feat: firewalled npm install path, gateway kill-switch, and model-catalog updates

app.yaml

@@ -30,6 +30,18 @@ env:
     value: "true"
   - name: CLAUDE_CODE_DISABLE_AUTO_MEMORY
     value: 0
+  # Firewalled network: install Claude Code from the npm registry instead of the
+  # claude.ai installer host. Same CLI (@anthropic-ai/claude-code); use when npm
+  # is allowlisted but claude.ai / its install CDN is blocked. Default (unset)
+  # keeps the curl claude.ai path.
+  - name: CLAUDE_INSTALL_METHOD
+    value: npm
+  # This workspace has no AI Gateway — force the legacy serving-endpoints path.
+  # Empty value skips the gateway probe and routes model calls to
+  # {DATABRICKS_HOST}/serving-endpoints. Do NOT set a non-empty value here unless
+  # the gateway actually exists (it's trusted with no reachability probe).
+  - name: DATABRICKS_GATEWAY_HOST
+    value: ""
   # Set CLAUDE_CODE_OTEL_ENABLED=true to send Claude Code OTEL signals to Unity
   # Catalog. Off by default; requires CLAUDE_CODE_OTEL_CATALOG_SCHEMA pointing at
   # a <catalog>.<schema> your app SP can write to.

setup_claude.py

@@ -262,6 +262,52 @@ def main():
 
 if os.environ.get("CODA_SKIP_CLAUDE_INSTALL", "").lower() == "true":
     print("Claude Code CLI install skipped")
+elif os.environ.get("CLAUDE_INSTALL_METHOD", "").strip().lower() == "npm":
+    # npm install path for firewalled networks where the claude.ai installer
+    # host (or the CDN its install.sh pulls from) is blocked but the npm
+    # registry is reachable. @anthropic-ai/claude-code is the same CLI as the
+    # curl installer produces. Mirrors setup_pi.py's hardened pattern:
+    # version cooldown (get_npm_version), NPM_REGISTRY override (npm_env),
+    # retries, and loud stderr. Unlike setup_pi.py we do NOT pass
+    # --ignore-scripts: Claude Code's postinstall (node install.cjs) is what
+    # places the native binary, so skipping scripts yields no working `claude`.
+    from utils import get_npm_version
+    from enterprise_config import npm_env
+
+    CLAUDE_PACKAGE = "@anthropic-ai/claude-code"
+    npm_prefix = str(home / ".local")
+    claude_version = get_npm_version(CLAUDE_PACKAGE)
+    claude_pkg = (
+        f"{CLAUDE_PACKAGE}@{claude_version}" if claude_version
+        else f"{CLAUDE_PACKAGE}@latest"
+    )
+
+    MAX_RETRIES = 3
+    RETRY_DELAY = 5  # seconds
+    for attempt in range(1, MAX_RETRIES + 1):
+        print(f"Installing {claude_pkg} via npm (attempt {attempt}/{MAX_RETRIES})...")
+        result = subprocess.run(
+            ["npm", "install", "-g", f"--prefix={npm_prefix}", claude_pkg],
+            capture_output=True, text=True,
+            env={**os.environ, "HOME": str(home), **npm_env()},
+        )
+        if result.returncode == 0 and claude_bin.exists():
+            print(f"Claude Code CLI installed to {claude_bin}")
+            break
+        else:
+            stderr = result.stderr.strip()
+            print(f"Claude Code npm install failed (attempt {attempt}/{MAX_RETRIES}, rc={result.returncode})")
+            if stderr:
+                print(f"  stderr: {stderr[:500]}")
+            if result.stdout.strip():
+                print(f"  stdout: {result.stdout.strip()[:500]}")
+            if attempt < MAX_RETRIES:
+                import time
+                print(f"  Retrying in {RETRY_DELAY}s...")
+                time.sleep(RETRY_DELAY)
+            else:
+                print(f"ERROR: Claude Code npm install failed after {MAX_RETRIES} attempts. "
+                      f"Run manually: npm install -g --prefix=$HOME/.local {CLAUDE_PACKAGE}")
 else:
     # Honour CLAUDE_INSTALLER_URL for enterprise environments where claude.ai is
     # firewalled — defaults to the public installer when unset. The URL is

