ultimate-pi 0.3.1 → 0.4.1

package/vendor/pi-vcc/src/core/search-entries.ts

@@ -0,0 +1,221 @@
+import type { Message } from "@mariozechner/pi-ai";
+import type { RenderedEntry } from "./render-entries";
+import { textOf } from "./content";
+
+export interface SearchHit extends RenderedEntry {
+  /** Context snippet around the first matched term (only when query provided) */
+  snippet?: string;
+  /** Number of query terms matched (for ranking) */
+  matchCount?: number;
+}
+
+const escapeRegex = (s: string): string =>
+  s.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+
+/** Try to compile as regex; fall back to escaped literal. */
+const safeRegex = (pattern: string): RegExp => {
+  try {
+    return new RegExp(pattern, "i");
+  } catch {
+    return new RegExp(escapeRegex(pattern), "i");
+  }
+};
+
+/** Detect if the query looks like a single regex pattern (contains regex metacharacters). */
+const looksLikeRegex = (query: string): boolean =>
+  /[|*+?{}()[\]\\^$.]/.test(query);
+
+/** Build a regex for snippet highlighting — matches first available term. */
+const snippetRegex = (terms: string[]): RegExp => {
+  const alts = terms.map((t) => {
+    try {
+      // Validate that it's a valid regex
+      new RegExp(t, "i");
+      return t;
+    } catch {
+      return escapeRegex(t);
+    }
+  });
+  return new RegExp(alts.join("|"), "i");
+};
+
+// ── Stopwords for natural language queries ──
+const STOPWORDS = new Set([
+  // English
+  "the", "a", "an", "is", "are", "was", "were", "be", "been", "being",
+  "have", "has", "had", "do", "does", "did", "will", "would", "could",
+  "should", "may", "might", "can", "shall", "of", "in", "to", "for",
+  "with", "on", "at", "from", "by", "as", "into", "through", "during",
+  "before", "after", "above", "below", "between", "out", "off", "over",
+  "under", "again", "further", "then", "once", "here", "there", "when",
+  "where", "why", "how", "all", "both", "each", "few", "more", "most",
+  "other", "some", "such", "no", "nor", "not", "only", "own", "same",
+  "so", "than", "too", "very", "just", "about", "it", "its", "that",
+  "this", "what", "which", "who", "whom", "these", "those",
+]);
+
+/** Remove stopwords, keep meaningful terms. */
+const filterStopwords = (terms: string[]): string[] => {
+  const meaningful = terms.filter((t) => !STOPWORDS.has(t.toLowerCase()) && t.length > 1);
+  // If all terms were stopwords, return original (don't lose everything)
+  return meaningful.length > 0 ? meaningful : terms;
+};
+
+/** Count how many distinct terms match the haystack. */
+const countMatches = (hay: string, terms: string[]): number => {
+  let count = 0;
+  for (const t of terms) {
+    if (safeRegex(t).test(hay)) count++;
+  }
+  return count;
+};
+
+// ── BM25-lite scoring ──
+const BM25_K = 1.2;
+const BM25_B = 0.75;
+
+/** Count occurrences of a regex pattern in text. */
+const termFreq = (text: string, pattern: RegExp): number => {
+  const matches = text.match(new RegExp(pattern.source, "gi"));
+  return matches ? matches.length : 0;
+};
+
+interface BM25Context {
+  n: number;         // total docs
+  avgDl: number;     // average doc length (words)
+  df: Map<string, number>; // term -> number of docs containing it
+}
+
+/** Precompute IDF and avgDl across all docs. */
+const buildBM25Context = (docs: string[], terms: string[]): BM25Context => {
+  const n = docs.length;
+  const df = new Map<string, number>();
+  let totalLen = 0;
+
+  for (const doc of docs) {
+    totalLen += doc.split(/\s+/).length;
+    for (const t of terms) {
+      if (safeRegex(t).test(doc)) {
+        df.set(t, (df.get(t) ?? 0) + 1);
+      }
+    }
+  }
+
+  return { n, avgDl: totalLen / Math.max(n, 1), df };
+};
+
