From 2a590e74cd063e0aaf0156c5beca377039652b74 Mon Sep 17 00:00:00 2001
From: Kaiss Bouali <k.bouali@outlook.com>
Date: Thu, 9 Apr 2026 13:04:40 -0400
Subject: [PATCH 1/2] feat: implement model auto-downgrade chain

- Rewrite model field in request body when threshold crossed
- Walk downgrade chain to find cheaper alternative
- Add X-Oberwatch-Downgraded response headers
- Set Downgraded and OriginalModel on cost records
---
 internal/proxy/provider_routing.go | 4 ++++
 internal/proxy/proxy.go            | 4 ++++
 2 files changed, 8 insertions(+)

diff --git a/internal/proxy/provider_routing.go b/internal/proxy/provider_routing.go
index ec3d14f..85525f8 100644
--- a/internal/proxy/provider_routing.go
+++ b/internal/proxy/provider_routing.go
@@ -266,6 +266,10 @@ func serveAnthropicOpenAICompat(w http.ResponseWriter, r *http.Request, route re
 					CreatedAt:     time.Now().UTC(),
 				})
 			}
+			if meta.downgraded && meta.originalModel != "" {
+				w.Header().Set("X-Oberwatch-Downgraded", "true")
+				w.Header().Set("X-Oberwatch-Original-Model", meta.originalModel)
+			}
 		}
 	}
 
diff --git a/internal/proxy/proxy.go b/internal/proxy/proxy.go
index 35547a1..a5e14e2 100644
--- a/internal/proxy/proxy.go
+++ b/internal/proxy/proxy.go
@@ -97,6 +97,10 @@ func New(cfg config.Config, hooks Hooks) (*Server, error) {
 			if !ok {
 				return nil
 			}
+			if meta.downgraded && meta.originalModel != "" {
+				response.Header.Set("X-Oberwatch-Downgraded", "true")
+				response.Header.Set("X-Oberwatch-Original-Model", meta.originalModel)
+			}
 			response.Body = newBudgetTrackingBody(
 				response.Body,
 				response.StatusCode,

From e453a36694b96a64865a2eae6d076525afa693ff Mon Sep 17 00:00:00 2001
From: Kaiss Bouali <k.bouali@outlook.com>
Date: Thu, 9 Apr 2026 13:05:44 -0400
Subject: [PATCH 2/2] test: add integration test script for model
 auto-downgrade

- Build binary and start mock upstream server
- Verify X-Oberwatch-Downgraded header present on downgraded request
- Verify X-Oberwatch-Original-Model header contains original model name
---
 scripts/dev/test-downgrade.sh | 155 ++++++++++++++++++++++++++++++++++
 1 file changed, 155 insertions(+)
 create mode 100644 scripts/dev/test-downgrade.sh

