@octocodeai/freellama-mcp-server 0.9.0 → 0.10.0

package/dist/index.js

@@ -25057,6 +25057,324 @@ var responseStore = new ResponseStore(
   envNumber("FREELLAMA_RESPONSE_TTL_DAYS", 7)
 );
 
+// src/handoffs.ts
+init_config();
+import { lstatSync, mkdirSync, readdirSync, realpathSync as realpathSync3, rmSync, statSync as statSync2, writeFileSync } from "node:fs";
+import path5 from "node:path";
+
+// src/task-files.ts
+import { globSync, realpathSync as realpathSync2, statSync, readFileSync } from "node:fs";
+import { homedir as homedir2, tmpdir } from "node:os";
+import path4 from "node:path";
+var MAX_TEXT_FILE_BYTES = 4 * 1024 * 1024;
+var MAX_IMAGE_FILE_BYTES = 8 * 1024 * 1024;
+var MAX_TOTAL_FILE_BYTES = 32 * 1024 * 1024;
+var MAX_FILES = 2e3;
+var IMAGE_EXTENSIONS = /* @__PURE__ */ new Set([".png", ".jpg", ".jpeg", ".webp", ".gif"]);
+var AUDIO_EXTENSIONS = /* @__PURE__ */ new Set([".wav"]);
+var EMBEDDING_CHUNK_BYTES = 1536;
+var CHUNK_OVERLAP_LINES = 8;
+function chunkLines(text3, limit = EMBEDDING_CHUNK_BYTES) {
+  const lines = text3.split("\n");
+  const chunks = [];
+  let current = [];
+  let bytes = 0;
+  let start2 = 1;
+  const flush = (end, overlap = 0) => {
+    if (current.length > 0 && current.join("\n").trim()) chunks.push({ text: current.join("\n"), lines: [start2, end] });
+    const carried = overlap > 0 ? current.slice(-overlap) : [];
+    const carriedBytes = carried.reduce((sum, line2) => sum + Buffer.byteLength(line2) + 1, 0);
+    current = carriedBytes < limit / 4 ? carried : [];
+    bytes = current === carried ? carriedBytes : 0;
+    start2 = end + 1 - current.length;
+  };
+  lines.forEach((line2, index) => {
+    const size2 = Buffer.byteLength(line2) + 1;
+    if (bytes + size2 > limit && current.length > 0) flush(index, CHUNK_OVERLAP_LINES);
+    if (size2 > limit) {
+      for (let at = 0; at < line2.length; at += Math.floor(limit / 2)) {
+        chunks.push({ text: line2.slice(at, at + Math.floor(limit / 2)), lines: [index + 1, index + 1] });
+      }
+      start2 = index + 2;
+      return;
+    }
+    current.push(line2);
+    bytes += size2;
+  });
+  flush(lines.length);
+  return chunks;
+}
+function fileRoots(env = process.env, cwd = process.cwd(), home = homedir2()) {
+  const configured = (env.FREELLAMA_FILE_ROOTS ?? "").split(path4.delimiter).filter(Boolean);
+  const broad = /* @__PURE__ */ new Set([path4.parse(path4.resolve(cwd)).root, resolvedRoot(home)]);
+  const project = broad.has(resolvedRoot(cwd)) ? [] : [cwd];
+  const handoffs = path4.join(env.FREELLAMA_HOME || path4.join(home, ".octocode", "freellama"), "handoffs");
+  const outputs = path4.join(env.FREELLAMA_HOME || path4.join(home, ".octocode", "freellama"), "tmp");
+  return [...project, tmpdir(), ...configured, handoffs, outputs].map(resolvedRoot);
+}
+function resolvedRoot(root) {
+  try {
+    return realpathSync2(root);
+  } catch {
+    const absolute = path4.resolve(root);
+    const parent = path4.dirname(absolute);
+    return parent === absolute ? absolute : path4.join(resolvedRoot(parent), path4.basename(absolute));
+  }
+}
+function within(file, root) {
+  const relative = path4.relative(root, file);
+  return relative === "" || !relative.startsWith("..") && !path4.isAbsolute(relative);
+}
+function isWav(bytes) {
+  return bytes.length >= 12 && bytes.subarray(0, 4).toString() === "RIFF" && bytes.subarray(8, 12).toString() === "WAVE";
+}
+function expandFileEntries(entries, cwd = process.cwd()) {
+  const out = [];
+  for (const entry of entries) {
+    if (!/[*?[]/.test(entry)) {
+      out.push(entry);
+      continue;
+    }
+    const matches = globSync(entry, { cwd }).map((match) => path4.resolve(cwd, match)).filter((match) => statSync(match).isFile()).sort();
+    if (matches.length === 0) throw new Error(`no files match ${entry} (globs resolve from ${cwd})`);
+    out.push(...matches);
+    if (out.length > MAX_FILES) break;
+  }
+  return out;
+}
+function readFiles(entries, givenRoots) {
+  const roots = givenRoots.map(resolvedRoot);
+  const paths = expandFileEntries(entries);
+  if (paths.length > MAX_FILES) throw new Error(`files names ${paths.length} paths; at most ${MAX_FILES} per call`);
+  let total = 0;
+  return paths.map((name) => {
+    const absolute = path4.resolve(name);
+    let real;
+    try {
+      real = realpathSync2(absolute);
+    } catch {
+      throw new Error(`file not found: ${name}`);
+    }
+    if (!roots.some((root) => within(real, root))) {
+      throw new Error(
+        `file ${name} is outside the readable roots (${roots.join(", ")}); copy it into the project or temp directory, or add its directory to FREELLAMA_FILE_ROOTS`
+      );
+    }
+    const stat3 = statSync(real);
+    if (!stat3.isFile()) throw new Error(`not a regular file: ${name}`);
+    const extension = path4.extname(real).toLowerCase();
+    let media = AUDIO_EXTENSIONS.has(extension) ? "audio" : IMAGE_EXTENSIONS.has(extension) ? "image" : void 0;
+    const limit = media ? MAX_IMAGE_FILE_BYTES : MAX_TEXT_FILE_BYTES;
+    if (stat3.size > limit) {
+      throw new Error(`file ${name} is ${stat3.size} bytes; the limit is ${limit} for ${media ?? "text"}; split it or send the part that matters`);
+    }
+    total += stat3.size;
+    if (total > MAX_TOTAL_FILE_BYTES) throw new Error(`files total more than ${MAX_TOTAL_FILE_BYTES} bytes; split the call`);
+    const bytes = readFileSync(real);
+    if (media === "audio" && !isWav(bytes)) {
+      throw new Error(`audio file ${name} must contain a WAV RIFF/WAVE header`);
+    }
+    if (isWav(bytes)) media = "audio";
+    return { path: name, bytes, media };
+  });
+}
+function shortPath(name) {
+  const relative = path4.relative(process.cwd(), path4.resolve(name));
+  return relative && !relative.startsWith("..") && !path4.isAbsolute(relative) ? relative : name;
+}
+function leadingComment(body) {
+  const lines = [];
+  for (const line2 of body.split("\n").slice(0, 20)) {
+    const trimmed = line2.trim();
+    if (!/^(\/\/|#(?!include)|\/\*|\*|"""|\s*$)/.test(trimmed) || trimmed.startsWith("#!")) break;
+    if (trimmed) lines.push(trimmed.replace(/^(\/\/[!/]?|#+|\/\*+|\*\/?|""")\s?/, ""));
+  }
+  return lines.join(" ").slice(0, 400);
+}
+var MAX_AUDIO_FILE_BYTES = 256 * 1024 * 1024;
+function readAllowedFile(name, limit, roots = fileRoots()) {
+  const allowed = roots.map(resolvedRoot);
+  let real;
+  try {
+    real = realpathSync2(path4.resolve(name));
+  } catch {
+    throw new Error(`file not found: ${name}`);
+  }
+  if (!allowed.some((root) => within(real, root))) {
+    throw new Error(`file ${name} is outside the readable roots (${allowed.join(", ")}); copy it into the project or temp directory, or add its directory to FREELLAMA_FILE_ROOTS`);
+  }
+  const stat3 = statSync(real);
+  if (!stat3.isFile()) throw new Error(`not a regular file: ${name}`);
+  if (stat3.size > limit) throw new Error(`file ${name} is ${stat3.size} bytes; the limit is ${limit}`);
+  return readFileSync(real);
+}
+function text(file) {
+  if (file.bytes.includes(0)) throw new Error(`file ${file.path} is binary; only images (${[...IMAGE_EXTENSIONS].join(", ")}) and text files can be read`);
+  return file.bytes.toString("utf8");
+}
+function applyTaskFiles(payload, paths, roots = fileRoots()) {
+  if (paths === void 0 || paths.length === 0) return { payload, files: [] };
+  const read = readFiles(paths, roots);
+  const receipts = [];
+  const media = read.filter((file) => file.media !== void 0);
+  const texts = read.filter((file) => file.media === void 0);
+  const next = { ...payload };
+  const task = payload.task ?? "completion";
+  if (task === "embedding") {
+    const given = payload.input === void 0 ? [] : Array.isArray(payload.input) ? payload.input : [payload.input];
+    const prefix = payload.documentPrefix ?? "";
+    const items = [...given];
+    for (const file of read) {
+      if (file.media) {
+        const encoded = file.bytes.toString("base64");
+        items.push(file.media === "audio" ? { audio: encoded } : { image: encoded });
+        receipts.push({ path: file.path, bytes: file.bytes.length, role: "input" });
+        continue;
+      }
+      const body = text(file);
+      const header = `${shortPath(file.path)}
+${leadingComment(body)}`.trimEnd();
+      for (const chunk of chunkLines(body)) {
+        items.push(`${prefix}${header}
+${chunk.text}`);
+        receipts.push({ path: file.path, bytes: Buffer.byteLength(chunk.text), role: "input", lines: chunk.lines });
+      }
+    }
+    next.input = items;
+    return { payload: next, files: receipts };
+  }
+  const audio = media.find((file) => file.media === "audio");
+  if (audio && task !== "decision") throw new Error(`audio file ${audio.path} is read only by tasks "decision" and "embedding" (models with audio input)`);
+  if (media.length > 0) {
+    next.images = [...payload.images ?? [], ...media.map((file) => file.bytes.toString("base64"))];
+    for (const file of media) receipts.push({ path: file.path, bytes: file.bytes.length, role: file.media });
+  }
+  if (task === "decision") {
+    if (texts.length > 0 && payload.state !== void 0) throw new Error("give the decision state either as `state` or as text `files`, not both");
+    if (texts.length === 1) {
+      const body = text(texts[0]);
+      next.state = jsonOrText(body);
+    } else if (texts.length > 1) {
+      next.state = Object.fromEntries(texts.map((file) => [file.path, text(file)]));
+    }
+    for (const file of texts) receipts.push({ path: file.path, bytes: file.bytes.length, role: "state" });
+    return { payload: next, files: receipts };
+  }
+  if (texts.length > 0) {
+    if (payload.prompt === void 0) throw new Error("text `files` are appended to `prompt`; give the instructions in `prompt`");
+    const sections = texts.map((file) => `--- file: ${file.path} ---
+${text(file)}`);
+    next.prompt = [payload.prompt, ...sections].join("\n\n");
+    for (const file of texts) receipts.push({ path: file.path, bytes: file.bytes.length, role: "prompt" });
+  }
+  return { payload: next, files: receipts };
+}
+function jsonOrText(body) {
+  try {
+    const value = JSON.parse(body);
+    if (value !== null && typeof value === "object") return value;
+  } catch {
+  }
+  return body;
+}
+function inputLabels(inputCount, files) {
+  const fromFiles = files.filter((file) => file.role === "input").map((file) => file.lines ? `${shortPath(file.path)}:L${file.lines[0]}-${file.lines[1]}` : shortPath(file.path));
+  const given = inputCount - fromFiles.length;
+  return [...Array.from({ length: given }, (_, index) => `input[${index}]`), ...fromFiles];
+}
+
+// src/handoffs.ts
+var HANDOFF_DIR = path5.join(FREELLAMA_HOME, "handoffs");
+var HANDOFF_NAME = /^[a-z0-9-]{1,60}$/;
+function summaryOf(text3, lines = 5) {
+  const kept = text3.split("\n").filter((line2) => line2.trim()).slice(0, lines).join("\n");
+  return kept.length > 600 ? `${kept.slice(0, 600)}\u2026` : kept;
+}
+function handoffBody(title, content) {
+  return `# ${title}
+
