feat(vision): cache vision-proxy image descriptions

README.md

@@ -339,6 +339,7 @@ Provider diagnostics also include the VS Code/extension versions, extension host
 | `opencodego.streamIdleTimeoutSeconds`     | `120`    | Cancel if stream goes idle                                                       |
 | `opencodego.showUsageStatusBar`           | `true`   | Show usage summary in status bar                                                 |
 | `opencodego.showProviderPrefix`           | `true`   | Include `OpenCode Go` / `OpenCode Zen` in model names                            |
+| `opencodego.visionProxyWholeConversation` | `false`  | Vision proxy: describe the whole conversation instead of only the message with a new image (more context, more tokens) |
 | `opencodego.freeOnly`                     | `true`   | Zen: free models only. `false` = include paid                                    |
 | `opencodego.agentsWindow`                 | `true`   | Expose agent-host model variants (`targetChatSessionType`) for the Agents window |
 | `opencodego.showAgentModelsInManagePanel` | `false`  | Show agent vendors in Manage Language Models panel                               |

docs/features/11-20260715-vision-proxy.md

@@ -79,6 +79,20 @@ Storage keys (in extension `globalState`):
 - `opencodego.visionProxyModelId` — target Copilot model ID (e.g. `copilot:gpt-5.5`)
 - `opencodego.visionProxyPrompt` — description instruction sent to the vision model
 
+### Description Cache & Whole-Conversation Mode
+
+Descriptions produced by the proxy are cached per image (SHA-256 of its bytes), so
+images already described in earlier turns are reused without calling the vision model
+again - saving Copilot quota and latency in multi-turn conversations (issue #119).
+
+A boolean setting controls how much context each description gets:
+
+- `opencodego.visionProxyWholeConversation` (default `false`). Off: the proxy
+  describes only the message that contains a new image and reuses cached
+  descriptions, keeping token usage low. On: the proxy sends the whole
+  conversation to the vision model so descriptions carry full context, at the
+  cost of more tokens. Descriptions are still cached in both modes.
+
 ### Graceful Fallback
 
 If the proxy fails (model not found, API error, empty description), images are stripped with a placeholder `[Image unavailable — vision proxy unavailable]` instead of forwarding raw image data to a text-only model (which would 400). Original text parts are preserved.

package.json

@@ -176,6 +176,11 @@
           "default": true,
           "description": "Show the OpenCode Go or OpenCode Zen provider prefix in model names."
         },
+        "opencodego.visionProxyWholeConversation": {
+          "type": "boolean",
+          "default": false,
+          "description": "Off (default): describe only the message with a new image and reuse cached image descriptions to save quota. On: describe the whole conversation for richer context (uses more tokens; descriptions are still cached)."
+        },
         "opencodego.freeOnly": {
           "type": "boolean",
           "default": true,