setup_hermes.py

@@ -130,14 +130,12 @@ def _run(cmd, **kwargs):
 
 model_catalog = [
     "databricks-claude-opus-4-8",
-    "databricks-claude-opus-4-6",
     "databricks-claude-sonnet-4-6",
     "databricks-claude-haiku-4-5",
     "databricks-gpt-5-3-codex",
     "databricks-gpt-5-1-codex-max",
     "databricks-gemini-2-5-flash",
     "databricks-gemini-2-5-pro",
-    "databricks-gemini-2-5-pro",
 ]
 
 lines = []

setup_opencode.py

@@ -15,6 +15,7 @@
 from pathlib import Path
 
 from utils import ensure_https, get_gateway_host, get_npm_version
+from enterprise_config import npm_env
 
 # Opt-out: allow operators to disable OpenCode bundling without removing the file.
 if os.environ.get("ENABLE_OPENCODE", "true").strip().lower() in ("false", "0", "no"):
@@ -55,7 +56,7 @@
         result = subprocess.run(
             ["npm", "install", "-g", f"--prefix={npm_prefix}", oc_pkg],
             capture_output=True, text=True,
-            env={**os.environ, "HOME": str(home)}
+            env={**os.environ, "HOME": str(home), **npm_env()}
         )
         if result.returncode == 0 and opencode_bin.exists():
             print(f"OpenCode CLI installed to {opencode_bin}")
@@ -82,7 +83,7 @@
     result = subprocess.run(
         ["npm", "install", "-g", f"--prefix={npm_prefix}", sdk_pkg],
         capture_output=True, text=True,
-        env={**os.environ, "HOME": str(home)}
+        env={**os.environ, "HOME": str(home), **npm_env()}
     )
     if result.returncode == 0:
         print(f"@ai-sdk/openai@{sdk_version or 'latest'} installed (Responses API support)")
@@ -150,17 +151,17 @@
                     "apiKey": "{env:DATABRICKS_TOKEN}"
                 },
                 "models": {
-                    "databricks-claude-opus-4-6": {
-                        "name": "Claude Opus 4.6 (Databricks)",
+                    "databricks-claude-opus-4-8": {
+                        "name": "Claude Opus 4.8 (Databricks)",
                         "limit": {
-                            "context": 200000,
+                            "context": 1000000,
                             "output": 16384
                         }
                     },
                     "databricks-claude-sonnet-4-6": {
                         "name": "Claude Sonnet 4.6 (Databricks)",
                         "limit": {
-                            "context": 200000,
+                            "context": 1000000,
                             "output": 8192
                         }
                     },
@@ -178,13 +179,6 @@
                             "output": 8192
                         }
                     },
-                    "databricks-gemini-2-5-pro": {
-                        "name": "Gemini 2.5 Pro (Databricks)",
-                        "limit": {
-                            "context": 1000000,
-                            "output": 8192
-                        }
-                    },
                 }
             },
             "databricks-openai": {
@@ -230,17 +224,17 @@
                     "apiKey": "{env:DATABRICKS_TOKEN}"
                 },
                 "models": {
-                    "databricks-claude-opus-4-6": {
-                        "name": "Claude Opus 4.6 (Databricks)",
+                    "databricks-claude-opus-4-8": {
+                        "name": "Claude Opus 4.8 (Databricks)",
                         "limit": {
-                            "context": 200000,
+                            "context": 1000000,
                             "output": 16384
                         }
                     },
                     "databricks-claude-sonnet-4-6": {
                         "name": "Claude Sonnet 4.6 (Databricks)",
                         "limit": {
-                            "context": 200000,
+                            "context": 1000000,
                             "output": 8192
                         }
                     },
@@ -258,13 +252,6 @@
                             "output": 8192
                         }
                     },
-                    "databricks-gemini-2-5-pro": {
-                        "name": "Gemini 2.5 Pro (Databricks)",
-                        "limit": {
-                            "context": 1000000,
-                            "output": 8192
-                        }
-                    },
                 }
             }
         },

