Vision Proxy for Text-Only Models + Context Overflow Fix + Output Popup Fix

CHANGELOG.md

@@ -2,6 +2,17 @@
 
 All notable changes to the **OpenCode Go BYOK Provider** extension are documented here.
 
+## [Unreleased]
+
+### Added
+
+- **`[Vision]` Transparent vision proxy for text-only models (#74).** Run **OpenCode Go: Configure Vision Proxy** from the Command Palette to pick a vision-capable model. When a non-vision OpenCode model receives an image, the extension forwards it to the configured model, receives a text description, and feeds that to the original model — so text-only models "see" images with zero extra steps. The picker shows only vision-capable models (filtered by `models.dev` metadata), with a **None** option to disable and a **Customize prompt** entry to edit the description instruction. No settings to toggle — if a model is configured, the proxy is on. Implemented with `vscode.lm.selectChatModels` + `sendRequest`.
+
+### Fixed
+
+- **`[Streaming]` Context overflow 400 when prompt + output approached the model's limit.** `estimateTokenCount` can underestimate by 0–2%, pushing payloads past the context window on large prompts. Added a 64-token safety margin to `promptReserve` in `modelLimits()`. Affects all models (reported with GLM-5.2).
+- **`[Streaming]` Empty response warning no longer steals focus to the Output pane (#67).** Removed the stray `.show(true)` call on the empty-response diagnostic log.
+
 ## [0.4.0] — 2026-07-13
 
 ### Added
@@ -15,7 +26,6 @@ All notable changes to the **OpenCode Go BYOK Provider** extension are documente
 
 - **`[Usage]` 5-hour rolling usage now isolates per profile in multi-profile mode.** The shared `opencode.db` SQLite database has no API key column, so reading it in multi-account mode mixed quota from all accounts. The tracker now skips SQLite entirely when multiple profiles exist and falls back to extension-tracked entries (which are namespaced per profile). This trades billed-accuracy for correct per-profile isolation.
 
-
 ## [0.3.7] — 2026-07-09
 
 ### Added

README.md

@@ -35,6 +35,7 @@
 | 🌍 **30+ frontier models** | DeepSeek V4, Kimi K2.6, GLM-5.1, Qwen3.7 Max, MiMo V2.5, MiniMax M2.7, Big Pickle, Nemotron — **all in one picker** |
 | 🤖 **Full Agent Mode** | Tool-calling (read files, edit, run terminal) works natively — not just chat. Models also appear in the **Agents window** (Copilot CLI session) |
 | 🧠 **Thinking controls** | Per-model reasoning effort (DeepSeek `max`, Qwen `thinking_budget`, MiniMax `on/off`, Mimo `low/med/high`) |
+| 🖼️ **Vision proxy** | Text-only models can "see" images via a configured vision model. Run **OpenCode Go: Configure Vision Proxy** from the Command Palette to set it up. |
 | 📊 **Live usage tracking** | Status bar shows Go subscription burn-rate across 5h / weekly / monthly tiers |
 | 🔌 **Dual providers** | OpenCode **Go** ($10/mo subscription) + OpenCode **Zen** (free + paid models) — run both at once, switch instantly |
 | 🎯 **Smart routing** | Each model family auto-routes to its native transport (`/responses`, `/messages`, `streamGenerateContent`, `/chat/completions`) |

package.json

@@ -105,6 +105,10 @@
       {
         "command": "opencodego.deleteProfile",
         "title": "OpenCode Go: Delete Profile"
+      },
+      {
+        "command": "opencodego.configureVisionProxy",
+        "title": "OpenCode Go: Configure Vision Proxy"
       }
     ],
     "configuration": {

src/metadata.ts

@@ -235,7 +235,7 @@ const MODELS_WITHOUT_TEMPERATURE = new Set([
   "kimi-k2.7-code",
 ]);
 
-const VISION_CAPABLE_MODELS = new Set([
+export const VISION_CAPABLE_MODELS = new Set([
   "minimax-m2.7",
   "minimax-m2.5",
   "minimax-m2.5-free",

src/streaming.ts

@@ -107,7 +107,8 @@ export async function streamChatCompletions(
     options.output?.appendLine(
       `[warn] empty response from model=${options.modelId} (no text, no tool calls, no reasoning). Try a different free model or enable opencodego.debugReasoning to inspect raw SSE.`,
     );
-    options.output?.show(true);
+    // Intentionally not calling .show(true) — the diagnostic log is
+    // available in the Output pane when the user opens it manually.
   }
 }
 

src/test/metadata.test.ts

@@ -1,6 +1,6 @@
 import { describe, it } from "node:test";
 import assert from "node:assert/strict";
-import { fallbackModelMetadata } from "../metadata.js";
+import { fallbackModelMetadata, VISION_CAPABLE_MODELS } from "../metadata.js";
 import { GO_VENDOR, ZEN_VENDOR } from "../providerTypes.js";
 