+${content.source}, ${(/* @__PURE__ */ new Date()).toISOString()}.
+
+## Summary
+
+${content.summary.trim() || "(empty)"}
+
+## Result
+
+${content.result.trim() || "(empty)"}
+`;
+}
+function writeHandoff(name, content, dir = HANDOFF_DIR) {
+  if (!HANDOFF_NAME.test(name)) throw new Error("saveAs is 1 to 60 characters of a-z, 0-9 and -");
+  mkdirSync(dir, { recursive: true, mode: 448 });
+  const file = path5.join(dir, `${name}.md`);
+  writeFileSync(file, handoffBody(name, content), { mode: 384 });
+  return file;
+}
+function resolveThroughSymlinks(file) {
+  try {
+    return realpathSync3(file);
+  } catch {
+    const parent = path5.dirname(file);
+    return parent === file ? file : path5.join(resolveThroughSymlinks(parent), path5.basename(file));
+  }
+}
+function checkOutputPath(name, roots = fileRoots()) {
+  const given = path5.resolve(name);
+  const file = resolveThroughSymlinks(given);
+  if ([given, file].some((candidate) => candidate.split(path5.sep).includes(".git"))) {
+    throw new Error(`outputPath ${name} is inside a .git directory; FreeLlama does not write there`);
+  }
+  if (!roots.some((root) => within(file, root))) {
+    throw new Error(`outputPath ${name} is outside the writable roots (${roots.join(", ")}); use a path inside the project or temp directory, or add its directory to FREELLAMA_FILE_ROOTS`);
+  }
+  let exists2 = true;
+  try {
+    lstatSync(file);
+  } catch {
+    exists2 = false;
+  }
+  if (exists2 && !statSync2(file).isFile()) throw new Error(`outputPath ${name} exists and is not a regular file`);
+  return file;
+}
+function writeOutputFile(name, content, roots = fileRoots()) {
+  const file = checkOutputPath(name, roots);
+  mkdirSync(path5.dirname(file), { recursive: true, mode: 448 });
+  writeFileSync(file, handoffBody(path5.basename(file), content), { mode: 384 });
+  return file;
+}
+var OUTPUT_TMP_DIR = path5.join(FREELLAMA_HOME, "tmp");
+var DAY_MS2 = 864e5;
+function outputTtlDays(env = process.env) {
+  const configured = Number(env.FREELLAMA_RESPONSE_TTL_DAYS ?? 7);
+  return Number.isFinite(configured) && configured >= 0 ? configured : 7;
+}
+function pruneOutputDir(dir, now, ttlDays) {
+  if (ttlDays <= 0) return;
+  let names;
+  try {
+    names = readdirSync(dir);
+  } catch {
+    return;
+  }
+  for (const name of names.filter((entry) => entry.endsWith(".md"))) {
+    const file = path5.join(dir, name);
+    try {
+      if (statSync2(file).mtimeMs < now - ttlDays * DAY_MS2) rmSync(file);
+    } catch {
+    }
+  }
+}
+var outputCounter = 0;
+function writeDefaultOutput(content, dir = OUTPUT_TMP_DIR, now = /* @__PURE__ */ new Date(), ttlDays = outputTtlDays()) {
+  mkdirSync(dir, { recursive: true, mode: 448 });
+  pruneOutputDir(dir, now.getTime(), ttlDays);
+  outputCounter += 1;
+  const file = path5.join(dir, `${now.toISOString().replace(/[:.]/g, "-")}-${process.pid}-${outputCounter}.md`);
+  writeFileSync(file, handoffBody(path5.basename(file), content), { mode: 384, flag: "wx" });
+  return file;
+}
+
 // src/tool-results.ts
 var nsToMs = (value) => typeof value === "number" ? Math.round(value / 1e6) : void 0;
 var object4 = (value) => value !== null && typeof value === "object" && !Array.isArray(value) ? value : void 0;
@@ -25113,8 +25431,6 @@ function compactReceipt(receipt) {
     "selected_model",
     "response",
     "output_check",
-    "metrics",
-    "telemetry",
     "scope",
     "warm",
     "job",
@@ -25131,13 +25447,15 @@ function compactReceipt(receipt) {
   }
   const response = withoutThinking(object4(receipt.response));
   if (response) compact.response = response;
+  const metrics = object4(receipt.metrics);
+  if (metrics) compact.metrics = Object.fromEntries(Object.entries(metrics).filter(([, value]) => value === null || typeof value !== "object"));
   const route3 = pick2(object4(receipt.route), ["selected_model", "task", "confidence", "reasons", "output_reserve_tokens"]);
   if (route3) compact.route = route3;
   const admission = pick2(object4(receipt.admission), ["mode", "queue_wait_ms", "transition_wait_ms", "resource_wait_ms", "total_wait_ms"]);
   if (admission) compact.admission = admission;
   if (execution) {
     compact.execution = {
-      ...pick2(execution, ["placement", "backend", "reason", "observation", "lifecycle", "context_sizing", "keep_alive", "preference_satisfied"]),
+      ...pick2(execution, ["placement", "backend", "reason", "observation", "keep_alive", "preference_satisfied"]),
       ...completionPlanning ? { completion_planning: completionPlanning } : {},
       ...reservation ? { memory_reservation: reservation } : {}
     };
@@ -25178,8 +25496,19 @@ function compactResearchResult(content) {
   return { shaped, sections };
 }
 var UNSAVED_TOOLS = /* @__PURE__ */ new Set(["read_response"]);
+function finishedJob(tool, args, content) {
+  if (tool !== "task_jobs" || args?.action !== "get") return void 0;
+  const job = object4(content.job);
+  const result = object4(job?.result);
+  return job && result ? { job, result } : void 0;
+}
 function shapeResult(tool, args, content, isError, saved) {
   const full = args?.receipt === "full" || args?.preview === true || !saved;
+  const finished = finishedJob(tool, args, content);
+  if (finished) {
+    if (full || args?.returnEmbeddings === true) return { shaped: content, sections: [] };
+    return { shaped: { ...content, job: { ...finished.job, result: compactTaskResult(finished.result) } }, sections: ["receipt"] };
+  }
   if (tool === "delegate_research") return full ? { shaped: content, sections: [] } : compactResearchResult(content);
   if (tool !== "run_task" || isError || content.deferred === true) return { shaped: content, sections: [] };
   let shaped = full ? content : compactTaskResult(content);
@@ -25194,6 +25523,72 @@ function shapeResult(tool, args, content, isError, saved) {
   }
   return { shaped, sections };
 }
+var AUTO_INLINE_CHARS = 8e3;
+var outputInput = {
+  output: _enum(["auto", "inline", "file"]).optional().describe("auto (default): inline up to 8,000 chars, else a summary plus check lines and where to read the rest; inline: always whole; file: always the summary view"),
+  outputPath: string2().min(1).optional().describe("also write the answer as Markdown here, creating or replacing it (inside the readable roots); default when shortened: FreeLlama home tmp/<timestamp>.md")
+};
+var SAVED_TEXT_SECTION = "/result/content/0/text";
+var KEPT_LINE = /^(check|handoff_file|verification|cited): /;
+var LABEL = /^\[[^\]\n]*\] /;
+function shortView(text3, where) {
+  const label = LABEL.exec(text3)?.[0].trim();
+  const lines = text3.replace(LABEL, "").split("\n");
+  const kept = lines.filter((line2) => KEPT_LINE.test(line2));
+  const summary = summaryOf(lines.filter((line2) => !KEPT_LINE.test(line2)).join("\n"));
+  return [label, summary, ...kept, where].filter((part) => part !== void 0 && part !== "").join("\n");
+}
+function viewText(args, text3, responseFile) {
+  const mode = typeof args?.output === "string" ? args.output : "auto";
+  const shorten = mode === "file" || mode === "auto" && text3.length > AUTO_INLINE_CHARS;
+  const answer = text3.replace(LABEL, "");
+  const content = { source: "From FreeLlama", summary: summaryOf(answer), result: answer };
+  let written;
+  const notes = [];
+  try {
+    if (typeof args?.outputPath === "string") written = writeOutputFile(args.outputPath, content);
+    else if (shorten) written = writeDefaultOutput(content);
+  } catch (error2) {
+    notes.push(`output_file not written: ${error2 instanceof Error ? error2.message : String(error2)}`);
+  }
+  const file = written ?? (shorten ? responseFile : void 0);
+  const tail = notes.length > 0 ? `
+${notes.join("\n")}` : "";
+  if (!shorten) {
+    return { text: `${text3}${written ? `
+output_file: ${written}` : ""}${tail}`, ...written ? { output_file: written } : {} };
+  }
+  if (!file) return { text: `${text3}${tail}
+not shortened (${text3.length} chars): no output file could be written and the response store is off, so this text is the only copy` };
+  const where = written ? `output_file: ${written} (the full answer, ${text3.length} chars)` : `full answer (${text3.length} chars): read_response{file:"${responseFile}", section:"${SAVED_TEXT_SECTION}"}`;
+  return {
+    text: `${shortView(text3, where)}${tail}`,
+    output: { mode, chars: text3.length, file },
+    ...written ? { output_file: written } : {},
+    where,
+    answer: text3
+  };
+}
+var MIN_COPY_CHARS = 400;
+function withoutAnswerCopies(value, answer, pointer2) {
+  if (typeof value === "string") return value.length >= MIN_COPY_CHARS && answer.includes(value.trim()) ? pointer2 : value;
+  if (Array.isArray(value)) return value.map((item) => withoutAnswerCopies(item, answer, pointer2));
+  const fields = object4(value);
+  return fields ? Object.fromEntries(Object.entries(fields).map(([key, item]) => [key, withoutAnswerCopies(item, answer, pointer2)])) : value;
+}
+function reshapedContent(result, args, responseFile) {
+  const parts = result.content;
+  const first = Array.isArray(parts) ? object4(parts[0]) : void 0;
+  if (!Array.isArray(parts) || first?.type !== "text" || typeof first.text !== "string") return void 0;
+  const view = viewText(args, first.text, responseFile);
+  return { view, content: [{ ...first, text: view.text }, ...parts.slice(1)] };
+}
+function hasAnswerText(tool, args, content, isError) {
+  if (isError) return false;
+  if (finishedJob(tool, args, content)) return true;
+  if (tool !== "run_task" && tool !== "delegate_research") return false;
+  return args?.preview !== true && args?.defer !== true && content.deferred !== true;
+}
 function resultModel(content) {
   const model = content?.selected_model ?? object4(content?.route)?.selected_model;
   return typeof model === "string" && model ? model : void 0;
@@ -25206,13 +25601,19 @@ async function finishResult(tool, args, result, elapsedMs, startedAt, store2 = r
     return file ? { ...timed, _meta: { ...timed._meta, "freellama/response_file": file } } : timed;
   }
   const { shaped, sections } = shapeResult(tool, object4(args), content, timed.isError === true, file !== void 0);
+  const reshaped = hasAnswerText(tool, object4(args), content, timed.isError === true) ? reshapedContent(timed, object4(args), file) : void 0;
+  const { where, answer } = reshaped?.view ?? {};
+  const lean = where !== void 0 && answer !== void 0 ? withoutAnswerCopies(shaped, answer, `answer moved: ${where}`) : shaped;
   return {
     ...timed,
+    ...reshaped ? { content: reshaped.content } : {},
     structuredContent: {
-      ...shaped,
+      ...lean,
       timing: content.timing,
       ...file ? { response_file: file } : {},
-      ...file && sections.length > 0 ? { sections } : {}
+      ...file && sections.length > 0 ? { sections } : {},
+      ...reshaped?.view.output ? { output: reshaped.view.output } : {},
+      ...reshaped?.view.output_file ? { output_file: reshaped.view.output_file } : {}
     }
   };
 }
@@ -25267,16 +25668,16 @@ function taskNotices(content) {
 }
 
