@smeltjs/core 0.2.0 → 0.3.0

package/dist/repomap/tags.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"tags.d.ts","sourceRoot":"","sources":["../../src/repomap/tags.ts"],"names":[],"mappings":"AAMA,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAE9C;;;;;;;;;;;;;;;GAeG;AAEH,wEAAwE;AACxE,MAAM,WAAW,aAAa;IAC5B,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,uFAAuF;IACvF,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;CACvB;AAED,4EAA4E;AAC5E,MAAM,WAAW,YAAY;IAC3B,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;CACxB;AAED,4DAA4D;AAC5D,MAAM,WAAW,QAAQ;IACvB,yBAAyB;IACzB,QAAQ,CAAC,IAAI,EAAE,SAAS,aAAa,EAAE,CAAC;IACxC,qEAAqE;IACrE,QAAQ,CAAC,IAAI,EAAE,SAAS,YAAY,EAAE,CAAC;CACxC;AAiCD;;;;;;;;GAQG;AACH,wBAAsB,WAAW,CAAC,IAAI,EAAE,MAAM,EAAE,QAAQ,EAAE,UAAU,GAAG,OAAO,CAAC,QAAQ,CAAC,CA8DvF"}
+{"version":3,"file":"tags.d.ts","sourceRoot":"","sources":["../../src/repomap/tags.ts"],"names":[],"mappings":"AAMA,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAE9C;;;;;;;;;;;;;;;GAeG;AAEH,wEAAwE;AACxE,MAAM,WAAW,aAAa;IAC5B,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,uFAAuF;IACvF,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;CACvB;AAED,4EAA4E;AAC5E,MAAM,WAAW,YAAY;IAC3B,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;CACxB;AAED,4DAA4D;AAC5D,MAAM,WAAW,QAAQ;IACvB,yBAAyB;IACzB,QAAQ,CAAC,IAAI,EAAE,SAAS,aAAa,EAAE,CAAC;IACxC,qEAAqE;IACrE,QAAQ,CAAC,IAAI,EAAE,SAAS,YAAY,EAAE,CAAC;CACxC;AAiCD;;;;;;;;GAQG;AACH,wBAAsB,WAAW,CAAC,IAAI,EAAE,MAAM,EAAE,QAAQ,EAAE,UAAU,GAAG,OAAO,CAAC,QAAQ,CAAC,CA8DvF","sourcesContent":["import { Parser } from 'web-tree-sitter';\nimport type { Node } from 'web-tree-sitter';\n\nimport { GrammarUnavailableError } from '../errors.ts';\nimport { profileFor } from '../lang/registry.ts';\nimport { loadGrammar } from '../plan/grammar.ts';\nimport type { LanguageId } from '../types.ts';\n\n/**\n * Definition/reference tag extraction for the repo map — one file at a time, through\n * the same bundled tree-sitter grammars the structural planner uses.\n *\n * Modelled on Aider's repo-map tags (https://aider.chat/docs/repomap.html): a *def* is\n * a named declaration read off the parse tree, a *ref* is an identifier occurrence.\n * Aider extracts both with per-language `.scm` query files; here the same facts come\n * from a manual tree walk over each language's `repomap` profile section\n * (`src/lang/`), so no query assets are added. The design is Aider's, not this\n * project's.\n *\n * Honest scope, stated rather than implied: refs count `identifier`/`type_identifier`\n * nodes only, so a member access like `config.load()` contributes `config`, not\n * `load` — method references travel through the object they are called on. A def whose\n * name is a destructuring pattern is skipped rather than guessed at.\n */\n\n/** One named definition, read off the parse tree. `line` is 1-based. */\nexport interface DefinitionTag {\n  readonly name: string;\n  /** Human word for the node kind, e.g. `'function'` — same register as the planners. */\n  readonly kind: string;\n  readonly line: number;\n}\n\n/** One referenced identifier and how many times this file references it. */\nexport interface ReferenceTag {\n  readonly name: string;\n  readonly count: number;\n}\n\n/** Everything the repo map needs to know about one file. */\nexport interface FileTags {\n  /** In document order. */\n  readonly defs: readonly DefinitionTag[];\n  /** Sorted by name, so the extraction is deterministic end to end. */\n  readonly refs: readonly ReferenceTag[];\n}\n\n/**\n * C/C++ specifier node types that name a *usage site* as readily as a definition:\n * `struct point p;` parses as a `struct_specifier` with a `name` and no `body`, same\n * as the `struct point { … };` that actually defines it. Only the bodied form is a\n * definition. Without this check a mere mention earns a `defined at` receipt, *and*\n * its name node lands in `defNameStarts` — so the real definition's cross-file\n * references silently vanish from the map.\n */\nconst BODY_REQUIRED_TYPES: ReadonlySet<string> = new Set([\n  'struct_specifier',\n  'union_specifier',\n  'enum_specifier',\n  'class_specifier',\n]);\n\n/**\n * A local `const x = …` inside a function body is not a map-worthy symbol. A\n * `variable_declarator` counts as a definition only at module top level — directly\n * under the program node, or under an `export` statement that is.\n */\nfunction isDefinitionSite(node: Node): boolean {\n  if (BODY_REQUIRED_TYPES.has(node.type) && node.childForFieldName('body') === null) {\n    return false; // a bodiless specifier is a usage or forward declaration, not a definition\n  }\n  if (node.type !== 'variable_declarator') return true;\n  const container = node.parent?.parent;\n  if (container === null || container === undefined) return false;\n  if (container.type === 'program') return true;\n  return container.type === 'export_statement' && container.parent?.type === 'program';\n}\n\n/**\n * Extract definition and reference tags from one file.\n *\n * Deterministic: same text, same language, identical tags — defs in document order,\n * refs sorted by name. A def's own name node is not counted as a reference to itself.\n *\n * @throws {GrammarUnavailableError} when the grammar cannot load or produces no tree.\n *   Never a lexical guess — same no-fallback rule as the structural planner.\n */\nexport async function extractTags(text: string, language: LanguageId): Promise<FileTags> {\n  const grammar = await loadGrammar(language);\n  // The per-language facts live on the language's profile (`src/lang/`). A profile\n  // without a repomap section contributes nothing — no symbols, never wrong ones.\n  const repomap = profileFor(language).repomap;\n  const defKinds = repomap?.defKinds ?? {};\n  const refTypes = repomap?.refTypes ?? [];\n\n  const parser = new Parser();\n  let tree = null;\n  try {\n    parser.setLanguage(grammar);\n    tree = parser.parse(text);\n    if (tree === null) {\n      throw new GrammarUnavailableError(\n        `smelt: the ${language} parser returned no tree, so no tags can be extracted. ` +\n          `Refusing to guess at definitions that were never parsed.`,\n      );\n    }\n\n    const defs: DefinitionTag[] = [];\n    const refCounts = new Map<string, number>();\n    /** Start indices of def-name nodes, so a definition never references itself. */\n    const defNameStarts = new Set<number>();\n\n    // Depth-first, document order, via an explicit stack — recursion depth on a real\n    // file is the tree's depth, which nothing bounds.\n    const stack: Node[] = [tree.rootNode];\n    while (stack.length > 0) {\n      const node = stack.pop()!;\n\n      const kind = defKinds[node.type];\n      if (kind !== undefined && isDefinitionSite(node)) {\n        const nameNode = node.childForFieldName('name');\n        // Only a plain identifier names a def; a destructuring pattern is skipped, not\n        // guessed at.\n        if (nameNode !== null && nameNode.type.endsWith('identifier')) {\n          defs.push({ name: nameNode.text, kind, line: node.startPosition.row + 1 });\n          defNameStarts.add(nameNode.startIndex);\n        }\n      }\n\n      if (refTypes.includes(node.type) && !defNameStarts.has(node.startIndex)) {\n        refCounts.set(node.text, (refCounts.get(node.text) ?? 0) + 1);\n      }\n\n      const children = node.namedChildren;\n      for (let i = children.length - 1; i >= 0; i -= 1) {\n        const child = children[i];\n        if (child !== null && child !== undefined) stack.push(child);\n      }\n    }\n\n    const refs = [...refCounts.entries()]\n      .toSorted(([a], [b]) => (a < b ? -1 : a > b ? 1 : 0))\n      .map(([name, count]) => ({ name, count }));\n\n    return { defs, refs };\n  } finally {\n    tree?.delete();\n    parser.delete();\n  }\n}\n"]}

