rag-rat: Skill for Claude Code

dream-review — triage and resolve the rag-rat dream worklist

The dream tool surfaces findings about the repo's memories and load-bearing code. It never mutates a memory — the deterministic passes and the optional model propose; a reviewer decides. This skill is that reviewer loop: for each open finding, investigate with the rag-rat MCP tools, then resolve it — preferring a root fix (repair the memory / close the coverage gap, so the finding resolves on its own next run) over a bare verdict.

Drive the whole loop through the rag-rat MCP tools — dream and dream_review for the worklist, and semantic_search / symbol_lookup / impact_surface / find_callers / history_for / memory_search / read_chunk to investigate — not grep, and not the rag-rat CLI (an agent has the MCP but may not have the binary on PATH). It is a batch chore meant to run a few times a day, not continuously.

dream and dream_review are in the MCP server's admin toolset, which is not listed by default. If they are not among your tools, tell the user to add [mcp] toolsets = ["admin"] to the repo's rag-rat.toml and reconnect the rag-rat MCP server (in Claude Code: /mcp), then continue.

1. Preflight

• Confirm the index is fresh: call index_status. If it reports drift, call heal_index to repair already-indexed files. (Discovering brand-new files and re-embedding is a CLI/cron job — rag-rat index --discover then rag-rat reconcile — outside this MCP loop; if findings look empty, that's the thing to ask a human to run.)
• The dream tool is deterministic-only. It recomputes coverage_gap + stale_reference on every call and surfaces any memory_divergence / memory_unverifiable findings a prior model run persisted — but it does not run the model itself. Generating fresh model findings is the CLI/cron rag-rat dream --verify --compact pass (it provisions a GPU and takes minutes, so it never runs from a tool call). If the divergence findings look stale, that CLI pass is what refreshes them.