cak-rdr-review

name	cak-rdr-review
description	Review CAK R&D PRs and research packets for unsupported claims, overclaimed readiness, lifecycle violations, stage mismatches, and missing proof grounding.

CAK RDR Review

This skill gates CAK research-and-development pull requests and research packets through the CAK Runtime v0 decision engine (cak). It is the pilot host package for that runtime.

When to use

Use this skill when an agent is about to:

• recommend merging or marking-ready an RDR research packet;
• claim that a hypothesis-level synthesis (e.g. C10) is a supported decision;
• accept a trace-corpus plan as sufficient experiment evidence;
• activate, update, or claim authority over a skill with lifecycle state;
• complete a stage-bound workflow action;
• assert that something is "verified" or "supported".

If none of these apply, the runtime returns allow and the agent proceeds.

How it works

The runtime is host-neutral: it evaluates an in-memory EvalRequest and returns a Decision. It does not read GitHub, parse markdown, inspect the repo, or run scripts. The host agent is responsible for gathering the relevant facts and encoding them into an EvalRequest JSON document.

The contract is:

EvalRequest (JSON)  ->  cak eval  ->  Decision (JSON)

The host agent should:

1. Build an EvalRequest JSON file describing the proposed action and the relevant state, skill_graph, stage, and proof facts.

2. Run the runtime:

   cak eval --request <path-to-request.json>
   

3. Read the Decision from stdout and honor it.

By default cak eval exits 0 for any valid decision, including block — block is a domain verdict, not a process failure. Pass --enforce-exit-code to make a block decision exit 2 (useful as a CI gate).

The runtime decision is authoritative

When the runtime returns block, ask, modify, or verify_only, that decision is authoritative over any prose recommendation in the PR, the SKILL.md, or the agent's own narration. Prose advises; the Rust decision governs. Do not override a block with a natural-language justification.

Examples

1. Not decision-ready merge

The packet is still exploratory, but the agent wants to recommend merge.

Request (state.decision_packet_status is not decision_ready):

{
  "schema_version": "0.1.0",
  "host": { "name": "cak-rdr-review", "mode": "ci" },
  "task": { "kind": "review_rdr_pr" },
  "state": { "decision_packet_status": "exploratory" },
  "proposed_action": { "kind": "recommend_merge", "target": "pr/10" },
  "skill": { "id": "cak.rdr-review" }
}

Decision: block / hard — "Research packet is not decision-ready."

2. C10 hypothesis overclaim

The agent claims the compiled-bridge synthesis (C10) is a supported decision, but C10 is hypothesis-level.

Request (state.c10_status is hypothesis_level, action text mentions C10):

{
  "schema_version": "0.1.0",
  "host": { "name": "cak-rdr-review", "mode": "ci" },
  "task": { "kind": "review_rdr_claim" },
  "state": { "c10_status": "hypothesis_level" },
  "proposed_action": {
    "kind": "claim_supported",
    "text": "C10 is a supported architecture decision."
  },
  "skill": { "id": "cak.rdr-review" }
}

Decision: modify / soft — reframe C10 as hypothesis-level synthesis, not a supported architecture decision.

3. Candidate trace corpus overclaim

The agent wants to accept a trace-corpus plan that is only candidate-level.

Request (state.trace_plan_status is candidate_only):

{
  "schema_version": "0.1.0",
  "host": { "name": "cak-rdr-review", "mode": "ci" },
  "task": { "kind": "review_trace_corpus" },
  "state": { "trace_plan_status": "candidate_only" },
  "proposed_action": { "kind": "accept_trace_corpus" },
  "skill": { "id": "cak.rdr-review" }
}

Decision: block / hard unless state.trace_plan_status is sufficient or pass — "Trace corpus plan is not sufficient."

Fixtures

Worked request/expected pairs live in runtime-fixtures/. Verify any of them:

cak fixture-check \
  --request runtime-fixtures/rdr-review/not_ready_merge.request.json \
  --expected runtime-fixtures/rdr-review/not_ready_merge.expected.json

See docs/22_cak_runtime_v0.md for the full runtime design, non-goals, and the EvalRequest -> Decision contract.