Configuration
Keep startup defaults, selected-run parameters, and session controls easy to distinguish.
One file, explicit settings
Configuration uses TOML. Paths resolve relative to the configuration file, not the working directory of an agent. Unknown sections and settings are rejected.
[project]
state_db = ".agenting/example.sqlite"
prompts = "prompts"
[orchestrator]
workers = 4
requests_per_minute = 0
worker_timeout_seconds = 45
[codex]
mode = "fake"
[run]
model = "fake_1"
reasoning = "medium"
[web]
port = 8766Three levels of execution settings
- New-run defaults:
[run]supplies model and reasoning defaults when a run is created. - Stored run parameters: the selected run's model and reasoning remain authoritative during execution.
- Session controls: dashboard worker and task-start limits apply to the current coordinator session.
Worker capacity and task admission
workers sets the startup concurrency limit. requests_per_minute limits runner starts across a rolling 60-second window; zero disables the limit. It does not count every HTTP request made internally by the CLI.
Lowering worker capacity lets active attempts finish. Waiting for a task-start slot leaves work pending. Session overrides do not rewrite the configuration or service database.
Real Codex CLI execution
Real mode additionally requires an installed CLI, saved authentication, and an explicitly configured HTTPS Responses endpoint compatible with that authentication. The framework does not derive its endpoint from a saved login.
The worker runner configures an HTTPS provider with WebSockets disabled. It inventories and disables configured MCP servers and additional integrations before each attempt; setup failures refuse execution.