@caupulican/pi-adaptative 0.80.96 → 0.80.98

package/dist/core/research/workspace-collector.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"workspace-collector.d.ts","sourceRoot":"","sources":["../../../src/core/research/workspace-collector.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,KAAK,iBAAiB,EAAY,MAAM,oBAAoB,CAAC;AAEtE;+FAC+F;AAC/F,MAAM,MAAM,mBAAmB,GAAG,CACjC,OAAO,EAAE,MAAM,EACf,IAAI,EAAE,SAAS,MAAM,EAAE,EACvB,OAAO,EAAE;IAAE,GAAG,CAAC,EAAE,MAAM,CAAC;IAAC,OAAO,CAAC,EAAE,MAAM,CAAC;IAAC,SAAS,CAAC,EAAE,MAAM,CAAC;IAAC,QAAQ,CAAC,EAAE,MAAM,CAAC;IAAC,WAAW,CAAC,EAAE,OAAO,CAAA;CAAE,EACzG,QAAQ,EAAE,CAAC,KAAK,EAAE,iBAAiB,GAAG,IAAI,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,KAAK,IAAI,KAC/E,OAAO,CAAC;AAEb,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,0BAA0B,CAAC;AAuG5D,MAAM,WAAW,2BAA2B;IAC3C,8EAA8E;IAC9E,KAAK,EAAE,MAAM,CAAC;IACd,0FAA0F;IAC1F,GAAG,EAAE,MAAM,CAAC;IACZ,mEAAmE;IACnE,UAAU,EAAE,MAAM,CAAC;IACnB,yDAAyD;IACzD,UAAU,CAAC,EAAE,mBAAmB,CAAC;CACjC;AAED,0GAA0G;AAC1G,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,EAAE,CAazD;AA6FD,wBAAsB,uBAAuB,CAAC,IAAI,EAAE,2BAA2B,GAAG,OAAO,CAAC,WAAW,EAAE,CAAC,CAmGvG","sourcesContent":["import { type ExecFileException, execFile } from \"node:child_process\";\n\n/** Structural DI seam: only the callback overload the collector actually uses — demanding\n * node's full `typeof execFile` (with `__promisify__`) makes plain test mocks unassignable. */\nexport type WorkspaceExecFileFn = (\n\tcommand: string,\n\targs: readonly string[],\n\toptions: { cwd?: string; timeout?: number; maxBuffer?: number; encoding?: string; windowsHide?: boolean },\n\tcallback: (error: ExecFileException | null, stdout: string, stderr: string) => void,\n) => unknown;\n\nimport type { EvidenceRef } from \"../autonomy/contracts.ts\";\n\n/**\n * Best-effort workspace research source collector.\n *\n * Feeds the autonomous research lane POINTER-FIRST sources: a repo-relative path, a bounded excerpt,\n * and (when known) a line number — never whole file bodies. It runs `rg` under the session cwd exactly\n * like the grep tool does, so it only surfaces content ripgrep already matched. Collection is bounded\n * (a shared wall-clock deadline, a candidate cap, ripgrep's own binary/oversize skipping) and never\n * throws: if `rg` is missing or errors, it returns `[]`, which is today's \"no collector\" behavior.\n *\n * The returned sources are `EvidenceRef`s (the runner's source type) tagged `kind: \"workspace\"`.\n */\n\n/** Search terms shorter than this are too noisy to be useful discriminators. */\nconst MIN_TERM_LEN = 3;\n/** Cap on derived search terms; keeps the discovery pattern small and the collector cheap. */\nconst MAX_TERMS = 4;\n/** Pointer excerpts are bounded so we never spill a whole line (or a whole file) into the prompt. */\nconst EXCERPT_MAX_CHARS = 200;\n/** Shared wall-clock budget for the whole collection pass (both ripgrep calls together). */\nconst COLLECTION_BUDGET_MS = 5000;\n/** Floor for any single ripgrep call so a nearly-spent budget still gives ripgrep a chance to run. */\nconst MIN_CALL_MS = 500;\n/** ripgrep skips files larger than this; oversized files never contribute a source. */\nconst MAX_FILESIZE = \"1M\";\n/** Upper bound on candidate files carried from discovery into the pointer pass. */\nconst CANDIDATE_CAP = 24;\n/** Generous stdout ceiling; overflow degrades to fewer sources rather than throwing. */\nconst MAX_BUFFER_BYTES = 8 * 1024 * 1024;\n\n/**\n * A small, deliberately conservative English stopword set plus the structural words that show up in\n * goal/requirement text. Anything not here that is >= MIN_TERM_LEN survives as a search term.\n */\nconst STOPWORDS = new Set<string>([\n\t\"the\",\n\t\"and\",\n\t\"for\",\n\t\"with\",\n\t\"that\",\n\t\"this\",\n\t\"from\",\n\t\"into\",\n\t\"are\",\n\t\"was\",\n\t\"were\",\n\t\"but\",\n\t\"not\",\n\t\"you\",\n\t\"your\",\n\t\"our\",\n\t\"all\",\n\t\"any\",\n\t\"can\",\n\t\"has\",\n\t\"have\",\n\t\"had\",\n\t\"will\",\n\t\"would\",\n\t\"should\",\n\t\"could\",\n\t\"its\",\n\t\"his\",\n\t\"her\",\n\t\"their\",\n\t\"them\",\n\t\"they\",\n\t\"she\",\n\t\"him\",\n\t\"who\",\n\t\"what\",\n\t\"when\",\n\t\"where\",\n\t\"which\",\n\t\"how\",\n\t\"why\",\n\t\"use\",\n\t\"using\",\n\t\"used\",\n\t\"via\",\n\t\"per\",\n\t\"out\",\n\t\"off\",\n\t\"over\",\n\t\"under\",\n\t\"then\",\n\t\"than\",\n\t\"add\",\n\t\"adds\",\n\t\"get\",\n\t\"gets\",\n\t\"set\",\n\t\"sets\",\n\t\"new\",\n\t\"old\",\n\t\"one\",\n\t\"two\",\n\t\"let\",\n\t\"run\",\n\t\"runs\",\n]);\n\nexport interface CollectWorkspaceSourcesArgs {\n\t/** Free text (goal + requirement text) that search terms are derived from. */\n\tquery: string;\n\t/** Session working directory; ripgrep runs here and paths are reported relative to it. */\n\tcwd: string;\n\t/** Hard cap on returned sources; also the lane's source budget. */\n\tmaxSources: number;\n\t/** Injected for tests; defaults to node's `execFile`. */\n\texecFileFn?: WorkspaceExecFileFn;\n}\n\n/** Split on non-word runs, lowercase, drop stopwords/short/dupes, keep source order, cap at MAX_TERMS. */\nexport function deriveSearchTerms(query: string): string[] {\n\tconst seen = new Set<string>();\n\tconst terms: string[] = [];\n\tfor (const raw of query.split(/[^\\w]+/)) {\n\t\tconst term = raw.toLowerCase();\n\t\tif (term.length < MIN_TERM_LEN) continue;\n\t\tif (STOPWORDS.has(term)) continue;\n\t\tif (seen.has(term)) continue;\n\t\tseen.add(term);\n\t\tterms.push(term);\n\t\tif (terms.length >= MAX_TERMS) break;\n\t}\n\treturn terms;\n}\n\n/** The most specific term (longest wins; ties keep