package/dist/repomap/tags.js.map

@@ -1 +1 @@
-{"version":3,"file":"tags.js","sourceRoot":"","sources":["../../src/repomap/tags.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,MAAM,EAAE,MAAM,iBAAiB,CAAC;AAGzC,OAAO,EAAE,uBAAuB,EAAE,MAAM,cAAc,CAAC;AACvD,OAAO,EAAE,UAAU,EAAE,MAAM,qBAAqB,CAAC;AACjD,OAAO,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AA0CjD;;;;;;;GAOG;AACH,MAAM,mBAAmB,GAAwB,IAAI,GAAG,CAAC;IACvD,kBAAkB;IAClB,iBAAiB;IACjB,gBAAgB;IAChB,iBAAiB;CAClB,CAAC,CAAC;AAEH;;;;GAIG;AACH,SAAS,gBAAgB,CAAC,IAAU;IAClC,IAAI,mBAAmB,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,IAAI,CAAC,iBAAiB,CAAC,MAAM,CAAC,KAAK,IAAI,EAAE,CAAC;QAClF,OAAO,KAAK,CAAC,CAAC,2EAA2E;IAC3F,CAAC;IACD,IAAI,IAAI,CAAC,IAAI,KAAK,qBAAqB;QAAE,OAAO,IAAI,CAAC;IACrD,MAAM,SAAS,GAAG,IAAI,CAAC,MAAM,EAAE,MAAM,CAAC;IACtC,IAAI,SAAS,KAAK,IAAI,IAAI,SAAS,KAAK,SAAS;QAAE,OAAO,KAAK,CAAC;IAChE,IAAI,SAAS,CAAC,IAAI,KAAK,SAAS;QAAE,OAAO,IAAI,CAAC;IAC9C,OAAO,SAAS,CAAC,IAAI,KAAK,kBAAkB,IAAI,SAAS,CAAC,MAAM,EAAE,IAAI,KAAK,SAAS,CAAC;AACvF,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,CAAC,KAAK,UAAU,WAAW,CAAC,IAAY,EAAE,QAAoB;IAClE,MAAM,OAAO,GAAG,MAAM,WAAW,CAAC,QAAQ,CAAC,CAAC;IAC5C,iFAAiF;IACjF,gFAAgF;IAChF,MAAM,OAAO,GAAG,UAAU,CAAC,QAAQ,CAAC,CAAC,OAAO,CAAC;IAC7C,MAAM,QAAQ,GAAG,OAAO,EAAE,QAAQ,IAAI,EAAE,CAAC;IACzC,MAAM,QAAQ,GAAG,OAAO,EAAE,QAAQ,IAAI,EAAE,CAAC;IAEzC,MAAM,MAAM,GAAG,IAAI,MAAM,EAAE,CAAC;IAC5B,IAAI,IAAI,GAAG,IAAI,CAAC;IAChB,IAAI,CAAC;QACH,MAAM,CAAC,WAAW,CAAC,OAAO,CAAC,CAAC;QAC5B,IAAI,GAAG,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;QAC1B,IAAI,IAAI,KAAK,IAAI,EAAE,CAAC;YAClB,MAAM,IAAI,uBAAuB,CAC/B,cAAc,QAAQ,yDAAyD;gBAC7E,0DAA0D,CAC7D,CAAC;QACJ,CAAC;QAED,MAAM,IAAI,GAAoB,EAAE,CAAC;QACjC,MAAM,SAAS,GAAG,IAAI,GAAG,EAAkB,CAAC;QAC5C,gFAAgF;QAChF,MAAM,aAAa,GAAG,IAAI,GAAG,EAAU,CAAC;QAExC,iFAAiF;QACjF,kDAAkD;QAClD,MAAM,KAAK,GAAW,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;QACtC,OAAO,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACxB,MAAM,IAAI,GAAG,KAAK,CAAC,GAAG,EAAG,CAAC;YAE1B,MAAM,IAAI,GAAG,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;YACjC,IAAI,IAAI,KAAK,SAAS,IAAI,gBAAgB,CAAC,IAAI,CAAC,EAAE,CAAC;gBACjD,MAAM,QAAQ,GAAG,IAAI,CAAC,iBAAiB,CAAC,MAAM,CAAC,CAAC;gBAChD,+EAA+E;gBAC/E,cAAc;gBACd,IAAI,QAAQ,KAAK,IAAI,IAAI,QAAQ,CAAC,IAAI,CAAC,QAAQ,CAAC,YAAY,CAAC,EAAE,CAAC;oBAC9D,IAAI,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,QAAQ,CAAC,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,CAAC,aAAa,CAAC,GAAG,GAAG,CAAC,EAAE,CAAC,CAAC;oBAC3E,aAAa,CAAC,GAAG,CAAC,QAAQ,CAAC,UAAU,CAAC,CAAC;gBACzC,CAAC;YACH,CAAC;YAED,IAAI,QAAQ,CAAC,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,aAAa,CAAC,GAAG,CAAC,IAAI,CAAC,UAAU,CAAC,EAAE,CAAC;gBACxE,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;YAChE,CAAC;YAED,MAAM,QAAQ,GAAG,IAAI,CAAC,aAAa,CAAC;YACpC,KAAK,IAAI,CAAC,GAAG,QAAQ,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC;gBACjD,MAAM,KAAK,GAAG,QAAQ,CAAC,CAAC,CAAC,CAAC;gBAC1B,IAAI,KAAK,KAAK,IAAI,IAAI,KAAK,KAAK,SAAS;oBAAE,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;YAC/D,CAAC;QACH,CAAC;QAED,MAAM,IAAI,GAAG,CAAC,GAAG,SAAS,CAAC,OAAO,EAAE,CAAC;aAClC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;aACpD,GAAG,CAAC,CAAC,CAAC,IAAI,EAAE,KAAK,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC;QAE7C,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC;IACxB,CAAC;YAAS,CAAC;QACT,IAAI,EAAE,MAAM,EAAE,CAAC;QACf,MAAM,CAAC,MAAM,EAAE,CAAC;IAClB,CAAC;AACH,CAAC"}
+{"version":3,"file":"tags.js","sourceRoot":"","sources":["../../src/repomap/tags.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,MAAM,EAAE,MAAM,iBAAiB,CAAC;AAGzC,OAAO,EAAE,uBAAuB,EAAE,MAAM,cAAc,CAAC;AACvD,OAAO,EAAE,UAAU,EAAE,MAAM,qBAAqB,CAAC;AACjD,OAAO,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AA0CjD;;;;;;;GAOG;AACH,MAAM,mBAAmB,GAAwB,IAAI,GAAG,CAAC;IACvD,kBAAkB;IAClB,iBAAiB;IACjB,gBAAgB;IAChB,iBAAiB;CAClB,CAAC,CAAC;AAEH;;;;GAIG;AACH,SAAS,gBAAgB,CAAC,IAAU;IAClC,IAAI,mBAAmB,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,IAAI,CAAC,iBAAiB,CAAC,MAAM,CAAC,KAAK,IAAI,EAAE,CAAC;QAClF,OAAO,KAAK,CAAC,CAAC,2EAA2E;IAC3F,CAAC;IACD,IAAI,IAAI,CAAC,IAAI,KAAK,qBAAqB;QAAE,OAAO,IAAI,CAAC;IACrD,MAAM,SAAS,GAAG,IAAI,CAAC,MAAM,EAAE,MAAM,CAAC;IACtC,IAAI,SAAS,KAAK,IAAI,IAAI,SAAS,KAAK,SAAS;QAAE,OAAO,KAAK,CAAC;IAChE,IAAI,SAAS,CAAC,IAAI,KAAK,SAAS;QAAE,OAAO,IAAI,CAAC;IAC9C,OAAO,SAAS,CAAC,IAAI,KAAK,kBAAkB,IAAI,SAAS,CAAC,MAAM,EAAE,IAAI,KAAK,SAAS,CAAC;AACvF,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,CAAC,KAAK,UAAU,WAAW,CAAC,IAAY,EAAE,QAAoB;IAClE,MAAM,OAAO,GAAG,MAAM,WAAW,CAAC,QAAQ,CAAC,CAAC;IAC5C,iFAAiF;IACjF,gFAAgF;IAChF,MAAM,OAAO,GAAG,UAAU,CAAC,QAAQ,CAAC,CAAC,OAAO,CAAC;IAC7C,MAAM,QAAQ,GAAG,OAAO,EAAE,QAAQ,IAAI,EAAE,CAAC;IACzC,MAAM,QAAQ,GAAG,OAAO,EAAE,QAAQ,IAAI,EAAE,CAAC;IAEzC,MAAM,MAAM,GAAG,IAAI,MAAM,EAAE,CAAC;IAC5B,IAAI,IAAI,GAAG,IAAI,CAAC;IAChB,IAAI,CAAC;QACH,MAAM,CAAC,WAAW,CAAC,OAAO,CAAC,CAAC;QAC5B,IAAI,GAAG,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;QAC1B,IAAI,IAAI,KAAK,IAAI,EAAE,CAAC;YAClB,MAAM,IAAI,uBAAuB,CAC/B,cAAc,QAAQ,yDAAyD;gBAC7E,0DAA0D,CAC7D,CAAC;QACJ,CAAC;QAED,MAAM,IAAI,GAAoB,EAAE,CAAC;QACjC,MAAM,SAAS,GAAG,IAAI,GAAG,EAAkB,CAAC;QAC5C,gFAAgF;QAChF,MAAM,aAAa,GAAG,IAAI,GAAG,EAAU,CAAC;QAExC,iFAAiF;QACjF,kDAAkD;QAClD,MAAM,KAAK,GAAW,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;QACtC,OAAO,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACxB,MAAM,IAAI,GAAG,KAAK,CAAC,GAAG,EAAG,CAAC;YAE1B,MAAM,IAAI,GAAG,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;YACjC,IAAI,IAAI,KAAK,SAAS,IAAI,gBAAgB,CAAC,IAAI,CAAC,EAAE,CAAC;gBACjD,MAAM,QAAQ,GAAG,IAAI,CAAC,iBAAiB,CAAC,MAAM,CAAC,CAAC;gBAChD,+EAA+E;gBAC/E,cAAc;gBACd,IAAI,QAAQ,KAAK,IAAI,IAAI,QAAQ,CAAC,IAAI,CAAC,QAAQ,CAAC,YAAY,CAAC,EAAE,CAAC;oBAC9D,IAAI,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,QAAQ,CAAC,IAAI,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,CAAC,aAAa,CAAC,GAAG,GAAG,CAAC,EAAE,CAAC,CAAC;oBAC3E,aAAa,CAAC,GAAG,CAAC,QAAQ,CAAC,UAAU,CAAC,CAAC;gBACzC,CAAC;YACH,CAAC;YAED,IAAI,QAAQ,CAAC,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,aAAa,CAAC,GAAG,CAAC,IAAI,CAAC,UAAU,CAAC,EAAE,CAAC;gBACxE,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;YAChE,CAAC;YAED,MAAM,QAAQ,GAAG,IAAI,CAAC,aAAa,CAAC;YACpC,KAAK,IAAI,CAAC,GAAG,QAAQ,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC;gBACjD,MAAM,KAAK,GAAG,QAAQ,CAAC,CAAC,CAAC,CAAC;gBAC1B,IAAI,KAAK,KAAK,IAAI,IAAI,KAAK,KAAK,SAAS;oBAAE,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;YAC/D,CAAC;QACH,CAAC;QAED,MAAM,IAAI,GAAG,CAAC,GAAG,SAAS,CAAC,OAAO,EAAE,CAAC;aAClC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;aACpD,GAAG,CAAC,CAAC,CAAC,IAAI,EAAE,KAAK,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC;QAE7C,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC;IACxB,CAAC;YAAS,CAAC;QACT,IAAI,EAAE,MAAM,EAAE,CAAC;QACf,MAAM,CAAC,MAAM,EAAE,CAAC;IAClB,CAAC;AACH,CAAC","sourcesContent":["import { Parser } from 'web-tree-sitter';\nimport type { Node } from 'web-tree-sitter';\n\nimport { GrammarUnavailableError } from '../errors.ts';\nimport { profileFor } from '../lang/registry.ts';\nimport { loadGrammar } from '../plan/grammar.ts';\nimport type { LanguageId } from '../types.ts';\n\n/**\n * Definition/reference tag extraction for the repo map — one file at a time, through\n * the same bundled tree-sitter grammars the structural planner uses.\n *\n * Modelled on Aider's repo-map tags (https://aider.chat/docs/repomap.html): a *def* is\n * a named declaration read off the parse tree, a *ref* is an identifier occurrence.\n * Aider extracts both with per-language `.scm` query files; here the same facts come\n * from a manual tree walk over each language's `repomap` profile section\n * (`src/lang/`), so no query assets are added. The design is Aider's, not this\n * project's.\n *\n * Honest scope, stated rather than implied: refs count `identifier`/`type_identifier`\n * nodes only, so a member access like `config.load()` contributes `config`, not\n * `load` — method references travel through the object they are called on. A def whose\n * name is a destructuring pattern is skipped rather than guessed at.\n */\n\n/** One named definition, read off the parse tree. `line` is 1-based. */\nexport interface DefinitionTag {\n  readonly name: string;\n  /** Human word for the node kind, e.g. `'function'` — same register as the planners. */\n  readonly kind: string;\n  readonly line: number;\n}\n\n/** One referenced identifier and how many times this file references it. */\nexport interface ReferenceTag {\n  readonly name: string;\n  readonly count: number;\n}\n\n/** Everything the repo map needs to know about one file. */\nexport interface FileTags {\n  /** In document order. */\n  readonly defs: readonly DefinitionTag[];\n  /** Sorted by name, so the extraction is deterministic end to end. */\n  readonly refs: readonly ReferenceTag[];\n}\n\n/**\n * C/C++ specifier node types that name a *usage site* as readily as a definition:\n * `struct point p;` parses as a `struct_specifier` with a `name` and no `body`, same\n * as the `struct point { … };` that actually defines it. Only the bodied form is a\n * definition. Without this check a mere mention earns a `defined at` receipt, *and*\n * its name node lands in `defNameStarts` — so the real definition's cross-file\n * references silently vanish from the map.\n */\nconst BODY_REQUIRED_TYPES: ReadonlySet<string> = new Set([\n  'struct_specifier',\n  'union_specifier',\n  'enum_specifier',\n  'class_specifier',\n]);\n\n/**\n * A local `const x = …` inside a function body is not a map-worthy symbol. A\n * `variable_declarator` counts as a definition only at module top level — directly\n * under the program node, or under an `export` statement that is.\n */\nfunction isDefinitionSite(node: Node): boolean {\n  if (BODY_REQUIRED_TYPES.has(node.type) && node.childForFieldName('body') === null) {\n    return false; // a bodiless specifier is a usage or forward declaration, not a definition\n  }\n  if (node.type !== 'variable_declarator') return true;\n  const container = node.parent?.parent;\n  if (container === null || container === undefined) return false;\n  if (container.type === 'program') return true;\n  return container.type === 'export_statement' && container.parent?.type === 'program';\n}\n\n/**\n * Extract definition and reference tags from one file.\n *\n * Deterministic: same text, same language, identical tags — defs in document order,\n * refs sorted by name. A def's own name node is not counted as a reference to itself.\n *\n * @throws {GrammarUnavailableError} when the grammar cannot load or produces no tree.\n *   Never a lexical guess — same no-fallback rule as the structural planner.\n */\nexport async function extractTags(text: string, language: LanguageId): Promise<FileTags> {\n  const grammar = await loadGrammar(language);\n  // The per-language facts live on the language's profile (`src/lang/`). A profile\n  // without a repomap section contributes nothing — no symbols, never wrong ones.\n  const repomap = profileFor(language).repomap;\n  const defKinds = repomap?.defKinds ?? {};\n  const refTypes = repomap?.refTypes ?? [];\n\n  const parser = new Parser();\n  let tree = null;\n  try {\n    parser.setLanguage(grammar);\n    tree = parser.parse(text);\n    if (tree === null) {\n      throw new GrammarUnavailableError(\n        `smelt: the ${language} parser returned no tree, so no tags can be extracted. ` +\n          `Refusing to guess at definitions that were never parsed.`,\n      );\n    }\n\n    const defs: DefinitionTag[] = [];\n    const refCounts = new Map<string, number>();\n    /** Start indices of def-name nodes, so a definition never references itself. */\n    const defNameStarts = new Set<number>();\n\n    // Depth-first, document order, via an explicit stack — recursion depth on a real\n    // file is the tree's depth, which nothing bounds.\n    const stack: Node[] = [tree.rootNode];\n    while (stack.length > 0) {\n      const node = stack.pop()!;\n\n      const kind = defKinds[node.type];\n      if (kind !== undefined && isDefinitionSite(node)) {\n        const nameNode = node.childForFieldName('name');\n        // Only a plain identifier names a def; a destructuring pattern is skipped, not\n        // guessed at.\n        if (nameNode !== null && nameNode.type.endsWith('identifier')) {\n          defs.push({ name: nameNode.text, kind, line: node.startPosition.row + 1 });\n          defNameStarts.add(nameNode.startIndex);\n        }\n      }\n\n      if (refTypes.includes(node.type) && !defNameStarts.has(node.startIndex)) {\n        refCounts.set(node.text, (refCounts.get(node.text) ?? 0) + 1);\n      }\n\n      const children = node.namedChildren;\n      for (let i = children.length - 1; i >= 0; i -= 1) {\n        const child = children[i];\n        if (child !== null && child !== undefined) stack.push(child);\n      }\n    }\n\n    const refs = [...refCounts.entries()]\n      .toSorted(([a], [b]) => (a < b ? -1 : a > b ? 1 : 0))\n      .map(([name, count]) => ({ name, count }));\n\n    return { defs, refs };\n  } finally {\n    tree?.delete();\n    parser.delete();\n  }\n}\n"]}

package/dist/retrieve.d.ts

@@ -9,6 +9,14 @@ export declare const RETRIEVE_TOOL_NAME = "smelt_retrieve";
  * confident answer built on material it never saw, and the retrieve rate reads 0% —
  * which is indistinguishable from perfect pruning. Encouraging retrieval keeps the
  * signal in {@link ElisionStore.stats} honest.
+ *
+ * **The schema is strict-mode shaped on purpose.** `additionalProperties: false`, and
+ * `required` naming every property — the two rules OpenAI's structured-outputs strict
+ * mode enforces before it will register a function at all. Without them a whole class
+ * of consumer simply cannot expose this tool. It says nothing new: `hash` was always
+ * the only key `invoke` reads, and an extra key was always ignored. A schema that is
+ * strictly more precise about the same shape is not a change to the wire surface the
+ * tool name and behaviour guarantee covers.
  */
 export declare function createRetrieveTool(store: ElisionStore): RetrieveTool;
 //# sourceMappingURL=retrieve.d.ts.map