 // src/docs-resources.ts
-import { existsSync as existsSync3, readdirSync } from "node:fs";
+import { existsSync as existsSync3, readdirSync as readdirSync2 } from "node:fs";
 import { readFile as readFile2 } from "node:fs/promises";
-import path4 from "node:path";
+import path6 from "node:path";
 import { fileURLToPath as fileURLToPath3 } from "node:url";
 function registerDocsResources(server2) {
-  const PACKAGED_DOCS_DIR = path4.resolve(path4.dirname(fileURLToPath3(import.meta.url)), "..", "docs");
-  if (!existsSync3(path4.join(PACKAGED_DOCS_DIR, "README.md"))) {
+  const PACKAGED_DOCS_DIR = path6.resolve(path6.dirname(fileURLToPath3(import.meta.url)), "..", "docs");
+  if (!existsSync3(path6.join(PACKAGED_DOCS_DIR, "README.md"))) {
     throw new Error(`FreeLlama MCP documentation index is missing at ${PACKAGED_DOCS_DIR}/README.md; run the package build.`);
   }
-  const packagedDocs = readdirSync(PACKAGED_DOCS_DIR, { recursive: true, encoding: "utf8" }).map((name) => name.split(path4.sep).join("/")).filter((name) => name.endsWith(".md")).sort();
+  const packagedDocs = readdirSync2(PACKAGED_DOCS_DIR, { recursive: true, encoding: "utf8" }).map((name) => name.split(path6.sep).join("/")).filter((name) => name.endsWith(".md")).sort();
   for (const name of packagedDocs) {
     const isIndex = name === "README.md";
     const slug = isIndex ? "index" : name.replace(/\.md$/, "");
@@ -25289,14 +25690,14 @@ function registerDocsResources(server2) {
         description: isIndex ? "Index of packaged FreeLlama operator and agent documentation. Read this first, then fetch one relevant document." : `Packaged FreeLlama documentation: ${name}.`
       },
       async () => ({
-        contents: [{ uri, mimeType: "text/markdown", text: await readFile2(path4.join(PACKAGED_DOCS_DIR, name), "utf8") }]
+        contents: [{ uri, mimeType: "text/markdown", text: await readFile2(path6.join(PACKAGED_DOCS_DIR, name), "utf8") }]
       })
     );
   }
 }
 
 // src/builtin-prompts.ts
-import path6 from "node:path";
+import path7 from "node:path";
 
 // ../rust-core/src/model_catalog.json
 var model_catalog_default = {
@@ -25800,235 +26201,15 @@ Verify: local answers are candidates. Each run_task result ends with check: line
 
 Coding: the answer is SEARCH/REPLACE blocks. ${APPLY_EDITS}
 
-Hand a result on: saveAs:"<name>" writes the answer to handoffs/<name>.md and returns handoff_file; a later run_task reads it with files:[handoff_file].
+Hand a result on: saveAs:"<name>" writes the answer to handoffs/<name>.md and returns handoff_file; a later run_task reads it with files:[handoff_file]. Answers over 8,000 characters come back as a summary with check: lines and an output_file holding the whole answer; output:"inline" returns the whole text, outputPath:"<path>" writes (or replaces) that file inside your project.
 
