How to use the RDD Advisory Transport Rules agent skill

What is an agent skill and what does this one do?

An agent skill is a folder containing a SKILL.md file and optional reference documents. The agent loads it when a request matches the skill's trigger text, and then follows the rules inside it. This skill's trigger text names reviewer transport, advisory transport, review adapter, lens prompt or schema, the OpenCode reviewer plugin, the Codex reviewer and ReviewProviderContract.

Once loaded, the skill tells the agent to keep prompts, frozen evidence, result schemas, byte budgets and JSON extraction inside one Go provider contract. Adapters are kept thin: they only invoke the reviewer and return raw output with any transport error. Runtime output is treated as advisory, so only Go admission can open a gate. The skill also gives a rollout order and an output format for reporting each change.

• Keeps the Go provider contract as the single owner of binding, evidence, prompt, schema, budget and admission.
• Forbids adapters from parsing bindings, rebuilding prompts, copying schemas, applying budgets, parsing JSON or holding retry state.
• Requires a byte budget that refuses input before invocation rather than truncating it.
• Requires Codex to stay unadvertised until both positive and fail-closed proofs pass.
• Gives a six-step rollout plan and a reporting format for each slice.

How do I install the rdd-advisory-transport skill?

Download the ZIP from this page. You can install it for yourself or for a single project. For a personal install, unzip it into ~/.claude/skills/. For a project-only install, unzip it into the .claude/skills/ folder inside the project. After unzipping, the folder should contain SKILL.md along with the references folder.

If you use the Claude apps instead of the command line, open Settings, go to Capabilities, and upload the skill under Skills. Either way, the agent needs to be able to read the whole folder, because the rules point to two reference documents for full detail.

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into .claude/skills/ inside your project
• Claude apps: Settings, then Capabilities, then Skills, then upload
• Check that SKILL.md and the references folder are both present after unzipping

How do I trigger the skill with natural requests?

You do not need a special command. Ask for work that touches the areas in the trigger text, and name the runtime or the component involved. The agent should load the skill when your request matches those terms. If it does not, mention the skill by name in your request so the agent knows to use it.

Good requests are specific about the task. They say what is changing and where, for example an adapter, a prompt, a schema or a runtime plugin. Requests that only ask for a general code review will not match this skill.

• Refactor the Codex reviewer adapter so it only returns raw output.
• Use the rdd-advisory-transport skill to review this change to the lens prompt and ReviewerResultSchema.
• Plan the cutover of the OpenCode reviewer plugin under the advisory transport rules.
• Check whether this new adapter parses bindings or applies its own byte budget.

What do the hard rules and decision gates mean in practice?

The hard rules are short, strict statements. For example, an adapter may not parse JSON, a malformed or truncated result may never become PASS or a clean receipt, and no OpenCode restart or special session variables are allowed. The decision gates are simple if-then rules. If a change fits prompt, evidence, schema or validation logic, it belongs in the Go contract, not in an adapter. If a cutover slice is over 1000 lines, it should be split into chained pull requests.

The skill also says the RDD semantic pipeline is off limits for this kind of change. Lens selection, severity meanings, refuter adjudication and terminal receipts are review-lifecycle concerns and are not changed by transport work.

• Adapter cannot expose raw final output: the runtime stays marked unavailable, with no workaround that parses artifacts.
• A shared prompt change that alters any semantic outcome means stop, and the slice is blocked until it passes regression.
• Transport failures never consume a correction budget.
• Historical receipts stay readable and their provider_command and runtime_interception descriptors are never rewritten.

What does the agent report after a change?

The skill's output contract asks for a report per slice. The report lists the contract surface touched, the adapter minimality guard status, the bench corpus diff result, the semantic regression result, the capability advertisement state, and any open maintainer decision. An example of an open decision given in the skill is the receipt provider_contract descriptor.

• Contract surface touched
• Adapter minimality guard status
• Bench corpus diff result
• Semantic regression result
• Capability advertisement state
• Unresolved maintainer decisions