package/dist/retrieve.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"retrieve.d.ts","sourceRoot":"","sources":["../src/retrieve.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,YAAY,EAAE,YAAY,EAAE,MAAM,YAAY,CAAC;AAE7D,yFAAyF;AACzF,eAAO,MAAM,kBAAkB,mBAAmB,CAAC;AAwBnD;;;;;;;;GAQG;AACH,wBAAgB,kBAAkB,CAAC,KAAK,EAAE,YAAY,GAAG,YAAY,CAWpE"}
+{"version":3,"file":"retrieve.d.ts","sourceRoot":"","sources":["../src/retrieve.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,YAAY,EAAE,YAAY,EAAE,MAAM,YAAY,CAAC;AAE7D,yFAAyF;AACzF,eAAO,MAAM,kBAAkB,mBAAmB,CAAC;AAwBnD;;;;;;;;;;;;;;;;GAgBG;AACH,wBAAgB,kBAAkB,CAAC,KAAK,EAAE,YAAY,GAAG,YAAY,CAcpE","sourcesContent":["import { defaultMarker } from './apply.ts';\nimport type { ElisionStore, RetrieveTool } from './types.ts';\n\n/** The tool name smelt's markers reference. Consumers hard-code it; do not rename it. */\nexport const RETRIEVE_TOOL_NAME = 'smelt_retrieve';\n\n/**\n * The example marker inside the tool description is *rendered by the real marker\n * builder*, never hand-written — the description is the one string a model reads to\n * recognize markers, and an example whose shape drifted from the wire format\n * (`<<smelt: …>>` when real markers say `<<smelt/v1: …>>`) would teach the model to\n * miss every marker it actually receives. `test/guards/marker-format.test.ts` pins\n * this to `MARKER_FORMAT_VERSION`.\n */\nconst EXAMPLE_MARKER = defaultMarker({\n  hash: 'a1b2c3d4e5f60718',\n  bytes: 412,\n  rule: 'sibling-collapse',\n  explanation: 'collapsed 3 sibling functions',\n});\n\nconst DESCRIPTION =\n  'Return the exact original text that was elided from a previous tool result. ' +\n  `Context you were given may contain markers like \\`${EXAMPLE_MARKER}\\`. ` +\n  'Call this with that hash to get those bytes ' +\n  'back verbatim. Nothing was deleted — it is all still here. Ask whenever the elided ' +\n  'material might matter; guessing at what a marker hid is never correct.';\n\n/**\n * Wrap a store as the tool a model calls.\n *\n * Why the description says \"ask whenever it might matter\": under-retrieval is the\n * failure mode that looks like success. A model that never calls this produces a\n * confident answer built on material it never saw, and the retrieve rate reads 0% —\n * which is indistinguishable from perfect pruning. Encouraging retrieval keeps the\n * signal in {@link ElisionStore.stats} honest.\n *\n * **The schema is strict-mode shaped on purpose.** `additionalProperties: false`, and\n * `required` naming every property — the two rules OpenAI's structured-outputs strict\n * mode enforces before it will register a function at all. Without them a whole class\n * of consumer simply cannot expose this tool. It says nothing new: `hash` was always\n * the only key `invoke` reads, and an extra key was always ignored. A schema that is\n * strictly more precise about the same shape is not a change to the wire surface the\n * tool name and behaviour guarantee covers.\n */\nexport function createRetrieveTool(store: ElisionStore): RetrieveTool {\n  return {\n    name: RETRIEVE_TOOL_NAME,\n    description: DESCRIPTION,\n    inputSchema: {\n      type: 'object',\n      properties: {\n        hash: { type: 'string', description: 'The hash from a marker\\'s retrieve(\"hash\").' },\n      },\n      required: ['hash'],\n      additionalProperties: false,\n    },\n    invoke: ({ hash }) => store.retrieve(hash),\n  };\n}\n"]}

package/dist/retrieve.js

@@ -28,6 +28,14 @@ const DESCRIPTION = 'Return the exact original text that was elided from a previ
  * confident answer built on material it never saw, and the retrieve rate reads 0% —
  * which is indistinguishable from perfect pruning. Encouraging retrieval keeps the
  * signal in {@link ElisionStore.stats} honest.
+ *
+ * **The schema is strict-mode shaped on purpose.** `additionalProperties: false`, and
+ * `required` naming every property — the two rules OpenAI's structured-outputs strict
+ * mode enforces before it will register a function at all. Without them a whole class
+ * of consumer simply cannot expose this tool. It says nothing new: `hash` was always
+ * the only key `invoke` reads, and an extra key was always ignored. A schema that is
+ * strictly more precise about the same shape is not a change to the wire surface the
+ * tool name and behaviour guarantee covers.
  */
 export function createRetrieveTool(store) {
     return {
@@ -35,8 +43,11 @@ export function createRetrieveTool(store) {
         description: DESCRIPTION,
         inputSchema: {
             type: 'object',
-            properties: { hash: { type: 'string' } },
+            properties: {
+                hash: { type: 'string', description: 'The hash from a marker\'s retrieve("hash").' },
+            },
             required: ['hash'],
+            additionalProperties: false,
         },
         invoke: ({ hash }) => store.retrieve(hash),
     };

package/dist/retrieve.js.map

@@ -1 +1 @@
-{"version":3,"file":"retrieve.js","sourceRoot":"","sources":["../src/retrieve.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,aAAa,EAAE,MAAM,YAAY,CAAC;AAG3C,yFAAyF;AACzF,MAAM,CAAC,MAAM,kBAAkB,GAAG,gBAAgB,CAAC;AAEnD;;;;;;;GAOG;AACH,MAAM,cAAc,GAAG,aAAa,CAAC;IACnC,IAAI,EAAE,kBAAkB;IACxB,KAAK,EAAE,GAAG;IACV,IAAI,EAAE,kBAAkB;IACxB,WAAW,EAAE,+BAA+B;CAC7C,CAAC,CAAC;AAEH,MAAM,WAAW,GACf,8EAA8E;IAC9E,qDAAqD,cAAc,MAAM;IACzE,8CAA8C;IAC9C,qFAAqF;IACrF,wEAAwE,CAAC;AAE3E;;;;;;;;GAQG;AACH,MAAM,UAAU,kBAAkB,CAAC,KAAmB;IACpD,OAAO;QACL,IAAI,EAAE,kBAAkB;QACxB,WAAW,EAAE,WAAW;QACxB,WAAW,EAAE;YACX,IAAI,EAAE,QAAQ;YACd,UAAU,EAAE,EAAE,IAAI,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,EAAE;YACxC,QAAQ,EAAE,CAAC,MAAM,CAAC;SACnB;QACD,MAAM,EAAE,CAAC,EAAE,IAAI,EAAE,EAAE,EAAE,CAAC,KAAK,CAAC,QAAQ,CAAC,IAAI,CAAC;KAC3C,CAAC;AACJ,CAAC"}
+{"version":3,"file":"retrieve.js","sourceRoot":"","sources":["../src/retrieve.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,aAAa,EAAE,MAAM,YAAY,CAAC;AAG3C,yFAAyF;AACzF,MAAM,CAAC,MAAM,kBAAkB,GAAG,gBAAgB,CAAC;AAEnD;;;;;;;GAOG;AACH,MAAM,cAAc,GAAG,aAAa,CAAC;IACnC,IAAI,EAAE,kBAAkB;IACxB,KAAK,EAAE,GAAG;IACV,IAAI,EAAE,kBAAkB;IACxB,WAAW,EAAE,+BAA+B;CAC7C,CAAC,CAAC;AAEH,MAAM,WAAW,GACf,8EAA8E;IAC9E,qDAAqD,cAAc,MAAM;IACzE,8CAA8C;IAC9C,qFAAqF;IACrF,wEAAwE,CAAC;AAE3E;;;;;;;;;;;;;;;;GAgBG;AACH,MAAM,UAAU,kBAAkB,CAAC,KAAmB;IACpD,OAAO;QACL,IAAI,EAAE,kBAAkB;QACxB,WAAW,EAAE,WAAW;QACxB,WAAW,EAAE;YACX,IAAI,EAAE,QAAQ;YACd,UAAU,EAAE;gBACV,IAAI,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,WAAW,EAAE,6CAA6C,EAAE;aACrF;YACD,QAAQ,EAAE,CAAC,MAAM,CAAC;YAClB,oBAAoB,EAAE,KAAK;SAC5B;QACD,MAAM,EAAE,CAAC,EAAE,IAAI,EAAE,EAAE,EAAE,CAAC,KAAK,CAAC,QAAQ,CAAC,IAAI,CAAC;KAC3C,CAAC;AACJ,CAAC","sourcesContent":["import { defaultMarker } from './apply.ts';\nimport type { ElisionStore, RetrieveTool } from './types.ts';\n\n/** The tool name smelt's markers reference. Consumers hard-code it; do not rename it. */\nexport const RETRIEVE_TOOL_NAME = 'smelt_retrieve';\n\n/**\n * The example marker inside the tool description is *rendered by the real marker\n * builder*, never hand-written — the description is the one string a model reads to\n * recognize markers, and an example whose shape drifted from the wire format\n * (`<<smelt: …>>` when real markers say `<<smelt/v1: …>>`) would teach the model to\n * miss every marker it actually receives. `test/guards/marker-format.test.ts` pins\n * this to `MARKER_FORMAT_VERSION`.\n */\nconst EXAMPLE_MARKER = defaultMarker({\n  hash: 'a1b2c3d4e5f60718',\n  bytes: 412,\n  rule: 'sibling-collapse',\n  explanation: 'collapsed 3 sibling functions',\n});\n\nconst DESCRIPTION =\n  'Return the exact original text that was elided from a previous tool result. ' +\n  `Context you were given may contain markers like \\`${EXAMPLE_MARKER}\\`. ` +\n  'Call this with that hash to get those bytes ' +\n  'back verbatim. Nothing was deleted — it is all still here. Ask whenever the elided ' +\n  'material might matter; guessing at what a marker hid is never correct.';\n\n/**\n * Wrap a store as the tool a model calls.\n *\n * Why the description says \"ask whenever it might matter\": under-retrieval is the\n * failure mode that looks like success. A model that never calls this produces a\n * confident answer built on material it never saw, and the retrieve rate reads 0% —\n * which is indistinguishable from perfect pruning. Encouraging retrieval keeps the\n * signal in {@link ElisionStore.stats} honest.\n *\n * **The schema is strict-mode shaped on purpose.** `additionalProperties: false`, and\n * `required` naming every property — the two rules OpenAI's structured-outputs strict\n * mode enforces before it will register a function at all. Without them a whole class\n * of consumer simply cannot expose this tool. It says nothing new: `hash` was always\n * the only key `invoke` reads, and an extra key was always ignored. A schema that is\n * strictly more precise about the same shape is not a change to the wire surface the\n * tool name and behaviour guarantee covers.\n */\nexport function createRetrieveTool(store: ElisionStore): RetrieveTool {\n  return {\n    name: RETRIEVE_TOOL_NAME,\n    description: DESCRIPTION,\n    inputSchema: {\n      type: 'object',\n      properties: {\n        hash: { type: 'string', description: 'The hash from a marker\\'s retrieve(\"hash\").' },\n      },\n      required: ['hash'],\n      additionalProperties: false,\n    },\n    invoke: ({ hash }) => store.retrieve(hash),\n  };\n}\n"]}

package/dist/smelter.d.ts

@@ -0,0 +1,79 @@
+import type { MarkerBuilder } from './apply.ts';
+import type { LexicalPlannerOptions } from './plan/lexical.ts';
+import type { Strategy } from './plan/planners.ts';
+import type { StructuralPlannerOptions } from './plan/structural.ts';
+import type { DetectedLanguage, ElisionStore, Measure, Planner, RetrieveStats, RetrieveTool, SmeltResult } from './types.ts';
+/**
+ * `createSmelter()` and the three types it is spelled in.
+ *
+ * It lives here rather than in `index.ts` so that nothing under `src/` has to import
+ * the package barrel to build a smelter. `cli/subcommands/smelt.ts` does exactly that,
+ * and while `index.ts` re-exports the CLI, a barrel that imports the CLI which imports
+ * the barrel is a cycle whose only symptom is a registry that evaluates to `undefined`
+ * in whichever module the loader happens to enter first. `index.ts` re-exports every
+ * name below, so the published surface is unchanged.
+ */
+export interface SmelterConfig {
+    /** Where elided bytes live. Defaults to a fresh {@link MemoryElisionStore}. */
+    readonly store?: ElisionStore;
+    /** Used when a `smelt()` call omits `budgetBytes`. No global default is assumed. */
+    readonly defaultBudgetBytes?: number;
+    /**
+     * A constructed planner instance. Wins over `strategy`: the registry is a
+     * convenience for the shipped planners, and an instance you built yourself is
+     * always more specific than a name.
+     */
+    readonly planner?: Planner;
+    readonly strategy?: Strategy;
+    readonly marker?: MarkerBuilder;
+    /**
+     * Your own counter, so results carry a number in your unit as well as in bytes.
+     * The budget stays in bytes — see {@link Measure} and `docs/ARCHITECTURE.md` § "Decision 1".
+     */
+    readonly measure?: Measure;
+    readonly lexical?: LexicalPlannerOptions;
+    readonly structural?: StructuralPlannerOptions;
+}
+/** Options for one `smelt()` call. `budgetBytes` may come from the smelter instead. */
+export interface SmeltCallOptions {
+    readonly budgetBytes?: number;
+    readonly path?: string;
+    readonly language?: DetectedLanguage;
+    readonly focus?: readonly string[];
+}
+/**
+ * One smelter, one store, one set of counters. The store is the reason this is an
+ * object rather than a free function: elisions are only reversible for as long as
+ * something holds them, so the thing that cuts and the thing that remembers have the
+ * same lifetime by construction.
+ */
+export interface Smelter {
+    /** Shrink one blob of text. Never mutates its input. */
+    smelt(text: string, options?: SmeltCallOptions): Promise<SmeltResult>;
+    /** The exact original text of a previous result. @throws {UnknownHashError} */
+    reconstruct(result: SmeltResult): string;
+    /** One elided run, counted as a retrieval. @throws {UnknownHashError} */
+    retrieve(hash: string): string;
+    /** The tool to expose to your model. See {@link RetrieveTool}. */
+    readonly tool: RetrieveTool;
+    /** Live counters, including the expansion rate. See {@link RetrieveStats}. */
+    stats(): RetrieveStats;
+    /** The underlying store, for consumers that persist or inspect it. */
+    readonly store: ElisionStore;
+}
+/**
+ * Build a smelter.
+ *
+ * ```ts
+ * const smelter = createSmelter();
+ * const result = await smelter.smelt(toolOutput, {
+ *   path: 'src/server.ts',
+ *   budgetBytes: 4_000,
+ *   focus: ['handleRequest'],
+ * });
+ * // result.text goes to the model; smelter.tool lets it ask for the rest back.
+ * // smelter.stats().expansionRate tells you whether you cut too much.
+ * ```
+ */
+export declare function createSmelter(config?: SmelterConfig): Smelter;
+//# sourceMappingURL=smelter.d.ts.map

