Something went wrong. Try again.
Review the work your AI agents do. CLI + MCP.
Something went wrong. Try again.
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197import type { ClaimEntities } from "./types.js";
// Deterministic entity extraction over claim text. These extracted anchors,// file paths, commands, function names, commit hashes, PR numbers, branches,// are what the matcher keys on, so this is the highest-value part of the parser.// Each extractor is small and independently testable.
// File extensions we treat as code/doc files when a token has no slash.const CODE_EXTENSIONS = [ "py", "ts", "tsx", "js", "jsx", "mjs", "cjs", "md", "json", "html", "htm", "css", "scss", "sql", "yml", "yaml", "toml", "ini", "cfg", "sh", "bash", "go", "rs", "java", "rb", "c", "h", "cpp", "hpp", "txt", "lock", "env", "xml", "svg", "vue", "php", "kt", "swift", "proto", "dockerfile",];
// First-token shell verbs that mark a line/backtick span as a command.const SHELL_VERBS = [ "npm", "pnpm", "yarn", "npx", "python", "python3", "py", "pytest", "git", "node", "make", "curl", "wget", "bash", "sh", "docker", "docker-compose", "cd", "ls", "grep", "rg", "sed", "awk", "cat", "flask", "vitest", "tsc", "tsx", "cargo", "go", "pip", "pip3", "poetry", "ruff", "eslint", "prettier", "mypy", "deno", "bun", "kubectl", "terraform", "psql", "sqlite3",];
// Closed list of action verbs used for classification and matching. The spec's// past-tense set, plus a few base/related forms so action detection is robust.export const ACTION_VERBS = [ "edited", "edit", "changed", "change", "added", "add", "removed", "remove", "deleted", "delete", "ran", "run", "tested", "test", "fixed", "fix", "replaced", "replace", "wrote", "write", "created", "create", "updated", "update", "refactored", "refactor", "renamed", "rename", "dropped", "drop", "moved", "move", "committed", "commit", "merged", "merge", "built", "build", "shipped", "ship", "implemented", "implement", "wired", "wire",];
const codeExtAlt = CODE_EXTENSIONS.join("|");const shellVerbAlt = SHELL_VERBS.join("|");
// A path with at least one slash and a final dotted extension, e.g. o27v2/web/app.pyconst SLASH_PATH_RE = new RegExp( String.raw`(?:[\w@.+-]+\/)+[\w@.+-]+\.(?:${codeExtAlt})\b`, "gi",);// A bare filename with a known code extension, e.g. app.py, db.py, README.mdconst BARE_FILE_RE = new RegExp( String.raw`\b[\w@+-]+\.(?:${codeExtAlt})\b`, "gi",);// Special-case known filenames without a code extension. Dotfiles need their// own pattern: a leading \b never matches before a dot.const SPECIAL_FILES_RE = /\b(?:Dockerfile|Makefile)\b|(?<![\w.])\.(?:replit|gitignore|env|dockerignore)\b/g;
const BACKTICK_RE = /`([^`]+)`/g;const DOUBLE_QUOTE_RE = /"([^"]+)"|“([^”]+)”/g;const PR_RE = /#\d+\b/g;// 7-40 char hex; bare form must contain a digit to avoid matching english words.const HEX_RE = /\b[0-9a-f]{7,40}\b/gi;const CLAUDE_BRANCH_RE = /\bclaude\/[\w./-]+/g;
function dedupe(items: string[]): string[] { return [...new Set(items.map((s) => s.trim()).filter(Boolean))];}
function stripPunct(token: string): string { // Strip trailing punctuation that commonly clings to inline tokens. return token.replace(/^[("'<]+/, "").replace(/[)"'>.,;:!?]+$/, "");}
export function extractFiles(text: string): string[] { const out: string[] = []; for (const m of text.matchAll(SLASH_PATH_RE)) out.push(m[0]); for (const m of text.matchAll(BARE_FILE_RE)) out.push(m[0]); for (const m of text.matchAll(SPECIAL_FILES_RE)) out.push(m[0]); // A file token may be written with a trailing line range, e.g. app.py:1995-1999. // The base regexes already stop at the extension, so nothing extra to strip, // but normalize away any accidental trailing colon-range that slipped in. return dedupe(out.map((f) => f.replace(/:[0-9-]+$/, "")));}
