Full-text search across the local Title 38 (VA) CFR index using ranked keyword matching. Returns section citations, headings, relevance snippets, bm25 scores (lower is better) and the query terms visible in each hit. The CANDIDATE SET is every provision matching AT LEAST ONE query term, so a provision that answers part of the question stays in the results instead of being excluded by the words it does not use; total counts those candidates (plus the pinned citation when the candidate set did not already hold it). The hits are ordered by: (1) a hit whose HEADING carries the query's content terms as a phrase (or the corpus phrase a synonym stands for), (2) a hit matching EVERY query term ahead of one matching only some, (3) bm25, lower is better. Use part to scope results to a single CFR part (e.g. "4" for ratings, "17" for prosthetics), kind to restrict results to one provision kind ("section" or "appendix"), and chapter to scope to one chapter by its Roman-numeral identifier (e.g. "I"; exact, case-sensitive). Follow up with get_section to read a hit in full. If the query itself is a citation (e.g. "4.130" or "38 CFR 4.16"), that exact section is pinned to the top of page 1 (totals include the pin on every page) and marked direct: true — the reliable client signal; the pinned hit's inResults flag reports whether it also matched the ranked results. Quoted spans (e.g. "exact regulatory language") are emitted as FTS5 phrase queries — the last token gets a prefix wildcard (so "language" also matches "languages"); earlier tokens are matched exactly. Hits are ranked by LEXICAL OVERLAP with the query, not by legal authority: the first hit is the provision sharing the most words with the query, which is often not the provision that governs the question. Read each hit's heading and fetch its full text with get_section before you cite it. Each response also carries matchedTerms and unmatchedTerms: the query's content words (stopwords, single characters and bare numerals removed), split into those the returned hits answered and those they did not. A non-empty unmatchedTerms means the returned text does not speak to that part of your question — the governing provision may be a later hit, may be reachable only under different wording, or may not be in this index. Never cite a provision you have not read in full, and never state that Title 38 does not address a question on the basis of this response alone.

Fetch the full text of one Title 38 CFR section or appendix as markdown. Accepts flexible citations: "4.125", "§ 4.125", "38 CFR 4.125", or appendix names like "Appendix A to Part 4". Returns heading, hierarchy path, issue date and an ecfr.gov permalink.

Navigate the Title 38 hierarchy (title → chapter → part → subpart → subject group) and list what exists at a level. With no arguments, returns the chapters. Use this to discover valid parts/subparts before searching. Most sections live inside a subject group: pass the subject_group identifier you saw in a previous response to list the sections it holds.

Report how current the local Title 38 index is: eCFR issue date, last sync time, row counts, the corpus receipt hash (a sha256 over the sorted per-provision (citation, text+heading) pairs at rebuild time, null on older indexes) and whether the index is stale relative to CFR_MAX_AGE_DAYS. Check this before relying on search results.

Enumerate every Title 38 CFR part present in the local index, with its structure-tree title and section/appendix counts. Use this to discover which parts are searchable.

One-call readiness verdict for the local index: "ready" (query away), "stale" (older than CFR_MAX_AGE_DAYS — verify results against the current eCFR), "degraded" (structure drift — provisions may be missing; verify against the current eCFR), or "empty" (no index yet — run cfr-free sync). Clients should gate queries on this before spending tool calls.

One Title 38 provision as it stood at a requested date (YYYY-MM-DD, required). eCFR serves the revision IN FORCE on that date, so a date that is not itself an issue date resolves to the revision that was in force then; the upstream does not report which issue date it served, so the response names the REQUESTED date and never a served-revision date. Stateless: the revision is fetched from eCFR on every call — no local index, no cache. eCFR's point-in-time archive starts at 2017-01-01; a date before that is rejected as not_found. Use this for date-of-claim and CUE questions ("what did this section say when the claim was filed?"); for the CURRENT text use get_section.

Provision-level diff of Title 38 between two eCFR issue dates: which citations changed, which were added, and which were removed between from_date and to_date (both YYYY-MM-DD, both required — this tool has no default-to; for "what changed since my index was built" run cfr-free diff <from> locally, which resolves to from the index issue date). Stateless: both dates' full-title XML is fetched from eCFR and hashed per citation on the fly — no local index needed, no cache. Each citation list is capped at 100 entries; truncated: true means at least one list was cut. Identical dates or an impossible calendar date (e.g. month 13) are rejected as errors before any fetch.

Line-level redline of one Title 38 provision between two eCFR issue dates: which lines of citation changed, were added, or were removed between from_date and to_date (both YYYY-MM-DD, both required — this tool has no default-to). The dates are SELECTORS — eCFR serves the revision in force on the requested date; they are not a verified stamp of which issue date was served. Returns the status (changed/added/removed/unchanged), both sides' raw text, and one hunk per line (same/add/del with 1-based anchors into each side's text). Identical or impossible dates (e.g. month 13) are rejected as errors before any fetch; a citation missing from BOTH dates is a not_found error after the fetches. Stateless: both dates' full-title XML is fetched from eCFR per call — no local index needed, no cache. When either side exceeds the core's line cap the per-line walk is replaced by two whole-text coarse blocks (coarse: true) — passed through uncapped here, the core's own bound.