setup_pi.py

@@ -149,7 +149,13 @@
     "apiKey": auth_token,
     "authHeader": True,
     "compat": {"supportsEagerToolInputStreaming": False},
-    "models": [{"id": active_model}],
+    # contextWindow is explicit: Pi defaults a custom provider's model to 131072
+    # (128K) when absent. The Databricks FMAPI-served Claude model has a ~1.05M
+    # total-token context window (per Databricks Foundation Model APIs supported-
+    # models docs), so the 128K default badly under-uses it. Set the real window.
+    # NB: the FMAPI *rate* limits are separate (e.g. ~200k input tokens/minute on
+    # the default tier) — that's throughput, not context, and is not fixed here.
+    "models": [{"id": active_model, "contextWindow": 1000000}],
 }
 
 models_path.write_text(json.dumps(config, indent=2))

tests/test_gateway_discovery.py

@@ -115,16 +115,25 @@ def test_empty_when_both_blank(self):
 
     @mock.patch("utils._probe_gateway", return_value=True)
     @mock.patch.dict(os.environ, {"DATABRICKS_GATEWAY_HOST": "  ", "DATABRICKS_WORKSPACE_ID": "12345"})
-    def test_whitespace_only_gateway_falls_through(self, mock_probe):
-        """Whitespace-only DATABRICKS_GATEWAY_HOST falls through to workspace ID."""
+    def test_empty_gateway_host_forces_off(self, mock_probe):
+        """An explicitly-set-but-empty DATABRICKS_GATEWAY_HOST forces the gateway
+        OFF (returns "") — it does NOT fall through to workspace-ID derivation.
+        This is the operator kill-switch: setting it empty on an Azure host must
+        route model calls to serving-endpoints, not to a derived ai-gateway URL.
+        """
         os.environ.pop("_GATEWAY_RESOLVED", None)
-        assert self._get_fn()() == "https://12345.ai-gateway.cloud.databricks.com"
+        assert self._get_fn()() == ""
 
     @mock.patch("utils._probe_gateway", return_value=True)
-    @mock.patch.dict(os.environ, {"DATABRICKS_GATEWAY_HOST": "", "DATABRICKS_WORKSPACE_ID": " 99999 "})
+    @mock.patch.dict(os.environ, {"DATABRICKS_WORKSPACE_ID": " 99999 "})
     def test_workspace_id_whitespace_stripped(self, mock_probe):
-        """Leading/trailing whitespace in workspace ID is stripped."""
+        """Leading/trailing whitespace in workspace ID is stripped (tier 2).
+
+        DATABRICKS_GATEWAY_HOST is left unset so derivation runs — setting it
+        empty would now force the gateway off before tier 2 is reached.
+        """
         os.environ.pop("_GATEWAY_RESOLVED", None)
+        os.environ.pop("DATABRICKS_GATEWAY_HOST", None)
         assert self._get_fn()() == "https://99999.ai-gateway.cloud.databricks.com"
 
 

utils.py

@@ -262,10 +262,17 @@ def get_gateway_host() -> str:
     if resolved is not None:
         return resolved
 
-    # Tier 1: explicit override (trusted, no probe)
-    explicit = os.environ.get("DATABRICKS_GATEWAY_HOST", "").strip().rstrip("/")
-    if explicit:
-        return ensure_https(explicit)
+    # Tier 1: explicit override. Presence of the env var is authoritative:
+    #   - set to a URL  -> trust it, no probe.
+    #   - set but empty  -> operator is explicitly disabling the gateway; return
+    #     "" so callers use serving-endpoints. Do NOT fall through to tier 2,
+    #     which would otherwise DERIVE a gateway URL from an Azure host
+    #     (adb-{id}.…azuredatabricks.net) and probe it — reintroducing a gateway
+    #     the operator meant to turn off.
+    #   - unset entirely -> fall through to tier 2 auto-derivation.
+    if "DATABRICKS_GATEWAY_HOST" in os.environ:
+        explicit = os.environ["DATABRICKS_GATEWAY_HOST"].strip().rstrip("/")
+        return ensure_https(explicit) if explicit else ""
 
     # Tier 2: auto-construct from workspace ID and probe for reachability
     host = os.environ.get("DATABRICKS_HOST", "")