diff --git a/scripts/dev/test-downgrade.sh b/scripts/dev/test-downgrade.sh
new file mode 100644
index 0000000..470febd
--- /dev/null
+++ b/scripts/dev/test-downgrade.sh
@@ -0,0 +1,155 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+export PATH=/usr/local/go/bin:$PATH
+export GOCACHE=/tmp/go-build
+
+echo "==> Testing model auto-downgrade chain"
+
+TEMP_DIR=$(mktemp -d)
+trap 'rm -rf "$TEMP_DIR"' EXIT
+
+CONFIG_FILE="$TEMP_DIR/oberwatch.toml"
+DB_FILE="$TEMP_DIR/oberwatch.db"
+BINARY="$TEMP_DIR/oberwatch"
+
+# Build binary first
+echo "==> Building oberwatch binary..."
+go build -o "$BINARY" ./cmd/oberwatch
+
+# Strategy: set gpt-4o pricing at $1,000,000/M so even 1 estimated token = $1.00.
+# Budget limit = $10. Threshold = 50% = $5.
+# First request body is ~50 bytes ≈ 12 tokens estimated → cost estimate ~$12 > 50% threshold.
+# This means the FIRST real request will trigger downgrade immediately.
+# We verify the downgrade headers are present.
+cat > "$CONFIG_FILE" <<EOF
+[server]
+port = 18080
+admin_token = "test-token"
+
+[upstream.openai]
+base_url = "http://localhost:18081"
+
+[upstream.anthropic]
+base_url = "http://localhost:18082"
+
+[trace]
+sqlite_path = "$DB_FILE"
+
+[gate]
+enabled = true
+downgrade_threshold_pct = 50.0
+
+[gate.default_budget]
+limit_usd = 10.0
+period = "hourly"
+action_on_exceed = "reject"
+
+[[gate.agents]]
+name = "test-agent"
+limit_usd = 10.0
+period = "hourly"
+action_on_exceed = "downgrade"
+downgrade_chain = ["gpt-4o", "gpt-4o-mini"]
+
+[[pricing]]
+model = "gpt-4o"
+input_per_million = 1000000.0
+output_per_million = 1000000.0
+
+[[pricing]]
+model = "gpt-4o-mini"
+input_per_million = 100000.0
+output_per_million = 100000.0
+EOF
+
+# Start mock upstream that returns usage
+python3 -c '
+import http.server
+import socketserver
+import json
+
+class Handler(http.server.BaseHTTPRequestHandler):
+    def do_POST(self):
+        content_length = int(self.headers["Content-Length"])
+        body = self.rfile.read(content_length)
+        
+        # Return response with usage
+        response = {
+            "id": "test-123",
+            "object": "chat.completion",
+            "created": 1234567890,
+            "model": "gpt-4o",
+            "choices": [{
+                "index": 0,
+                "message": {"role": "assistant", "content": "Hello"},
+                "finish_reason": "stop"
+            }],
+            "usage": {
+                "prompt_tokens": 100,
+                "completion_tokens": 50,
+                "total_tokens": 150
+            }
+        }
+        
+        self.send_response(200)
+        self.send_header("Content-Type", "application/json")
+        self.end_headers()
+        self.wfile.write(json.dumps(response).encode())
+    
+    def log_message(self, format, *args):
+        pass
+
+with socketserver.TCPServer(("", 18081), Handler) as httpd:
+    httpd.serve_forever()
+' &
+MOCK_PID=$!
+trap 'kill $MOCK_PID 2>/dev/null || true; rm -rf "$TEMP_DIR"' EXIT
+
+# Wait for mock server
+sleep 1
+
+# Start oberwatch
+"$BINARY" serve --config "$CONFIG_FILE" &
+OBERWATCH_PID=$!
+trap 'kill $OBERWATCH_PID 2>/dev/null || true; kill $MOCK_PID 2>/dev/null || true; rm -rf "$TEMP_DIR"' EXIT
+
+# Wait for oberwatch to start
+sleep 2
+
+# Verify oberwatch is up
+if ! curl -sf http://localhost:18080/_oberwatch/api/v1/health > /dev/null 2>&1; then
+  echo "FAIL: Oberwatch did not start"
+  exit 1
+fi
+echo "✓ Oberwatch is running"
+
+# Send request — pricing is $1M/M so even body byte estimate crosses 50% of $10 threshold.
+# Gate checks estimated cost before forwarding → ActionDowngrade → rewrite model.
+RESPONSE2=$(curl -s -i -X POST http://localhost:18080/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -H "X-Oberwatch-Agent: test-agent" \
+  -H "Authorization: Bearer sk-test" \
+  -d '{"model": "gpt-4o", "messages": [{"role": "user", "content": "hello"}]}')
+
+echo "--- Response headers ---"
+echo "$RESPONSE2" | head -20
+echo "---"
+
+# Check for downgrade headers
+if ! echo "$RESPONSE2" | grep -qi "X-Oberwatch-Downgraded: true"; then
+  echo "FAIL: Second request should have X-Oberwatch-Downgraded: true header"
+  echo "--- Response ---"
+  echo "$RESPONSE2"
+  exit 1
+fi
+
+if ! echo "$RESPONSE2" | grep -qi "X-Oberwatch-Original-Model: gpt-4o"; then
+  echo "FAIL: Second request should have X-Oberwatch-Original-Model: gpt-4o header"
+  echo "--- Response ---"
+  echo "$RESPONSE2"
+  exit 1
+fi
+
+echo "✓ Second request triggered downgrade with correct headers"
+echo "==> All downgrade tests passed!"