wiki-querylisted

# wiki-query — cited retrieval from a hallouminate wiki Answer a question **strictly from the wiki**, with a citation on every claim. The model is a synthesizer over retrieved chunks, never a substitute for them. If the corpus does not support a claim, say so — do not fall back to training data. **Agent topology (required):** - **Root = opus-tier** (the strongest model the harness offers). Plans the retrieval, decides what's a distinct sub-question, synthesizes the final answer, and verifies every citation. Reasoning lives here. - **Fan-out = haiku-tier** (the cheapest model). One sub-agent per sub-question. Each runs `ground`, reads the top chunks, and returns a compact cited evidence digest — never prose for the user. Retrieval noise stays in the sub-agent's context, not the root's. The root NEVER answers from memory of the codebase. It answers from what the haiku digests bring back. ## Flow ### 1. Plan (root / opus) - Restate the question and name loaded assumptions in it. - Decompose into **2–5 orthogonal sub-questions**. One retrieval angle each — splitting "how does auth work and where are tokens stored" into two beats one blurry search. Single, narrow questions skip decomposition. - Pick the corpus. If unspecified and >1 corpus exists, call `list_corpora` and ask which, or default to the repo's `repo:{name}:wiki`. - Optionally `list_tree` once to see the wiki's shape — use it to phrase searches toward the right area (progressive disclosure: navig