 Approval: pull a model only after the user approves one exact tag and its size (ollama_pull approvedSizeBytes). Unload only a model you warmed or the user named; delete only a tag the user names. A memory refusal names what holds memory: tell the user and never stop their processes.`;
 
-// src/task-files.ts
-import { globSync, realpathSync as realpathSync2, statSync, readFileSync } from "node:fs";
-import { homedir as homedir2, tmpdir } from "node:os";
-import path5 from "node:path";
-var MAX_TEXT_FILE_BYTES = 4 * 1024 * 1024;
-var MAX_IMAGE_FILE_BYTES = 8 * 1024 * 1024;
-var MAX_TOTAL_FILE_BYTES = 32 * 1024 * 1024;
-var MAX_FILES = 2e3;
-var IMAGE_EXTENSIONS = /* @__PURE__ */ new Set([".png", ".jpg", ".jpeg", ".webp", ".gif"]);
-var AUDIO_EXTENSIONS = /* @__PURE__ */ new Set([".wav"]);
-var EMBEDDING_CHUNK_BYTES = 1536;
-var CHUNK_OVERLAP_LINES = 8;
-function chunkLines(text3, limit = EMBEDDING_CHUNK_BYTES) {
-  const lines = text3.split("\n");
-  const chunks = [];
-  let current = [];
-  let bytes = 0;
-  let start2 = 1;
-  const flush = (end, overlap = 0) => {
-    if (current.length > 0 && current.join("\n").trim()) chunks.push({ text: current.join("\n"), lines: [start2, end] });
-    const carried = overlap > 0 ? current.slice(-overlap) : [];
-    const carriedBytes = carried.reduce((sum, line2) => sum + Buffer.byteLength(line2) + 1, 0);
-    current = carriedBytes < limit / 4 ? carried : [];
-    bytes = current === carried ? carriedBytes : 0;
-    start2 = end + 1 - current.length;
-  };
-  lines.forEach((line2, index) => {
-    const size2 = Buffer.byteLength(line2) + 1;
-    if (bytes + size2 > limit && current.length > 0) flush(index, CHUNK_OVERLAP_LINES);
-    if (size2 > limit) {
-      for (let at = 0; at < line2.length; at += Math.floor(limit / 2)) {
-        chunks.push({ text: line2.slice(at, at + Math.floor(limit / 2)), lines: [index + 1, index + 1] });
-      }
-      start2 = index + 2;
-      return;
-    }
-    current.push(line2);
-    bytes += size2;
-  });
-  flush(lines.length);
-  return chunks;
-}
-function fileRoots(env = process.env, cwd = process.cwd(), home = homedir2()) {
-  const configured = (env.FREELLAMA_FILE_ROOTS ?? "").split(path5.delimiter).filter(Boolean);
-  const broad = /* @__PURE__ */ new Set([path5.parse(path5.resolve(cwd)).root, resolvedRoot(home)]);
-  const project = broad.has(resolvedRoot(cwd)) ? [] : [cwd];
-  const handoffs = path5.join(env.FREELLAMA_HOME || path5.join(home, ".octocode", "freellama"), "handoffs");
-  return [...project, tmpdir(), ...configured, handoffs].map(resolvedRoot);
-}
-function resolvedRoot(root) {
-  try {
-    return realpathSync2(root);
-  } catch {
-    const absolute = path5.resolve(root);
-    const parent = path5.dirname(absolute);
-    return parent === absolute ? absolute : path5.join(resolvedRoot(parent), path5.basename(absolute));
-  }
-}
-function within(file, root) {
-  const relative = path5.relative(root, file);
-  return relative === "" || !relative.startsWith("..") && !path5.isAbsolute(relative);
-}
-function isWav(bytes) {
-  return bytes.length >= 12 && bytes.subarray(0, 4).toString() === "RIFF" && bytes.subarray(8, 12).toString() === "WAVE";
-}
-function expandFileEntries(entries, cwd = process.cwd()) {
-  const out = [];
-  for (const entry of entries) {
-    if (!/[*?[]/.test(entry)) {
-      out.push(entry);
-      continue;
-    }
-    const matches = globSync(entry, { cwd }).map((match) => path5.resolve(cwd, match)).filter((match) => statSync(match).isFile()).sort();
-    if (matches.length === 0) throw new Error(`no files match ${entry} (globs resolve from ${cwd})`);
-    out.push(...matches);
-    if (out.length > MAX_FILES) break;
-  }
-  return out;
-}
-function readFiles(entries, givenRoots) {
-  const roots = givenRoots.map(resolvedRoot);
-  const paths = expandFileEntries(entries);
-  if (paths.length > MAX_FILES) throw new Error(`files names ${paths.length} paths; at most ${MAX_FILES} per call`);
-  let total = 0;
-  return paths.map((name) => {
-    const absolute = path5.resolve(name);
-    let real;
-    try {
-      real = realpathSync2(absolute);
-    } catch {
-      throw new Error(`file not found: ${name}`);
-    }
-    if (!roots.some((root) => within(real, root))) {
-      throw new Error(
-        `file ${name} is outside the readable roots (${roots.join(", ")}); copy it into the project or temp directory, or add its directory to FREELLAMA_FILE_ROOTS`
-      );
-    }
-    const stat3 = statSync(real);
-    if (!stat3.isFile()) throw new Error(`not a regular file: ${name}`);
-    const extension = path5.extname(real).toLowerCase();
-    let media = AUDIO_EXTENSIONS.has(extension) ? "audio" : IMAGE_EXTENSIONS.has(extension) ? "image" : void 0;
-    const limit = media ? MAX_IMAGE_FILE_BYTES : MAX_TEXT_FILE_BYTES;
-    if (stat3.size > limit) {
-      throw new Error(`file ${name} is ${stat3.size} bytes; the limit is ${limit} for ${media ?? "text"}; split it or send the part that matters`);
-    }
-    total += stat3.size;
-    if (total > MAX_TOTAL_FILE_BYTES) throw new Error(`files total more than ${MAX_TOTAL_FILE_BYTES} bytes; split the call`);
-    const bytes = readFileSync(real);
-    if (media === "audio" && !isWav(bytes)) {
-      throw new Error(`audio file ${name} must contain a WAV RIFF/WAVE header`);
-    }
-    if (isWav(bytes)) media = "audio";
-    return { path: name, bytes, media };
-  });
-}
-function shortPath(name) {
-  const relative = path5.relative(process.cwd(), path5.resolve(name));
-  return relative && !relative.startsWith("..") && !path5.isAbsolute(relative) ? relative : name;
-}
-function leadingComment(body) {
-  const lines = [];
-  for (const line2 of body.split("\n").slice(0, 20)) {
-    const trimmed = line2.trim();
-    if (!/^(\/\/|#(?!include)|\/\*|\*|"""|\s*$)/.test(trimmed) || trimmed.startsWith("#!")) break;
-    if (trimmed) lines.push(trimmed.replace(/^(\/\/[!/]?|#+|\/\*+|\*\/?|""")\s?/, ""));
-  }
-  return lines.join(" ").slice(0, 400);
-}
-var MAX_AUDIO_FILE_BYTES = 256 * 1024 * 1024;
-function readAllowedFile(name, limit, roots = fileRoots()) {
-  const allowed = roots.map(resolvedRoot);
-  let real;
-  try {
-    real = realpathSync2(path5.resolve(name));
-  } catch {
-    throw new Error(`file not found: ${name}`);
-  }
-  if (!allowed.some((root) => within(real, root))) {
-    throw new Error(`file ${name} is outside the readable roots (${allowed.join(", ")}); copy it into the project or temp directory, or add its directory to FREELLAMA_FILE_ROOTS`);
-  }
-  const stat3 = statSync(real);
-  if (!stat3.isFile()) throw new Error(`not a regular file: ${name}`);
-  if (stat3.size > limit) throw new Error(`file ${name} is ${stat3.size} bytes; the limit is ${limit}`);
-  return readFileSync(real);
-}
-function text(file) {
-  if (file.bytes.includes(0)) throw new Error(`file ${file.path} is binary; only images (${[...IMAGE_EXTENSIONS].join(", ")}) and text files can be read`);
-  return file.bytes.toString("utf8");
-}
-function applyTaskFiles(payload, paths, roots = fileRoots()) {
-  if (paths === void 0 || paths.length === 0) return { payload, files: [] };
-  const read = readFiles(paths, roots);
-  const receipts = [];
-  const media = read.filter((file) => file.media !== void 0);
-  const texts = read.filter((file) => file.media === void 0);
-  const next = { ...payload };
-  const task = payload.task ?? "completion";
-  if (task === "embedding") {
-    const given = payload.input === void 0 ? [] : Array.isArray(payload.input) ? payload.input : [payload.input];
-    const prefix = payload.documentPrefix ?? "";
-    const items = [...given];
-    for (const file of read) {
-      if (file.media) {
-        const encoded = file.bytes.toString("base64");
-        items.push(file.media === "audio" ? { audio: encoded } : { image: encoded });
-        receipts.push({ path: file.path, bytes: file.bytes.length, role: "input" });
-        continue;
-      }
-      const body = text(file);
-      const header = `${shortPath(file.path)}
-${leadingComment(body)}`.trimEnd();
-      for (const chunk of chunkLines(body)) {
-        items.push(`${prefix}${header}
-${chunk.text}`);
-        receipts.push({ path: file.path, bytes: Buffer.byteLength(chunk.text), role: "input", lines: chunk.lines });
-      }
-    }
-    next.input = items;
-    return { payload: next, files: receipts };
-  }
-  const audio = media.find((file) => file.media === "audio");
-  if (audio && task !== "decision") throw new Error(`audio file ${audio.path} is read only by tasks "decision" and "embedding" (models with audio input)`);
-  if (media.length > 0) {
-    next.images = [...payload.images ?? [], ...media.map((file) => file.bytes.toString("base64"))];
-    for (const file of media) receipts.push({ path: file.path, bytes: file.bytes.length, role: file.media });
-  }
-  if (task === "decision") {
-    if (texts.length > 0 && payload.state !== void 0) throw new Error("give the decision state either as `state` or as text `files`, not both");
-    if (texts.length === 1) {
-      const body = text(texts[0]);
-      next.state = jsonOrText(body);
-    } else if (texts.length > 1) {
-      next.state = Object.fromEntries(texts.map((file) => [file.path, text(file)]));
-    }
-    for (const file of texts) receipts.push({ path: file.path, bytes: file.bytes.length, role: "state" });
-    return { payload: next, files: receipts };
-  }
-  if (texts.length > 0) {
-    if (payload.prompt === void 0) throw new Error("text `files` are appended to `prompt`; give the instructions in `prompt`");
-    const sections = texts.map((file) => `--- file: ${file.path} ---
-${text(file)}`);
-    next.prompt = [payload.prompt, ...sections].join("\n\n");
-    for (const file of texts) receipts.push({ path: file.path, bytes: file.bytes.length, role: "prompt" });
-  }
-  return { payload: next, files: receipts };
-}
-function jsonOrText(body) {
-  try {
-    const value = JSON.parse(body);
-    if (value !== null && typeof value === "object") return value;
-  } catch {
-  }
-  return body;
-}
-function inputLabels(inputCount, files) {
-  const fromFiles = files.filter((file) => file.role === "input").map((file) => file.lines ? `${shortPath(file.path)}:L${file.lines[0]}-${file.lines[1]}` : shortPath(file.path));
-  const given = inputCount - fromFiles.length;
-  return [...Array.from({ length: given }, (_, index) => `input[${index}]`), ...fromFiles];
-}
-
 // src/builtin-prompts.ts
 var list = (value) => value.split(/[,\n]/).map((entry) => entry.trim()).filter(Boolean);
 function short(file) {
-  const relative = path6.relative(process.cwd(), path6.resolve(file));
-  return relative && !relative.startsWith("..") && !path6.isAbsolute(relative) ? relative : file;
+  const relative = path7.relative(process.cwd(), path7.resolve(file));
+  return relative && !relative.startsWith("..") && !path7.isAbsolute(relative) ? relative : file;
 }
 function perFile(files, shared, item = {}) {
   let paths = list(files);
@@ -26236,7 +26417,7 @@ function installedModelsText(data) {
       ].filter(Boolean).join("  ")
     );
   }
-  lines.push('Context, policy rank and benchmark per model: structuredContent; one model: models{view:"detail", model}.');
+  lines.push('Context, policy rank and benchmark per model: models{view:"installed", includeVerbose:true}; one model: models{view:"detail", model}.');
   return lines.join("\n");
 }
 function residentModelsText(data) {
@@ -27154,6 +27335,32 @@ var modelsResultSchema = object2({
   knowledge: unknown().optional(),
   library: unknown().optional()
 }).loose();
+var SLIM_MODEL_KEYS = [
+  "name",
+  "model",
+  "size",
+  "capabilities",
+  "resident",
+  "resident_size",
+  "resident_vram",
+  "resident_context",
+  "advertised_context",
+  "placement"
+];
+function slimModel(entry, includeKnowledge) {
+  const slim = {};
+  for (const key of SLIM_MODEL_KEYS) if (entry[key] !== void 0) slim[key] = entry[key];
+  const execution = entry.execution;
+  if (execution && typeof execution === "object") {
+    slim.execution = {
+      ...execution.placement !== void 0 ? { placement: execution.placement } : {},
+      ...execution.upstream !== void 0 ? { upstream: execution.upstream } : {},
+      ...execution.observation !== void 0 ? { observation: execution.observation } : {}
+    };
+  }
+  if (includeKnowledge && entry.knowledge !== void 0) slim.knowledge = entry.knowledge;
+  return slim;
+}
 var HOST_MEMORY_FIT_FRACTION = 0.6;
 var HUB_TASKS = /* @__PURE__ */ new Map([
   ["embedding", ["feature-extraction", "sentence-similarity"]],
@@ -27382,8 +27589,8 @@ function registerModelsTool(server2) {
         if (selectedView !== "library" && selectedView !== "raw" && [limit, cursor].some((value) => value !== void 0)) {
           return errorResult(new Error(`\`limit\` and \`cursor\` apply only to views "library" and "raw", not "${selectedView}".`));
         }
-        if (selectedView !== "detail" && includeVerbose !== void 0) {
-          return errorResult(new Error('`includeVerbose` is valid only for view "detail".'));
+        if (selectedView !== "detail" && selectedView !== "installed" && selectedView !== "resident" && includeVerbose !== void 0) {
+          return errorResult(new Error('`includeVerbose` is valid only for views "installed", "resident" and "detail".'));
         }
         if (selectedView !== "detail" && selectedView !== "library" && model !== void 0) {
           return errorResult(new Error('`model` is valid only for views "detail" and "library".'));
@@ -27417,7 +27624,8 @@ function registerModelsTool(server2) {
                 }
               };
             });
-            return structuredResult({ ...data, models }, { text: residentModelsText({ models }) });
+            const shown2 = includeVerbose ? models : models.map((entry) => slimModel(entry, false));
+            return structuredResult({ ...data, models: shown2 }, { text: residentModelsText({ models }) });
           }
           case "detail": {
             if (!model) {
@@ -27463,7 +27671,8 @@ function registerModelsTool(server2) {
               includeLibrary ?? false,
               async (upstream) => (await ollamaFetch(upstream ?? ollamaEndpoint, "/api/tags")).models ?? []
             ) };
-            return structuredResult(installed, { text: installedModelsText(installed) });
+            const shown2 = includeVerbose ? installed : { ...installed, models: installed.models.map((entry) => slimModel(entry, includeLibrary === true)) };
+            return structuredResult(shown2, { text: installedModelsText(installed) });
           }
         }
       } catch (error2) {
@@ -27721,36 +27930,6 @@ function mergeProfiles(job, profiles) {
   return { job, model: models.length === 1 ? models[0] : models.join(", ") || void 0, applied: [...new Set(profiles.flatMap((profile) => profile.applied))] };
 }
 