the earliest) drives the line-level pointer pass. */\nfunction pickBestTerm(terms: readonly string[]): string {\n\treturn terms.reduce((best, term) => (term.length > best.length ? term : best), terms[0]);\n}\n\nfunction truncateExcerpt(text: string): string | undefined {\n\tconst trimmed = text.trim();\n\tif (trimmed.length === 0) return undefined;\n\tif (trimmed.length <= EXCERPT_MAX_CHARS) return trimmed;\n\treturn `${trimmed.slice(0, EXCERPT_MAX_CHARS - 1)}…`;\n}\n\ninterface RgOutcome {\n\t/** True when ripgrep ran to a usable result (matches found, or a clean \"no matches\"). */\n\tok: boolean;\n\tstdout: string;\n\t/** True when the `rg` binary could not be spawned at all — the collector bails entirely. */\n\tmissing: boolean;\n}\n\nfunction runRg(\n\texecFileFn: WorkspaceExecFileFn,\n\targs: readonly string[],\n\tcwd: string,\n\ttimeoutMs: number,\n): Promise<RgOutcome> {\n\treturn new Promise((resolve) => {\n\t\tlet settled = false;\n\t\tconst done = (outcome: RgOutcome): void => {\n\t\t\tif (settled) return;\n\t\t\tsettled = true;\n\t\t\tresolve(outcome);\n\t\t};\n\t\ttry {\n\t\t\texecFileFn(\n\t\t\t\t\"rg\",\n\t\t\t\t[...args],\n\t\t\t\t{ cwd, timeout: timeoutMs, maxBuffer: MAX_BUFFER_BYTES, encoding: \"utf8\", windowsHide: true },\n\t\t\t\t(error: ExecFileException | null, stdout: string) => {\n\t\t\t\t\tconst out = typeof stdout === \"string\" ? stdout : \"\";\n\t\t\t\t\tif (!error) {\n\t\t\t\t\t\tdone({ ok: true, stdout: out, missing: false });\n\t\t\t\t\t\treturn;\n\t\t\t\t\t}\n\t\t\t\t\t// Exit code 1 is ripgrep's \"no matches\" — a clean, usable result, not a failure.\n\t\t\t\t\tif (error.code === 1) {\n\t\t\t\t\t\tdone({ ok: true, stdout: out, missing: false });\n\t\t\t\t\t\treturn;\n\t\t\t\t\t}\n\t\t\t\t\tdone({ ok: false, stdout: \"\", missing: error.code === \"ENOENT\" });\n\t\t\t\t},\n\t\t\t);\n\t\t} catch {\n\t\t\t// A synchronous spawn failure (e.g. rg entirely absent) is treated as \"missing\".\n\t\t\tdone({ ok: false, stdout: \"\", missing: true });\n\t\t}\n\t});\n}\n\n/** rg prints `./foo` when the search root is `.`; keep sources cleanly repo-relative. */\nfunction normalizePath(path: string): string {\n\treturn path.startsWith(\"./\") ? path.slice(2) : path;\n}\n\nfunction parseFileList(stdout: string): string[] {\n\tconst files: string[] = [];\n\tfor (const line of stdout.split(/\\r?\\n/)) {\n\t\tconst path = line.trim();\n\t\tif (path.length > 0) files.push(normalizePath(path));\n\t}\n\treturn files;\n}\n\ninterface ContentMatch {\n\tpath: string;\n\tline: number;\n\ttext: string;\n}\n\nfunction parseContentMatches(stdout: string): ContentMatch[] {\n\tconst matches: ContentMatch[] = [];\n\tfor (const raw of stdout.split(/\\r?\\n/)) {\n\t\tif (raw.length === 0) continue;\n\t\t// `-H -n --no-heading` yields `path:line:text`; text may itself contain colons.\n\t\tconst parsed = /^(.+?):(\\d+):(.*)$/.exec(raw);\n\t\tif (!parsed) continue;\n\t\tmatches.push({ path: normalizePath(parsed[1]), line: Number(parsed[2]), text: parsed[3] });\n\t}\n\treturn matches;\n}\n\nexport async function collectWorkspaceSources(args: CollectWorkspaceSourcesArgs): Promise<EvidenceRef[]> {\n\tconst { query, cwd, maxSources } = args;\n\tconst execFileFn = args.execFileFn ?? (execFile as unknown as WorkspaceExecFileFn);\n\tif (!cwd || maxSources <= 0) return [];\n\n\tconst terms = deriveSearchTerms(query);\n\tif (terms.length === 0) return [];\n\n\tconst deadline = Date.now() + COLLECTION_BUDGET_MS;\n\tconst remainingBudget = (): number => Math.max(MIN_CALL_MS, deadline - Date.now());\n\n\t// Phase 1 (discovery): which files match ANY term. `--max-count 1` stops at the first hit per file;\n\t// ripgrep skips binary and oversized files by default / via --max-filesize.\n\tconst discoveryArgs = [\n\t\t\"--files-with-matches\",\n\t\t\"--max-count\",\n\t\t\"1\",\n\t\t\"--fixed-strings\",\n\t\t\"--smart-case\",\n\t\t\"--no-messages\",\n\t\t\"--max-filesize\",\n\t\tMAX_FILESIZE,\n\t\t\"--color\",\n\t\t\"never\",\n\t\t...terms.flatMap((term) => [\"-e\", term]),\n\t\t// Explicit search root: execFile hands rg a piped stdin, and rg with no path argument would\n\t\t// read (and block on) that pipe instead of scanning the tree. \".\" keeps output repo-relative.\n\t\t\"--\",\n\t\t\".\",\n\t];\n\tconst discovery = await runRg(execFileFn, discoveryArgs, cwd, remainingBudget());\n\tif (discovery.missing || !discovery.ok) return [];\n\n\tconst candidateFiles = parseFileList(discovery.stdout).slice(0, CANDIDATE_CAP);\n\tif (candidateFiles.length === 0) return [];\n\n\t// Phase 2 (pointers): line-level hits for the single best term, scanned only over files discovery\n\t// already matched — so we never read a file ripgrep did not surface.\n\tconst bestTerm = pickBestTerm(terms);\n\tconst contentArgs = [\n\t\t\"-H\",\n\t\t\"-n\",\n\t\t\"--no-heading\",\n\t\t\"-m\",\n\t\t\"2\",\n\t\t\"--fixed-strings\",\n\t\t\"--smart-case\",\n\t\t\"--no-messages\",\n\t\t\"--color\",\n\t\t\"never\",\n\t\t\"-e\",\n\t\tbestTerm,\n\t\t\"--\",\n\t\t...candidateFiles,\n\t];\n\tconst content = await runRg(execFileFn, contentArgs, cwd, remainingBudget());\n\n\tconst sources: EvidenceRef[] = [];\n\tconst seenPaths = new Set<string>();\n\tconst seenLineKeys = new Set<string>();\n\tlet counter = 0;\n\n\tif (content.ok) {\n\t\tfor (const match of parseContentMatches(content.stdout)) {\n\t\t\tif (sources.length >= maxSources) break;\n\t\t\tconst key = `${match.path}:${match.line}`;\n\t\t\tif (seenLineKeys.has(key)) continue;\n\t\t\tseenLineKeys.add(key);\n\t\t\tseenPaths.add(match.path);\n\t\t\tconst excerpt = truncateExcerpt(match.text);\n\t\t\tsources.push({\n\t\t\t\tid: `ws-${++counter}`,\n\t\t\t\tkind: \"workspace\",\n\t\t\t\ttitle: `${match.path}:${match.line}`,\n\t\t\t\turi: match.path,\n\t\t\t\ttrusted: true,\n\t\t\t\t...