Mimir: introduce AI companion plugin

[Ahmed5Emad]

Plugin

• Id: alexander/mimir
• New plugin

What it does

AI companion for Noctalia — LLM-powered chat in a side panel. Type questions, get answers with basic markdown rendering (code blocks in shaded boxes). Switch models on the fly, auto-discovers models from the API.

Future

Designed for future tool-use (terminal commands, file search, code editing).

External dependencies

None. Communicates with any OpenAI-compatible HTTP API. Default endpoint is OpenCode Go (opencode.ai/zen/go/v1). API key is user-configured in settings or auto-detected from ~/.local/share/opencode/auth.json.

Testing

• Clicked the brain bar widget — opens/closes the panel.

• Typed messages, received streaming-style replies.

• Refreshed model list, switched between models.

• Cleared conversation history.

• Toggled panel via IPC: noctalia msg panel-toggle alexander/mimir:chat

• Tested with both OpenCode Go API and empty endpoint (graceful error).

• Tested on Niri

• Tested on Hyprland

• Tested on Sway

• Noctalia version tested against: 5.0.0 beta 6

• Plugin API level: 16

note: tested only with opencode go api , sorry don't have 5 AI supscriptions

Screenshots / Videos

Checklist

• The directory name (mimir) matches id after / (alexander/mimir).
• It ships plugin.toml, README.md, thumbnail.webp, and translations/en.json.
• README.md follows the README template, documents every entry id and dependency, and includes the exact IPC command.
• I created thumbnail.webp with the thumbnail generator.
• Version 0.1.0 follows semver; plugin_api is 16.
• No non-English translations.
• I did not edit catalog.toml.
• This PR touches exactly one plugin directory.

Code review attestation

• The code is readable and not obfuscated, minified, or generated.
• It does not download and execute remote code.
• Every network call (POST /chat/completions, GET /models) and filesystem read (auth.json) is accounted for.
• I have the right to publish this code under the MIT license.

[ItsLemmy]

1. blocking - mimir/service.luau:5

When api_key is empty, loadApiKey unconditionally reads the OpenCode Go credential from ~/.local/share/opencode/auth.json.
The endpoint is configured independently at mimir/service.luau:18, and mimir/service.luau:59 sends the detected credential as a Bearer token to that endpoint.

A user who selects another OpenAI-compatible provider and leaves api_key empty can disclose their OpenCode Go credential to that provider. Auto-detection must be restricted to the matching OpenCode origin, or stored credentials must never be reused across configured origins.

2. blocking - mimir/service.luau:60

Chat requests are rejected whenever no API key or auto-detected OpenCode credential exists. This contradicts the documented support for unauthenticated local servers at mimir/README.md:22 and mimir/README.md:87, including Ollama.

Users without an OpenCode credential cannot use the advertised local, no-key configuration. Allow requests without Authorization when api_key is empty.

3. blocking - mimir/plugin.toml:8

The catalog description says the plugin helps with terminal commands and file operations, but the current implementation only provides LLM chat. mimir/README.md:122 explicitly lists command execution, file search, and editing as future work.

The manifest therefore advertises capabilities the submitted plugin does not provide. Update the description to describe the current chat functionality.

4. non-blocking - mimir/service.luau:47

The input watcher records the current user message at mimir/service.luau:111 before calling think. think then copies that
conversation and appends the same input again at mimir/service.luau:50.

Every chat request sends the newest user prompt twice, which can alter responses and consume extra tokens. Append the message only once.

5. non-blocking - mimir/plugin.toml:52

max_history is declared and documented, but service.luau never reads the setting or trims M.conversation. Long sessions therefore grow the transmitted context without bound, and changing the setting has no effect.

Apply the configured limit before building and publishing conversation history.

[blackbartblues]

Isnt it better to connect it to hermes instead of openai?
With that you have all tools you need to use it as your Commander terminal helper

[Ahmed5Emad]

Isnt it better to connect it to hermes instead of openai? With that you have all tools you need to use it as your Commander terminal helper

tryed to do something like that with open code but I couldn't any way , I finshed makeing the tool call system anyway

[Ahmed5Emad]

second , I want it capable but simple hermes take 40k to 90K token just to call tools

[Ahmed5Emad]

I updated version to 0.3.0 add tool calling now it can search files and run terminal commands

[spiros132]

@Ahmed5Emad Would it be possible to keep each PR specifically to one plugin instead of two? You've also added the game-launcher changes in this PR.

[Ahmed5Emad]

@Ahmed5Emad Would it be possible to keep each PR specifically to one plugin instead of two? You've also added the game-launcher changes in this PR.

fixed sorry

[ItsLemmy]

1. blocking - mimir/service.luau:33

The prior credential-origin finding is not fully resolved. Auto-detection now trusts any configured endpoint containing the substring "opencode" at mimir/service.luau:37, then sends the detected OpenCode Go credential to that endpoint at mimir/service.luau:131.

An unrelated or attacker-controlled host whose URL contains "opencode" still receives the credential. Parse the endpoint and require the exact trusted OpenCode scheme, host, and port before reading the auth file.

2. blocking - mimir/service.luau:58

The command blocklist checks only the first whitespace-delimited token, while noctalia.runAsync executes the complete string through /bin/sh -c, as confirmed at ../noctalia-shell/src/scripting/luau_host.cpp:1866. Simple destructive commands such as rm are not in the default list, and blocked commands can be bypassed using shell composition, an absolute executable path, env, sh -c, or another interpreter.

In allow mode, mimir/service.luau:240 executes these model-supplied commands without confirmation. A compromised or incorrect API response can therefore delete files, read credentials, or upload data despite the documented blocklist. Raw shell commands should remain subject to per-command approval, or be replaced with narrowly scoped, structured, allowlisted operations. The current token blocklist is not a security
boundary.

3. non-blocking - mimir/panel.luau:167

The README says commands can be edited before approval, but the approval UI renders each command as a label and only exposes Approve or Deny. mimir/panel.luau:279 sends only a boolean, and mimir/service.luau:314 executes the original model-provided command.

Either implement the documented editable command field and pass the edited value to the service, or remove the editing claims from mimir/README.md:24, mimir/README.md:46, and mimir/README.md:69.

4. non-blocking - mimir/panel.luau:160

The documented copy button does not copy message content to the clipboard. It toggles a focused multiline input containing the message at mimir/panel.luau:191, leaving the user to copy it manually.

Implement clipboard copying or describe this as exposing selectable text rather than copy-to-clipboard functionality.

[Ahmed5Emad]

for number 3 sorry , I removed this and forgot to edit it (will be fixed)
for number 4 it is UX choice because I really didn't find coping all the thing useful , and for some reason , there is no way to make text copy able in the panel to just choose and copy

[Ahmed5Emad]

that should be working , add much more restricted defaults , anyway sorry for security issues , usually I use plugins for myself so I don't usually check for security 😅😅