Skills Factory.

Start with a job worth repeating.

An agent skill is a package of instructions and supporting resources for a particular task. It can explain which evidence to gather, how to use approved tools, how to handle missing information and what a useful result should contain.

Skills Factory is the engineering work around those packages: define the contract, build the skill, review its boundaries, test realistic cases and keep its assumptions current. It is not a hosted skill-running service. Security Skills Architect is the public builder preview available today.

What you can download.

The Security Skills Architect repository contains two reusable skills:

• skill-architect helps design, build, review and evaluate a skill, with a specification, test cases and a release decision grounded in evidence.
• application-defence learns one application and helps generate its dedicated defence skill, evidence profile, repeatable checks, finding rules and operator handoff.

The application workflow prioritises emerging vulnerabilities: identify deployed dependencies, consult fresh advisory and exploitation sources, review affected code paths and prepare vendor-supported patch or mitigation guidance. Version matches need exposure review; they do not prove exploitability or discover every undisclosed zero-day.

The reviewed preview release includes versioned ZIPs, a manifest, checksums and installation guidance. Public reuse licensing is still awaiting an owner decision. Native host acceptance and broad behavioural validation remain incomplete.

From the builder to a running application skill.

1. Download the builder and follow the repository instructions for your agent.
2. Identify an application you are authorised to review. Supply its repository, environments and a private evidence location.
3. Generate its dedicated skill and review the application profile, checks, source freshness, missing coverage and response rules.
4. Connect inspected collectors, a recurring executor and an operator destination. Observe real runs and test loss-of-coverage handling before relying on the monitor.
5. Retest changes against the affected deployment. Keep stale evidence and unresolved findings visible.

Installing the instructions alone does not start background monitoring. Our separate private pilot uses a local model to interpret deterministic observations; its executor and raw application evidence are not bundled in the public download. A working feedback connection to krnali Ops is still integration work.

Online advisory queries can send package names and versions to their source. A local model does not make connected tools offline. The agent, executor and data sources need their own access and data-handling review.

How skills fit with Sigil, VE and credentials.

A skill describes how to do a job. krnali Sigil, previously presented here as Secure Agent Runtime (SAR), addresses the authority to act: approved scope, policy decisions and credential lifecycle. A skill cannot approve its own privileges.

Krnali VE addresses the evidence: signed statements, transparency-log receipts and independently checkable records of selected events. Recording a statement does not establish that it is true or grant authority for an action.

Verifiable Credentials carry issuer claims that a holder can present to a verifier. They need issuer trust and acceptance policy. Credentials, authority decisions and retained evidence have distinct roles; the development integrations do not establish complete end-to-end verification for every workflow.

The PHP review skill remains a separate candidate.

PHP Composer Risk helps an agent review composer.json and composer.lock, relate dependency versions to advisory evidence and prepare an upgrade plan. The review candidate also covers proposed patch and targeted verification output.

Its distribution preparation and controlled-text checks are separate from the Architect release. Real-application acceptance, native installation and licensing remain pending; it has no confirmed public download or marketplace listing. The public GitHub links on this page lead to the available Architect packages.

What the preview evidence establishes.

The Architect tooling passes 34 local tests, and both packages pass selected structural checks. The current application-defence suite contains planned scenario and trigger cases; its evaluation plan was validated, but those model evaluations have not been executed.

A separate private application pilot has exercised recurring local-model reviews and advisory matching. That experience has improved the instructions, but it does not certify every generated skill, native host, reasoning result or operational integration. Read the repository review and validation records for the remaining gates.