(excerpt !== undefined ? { excerpt } : {}),\n\t\t\t\tmetadata: { line: match.line, term: bestTerm },\n\t\t\t});\n\t\t}\n\t}\n\n\t// Fill the remaining budget with file-level pointers for candidates that matched a secondary term\n\t// (and so produced no best-term line). Still pointer-first: a path, never a body.\n\tfor (const file of candidateFiles) {\n\t\tif (sources.length >= maxSources) break;\n\t\tif (seenPaths.has(file)) continue;\n\t\tseenPaths.add(file);\n\t\tsources.push({\n\t\t\tid: `ws-${++counter}`,\n\t\t\tkind: \"workspace\",\n\t\t\ttitle: file,\n\t\t\turi: file,\n\t\t\ttrusted: true,\n\t\t\tmetadata: { matchedBy: \"discovery\" },\n\t\t});\n\t}\n\n\treturn sources.slice(0, maxSources);\n}\n"]}

package/dist/core/research/workspace-collector.js

@@ -0,0 +1,286 @@
+import { execFile } from "node:child_process";
+/**
+ * Best-effort workspace research source collector.
+ *
+ * Feeds the autonomous research lane POINTER-FIRST sources: a repo-relative path, a bounded excerpt,
+ * and (when known) a line number — never whole file bodies. It runs `rg` under the session cwd exactly
+ * like the grep tool does, so it only surfaces content ripgrep already matched. Collection is bounded
+ * (a shared wall-clock deadline, a candidate cap, ripgrep's own binary/oversize skipping) and never
+ * throws: if `rg` is missing or errors, it returns `[]`, which is today's "no collector" behavior.
+ *
+ * The returned sources are `EvidenceRef`s (the runner's source type) tagged `kind: "workspace"`.
+ */
+/** Search terms shorter than this are too noisy to be useful discriminators. */
+const MIN_TERM_LEN = 3;
+/** Cap on derived search terms; keeps the discovery pattern small and the collector cheap. */
+const MAX_TERMS = 4;
+/** Pointer excerpts are bounded so we never spill a whole line (or a whole file) into the prompt. */
+const EXCERPT_MAX_CHARS = 200;
+/** Shared wall-clock budget for the whole collection pass (both ripgrep calls together). */
+const COLLECTION_BUDGET_MS = 5000;
+/** Floor for any single ripgrep call so a nearly-spent budget still gives ripgrep a chance to run. */
+const MIN_CALL_MS = 500;
+/** ripgrep skips files larger than this; oversized files never contribute a source. */
+const MAX_FILESIZE = "1M";
+/** Upper bound on candidate files carried from discovery into the pointer pass. */
+const CANDIDATE_CAP = 24;
+/** Generous stdout ceiling; overflow degrades to fewer sources rather than throwing. */
+const MAX_BUFFER_BYTES = 8 * 1024 * 1024;
+/**
+ * A small, deliberately conservative English stopword set plus the structural words that show up in
+ * goal/requirement text. Anything not here that is >= MIN_TERM_LEN survives as a search term.
+ */
+const STOPWORDS = new Set([
+    "the",
+    "and",
+    "for",
+    "with",
+    "that",
+    "this",
+    "from",
+    "into",
+    "are",
+    "was",
+    "were",
+    "but",
+    "not",
+    "you",
+    "your",
+    "our",
+    "all",
+    "any",
+    "can",
+    "has",
+    "have",
+    "had",
+    "will",
+    "would",
+    "should",
+    "could",
+    "its",
+    "his",
+    "her",
+    "their",
+    "them",
+    "they",
+    "she",
+    "him",
+    "who",
+    "what",
+    "when",
+    "where",
+    "which",
+    "how",
+    "why",
+    "use",
+    "using",
+    "used",
+    "via",
+    "per",
+    "out",
+    "off",
+    "over",
+    "under",
+    "then",
+    "than",
+    "add",
+    "adds",
+    "get",
+    "gets",
+    "set",
+    "sets",
+    "new",
+    "old",
+    "one",
+    "two",
+    "let",
+    "run",
+    "runs",
+]);
+/** Split on non-word runs, lowercase, drop stopwords/short/dupes, keep source order, cap at MAX_TERMS. */
+export function deriveSearchTerms(query) {
+    const seen = new Set();
+    const terms = [];
+    for (const raw of query.split(/[^\w]+/)) {
+        const term = raw.toLowerCase();
+        if (term.length < MIN_TERM_LEN)
+            continue;
+        if (STOPWORDS.has(term))
+            continue;
+        if (seen.has(term))
+            continue;
+        seen.add(term);
+        terms.push(term);
+        if (terms.length >= MAX_TERMS)
+            break;
+    }
+    return terms;
+}
+/** The most specific term (longest wins; ties keep the earliest) drives the line-level pointer pass. */
+function pickBestTerm(terms) {
+    return terms.reduce((best, term) => (term.length > best.length ? term : best), terms[0]);
+}
+function truncateExcerpt(text) {
+    const trimmed = text.trim();
+    if (trimmed.length === 0)
+        return undefined;
+    if (trimmed.length <= EXCERPT_MAX_CHARS)
+        return trimmed;
+    return `${trimmed.slice(0, EXCERPT_MAX_CHARS - 1)}…`;
+}
+function runRg(execFileFn, args, cwd, timeoutMs) {
+    return new Promise((resolve) => {
+        let settled = false;
+        const done = (outcome) => {
+            if (settled)
+                return;
+            settled = true;
+            resolve(outcome);
+        };
+        try {
+            execFileFn("rg", [...args], { cwd, timeout: timeoutMs, maxBuffer: MAX_BUFFER_BYTES, encoding: "utf8", windowsHide: true }, (error, stdout) => {
+                const out = typeof stdout === "string" ? stdout : "";
+                if (!error) {
+                    done({ ok: true, stdout: out, missing: false });
+                    return;
+                }
+                // Exit code 1 is ripgrep's "no matches" — a clean, usable result, not a failure.
