opencode-agent-plus

Per-agent reminders, request settings and safe-command auto-allow for OpenCode v2.

Set chat_template_kwargs (like enable_thinking), sampling settings (temperature, top_p, top_k, min_p, penalties) and a system reminder for each OpenCode agent. Works with llama.cpp, LiteLLM and other OpenAI-compatible servers, and with thinking models like Qwen.

Auto-allows read-only tool calls and safe shell commands (ls, cat, git status, docker compose ps, …), so with shell: ask in your permissions you get no prompt window for them — everything else still asks.

OpenCode v2 drops request settings such as temperature for OpenAI-compatible providers, and has no per-agent reminder. agent-plus reads extra keys from the agent's .md file and applies them to every model request.

Tested with OpenCode 2.2.0 and llama.cpp behind LiteLLM.

Install

Add it to the plugin list in opencode.json:

{
  "plugin": ["opencode-agent-plus"]
}

Or, from GitHub: "plugin": ["github:Anooshiravan/opencode-agent-plus"]. Agents: follow AGENTS.md.

Example

---
description: Writing and editing text
mode: primary
model: litellm/my-model
temperature: 1.0
enable_thinking: false
reminder: |
  - Keep replies short, simple words
  - If unclear, say so — don't guess
---

You help the user write and edit text.

Keys

Key	Value	Effect
reminder	text, one line or a | block	Added as the last message of every model request
reminder_file	path, relative to the agent's .md file	Like reminder, read from a file
on_error_command	shell command, one line or a | block	Run when a request fails with a connection error (server off), then retry
safe_commands	none (default: on)	Turn safe-command auto-allow off for this agent
temperature	number	Request body
top_p	number	Request body
top_k	whole number	Request body
min_p	number	Request body
presence_penalty	number	Request body
repeat_penalty	number	Request body
max_completion_tokens	whole number	Request body
reasoning_budget_tokens	whole number	Request body (llama.cpp)
reasoning_effort	none, minimal, low, medium, high or xhigh	Request body
enable_thinking	true or false	Request body, under chat_template_kwargs

Names are sent as written. A setting with a wrong value is not sent.

Reminders

Why

The system prompt sits at the start of the conversation. As the conversation grows with messages and tool output, the model pays less attention to that early text and drifts: longer answers, extra work, guessing. Models weigh recent text most, so a short reminder at the end puts the rules right where the model starts writing. Small local models benefit the most.

How

Before every model request, agent-plus adds one message at the end of the conversation with, in this order:

1. the agent's reminder
2. the agent's reminder_file
3. reminder.txt in the project folder, for every agent

The text is wrapped in <system-reminder> tags with a note that it is not from the user. Without this, the model may take the reminder as your message and reply to it ("Understood. Short replies...") instead of answering you.

The message is only added to the outgoing request, NOT saved in the chat history. Every request has exactly one reminder at the end, so reminders never pile up.

Prompt cache

The impact on the server's prompt cache (KV cache) is small and does not grow with the conversation:

• Request 1 is history + reminder; the model writes its reply after the reminder.
• Request 2 is history + reply + new message + reminder.

The cache matches up to the end of history, where the old reminder was. The server then processes the model's last reply again, plus the new message and the new reminder, which are new anyway. The extra cost per request is the length of the last reply (one answer or one tool call), never the whole conversation. Servers that can reuse shifted cache chunks, like llama.cpp with --cache-reuse, may recover part of it.

Warnings

At start, a toast lists problems in agent files:

• misspelled keys, like temprature → temperature
• wrong values, like top_k: must be a whole number
• a reminder_file that is missing or empty

Unknown keys that don't look like a misspelling are left alone, so other plugins can add their own keys. With no problems, the toast says "agent-plus is loaded".

Sidebar

The session sidebar shows an "Agent-plus" section with the settings of the session's agent. It updates when you switch agents or save an agent file.

Wake the server

on_error_command runs a shell command when a model request fails with a connection error (the server is off or unreachable), then asks OpenCode to retry after a short delay. HTTP errors (like 500 or 429) never trigger it.

---
on_error_command: |
  server-on
  curl -sf http://litellm:4000/health
---

It fires once per outage per session: after a request gets through, the next connection error can run it again.

Safe commands

With shell: ask in your permissions, every shell command would prompt. agent-plus skips the prompt for calls that can only read:

• Tools: read, glob, grep, websearch
• Commands: ls, cat, head, tail, wc, find, grep, rg, sort, diff, pwd, echo, du, df, git status/log/diff/show/branch, docker ps/logs/inspect, docker compose ps/logs, npm ls/view/outdated, node --version, psql/sqlite3 with read-only queries, and ssh host <safe command>
• Anything with ;, |, &&, $( ) or redirections is checked per part: all parts must be safe

It never overrides a configured deny, and only turns ask into allow. Doubtful = asked.

Known unsafe flags are refused, e.g. find -exec, env <cmd>, rg --pre, sort -o, git branch -D, git log --output=, ssh -o ProxyCommand=.

Turn it off for one agent with safe_commands: none in its frontmatter.

Notes

• Use max_completion_tokens, not max_tokens: OpenCode already sends max_completion_tokens, and servers use that one.
• Do not give reminder_file a .md name: OpenCode loads every .md in the agent folder as an agent.
• enable_thinking and reasoning_budget_tokens need a server that supports them, like llama.cpp.

License

MIT