-// src/handoffs.ts
-init_config();
-import { mkdirSync, writeFileSync } from "node:fs";
-import path7 from "node:path";
-var HANDOFF_DIR = path7.join(FREELLAMA_HOME, "handoffs");
-var HANDOFF_NAME = /^[a-z0-9-]{1,60}$/;
-function summaryOf(text3, lines = 5) {
-  const kept = text3.split("\n").filter((line2) => line2.trim()).slice(0, lines).join("\n");
-  return kept.length > 600 ? `${kept.slice(0, 600)}\u2026` : kept;
-}
-function writeHandoff(name, content, dir = HANDOFF_DIR) {
-  if (!HANDOFF_NAME.test(name)) throw new Error("saveAs is 1 to 60 characters of a-z, 0-9 and -");
-  mkdirSync(dir, { recursive: true, mode: 448 });
-  const file = path7.join(dir, `${name}.md`);
-  const body = `# ${name}
-
-${content.source}, ${(/* @__PURE__ */ new Date()).toISOString()}.
-
-## Summary
-
-${content.summary.trim() || "(empty)"}
-
-## Result
-
-${content.result.trim() || "(empty)"}
-`;
-  writeFileSync(file, body, { mode: 384 });
-  return file;
-}
-
 // src/result-checks.ts
 import { readFileSync as readFileSync2 } from "node:fs";
 var object5 = (value) => value !== null && typeof value === "object" && !Array.isArray(value) ? value : void 0;
@@ -28469,6 +28648,7 @@ function registerRunTaskTool(server2) {
         receipt: _enum(["compact", "full"]).optional().describe("compact (default): full receipt at response_file; full: inline"),
         defer: boolean2().optional(),
         saveAs: string2().regex(HANDOFF_NAME).optional().describe("answer to handoffs/<name>.md, returned as handoff_file"),
+        ...outputInput,
         returnEmbeddings: boolean2().optional(),
         nearest: object2({
           query: string2().min(1).optional(),
@@ -28491,6 +28671,7 @@ function registerRunTaskTool(server2) {
       job,
       workflow,
       saveAs,
+      outputPath,
       objective,
       model,
       sessionId,
@@ -28528,6 +28709,10 @@ function registerRunTaskTool(server2) {
       maxParallelism
     }, extra) => {
       try {
+        if (outputPath !== void 0) {
+          if (preview || defer) return errorResult(new Error("`outputPath` saves a finished answer; it is not valid with `preview` or `defer`."));
+          checkOutputPath(outputPath);
+        }
         if (workflow !== void 0) {
           const flowJob = WORKFLOWS[workflow].job;
           if (job !== void 0 && job !== flowJob) return errorResult(new Error(`workflow "${workflow}" runs job "${flowJob}"; remove \`job\`.`));
@@ -28832,7 +29017,8 @@ function registerTaskJobsTool(server2) {
         endpoint: endpointParam,
         action: _enum(["list", "get", "cancel", "remove"]),
         jobId: string2().uuid().optional(),
-        returnEmbeddings: boolean2().optional()
+        returnEmbeddings: boolean2().optional(),
+        output: outputInput.output.describe("get only; as run_task")
       }).strict(),
       outputSchema: objectResultSchema,
       // One tool for reads and changes, so not read-only (the spec default, not restated). Not destructive either: cancel and remove
@@ -28840,13 +29026,13 @@ function registerTaskJobsTool(server2) {
       // outside FreeLlama, and a job's input can be submitted again.
       annotations: { destructiveHint: false }
     },
-    async ({ endpoint, action, jobId, returnEmbeddings }) => {
+    async ({ endpoint, action, jobId, returnEmbeddings, output }) => {
       try {
         if (action === "list" ? jobId !== void 0 : jobId === void 0) {
           return errorResult(new Error("Use jobId only with get/cancel/remove; all require a jobId."));
         }
-        if (returnEmbeddings !== void 0 && action !== "get") {
-          return errorResult(new Error("returnEmbeddings is valid only with get."));
+        if ((returnEmbeddings !== void 0 || output !== void 0) && action !== "get") {
+          return errorResult(new Error("returnEmbeddings and output are valid only with get."));
         }
         const result = parsedResult(action === "list" ? await listTaskJobs2(endpoint) : action === "get" ? await getTaskJob2(endpoint, jobId) : action === "cancel" ? await cancelTaskJob2(endpoint, jobId) : await removeTaskJob2(endpoint, jobId));
         if (!returnEmbeddings && isStructuredSuccess(result)) {
@@ -29607,7 +29793,7 @@ function cancelOpenJobs(openJobs, cancel) {
 }
 
