Agentic Kit v2

Agentic Kit is a vendor-independent repository control layer for coding agents. It defines project commands, capability policy, deterministic verification, and machine evidence, then renders only the native configuration required by selected tools.

It does not implement an agent runtime, orchestrator, worktree manager, sandbox, permission prompt system, or generic code reviewer.

Build

cargo build --release
cargo test --all-targets

The resulting executable is agentic-kit.

Initialize a project

agentic-kit init --tools codex

The initial verification check deliberately references agentic-kit-unconfigured-check. Replace it with the project's real structured argv before verification can pass:

checks:
  unit:
    argv: ["cargo", "test", "--all-targets"]
    required: true
    timeout: 10m

Then run:

agentic-kit validate
agentic-kit doctor
agentic-kit policy lint
agentic-kit render --check
agentic-kit verify --profile handoff

Migration

Migration is dry-run by default:

agentic-kit migrate
agentic-kit migrate --apply

Shell fragments, locally modified files, and unsupported capability mappings require explicit manual resolution. Generated cleanup is also dry-run-first:

agentic-kit clean
agentic-kit clean --apply

Optional domain behavior is installed explicitly and rendered through the selected vendor adapters:

agentic-kit pack add database-migration
agentic-kit render

Repository layout

.agentic/                 Canonical project, policy, lock, and evidence
crates/agentic-kit-cli/   Rust CLI
schemas/                  Public versioned schemas
compatibility/            Tool/version/dialect support registry
packs/                    Optional domain-specific packs
fixtures/ and evals/      Compatibility fixtures and executable evaluation
docs/                     Architecture, plans, and runbooks

Generated vendor config is conditional. Projects that do not select Claude, Codex, or OpenCode do not receive those tools' files.

Evaluation

agentic-kit eval validate validates the representative task corpus. agentic-kit eval compare --observations <path> requires a complete four-profile matrix before reporting accepted-change rate, time, intervention, rework, escaped-defect, false-positive and unverified-claim metrics.

Release and compatibility

Tagged releases build native archives on Linux, macOS and Windows, include a CycloneDX SBOM and SHA-256 checksum, and receive GitHub build-provenance attestations. Compatibility fixtures run weekly; support and deprecation rules are defined in docs/compatibility-policy.md.