+/** BM25 score for a single doc against query terms. */
+const bm25Score = (doc: string, terms: string[], ctx: BM25Context): number => {
+  const dl = doc.split(/\s+/).length;
+  let score = 0;
+
+  for (const t of terms) {
+    const tf = termFreq(doc, safeRegex(t));
+    if (tf === 0) continue;
+
+    const docFreq = ctx.df.get(t) ?? 0;
+    // IDF: log((N - df + 0.5) / (df + 0.5) + 1)
+    const idf = Math.log((ctx.n - docFreq + 0.5) / (docFreq + 0.5) + 1);
+    // TF saturation with length normalization
+    const tfNorm = (tf * (BM25_K + 1)) / (tf + BM25_K * (1 - BM25_B + BM25_B * dl / ctx.avgDl));
+    score += idf * tfNorm;
+  }
+
+  return score;
+};
+
+/** Line-based snippet: ±contextLines around first regex match. */
+const lineSnippet = (text: string, regex: RegExp, contextLines = 2): string | undefined => {
+  const lines = text.split("\n");
+  let matchIdx = -1;
+  for (let i = 0; i < lines.length; i++) {
+    if (regex.test(lines[i])) {
+      matchIdx = i;
+      break;
+    }
+  }
+  if (matchIdx === -1) return undefined;
+
+  const start = Math.max(0, matchIdx - contextLines);
+  const end = Math.min(lines.length, matchIdx + contextLines + 1);
+  const slice = lines.slice(start, end);
+
+  const parts: string[] = [];
+  if (start > 0) parts.push(`...(${start} lines above)`);
+  parts.push(...slice);
+  if (end < lines.length) parts.push(`...(${lines.length - end} lines below)`);
+  return parts.join("\n");
+};
+
+/** Build full searchable text for a message. */
+const fullText = (msg: Message): string => {
+  if ((msg as any).role === "bashExecution") {
+    return `${(msg as any).command ?? ""} ${(msg as any).output ?? ""}`;
+  }
+  return textOf(msg.content);
+};
+
+export const searchEntries = (
+  entries: RenderedEntry[],
+  messages: Message[],
+  query?: string,
+): SearchHit[] => {
+  if (!query?.trim()) return entries;
+
+  const rawQuery = query.trim();
+
+  // If query looks like a single regex pattern (contains metacharacters),
+  // treat the whole thing as one pattern — don't split into terms
+  if (looksLikeRegex(rawQuery)) {
+    const regex = safeRegex(rawQuery);
+    const hits: SearchHit[] = [];
+    for (let i = 0; i < entries.length; i++) {
+      const e = entries[i];
+      const msg = messages[i];
+      const text = msg ? fullText(msg) : e.summary;
+      const filePart = e.files?.join(" ") ?? "";
+      const hay = `${e.role} ${text} ${filePart}`;
+      if (regex.test(hay)) {
+        const snip = lineSnippet(text, regex);
+        hits.push({ ...e, snippet: snip, matchCount: 1 });
+      }
+    }
+    return hits;
+  }
+
+  // Natural language / multi-word query: BM25 scoring
+  const rawTerms = rawQuery.split(/\s+/);
+  const terms = filterStopwords(rawTerms);
+  const snipRe = snippetRegex(terms);
+
+  // Build all docs for BM25 context
+  const docs: string[] = [];
+  for (let i = 0; i < entries.length; i++) {
+    const e = entries[i];
+    const msg = messages[i];
+    const text = msg ? fullText(msg) : e.summary;
+    const filePart = e.files?.join(" ") ?? "";
+    docs.push(`${e.role} ${text} ${filePart}`);
+  }
+
+  const ctx = buildBM25Context(docs, terms);
+
+  const scored: Array<{ hit: SearchHit; score: number }> = [];
+  for (let i = 0; i < entries.length; i++) {
+    const e = entries[i];
+    const hay = docs[i];
+    const mc = countMatches(hay, terms);
+    if (mc === 0) continue;
+    const score = bm25Score(hay, terms, ctx);
+    const text = messages[i] ? fullText(messages[i]) : e.summary;
+    const snip = lineSnippet(text, snipRe);
+    scored.push({
+      hit: { ...e, snippet: snip, matchCount: mc },
+      score,
+    });
+  }
+
+  // Sort by BM25 score desc
+  scored.sort((a, b) => b.score - a.score);
+  return scored.map((s) => s.hit);
+};

