Configuration (`settings.toml`) & Profiles

CURD v0.7.1 introduces a hierarchical configuration model designed for organizational governance, safely bringing advanced features to both indie developers and enterprise teams. Configuration is stored in .curd/settings.toml.

Runtime Ceilings

CURD provides runtime ceilings that act as hard bounds on what operations are possible for any agent, regardless of context:

full: Exposes the complete tool surface, including code edits, shell commands, and plugins. Best for mature workflows or human-supervised local use.
lite: Intentionally restrictive. Ideal for purely exploratory or read-only agents. It limits tools to:
- workspace (restricted to status, list, dependencies)
- search
- read
- edit
- graph
- protocol basics


[runtime]
ceiling = "full" # or "lite"

Profiles

Profiles define actual behavior and capability within the chosen ceiling. A single binary can power multiple agent roles seamlessly:


[profiles.default]
role = "human_core"
capabilities = ["lookup", "traverse", "read", "change.apply", "session.begin", "session.verify", "session.commit"]
promotion = "user_only"
 
[profiles.assist]
role = "assist_agent"
capabilities = ["lookup", "traverse", "read", "change.prepare", "session.begin", "session.verify", "review.run"]
promotion = "forbidden"
 
[profiles.supervised]
role = "supervised_agent"
capabilities = ["lookup", "traverse", "read", "change.apply", "session.begin", "session.verify", "exec.task", "plan.execute", "review.run"]
session_required_for_change = true
promotion = "approval_required"
 
[profiles.autonomous]
role = "autonomous_agent"
capabilities = ["lookup", "traverse", "read", "change.apply", "session.begin", "session.verify", "session.commit", "exec.task", "plan.execute", "plan.parallel", "review.run", "hook.run"]
session_required_for_change = true
promotion = "policy_gated"

Ceiling gates what is physically possible.
Profile gates what the actor is allowed to do via capability atoms (e.g. search maps to lookup, edit maps to change.apply).
Policy (defined below) can still deny a valid request.

Organizational Guardrails (`[policy]`)

The [policy] block is the central firewall for all agent interactions, enforcing organizational rules across all interfaces (CLI, REPL, MCP).

Key	Type	Description
`mode`	`string`	`strict` (default-deny), `permissive` (default-allow), or `audit` (log only).
`block_files`	`array<glob>`	Explicit list of files/directories agents can never touch.
`allow_files`	`array<glob>`	Explicit safe zones. In `strict` mode, ONLY these are accessible.
`allowed_binaries`	`array<string>`	List of binaries permitted via the `shell` tool.
`block_shell_metachars`	`bool`	If true, blocks command chaining (`;`, `&`, `
`require_plan_for_mutations`	`bool`	If true, agents must register a plan before calling `edit` or `shell`.
`protected_symbols`	`array<uri>`	List of symbol IDs (e.g., `main.rs::init`) that cannot be modified.

Build Tasks

You can define pixi-style hooks for standard project operations directly within settings:


[build.tasks]
build = "cargo build"
test = "pytest tests/"
release = "make release"

CURD’s new adapter coverage allows automatic execution via curd build or curd test across poetry, pip, conda, mamba, cargo, npm, uv, and more.

AST Safety Valves (`[edit]`)

Prevents agents from performing non-semantic “oversteering”:

Key	Type	Description
`churn_limit`	`float`	Max percentage of AST nodes modifiable in one call (default: `0.30`).
`enforce_transactional`	`bool`	Require a `workspace` session for all edits.

Verification Integrity

CURD computes a SHA-256 Config Hash of these settings. This hash is provided to front-ends during authentication to ensure the agent is adhering to the verifiable human-defined policy.

Configuration (settings.toml) & Profiles