package/dist/smelter.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"smelter.d.ts","sourceRoot":"","sources":["../src/smelter.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAgB,aAAa,EAAE,MAAM,YAAY,CAAC;AAG9D,OAAO,KAAK,EAAE,qBAAqB,EAAE,MAAM,mBAAmB,CAAC;AAE/D,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,oBAAoB,CAAC;AACnD,OAAO,KAAK,EAAE,wBAAwB,EAAE,MAAM,sBAAsB,CAAC;AAGrE,OAAO,KAAK,EACV,gBAAgB,EAChB,YAAY,EACZ,OAAO,EAEP,OAAO,EACP,aAAa,EACb,YAAY,EACZ,WAAW,EACZ,MAAM,YAAY,CAAC;AAEpB;;;;;;;;;GASG;AAEH,MAAM,WAAW,aAAa;IAC5B,+EAA+E;IAC/E,QAAQ,CAAC,KAAK,CAAC,EAAE,YAAY,CAAC;IAC9B,oFAAoF;IACpF,QAAQ,CAAC,kBAAkB,CAAC,EAAE,MAAM,CAAC;IACrC;;;;OAIG;IACH,QAAQ,CAAC,OAAO,CAAC,EAAE,OAAO,CAAC;IAC3B,QAAQ,CAAC,QAAQ,CAAC,EAAE,QAAQ,CAAC;IAC7B,QAAQ,CAAC,MAAM,CAAC,EAAE,aAAa,CAAC;IAChC;;;OAGG;IACH,QAAQ,CAAC,OAAO,CAAC,EAAE,OAAO,CAAC;IAC3B,QAAQ,CAAC,OAAO,CAAC,EAAE,qBAAqB,CAAC;IACzC,QAAQ,CAAC,UAAU,CAAC,EAAE,wBAAwB,CAAC;CAChD;AAED,uFAAuF;AACvF,MAAM,WAAW,gBAAgB;IAC/B,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,IAAI,CAAC,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,QAAQ,CAAC,EAAE,gBAAgB,CAAC;IACrC,QAAQ,CAAC,KAAK,CAAC,EAAE,SAAS,MAAM,EAAE,CAAC;CACpC;AAED;;;;;GAKG;AACH,MAAM,WAAW,OAAO;IACtB,wDAAwD;IACxD,KAAK,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,CAAC,EAAE,gBAAgB,GAAG,OAAO,CAAC,WAAW,CAAC,CAAC;IACtE,+EAA+E;IAC/E,WAAW,CAAC,MAAM,EAAE,WAAW,GAAG,MAAM,CAAC;IACzC,yEAAyE;IACzE,QAAQ,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAAC;IAC/B,kEAAkE;IAClE,QAAQ,CAAC,IAAI,EAAE,YAAY,CAAC;IAC5B,8EAA8E;IAC9E,KAAK,IAAI,aAAa,CAAC;IACvB,sEAAsE;IACtE,QAAQ,CAAC,KAAK,EAAE,YAAY,CAAC;CAC9B;AAED;;;;;;;;;;;;;GAaG;AACH,wBAAgB,aAAa,CAAC,MAAM,GAAE,aAAkB,GAAG,OAAO,CA8CjE","sourcesContent":["import { applyPlan, markerForLanguage, markerPricing, reconstruct } from './apply.ts';\nimport type { ApplyOptions, MarkerBuilder } from './apply.ts';\nimport { detectLanguage } from './detect.ts';\nimport { SmeltError } from './errors.ts';\nimport type { LexicalPlannerOptions } from './plan/lexical.ts';\nimport { DEFAULT_STRATEGY, PLANNERS } from './plan/planners.ts';\nimport type { Strategy } from './plan/planners.ts';\nimport type { StructuralPlannerOptions } from './plan/structural.ts';\nimport { createRetrieveTool } from './retrieve.ts';\nimport { MemoryElisionStore } from './store.ts';\nimport type {\n  DetectedLanguage,\n  ElisionStore,\n  Measure,\n  PlanInput,\n  Planner,\n  RetrieveStats,\n  RetrieveTool,\n  SmeltResult,\n} from './types.ts';\n\n/**\n * `createSmelter()` and the three types it is spelled in.\n *\n * It lives here rather than in `index.ts` so that nothing under `src/` has to import\n * the package barrel to build a smelter. `cli/subcommands/smelt.ts` does exactly that,\n * and while `index.ts` re-exports the CLI, a barrel that imports the CLI which imports\n * the barrel is a cycle whose only symptom is a registry that evaluates to `undefined`\n * in whichever module the loader happens to enter first. `index.ts` re-exports every\n * name below, so the published surface is unchanged.\n */\n\nexport interface SmelterConfig {\n  /** Where elided bytes live. Defaults to a fresh {@link MemoryElisionStore}. */\n  readonly store?: ElisionStore;\n  /** Used when a `smelt()` call omits `budgetBytes`. No global default is assumed. */\n  readonly defaultBudgetBytes?: number;\n  /**\n   * A constructed planner instance. Wins over `strategy`: the registry is a\n   * convenience for the shipped planners, and an instance you built yourself is\n   * always more specific than a name.\n   */\n  readonly planner?: Planner;\n  readonly strategy?: Strategy;\n  readonly marker?: MarkerBuilder;\n  /**\n   * Your own counter, so results carry a number in your unit as well as in bytes.\n   * The budget stays in bytes — see {@link Measure} and `docs/ARCHITECTURE.md` § \"Decision 1\".\n   */\n  readonly measure?: Measure;\n  readonly lexical?: LexicalPlannerOptions;\n  readonly structural?: StructuralPlannerOptions;\n}\n\n/** Options for one `smelt()` call. `budgetBytes` may come from the smelter instead. */\nexport interface SmeltCallOptions {\n  readonly budgetBytes?: number;\n  readonly path?: string;\n  readonly language?: DetectedLanguage;\n  readonly focus?: readonly string[];\n}\n\n/**\n * One smelter, one store, one set of counters. The store is the reason this is an\n * object rather than a free function: elisions are only reversible for as long as\n * something holds them, so the thing that cuts and the thing that remembers have the\n * same lifetime by construction.\n */\nexport interface Smelter {\n  /** Shrink one blob of text. Never mutates its input. */\n  smelt(text: string, options?: SmeltCallOptions): Promise<SmeltResult>;\n  /** The exact original text of a previous result. @throws {UnknownHashError} */\n  reconstruct(result: SmeltResult): string;\n  /** One elided run, counted as a retrieval. @throws {UnknownHashError} */\n  retrieve(hash: string): string;\n  /** The tool to expose to your model. See {@link RetrieveTool}. */\n  readonly tool: RetrieveTool;\n  /** Live counters, including the expansion rate. See {@link RetrieveStats}. */\n  stats(): RetrieveStats;\n  /** The underlying store, for consumers that persist or inspect it. */\n  readonly store: ElisionStore;\n}\n\n/**\n * Build a smelter.\n *\n * ```ts\n * const smelter = createSmelter();\n * const result = await smelter.smelt(toolOutput, {\n *   path: 'src/server.ts',\n *   budgetBytes: 4_000,\n *   focus: ['handleRequest'],\n * });\n * // result.text goes to the model; smelter.tool lets it ask for the rest back.\n * // smelter.stats().expansionRate tells you whether you cut too much.\n * ```\n */\nexport function createSmelter(config: SmelterConfig = {}): Smelter {\n  const store = config.store ?? new MemoryElisionStore();\n  // A constructed instance wins over a strategy name; the registry serves the names.\n  const planner: Planner = config.planner ?? PLANNERS[config.strategy ?? DEFAULT_STRATEGY](config);\n  const applyOptions: ApplyOptions = {\n    ...(config.marker === undefined ? {} : { marker: config.marker }),\n    ...(config.measure === undefined ? {} : { measure: config.measure }),\n  };\n\n  return {\n    store,\n    tool: createRetrieveTool(store),\n    stats: () => store.stats(),\n    retrieve: (hash) => store.retrieve(hash),\n    reconstruct: (result) => reconstruct(result, store),\n    async smelt(text, options = {}) {\n      const budgetBytes = options.budgetBytes ?? config.defaultBudgetBytes;\n      if (budgetBytes === undefined) {\n        throw new SmeltError(\n          'smelt: no budget. Pass `budgetBytes` to smelt() or `defaultBudgetBytes` to ' +\n            'createSmelter(). There is no built-in default, because a budget smelt ' +\n            'invented would silently decide how much of your context to throw away.',\n        );\n      }\n      const language = options.language ?? detectLanguage(options.path);\n      const input: PlanInput = {\n        text,\n        language,\n        budgetBytes,\n        // The MarkerPricing seam, constructed centrally — here, and nowhere else in\n        // the shipped pipeline — from the exact builder the applyPlan call below will\n        // use: a caller-supplied `config.marker` prices with its own rendering (a\n        // longer custom marker makes small cuts unprofitable, and the planner must\n        // see that), otherwise the language's leader-wrapped default.\n        pricing: markerPricing(language, config.marker),\n        ...(options.focus === undefined ? {} : { focus: options.focus }),\n      };\n      const plan = await planner.plan(input);\n      // The marker follows the *result's* language: it lands behind the language's\n      // line-comment leader (see MARKER_LINE_COMMENT_LEADERS), because a bare marker\n      // line breaks the survivor's syntax in every grammar tested. A caller-supplied\n      // marker builder always wins.\n      const marker = config.marker ?? markerForLanguage(plan.language);\n      return applyPlan(text, plan, store, { ...applyOptions, marker });\n    },\n  };\n}\n"]}

package/dist/smelter.js

@@ -0,0 +1,65 @@
+import { applyPlan, markerForLanguage, markerPricing, reconstruct } from './apply.js';
+import { detectLanguage } from './detect.js';
+import { SmeltError } from './errors.js';
+import { DEFAULT_STRATEGY, PLANNERS } from './plan/planners.js';
+import { createRetrieveTool } from './retrieve.js';
+import { MemoryElisionStore } from './store.js';
+/**
+ * Build a smelter.
+ *
+ * ```ts
+ * const smelter = createSmelter();
+ * const result = await smelter.smelt(toolOutput, {
+ *   path: 'src/server.ts',
+ *   budgetBytes: 4_000,
+ *   focus: ['handleRequest'],
+ * });
+ * // result.text goes to the model; smelter.tool lets it ask for the rest back.
+ * // smelter.stats().expansionRate tells you whether you cut too much.
+ * ```
+ */
+export function createSmelter(config = {}) {
+    const store = config.store ?? new MemoryElisionStore();
+    // A constructed instance wins over a strategy name; the registry serves the names.
+    const planner = config.planner ?? PLANNERS[config.strategy ?? DEFAULT_STRATEGY](config);
+    const applyOptions = {
+        ...(config.marker === undefined ? {} : { marker: config.marker }),
+        ...(config.measure === undefined ? {} : { measure: config.measure }),
+    };
+    return {
+        store,
+        tool: createRetrieveTool(store),
+        stats: () => store.stats(),
+        retrieve: (hash) => store.retrieve(hash),
+        reconstruct: (result) => reconstruct(result, store),
+        async smelt(text, options = {}) {
+            const budgetBytes = options.budgetBytes ?? config.defaultBudgetBytes;
+            if (budgetBytes === undefined) {
+                throw new SmeltError('smelt: no budget. Pass `budgetBytes` to smelt() or `defaultBudgetBytes` to ' +
+                    'createSmelter(). There is no built-in default, because a budget smelt ' +
+                    'invented would silently decide how much of your context to throw away.');
+            }
+            const language = options.language ?? detectLanguage(options.path);
+            const input = {
+                text,
+                language,
+                budgetBytes,
+                // The MarkerPricing seam, constructed centrally — here, and nowhere else in
+                // the shipped pipeline — from the exact builder the applyPlan call below will
+                // use: a caller-supplied `config.marker` prices with its own rendering (a
+                // longer custom marker makes small cuts unprofitable, and the planner must
+                // see that), otherwise the language's leader-wrapped default.
+                pricing: markerPricing(language, config.marker),
+                ...(options.focus === undefined ? {} : { focus: options.focus }),
+            };
+            const plan = await planner.plan(input);
+            // The marker follows the *result's* language: it lands behind the language's
+            // line-comment leader (see MARKER_LINE_COMMENT_LEADERS), because a bare marker
+            // line breaks the survivor's syntax in every grammar tested. A caller-supplied
+            // marker builder always wins.
+            const marker = config.marker ?? markerForLanguage(plan.language);
+            return applyPlan(text, plan, store, { ...applyOptions, marker });
+        },
+    };
+}
+//# sourceMappingURL=smelter.js.map

