c7aea0db59
The sandbox escape mechanism is a bypass, not privilege elevation. The old name caused the agent to confuse it with sudo. - elevate/ → bypass/ (package, types, tests) - elevate_notify.go → bypass_notify.go - Namespace: /elevate → /bypass, session/*/elevate → session/*/bypass - Tool arg: "elevated" → "bypass" - Env: OLLIE_ELEVATE_SOCKET → OLLIE_BYPASS_SOCKET - File: elevate-policy.yaml → bypass-policy.yaml - All docs, prompts, and scripts updated |
||
|---|---|---|
| .. | ||
| README.md | ||
| copilot.json | ||
| default.json | ||
| driver.json | ||
| explorer.json | ||
| librarian.json | ||
| navigator.json | ||
| taskmanager.json | ||
| theo.json | ||
README.md
Agents
Each .json file defines an agent configuration. Agent configs are installed to ~/.config/ollie/agents/ and selected at runtime via /agent <name>.
default
Base agent with no specializations. Provides the standard ollie toolset and prompt assembly. Also serves as the base for JIT-generated subagents — subagent_generate copies default.json and replaces the prompt with SYSTEM_PROMPT + the identity text.
coding
Coding-focused agent. Extends default with agent instructions covering code modification discipline, security practices, and exploratory reading habits. Also enables the execute_bypass_bash trusted tool for operations requiring bypass privileges, and configures an MCP server for bypass operations.
System prompt
The system prompt is assembled from the prompt array in the agent config. Each entry is a shell command whose stdout is appended to the prompt. Commands run in the session's working directory with access to $PWD.
The base system prompt is embedded in the binary and automatically prepended.
The fully assembled system prompt is visible at session/{sname}/agent/{aname}/systemprompt.
Hooks
Hooks are lifecycle callbacks executed at specific points:
| Hook | When | Use case |
|---|---|---|
agentSpawn |
Session creation | Start bypass adapter, initialize state |
postTurn |
After each turn | Task list display, logging |
preCompact |
Before context compaction | Save state |
postCompact |
After context compaction | Notify user |
turnError |
On turn error (rate limit, transient failure) | Fallback handling, model switching |
Hook commands run with OLLIE_SESSION_ID set to the current session ID.
MCP Servers
Agent configs can specify MCP (Model Context Protocol) servers for external tool providers:
{
"mcpServers": {
"server-name": {
"command": "server-binary",
"args": [],
"env": {"KEY": "value"}
}
}
}
Tools from MCP servers are registered alongside built-in tools and can be invoked by the agent.
Adding a new agent
Copy default.json, rename it, and adjust the prompt array to include the prompts you want. Add hooks, trusted tools, MCP servers, and generation parameters as needed.