From f7a7af4e11a4817ae342b29ddda7593d26e3aaf0 Mon Sep 17 00:00:00 2001
From: Craig Allan-McWilliams <craig@epic28.com>
Date: Tue, 16 Jun 2026 09:01:27 +0100
Subject: [PATCH] fix(model): default synthetic models to text-only input to
 avoid image 400s

Synthetic models (built for ids not in the pi-ai registry, e.g. deepseek-v4-pro
via Ollama) hard-coded `input: ['text', 'image']`, falsely claiming vision
support. Because the model advertised image input, the openai-completions
provider did not filter image content, so screenshots from the GUI/computer-use
tools were sent to text-only endpoints. Ollama rejects these with HTTP 400
"this model does not support image input", surfaced to users as an opaque
"invalid message format" error.

Default synthetic models to text-only input. Vision-capable models resolved from
the pi-ai registry keep their real modalities; only synthetic fallbacks change.
For a custom vision endpoint resolved as synthetic this drops images gracefully
instead of hard-failing the whole request.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---
 src/main/claude/pi-model-resolution.ts | 10 ++++++++-
 tests/synthetic-model-input.test.ts    | 29 ++++++++++++++++++++++++++
 2 files changed, 38 insertions(+), 1 deletion(-)
 create mode 100644 tests/synthetic-model-input.test.ts

diff --git a/src/main/claude/pi-model-resolution.ts b/src/main/claude/pi-model-resolution.ts
index ca6259996..f1b7ffcf3 100644
--- a/src/main/claude/pi-model-resolution.ts
+++ b/src/main/claude/pi-model-resolution.ts
@@ -139,7 +139,15 @@ export function buildSyntheticPiModel(
     provider,
     baseUrl: baseUrl || '',
     reasoning: autoReasoning,
-    input: ['text', 'image'],
+    // Default unknown/synthetic models to text-only input. We cannot know whether
+    // an arbitrary model supports image input, and falsely claiming vision support
+    // causes hard request failures rather than gracefully dropping images — e.g.
+    // Ollama returns HTTP 400 "this model does not support image input" for a
+    // text-only model like deepseek-v4-pro, which surfaces to the user as an
+    // opaque "invalid message format" error. Vision-capable models resolved from
+    // the pi-ai registry keep their real input modalities; only synthetic
+    // fallbacks are affected here.
+    input: ['text'],
     cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
     contextWindow: contextWindow || knownSpecs?.contextWindow || 128000,
     maxTokens: maxTokens || knownSpecs?.maxTokens || 16384,
diff --git a/tests/synthetic-model-input.test.ts b/tests/synthetic-model-input.test.ts
new file mode 100644
index 000000000..c9c066d62
--- /dev/null
+++ b/tests/synthetic-model-input.test.ts
@@ -0,0 +1,29 @@
+import { describe, it, expect } from 'vitest';
+import { buildSyntheticPiModel } from '../src/main/claude/pi-model-resolution';
+
+/**
+ * Regression guard for the text-only-model image bug.
+ *
+ * Synthetic models (built for ids not in the pi-ai registry) must NOT claim
+ * image input. Falsely advertising vision support means image content — e.g.
+ * screenshots from the GUI/computer-use tools — is sent to text-only endpoints,
+ * which hard-fail instead of gracefully dropping the images. For example Ollama
+ * returns HTTP 400 "this model does not support image input" for a text-only
+ * model like deepseek-v4-pro, which the app surfaces as an opaque
+ * "invalid message format" error.
+ */
+describe('buildSyntheticPiModel input modalities', () => {
+  it('defaults an unknown synthetic model to text-only input', () => {
+    const m = buildSyntheticPiModel('deepseek-v4-pro', 'openai', 'openai', 'https://ollama.com/v1');
+    expect(m.input).toEqual(['text']);
+    expect(m.input).not.toContain('image');
+  });
+
+  it('still resolves other core model fields normally', () => {
+    const m = buildSyntheticPiModel('some-custom-model', 'openai', 'openai', 'https://example.com/v1');
+    expect(m.id).toBe('some-custom-model');
+    expect(m.api).toBeTruthy();
+    expect(m.contextWindow).toBeGreaterThan(0);
+    expect(m.maxTokens).toBeGreaterThan(0);
+  });
+});