 /**
@@ -72,3 +72,25 @@ describe("fallbackModelMetadata — non-kimi models unaffected", () => {
     assert.notEqual(meta?.temperature, false);
   });
 });
+
+describe("VISION_CAPABLE_MODELS", () => {
+  it("includes known vision models (minimax-m2.7, kimi-k2.6, mimo-v2.5)", () => {
+    assert.ok(VISION_CAPABLE_MODELS.has("minimax-m2.7"));
+    assert.ok(VISION_CAPABLE_MODELS.has("kimi-k2.6"));
+    assert.ok(VISION_CAPABLE_MODELS.has("mimo-v2.5"));
+    assert.ok(VISION_CAPABLE_MODELS.has("glm-5.1"));
+    assert.ok(VISION_CAPABLE_MODELS.has("mimo-v2.5-pro"));
+  });
+
+  it("does NOT include text-only models (deepseek-v4-flash, hy3-preview, big-pickle)", () => {
+    assert.ok(!VISION_CAPABLE_MODELS.has("deepseek-v4-flash"));
+    assert.ok(!VISION_CAPABLE_MODELS.has("deepseek-v4-pro"));
+    assert.ok(!VISION_CAPABLE_MODELS.has("hy3-preview"));
+    assert.ok(!VISION_CAPABLE_MODELS.has("big-pickle"));
+  });
+
+  it("is an exported Set", () => {
+    assert.ok(VISION_CAPABLE_MODELS instanceof Set);
+    assert.ok(VISION_CAPABLE_MODELS.size > 10);
+  });
+});

src/test/visionProxy.test.ts

@@ -0,0 +1,90 @@
+import { describe, it, mock } from "node:test";
+import assert from "node:assert/strict";
+
+/**
+ * Vision proxy condition tests.
+ *
+ * The core fix for #74 is caching `metadata.supportsVision` in
+ * `actuallySupportsVision` BEFORE `modelCapabilities` overrides it.
+ * The proxy condition is:
+ *
+ *   hasImageInput && !actuallySupportsVision && visionProxyModelId
+ *
+ * where `actuallySupportsVision` is the RAW model metadata (true =
+ * model natively supports images), NOT the enhanced capabilities.
+ */
+
+type ProxyConditionInput = {
+  hasImageInput: boolean;
+  actuallySupportsVision: boolean;
+  visionProxyModelId: string;
+};
+
+function shouldProxy({ hasImageInput, actuallySupportsVision, visionProxyModelId }: ProxyConditionInput): boolean {
+  return Boolean(hasImageInput && !actuallySupportsVision && visionProxyModelId);
+}
+
+describe("vision proxy condition (shouldProxy)", () => {
+  it("enters proxy when text-only model receives images with proxy configured", () => {
+    assert.ok(shouldProxy({ hasImageInput: true, actuallySupportsVision: false, visionProxyModelId: "gpt-5.5" }));
+  });
+
+  it("skips proxy when no images present", () => {
+    assert.ok(!shouldProxy({ hasImageInput: false, actuallySupportsVision: false, visionProxyModelId: "gpt-5.5" }));
+  });
+
+  it("skips proxy when model natively supports vision", () => {
+    assert.ok(!shouldProxy({ hasImageInput: true, actuallySupportsVision: true, visionProxyModelId: "gpt-5.5" }));
+  });
+
+  it("skips proxy when no vision model is configured (empty string)", () => {
+    assert.ok(!shouldProxy({ hasImageInput: true, actuallySupportsVision: false, visionProxyModelId: "" }));
+  });
+
+  it("skips proxy when all conditions are false", () => {
+    assert.ok(!shouldProxy({ hasImageInput: false, actuallySupportsVision: true, visionProxyModelId: "" }));
+  });
+
+  it("cached supportsVision (actuallySupportsVision) prevents circular regression", () => {
+    // This is the fix for #74: even if modelCapabilities overrides
+    // metadata.supportsVision to true (because proxy is enabled),
+    // the CACHED value (actuallySupportsVision) stays false for
+    // text-only models — so the proxy fires correctly.
+    const textOnlyModel = { hasImageInput: true, actuallySupportsVision: false, visionProxyModelId: "gpt-5.5" };
+    const visionModel = { hasImageInput: true, actuallySupportsVision: true, visionProxyModelId: "gpt-5.5" };
+
+    // Before fix: visionModel.actuallySupportsVision was false → proxy fired
+    // After fix: both behave correctly
+    assert.ok(shouldProxy(textOnlyModel), "text-only model: proxy fires");
+    assert.ok(!shouldProxy(visionModel), "vision model: proxy does NOT fire");
+  });
+});
+
+describe("modelCapabilities vision proxy flag", () => {
+  // modelCapabilities() returns imageInput: true when:
+  //   metadata.supportsVision (native) OR isVisionProxyEnabled()
+  // This tells VS Code NOT to strip images from requests.
+
+  it("returns imageInput: true when proxy is enabled on text-only models", () => {
+    const capabilities = simulateModelCapabilities(false, true);
+    assert.equal(capabilities.imageInput, true);
+  });
+
+  it("returns imageInput: true when model natively supports vision", () => {
+    const capabilities = simulateModelCapabilities(true, false);
+    assert.equal(capabilities.imageInput, true);
+  });
+
+  it("returns imageInput: false only when no vision support and no proxy", () => {
+    const capabilities = simulateModelCapabilities(false, false);
+    assert.equal(capabilities.imageInput, false);
+  });
+});
+
+function simulateModelCapabilities(
+  metadataSupportsVision: boolean,
+  visionProxyEnabled: boolean,
+): { imageInput: boolean } {
+  const supportsVision = metadataSupportsVision || visionProxyEnabled;
+  return { imageInput: supportsVision };
+}