+                if (error.code === 1) {
+                    done({ ok: true, stdout: out, missing: false });
+                    return;
+                }
+                done({ ok: false, stdout: "", missing: error.code === "ENOENT" });
+            });
+        }
+        catch {
+            // A synchronous spawn failure (e.g. rg entirely absent) is treated as "missing".
+            done({ ok: false, stdout: "", missing: true });
+        }
+    });
+}
+/** rg prints `./foo` when the search root is `.`; keep sources cleanly repo-relative. */
+function normalizePath(path) {
+    return path.startsWith("./") ? path.slice(2) : path;
+}
+function parseFileList(stdout) {
+    const files = [];
+    for (const line of stdout.split(/\r?\n/)) {
+        const path = line.trim();
+        if (path.length > 0)
+            files.push(normalizePath(path));
+    }
+    return files;
+}
+function parseContentMatches(stdout) {
+    const matches = [];
+    for (const raw of stdout.split(/\r?\n/)) {
+        if (raw.length === 0)
+            continue;
+        // `-H -n --no-heading` yields `path:line:text`; text may itself contain colons.
+        const parsed = /^(.+?):(\d+):(.*)$/.exec(raw);
+        if (!parsed)
+            continue;
+        matches.push({ path: normalizePath(parsed[1]), line: Number(parsed[2]), text: parsed[3] });
+    }
+    return matches;
+}
+export async function collectWorkspaceSources(args) {
+    const { query, cwd, maxSources } = args;
+    const execFileFn = args.execFileFn ?? execFile;
+    if (!cwd || maxSources <= 0)
+        return [];
+    const terms = deriveSearchTerms(query);
+    if (terms.length === 0)
+        return [];
+    const deadline = Date.now() + COLLECTION_BUDGET_MS;
+    const remainingBudget = () => Math.max(MIN_CALL_MS, deadline - Date.now());
+    // Phase 1 (discovery): which files match ANY term. `--max-count 1` stops at the first hit per file;
+    // ripgrep skips binary and oversized files by default / via --max-filesize.
+    const discoveryArgs = [
+        "--files-with-matches",
+        "--max-count",
+        "1",
+        "--fixed-strings",
+        "--smart-case",
+        "--no-messages",
+        "--max-filesize",
+        MAX_FILESIZE,
+        "--color",
+        "never",
+        ...terms.flatMap((term) => ["-e", term]),
+        // Explicit search root: execFile hands rg a piped stdin, and rg with no path argument would
+        // read (and block on) that pipe instead of scanning the tree. "." keeps output repo-relative.
+        "--",
+        ".",
+    ];
+    const discovery = await runRg(execFileFn, discoveryArgs, cwd, remainingBudget());
+    if (discovery.missing || !discovery.ok)
+        return [];
+    const candidateFiles = parseFileList(discovery.stdout).slice(0, CANDIDATE_CAP);
+    if (candidateFiles.length === 0)
+        return [];
+    // Phase 2 (pointers): line-level hits for the single best term, scanned only over files discovery
+    // already matched — so we never read a file ripgrep did not surface.
+    const bestTerm = pickBestTerm(terms);
+    const contentArgs = [
+        "-H",
+        "-n",
+        "--no-heading",
+        "-m",
+        "2",
+        "--fixed-strings",
+        "--smart-case",
+        "--no-messages",
+        "--color",
+        "never",
+        "-e",
+        bestTerm,
+        "--",
+        ...candidateFiles,
+    ];
+    const content = await runRg(execFileFn, contentArgs, cwd, remainingBudget());
+    const sources = [];
+    const seenPaths = new Set();
+    const seenLineKeys = new Set();
+    let counter = 0;
+    if (content.ok) {
+        for (const match of parseContentMatches(content.stdout)) {
+            if (sources.length >= maxSources)
+                break;
+            const key = `${match.path}:${match.line}`;
+            if (seenLineKeys.has(key))
+                continue;
+            seenLineKeys.add(key);
+            seenPaths.add(match.path);
+            const excerpt = truncateExcerpt(match.text);
+            sources.push({
+                id: `ws-${++counter}`,
+                kind: "workspace",
+                title: `${match.path}:${match.line}`,
+                uri: match.path,
+                trusted: true,
+                ...(excerpt !== undefined ? { excerpt } : {}),
+                metadata: { line: match.line, term: bestTerm },
+            });
+        }
+    }
+    // Fill the remaining budget with file-level pointers for candidates that matched a secondary term
+    // (and so produced no best-term line). Still pointer-first: a path, never a body.
