Sayri Subagents Specification

In Sayri, a Subagent is an autonomous AI worker instance configured with its own system prompt, isolated model parameters, fine-grained Bubblewrap sandbox policy, and allowed tools/skills.

Subagent profiles are stored persistently as JSON files under:

~/.config/sayri/agents/<agent-id>.json

---

1. Agent Profile JSON Schema

Below is the complete specification for a subagent definition:

{
  "id": "code-debugger",
  "name": "Code Debugger & Tester",
  "description": "Autonomous Python and Rust code analyzer running in isolated Bubblewrap sandbox.",
  "system_prompt": "You are a code debugging subagent. When given code or error tracebacks, isolate issues, generate tests, and verify outputs inside your sandbox.",
  "model": {
    "provider": "default",
    "model_name": "mistral-small-latest",
    "base_url": null,
    "api_key": null,
    "temperature": 0.2,
    "max_tokens": 4096
  },
  "sandbox": {
    "level": "LEVEL_2_ISOLATED_DEV",
    "timeout_seconds": 30,
    "isolated_dir": "~/.local/share/sayri/sandboxes/code-debugger",
    "allow_network": false,
    "allowed_binaries": ["python3", "pytest", "rustc", "cargo"],
    "blocked_binaries": ["dd", "mkfs", "shutdown", "reboot", "rm"]
  },
  "allowed_skills": [
    "sayri-skill-web-search"
  ],
  "allowed_tools": [
    "bash",
    "read_file",
    "write_file"
  ],
  "is_builtin": false
}

---

2. Configuration Fields Reference

Model Settings (model)

• provider: Model provider client (default, openai, groq, ollama, openrouter).
• model_name: LLM identifier (e.g. mistral-small-latest, llama-3.3-70b-versatile, gpt-4o).
• temperature: Sampling temperature (0.0 to 1.0). Use lower values (0.1–0.3) for deterministic code generation.

Sandbox Isolation Settings (sandbox)

• level:
  • LEVEL_0_NO_EXEC: Total prohibition of commands.
  • LEVEL_1_READONLY: Read-only host filesystem (/), ephemeral /tmp.
  • LEVEL_2_ISOLATED_DEV: Read-only system root, private read-write workspace in ~/.local/share/sayri/sandboxes/<id>.
  • LEVEL_3_HOST_USER: Host user permissions ($HOME).
  • LEVEL_4_HOST_ROOT: Elevated via Polkit graphical confirmation (pkexec).
• allow_network: true or false (uses --unshare-net in Bubblewrap).
• timeout_seconds: Maximum process execution runtime before automatic SIGKILL.
• blocked_binaries: Array of banned command names intercepted prior to execution.

---

3. Managing Subagents via UI Cajita

You can manage agents graphically from the Sayri Cajita widget:

1. Press the Sayri hotkey or tray indicator to open Cajita.
2. Switch to the Agents tab.
3. Click + New Subagent to declare its name, system prompt, model, and sandbox level.
4. Existing subagents can be inspected, tested, or deleted with one click.