fix(vision-bridge): reroute auto/ prefix to vision model when images present (#7871)

* fix(vision-bridge): reroute auto/ prefix to vision model when images present Rebuilt clean on release/v3.8.49 (the original branch forked from an old main and dragged ~70 unrelated files of base drift). Reconciled with the newer VibeProxy credential guards on the tip: the reroute now also fires for auto/ models, the keep-credentialed-model skip does not apply to auto (keeping auto would land on a text-only candidate), and the reroute-target credential guard is preserved. Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> * test(guardrails): compact image_url literals to fit the 800-line test cap Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com> --------- Co-authored-by: herjarsa <204746071+herjarsa@users.noreply.github.com> Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
2026-07-26 09:52:11 +03:00 · 2026-07-20 20:56:54 +02:00
parent 74204911c7
commit 6fa748e321
2 changed files with 64 additions and 47 deletions
--- a/src/lib/guardrails/visionBridge.ts
+++ b/src/lib/guardrails/visionBridge.ts
@@ -196,37 +196,48 @@ export class VisionBridgeGuardrail extends BaseGuardrail {
      return { block: false };
    }

-    // 3b. Auto/ prefix → skip guardrail (auto-combo resolver handles vision-capable model selection)
-    if (model === "auto" || model.startsWith("auto/")) {
-      return { block: false };
-    }
+    // 3b. Auto/ prefix — don't skip guardrail entirely. Images still need to be
+    // described or rerouted to a vision-capable model. The auto-combo resolver
+    // does NOT currently filter models by vision capability, so without the
+    // guardrail an image-bearing request assigned to a text-only model will
+    // fail upstream with "does not support images".
+    const isAuto = model === "auto" || model.startsWith("auto/");

-    const forceVisionBridge = isVisionBridgeForcedModel(model);
+    // Declare before the conditional so they're available to the rest of preCall
+    let forceVisionBridge = false;
+    let comboVisionBridgeDecision: ComboVisionBridgeDecision | undefined;

-    // 4. Check if model supports vision
-    const capabilities = getResolvedModelCapabilities(model);
-    const comboVisionBridgeDecision = forceVisionBridge
-      ? "process"
-      : this.deps.checkModelHasComboMapping
-        ? (await this.deps.checkModelHasComboMapping(model))
-          ? "process"
-          : "skip"
-        : await getComboVisionBridgeDecision(model);
+    if (!isAuto) {
+      forceVisionBridge = isVisionBridgeForcedModel(model);

-    if (comboVisionBridgeDecision === "skip") {
-      return { block: false };
-    }
+      // 4. Check if model supports vision
+      const capabilities = getResolvedModelCapabilities(model);
+      comboVisionBridgeDecision = forceVisionBridge
+        ? "process"
+        : this.deps.checkModelHasComboMapping
+          ? (await this.deps.checkModelHasComboMapping(model))
+            ? "process"
+            : "skip"
+          : await getComboVisionBridgeDecision(model);

-    if (capabilities.supportsVision === true && !forceVisionBridge) {
-      // The request model supports vision natively, but check if a
-      // model-combo mapping routes this model through a combo where
-      // some targets may NOT support vision. In that case, the vision
-      // bridge must process images so combo targets can describe them.
-      if (comboVisionBridgeDecision !== "process") {
+      if (comboVisionBridgeDecision === "skip") {
        return { block: false };
      }
-      // Combo mapping found — fall through to process images
+
+      if (capabilities?.supportsVision === true && !forceVisionBridge) {
+        // The request model supports vision natively, but check if a
+        // model-combo mapping routes this model through a combo where
+        // some targets may NOT support vision. In that case, the vision
+        // bridge must process images so combo targets can describe them.
+        if (comboVisionBridgeDecision !== "process") {
+          return { block: false };
+        }
+        // Combo mapping found — fall through to process images
+      }
    }
+    // For auto models (isAuto=true), force remains false and combo decision
+    // remains undefined, which makes the reroute check on line ~189 treat it
+    // like a non-combo model — exactly what we want: reroute to a vision model.