+    for (const file of candidateFiles) {
+        if (sources.length >= maxSources)
+            break;
+        if (seenPaths.has(file))
+            continue;
+        seenPaths.add(file);
+        sources.push({
+            id: `ws-${++counter}`,
+            kind: "workspace",
+            title: file,
+            uri: file,
+            trusted: true,
+            metadata: { matchedBy: "discovery" },
+        });
+    }
+    return sources.slice(0, maxSources);
+}
+//# sourceMappingURL=workspace-collector.js.map

package/dist/core/research/workspace-collector.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"workspace-collector.js","sourceRoot":"","sources":["../../../src/core/research/workspace-collector.ts"],"names":[],"mappings":"AAAA,OAAO,EAA0B,QAAQ,EAAE,MAAM,oBAAoB,CAAC;AAatE;;;;;;;;;;GAUG;AAEH,gFAAgF;AAChF,MAAM,YAAY,GAAG,CAAC,CAAC;AACvB,8FAA8F;AAC9F,MAAM,SAAS,GAAG,CAAC,CAAC;AACpB,qGAAqG;AACrG,MAAM,iBAAiB,GAAG,GAAG,CAAC;AAC9B,4FAA4F;AAC5F,MAAM,oBAAoB,GAAG,IAAI,CAAC;AAClC,sGAAsG;AACtG,MAAM,WAAW,GAAG,GAAG,CAAC;AACxB,uFAAuF;AACvF,MAAM,YAAY,GAAG,IAAI,CAAC;AAC1B,mFAAmF;AACnF,MAAM,aAAa,GAAG,EAAE,CAAC;AACzB,wFAAwF;AACxF,MAAM,gBAAgB,GAAG,CAAC,GAAG,IAAI,GAAG,IAAI,CAAC;AAEzC;;;GAGG;AACH,MAAM,SAAS,GAAG,IAAI,GAAG,CAAS;IACjC,KAAK;IACL,KAAK;IACL,KAAK;IACL,MAAM;IACN,MAAM;IACN,MAAM;IACN,MAAM;IACN,MAAM;IACN,KAAK;IACL,KAAK;IACL,MAAM;IACN,KAAK;IACL,KAAK;IACL,KAAK;IACL,MAAM;IACN,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,MAAM;IACN,KAAK;IACL,MAAM;IACN,OAAO;IACP,QAAQ;IACR,OAAO;IACP,KAAK;IACL,KAAK;IACL,KAAK;IACL,OAAO;IACP,MAAM;IACN,MAAM;IACN,KAAK;IACL,KAAK;IACL,KAAK;IACL,MAAM;IACN,MAAM;IACN,OAAO;IACP,OAAO;IACP,KAAK;IACL,KAAK;IACL,KAAK;IACL,OAAO;IACP,MAAM;IACN,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,MAAM;IACN,OAAO;IACP,MAAM;IACN,MAAM;IACN,KAAK;IACL,MAAM;IACN,KAAK;IACL,MAAM;IACN,KAAK;IACL,MAAM;IACN,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,MAAM;CACN,CAAC,CAAC;AAaH,0GAA0G;AAC1G,MAAM,UAAU,iBAAiB,CAAC,KAAa,EAAY;IAC1D,MAAM,IAAI,GAAG,IAAI,GAAG,EAAU,CAAC;IAC/B,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,KAAK,MAAM,GAAG,IAAI,KAAK,CAAC,KAAK,CAAC,QAAQ,CAAC,EAAE,CAAC;QACzC,MAAM,IAAI,GAAG,GAAG,CAAC,WAAW,EAAE,CAAC;QAC/B,IAAI,IAAI,CAAC,MAAM,GAAG,YAAY;YAAE,SAAS;QACzC,IAAI,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC;YAAE,SAAS;QAClC,IAAI,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC;YAAE,SAAS;QAC7B,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;QACf,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACjB,IAAI,KAAK,CAAC,MAAM,IAAI,SAAS;YAAE,MAAM;IACtC,CAAC;IACD,OAAO,KAAK,CAAC;AAAA,CACb;AAED,wGAAwG;AACxG,SAAS,YAAY,CAAC,KAAwB,EAAU;IACvD,OAAO,KAAK,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,IAAI,EAAE,EAAE,CAAC,CAAC,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC;AAAA,CACzF;AAED,SAAS,eAAe,CAAC,IAAY,EAAsB;IAC1D,MAAM,OAAO,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;IAC5B,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,SAAS,CAAC;IAC3C,IAAI,OAAO,CAAC,MAAM,IAAI,iBAAiB;QAAE,OAAO,OAAO,CAAC;IACxD,OAAO,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,iBAAiB,GAAG,CAAC,CAAC,KAAG,CAAC;AAAA,CACrD;AAUD,SAAS,KAAK,CACb,UAA+B,EAC/B,IAAuB,EACvB,GAAW,EACX,SAAiB,EACI;IACrB,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC;QAC/B,IAAI,OAAO,GAAG,KAAK,CAAC;QACpB,MAAM,IAAI,GAAG,CAAC,OAAkB,EAAQ,EAAE,CAAC;YAC1C,IAAI,OAAO;gBAAE,OAAO;YACpB,OAAO,GAAG,IAAI,CAAC;YACf,OAAO,CAAC,OAAO,CAAC,CAAC;QAAA,CACjB,CAAC;QACF,IAAI,CAAC;YACJ,UAAU,CACT,IAAI,EACJ,CAAC,GAAG,IAAI,CAAC,EACT,EAAE,GAAG,EAAE,OAAO,EAAE,SAAS,EAAE,SAAS,EAAE,gBAAgB,EAAE,QAAQ,EAAE,MAAM,EAAE,WAAW,EAAE,IAAI,EAAE,EAC7F,CAAC,KAA+B,EAAE,MAAc,EAAE,EAAE,CAAC;gBACpD,MAAM,GAAG,GAAG,OAAO,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC;gBACrD,IAAI,CAAC,KAAK,EAAE,CAAC;oBACZ,IAAI,CAAC,EAAE,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,GAAG,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC,CAAC;oBAChD,OAAO;gBACR,CAAC;gBACD,mFAAiF;gBACjF,IAAI,KAAK,CAAC,IAAI,KAAK,CAAC,EAAE,CAAC;oBACtB,IAAI,CAAC,EAAE,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,GAAG,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC,CAAC;oBAChD,OAAO;gBACR,CAAC;gBACD,IAAI,CAAC,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,EAAE,EAAE,OAAO,EAAE,KAAK,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC,CAAC;YAAA,CAClE,CACD,CAAC;QACH,CAAC;QAAC,MAAM,CAAC;YACR,iFAAiF;YACjF,IAAI,CAAC,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,EAAE,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC;QAChD,CAAC;IAAA,CACD,CAAC,CAAC;AAAA,CACH;AAED,yFAAyF;AACzF,SAAS,aAAa,CAAC,IAAY,EAAU;IAC5C,OAAO,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC;AAAA,CACpD;AAED,SAAS,aAAa,CAAC,MAAc,EAAY;IAChD,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,KAAK,MAAM,IAAI,IAAI,MAAM,CAAC,KAAK,CAAC,OAAO,CAAC,EAAE,CAAC;QAC1C,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,EAAE,CAAC;QACzB,IAAI,IAAI,CAAC,MAAM,GAAG,CAAC;YAAE,KAAK,CAAC,IAAI,CAAC,aAAa,CAAC,IAAI,CAAC,CAAC,CAAC;IACtD,CAAC;IACD,OAAO,KAAK,CAAC;AAAA,CACb;AAQD,SAAS,mBAAmB,CAAC,MAAc,EAAkB;IAC5D,MAAM,OAAO,GAAmB,EAAE,CAAC;IACnC,KAAK,MAAM,GAAG,IAAI,MAAM,CAAC,KAAK,CAAC,OAAO,CAAC,EAAE,CAAC;QACzC,IAAI,GAAG,CAAC,MAAM,KAAK,CAAC;YAAE,SAAS;QAC/B,gFAAgF;QAChF,MAAM,MAAM,GAAG,oBAAoB,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;QAC9C,IAAI,CAAC,MAAM;YAAE,SAAS;QACtB,OAAO,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,aAAa,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;IAC5F,CAAC;IACD,OAAO,OAAO,CAAC;AAAA,CACf;AAED,MAAM,CAAC,KAAK,UAAU,uBAAuB,CAAC,IAAiC,EAA0B;IACxG,MAAM,EAAE,KAAK,EAAE,GAAG,EAAE,UAAU,EAAE,GAAG,IAAI,CAAC;IACxC,MAAM,UAAU,GAAG,IAAI,CAAC,UAAU,IAAK,QAA2C,CAAC;IACnF,IAAI,CAAC,GAAG,IAAI,UAAU,IAAI,CAAC;QAAE,OAAO,EAAE,CAAC;IAEvC,MAAM,KAAK,GAAG,iBAAiB,CAAC,KAAK,CAAC,CAAC;IACvC,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,EAAE,CAAC;IAElC,MAAM,QAAQ,GAAG,IAAI,CAAC,GAAG,EAAE,GAAG,oBAAoB,CAAC;IACnD,MAAM,eAAe,GAAG,GAAW,EAAE,CAAC,IAAI,CAAC,GAAG,CAAC,WAAW,EAAE,QAAQ,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC;IAEnF,oGAAoG;IACpG,4EAA4E;IAC5E,MAAM,aAAa,GAAG;QACrB,sBAAsB;QACtB,aAAa;QACb,GAAG;QACH,iBAAiB;QACjB,cAAc;QACd,eAAe;QACf,gBAAgB;QAChB,YAAY;QACZ,SAAS;QACT,OAAO;QACP,GAAG,KAAK,CAAC,OAAO,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;QACxC,4FAA4F;QAC5F,8FAA8F;QAC9F,IAAI;QACJ,GAAG;KACH,CAAC;IACF,MAAM,SAAS,GAAG,MAAM,KAAK,CAAC,UAAU,EAAE,aAAa,EAAE,GAAG,EAAE,eAAe,EAAE,CAAC,CAAC;IACjF,IAAI,SAAS,CAAC,OAAO,IAAI,CAAC,SAAS,CAAC,EAAE;QAAE,OAAO,EAAE,CAAC;IAElD,MAAM,cAAc,GAAG,aAAa,CAAC,SAAS,CAAC,MAAM,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,aAAa,CAAC,CAAC;IAC/E,IAAI,cAAc,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,EAAE,CAAC;IAE3C,kGAAkG;IAClG,uEAAqE;IACrE,MAAM,QAAQ,GAAG,YAAY,CAAC,KAAK,CAAC,CAAC;IACrC,MAAM,WAAW,GAAG;QACnB,IAAI;QACJ,IAAI;QACJ,cAAc;QACd,IAAI;QACJ,GAAG;QACH,iBAAiB;QACjB,cAAc;QACd,eAAe;QACf,SAAS;QACT,OAAO;QACP,IAAI;QACJ,QAAQ;QACR,IAAI;QACJ,GAAG,cAAc;KACjB,CAAC;IACF,MAAM,OAAO,GAAG,MAAM,KAAK,CAAC,UAAU,EAAE,WAAW,EAAE,GAAG,EAAE,eAAe,EAAE,CAAC,CAAC;IAE7E,MAAM,OAAO,GAAkB,EAAE,CAAC;IAClC,MAAM,SAAS,GAAG,IAAI,GAAG,EAAU,CAAC;IACpC,MAAM,YAAY,GAAG,IAAI,GAAG,EAAU,CAAC;IACvC,IAAI,OAAO,GAAG,CAAC,CAAC;IAEhB,IAAI,OAAO,CAAC,EAAE,EAAE,CAAC;QAChB,KAAK,MAAM,KAAK,IAAI,mBAAmB,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE,CAAC;YACzD,IAAI,OAAO,CAAC,MAAM,IAAI,UAAU;gBAAE,MAAM;YACxC,MAAM,GAAG,GAAG,GAAG,KAAK,CAAC,IAAI,IAAI,KAAK,CAAC,IAAI,EAAE,CAAC;YAC1C,IAAI,YAAY,CAAC,GAAG,CAAC,GAAG,CAAC;gBAAE,SAAS;YACpC,YAAY,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;YACtB,SAAS,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;YAC1B,MAAM,OAAO,GAAG,eAAe,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;YAC5C,OAAO,CAAC,IAAI,CAAC;gBACZ,EAAE,EAAE,MAAM,EAAE,OAAO,EAAE;gBACrB,IAAI,EAAE,WAAW;gBACjB,KAAK,EAAE,GAAG,KAAK,CAAC,IAAI,IAAI,KAAK,CAAC,IAAI,EAAE;gBACpC,GAAG,EAAE,KAAK,CAAC,IAAI;gBACf,OAAO,EAAE,IAAI;gBACb,GAAG,CAAC,OAAO,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBAC7C,QAAQ,EAAE,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,EAAE,IAAI,EAAE,QAAQ,EAAE;aAC9C,CAAC,CAAC;QACJ,CAAC;IACF,CAAC;IAED,kGAAkG;IAClG,kFAAkF;IAClF,KAAK,MAAM,IAAI,IAAI,cAAc,EAAE,CAAC;QACnC,IAAI,OAAO,CAAC,MAAM,IAAI,UAAU;YAAE,MAAM;QACxC,IAAI,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC;YAAE,SAAS;QAClC,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;QACpB,OAAO,CAAC,IAAI,CAAC;YACZ,EAAE,EAAE,MAAM,EAAE,OAAO,EAAE;YACrB,IAAI,EAAE,WAAW;YACjB,KAAK,EAAE,IAAI;YACX,GAAG,EAAE,IAAI;YACT,OAAO,EAAE,IAAI;YACb,QAAQ,EAAE,EAAE,SAAS,EAAE,WAAW,EAAE;SACpC,CAAC,CAAC;IACJ,CAAC;IAED,OAAO,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,UAAU,CAAC,CAAC;AAAA,CACpC","sourcesContent":["import { type ExecFileException, execFile } from \"node:child_process\";\n\n/** Structural DI seam: only the callback overload the collector actually uses — demanding\n * node's full `typeof execFile` (with `__promisify__`) makes plain test mocks unassignable. */\nexport type WorkspaceExecFileFn = (\n\tcommand: string,\n\targs: readonly string[],\n\toptions: { cwd?: string; timeout?: number; maxBuffer?: number; encoding?: string; windowsHide?: boolean },\n\tcallback: (error: ExecFileException | null, stdout: string, stderr: string) => void,\n) => unknown;\n\nimport type { EvidenceRef } from \"../autonomy/contracts.ts\";\n\n/**\n * Best-effort workspace research source collector.\n *\n * Feeds the autonomous research lane POINTER-FIRST sources: a repo-relative path, a bounded excerpt,\n * and (when known) a line number — never whole file bodies. It runs `rg` under the session cwd exactly\n * like the grep tool does, so it only surfaces content ripgrep already matched. Collection is bounded\n * (a shared wall-clock deadline, a candidate cap, ripgrep's own binary/oversize skipping) and never\n * throws: if `rg` is missing or errors, it returns `[]`, which is today's \"no collector\" behavior.\n *\n * The returned sources are `EvidenceRef`s (the runner's source type) tagged `kind: \"workspace\"`.