package/vendor/pi-vcc/src/core/settings.ts

@@ -0,0 +1,8 @@
+/**
+ * ultimate-pi harness settings (env-only). Re-exported for vendored pi-vcc layout.
+ */
+export type { PiVccSettings } from "../../../../.pi/extensions/lib/harness-vcc-settings.js";
+export {
+	loadSettings,
+	scaffoldSettings,
+} from "../../../../.pi/extensions/lib/harness-vcc-settings.js";

package/vendor/pi-vcc/src/core/skill-collapse.ts

@@ -0,0 +1,35 @@
+/** Shared skill-tag collapse utilities */
+
+const SKILL_TAG_RE = /^-?\s*<skill\s+name="([^"]+)"/;
+const SKILL_CLOSE_RE = /^-?\s*<\/skill>/;
+
+/** Collapse skill tags in an array of lines — dedup by name, drop all content inside block */
+export const collapseSkillLines = (lines: string[]): string[] => {
+  const result: string[] = [];
+  const seenSkills = new Set<string>();
+  let insideSkill = false;
+
+  for (const line of lines) {
+    const skillMatch = line.match(SKILL_TAG_RE);
+    if (skillMatch) {
+      insideSkill = true;
+      const name = skillMatch[1];
+      if (!seenSkills.has(name)) {
+        seenSkills.add(name);
+        result.push(`[skill: ${name}]`);
+      }
+      continue;
+    }
+    if (insideSkill) {
+      if (SKILL_CLOSE_RE.test(line)) insideSkill = false;
+      continue;
+    }
+    result.push(line);
+  }
+  return result;
+};
+
+/** Collapse <skill name="X" ...>...</skill> blocks in raw text */
+const SKILL_BLOCK_RE = /<skill\s+name="([^"]+)"[^>]*>[\s\S]*?(?:<\/skill>|$)/g;
+export const collapseSkillText = (text: string): string =>
+  text.replace(SKILL_BLOCK_RE, (_, name) => `[skill: ${name}]`);