package/dist/smelter.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"smelter.js","sourceRoot":"","sources":["../src/smelter.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,SAAS,EAAE,iBAAiB,EAAE,aAAa,EAAE,WAAW,EAAE,MAAM,YAAY,CAAC;AAEtF,OAAO,EAAE,cAAc,EAAE,MAAM,aAAa,CAAC;AAC7C,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAEzC,OAAO,EAAE,gBAAgB,EAAE,QAAQ,EAAE,MAAM,oBAAoB,CAAC;AAGhE,OAAO,EAAE,kBAAkB,EAAE,MAAM,eAAe,CAAC;AACnD,OAAO,EAAE,kBAAkB,EAAE,MAAM,YAAY,CAAC;AA0EhD;;;;;;;;;;;;;GAaG;AACH,MAAM,UAAU,aAAa,CAAC,MAAM,GAAkB,EAAE;IACtD,MAAM,KAAK,GAAG,MAAM,CAAC,KAAK,IAAI,IAAI,kBAAkB,EAAE,CAAC;IACvD,mFAAmF;IACnF,MAAM,OAAO,GAAY,MAAM,CAAC,OAAO,IAAI,QAAQ,CAAC,MAAM,CAAC,QAAQ,IAAI,gBAAgB,CAAC,CAAC,MAAM,CAAC,CAAC;IACjG,MAAM,YAAY,GAAiB;QACjC,GAAG,CAAC,MAAM,CAAC,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,CAAC;QACjE,GAAG,CAAC,MAAM,CAAC,OAAO,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,MAAM,CAAC,OAAO,EAAE,CAAC;KACrE,CAAC;IAEF,OAAO;QACL,KAAK;QACL,IAAI,EAAE,kBAAkB,CAAC,KAAK,CAAC;QAC/B,KAAK,EAAE,GAAG,EAAE,CAAC,KAAK,CAAC,KAAK,EAAE;QAC1B,QAAQ,EAAE,CAAC,IAAI,EAAE,EAAE,CAAC,KAAK,CAAC,QAAQ,CAAC,IAAI,CAAC;QACxC,WAAW,EAAE,CAAC,MAAM,EAAE,EAAE,CAAC,WAAW,CAAC,MAAM,EAAE,KAAK,CAAC;QACnD,KAAK,CAAC,KAAK,CAAC,IAAI,EAAE,OAAO,GAAG,EAAE;YAC5B,MAAM,WAAW,GAAG,OAAO,CAAC,WAAW,IAAI,MAAM,CAAC,kBAAkB,CAAC;YACrE,IAAI,WAAW,KAAK,SAAS,EAAE,CAAC;gBAC9B,MAAM,IAAI,UAAU,CAClB,6EAA6E;oBAC3E,wEAAwE;oBACxE,wEAAwE,CAC3E,CAAC;YACJ,CAAC;YACD,MAAM,QAAQ,GAAG,OAAO,CAAC,QAAQ,IAAI,cAAc,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;YAClE,MAAM,KAAK,GAAc;gBACvB,IAAI;gBACJ,QAAQ;gBACR,WAAW;gBACX,4EAA4E;gBAC5E,8EAA8E;gBAC9E,0EAA0E;gBAC1E,2EAA2E;gBAC3E,8DAA8D;gBAC9D,OAAO,EAAE,aAAa,CAAC,QAAQ,EAAE,MAAM,CAAC,MAAM,CAAC;gBAC/C,GAAG,CAAC,OAAO,CAAC,KAAK,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,OAAO,CAAC,KAAK,EAAE,CAAC;aACjE,CAAC;YACF,MAAM,IAAI,GAAG,MAAM,OAAO,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;YACvC,6EAA6E;YAC7E,+EAA+E;YAC/E,+EAA+E;YAC/E,8BAA8B;YAC9B,MAAM,MAAM,GAAG,MAAM,CAAC,MAAM,IAAI,iBAAiB,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;YACjE,OAAO,SAAS,CAAC,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,EAAE,GAAG,YAAY,EAAE,MAAM,EAAE,CAAC,CAAC;QACnE,CAAC;KACF,CAAC;AACJ,CAAC","sourcesContent":["import { applyPlan, markerForLanguage, markerPricing, reconstruct } from './apply.ts';\nimport type { ApplyOptions, MarkerBuilder } from './apply.ts';\nimport { detectLanguage } from './detect.ts';\nimport { SmeltError } from './errors.ts';\nimport type { LexicalPlannerOptions } from './plan/lexical.ts';\nimport { DEFAULT_STRATEGY, PLANNERS } from './plan/planners.ts';\nimport type { Strategy } from './plan/planners.ts';\nimport type { StructuralPlannerOptions } from './plan/structural.ts';\nimport { createRetrieveTool } from './retrieve.ts';\nimport { MemoryElisionStore } from './store.ts';\nimport type {\n  DetectedLanguage,\n  ElisionStore,\n  Measure,\n  PlanInput,\n  Planner,\n  RetrieveStats,\n  RetrieveTool,\n  SmeltResult,\n} from './types.ts';\n\n/**\n * `createSmelter()` and the three types it is spelled in.\n *\n * It lives here rather than in `index.ts` so that nothing under `src/` has to import\n * the package barrel to build a smelter. `cli/subcommands/smelt.ts` does exactly that,\n * and while `index.ts` re-exports the CLI, a barrel that imports the CLI which imports\n * the barrel is a cycle whose only symptom is a registry that evaluates to `undefined`\n * in whichever module the loader happens to enter first. `index.ts` re-exports every\n * name below, so the published surface is unchanged.\n */\n\nexport interface SmelterConfig {\n  /** Where elided bytes live. Defaults to a fresh {@link MemoryElisionStore}. */\n  readonly store?: ElisionStore;\n  /** Used when a `smelt()` call omits `budgetBytes`. No global default is assumed. */\n  readonly defaultBudgetBytes?: number;\n  /**\n   * A constructed planner instance. Wins over `strategy`: the registry is a\n   * convenience for the shipped planners, and an instance you built yourself is\n   * always more specific than a name.\n   */\n  readonly planner?: Planner;\n  readonly strategy?: Strategy;\n  readonly marker?: MarkerBuilder;\n  /**\n   * Your own counter, so results carry a number in your unit as well as in bytes.\n   * The budget stays in bytes — see {@link Measure} and `docs/ARCHITECTURE.md` § \"Decision 1\".\n   */\n  readonly measure?: Measure;\n  readonly lexical?: LexicalPlannerOptions;\n  readonly structural?: StructuralPlannerOptions;\n}\n\n/** Options for one `smelt()` call. `budgetBytes` may come from the smelter instead. */\nexport interface SmeltCallOptions {\n  readonly budgetBytes?: number;\n  readonly path?: string;\n  readonly language?: DetectedLanguage;\n  readonly focus?: readonly string[];\n}\n\n/**\n * One smelter, one store, one set of counters. The store is the reason this is an\n * object rather than a free function: elisions are only reversible for as long as\n * something holds them, so the thing that cuts and the thing that remembers have the\n * same lifetime by construction.\n */\nexport interface Smelter {\n  /** Shrink one blob of text. Never mutates its input. */\n  smelt(text: string, options?: SmeltCallOptions): Promise<SmeltResult>;\n  /** The exact original text of a previous result. @throws {UnknownHashError} */\n  reconstruct(result: SmeltResult): string;\n  /** One elided run, counted as a retrieval. @throws {UnknownHashError} */\n  retrieve(hash: string): string;\n  /** The tool to expose to your model. See {@link RetrieveTool}. */\n  readonly tool: RetrieveTool;\n  /** Live counters, including the expansion rate. See {@link RetrieveStats}. */\n  stats(): RetrieveStats;\n  /** The underlying store, for consumers that persist or inspect it. */\n  readonly store: ElisionStore;\n}\n\n/**\n * Build a smelter.\n *\n * ```ts\n * const smelter = createSmelter();\n * const result = await smelter.smelt(toolOutput, {\n *   path: 'src/server.ts',\n *   budgetBytes: 4_000,\n *   focus: ['handleRequest'],\n * });\n * // result.text goes to the model; smelter.tool lets it ask for the rest back.\n * // smelter.stats().expansionRate tells you whether you cut too much.\n * ```\n */\nexport function createSmelter(config: SmelterConfig = {}): Smelter {\n  const store = config.store ?? new MemoryElisionStore();\n  // A constructed instance wins over a strategy name; the registry serves the names.\n  const planner: Planner = config.planner ?? PLANNERS[config.strategy ?? DEFAULT_STRATEGY](config);\n  const applyOptions: ApplyOptions = {\n    ...(config.marker === undefined ? {} : { marker: config.marker }),\n    ...(config.measure === undefined ? {} : { measure: config.measure }),\n  };\n\n  return {\n    store,\n    tool: createRetrieveTool(store),\n    stats: () => store.stats(),\n    retrieve: (hash) => store.retrieve(hash),\n    reconstruct: (result) => reconstruct(result, store),\n    async smelt(text, options = {}) {\n      const budgetBytes = options.budgetBytes ?? config.defaultBudgetBytes;\n      if (budgetBytes === undefined) {\n        throw new SmeltError(\n          'smelt: no budget. Pass `budgetBytes` to smelt() or `defaultBudgetBytes` to ' +\n            'createSmelter(). There is no built-in default, because a budget smelt ' +\n            'invented would silently decide how much of your context to throw away.',\n        );\n      }\n      const language = options.language ?? detectLanguage(options.path);\n      const input: PlanInput = {\n        text,\n        language,\n        budgetBytes,\n        // The MarkerPricing seam, constructed centrally — here, and nowhere else in\n        // the shipped pipeline — from the exact builder the applyPlan call below will\n        // use: a caller-supplied `config.marker` prices with its own rendering (a\n        // longer custom marker makes small cuts unprofitable, and the planner must\n        // see that), otherwise the language's leader-wrapped default.\n        pricing: markerPricing(language, config.marker),\n        ...(options.focus === undefined ? {} : { focus: options.focus }),\n      };\n      const plan = await planner.plan(input);\n      // The marker follows the *result's* language: it lands behind the language's\n      // line-comment leader (see MARKER_LINE_COMMENT_LEADERS), because a bare marker\n      // line breaks the survivor's syntax in every grammar tested. A caller-supplied\n      // marker builder always wins.\n      const marker = config.marker ?? markerForLanguage(plan.language);\n      return applyPlan(text, plan, store, { ...applyOptions, marker });\n    },\n  };\n}\n"]}

package/dist/stages.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"stages.d.ts","sourceRoot":"","sources":["../src/stages.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,YAAY,EAAmB,WAAW,EAAqB,MAAM,YAAY,CAAC;AAEhG;;;;;;;;GAQG;AACH,eAAO,MAAM,uBAAuB,EAAE,WAYrC,CAAC;AAEF;;;;;;;;GAQG;AACH,eAAO,MAAM,wBAAwB,EAAE,YAQtC,CAAC"}
+{"version":3,"file":"stages.d.ts","sourceRoot":"","sources":["../src/stages.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,YAAY,EAAmB,WAAW,EAAqB,MAAM,YAAY,CAAC;AAEhG;;;;;;;;GAQG;AACH,eAAO,MAAM,uBAAuB,EAAE,WAYrC,CAAC;AAEF;;;;;;;;GAQG;AACH,eAAO,MAAM,wBAAwB,EAAE,YAQtC,CAAC","sourcesContent":["import { NotImplementedError } from './errors.ts';\nimport type { DistillStage, RerankCandidate, RerankStage, RerankedCandidate } from './types.ts';\n\n/**\n * The reranker that ships with smelt: one that refuses.\n *\n * It exists so that \"reranking is a seam, not a feature\" is enforced rather than\n * promised. Wire this in and you get an exception naming the interface you were\n * supposed to implement. There is no default hosted reranker, no bundled key handling,\n * and no `SMELT_RERANK_API_KEY` — the first of those to appear breaks Law 1 for every\n * consumer at once, including the ones who never read the changelog.\n */\nexport const unconfiguredRerankStage: RerankStage = {\n  id: 'rerank/unconfigured',\n  rerank(\n    _candidates: readonly RerankCandidate[],\n    _query: string,\n  ): Promise<readonly RerankedCandidate[]> {\n    throw new NotImplementedError(\n      'reranking',\n      'docs/ARCHITECTURE.md § \"Explicitly out of scope\" — implement `RerankStage` in your own ' +\n        'code, with your own key, so the network call is visible in your source',\n    );\n  },\n};\n\n/**\n * Learned distillation, same treatment, for a different reason.\n *\n * Distillation is out of v1 because a model-written summary cannot satisfy Law 2. \"The\n * model condensed this\" does not say what was removed, and once the text is rewritten\n * there is nothing left to store under a hash. If this ever ships it will store the\n * original, explain itself in the same rule-named terms every other elision uses, and\n * be reversible — or it will not ship.\n */\nexport const unconfiguredDistillStage: DistillStage = {\n  id: 'distill/unconfigured',\n  distill(_text: string, _budgetBytes: number): Promise<string> {\n    throw new NotImplementedError(\n      'learned distillation',\n      'docs/ARCHITECTURE.md § \"Explicitly out of scope\"',\n    );\n  },\n};\n"]}

package/dist/stages.js

@@ -11,7 +11,7 @@ import { NotImplementedError } from './errors.js';
 export const unconfiguredRerankStage = {
     id: 'rerank/unconfigured',
     rerank(_candidates, _query) {
-        throw new NotImplementedError('reranking', 'docs/HANDOFF.md § "Explicitly out of v1" — implement `RerankStage` in your own ' +
+        throw new NotImplementedError('reranking', 'docs/ARCHITECTURE.md § "Explicitly out of scope" — implement `RerankStage` in your own ' +
             'code, with your own key, so the network call is visible in your source');
     },
 };
@@ -27,7 +27,7 @@ export const unconfiguredRerankStage = {
 export const unconfiguredDistillStage = {
     id: 'distill/unconfigured',
     distill(_text, _budgetBytes) {
-        throw new NotImplementedError('learned distillation', 'docs/HANDOFF.md § "Explicitly out of v1"');
+        throw new NotImplementedError('learned distillation', 'docs/ARCHITECTURE.md § "Explicitly out of scope"');
     },
 };
 //# sourceMappingURL=stages.js.map

package/dist/stages.js.map

@@ -1 +1 @@
-{"version":3,"file":"stages.js","sourceRoot":"","sources":["../src/stages.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,mBAAmB,EAAE,MAAM,aAAa,CAAC;AAGlD;;;;;;;;GAQG;AACH,MAAM,CAAC,MAAM,uBAAuB,GAAgB;IAClD,EAAE,EAAE,qBAAqB;IACzB,MAAM,CACJ,WAAuC,EACvC,MAAc;QAEd,MAAM,IAAI,mBAAmB,CAC3B,WAAW,EACX,iFAAiF;YAC/E,wEAAwE,CAC3E,CAAC;IACJ,CAAC;CACF,CAAC;AAEF;;;;;;;;GAQG;AACH,MAAM,CAAC,MAAM,wBAAwB,GAAiB;IACpD,EAAE,EAAE,sBAAsB;IAC1B,OAAO,CAAC,KAAa,EAAE,YAAoB;QACzC,MAAM,IAAI,mBAAmB,CAC3B,sBAAsB,EACtB,0CAA0C,CAC3C,CAAC;IACJ,CAAC;CACF,CAAC"}
+{"version":3,"file":"stages.js","sourceRoot":"","sources":["../src/stages.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,mBAAmB,EAAE,MAAM,aAAa,CAAC;AAGlD;;;;;;;;GAQG;AACH,MAAM,CAAC,MAAM,uBAAuB,GAAgB;IAClD,EAAE,EAAE,qBAAqB;IACzB,MAAM,CACJ,WAAuC,EACvC,MAAc;QAEd,MAAM,IAAI,mBAAmB,CAC3B,WAAW,EACX,yFAAyF;YACvF,wEAAwE,CAC3E,CAAC;IACJ,CAAC;CACF,CAAC;AAEF;;;;;;;;GAQG;AACH,MAAM,CAAC,MAAM,wBAAwB,GAAiB;IACpD,EAAE,EAAE,sBAAsB;IAC1B,OAAO,CAAC,KAAa,EAAE,YAAoB;QACzC,MAAM,IAAI,mBAAmB,CAC3B,sBAAsB,EACtB,kDAAkD,CACnD,CAAC;IACJ,CAAC;CACF,CAAC","sourcesContent":["import { NotImplementedError } from './errors.ts';\nimport type { DistillStage, RerankCandidate, RerankStage, RerankedCandidate } from './types.ts';\n\n/**\n * The reranker that ships with smelt: one that refuses.\n *\n * It exists so that \"reranking is a seam, not a feature\" is enforced rather than\n * promised. Wire this in and you get an exception naming the interface you were\n * supposed to implement. There is no default hosted reranker, no bundled key handling,\n * and no `SMELT_RERANK_API_KEY` — the first of those to appear breaks Law 1 for every\n * consumer at once, including the ones who never read the changelog.\n */\nexport const unconfiguredRerankStage: RerankStage = {\n  id: 'rerank/unconfigured',\n  rerank(\n    _candidates: readonly RerankCandidate[],\n    _query: string,\n  ): Promise<readonly RerankedCandidate[]> {\n    throw new NotImplementedError(\n      'reranking',\n      'docs/ARCHITECTURE.md § \"Explicitly out of scope\" — implement `RerankStage` in your own ' +\n        'code, with your own key, so the network call is visible in your source',\n    );\n  },\n};\n\n/**\n * Learned distillation, same treatment, for a different reason.\n *\n * Distillation is out of v1 because a model-written summary cannot satisfy Law 2. \"The\n * model condensed this\" does not say what was removed, and once the text is rewritten\n * there is nothing left to store under a hash. If this ever ships it will store the\n * original, explain itself in the same rule-named terms every other elision uses, and\n * be reversible — or it will not ship.\n */\nexport const unconfiguredDistillStage: DistillStage = {\n  id: 'distill/unconfigured',\n  distill(_text: string, _budgetBytes: number): Promise<string> {\n    throw new NotImplementedError(\n      'learned distillation',\n      'docs/ARCHITECTURE.md § \"Explicitly out of scope\"',\n    );\n  },\n};\n"]}