    // 5. Get body and check for messages
    const body = payload as Record<string, unknown>;
@@ -266,12 +277,18 @@ export class VisionBridgeGuardrail extends BaseGuardrail {
    // exists, which produced: HTTP log zai → Guardrail reroute → opencode-zen 401
    // "Missing API key" while the combo UI still showed body=zai. Fall through to
    // the image-describe path so the user's chosen model still answers.
-    if (comboVisionBridgeDecision === "not-combo" && !forceVisionBridge) {
+    //
+    // Reroute also fires for the auto/ prefix (isAuto): the auto-combo resolver
+    // does not filter candidates for vision capability, so an image-bearing
+    // request with model=auto would land on a text-only model (#7871). Keeping
+    // "auto" is never the answer there, so the keep-credentialed-model skip
+    // below does not apply to auto — only the reroute-target credential guard.
+    if ((comboVisionBridgeDecision === "not-combo" || isAuto) && !forceVisionBridge) {
      const checkCreds =
        this.deps.hasUsableCredentials ?? hasUsableCredentialsForModel;
      const originalUsable = await checkCreds(model);

-      if (originalUsable === true) {
+      if (originalUsable === true && !isAuto) {
        // Keep the credentialed model; describe images below if needed.
        context.log?.debug?.(
          "VISION_BRIDGE",
--- a/tests/unit/guardrails/visionBridge.test.ts
+++ b/tests/unit/guardrails/visionBridge.test.ts
@@ -107,10 +107,7 @@ test("VB-S05: passthroughs when visionBridgeEnabled is false", async () => {
        role: "user",
        content: [
          { type: "text", text: "What is in this image?" },
-          {
-            type: "image_url",
-            image_url: { url: "https://example.com/image.png" },
-          },
+          { type: "image_url", image_url: { url: "https://example.com/image.png" } },
        ],
      },
    ],
@@ -237,7 +234,7 @@ test("VB-S04: passthroughs when messages array is empty", async () => {

 // ── VB-S12: Auto-prefix skip ────────────────────────────────────────────────

-test("VB-S12: skips guardrail for auto/ prefix model (auto/vision)", async () => {
+test("VB-S12: reroutes auto/ prefix model to vision model (auto/vision)", async () => {
  const guardrail = createGuardrail();

  const payload = createPayload({
@@ -247,10 +244,7 @@ test("VB-S12: skips guardrail for auto/ prefix model (auto/vision)", async () =>
        role: "user",
        content: [
          { type: "text", text: "What is in this image?" },
-          {
-            type: "image_url",
-            image_url: { url: "https://example.com/image.png" },
-          },
+          { type: "image_url", image_url: { url: "https://example.com/image.png" } },
        ],
      },
    ],
@@ -258,11 +252,17 @@ test("VB-S12: skips guardrail for auto/ prefix model (auto/vision)", async () =>

  const result = await guardrail.preCall(payload, createContext({ model: "auto/vision" }));
  assert.strictEqual(result.block, false);
-  assert.strictEqual(result.modifiedPayload, undefined, "auto/vision should passthrough");
-  assert.strictEqual(visionCallCount, 0, "should NOT call vision API for auto prefix");
+  assert.ok(result.modifiedPayload, "auto/vision should reroute to vision model");
+  assert.strictEqual(
+    result.modifiedPayload?.model,
+    "openai/gpt-4o-mini",
+    "should reroute to configured vision model"
+  );
+  assert.strictEqual(result.meta?.rerouted, true, "rerouted meta should be set");
+  assert.strictEqual(visionCallCount, 0, "should NOT call vision API (reroute, not describe)");
 });

-test("VB-S12b: skips guardrail for bare auto prefix", async () => {
+test("VB-S12b: reroutes auto prefix to best vision model when images present", async () => {
  const guardrail = createGuardrail();

  const payload = createPayload({
@@ -272,10 +272,7 @@ test("VB-S12b: skips guardrail for bare auto prefix", async () => {
        role: "user",
        content: [
          { type: "text", text: "What is in this image?" },
-          {
-            type: "image_url",
-            image_url: { url: "https://example.com/image.png" },
-          },
+          { type: "image_url", image_url: { url: "https://example.com/image.png" } },
        ],
      },
    ],
@@ -283,7 +280,13 @@ test("VB-S12b: skips guardrail for bare auto prefix", async () => {

  const result = await guardrail.preCall(payload, createContext({ model: "auto" }));
  assert.strictEqual(result.block, false);
-  assert.strictEqual(result.modifiedPayload, undefined, "auto should passthrough");
+  assert.ok(result.modifiedPayload, "auto should reroute to vision model");
+  assert.strictEqual(
+    result.modifiedPayload?.model,
+    "openai/gpt-4o-mini",
+    "should reroute to configured vision model"
+  );
+  assert.strictEqual(result.meta?.rerouted, true, "rerouted meta should be set");
 });

 // ── VB-S01: Single image → reroute (individual non-vision model) ───────────
@@ -298,10 +301,7 @@ test("VB-S01: reroutes non-vision model with images to best vision model", async
        role: "user",
        content: [
          { type: "text", text: "What is in this image?" },
-          {
-            type: "image_url",
-            image_url: { url: "https://example.com/image.png" },
-          },
+          { type: "image_url", image_url: { url: "https://example.com/image.png" } },
        ],
      },
    ],