package/vendor/pi-vcc/src/core/summarize.ts

@@ -0,0 +1,157 @@
+import type { Message } from "@mariozechner/pi-ai";
+import type { FileOps } from "../types";
+import { normalize } from "./normalize";
+import { filterNoise } from "./filter-noise";
+import { buildSections } from "./build-sections";
+import { formatSummary, capBrief, RECALL_NOTE } from "./format";
+
+export interface CompileInput {
+  messages: Message[];
+  previousSummary?: string;
+  fileOps?: FileOps;
+}
+
+const HEADER_NAMES = ["Session Goal", "Files And Changes", "Commits", "Outstanding Context", "User Preferences"];
+
+const SEPARATOR = "\n\n---\n\n";
+
+/** Extract a named section from summary text */
+const sectionOf = (text: string, header: string): string => {
+  const tag = `[${header}]`;
+  const start = text.indexOf(tag);
+  if (start < 0) return "";
+  const after = text.slice(start);
+  // Find next section header or separator
+  const nextSection = HEADER_NAMES
+    .filter((h) => h !== header)
+    .map((h) => after.indexOf(`[${h}]`))
+    .filter((n) => n > 0);
+  const nextSep = after.indexOf("\n\n---\n\n");
+  const candidates = [...nextSection, ...(nextSep > 0 ? [nextSep] : [])].sort((a, b) => a - b);
+  const end = candidates[0];
+  return (end ? after.slice(0, end) : after).trim();
+};
+
+/** Extract the brief transcript part (everything after ---) */
+const briefOf = (text: string): string => {
+  const idx = text.indexOf(SEPARATOR);
+  if (idx < 0) return "";
+  return text.slice(idx + SEPARATOR.length).trim();
+};
+
+/** Merge a header section */
+const mergeHeaderSection = (header: string, prev: string, fresh: string): string => {
+  // Outstanding Context is volatile -- always use fresh only
+  if (header === "Outstanding Context") return fresh;
+  if (!prev) return fresh;
+  if (!fresh) return prev;
+
+  // Files And Changes: merge by category (Modified/Created/Read), dedup paths
+  if (header === "Files And Changes") {
+    return mergeFileLines(prev, fresh);
+  }
+
+  // Session Goal, User Preferences: line-level dedup, cap
+  const isClean = (l: string) => l.startsWith("- ") && !l.includes("<skill") && !l.includes("</skill");
+  const prevLines = prev.split("\n").filter(isClean);
+  const freshLines = fresh.split("\n").filter(isClean);
+  const combined = [...new Set([...prevLines, ...freshLines])];
+  const CAP = header === "Session Goal" ? 8 : header === "Commits" ? 8 : 15;
+  const capped = combined.length > CAP ? combined.slice(-CAP) : combined;
+  if (capped.length === 0) return "";
+  return `[${header}]\n${capped.join("\n")}`;
+};
+
+/** Merge Files And Changes by category, dedup paths across compactions */
+const mergeFileLines = (prev: string, fresh: string): string => {
+  const categories = ["Modified", "Created", "Read"] as const;
+  const merged: Record<string, Set<string>> = {};
+  for (const cat of categories) merged[cat] = new Set();
+
+  // Parse "- Modified: a, b, c (+N more)" lines from both prev and fresh
+  for (const text of [prev, fresh]) {
+    for (const line of text.split("\n")) {
+      for (const cat of categories) {
+        const prefix = `- ${cat}: `;
+        if (!line.startsWith(prefix)) continue;
+        let rest = line.slice(prefix.length);
+        // Strip "(+N more)" suffix
+        rest = rest.replace(/\s*\(\+\d+ more\)\s*$/, "");
+        for (const p of rest.split(",")) {
+          const trimmed = p.trim();
+          if (trimmed) merged[cat].add(trimmed);
+        }
+      }
+    }
+  }
+
+  // Dedup: if already in Modified, drop from Created (file existed before)
+  for (const p of merged.Modified) merged.Created.delete(p);
+
+  const cap = (set: Set<string>, limit: number) => {
+    const arr = [...set];
+    if (arr.length <= limit) return arr.join(", ");
+    return arr.slice(0, limit).join(", ") + ` (+${arr.length - limit} more)`;
+  };
+
+  const lines: string[] = [];
+  if (merged.Modified.size > 0) lines.push(`- Modified: ${cap(merged.Modified, 10)}`);
+  if (merged.Created.size > 0) lines.push(`- Created: ${cap(merged.Created, 10)}`);
+  if (merged.Read.size > 0) lines.push(`- Read: ${cap(merged.Read, 10)}`);
+  if (lines.length === 0) return "";
+  return `[Files And Changes]\n${lines.join("\n")}`;
+};
+
+const mergeBriefTranscript = (prev: string, fresh: string): string => {
+  if (!prev) return fresh;
+  if (!fresh) return prev;
+  return prev + "\n\n" + fresh;
+};
+
+const mergePrevious = (prev: string, fresh: string): string => {
+  // Merge header sections
+  const headers = HEADER_NAMES
+    .map((header) => {
+      const freshSec = sectionOf(fresh, header);
+      const prevSec = sectionOf(prev, header);
+      return mergeHeaderSection(header, prevSec, freshSec);
+    })
+    .filter(Boolean);
+
+  // Merge brief transcript
+  const prevBrief = briefOf(prev);
+  const freshBrief = briefOf(fresh);
+  const mergedBrief = mergeBriefTranscript(prevBrief, freshBrief);
+
+  const parts: string[] = [];
+  if (headers.length > 0) {
+    parts.push(headers.join("\n\n"));
+  }
+  if (mergedBrief) {
+    parts.push(capBrief(mergedBrief));
+  }
+
+  return parts.join(SEPARATOR);
+};
+
+export const compile = (input: CompileInput): string => {
+  const blocks = filterNoise(normalize(input.messages));
+  const data = buildSections({ blocks });
+  const fresh = formatSummary(data);
+  // Strip any legacy RECALL_NOTE baked into prev summary (pre-fix format)
+  // so merge doesn't re-stack it inside the brief.
+  const prev = input.previousSummary
+    ? stripRecallNote(input.previousSummary)
+    : undefined;
+  const merged = prev ? mergePrevious(prev, fresh) : fresh;
+  if (!merged) return "";
+  return merged + SEPARATOR + RECALL_NOTE;
+};
+
+const stripRecallNote = (text: string): string => {
+  // Remove trailing RECALL_NOTE (and any separators surrounding it) if present.
+  // Handles both current format (---\n\nNOTE) and bare trailing NOTE.
+  const idx = text.lastIndexOf(RECALL_NOTE);
+  if (idx < 0) return text;
+  return text.slice(0, idx).replace(/\s*(?:\n\n---\n\n)?\s*$/, "").trimEnd();
+};