package/dist/stats.d.ts

@@ -11,7 +11,8 @@ export type RawRetrieveCounters = Omit<RetrieveStats, 'expansionRate' | 'allElis
  * The one derivation of the honesty arithmetic, shared by every store.
  *
  * `expansionRate` is the number this project exists to keep honest, and
- * `allElisionsRetrieved` is the one degenerate outcome it names (HANDOFF Decision 4).
+ * `allElisionsRetrieved` is the one degenerate outcome it names (`docs/ARCHITECTURE.md`
+ * § Decision 4).
  * When each store derived them privately, the two copies could drift — and a store
  * whose arithmetic drifted flattering-ward would be the exact silent failure Law 3's
  * counters exist to refuse. So the seam between a store and its stats is narrowed to

package/dist/stats.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"stats.d.ts","sourceRoot":"","sources":["../src/stats.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,YAAY,CAAC;AAEhD;;;;;;GAMG;AACH,MAAM,MAAM,mBAAmB,GAAG,IAAI,CAAC,aAAa,EAAE,eAAe,GAAG,sBAAsB,CAAC,CAAC;AAEhG;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4BG;AACH,wBAAgB,aAAa,CAAC,GAAG,EAAE,mBAAmB,GAAG,aAAa,CAMrE"}
+{"version":3,"file":"stats.d.ts","sourceRoot":"","sources":["../src/stats.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,YAAY,CAAC;AAEhD;;;;;;GAMG;AACH,MAAM,MAAM,mBAAmB,GAAG,IAAI,CAAC,aAAa,EAAE,eAAe,GAAG,sBAAsB,CAAC,CAAC;AAEhG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6BG;AACH,wBAAgB,aAAa,CAAC,GAAG,EAAE,mBAAmB,GAAG,aAAa,CAMrE","sourcesContent":["import type { RetrieveStats } from './types.ts';\n\n/**\n * The directly-observed half of {@link RetrieveStats}: the five counts a store reads\n * off its own records — a map size, a directory scan, a journal fold. Nothing in here\n * is derived; every field is a fact the store witnessed. The derived half —\n * `expansionRate` and `allElisionsRetrieved`, the honesty arithmetic of Law 3 — is\n * computed from these by {@link retrieveStats}, in exactly one place.\n */\nexport type RawRetrieveCounters = Omit<RetrieveStats, 'expansionRate' | 'allElisionsRetrieved'>;\n\n/**\n * The one derivation of the honesty arithmetic, shared by every store.\n *\n * `expansionRate` is the number this project exists to keep honest, and\n * `allElisionsRetrieved` is the one degenerate outcome it names (`docs/ARCHITECTURE.md`\n * § Decision 4).\n * When each store derived them privately, the two copies could drift — and a store\n * whose arithmetic drifted flattering-ward would be the exact silent failure Law 3's\n * counters exist to refuse. So the seam between a store and its stats is narrowed to\n * raw counters: an {@link ElisionStore} adapter *supplies counts*\n * ({@link RawRetrieveCounters}) and never derives the metric.\n *\n * **Contract for adapter authors.** A custom store implements a\n * `rawCounters(): RawRetrieveCounters` method (or any equivalent that gathers the five\n * counts from its own records) and delegates its public `stats()` to this function:\n *\n * ```ts\n * stats(): RetrieveStats {\n *   return retrieveStats(this.rawCounters());\n * }\n * ```\n *\n * This is a free wrapper function rather than an abstract base class, deliberately —\n * it is the least-breaking shape. The stores share no storage machinery (one is a\n * `Map`, one is a directory), so a base class would couple every adapter's inheritance\n * chain to smelt's for the sake of two lines of arithmetic; an existing `ElisionStore`\n * implementation keeps its own hierarchy and adopts this contract by changing only its\n * `stats()` body. The {@link ElisionStore} interface itself is unchanged: consumers\n * still call `stats()` and never see this seam.\n */\nexport function retrieveStats(raw: RawRetrieveCounters): RetrieveStats {\n  return {\n    ...raw,\n    expansionRate: raw.elisionsStored === 0 ? 0 : raw.uniqueRetrieved / raw.elisionsStored,\n    allElisionsRetrieved: raw.elisionsStored > 0 && raw.uniqueRetrieved === raw.elisionsStored,\n  };\n}\n"]}

package/dist/stats.js

@@ -2,7 +2,8 @@
  * The one derivation of the honesty arithmetic, shared by every store.
  *
  * `expansionRate` is the number this project exists to keep honest, and
- * `allElisionsRetrieved` is the one degenerate outcome it names (HANDOFF Decision 4).
+ * `allElisionsRetrieved` is the one degenerate outcome it names (`docs/ARCHITECTURE.md`
+ * § Decision 4).
  * When each store derived them privately, the two copies could drift — and a store
  * whose arithmetic drifted flattering-ward would be the exact silent failure Law 3's
  * counters exist to refuse. So the seam between a store and its stats is narrowed to

package/dist/stats.js.map

@@ -1 +1 @@
-{"version":3,"file":"stats.js","sourceRoot":"","sources":["../src/stats.ts"],"names":[],"mappings":"AAWA;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4BG;AACH,MAAM,UAAU,aAAa,CAAC,GAAwB;IACpD,OAAO;QACL,GAAG,GAAG;QACN,aAAa,EAAE,GAAG,CAAC,cAAc,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,eAAe,GAAG,GAAG,CAAC,cAAc;QACtF,oBAAoB,EAAE,GAAG,CAAC,cAAc,GAAG,CAAC,IAAI,GAAG,CAAC,eAAe,KAAK,GAAG,CAAC,cAAc;KAC3F,CAAC;AACJ,CAAC"}
+{"version":3,"file":"stats.js","sourceRoot":"","sources":["../src/stats.ts"],"names":[],"mappings":"AAWA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6BG;AACH,MAAM,UAAU,aAAa,CAAC,GAAwB;IACpD,OAAO;QACL,GAAG,GAAG;QACN,aAAa,EAAE,GAAG,CAAC,cAAc,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,eAAe,GAAG,GAAG,CAAC,cAAc;QACtF,oBAAoB,EAAE,GAAG,CAAC,cAAc,GAAG,CAAC,IAAI,GAAG,CAAC,eAAe,KAAK,GAAG,CAAC,cAAc;KAC3F,CAAC;AACJ,CAAC","sourcesContent":["import type { RetrieveStats } from './types.ts';\n\n/**\n * The directly-observed half of {@link RetrieveStats}: the five counts a store reads\n * off its own records — a map size, a directory scan, a journal fold. Nothing in here\n * is derived; every field is a fact the store witnessed. The derived half —\n * `expansionRate` and `allElisionsRetrieved`, the honesty arithmetic of Law 3 — is\n * computed from these by {@link retrieveStats}, in exactly one place.\n */\nexport type RawRetrieveCounters = Omit<RetrieveStats, 'expansionRate' | 'allElisionsRetrieved'>;\n\n/**\n * The one derivation of the honesty arithmetic, shared by every store.\n *\n * `expansionRate` is the number this project exists to keep honest, and\n * `allElisionsRetrieved` is the one degenerate outcome it names (`docs/ARCHITECTURE.md`\n * § Decision 4).\n * When each store derived them privately, the two copies could drift — and a store\n * whose arithmetic drifted flattering-ward would be the exact silent failure Law 3's\n * counters exist to refuse. So the seam between a store and its stats is narrowed to\n * raw counters: an {@link ElisionStore} adapter *supplies counts*\n * ({@link RawRetrieveCounters}) and never derives the metric.\n *\n * **Contract for adapter authors.** A custom store implements a\n * `rawCounters(): RawRetrieveCounters` method (or any equivalent that gathers the five\n * counts from its own records) and delegates its public `stats()` to this function:\n *\n * ```ts\n * stats(): RetrieveStats {\n *   return retrieveStats(this.rawCounters());\n * }\n * ```\n *\n * This is a free wrapper function rather than an abstract base class, deliberately —\n * it is the least-breaking shape. The stores share no storage machinery (one is a\n * `Map`, one is a directory), so a base class would couple every adapter's inheritance\n * chain to smelt's for the sake of two lines of arithmetic; an existing `ElisionStore`\n * implementation keeps its own hierarchy and adopts this contract by changing only its\n * `stats()` body. The {@link ElisionStore} interface itself is unchanged: consumers\n * still call `stats()` and never see this seam.\n */\nexport function retrieveStats(raw: RawRetrieveCounters): RetrieveStats {\n  return {\n    ...raw,\n    expansionRate: raw.elisionsStored === 0 ? 0 : raw.uniqueRetrieved / raw.elisionsStored,\n    allElisionsRetrieved: raw.elisionsStored > 0 && raw.uniqueRetrieved === raw.elisionsStored,\n  };\n}\n"]}

package/dist/store-dir.d.ts

@@ -40,10 +40,30 @@ export interface DirectoryElisionStoreOptions {
  * - **Writes are crash-safe.** A blob is written to `tmp/`, `fsync`ed, then `link(2)`ed
  *   into `blobs/` — an atomic, no-clobber publish. A torn write dies in `tmp/`, where
  *   nothing looks; a name in `blobs/` always refers to a fully written file.
- * - **Reads verify.** `retrieve()` and `peek()` re-hash the bytes and refuse a mismatch
- *   with {@link StoreCorruptionError} — a damaged blob is never handed back as a
- *   retrieval, and "we hold damaged bytes" is distinct from {@link UnknownHashError}'s
- *   "never existed". The guard in `test/guards/persistent-store.test.ts` watches this.
+ * - **`fsync` is only as strong as the platform makes it.** Every flush here is Node's
+ *   `fsyncSync`, which is libuv's `uv_fs_fsync`, and what that reaches the hardware with
+ *   differs by platform. On Apple it is strong: libuv knows macOS's own `fsync(2)` only
+ *   hands the write to the drive, so its `__APPLE__` branch issues
+ *   `fcntl(fd, F_FULLFSYNC)` — a real drive-cache flush — before falling back to
+ *   `F_BARRIERFSYNC` and then plain `fsync(2)`. Measured here (Node 26, libuv 1.52,
+ *   internal APFS SSD): an 11-byte append costs 0.01 ms unflushed, 5.5 ms through
+ *   `fsyncSync`, and 0.10 ms through a raw `fsync(2)` against 4.8 ms through a raw
+ *   `F_FULLFSYNC` — the cost says which syscall is being made. Everywhere else libuv
+ *   calls plain `fsync(2)`, which is as durable as the drive's honesty about its own
+ *   write cache. So a **power loss** can lose a blob or a journal line this code has
+ *   already `fsync`ed and reported as written on any non-Apple platform whose drive
+ *   lies, and on Apple only where `F_FULLFSYNC` itself fails and libuv degrades
+ *   silently — some non-APFS and network mounts. A **process** crash cannot lose one
+ *   anywhere: the bytes are in the page cache and the publish is still atomic. Nothing
+ *   is ever handed back unverified either way, so the worst a lost blob can produce is
+ *   {@link UnknownHashError} — never wrong bytes presented as right ones. "Crash-safe"
+ *   is the claim, deliberately not "power-loss-proof": the durability is real, it is
+ *   just not unconditional.
+ * - **Reads verify.** `retrieve()`, `peek()` and `has()` re-hash the bytes and refuse a
+ *   mismatch with {@link StoreCorruptionError} — a damaged blob is never handed back as
+ *   a retrieval nor reported as present, and "we hold damaged bytes" is distinct from
+ *   {@link UnknownHashError}'s "never existed". The guard in
+ *   `test/guards/persistent-store.test.ts` watches this.
  * - **Counters survive a restart.** Every `retrieve()` appends one `fsync`ed line to
  *   `retrievals.log`, and `stats()` is a fold over it — so `expansionRate` stays
  *   meaningful across a whole session, not just one process. A crash in the middle of
@@ -86,6 +106,25 @@ export declare class DirectoryElisionStore implements ElisionStore {
     put(content: string): string;
     peek(hash: string): string | undefined;
     retrieve(hash: string): string;
+    /**
+     * Whether this hash can be **retrieved** — verified, exactly as {@link peek} and
+     * {@link retrieve} verify, because it is `peek()`.
+     *
+     * `has()` used to be the one read that skipped verification: a damaged blob answered
+     * `true` and then threw {@link StoreCorruptionError} on the very next line, so a
+     * consumer that checked before retrieving was told a lie by the cheaper call. The two
+     * answers now come from one place and cannot drift: `true` means the bytes are there
+     * and hash to their name, `false` means this store never held them, and damage is
+     * raised rather than hidden behind a boolean — the same distinction `peek()` draws
+     * between "we hold damaged bytes" and "never existed".
+     *
+     * It stays uncounted: a check is not the model asking for material back, and counting
+     * one would inflate `retrieveCalls` and with it the expansion rate, which is the one
+     * number this library exists to keep honest. So no journal line is written here, not
+     * even for the corrupt case — `retrieve()` journals that when the model asks.
+     *
+     * @throws {StoreCorruptionError} when the stored bytes do not hash to their name.
+     */
     has(hash: string): boolean;
     /**
      * The five directly-observed counts, every one read off the disk — a scan of

