DeepSeek V4 for Copilot Chat
Native DeepSeek V4 (Pro / Flash) provider for VS Code Copilot Chat — with full extended-thinking support, agent-mode tool calling, and built-in cost tracking.
Features
- Four model variants in the Copilot model picker: Pro (thinking), Pro, Flash (thinking), Flash
- Extended thinking with configurable effort (
high / max, see Settings) and full reasoning chain preserved across multi-turn agent loops
- Agent-mode tool calling that does not break on the second turn
- Status bar with live account balance and session spend (auto-detects CNY / USD)
- Background balance refresh after each chat (debounced, silent)
- Persistent reasoning cache (survives VS Code restarts)
- Actionable error notifications for 400 (reasoning) / 401 / 402 / 422 / 429, plus mid-stream
insufficient_system_resource truncation handling
- Retry on transient failures (5xx, 429, network jitter)
- Adaptive token estimator (EMA-calibrated chars/token from real
usage data)
- First-run walkthrough and a "key required" warning state in the picker so the model entries are always discoverable
Text-only, by design
DeepSeek V4 is in preview and only accepts text input. A common workaround
in other extensions is a "vision proxy": when you drop an image into chat,
it gets forwarded to a different multimodal model (GPT-4o, Claude, etc.)
and the resulting description is fed to DeepSeek as if it had seen the
picture.
We deliberately do not do this. The proxy introduces a silent fidelity
gap — DeepSeek answers based on another model's description of the image,
not the image itself. Any hallucination, omission, or bias in the
intermediate description leaks straight into DeepSeek's reply, and the UI
gives the user no signal that the model never actually saw what they
attached.
When DeepSeek itself ships native multimodal support, we will enable it
through this provider. Until then, image attachments are intentionally
unsupported here.
What this plugin uniquely solves
Generic OAI-compatible bridges drop the reasoning_content field when forwarding to DeepSeek, which causes a hard 400 on the second turn of any tools-enabled thinking conversation:
The reasoning_content in the thinking mode must be passed back to the API.
This extension is a native VS Code Language Model Provider — it intercepts each request, restores the prior reasoning_content from a local cache, and re-attaches it to every assistant turn in history (not just the ones that called a tool, which is what DeepSeek actually requires when tools is present). Result: agent loops, multi-turn refactors and long thinking sessions work end-to-end.
Quick start
- Install DeepSeek V4 for Copilot Chat from the VS Code Marketplace.
- Run
Manage DeepSeek V4 Provider from the command palette and paste your DeepSeek API key.
- Open Copilot Chat, pick a DeepSeek V4 variant from the model picker, start chatting.
Commands
| Command |
Description |
Manage DeepSeek V4 Provider |
Set or update your API key |
Refresh DeepSeek V4 Balance |
Fetch the latest account balance |
Show DeepSeek V4 Log |
Open the runtime log channel |
Show DeepSeek V4 Reasoning Cache Stats |
Diagnostics for the reasoning cache |
Clear DeepSeek V4 Session Counter |
Reset the session spend display |
Settings
| Setting |
Values |
Default |
Description |
deepseekv4.reasoningEffort |
high | max |
max |
Reasoning depth for (thinking) model variants. high is faster with shorter reasoning chains; max is the deepest setting. No effect on non-thinking variants. Picked up at request time. |
License
MIT. See LICENSE. Forked from huggingface-vscode-chat; the protocol layer was rewritten for DeepSeek V4.