package/vendor/pi-vcc/src/core/tool-args.ts

@@ -0,0 +1,14 @@
+export const extractPath = (args: Record<string, unknown>): string | null => {
+  for (const key of ["path", "file_path", "filePath", "file"]) {
+    if (typeof args[key] === "string") return args[key] as string;
+  }
+  return null;
+};
+
+export const summarizeToolArgs = (args: Record<string, unknown>): string => {
+  const path = extractPath(args);
+  if (path) return `path=${path}`;
+  if (typeof args.command === "string") return `command=${args.command}`;
+  if (typeof args.query === "string") return `query=${args.query}`;
+  return Object.keys(args).join(", ");
+};

package/vendor/pi-vcc/src/details.ts

@@ -0,0 +1,7 @@
+export interface PiVccCompactionDetails {
+  compactor: "pi-vcc" | "ultimate-pi-vcc";
+  version: number;
+  sections: string[];
+  sourceMessageCount: number;
+  previousSummaryUsed: boolean;
+}

package/vendor/pi-vcc/src/extract/commits.ts

@@ -0,0 +1,69 @@
+import type { NormalizedBlock } from "../types";
+
+interface CommitInfo {
+  hash?: string;
+  message: string;
+}
+
+const COMMIT_MSG_RE = /git\s+commit[^\n]*?-m\s+(?:"((?:[^"\\]|\\.)*)"|'((?:[^'\\]|\\.)*)'|\$?'((?:[^'\\]|\\.)*)')/;
+// Match short hash from git output: "[branch hash]" or "main hash" or 7-12 hex
+const HASH_RE = /\b([0-9a-f]{7,12})\b/;
+
+const firstLineOf = (text: string): string => {
+  const line = text.split(/\\n|\n/)[0] ?? "";
+  return line.trim();
+};
+
+const cleanMessage = (msg: string): string =>
+  msg.replace(/\\"/g, '"').replace(/\\'/g, "'").trim();
+
+/**
+ * Extract git commits from bash tool calls (`git commit -m "..."`) and pair
+ * with hash from the immediately following tool_result.
+ */
+export const extractCommits = (blocks: NormalizedBlock[]): CommitInfo[] => {
+  const commits: CommitInfo[] = [];
+
+  for (let i = 0; i < blocks.length; i++) {
+    const b = blocks[i];
+    if (b.kind !== "tool_call" || b.name !== "bash") continue;
+    const cmd = typeof b.args.command === "string" ? b.args.command : "";
+    if (!/\bgit\s+commit\b/.test(cmd)) continue;
+    const m = cmd.match(COMMIT_MSG_RE);
+    if (!m) continue;
+    const message = firstLineOf(cleanMessage(m[1] ?? m[2] ?? m[3] ?? ""));
+    if (!message) continue;
+
+    let hash: string | undefined;
+    // Look at next tool_result for hash
+    for (let j = i + 1; j < Math.min(blocks.length, i + 3); j++) {
+      const r = blocks[j];
+      if (r.kind !== "tool_result") continue;
+      // Common git commit output: `[branch <hash>] message` or `<branch> <hash>..<hash>`
+      const bracket = r.text.match(/\[\S+\s+([0-9a-f]{7,12})\]/);
+      if (bracket) { hash = bracket[1]; break; }
+      const range = r.text.match(/\b([0-9a-f]{7,12})\.\.([0-9a-f]{7,12})\b/);
+      if (range) { hash = range[2]; break; }
+      const plain = r.text.match(HASH_RE);
+      if (plain) { hash = plain[1]; break; }
+    }
+
+    // Dedup by message+hash
+    const key = `${hash ?? ""}::${message}`;
+    if (!commits.some((c) => `${c.hash ?? ""}::${c.message}` === key)) {
+      commits.push({ hash, message });
+    }
+  }
+
+  return commits;
+};
+
+export const formatCommits = (commits: CommitInfo[], limit = 8): string[] => {
+  const lines: string[] = [];
+  const items = commits.slice(-limit); // keep most recent
+  for (const c of items) {
+    const prefix = c.hash ? `${c.hash}: ` : "";
+    lines.push(`${prefix}${c.message}`);
+  }
+  return lines;
+};