\n */\n\n/** Search terms shorter than this are too noisy to be useful discriminators. */\nconst MIN_TERM_LEN = 3;\n/** Cap on derived search terms; keeps the discovery pattern small and the collector cheap. */\nconst MAX_TERMS = 4;\n/** Pointer excerpts are bounded so we never spill a whole line (or a whole file) into the prompt. */\nconst EXCERPT_MAX_CHARS = 200;\n/** Shared wall-clock budget for the whole collection pass (both ripgrep calls together). */\nconst COLLECTION_BUDGET_MS = 5000;\n/** Floor for any single ripgrep call so a nearly-spent budget still gives ripgrep a chance to run. */\nconst MIN_CALL_MS = 500;\n/** ripgrep skips files larger than this; oversized files never contribute a source. */\nconst MAX_FILESIZE = \"1M\";\n/** Upper bound on candidate files carried from discovery into the pointer pass. */\nconst CANDIDATE_CAP = 24;\n/** Generous stdout ceiling; overflow degrades to fewer sources rather than throwing. */\nconst MAX_BUFFER_BYTES = 8 * 1024 * 1024;\n\n/**\n * A small, deliberately conservative English stopword set plus the structural words that show up in\n * goal/requirement text. Anything not here that is >= MIN_TERM_LEN survives as a search term.\n */\nconst STOPWORDS = new Set<string>([\n\t\"the\",\n\t\"and\",\n\t\"for\",\n\t\"with\",\n\t\"that\",\n\t\"this\",\n\t\"from\",\n\t\"into\",\n\t\"are\",\n\t\"was\",\n\t\"were\",\n\t\"but\",\n\t\"not\",\n\t\"you\",\n\t\"your\",\n\t\"our\",\n\t\"all\",\n\t\"any\",\n\t\"can\",\n\t\"has\",\n\t\"have\",\n\t\"had\",\n\t\"will\",\n\t\"would\",\n\t\"should\",\n\t\"could\",\n\t\"its\",\n\t\"his\",\n\t\"her\",\n\t\"their\",\n\t\"them\",\n\t\"they\",\n\t\"she\",\n\t\"him\",\n\t\"who\",\n\t\"what\",\n\t\"when\",\n\t\"where\",\n\t\"which\",\n\t\"how\",\n\t\"why\",\n\t\"use\",\n\t\"using\",\n\t\"used\",\n\t\"via\",\n\t\"per\",\n\t\"out\",\n\t\"off\",\n\t\"over\",\n\t\"under\",\n\t\"then\",\n\t\"than\",\n\t\"add\",\n\t\"adds\",\n\t\"get\",\n\t\"gets\",\n\t\"set\",\n\t\"sets\",\n\t\"new\",\n\t\"old\",\n\t\"one\",\n\t\"two\",\n\t\"let\",\n\t\"run\",\n\t\"runs\",\n]);\n\nexport interface CollectWorkspaceSourcesArgs {\n\t/** Free text (goal + requirement text) that search terms are derived from. */\n\tquery: string;\n\t/** Session working directory; ripgrep runs here and paths are reported relative to it. */\n\tcwd: string;\n\t/** Hard cap on returned sources; also the lane's source budget. */\n\tmaxSources: number;\n\t/** Injected for tests; defaults to node's `execFile`. */\n\texecFileFn?: WorkspaceExecFileFn;\n}\n\n/** Split on non-word runs, lowercase, drop stopwords/short/dupes, keep source order, cap at MAX_TERMS. */\nexport function deriveSearchTerms(query: string): string[] {\n\tconst seen = new Set<string>();\n\tconst terms: string[] = [];\n\tfor (const raw of query.split(/[^\\w]+/)) {\n\t\tconst term = raw.toLowerCase();\n\t\tif (term.length < MIN_TERM_LEN) continue;\n\t\tif (STOPWORDS.has(term)) continue;\n\t\tif (seen.has(term)) continue;\n\t\tseen.add(term);\n\t\tterms.push(term);\n\t\tif (terms.length >= MAX_TERMS) break;\n\t}\n\treturn terms;\n}\n\n/** The most specific term (longest wins; ties keep the earliest) drives the line-level pointer pass. */\nfunction pickBestTerm(terms: readonly string[]): string {\n\treturn terms.reduce((best, term) => (term.length > best.length ? term : best), terms[0]);\n}\n\nfunction truncateExcerpt(text: string): string | undefined {\n\tconst trimmed = text.trim();\n\tif (trimmed.length === 0) return undefined;\n\tif (trimmed.length <= EXCERPT_MAX_CHARS) return trimmed;\n\treturn `${trimmed.slice(0, EXCERPT_MAX_CHARS - 1)}…`;\n}\n\ninterface RgOutcome {\n\t/** True when ripgrep ran to a usable result (matches found, or a clean \"no matches\"). */\n\tok: boolean;\n\tstdout: string;\n\t/** True when the `rg` binary could not be spawned at all — the collector bails entirely. */\n\tmissing: boolean;\n}\n\nfunction runRg(\n\texecFileFn: WorkspaceExecFileFn,\n\targs: readonly string[],\n\tcwd: string,\n\ttimeoutMs: number,\n): Promise<RgOutcome> {\n\treturn new Promise((resolve) => {\n\t\tlet settled = false;\n\t\tconst done = (outcome: RgOutcome): void => {\n\t\t\tif (settled) return;\n\t\t\tsettled = true;\n\t\t\tresolve(outcome);\n\t\t};\n\t\ttry {\n\t\t\texecFileFn(\n\t\t\t\t\"rg\",\n\t\t\t\t[...args],\n\t\t\t\t{ cwd, timeout: timeoutMs, maxBuffer: MAX_BUFFER_BYTES, encoding: \"utf8\", windowsHide: true },\n\t\t\t\t(error: ExecFileException | null, stdout: string) => {\n\t\t\t\t\tconst out = typeof stdout === \"string\" ? stdout : \"\";\n\t\t\t\t\tif (!error) {\n\t\t\t\t\t\tdone({ ok: true, stdout: out, missing: false });\n\t\t\t\t\t\treturn;\n\t\t\t\t\t}\n\t\t\t\t\t// Exit code 1 is ripgrep's \"no matches\" — a clean, usable result, not a failure.\n\t\t\t\t\tif (error.code === 1) {\n\t\t\t\t\t\tdone({ ok: true, stdout: out, missing: false });\n\t\t\t\t\t\treturn;\n\t\t\t\t\t}\n\t\t\t\t\tdone({ ok: false, stdout: \"\", missing: error.code === \"ENOENT\" });\n\t\t\t\t},\n\t\t\t);\n\t\t} catch {\n\t\t\t// A synchronous spawn failure (e.g. rg entirely absent) is treated as \"missing\".\n\t\t\tdone({ ok: false, stdout: \"\", missing: true });\n\t\t}\n\t});\n}\n\n/** rg prints `./foo` when the search root is `.