 // src/research/source-evidence.ts
-import { realpathSync as realpathSync3, statSync as statSync2 } from "node:fs";
+import { realpathSync as realpathSync4, statSync as statSync3 } from "node:fs";
 import path9 from "node:path";
 
 // src/research/canonical-json.ts
@@ -29743,11 +29929,11 @@ var READS = /* @__PURE__ */ new Set(["cat", "head", "tail", "nl", "wc", "diff",
 function sourcePath(value, workspace) {
   if (!value || value.startsWith("-") || /[*?$`]/.test(value)) return false;
   try {
-    const root = realpathSync3(workspace);
-    const source = realpathSync3(path9.resolve(root, value));
+    const root = realpathSync4(workspace);
+    const source = realpathSync4(path9.resolve(root, value));
     const relative = path9.relative(root, source);
     if (relative === ".." || relative.startsWith(`..${path9.sep}`) || path9.isAbsolute(relative)) return false;
-    const stat3 = statSync2(source);
+    const stat3 = statSync3(source);
     return stat3.isFile() || stat3.isDirectory();
   } catch {
     return false;
@@ -29826,7 +30012,7 @@ function searchHitCitations(steps, answer, workspace) {
 }
 
 // src/delegate.ts
-import { existsSync as existsSync5, readFileSync as readFileSync3, realpathSync as realpathSync4, statSync as statSync3 } from "node:fs";
+import { existsSync as existsSync5, readFileSync as readFileSync3, realpathSync as realpathSync5, statSync as statSync4 } from "node:fs";
 import path11 from "node:path";
 init_config();
 function researchErrorResult(error2, diagnostic) {
@@ -29962,7 +30148,7 @@ function assessDelegatedAnswer(question, evidenceCount, model, verifiedDigest) {
 function extractExistingWorkspacePath(command, workspace) {
   let root;
   try {
-    root = realpathSync4(workspace);
+    root = realpathSync5(workspace);
   } catch {
     return null;
   }
@@ -29976,10 +30162,10 @@ function extractExistingWorkspacePath(command, workspace) {
     const candidate = path11.isAbsolute(token) ? token : path11.resolve(root, token);
     if (!existsSync5(candidate)) continue;
     try {
-      const resolved = realpathSync4(candidate);
+      const resolved = realpathSync5(candidate);
       const relative = path11.relative(root, resolved);
       const withinRoot = relative && relative !== ".." && !relative.startsWith(`..${path11.sep}`);
-      if (withinRoot && statSync3(resolved).isFile()) return relative;
+      if (withinRoot && statSync4(resolved).isFile()) return relative;
     } catch {
     }
   }
@@ -30022,7 +30208,7 @@ function parseAdapterResult(raw) {
 }
 
