review-query

The ONE token-query engine behind both review lists — scanner/parser/serializer, token surgery, last-wins matcher, legacy-param replay, and bounded inline autocomplete — pure JS with named exports any consumer can reuse.

raw source

GitHub's issues UI (measured live on github.com/microsoft/vscode in a real Chromium) keeps ONE source of truth — the raw query text; tabs, menus, and suggestions only rewrite tokens inside it, and the URL carries the text verbatim. SpexCode's two review lists previously split state between a substring input and structured hash params — the exact fork GitHub's model eliminates. This engine is that model as a standalone pure-JS module: no React, no DOM, importable by the pages, the route layer, and any future consumer that needs canonical review addresses (e.g. eval anchors on other boards).

expanded spec

• Scan loses nothing. The scanner splits text into whitespace runs and tokens with exact [start,end) offsets — a " swallows spaces until it closes, as a whole phrase ("long title") or a qualifier value (node:"a b") — so the highlight overlay mirrors the input glyph-for-glyph and the caret's token is findable. A token is key:value (key case-insensitive) or a bare word; serialize(tokenize(text)) is the canonical form and unknown tokens survive verbatim.
• Surgery is the only write. setToken rewrites a key at its first position, drops later duplicates, appends when absent, and removes the key on an empty value; every other token — known or not — keeps its order and spelling. Reads and the matcher agree on LAST-occurrence-wins for a duplicated qualifier.
• Matching is delegated, never duplicated. This module owns TEXT — scan/serialize, surgery, the canonical-address discipline, suggestions, and last-wins duplicate resolution (effectiveTokens). The conjunctive field matching — bare words as conjunctive substrings, the unknown-qualifier IMPOSSIBLE state that keeps the token verbatim and honestly shows nothing — lives in the ONE review-filters engine, reached through its token→state bridge; no second predicate exists here.
• Default view = bare address. queryParam compares token-normalized text against the page default (Issues is:issue state:open, Evals is:eval) — equal or empty means the bare page address, anything else exactly ?q=<raw text>. The evidence default is all with no hidden data-dependent fallback. Evals deliberately defaults to the honest whole scenario population: Fail/Pass/Unmeasured are non-exhaustive verdict: quick filters (unmeasured means declared without a reading), while state:current|reviewed remains the visible human-review lifecycle token and secondary builder. Blind, unscored, and unknown verdicts therefore remain reachable when no verdict token is active.
• Legacy params replay as the full visible state. An old structured list address rebuilds the default text with each param surgically applied — state/concluded→state:, ok=1→state:reviewed, kind→evidence: (kind=all → nothing), store/author/node/filer/verdict/freshness→their token, live=1→session:present, session=<id>→scope:<id> — and the free q appended as ONE text token, quoted whenever the tokenizer would misread it (spaces, a colon like q=drift:check, a stray quote) so the old single-substring search survives verbatim. The route layer applies this as a REPLACE; the old shape is never re-minted. One boundary is UNDECIDABLE and decided by fiat: a bare ?q= with no structured param sits byte-identical in both grammars, so it always reads as the NEW token grammar — an old bare ?q=drift:check deep link becomes an unknown-qualifier honest zero. No heuristic and no stored state may reintroduce the distinction.
• Autocomplete is bounded. A bare prefix completes qualifier keys (insert key:, keep typing); a key:prefix completes values from the page-supplied candidates only — data-derived sets, and for scope: only sessions on the current board — capped at 8. Unknown or historical values remain hand-typable and submit verbatim.