ollie/data/agents/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>.

Config format

{
  "prompt": ["$XDG_CONFIG_HOME/ollie/prompts/agent-coding.md"],
  "userPrompts": ["$XDG_CONFIG_HOME/ollie/prompts/rules.md"],
  "tools": ["shell", "file_read", "file_edit", "file_glob", "file_grep"],
  "temperature": 0.5,
  "maxTokens": 16384,
  "reasoningEffort": "medium",
  "backend": "openrouter",
  "model": "deepseek/deepseek-v4-flash",
  "systemPrompt": "/path/to/override.md",
  "compactionModel": "cheap-model-name"
}

Field	Purpose
prompt	File paths to concatenate as the agent prompt (system preamble). Env vars expanded.
userPrompts	File paths to concatenate as active global rules. Prepended to every user message at turn start. Env vars expanded.
tools	Tools loaded at session start. Add more at runtime with the /tool_load ctl command.
temperature	Sampling temperature.
maxTokens	Max output tokens per response.
reasoningEffort	How hard a reasoning-capable model thinks: low, medium, or high. See below.
thinkingBudget	Advanced: explicit thinking-token budget. Overrides reasoningEffort. See below.
backend	Override backend for this agent.
model	Override model for this agent.
systemPrompt	Override the embedded system prompt with a file path.
compactionModel	Use a cheaper model for context compaction.

Reasoning effort

reasoningEffort is the primary, human-facing control for how much a reasoning-capable model deliberates before answering. It takes one of three levels — set it and forget it; you do not need to reason about token counts.

Level	Use for	Thinking budget
low	Quick lookups, inline assistance, read-only monitoring	4096 tokens
medium	General coding, exploration, writing	8192 tokens
high	Planning, orchestration, security/code review	16384 tokens

How each backend applies it:

• OpenAI, OpenRouter, Copilot, Gemini — sent as the discrete reasoning_effort level (low/medium/high).
• Kiro — enables summarized adaptive thinking and maps the level to its output_config.effort.
• Anthropic — has no discrete level, so the level is translated to the thinking-token budget in the table above (extended thinking). Because thinking tokens are drawn from the output budget, the runtime automatically grows maxTokens above the thinking budget and forces temperature to 1 (both are Anthropic requirements when thinking is enabled).

An invalid reasoningEffort (anything other than low/medium/high) is a config error and the agent will fail to load, rather than silently doing nothing.

thinkingBudget (advanced override)

thinkingBudget sets an explicit thinking-token budget and takes precedence over reasoningEffort. Use it only when you need a specific budget that the three levels don't provide; most agents should use reasoningEffort instead. It only affects backends that accept a numeric budget (Anthropic).

Agents

Name	Role
default	Base coding agent. Also the base for JIT subagents.
driver	Full-featured agent with all tools, LSP, GUI, memory, subagents.
copilot	Read-only code copilot. Suggests changes as code fences.
navigator	Pair programming partner. Reads and advises, doesn't edit.
explorer	Code exploration and explanation. Read-only.
librarian	Document search and knowledge curation. Read-only.
taskmanager	TODO list management and planning.
theo	Security auditor (Theo de Raadt voice). Read-only.

System prompt assembly

The final system prompt sent to the model is assembled in sections:

1. system — embedded base prompt (prompts/system_prompt.md)
2. env — runtime environment (cwd, platform, date, git status)
3. agent — concatenated files from the prompt array
4. tools — auto-generated from loaded tools (name + .meta prompt)

The fully assembled prompt is readable at session/{sname}/agent/{aname}/systemprompt.

User prompts (active global rules)

Files listed in userPrompts are resolved the same way as prompt (file paths with env var expansion), but instead of being injected into the system preamble, they are prepended to every user message at turn start. This lets you define persistent "global rules" that appear alongside each user request without bloating the system prompt.

Adding a new agent

1. Copy default.json, rename it.
2. Adjust the prompt array to reference your prompt files in data/prompts/.
3. Set tools to the tools this agent needs.
4. Set generation parameters as needed.
5. Run make install-data to install.