:
+        stop.set()
+
+
+def _body(upstream: str) -> str:
+    """Convenience: body text of a normal streaming POST through the proxy."""
+    return _post_through_proxy(upstream)[1]
+
+
+class TestFinishReasonInjection:
+    def test_injects_finish_reason_when_absent(self):
+        upstream, server = _make_fake_gateway(_STREAM_NO_FINISH)
+        try:
+            out = _body(upstream)
+        finally:
+            server.shutdown()
+        assert '"finish_reason": "stop"' in out or '"finish_reason":"stop"' in out
+        # The injected chunk must come before the terminal [DONE].
+        assert out.index("finish_reason") < out.index("[DONE]")
+        assert out.rstrip().endswith("[DONE]")
+
+    def test_passes_through_when_finish_reason_present(self):
+        upstream, server = _make_fake_gateway(_STREAM_WITH_FINISH)
+        try:
+            out = _body(upstream)
+        finally:
+            server.shutdown()
+        # Exactly one finish_reason — the proxy did not add a second.
+        assert out.count("finish_reason") == 1
+
+    def test_forwards_content_unchanged(self):
+        upstream, server = _make_fake_gateway(_STREAM_NO_FINISH)
+        try:
+            out = _body(upstream)
+        finally:
+            server.shutdown()
+        assert '"content":"ok"' in out
+
+    def test_repairs_abrupt_end_with_finish_and_done(self):
+        # Upstream stopped after content with no finish_reason and no [DONE]
+        # (mid-stream 429/drop). Proxy must synthesize BOTH so a strict client
+        # doesn't error on a truncated stream.
+        upstream, server = _make_fake_gateway(_STREAM_ABRUPT_END)
+        try:
+            out = _body(upstream)
+        finally:
+            server.shutdown()
+        assert "finish_reason" in out
+        assert out.rstrip().endswith("[DONE]")
+        assert '"content":"ok"' in out
+
+    def test_parses_data_line_without_space(self):
+        # SSE allows `data:{...}` (no space). The already-present finish_reason
+        # must be detected so no duplicate is injected.
+        upstream, server = _make_fake_gateway(_STREAM_NO_SPACE)
+        try:
+            out = _body(upstream)
+        finally:
+            server.shutdown()
+        # One synthetic finish (none was present) — not two.
+        assert out.count("finish_reason") == 1
+
+
+class TestNonStreamingAndErrors:
+    def test_non_streaming_json_relayed_unchanged(self):
+        # A non-SSE 200 must pass through verbatim with its content-type — no
+        # finish_reason injection, no SSE relabeling.
+        upstream, server = _make_fake_gateway(_JSON_BODY, content_type="application/json")
+        try:
+            status, body, ctype = _post_through_proxy(upstream)
+        finally:
+            server.shutdown()
+        assert status == 200
+        assert "application/json" in ctype
+        assert body == _JSON_BODY.decode()
+        assert "chat.completion.chunk" not in body  # no injected SSE chunk
+
+    def test_gateway_error_relayed_with_status(self):
+        # A gateway 400 (e.g. max_tokens exceeded) is relayed with its status.
+        err = b'{"error":"max_tokens cannot exceed 8192"}'
+        upstream, server = _make_fake_gateway(err, content_type="application/json", status=400)
+        try:
+            status, body, _ = _post_through_proxy(upstream)
+        finally:
+            server.shutdown()
+        assert status == 400
+        assert "max_tokens" in body
+
+    def test_connection_failure_returns_502(self):
+        # Point the proxy at a dead port: urlopen raises URLError, and the proxy
+        # must return a clean 502, not a dead socket.
+        status, body, _ = _post_through_proxy("http://127.0.0.1:1")
+        assert status == 502
+        assert "upstream unreachable" in body
+
+    def test_disallowed_path_returns_404(self):
+        # Only the mlflow route may be forwarded (SSRF guard).
+        upstream, server = _make_fake_gateway(_JSON_BODY, content_type="application/json")
+        try:
+            status, body, _ = _post_through_proxy(upstream, path="/api/2.0/secrets/get")
+        finally:
+            server.shutdown()
+        assert status == 404
+        assert "not allowed" in body
+
+
+class TestFinishChunk:
+    def test_shape(self):
+
+        chunk = json.loads(_mlflow_proxy._finish_chunk("abc"))
+        assert chunk["id"] == "abc"
+        assert chunk["choices"][0]["finish_reason"] == "stop"
+        assert chunk["object"] == "chat.completion.chunk"
+
+
+class TestStart:
+    def test_returns_local_base_url_and_stop_event(self):
+        upstream, server = _make_fake_gateway(_STREAM_NO_FINISH)
+        try:
+            base, stop = _mlflow_proxy.start(upstream)
+            assert base.startswith("http://127.0.0.1:")
+            assert isinstance(stop, threading.Event)
+            stop.set()
+        finally:
+            server.shutdown()
diff --git a/uv.lock b/uv.lock
index 1f599a0..512d256 100644
--- a/uv.lock
+++ b/uv.lock
@@ -2927,7 +2927,7 @@ wheels = [
 
 [[package]]
 name = "ucode"
-version = "0.1.0"
+version = "0.2.0"
 source = { editable = "." }
 dependencies = [
     { name = "databricks-sql-connector" },