 // src/tools/coder.ts
-import { existsSync as existsSync7, rmSync as rmSync2 } from "node:fs";
+import { existsSync as existsSync7, rmSync as rmSync3 } from "node:fs";
 import { mkdtemp, realpath as realpath2 } from "node:fs/promises";
 import { tmpdir as tmpdir2 } from "node:os";
 import path13 from "node:path";
@@ -30032,13 +30218,13 @@ import { spawnSync } from "node:child_process";
 import {
   cpSync,
   existsSync as existsSync6,
-  lstatSync,
+  lstatSync as lstatSync2,
   mkdirSync as mkdirSync2,
   mkdtempSync,
   readFileSync as readFileSync4,
-  readdirSync as readdirSync2,
-  realpathSync as realpathSync5,
-  rmSync,
+  readdirSync as readdirSync3,
+  realpathSync as realpathSync6,
+  rmSync as rmSync2,
   symlinkSync,
   writeFileSync as writeFileSync2
 } from "node:fs";
@@ -30069,30 +30255,30 @@ function repositoryTop(dir) {
   const top = git(["rev-parse", "--show-toplevel"], dir);
   if (top.status !== 0) return null;
   if (git(["rev-parse", "--verify", "--quiet", "HEAD"], dir).status !== 0) return null;
-  return realpathSync5(line(top.stdout));
+  return realpathSync6(line(top.stdout));
 }
 function commonDirOf(top) {
   const raw = line(gitOk(["rev-parse", "--git-common-dir"], top));
-  return realpathSync5(path12.resolve(top, raw));
+  return realpathSync6(path12.resolve(top, raw));
 }
 function removeIsolatedCopy(workspace, container) {
   const work = path12.join(container, WORKTREE_NAME);
   let commonDir = null;
   try {
-    const top = repositoryTop(realpathSync5(workspace));
+    const top = repositoryTop(realpathSync6(workspace));
     commonDir = top ? commonDirOf(top) : null;
     if (top && existsSync6(work)) git(["worktree", "remove", "--force", work], top);
   } catch {
   }
   try {
-    rmSync(container, { recursive: true, force: true });
+    rmSync2(container, { recursive: true, force: true });
   } catch {
   }
   if (!commonDir) return;
   const entries = path12.join(commonDir, "worktrees");
   let names;
   try {
-    names = readdirSync2(entries);
+    names = readdirSync3(entries);
   } catch {
     return;
   }
@@ -30100,7 +30286,7 @@ function removeIsolatedCopy(workspace, container) {
   for (const name of names) {
     try {
       if (ours.has(readFileSync4(path12.join(entries, name, "gitdir"), "utf8").trim())) {
-        rmSync(path12.join(entries, name), { recursive: true, force: true });
+        rmSync2(path12.join(entries, name), { recursive: true, force: true });
       }
     } catch {
     }
@@ -30140,7 +30326,7 @@ async function prepareCoderRun(commands) {
     patchPath,
     cleanup: (workspace) => {
       removeIsolatedCopy(workspace, container);
-      if (!existsSync7(patchPath)) rmSync2(patchDir, { recursive: true, force: true });
+      if (!existsSync7(patchPath)) rmSync3(patchDir, { recursive: true, force: true });
     }
   };
 }
@@ -30216,6 +30402,25 @@ var researchResultSchema = object2({
   usage: unknown().optional(),
   telemetry: unknown().optional()
 }).loose();
+var RESEARCH_ROLES = ["stronger", "default", "small"];
+var GITHUB_RESEARCH_ROLES = ["stronger", "default"];
+async function catalogResearchModel(endpoint, adapter) {
+  try {
+    const [installed, available] = await catalogLiveState(endpoint, void 0);
+    const picks = catalogView("research-agent", installed, available).jobs[0]?.picks ?? [];
+    const nameOf = (pick3) => "installed_as" in pick3 && typeof pick3.installed_as === "string" ? pick3.installed_as : pick3.tag;
+    for (const role of adapter === "github" ? GITHUB_RESEARCH_ROLES : RESEARCH_ROLES) {
+      const pick3 = picks.find((entry) => entry.role === role && entry.installed === true && entry.fits_now === true);
+      if (pick3) return { model: nameOf(pick3), reason: `catalog ${role} pick, installed, fits now` };
+    }
+    for (const role of ["default", "stronger"]) {
+      const pick3 = picks.find((entry) => entry.role === role && entry.installed === true);
+      if (pick3) return { model: nameOf(pick3), reason: `catalog ${role} pick, installed; not fitting now, so serve's memory gate decides` };
+    }
+  } catch {
+  }
+  return void 0;
+}
 async function routedResearchModel(endpoint) {
   const decision = JSON.parse(
     await route2(endpoint, "coding", "balanced", void 0, void 0, void 0, [], void 0, void 0, void 0)
@@ -30231,7 +30436,7 @@ function registerDelegateResearchTool(server2) {
   server2.registerTool(
     "delegate_research",
     {
-      description: 'Use when: a read-only question about files under workspacePath (where is X, what calls Y) would take several searches; it needs a local model that calls tools well: tools metadata does not prove it, so take the catalog\'s research-agent pick. Do not use when: editing (adapter:"coder" edits an isolated copy, if the operator enabled it), facts outside the workspace (GitHub excepted: adapter:"github", which needs no `workspacePath`), final judgment, or a question one search answers; run_task embedding with `nearest` finds files faster. Up to 8 turns, killed after 180 s by default (longer for bigger models, or agent.timeoutSeconds); no defer (a host that cuts calls near 60 s may drop it). Returns: citations/answer; discard verification.recommendation=escalate, otherwise verify citations. coder: review diff, then run its applyCommand.',
+      description: 'Use when: a read-only question about files under workspacePath (where is X, what calls Y) would take several searches; without `model` it uses the catalog\'s research-agent pick that is installed and fits now. Do not use when: editing (adapter:"coder" edits an isolated copy, if the operator enabled it), facts outside the workspace (GitHub excepted: adapter:"github", which needs no `workspacePath`), final judgment, or a question one search answers; run_task embedding with `nearest` finds files faster. Up to 8 turns, killed after 180 s by default (longer for bigger models, or agent.timeoutSeconds); no defer (a host that cuts calls near 60 s may drop it). Returns: citations/answer; discard verification.recommendation=escalate, otherwise verify citations. coder: review diff, then run its applyCommand.',
       inputSchema: object2({
         question: string2().min(1),
         workspacePath: string2().min(1).optional().describe(
@@ -30247,7 +30452,8 @@ function registerDelegateResearchTool(server2) {
           timeoutSeconds: number2().int().positive().max(MAX_DELEGATE_TIMEOUT_SECONDS).optional().describe("default scales with the model's size"),
           commands: coderCommandsSchema
         }).strict().optional(),
-        receipt: _enum(["compact", "full"]).optional().describe("compact (default): trail at response_file; full: inline")
+        receipt: _enum(["compact", "full"]).optional().describe("compact (default): trail at response_file; full: inline"),
+        ...outputInput
       }).strict(),
       outputSchema: researchResultSchema,
       // Read-only lookup; adapter "github" searches GitHub over the internet, so the tool is
@@ -30255,18 +30461,33 @@ function registerDelegateResearchTool(server2) {
       annotations: { destructiveHint: false }
     },
     // `receipt` only shapes the inline view (tool-results.ts); the tool always returns everything.
-    async ({ question, workspacePath, adapter, model, endpoint, agent }, extra) => {
+    async ({ question, workspacePath, adapter, model, endpoint, agent, outputPath }, extra) => {
       const chosenAdapter = adapter ?? DEFAULT_RESEARCH_ADAPTER;
       let resolvedWorkspace;
       let chosenModel;
+      let modelReason = "explicit model";
       try {
+        if (outputPath !== void 0) checkOutputPath(outputPath);
         if (workspacePath === void 0 && chosenAdapter !== "github") {
           throw new Error(`workspacePath is required for adapter "${chosenAdapter}": the absolute directory to research (or one relative to the server's working directory). Only adapter "github" runs without one.`);
         }
         const refusal = coderRefusal(chosenAdapter, agent?.commands);
         if (refusal) throw refusal;
         resolvedWorkspace = workspacePath === void 0 ? void 0 : await assertAllowedWorkspace(workspacePath);
-        chosenModel = model ?? DEFAULT_DELEGATE_MODEL ?? await routedResearchModel(endpoint);
+        if (model) chosenModel = model;
+        else if (DEFAULT_DELEGATE_MODEL) {
+          chosenModel = DEFAULT_DELEGATE_MODEL;
+          modelReason = "FREELLAMA_MCP_DEFAULT_MODEL";
+        } else {
+          const picked = await catalogResearchModel(endpoint, chosenAdapter);
+          if (picked) {
+            chosenModel = picked.model;
+            modelReason = picked.reason;
+          } else {
+            chosenModel = await routedResearchModel(endpoint);
+            modelReason = "routing (no catalog research pick installed)";
+          }
+        }
       } catch (error2) {
         return researchErrorResult(error2);
       }
@@ -30422,9 +30643,10 @@ ${evidenceText}` : "");
         }
         const digest = verifiedResearchDigest(result.model_metadata?.execution_receipts, chosenModel);
         const verification = assessDelegatedAnswer(question, sources.length, chosenModel, digest);
-        const summary = `Delegated answer ready: ${result.tool_calls.length} tool call(s)` + (failed > 0 ? `, ${failed} of which did not succeed` : "") + `; ${result.usage.input_tokens ?? "?"} input / ${result.usage.output_tokens ?? "?"} output local tokens; verification=${verification.recommendation}. Read structuredContent.answer and citations.`;
+        const summary = `Delegated answer ready: ${result.tool_calls.length} tool call(s)` + (failed > 0 ? `, ${failed} of which did not succeed` : "") + `; ${result.usage.input_tokens ?? "?"} input / ${result.usage.output_tokens ?? "?"} output local tokens; verification=${verification.recommendation}; research model: ${chosenModel} (${modelReason}). Read structuredContent.answer and citations.`;
         const payload = {
           adapter: chosenAdapter,
+          researchModel: { model: chosenModel, reason: modelReason },
           verification,
           answer: result.final_answer,
           // recommendation + why live under `verification`; `citations` completes the triple so a
@@ -30452,7 +30674,7 @@ ${evidenceText}` : "");
 
 verification: ${verification.recommendation}` + (cited.length ? `
 cited: ${cited.join(", ")}` : "") + `
-(${chosenModel}, ${result.tool_calls.length} tool call(s))`;
+(research model: ${chosenModel} (${modelReason}), ${result.tool_calls.length} tool call(s))`;
         return structuredResult(payload, { text: answerText });
       } catch (error2) {
         return researchErrorResult(error2);
@@ -30592,7 +30814,7 @@ function registerReadResponseTool(server2) {
 }
 
 // src/tools/audio.ts
-import { existsSync as existsSync9, realpathSync as realpathSync6, writeFileSync as writeFileSync3 } from "node:fs";
+import { existsSync as existsSync9, realpathSync as realpathSync7, writeFileSync as writeFileSync3 } from "node:fs";
 import { tmpdir as tmpdir4 } from "node:os";
 import path16 from "node:path";
 init_config();
@@ -30663,7 +30885,7 @@ function registerAudioTool(server2) {
           const target = path16.resolve(outputFile ?? path16.join(tmpdir4(), `freellama-speech-${Date.now()}.wav`));
           let directory;
           try {
-            directory = realpathSync6(path16.dirname(target));
+            directory = realpathSync7(path16.dirname(target));
           } catch {
             return errorResult(new Error(`outputFile directory does not exist: ${path16.dirname(target)}`));
           }
@@ -30923,10 +31145,18 @@ ${chunk.text}`
 