export function extractCommands(text: string): string[] { const out: string[] = [];
// 1. Backtick spans whose first word is a shell verb. for (const m of text.matchAll(BACKTICK_RE)) { const inner = m[1].trim(); const first = inner.split(/\s+/)[0]; if (SHELL_VERBS.includes(first.toLowerCase())) out.push(inner); }
// 2. Lines (e.g. inside fenced blocks the caller may pass) that start with a // shell verb. Operates line by line so prose mentioning "git" mid-sentence // is not captured. const lineRe = new RegExp(String.raw`^\s*((?:${shellVerbAlt})\b[^\n\r]*)$`, "gim"); for (const m of text.matchAll(lineRe)) { const cmd = m[1].trim().replace(/`/g, ""); out.push(cmd); }
return dedupe(out);}
export function extractTools(text: string): string[] { const out: string[] = []; for (const m of text.matchAll(BACKTICK_RE)) { let inner = m[1].trim(); // Allow a single trailing () to denote a function. const fn = /^_?[A-Za-z][A-Za-z0-9_]*(\(\))?$/.test(inner); if (!fn) continue; // Skip things that are really files (have an extension) or commands. const base = inner.replace(/\(\)$/, ""); if (new RegExp(`\\.(?:${codeExtAlt})$`, "i").test(base)) continue; if (SHELL_VERBS.includes(base.toLowerCase())) continue; // Qualify as a symbol only if it looks like an identifier: has an // underscore, has parens, or is camelCase (an internal capital). const looksSymbol = base.includes("_") || /\(\)$/.test(inner) || /[a-z][A-Z]/.test(base); if (!looksSymbol) continue; out.push(inner.replace(/\(\)$/, "")); } return dedupe(out);}
export function extractCommits(text: string): string[] { const out: string[] = []; // Backticked hex tokens count even without a digit (they are explicitly marked). for (const m of text.matchAll(BACKTICK_RE)) { const inner = m[1].trim(); if (/^[0-9a-f]{7,40}$/i.test(inner)) out.push(inner.toLowerCase()); } // Bare hex tokens must contain a digit to avoid matching english words like "deedface". for (const m of text.matchAll(HEX_RE)) { const tok = m[0].toLowerCase(); if (/\d/.test(tok)) out.push(tok); } return dedupe(out);}
export function extractPrs(text: string): string[] { const out: string[] = []; for (const m of text.matchAll(PR_RE)) out.push(m[0]); return dedupe(out);}
export function extractBranches(text: string): string[] { const out: string[] = []; for (const m of text.matchAll(CLAUDE_BRANCH_RE)) out.push(stripPunct(m[0])); // Token following the word "branch". const afterBranch = /\bbranch(?:\s+name)?\s+`?([\w./-]+)`?/gi; for (const m of text.matchAll(afterBranch)) { const tok = stripPunct(m[1]); if (tok) out.push(tok); } // Backticked feature-branch shapes (a/b style without a file extension). for (const m of text.matchAll(BACKTICK_RE)) { const inner = m[1].trim(); if ( /^[\w-]+\/[\w./-]+$/.test(inner) && !new RegExp(`\\.(?:${codeExtAlt})$`, "i").test(inner) ) { out.push(inner); } } return dedupe(out);}
export function extractQuoted(text: string): string[] { const out: string[] = []; for (const m of text.matchAll(DOUBLE_QUOTE_RE)) { const inner = (m[1] ?? m[2] ?? "").trim(); if (inner) out.push(inner); } return dedupe(out);}
export function extractVerbs(text: string): string[] { const lower = text.toLowerCase(); const out: string[] = []; for (const v of ACTION_VERBS) { if (new RegExp(`\\b${v}\\b`).test(lower)) out.push(v); } return dedupe(out);}
/** Run every extractor over one piece of claim text. */export function extractEntities(text: string): ClaimEntities { return { files: extractFiles(text), commands: extractCommands(text), tools: extractTools(text), commits: extractCommits(text), prs: extractPrs(text), branches: extractBranches(text), quoted: extractQuoted(text), verbs: extractVerbs(text), };}