package/vendor/pi-vcc/src/extract/files.ts

@@ -0,0 +1,80 @@
+import type { FileOps, NormalizedBlock } from "../types";
+import { extractPath } from "../core/tool-args";
+
+interface FileActivity {
+  read: Set<string>;
+  modified: Set<string>;
+  created: Set<string>;
+}
+
+const FILE_READ_TOOLS = new Set([
+  "Read", "read_file", "View",
+]);
+
+const FILE_WRITE_TOOLS = new Set([
+  "Edit", "Write", "edit", "write", "edit_file", "write_file",
+  "MultiEdit",
+]);
+
+const FILE_CREATE_TOOLS = new Set([
+  "Write", "write", "write_file",
+]);
+
+/**
+ * Find the longest common directory prefix among absolute paths.
+ * Returns "" if fewer than 2 absolute paths or no meaningful common prefix.
+ */
+const longestCommonDirPrefix = (paths: string[]): string => {
+  const abs = paths.filter((p) => p.startsWith("/"));
+  if (abs.length < 2) return "";
+  const split = abs.map((p) => p.split("/"));
+  const min = Math.min(...split.map((s) => s.length));
+  let i = 0;
+  while (i < min - 1) {
+    const seg = split[0][i];
+    if (!split.every((s) => s[i] === seg)) break;
+    i++;
+  }
+  if (i < 2) return ""; // require at least /a/b common
+  return split[0].slice(0, i).join("/") + "/";
+};
+
+const trimPaths = (set: Set<string>, prefix: string): Set<string> => {
+  if (!prefix) return set;
+  const out = new Set<string>();
+  for (const p of set) {
+    out.add(p.startsWith(prefix) ? p.slice(prefix.length) : p);
+  }
+  return out;
+};
+
+export const extractFiles = (
+  blocks: NormalizedBlock[],
+  fileOps?: FileOps,
+): FileActivity => {
+  const act: FileActivity = {
+    read: new Set(fileOps?.readFiles ?? []),
+    modified: new Set(fileOps?.modifiedFiles ?? []),
+    created: new Set(fileOps?.createdFiles ?? []),
+  };
+
+  for (const b of blocks) {
+    if (b.kind !== "tool_call") continue;
+    const p = extractPath(b.args);
+    if (!p) continue;
+
+    if (FILE_READ_TOOLS.has(b.name)) act.read.add(p);
+    if (FILE_WRITE_TOOLS.has(b.name)) act.modified.add(p);
+    if (FILE_CREATE_TOOLS.has(b.name)) act.created.add(p);
+  }
+
+  const all = [...act.read, ...act.modified, ...act.created];
+  const prefix = longestCommonDirPrefix(all);
+  if (prefix) {
+    act.read = trimPaths(act.read, prefix);
+    act.modified = trimPaths(act.modified, prefix);
+    act.created = trimPaths(act.created, prefix);
+  }
+
+  return act;
+};

package/vendor/pi-vcc/src/extract/goals.ts

@@ -0,0 +1,79 @@
+import type { NormalizedBlock } from "../types";
+import { nonEmptyLines, clip } from "../core/content";
+import { collapseSkillLines } from "../core/skill-collapse";
+
+const SCOPE_CHANGE_RE =
+  /\b(instead|actually|change of plan|forget that|new task|switch to|now I want|pivot|let'?s do|stop .* and)\b/i;
+
+const TASK_RE =
+  /\b(fix|implement|add|create|build|refactor|debug|investigate|update|remove|delete|migrate|deploy|test|write|set up)\b/i;
+
+const NOISE_SHORT_RE = /^(ok|yes|no|sure|yeah|yep|go|hi|hey|thx|thanks|ok\b.*|y|n|k)\s*[.!?]*$/i;
+
+// Reject lines that are clearly not user goals (pasted output, code, paths, tool dumps)
+// or meta-prompt boilerplate (command templates like `/issues` that start with "For each issue:"
+// followed by numbered "Read the issue in full..." steps).
+const NON_GOAL_RE =
+  /^\s*[\[│├└─╭╰]|```|^\s*(=[A-Z]+\(|function |const |let |var |import |export |class )|^(https?:|file:|\/[A-Za-z])|\\n|^\s*For each\b|\bin full\b[^\n]*\b(comments|issue|issues|PRs?|linked)\b/;
+
+// Signals that the rest of the user message is a command template (e.g. /issues),
+// in which case we should stop collecting goals at the signal line.
+const TEMPLATE_SIGNAL_RE =
+  /^\s*(For each\b|Do NOT implement\b|Analyze and propose\b|If Task\/context\b|Output:\s*$)/i;
+
+const truncateAtTemplate = (lines: string[]): string[] => {
+  const idx = lines.findIndex((l) => TEMPLATE_SIGNAL_RE.test(l));
+  return idx >= 0 ? lines.slice(0, idx) : lines;
+};
+
+const stripLeadingBullet = (line: string): string =>
+  line.replace(/^\s*(?:[-*+]|\d+\.)\s+/, "").trim();
+
+const MAX_GOAL_CHARS = 200;
+
+const isSubstantiveGoal = (text: string): boolean => {
+  const t = text.trim();
+  if (t.length <= 5) return false;
+  if (t.length > MAX_GOAL_CHARS) return false;
+  if (NOISE_SHORT_RE.test(t)) return false;
+  if (NON_GOAL_RE.test(t)) return false;
+  return true;
+};
+
+// Test scope-change / task intent only on the leading portion of a user block
+// so that pasted outputs below the actual instruction do not trigger matches.
+const LEADING_CHARS = 200;
+
+export const extractGoals = (blocks: NormalizedBlock[]): string[] => {
+  const goals: string[] = [];
+  let latestScopeChange: string[] | null = null;
+
+  for (const b of blocks) {
+    if (b.kind !== "user") continue;
+    const rawLines = nonEmptyLines(b.text);
+    const truncated = truncateAtTemplate(rawLines);
+    const lines = collapseSkillLines(truncated.filter(isSubstantiveGoal))
+      .map(stripLeadingBullet)
+      .filter((l) => l.length > 5);
+    if (lines.length === 0) continue;
+
+    if (goals.length === 0) {
+      goals.push(...lines.slice(0, 6));
+      continue;
+    }
+
+    const leading = b.text.slice(0, LEADING_CHARS);
+    if (SCOPE_CHANGE_RE.test(leading)) {
+      latestScopeChange = lines.slice(0, 3).map((l) => clip(l, MAX_GOAL_CHARS));
+    } else if (TASK_RE.test(leading) && lines[0].length > 15) {
+      latestScopeChange = lines.slice(0, 2).map((l) => clip(l, MAX_GOAL_CHARS));
+    }
+  }
+
+  // Only emit the [Scope change] marker when we actually captured bullets.
+  if (latestScopeChange && latestScopeChange.length > 0) {
+    goals.push("[Scope change]", ...latestScopeChange);
+  }
+
+  return goals.slice(0, 8);
+};