`; keep sources cleanly repo-relative. */\nfunction normalizePath(path: string): string {\n\treturn path.startsWith(\"./\") ? path.slice(2) : path;\n}\n\nfunction parseFileList(stdout: string): string[] {\n\tconst files: string[] = [];\n\tfor (const line of stdout.split(/\\r?\\n/)) {\n\t\tconst path = line.trim();\n\t\tif (path.length > 0) files.push(normalizePath(path));\n\t}\n\treturn files;\n}\n\ninterface ContentMatch {\n\tpath: string;\n\tline: number;\n\ttext: string;\n}\n\nfunction parseContentMatches(stdout: string): ContentMatch[] {\n\tconst matches: ContentMatch[] = [];\n\tfor (const raw of stdout.split(/\\r?\\n/)) {\n\t\tif (raw.length === 0) continue;\n\t\t// `-H -n --no-heading` yields `path:line:text`; text may itself contain colons.\n\t\tconst parsed = /^(.+?):(\\d+):(.*)$/.exec(raw);\n\t\tif (!parsed) continue;\n\t\tmatches.push({ path: normalizePath(parsed[1]), line: Number(parsed[2]), text: parsed[3] });\n\t}\n\treturn matches;\n}\n\nexport async function collectWorkspaceSources(args: CollectWorkspaceSourcesArgs): Promise<EvidenceRef[]> {\n\tconst { query, cwd, maxSources } = args;\n\tconst execFileFn = args.execFileFn ?? (execFile as unknown as WorkspaceExecFileFn);\n\tif (!cwd || maxSources <= 0) return [];\n\n\tconst terms = deriveSearchTerms(query);\n\tif (terms.length === 0) return [];\n\n\tconst deadline = Date.now() + COLLECTION_BUDGET_MS;\n\tconst remainingBudget = (): number => Math.max(MIN_CALL_MS, deadline - Date.now());\n\n\t// Phase 1 (discovery): which files match ANY term. `--max-count 1` stops at the first hit per file;\n\t// ripgrep skips binary and oversized files by default / via --max-filesize.\n\tconst discoveryArgs = [\n\t\t\"--files-with-matches\",\n\t\t\"--max-count\",\n\t\t\"1\",\n\t\t\"--fixed-strings\",\n\t\t\"--smart-case\",\n\t\t\"--no-messages\",\n\t\t\"--max-filesize\",\n\t\tMAX_FILESIZE,\n\t\t\"--color\",\n\t\t\"never\",\n\t\t...terms.flatMap((term) => [\"-e\", term]),\n\t\t// Explicit search root: execFile hands rg a piped stdin, and rg with no path argument would\n\t\t// read (and block on) that pipe instead of scanning the tree. \".\" keeps output repo-relative.\n\t\t\"--\",\n\t\t\".\",\n\t];\n\tconst discovery = await runRg(execFileFn, discoveryArgs, cwd, remainingBudget());\n\tif (discovery.missing || !discovery.ok) return [];\n\n\tconst candidateFiles = parseFileList(discovery.stdout).slice(0, CANDIDATE_CAP);\n\tif (candidateFiles.length === 0) return [];\n\n\t// Phase 2 (pointers): line-level hits for the single best term, scanned only over files discovery\n\t// already matched — so we never read a file ripgrep did not surface.\n\tconst bestTerm = pickBestTerm(terms);\n\tconst contentArgs = [\n\t\t\"-H\",\n\t\t\"-n\",\n\t\t\"--no-heading\",\n\t\t\"-m\",\n\t\t\"2\",\n\t\t\"--fixed-strings\",\n\t\t\"--smart-case\",\n\t\t\"--no-messages\",\n\t\t\"--color\",\n\t\t\"never\",\n\t\t\"-e\",\n\t\tbestTerm,\n\t\t\"--\",\n\t\t...candidateFiles,\n\t];\n\tconst content = await runRg(execFileFn, contentArgs, cwd, remainingBudget());\n\n\tconst sources: EvidenceRef[] = [];\n\tconst seenPaths = new Set<string>();\n\tconst seenLineKeys = new Set<string>();\n\tlet counter = 0;\n\n\tif (content.ok) {\n\t\tfor (const match of parseContentMatches(content.stdout)) {\n\t\t\tif (sources.length >= maxSources) break;\n\t\t\tconst key = `${match.path}:${match.line}`;\n\t\t\tif (seenLineKeys.has(key)) continue;\n\t\t\tseenLineKeys.add(key);\n\t\t\tseenPaths.add(match.path);\n\t\t\tconst excerpt = truncateExcerpt(match.text);\n\t\t\tsources.push({\n\t\t\t\tid: `ws-${++counter}`,\n\t\t\t\tkind: \"workspace\",\n\t\t\t\ttitle: `${match.path}:${match.line}`,\n\t\t\t\turi: match.path,\n\t\t\t\ttrusted: true,\n\t\t\t\t...(excerpt !== undefined ? { excerpt } : {}),\n\t\t\t\tmetadata: { line: match.line, term: bestTerm },\n\t\t\t});\n\t\t}\n\t}\n\n\t// Fill the remaining budget with file-level pointers for candidates that matched a secondary term\n\t// (and so produced no best-term line). Still pointer-first: a path, never a body.\n\tfor (const file of candidateFiles) {\n\t\tif (sources.length >= maxSources) break;\n\t\tif (seenPaths.has(file)) continue;\n\t\tseenPaths.add(file);\n\t\tsources.push({\n\t\t\tid: `ws-${++counter}`,\n\t\t\tkind: \"workspace\",\n\t\t\ttitle: file,\n\t\t\turi: file,\n\t\t\ttrusted: true,\n\t\t\tmetadata: { matchedBy: \"discovery\" },\n\t\t});\n\t}\n\n\treturn sources.slice(0, maxSources);\n}\n"]}

package/dist/core/settings-manager.d.ts

@@ -141,6 +141,7 @@ export interface ModelRouterSettings {
     learningModel?: string;
     judgeEnabled?: boolean;
     judgeModel?: string;
+    executorModel?: string;
 }
 export declare const DEFAULT_RESEARCH_LANE_ENABLED = false;
 export declare const DEFAULT_RESEARCH_LANE_MAX_USD = 0.25;
@@ -524,6 +525,7 @@ export declare class SettingsManager {
         learningModel?: string;
         judgeEnabled: boolean;
         judgeModel?: string;
+        executorModel?: string;
     };
     setModelRouterSettings(settings: ModelRouterSettings, scope?: SettingsScope): void;
     /** Configured auxiliary summarizer model id, or "auto" (default) to pick the cheapest authed model. */