[GH-ISSUE #228] This plugin use my github copilot tikens too!! #82

New issue

Open

opened 2026-10-26 09:45:08 +00:00 by bsakaguchi · 3 comments

bsakaguchi commented

2026-10-26 09:45:08 +00:00

Owner

Originally created by @mehdijamshidi9 on GitHub (Jun 17, 2026).
Original GitHub issue: https://github.com/deepseek-ai/awesome-deepseek-agent/issues/228

i install this pluging for use deepseek v4pro model into github copilot. It`s work but after update vscode yesterday, all of my remaining github copilot max tokens gone!
when i use deepseek model into github copilot, both of them use tokens: deepseek and copilot

Originally created by @mehdijamshidi9 on GitHub (Jun 17, 2026). Original GitHub issue: https://github.com/deepseek-ai/awesome-deepseek-agent/issues/228 i install this pluging for use deepseek v4pro model into github copilot. It`s work but after update vscode yesterday, all of my remaining github copilot max tokens gone! when i use deepseek model into github copilot, both of them use tokens: deepseek and copilot

bsakaguchi commented

2026-10-26 09:45:10 +00:00

Author

Owner

@Kletternaut commented on GitHub (Jun 17, 2026):

I can confirm that too. I noticed it as well, but about 10 days ago already. I’ve also experienced the model repeatedly reverting automatically to GHCP models. (VSCode, not Insiders)

However, since I cancelled my GHCP Pro+ subscription and was downgraded to GHCP Free, the DeepSeek integration still works without issues (VSCode 1.124.2)—even with the free credit exhausted (chat quota reached).

A brief but important note:
How to Force High/Max Reasoning in VS Code (BYOK)VS Code often resets custom models to medium by default. You can bypass this restriction by naming the model ID after Anthropic's claude-opus, which VS Code naturally allows to use higher reasoning levels. The DeepSeek API will catch this name and route it to deepseek-v4-pro on their server.Step 1: Update your model configurationOpen your model settings file in VS Code and replace your deepseek-v4-pro block with the following configuration. By removing "low" and "medium" from the list, the UI is forced to stay on higher tiers:

.config/Code/User/chatLanguageModels.json:

    {
            "name": "DeepSeek Official",
            "vendor": "customendpoint",
            "apiKey": "${input:chat.lm.secret.xxxxx}",
            "apiType": "messages",
            "models": [
                    {
                            "id": "deepseek-chat",
                            "name": "DeepSeek V3",
                            "url": "https://api.deepseek.com/anthropic",
                            "toolCalling": true,
                            "vision": true,
                            "maxInputTokens": 128000,
                            "maxOutputTokens": 8192
                    },
                    {
                            "id": "claude-opus-deepseek",
                            "name": "DeepSeek V4 Pro",
                            "url": "https://api.deepseek.com/anthropic",
                            "toolCalling": true,
                            "vision": true,
                            "thinking": true,
                            "supportsReasoningEffort": [
                                    "high",
                                    "max"
                            ],
                            "includeReasoning": true,
                            "maxInputTokens": 128000,
                            "maxOutputTokens": 16384
                    }
            ]
    }

@Kletternaut commented on GitHub (Jun 17, 2026): I can confirm that too. I noticed it as well, but about 10 days ago already. I’ve also experienced the model repeatedly[ reverting automatically to GHCP models](https://www.reddit.com/r/GithubCopilot/comments/1sucfpo/thinking_mode_broken_again_after_vs_code_update/). (VSCode, not Insiders) However, since I cancelled my GHCP Pro+ subscription and was downgraded to GHCP Free, the DeepSeek integration still works without issues (VSCode 1.124.2)—even with the free credit exhausted (chat quota reached). A brief but important note: How to Force High/Max Reasoning in VS Code (BYOK)VS Code often resets custom models to medium by default. You can bypass this restriction by naming the model ID after Anthropic's claude-opus, which VS Code naturally allows to use higher reasoning levels. The DeepSeek API will catch this name and route it to deepseek-v4-pro on their server.Step 1: Update your model configurationOpen your model settings file in VS Code and replace your deepseek-v4-pro block with the following configuration. By removing "low" and "medium" from the list, the UI is forced to stay on higher tiers: .config/Code/User/chatLanguageModels.json: { "name": "DeepSeek Official", "vendor": "customendpoint", "apiKey": "${input:chat.lm.secret.xxxxx}", "apiType": "messages", "models": [ { "id": "deepseek-chat", "name": "DeepSeek V3", "url": "https://api.deepseek.com/anthropic", "toolCalling": true, "vision": true, "maxInputTokens": 128000, "maxOutputTokens": 8192 }, { "id": "claude-opus-deepseek", "name": "DeepSeek V4 Pro", "url": "https://api.deepseek.com/anthropic", "toolCalling": true, "vision": true, "thinking": true, "supportsReasoningEffort": [ "high", "max" ], "includeReasoning": true, "maxInputTokens": 128000, "maxOutputTokens": 16384 } ] }