package/dist/store-dir.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"store-dir.d.ts","sourceRoot":"","sources":["../src/store-dir.ts"],"names":[],"mappings":"AAyBA,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,YAAY,CAAC;AACtD,OAAO,KAAK,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,YAAY,CAAC;AAE9D;;;;GAIG;AACH,eAAO,MAAM,sBAAsB,wBAAwB,CAAC;AAC5D,eAAO,MAAM,uBAAuB,IAAI,CAAC;AAYzC,4FAA4F;AAC5F,MAAM,WAAW,4BAA4B;IAC3C;;;OAGG;IACH,QAAQ,CAAC,IAAI,CAAC,EAAE,CAAC,OAAO,EAAE,MAAM,KAAK,MAAM,CAAC;CAC7C;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgEG;AACH,qBAAa,qBAAsB,YAAW,YAAY;;IAMxD,YAAY,IAAI,EAAE,MAAM,EAAE,OAAO,GAAE,4BAAiC,EAiBnE;IAED,GAAG,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAoC3B;IAED,IAAI,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS,CAKrC;IAED,QAAQ,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAY7B;IAED,GAAG,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAEzB;IAED;;;;OAIG;IACH,WAAW,IAAI,mBAAmB,CAqBjC;IAED,KAAK,IAAI,aAAa,CAErB;CA6IF"}
+{"version":3,"file":"store-dir.d.ts","sourceRoot":"","sources":["../src/store-dir.ts"],"names":[],"mappings":"AAyBA,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,YAAY,CAAC;AACtD,OAAO,KAAK,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,YAAY,CAAC;AAE9D;;;;GAIG;AACH,eAAO,MAAM,sBAAsB,wBAAwB,CAAC;AAC5D,eAAO,MAAM,uBAAuB,IAAI,CAAC;AAYzC,4FAA4F;AAC5F,MAAM,WAAW,4BAA4B;IAC3C;;;OAGG;IACH,QAAQ,CAAC,IAAI,CAAC,EAAE,CAAC,OAAO,EAAE,MAAM,KAAK,MAAM,CAAC;CAC7C;AAED;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoFG;AACH,qBAAa,qBAAsB,YAAW,YAAY;;IAMxD,YAAY,IAAI,EAAE,MAAM,EAAE,OAAO,GAAE,4BAAiC,EAiBnE;IAED,GAAG,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAoC3B;IAED,IAAI,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS,CAKrC;IAED,QAAQ,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAY7B;IAED;;;;;;;;;;;;;;;;;;OAkBG;IACH,GAAG,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAEzB;IAED;;;;OAIG;IACH,WAAW,IAAI,mBAAmB,CAqBjC;IAED,KAAK,IAAI,aAAa,CAErB;CA6IF","sourcesContent":["import {\n  closeSync,\n  fsyncSync,\n  linkSync,\n  mkdirSync,\n  openSync,\n  readdirSync,\n  readFileSync,\n  statSync,\n  unlinkSync,\n  writeSync,\n} from 'node:fs';\nimport { join, resolve } from 'node:path';\nimport { randomBytes } from 'node:crypto';\nimport process from 'node:process';\n\nimport {\n  HashCollisionError,\n  SmeltError,\n  StoreCorruptionError,\n  StoreFormatError,\n  UnknownHashError,\n} from './errors.ts';\nimport { contentHash } from './hash.ts';\nimport { retrieveStats } from './stats.ts';\nimport type { RawRetrieveCounters } from './stats.ts';\nimport type { ElisionStore, RetrieveStats } from './types.ts';\n\n/**\n * The format marker every store directory carries, and the one version this code\n * understands. A future layout is a new version, refused loudly by old code — never a\n * quiet reinterpretation of someone's stored bytes.\n */\nexport const DIRECTORY_STORE_FORMAT = 'smelt-elision-store';\nexport const DIRECTORY_STORE_VERSION = 1;\n\n/**\n * What a storage key may look like. `contentHash` produces 16 lowercase hex characters;\n * the pattern is wider so an injected test hash still works, and strict enough that a\n * key can never traverse out of `blobs/` or collide with `format.json`.\n */\nconst KEY_PATTERN = /^[0-9a-f]{4,128}$/;\n\n/** One journal line: a kind, a space, and the hash as a JSON string literal. */\nconst LOG_LINE = /^(hit|miss|corrupt) (\"(?:[^\"\\\\]|\\\\.)*\")$/;\n\n/** See {@link MemoryElisionStoreOptions} in `store.ts` — same escape hatch, same reason. */\nexport interface DirectoryElisionStoreOptions {\n  /**\n   * Override the hash function, so the collision branch — unreachable with sha256 —\n   * can be tested. Production has no reason to pass this.\n   */\n  readonly hash?: (content: string) => string;\n}\n\n/**\n * A persistent {@link ElisionStore} over a content-addressed directory. `node:fs` only —\n * no SQLite, no new dependency, nothing that phones home. Elisions put here outlive the\n * process, so a long-lived agent session can `retrieve()` across restarts.\n *\n * ## Storage layout\n *\n * ```text\n * <root>/\n *   format.json      { \"format\": \"smelt-elision-store\", \"version\": 1 } — refused if unknown\n *   blobs/<hash>     one file per elision: the exact UTF-8 bytes, named by their content hash\n *   tmp/             staging for atomic writes; never read, safe to sweep\n *   retrievals.log   append-only journal: `hit \"<hash>\"` | `miss \"<hash>\"` | `corrupt \"<hash>\"`\n * ```\n *\n * **Nothing lives in memory.** Every read — `stats()` included — comes off the disk, so\n * two instances over the same directory (two processes, or one process before and after\n * a restart) always agree. `stats()` is a scan; elision counts per session are small and\n * retrieval is the model asking for material back, which is rare by design.\n *\n * ## Durability\n *\n * - **Writes are crash-safe.** A blob is written to `tmp/`, `fsync`ed, then `link(2)`ed\n *   into `blobs/` — an atomic, no-clobber publish. A torn write dies in `tmp/`, where\n *   nothing looks; a name in `blobs/` always refers to a fully written file.\n * - **`fsync` is only as strong as the platform makes it.** Every flush here is Node's\n *   `fsyncSync`, which is libuv's `uv_fs_fsync`, and what that reaches the hardware with\n *   differs by platform. On Apple it is strong: libuv knows macOS's own `fsync(2)` only\n *   hands the write to the drive, so its `__APPLE__` branch issues\n *   `fcntl(fd, F_FULLFSYNC)` — a real drive-cache flush — before falling back to\n *   `F_BARRIERFSYNC` and then plain `fsync(2)`. Measured here (Node 26, libuv 1.52,\n *   internal APFS SSD): an 11-byte append costs 0.01 ms unflushed, 5.5 ms through\n *   `fsyncSync`, and 0.10 ms through a raw `fsync(2)` against 4.8 ms through a raw\n *   `F_FULLFSYNC` — the cost says which syscall is being made. Everywhere else libuv\n *   calls plain `fsync(2)`, which is as durable as the drive's honesty about its own\n *   write cache. So a **power loss** can lose a blob or a journal line this code has\n *   already `fsync`ed and reported as written on any non-Apple platform whose drive\n *   lies, and on Apple only where `F_FULLFSYNC` itself fails and libuv degrades\n *   silently — some non-APFS and network mounts. A **process** crash cannot lose one\n *   anywhere: the bytes are in the page cache and the publish is still atomic. Nothing\n *   is ever handed back unverified either way, so the worst a lost blob can produce is\n *   {@link UnknownHashError} — never wrong bytes presented as right ones. \"Crash-safe\"\n *   is the claim, deliberately not \"power-loss-proof\": the durability is real, it is\n *   just not unconditional.\n * - **Reads verify.** `retrieve()`, `peek()` and `has()` re-hash the bytes and refuse a\n *   mismatch with {@link StoreCorruptionError} — a damaged blob is never handed back as\n *   a retrieval nor reported as present, and \"we hold damaged bytes\" is distinct from\n *   {@link UnknownHashError}'s \"never existed\". The guard in\n *   `test/guards/persistent-store.test.ts` watches this.\n * - **Counters survive a restart.** Every `retrieve()` appends one `fsync`ed line to\n *   `retrievals.log`, and `stats()` is a fold over it — so `expansionRate` stays\n *   meaningful across a whole session, not just one process. A crash in the middle of\n *   an append can tear at most that one line; a torn tail is skipped, costing at most\n *   the single count that was being written when the process died.\n * - **Concurrent writers are safe.** `link(2)` refuses to clobber, so two processes\n *   putting at once race to publish and the loser verifies byte-for-byte agreement with\n *   the winner — identical content dedupes, different content under one hash is a\n *   {@link HashCollisionError}. Journal appends use `O_APPEND`. Tested with two real\n *   processes in `test/store-dir.test.ts`.\n *\n * ## No eviction\n *\n * Same rule as {@link MemoryElisionStore}: no cap, no LRU, no `clear()`. A store that\n * can forget turns Law 3 into \"reversible, usually\". Elided text is smaller than the\n * session that produced it; if disk pressure ever forces a cap, retrieval of an evicted\n * hash must throw a distinct \"evicted\" error — never {@link UnknownHashError} — so the\n * model can tell \"we lost it\" from \"never existed\". Today there is no such error because\n * there is no such cap.\n *\n * ## Two deliberate choices around the edges\n *\n * - **The root is resolved to an absolute path at construction.** Every later path is\n *   joined from that, so a `process.chdir()` after construction cannot silently\n *   re-target the store — the bytes a relative-rooted store put before a chdir would\n *   otherwise be unreachable after it, which reads exactly like data loss.\n * - **A failed journal append never withholds intact bytes.** `retrieve()`'s order of\n *   business is: read, verify, count, return. When the *count* cannot be written (a\n *   read-only journal, a full disk), the bytes are still returned — they are verified\n *   and the caller asked for them; refusing would turn a bookkeeping failure into\n *   Law 3 breaking. The failure is surfaced distinctly instead: a\n *   `process.emitWarning` with name `SmeltCounterWriteFailure`, so \"your retrieval\n *   worked\" and \"your counters just went quiet\" stay two separate facts. The same\n *   applies to the `miss`/`corrupt` journal lines: the store's own error for the\n *   lookup still wins over the journal's I/O error.\n */\nexport class DirectoryElisionStore implements ElisionStore {\n  readonly #blobsDir: string;\n  readonly #tmpDir: string;\n  readonly #logPath: string;\n  readonly #hash: (content: string) => string;\n\n  constructor(root: string, options: DirectoryElisionStoreOptions = {}) {\n    this.#hash = options.hash ?? contentHash;\n    // Resolve NOW, against the working directory the caller constructed with — a\n    // later chdir must never re-point an already-constructed store. See the class doc.\n    const absoluteRoot = resolve(root);\n    this.#blobsDir = join(absoluteRoot, 'blobs');\n    this.#tmpDir = join(absoluteRoot, 'tmp');\n    this.#logPath = join(absoluteRoot, 'retrievals.log');\n    const markerPath = join(absoluteRoot, 'format.json');\n    // Validate before mutating: a directory carrying a marker this code does not\n    // understand is refused with the directory exactly as it was found — no blobs/,\n    // no tmp/, no staged temp file created inside someone else's layout.\n    const existing = this.#readMarker(markerPath);\n    if (existing !== undefined) this.#verifyMarker(markerPath, existing);\n    mkdirSync(this.#blobsDir, { recursive: true });\n    mkdirSync(this.#tmpDir, { recursive: true });\n    this.#claimFormat(markerPath);\n  }\n\n  put(content: string): string {\n    const hash = this.#hash(content);\n    if (!KEY_PATTERN.test(hash)) {\n      throw new SmeltError(\n        `smelt: hash \"${hash}\" is not usable as a storage key — it must match ` +\n          `${String(KEY_PATTERN)} so it can name a file inside blobs/ and nothing else.`,\n      );\n    }\n    const existing = this.#readBlob(hash);\n    if (existing !== undefined) {\n      // Verify the stored bytes before comparing: a damaged blob is corruption, not a\n      // collision. Only intact bytes that still differ earn HashCollisionError.\n      if (this.#hash(existing) !== hash) throw new StoreCorruptionError(hash);\n      if (existing !== content) throw new HashCollisionError(hash);\n      return hash;\n    }\n    const tmpPath = this.#writeTemp(content);\n    try {\n      // link(2) is the atomic, no-clobber publish: it fails with EEXIST rather than\n      // overwrite, so a concurrent writer can never silently replace someone's bytes.\n      linkSync(tmpPath, join(this.#blobsDir, hash));\n    } catch (error) {\n      if ((error as { code?: string }).code !== 'EEXIST') throw error;\n      // Another writer published this hash between our existence check and our link.\n      // Same bytes: idempotent put, done. Damaged or vanished bytes: corruption — the\n      // store was torn or edited outside smelt. Intact different bytes: a collision.\n      const winner = this.#readBlob(hash);\n      if (winner === undefined || this.#hash(winner) !== hash) {\n        throw new StoreCorruptionError(hash);\n      }\n      if (winner !== content) throw new HashCollisionError(hash);\n    } finally {\n      unlinkSync(tmpPath);\n    }\n    fsyncDirBestEffort(this.#blobsDir);\n    return hash;\n  }\n\n  peek(hash: string): string | undefined {\n    const content = this.#readBlob(hash);\n    if (content === undefined) return undefined;\n    if (this.#hash(content) !== hash) throw new StoreCorruptionError(hash);\n    return content;\n  }\n\n  retrieve(hash: string): string {\n    const content = this.#readBlob(hash);\n    if (content === undefined) {\n      this.#appendLogCounting('miss', hash);\n      throw new UnknownHashError(hash);\n    }\n    if (this.#hash(content) !== hash) {\n      this.#appendLogCounting('corrupt', hash);\n      throw new StoreCorruptionError(hash);\n    }\n    this.#appendLogCounting('hit', hash);\n    return content;\n  }\n\n  /**\n   * Whether this hash can be **retrieved** — verified, exactly as {@link peek} and\n   * {@link retrieve} verify, because it is `peek()`.\n   *\n   * `has()` used to be the one read that skipped verification: a damaged blob answered\n   * `true` and then threw {@link StoreCorruptionError} on the very next line, so a\n   * consumer that checked before retrieving was told a lie by the cheaper call. The two\n   * answers now come from one place and cannot drift: `true` means the bytes are there\n   * and hash to their name, `false` means this store never held them, and damage is\n   * raised rather than hidden behind a boolean — the same distinction `peek()` draws\n   * between \"we hold damaged bytes\" and \"never existed\".\n   *\n   * It stays uncounted: a check is not the model asking for material back, and counting\n   * one would inflate `retrieveCalls` and with it the expansion rate, which is the one\n   * number this library exists to keep honest. So no journal line is written here, not\n   * even for the corrupt case — `retrieve()` journals that when the model asks.\n   *\n   * @throws {StoreCorruptionError} when the stored bytes do not hash to their name.\n   */\n  has(hash: string): boolean {\n    return this.peek(hash) !== undefined;\n  }\n\n  /**\n   * The five directly-observed counts, every one read off the disk — a scan of\n   * `blobs/` plus a fold over `retrievals.log`. See {@link RawRetrieveCounters}; the\n   * derived half of the stats comes from the shared `retrieveStats()`, never here.\n   */\n  rawCounters(): RawRetrieveCounters {\n    let elisionsStored = 0;\n    let bytesStored = 0;\n    for (const entry of readdirSync(this.#blobsDir)) {\n      if (!KEY_PATTERN.test(entry)) continue; // `.DS_Store` and friends are not blobs\n      elisionsStored += 1;\n      bytesStored += statSync(join(this.#blobsDir, entry)).size;\n    }\n\n    let retrieveCalls = 0;\n    let misses = 0;\n    const hits = new Set<string>();\n    for (const line of this.#readLog().split('\\n')) {\n      const match = LOG_LINE.exec(line);\n      if (match === null) continue; // a torn tail from a crash mid-append, or blank\n      retrieveCalls += 1;\n      if (match[1] === 'miss') misses += 1;\n      else if (match[1] === 'hit') hits.add(JSON.parse(match[2]!) as string);\n    }\n\n    return { elisionsStored, bytesStored, retrieveCalls, uniqueRetrieved: hits.size, misses };\n  }\n\n  stats(): RetrieveStats {\n    return retrieveStats(this.rawCounters());\n  }\n\n  /** The blob's exact content, or `undefined` when no such blob is stored. */\n  #readBlob(hash: string): string | undefined {\n    if (!KEY_PATTERN.test(hash)) return undefined; // never a path component\n    try {\n      return readFileSync(join(this.#blobsDir, hash), 'utf8');\n    } catch (error) {\n      if ((error as { code?: string }).code === 'ENOENT') return undefined;\n      throw error;\n    }\n  }\n\n  /** Write content to a unique file in `tmp/`, fsynced, and return its path. */\n  #writeTemp(content: string): string {\n    const tmpPath = join(this.#tmpDir, `${String(process.pid)}-${randomBytes(8).toString('hex')}`);\n    const fd = openSync(tmpPath, 'wx');\n    try {\n      // writeSync may write fewer bytes than asked; loop, or a short write would be\n      // fsynced and published under the full content's hash as a torn blob.\n      const bytes = Buffer.from(content, 'utf8');\n      let written = 0;\n      while (written < bytes.length) {\n        written += writeSync(fd, bytes, written);\n      }\n      fsyncSync(fd);\n    } finally {\n      closeSync(fd);\n    }\n    return tmpPath;\n  }\n\n  /**\n   * A journal append on the `retrieve()` path — counting, not custody. A failure here\n   * must never decide whether the caller gets its verified bytes (or its true error),\n   * so it is caught and surfaced as a distinct `process.emitWarning` — see the class\n   * doc, and the read-only-journal case in `test/store-dir.test.ts`.\n   */\n  #appendLogCounting(kind: 'hit' | 'miss' | 'corrupt', hash: string): void {\n    try {\n      this.#appendLog(kind, hash);\n    } catch (error) {\n      process.emitWarning(\n        `smelt: could not journal a \"${kind}\" for hash \"${hash}\" in ${this.#logPath} ` +\n          `(${error instanceof Error ? error.message : String(error)}). The retrieval ` +\n          `itself is unaffected, but this count is lost — retrieveCalls and ` +\n          `expansionRate now UNDER-report until the journal is writable again.`,\n        'SmeltCounterWriteFailure',\n      );\n    }\n  }\n\n  /**\n   * One durable journal line. The hash is JSON-encoded because `retrieve()` takes it\n   * from the model verbatim — a hash containing a newline must not forge a second line.\n   * The record starts with its own newline so a torn tail from an earlier crash — a\n   * partial record with no trailing newline — can never bleed into this one: the tear\n   * stays on its own line and is skipped by `stats()`, as blank lines are.\n   */\n  #appendLog(kind: 'hit' | 'miss' | 'corrupt', hash: string): void {\n    const fd = openSync(this.#logPath, 'a');\n    try {\n      const record = Buffer.from(`\\n${kind} ${JSON.stringify(hash)}\\n`, 'utf8');\n      let written = 0;\n      while (written < record.length) {\n        written += writeSync(fd, record, written);\n      }\n      fsyncSync(fd);\n    } finally {\n      closeSync(fd);\n    }\n  }\n\n  #readLog(): string {\n    try {\n      return readFileSync(this.#logPath, 'utf8');\n    } catch (error) {\n      if ((error as { code?: string }).code === 'ENOENT') return '';\n      throw error;\n    }\n  }\n\n  /**\n   * Write the format marker if this directory has none, or verify the one it has.\n   * Creation is atomic (write to `tmp/`, then `link`), so a concurrent creator never\n   * observes a half-written marker. The constructor pre-verified any pre-existing\n   * marker; the verify here catches only a concurrent creator's claim.\n   */\n  #claimFormat(markerPath: string): void {\n    const claim = (): string | undefined => {\n      const body = `${JSON.stringify({\n        format: DIRECTORY_STORE_FORMAT,\n        version: DIRECTORY_STORE_VERSION,\n      })}\\n`;\n      const tmpPath = this.#writeTemp(body);\n      try {\n        linkSync(tmpPath, markerPath);\n        return undefined; // claimed by us; nothing to verify\n      } catch (error) {\n        if ((error as { code?: string }).code !== 'EEXIST') throw error;\n        return readFileSync(markerPath, 'utf8');\n      } finally {\n        unlinkSync(tmpPath);\n      }\n    };\n\n    const existing = claim();\n    if (existing === undefined) return;\n    this.#verifyMarker(markerPath, existing);\n  }\n\n  /** The marker's body, or `undefined` when the directory carries none. */\n  #readMarker(markerPath: string): string | undefined {\n    try {\n      return readFileSync(markerPath, 'utf8');\n    } catch (error) {\n      if ((error as { code?: string }).code === 'ENOENT') return undefined;\n      throw error;\n    }\n  }\n\n  /** Refuse a marker this version of smelt does not understand. */\n  #verifyMarker(markerPath: string, existing: string): void {\n    let parsed: { format?: unknown; version?: unknown };\n    try {\n      parsed = JSON.parse(existing) as { format?: unknown; version?: unknown };\n    } catch {\n      throw new StoreFormatError(\n        `smelt: \"${markerPath}\" is not parseable JSON, so this directory cannot be ` +\n          `trusted as an elision store. Refusing to read or write it.`,\n      );\n    }\n    if (parsed.format !== DIRECTORY_STORE_FORMAT || parsed.version !== DIRECTORY_STORE_VERSION) {\n      throw new StoreFormatError(\n        `smelt: \"${markerPath}\" declares format ${JSON.stringify(parsed.format)} ` +\n          `version ${JSON.stringify(parsed.version)}; this code understands ` +\n          `\"${DIRECTORY_STORE_FORMAT}\" version ${String(DIRECTORY_STORE_VERSION)}. ` +\n          `Refusing to reinterpret someone else's layout.`,\n      );\n    }\n  }\n}\n\n/**\n * Flush the directory entry after a publish, so the *name* survives a crash as well as\n * the bytes. Where the platform refuses to fsync a directory (Windows does), the publish\n * is still atomic — only the durability of the directory entry falls back to the OS's\n * own schedule. Only that refusal is swallowed: a real I/O failure (`EIO`) propagates,\n * because \"the disk could not flush\" must never be reported as a successful put.\n *\n * Where it does work it is the same `fsyncSync` the blob's own flush uses, and so is\n * exactly as strong as that — see the durability note on {@link DirectoryElisionStore}\n * for what that means per platform.\n */\nfunction fsyncDirBestEffort(path: string): void {\n  let fd: number;\n  try {\n    fd = openSync(path, 'r');\n  } catch {\n    return; // the platform refuses to even open a directory for reading (Windows)\n  }\n  try {\n    fsyncSync(fd);\n  } catch (error) {\n    const code = (error as { code?: string }).code;\n    // EINVAL/ENOTSUP/EPERM/EBADF: the platform refuses to fsync a directory — see the\n    // doc comment. Anything else (EIO above all) is a genuine write failure.\n    if (code !== 'EINVAL' && code !== 'ENOTSUP' && code !== 'EPERM' && code !== 'EBADF') {\n      throw error;\n    }\n  } finally {\n    closeSync(fd);\n  }\n}\n"]}

