Agent-owned context.d scripts for Marrow running on marrow-core.
In marrow-core, any executable script inside a context_dirs directory is run on every heartbeat tick. The script's stdout is appended verbatim to the agent prompt — no JSON protocol, just plain text.
context.d/
├── queue.py # Read tasks/queue and print pending tasks to prompt
├── health.py # Check local services when queue is idle
└── explore.py # Fallback exploration prompt when queue is empty
Reads ~/tasks/queue/ and injects pending task files into the prompt. Also shows delegation status between scout and artisan agents.
When the task queue is empty, probes local HTTP services (Caddy, web UI, etc.) and reports their status. Gives the agent situational awareness during idle time.
Fallback context when no tasks are queued. Suggests the agent should explore the codebase, check for improvements, or look at owner's GitHub repos.
# Install as context provider for marrow-core
cp context.d/*.py ~/context.d/
chmod +x ~/context.d/*.py
# Or clone directly
git clone https://github.com/marrow-bot/marrow-context.git
cp marrow-context/*.py ~/context.d/Configure marrow.toml to point to the directory:
[[agents]]
name = "scout"
context_dirs = ["/Users/marrow/context.d"]Add new .py scripts to context.d/. They must:
- Be executable (
chmod +x) - Print useful context to stdout only
- Exit cleanly (exit code 0)
- Run quickly (< 1s ideally)
MIT