☀ SolarOS

SolarOS 4.6.8 manual · app

agent

Native Responses/Chat-Completions LLM client and SolarOS agent control plane. It streams model text directly to the active shell and exposes typed system-status, storage-listing, job-listing, display-discovery, and optional Python/Lua execution tools. agent tools shows risk, policy, and runtime availability.

Configure the full endpoint and model, then ask a question:

agent config endpoint https://api.openai.com/v1/responses
agent config model gpt-model
agent config key api-key
agent config reasoning medium
agent config tools confirm
agent config max-tools 16
agent
agent list
agent resume SLOT
agent ask How much memory is free on this device?
agent script python -c "print(6 * 7)"
agent script lua /script.lua argument

Bare agent and agent new open a new foreground prompt loop. Completed turns are saved, and agent list plus agent resume SLOT restore a selected transcript after leaving the app or rebooting. agent delete SLOT removes one conversation without changing provider credentials. Responses uses its saved provider continuation ID; Chat Completions uses bounded local history. Slots are 1 to 3 on internal flash or 1 to 8 on SD; the oldest is reused when full. agent ask performs one unsaved request but likewise waits for Esc or the app-exit key after completion, so display-shell output is not immediately replaced by the shell screen.

Use agent config key clear for an endpoint that does not require bearer authentication. agent status shows configuration, request counters, HTTP status, reasoning effort, duration, traffic, and internal-RAM/PSRAM measurements from the last request. The API key itself is never printed. A Chat-Completions-compatible endpoint can still be configured explicitly.

The agent permits 16 sequential tool calls by default, followed by a separately reserved final provider turn. agent config max-tools COUNT stores a per-request limit from 1 through 32 in NVS. The final turn advertises no tools after that budget is consumed, so the model concludes from the collected results instead of failing on one more tool request. Output is bounded to 16 KiB, protocol buffers and queues prefer PSRAM, and the foreground worker uses a declared 16 KiB internal stack. Full builds can reuse that worker for bounded Python or Lua source/file execution. The manual script path captures at most 4095 output bytes, has a 30-second deadline, and supports cancellation with Esc or the app-exit key. Model-generated source is capped at 640 bytes and captures 383 output bytes.

The storage registry includes bounded listing, sensitive text-file reads, and text-file replacement. Reads and writes are capped at 3072 bytes and paths below .ssh are unavailable to protect SSH identity files.

The default confirm policy runs read-only tools automatically and shows the exact arguments of sensitive, mutating, disruptive, and script calls before waiting for a local y/N decision. off disables tools, readonly excludes all protected risk classes, and explicit all allows every available tool without prompting. See Native Agent Service for the provider contract and current limits.

Controls: