ferrify-application

ferrify-application orchestrates a governed Ferrify run.

This crate is the layer that wires repository modeling, policy resolution, patch planning, verification, trace grading, and final reporting into one coherent flow. It does not own repository parsing or process execution itself. Instead, it coordinates the surrounding crates and returns a structured RunResult.

What This Crate Owns

RunRequest
RunResult
GovernedAgent
ApplicationError

Execution Stages

A typical run goes through these stages:

classify task input and untrusted text
build a repository model
resolve mode and approval policy
create a bounded change plan
derive a patch plan
run verification
build an evidence-backed final report

Example

Add the packages:

[dependencies]
ferrify-application = "0.1.1"
ferrify-domain = "0.1.1"
ferrify-infra = "0.1.1"
ferrify-policy = "0.1.1"

Run the orchestrator:

use std::collections::BTreeSet;
use std::path::PathBuf;

use ferrify_application::{GovernedAgent, RunRequest};
use ferrify_domain::{ApprovalProfileSlug, Capability, TaskKind};
use ferrify_infra::ProcessVerificationBackend;
use ferrify_policy::{PolicyEngine, PolicyRepository};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let repository = PolicyRepository::load_from_root(std::path::Path::new("."))?;
    let engine = PolicyEngine::new(repository);
    let agent = GovernedAgent::new(engine, ProcessVerificationBackend);

    let result = agent.run(RunRequest {
        root: PathBuf::from("."),
        goal: "tighten CLI reporting surface".to_owned(),
        task_kind: TaskKind::CliEnhancement,
        in_scope: vec!["crates/agent-cli/src/main.rs".to_owned()],
        out_of_scope: Vec::new(),
        approval_profile: ApprovalProfileSlug::new("default")?,
        approval_grants: [Capability::EditWorkspace].into_iter().collect::<BTreeSet<_>>(),
        untrusted_texts: Vec::new(),
    })?;

    println!("{}", result.final_report.outcome.headline);
    Ok(())
}

Design Notes

The application layer should stay orchestration-focused.
Policy and trust decisions must remain visible in returned data.
A successful run is a verified plan unless the runtime actually performs edits.

Relationship To The Workspace