Ask what the index could see before trusting a search or calling absence. Inventory of the root and what the index did with each file: indexed (in the index, matches disk), changed (differs from the index), excluded (an ignore rule or a type the index does not read; reason says which), failed (read or conversion error), unconverted (a PDF whose text layer the index could not read, or an office file with no extracted .md beside it; reason says which). A PDF is indexed from its own text layer, one section per page, unless a .md sits beside it — that sidecar is indexed instead, and both rows carry source_state: stale means the original changed after the conversion, unverifiable that the sidecar records no hash of it or its original is not under the root. index.state is cold when nothing was ever indexed: an empty listing there is not "no files".

For a question that names no identifier (what a behaviour is, where it is decided, why), call this before any grep (a Grep tool, or grep or rg in a shell), Glob or Read: it ranks passages and, in a source tree, the functions, methods, classes and constants themselves, by the question's words and their vectors. Do not grep for the words of a question; grep for an exact string or a symbol name you already know. Each hit carries its densest width-char window and line; a PDF's section is a page, so on a PDF hit widen width (1200-1500) rather than reading the page. matched_terms are the query terms in the hit; coverage is the share of the query's idf mass they carry (near 1 = every rare term matched, near 0 = only the common words); terms_absent are query terms found nowhere in the corpus. No boolean "no answer" exists: read coverage and matched_terms and decide to stop or rephrase. A hit the dense leg alone found carries dense and coverage 0. In a source tree a hit's section is the symbol, span its lines, and silica_read(path, section=<symbol>) serves its body.

Read a located slice before citing it: lines start..end, one section by heading, a PDF page, or the whole file when it fits max_chars, always with the outline and a version; truncated and next_start say how to continue. A PDF is served from its own text layer, page by page: section="p. 8" serves one page — the whole page, ~1.2k tokens (measured 2026-09-09), so ask for it to cite a passage a search located, and widen the search's width to explore — pages counts them, page_map gives the pages the slice covers, and extract_path is that text on disk, for grep and the harness's own reader (a cache, rebuilt when the PDF changes, under no version guard). A PDF or office file with a .md extracted beside it is served from that instead, and source_state says whether the original still hashes to what the conversion recorded: stale means the PDF changed and the .md did not, unverifiable that the sidecar records no hash or its original is gone. Reading the .md itself, the path a search hit carries, reports the same source and source_state. Either way source names the original, and a PDF with no readable text layer is error.unconverted. Carry version forward as expect_version to be refused instead of reading different bytes under an old citation; version is the served text's, never the original's.

Use before editing or explaining a source file, or when a traceback or a diff names file:line: one call serves the file, a #Class.member, or the symbol at #L<line>, with its supertypes, extenders, callers, the signatures it names, external dependencies and importers, inside a character budget. Static AST only, same repo state, same bytes. Check truncated before treating the target as complete; dropped names what did not fit. languages lists the parsers this install has.

Write one note atomically, as given, then lint it: lint lists structural violations and unresolved wikilinks. Refuses to overwrite when expect_version does not match or when create_only is set and the file exists. Undo is git; the core keeps no journal.