[General] # Issue

**Mood:** 😔
**Category:** General

# Issue
Copilot is token inefficient with requests.

# Settings
Model settings: GPT 5.4 mini, Medium thinking.
Input: `test` x 2 (each in a separate session)
Expected: `Acknowledged` or whatever. It's never 100% consistent, but it is quick and short.
No global instructions set (which obviously use token).
Skills all disabled (which also use tokens for the reference loading).
This usage is coming from whatever base prompts are injected per request.
I cannot see the exact token using in Copilot, but I can see my credits being used up.

# OpenCode
This harness uses about 9K tokens per request which mostly come from the system prompts the user cannot change.

For further context, I primary use OpenCode and have done so for many months. It works solidly with the Copilot subscription and I do not have coding issues. I am therefore happy with whatever it requires to have high success rates (although I wish it used less)

## Usage
Test request x 2 = 1 credit
This is reproducible.

# Copilot desktop app
Starting testing this out. Since you moved to token based usage, I wanted to see what the difference was.

## Usage
Test request x 2 = 3-4 credits
This is reproducible.

As stated above, I cannot measure exact token counts but I can see my credits

# Summary
Copilot uses 3 to 4x the credits that OpenCode does. I was really hoping you guys had put lots of effort into cost efficiency and you'd beat an opensource harness. That is not the case.

Until this is fixed, this issue is an **adoption blocker**. There are no advantages to using Copilot desktop app over OpenCode, therefore cost efficiency or allowing a user to tweak your hidden system prompts to allow more customisation is a must-have feature.

# Related data
## opencode.jsonc
```
{
"$schema": "https://opencode.ai/config.json",
// Prevent accidentally sharing code on a remote edge server
"share": "disabled",
// OpenCode is in beta
"autoupdate": false,
// Global permissions - ensure all agents and sub-agents play safely
"permission": {
"bash": {
// Default: always prompt for ANY shell command
"*": "ask",
// Remote-damaging Git operations
"git push": "deny",
"git push *": "deny",
// Changing remotes can enable accidental pushes to the wrong place
"git remote add *": "deny",
"git remote remove *": "deny",
"git remote set-url *": "deny",
// Local destructive operations
"git clean -f*": "deny",
"git clean -x*": "deny",
"git clean *-f*": "deny", // catches reordered flags
"git reset --hard": "deny",
"git reset --hard *": "deny",
// Directory removal
"rm -rf*": "deny",
"rm -fr*": "deny",
"rm -r *": "deny",
"rm --recursive*": "deny",
"rmdir *": "deny",
// Repo hosting CLIs (if installed)
"gh repo delete*": "deny",
"glab repo delete*": "deny",
// Safe commonly requested commands
"dotnet *": "allow",
"git diff *": "allow",
"git log *": "allow",
"git ls-files *": "allow",
"git status *": "allow",
"javap *": "allow",
"ls *": "allow",
"mkdir *": "allow",
"mvn *": "allow",
"npm *": "allow",
"wc *": "allow"
},
"external_directory": "ask",
// An agent can ask sub-agents with broader permissions to do work on their behalf
"task": {
"*": "ask",
// Explore is read-only so safe to allow
"explore": "allow",
"general": "ask"
},
"skill": "deny",
"lsp": "deny"
},
// Agent-specific overrides (agent rules take precedence)
"agent": {
// Prevent Plan agent from doing anything potentially dangerous at all
"plan": {
"permission": {
"bash": "deny",
"edit": "deny"
}
}
}
}
```

---
| Field | Value |
| --- | --- |
| App version | 0.2.27 |
| OS | Windows 10.0.26200 |
| Theme | GitHub |
| Path | /chat |
| Tenure | Day 3 (Week 1) |

Bắt đầu bằng cách tái hiện hai yêu cầu `test` trong ứng dụng desktop Copilot ở phiên bản 0.2.27 và so sánh mức sử dụng credit của chúng với cấu hình OpenCode được mô tả trong `opencode.jsonc`. Issue được hoàn thành khi chênh lệch credit được giải thích hoặc cơ chế tính token nền tảng hay phần overhead của prompt được sửa và kết quả có thể tái hiện.

Do mô hình lập chỉ mục viết ra từ nội dung của issue.