offhands 0.1.11 → 0.1.13

package/dist/daemon.mjs

@@ -6749,8 +6749,66 @@ var STOPWORDS = /* @__PURE__ */ new Set([
   "into",
   "if",
   "then",
-  "than"
+  "than",
+  // Instruction-shaped filler that says how to answer, not what to look at —
+  // by length alone these outranked the task's real nouns ("requirements",
+  // "mirroring" pushed "closing" out of the term list).
+  "requirements",
+  "requirement",
+  "whatever",
+  "existing",
+  "which",
+  "tell",
+  "bullets",
+  "bullet",
+  "fewer",
+  "five",
+  "only",
+  "why",
+  "any",
+  "following",
+  "something",
+  "thing",
+  "things",
+  "just",
+  "about",
+  "there",
+  "their",
+  "them",
+  "they",
+  "these",
+  "those",
+  "your",
+  "back",
+  "let",
+  "lets",
+  "read",
+  "write",
+  "show",
+  "give",
+  "explain",
+  "describe",
+  "files",
+  "file",
+  // Generic action verbs: in code these are everywhere (createX, deleteY), so
+  // "do not create, edit, or delete any files" pulled in guildCreate.ts and
+  // guildDelete.ts. The nouns in a prompt carry the signal; the verb rarely does.
+  "create",
+  "creates",
+  "edit",
+  "edits",
+  "delete",
+  "deletes",
+  "change",
+  "changes",
+  "modify",
+  "update",
+  "updates",
+  "implement",
+  "implementation"
 ]);
+var MAX_TERMS = 24;
+var RIPGREP_TERMS = 8;
 function extractTerms(prompt, maxTerms = 8) {
   const raw = prompt.toLowerCase().split(/[^a-z0-9._-]+/).map((t2) => t2.replace(/^[._-]+|[._-]+$/g, "")).filter((t2) => t2.length >= 3 && !STOPWORDS.has(t2) && !/^\d+$/.test(t2));
   const seen = /* @__PURE__ */ new Set();
@@ -6813,6 +6871,10 @@ function pathWords(path) {
 }
 var NOISE_FILES = /* @__PURE__ */ new Set(["package-lock.json", "pnpm-lock.yaml", "yarn.lock", "bun.lockb", "bun.lock", "composer.lock", "cargo.lock", "poetry.lock"]);
 var isNoise = (p2) => NOISE_FILES.has((p2.split("/").pop() ?? "").toLowerCase());
+var MAX_MATCHED_SHOWN = 14;
+var MAX_CONTENT_PROBES = 10;
+var MAX_CONTENT_MATCHES = 6;
+var MAX_UNMATCHED_WITH_MATCHES = 3;
 var MAX_NAME_MATCHES = 10;
 var MAX_DIRTY_ONLY = 8;
 var MAX_RECENT = 6;
@@ -6843,7 +6905,14 @@ async function buildContext(workspace, prompt, opts = {}) {
   if (findings.length === 0) return null;
   const rank = (f2) => f2.matchScore > 0 ? f2.dirty ? 5 : 4 : f2.dirty ? 2 : 1;
   findings.sort((a2, b2) => rank(b2) - rank(a2) || b2.matchScore - a2.matchScore);
-  const { text: block, selectedCount } = formatBlock(prompt, findings, gitState, maxChars);
+  const strongMatches = findings.filter((f2) => f2.matchScore > 0).length;
+  const unmatchedCap = strongMatches >= 3 ? MAX_UNMATCHED_WITH_MATCHES : Infinity;
+  let unmatched = 0;
+  let matchedShown = 0;
+  const shown = findings.filter(
+    (f2) => f2.matchScore > 0 ? ++matchedShown <= MAX_MATCHED_SHOWN : ++unmatched <= unmatchedCap
+  );
+  const { text: block, selectedCount } = formatBlock(prompt, shown, gitState, maxChars);
   if (selectedCount === 0) return null;
   return {
     block,
@@ -6916,11 +6985,13 @@ async function gatherFindings(workspace, prompt) {
   }
   const ripgrepAvailable = await detectRipgrep();
   retrievalOpsRun++;
-  const terms = extractTerms(prompt);
+  const terms = extractTerms(prompt, MAX_TERMS);
+  let listedFileCount = 0;
   if (terms.length > 0) {
     const listing = await git(workspace, "ls-files", "-co", "--exclude-standard", "-z");
     retrievalOpsRun++;
     const all = (listing ?? "").split("\0").filter(Boolean);
+    listedFileCount = all.length;
     if (all.length > 0 && all.length <= MAX_LS_FILES) {
       const stems = /* @__PURE__ */ new Map();
       const literals = [];
@@ -6971,8 +7042,45 @@ async function gatherFindings(workspace, prompt) {
       }
     }
   }
+  if (!ripgrepAvailable && terms.length > 0) {
+    const probes = /* @__PURE__ */ new Map();
+    for (const t2 of terms) {
+      const s2 = stem(t2);
+      if (s2.length >= 4 && !probes.has(s2)) probes.set(s2, t2);
+      if (probes.size >= MAX_CONTENT_PROBES) break;
+    }
+    const results = await Promise.all(
+      [...probes].map(async ([s2, wording]) => {
+        const out = await git(workspace, "grep", "-I", "-i", "-F", "-l", "--untracked", "-e", s2);
+        retrievalOpsRun++;
+        return { wording, files: (out ?? "").split("\n").map((l2) => l2.trim()).filter(Boolean) };
+      })
+    );
+    const corpus = Math.max(listedFileCount, 1);
+    const contentScore = /* @__PURE__ */ new Map();
+    for (const { wording, files } of results) {
+      if (files.length === 0 || files.length > corpus / 2) continue;
+      const weight = 0.5 * Math.log(1 + corpus / files.length);
+      for (const file of files.slice(0, 300)) {
+        const entry = contentScore.get(file) ?? { score: 0, words: [] };
+        entry.score += weight;
+        entry.words.push(wording);
+        contentScore.set(file, entry);
+      }
+    }
+    let taken = 0;
+    for (const [file, { score, words }] of [...contentScore].sort((a2, b2) => b2[1].score - a2[1].score)) {
+      if (taken >= MAX_CONTENT_MATCHES) break;
+      if (isSecretLike(file) || !existsSync7(join6(workspace, file))) continue;
+      const f2 = touch(file);
+      f2.matchScore += score;
+      if (isDirty2(file)) f2.dirty = true;
+      note(f2, `content mentions ${words.slice(0, 3).map((w2) => `"${w2}"`).join(", ")}`);
+      taken++;
+    }
+  }
   if (ripgrepAvailable && terms.length > 0) {
-    const hits = await searchTerms(workspace, terms);
+    const hits = await searchTerms(workspace, terms.slice(0, RIPGREP_TERMS));
     retrievalOpsRun++;
     for (const hit of hits) {
       const file = hit.file.replace(/\\/g, "/").replace(/^\.\//, "");

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "offhands",
-  "version": "0.1.11",
+  "version": "0.1.13",
   "description": "Phone → coding-agent relay. Daemon runs on your laptop, PWA on your phone. E2E encrypted.",
   "license": "UNLICENSED",
   "type": "module",