bsakaguchi commented

2026-10-26 09:45:10 +00:00

Author

Owner

@Kletternaut commented on GitHub (Jun 17, 2026):

Oh, sorry, I just noticed you're using a plugin. That's not the case for me; I use GHCP in VS Code Copilot Chat with bjok. (api key)

@Kletternaut commented on GitHub (Jun 17, 2026): Oh, sorry, I just noticed you're using a plugin. That's not the case for me; I use GHCP in VS Code Copilot Chat with bjok. (api key)

bsakaguchi commented

2026-10-26 09:45:10 +00:00

Author

Owner

@tianhao8687 commented on GitHub (Aug 17, 2026):

There are two separate paths that can consume Copilot quota here:

Vision proxying. The extension's own documentation says that image attachments are sent to another installed Copilot model (Claude/GPT-4o/etc.) for description before the text is sent to DeepSeek. That delegated call can consume Copilot quota.
Fallback/model reset after a VS Code update. The extension uses non-public Copilot provider APIs and explicitly warns that newer VS Code releases may break them. If the picker silently falls back to a GitHub-hosted model, those turns can be billed by Copilot while other turns still hit DeepSeek.

A clean isolation test would be:

start a new text-only chat with no image attachment;
explicitly select DeepSeek (not Auto) for that new chat;
set deepseek-copilot.debugMode to metadata and confirm the turn appears in the DeepSeek usage/log output;
compare the Copilot premium-request counter before and after exactly one turn.

If only image turns consume Copilot quota, the vision proxy explains it. If a text-only DeepSeek turn also decrements the counter, this is likely a VS Code/extension regression. Please report it to the extension's issue tracker with the VS Code version, extension version, selected model, and before/after counters. Do not include API keys or verbose request dumps containing private prompts.

@tianhao8687 commented on GitHub (Aug 17, 2026): There are two separate paths that can consume Copilot quota here: 1. **Vision proxying.** The extension's own documentation says that image attachments are sent to another installed Copilot model (Claude/GPT-4o/etc.) for description before the text is sent to DeepSeek. That delegated call can consume Copilot quota. 2. **Fallback/model reset after a VS Code update.** The extension uses non-public Copilot provider APIs and explicitly warns that newer VS Code releases may break them. If the picker silently falls back to a GitHub-hosted model, those turns can be billed by Copilot while other turns still hit DeepSeek. A clean isolation test would be: - start a new **text-only** chat with no image attachment; - explicitly select DeepSeek (not Auto) for that new chat; - set `deepseek-copilot.debugMode` to `metadata` and confirm the turn appears in the DeepSeek usage/log output; - compare the Copilot premium-request counter before and after exactly one turn. If only image turns consume Copilot quota, the vision proxy explains it. If a text-only DeepSeek turn also decrements the counter, this is likely a VS Code/extension regression. Please report it to the [extension's issue tracker](https://github.com/Vizards/deepseek-v4-for-copilot/issues) with the VS Code version, extension version, selected model, and before/after counters. Do not include API keys or verbose request dumps containing private prompts.

bsakaguchi referenced this issue

2026-10-26 09:46:39 +00:00

[PR #82] [CLOSED] docs: add dscode integration guide (bilingual) #190