package/dist/store-dir.js

@@ -45,10 +45,30 @@ const LOG_LINE = /^(hit|miss|corrupt) ("(?:[^"\\]|\\.)*")$/;
  * - **Writes are crash-safe.** A blob is written to `tmp/`, `fsync`ed, then `link(2)`ed
  *   into `blobs/` — an atomic, no-clobber publish. A torn write dies in `tmp/`, where
  *   nothing looks; a name in `blobs/` always refers to a fully written file.
- * - **Reads verify.** `retrieve()` and `peek()` re-hash the bytes and refuse a mismatch
- *   with {@link StoreCorruptionError} — a damaged blob is never handed back as a
- *   retrieval, and "we hold damaged bytes" is distinct from {@link UnknownHashError}'s
- *   "never existed". The guard in `test/guards/persistent-store.test.ts` watches this.
+ * - **`fsync` is only as strong as the platform makes it.** Every flush here is Node's
+ *   `fsyncSync`, which is libuv's `uv_fs_fsync`, and what that reaches the hardware with
+ *   differs by platform. On Apple it is strong: libuv knows macOS's own `fsync(2)` only
+ *   hands the write to the drive, so its `__APPLE__` branch issues
+ *   `fcntl(fd, F_FULLFSYNC)` — a real drive-cache flush — before falling back to
+ *   `F_BARRIERFSYNC` and then plain `fsync(2)`. Measured here (Node 26, libuv 1.52,
+ *   internal APFS SSD): an 11-byte append costs 0.01 ms unflushed, 5.5 ms through
+ *   `fsyncSync`, and 0.10 ms through a raw `fsync(2)` against 4.8 ms through a raw
+ *   `F_FULLFSYNC` — the cost says which syscall is being made. Everywhere else libuv
+ *   calls plain `fsync(2)`, which is as durable as the drive's honesty about its own
+ *   write cache. So a **power loss** can lose a blob or a journal line this code has
+ *   already `fsync`ed and reported as written on any non-Apple platform whose drive
+ *   lies, and on Apple only where `F_FULLFSYNC` itself fails and libuv degrades
+ *   silently — some non-APFS and network mounts. A **process** crash cannot lose one
+ *   anywhere: the bytes are in the page cache and the publish is still atomic. Nothing
+ *   is ever handed back unverified either way, so the worst a lost blob can produce is
+ *   {@link UnknownHashError} — never wrong bytes presented as right ones. "Crash-safe"
+ *   is the claim, deliberately not "power-loss-proof": the durability is real, it is
+ *   just not unconditional.
+ * - **Reads verify.** `retrieve()`, `peek()` and `has()` re-hash the bytes and refuse a
+ *   mismatch with {@link StoreCorruptionError} — a damaged blob is never handed back as
+ *   a retrieval nor reported as present, and "we hold damaged bytes" is distinct from
+ *   {@link UnknownHashError}'s "never existed". The guard in
+ *   `test/guards/persistent-store.test.ts` watches this.
  * - **Counters survive a restart.** Every `retrieve()` appends one `fsync`ed line to
  *   `retrievals.log`, and `stats()` is a fold over it — so `expansionRate` stays
  *   meaningful across a whole session, not just one process. A crash in the middle of
@@ -171,8 +191,27 @@ export class DirectoryElisionStore {
         this.#appendLogCounting('hit', hash);
         return content;
     }
+    /**
+     * Whether this hash can be **retrieved** — verified, exactly as {@link peek} and
+     * {@link retrieve} verify, because it is `peek()`.
+     *
+     * `has()` used to be the one read that skipped verification: a damaged blob answered
+     * `true` and then threw {@link StoreCorruptionError} on the very next line, so a
+     * consumer that checked before retrieving was told a lie by the cheaper call. The two
+     * answers now come from one place and cannot drift: `true` means the bytes are there
+     * and hash to their name, `false` means this store never held them, and damage is
+     * raised rather than hidden behind a boolean — the same distinction `peek()` draws
+     * between "we hold damaged bytes" and "never existed".
+     *
+     * It stays uncounted: a check is not the model asking for material back, and counting
+     * one would inflate `retrieveCalls` and with it the expansion rate, which is the one
+     * number this library exists to keep honest. So no journal line is written here, not
+     * even for the corrupt case — `retrieve()` journals that when the model asks.
+     *
+     * @throws {StoreCorruptionError} when the stored bytes do not hash to their name.
+     */
     has(hash) {
-        return this.#readBlob(hash) !== undefined;
+        return this.peek(hash) !== undefined;
     }
     /**
      * The five directly-observed counts, every one read off the disk — a scan of
@@ -352,6 +391,10 @@ export class DirectoryElisionStore {
  * is still atomic — only the durability of the directory entry falls back to the OS's
  * own schedule. Only that refusal is swallowed: a real I/O failure (`EIO`) propagates,
  * because "the disk could not flush" must never be reported as a successful put.
+ *
+ * Where it does work it is the same `fsyncSync` the blob's own flush uses, and so is
+ * exactly as strong as that — see the durability note on {@link DirectoryElisionStore}
+ * for what that means per platform.
  */
 function fsyncDirBestEffort(path) {
     let fd;