Groq requests exceed TPM limit even in an empty directory

Description

Bug: Groq requests exceed TPM limit even in an empty directory

Environment

• OpenCode version: 1.18.4

• OS: Linux 6.17.0-PRoot-Distro (Termux + Ubuntu PRoot, arm64)

• Provider: Groq

• Models tested:

  • groq/llama-3.1-8b-instant
  • groq/openai/gpt-oss-20b

Description

OpenCode consistently sends requests that exceed Groq's Tokens Per Minute (TPM) limit, even when started in an empty directory with a brand-new session.

The problem occurs before any meaningful user prompt is sent. Even the hidden title request exceeds the TPM limit.

This makes OpenCode unusable with Groq's free tier.

Steps to reproduce

1. Install OpenCode 1.18.4.
2. Configure a valid Groq API key.
3. Create an empty directory:

mkdir empty
cd empty

4. Run:

opencode run hi --print-logs

Expected behavior

The prompt for a one-word request (hi) should be small enough to fit within Groq's TPM limits.

Actual behavior

The request is rejected immediately:

For llama-3.1-8b-instant:

Request too large for model `llama-3.1-8b-instant`
Limit 6000 TPM
Requested 42781

For openai/gpt-oss-20b:

Request too large for model `openai/gpt-oss-20b`
Limit 8000 TPM
Requested 38369

Diagnostics performed

I verified all of the following:

• Fresh Groq API key
• Fresh Groq account
• Empty directory
• New OpenCode session
• opencode run
• opencode --pure
• No project files
• No AGENTS.md
• No plugins
• Default configuration
• Multiple Groq models

Resolved configuration:

{
  "$schema": "https://opencode.ai/config.json",
  "compaction": {
    "auto": false,
    "prune": false
  },
  "agent": {},
  "mode": {},
  "plugin": [],
  "command": {},
  "username": "root"
}

opencode debug info:

opencode version: 1.18.4
plugins: none

Additional observations

The requests appear to have nearly constant sizes across completely fresh sessions:

Title agent

Approximately 32.5k tokens

agent=title
Requested 32550 tokens

Build agent

Approximately 42.7k tokens

agent=build
Requested 42781 tokens

GPT-OSS

agent=title
Requested 32589 tokens

agent=build
Requested 38369 tokens

The consistency suggests the large token count is not coming from my project or conversation history.

Question

Is OpenCode intentionally sending a very large internal prompt/tool schema to Groq, or could this be a regression in prompt construction or the Groq provider integration?

I'm happy to run additional diagnostics or test a development build if that would help.

log.txt

Plugins

none

OpenCode version

1.18.4

Steps to reproduce

just said hi in a new chat using groq gpt oss or any groq model

Screenshot and/or share link

Operating System

proot-distro ubuntu

Terminal

termux