Viridis Agent Reliability

Inspect an agent service's scope, MCP compatibility and security evidence. This package connects to the hosted SecurityPreflight MCP server and includes the agent-reliability-check diagnostic skill. It does not certify runtime security.

Installation

For Claude Code, add the public repository marketplace and install locally in an intended project:

claude plugin marketplace add jdhart81/viridis-agent-fleet
claude plugin install viridis-agent-reliability@viridis-agent-fleet --scope local

For a temporary Claude Code session, load the checked-out plugin directory:

claude --plugin-dir ./plugins/viridis-agent-reliability

Codex supports the repository's .agents/plugins/marketplace.json and the plugin's .codex-plugin/plugin.json compatibility manifest. Choose an explicit project scope. ChatGPT and Claude hosted account installation are separate from local CLI installation; neither platform directory lists this package yet.

What is available

The HTTP server advertises five tools. describe_agent explains its scope; get_security_receipt reads an unsigned filtered public summary. The diagnostic skill uses only these two tools. Three assessment tools also exist: security_preflight, scan_source, and screen_injection. They require an existing registered caller and a verified fixed-order session belonging to that caller. Installation does not create an account or an entitlement.

A diagnostic is not an assessment. To run an assessment, retain the original intake and use its existing paid session under the same authorized caller. Another caller, subject, input or second execution cannot consume that order. Credential setup needs an explicitly approved scope; never paste secrets into a conversation or plugin package. No OAuth or self-service signup is implemented.

Use exact host tool controls to restrict a diagnostic session to the two read operations. Skill instructions and MCP annotations are not authorization or server-side tool filtering. Do not automatically approve assessment tools.

Privacy and evidence

Public receipt views contain an opaque ID, allowlisted verdict/counts and hashes of the original record. They omit subject IDs, input digests, detailed findings, tool/policy names, scanner metadata and original signatures. Keep the original authorized delivery for signature verification; the public summary is unsigned. Original signed records remain stored and may repeat submitted names. Never submit secrets. Anonymity and a deletion period are not promised. A published service privacy policy covering retention and user controls is still required before directory submission; this technical description is not that policy.

Verified release status

The caller-bound fixed-order boundary and filtered receipt contract were published and deployed on 2026-10-06. The earlier shared-credit/raw-public-receipt behavior is closed. Native versions 0.1.1 and 0.1.2 installation/discovery passed in bundled Codex 0.160.0 and Claude Code 2.1.241. Codex invoked live describe_agent and a local synthetic receipt read; Claude discovered the skill and all five tools through its zero-model native control interface. No real paid assessment or hosted ChatGPT/Claude account workflow was tested.

Version 0.1.2 adds distribution packaging, documentation and adapter display titles. The titles are now deployed through the coordinated service pin and immutable image release. Model-driven Claude Code use is blocked by its current logged-out state. Hosted authenticated use and directory review are not complete.

See ROLLOUT_STATUS.md for the completed metadata deployment, and COMPATIBILITY.md for each platform's evidence and gates, and REVIEW.md for proposed reviewer cases. Licensed under Apache-2.0.