 // src/sweep/store.ts
 init_config();
-import { appendFileSync, existsSync as existsSync10, mkdirSync as mkdirSync3, readdirSync as readdirSync3, readFileSync as readFileSync5, renameSync, writeFileSync as writeFileSync4 } from "node:fs";
+import { appendFileSync, existsSync as existsSync10, mkdirSync as mkdirSync3, readdirSync as readdirSync4, readFileSync as readFileSync5, renameSync, writeFileSync as writeFileSync4 } from "node:fs";
 import path18 from "node:path";
 import { createHash as createHash2 } from "node:crypto";
 var SWEEP_ROOT = process.env.FREELLAMA_SWEEP_DIR || path18.join(FREELLAMA_HOME, "sweeps");
+function processAlive(pid) {
+  try {
+    process.kill(pid, 0);
+    return true;
+  } catch (error2) {
+    return error2.code === "EPERM";
+  }
+}
 var hashText = (text3) => createHash2("sha256").update(text3).digest("hex").slice(0, 32);
 var SweepStore = class {
   constructor(root = SWEEP_ROOT) {
@@ -30979,7 +31209,7 @@ var SweepStore = class {
   }
   list() {
     if (!existsSync10(this.root)) return [];
-    return readdirSync3(this.root).filter((name) => /^[0-9a-f-]{36}$/.test(name) && existsSync10(path18.join(this.root, name, "manifest.json")));
+    return readdirSync4(this.root).filter((name) => /^[0-9a-f-]{36}$/.test(name) && existsSync10(path18.join(this.root, name, "manifest.json")));
   }
 };
 function jsonl(file) {
@@ -31088,6 +31318,8 @@ var SweepRunner = class {
   resume(id) {
     if (this.active.has(id)) return this.view(id);
     const manifest = this.store.manifest(id);
+    const other = this.foreignOwner(manifest);
+    if (other !== void 0) throw new Error(`sweep ${id} is still running in server process ${other}; wait for it or stop that process first`);
     if (manifest.phase === "collecting") throw new Error("this sweep was interrupted while collecting pages; start a new sweep");
     manifest.phase = "running";
     delete manifest.finishedAt;
@@ -31144,7 +31376,9 @@ var SweepRunner = class {
       if (result.change === "changed") counts.changed += 1;
       if (result.change === "new") counts.added += 1;
     }
-    const phase = (manifest.phase === "running" || manifest.phase === "collecting") && !this.active.has(id) ? "interrupted" : manifest.phase;
+    const ownerPid = manifest.phase === "running" || manifest.phase === "collecting" ? this.foreignOwner(manifest) : void 0;
+    const phase = (manifest.phase === "running" || manifest.phase === "collecting") && !this.active.has(id) && ownerPid === void 0 ? "interrupted" : manifest.phase;
+    const resumable = phase === "interrupted" || phase === "cancelled" || phase === "completed" && counts.failed > 0 || phase === "failed" && counts.total > 0;
     const elapsedSeconds = Math.round(((manifest.finishedAt ? Date.parse(manifest.finishedAt) : Date.now()) - Date.parse(manifest.startedAt)) / 1e3);
     const pending = Math.max(0, counts.total - counts.done - counts.failed);
     const ms = this.active.get(id)?.ms ?? [];
@@ -31158,6 +31392,8 @@ var SweepRunner = class {
       pages: manifest.pages,
       skipped: manifest.skipped.length,
       error: manifest.error,
+      ...resumable && !(manifest.phase === "collecting") ? { resume: { action: "resume", sweepId: id } } : {},
+      ...ownerPid !== void 0 ? { ownerPid } : {},
       startedAt: manifest.startedAt,
       finishedAt: manifest.finishedAt,
       elapsedSeconds,
@@ -31165,10 +31401,18 @@ var SweepRunner = class {
       files: this.store.files(id)
     };
   }
+  /** The pid of another live server process that owns this sweep's run, if any. */
+  foreignOwner(manifest) {
+    if (manifest.phase !== "running" && manifest.phase !== "collecting") return void 0;
+    const pid = manifest.owner?.pid;
+    return pid !== void 0 && pid !== process.pid && !this.active.has(manifest.config.id) && processAlive(pid) ? pid : void 0;
+  }
   emit(id, message2) {
     for (const listener of this.listeners.get(id) ?? []) listener(message2);
   }
   launch(manifest, resuming = false) {
+    manifest.owner = { pid: process.pid };
+    this.store.saveManifest(manifest);
     const abort = new AbortController();
     const active = { abort, done: Promise.resolve(), ms: [] };
     this.active.set(manifest.config.id, active);
@@ -31382,11 +31626,11 @@ function runnerOf(id) {
   for (const runner of runners.values()) if (runner.isActive(id)) return runner;
   return runnerFor(store.manifest(id).config.endpoint);
 }
-function viewText(view) {
+function viewText2(view) {
   const c = view.counts;
   const changes = c.reused + c.changed + c.added > 0 || c.removed > 0 ? `; since baseline: ${c.added} new, ${c.changed} changed, ${c.reused} unchanged (reused), ${c.removed} removed` : "";
   const eta = view.etaSeconds !== void 0 ? `, about ${Math.max(1, Math.round(view.etaSeconds / 60))} min left` : "";
-  const next = view.phase === "running" || view.phase === "collecting" ? `sweep{action:"wait", sweepId:"${view.id}"} to follow it` : view.phase === "interrupted" || view.phase === "cancelled" || c.failed > 0 ? `sweep{action:"resume", sweepId:"${view.id}"} continues it` : `sweep{action:"results", sweepId:"${view.id}"} lists flagged chunks`;
+  const next = view.phase === "running" || view.phase === "collecting" ? `sweep{action:"wait", sweepId:"${view.id}"} to follow it` : view.resume ? `${view.phase === "interrupted" ? "its server process exited; " : ""}sweep{action:"resume", sweepId:"${view.id}"} continues it, reusing saved chunk results` : view.phase === "interrupted" ? "it stopped while collecting pages and cannot resume: start a new sweep" : c.failed > 0 ? `sweep{action:"resume", sweepId:"${view.id}"} retries the failed chunks` : `sweep{action:"results", sweepId:"${view.id}"} lists flagged chunks`;
   return `sweep ${view.id} ${view.phase}${view.error ? ` (${view.error})` : ""}: ${c.done + c.failed}/${c.total} chunks${view.pages !== void 0 ? ` from ${view.pages} pages` : ""}, ${c.flagged} flagged, ${c.unsure} unsure, ${c.clear} clear, ${c.failed} failed${changes}${eta}. Next: ${next}.`;
 }
 function answerLine(answers2) {
@@ -31447,7 +31691,7 @@ function registerSweepTool(server2) {
         if (action === "list") {
           const sweeps = store.list().map((id) => runnerOf(id).view(id)).sort((left, right) => right.startedAt.localeCompare(left.startedAt)).slice(0, 50);
           return structuredResult({ sweeps }, {
-            text: sweeps.length === 0 ? "no sweeps yet" : sweeps.map((view2) => `- ${view2.name ?? "(unnamed)"}: ${viewText(view2)}`).join("\n")
+            text: sweeps.length === 0 ? "no sweeps yet" : sweeps.map((view2) => `- ${view2.name ?? "(unnamed)"}: ${viewText2(view2)}`).join("\n")
           });
         }
         if (action === "start") return await start(args);
@@ -31456,11 +31700,11 @@ function registerSweepTool(server2) {
         const runner = runnerOf(sweepId);
         if (action === "cancel") {
           const view2 = await runner.cancel(sweepId);
-          return structuredResult(view2, { text: viewText(view2) });
+          return structuredResult(view2, { text: viewText2(view2) });
         }
         if (action === "resume") {
           const view2 = runner.resume(sweepId);
-          return structuredResult(view2, { text: viewText(view2) });
+          return structuredResult(view2, { text: viewText2(view2) });
         }
         if (action === "wait") {
           const report = progressSink(extra);
@@ -31473,7 +31717,7 @@ function registerSweepTool(server2) {
         }
         if (action === "results") return results(sweepId, args.filter ?? "flagged", 50, args.cursor, args.saveAs ? { name: args.saveAs, view: runner.view(sweepId) } : void 0);
         const view = runner.view(sweepId);
-        return structuredResult(view, { text: viewText(view) });
+        return structuredResult(view, { text: viewText2(view) });
       } catch (error2) {
         return errorResult(error2);
       }
@@ -31527,7 +31771,7 @@ async function start(args) {
     }
   };
   const view = await runnerFor(args.endpoint).start(config2);
-  return structuredResult(view, { text: viewText(view) });
+  return structuredResult(view, { text: viewText2(view) });
 }
 function rowLine(row) {
   const verified = row.verified ? ` \u2192 ${row.verified.model}: ${row.verified.verdict} (${answerLine(row.verified.answers)})` : "";
@@ -31571,7 +31815,7 @@ function results(id, filter, limit, cursor, handoff2) {
 (${rows.length - offset - page2.length} more: cursor "${next}")` : "") + "\nThese are a small model's candidates: read 3 to 5 of the cited ranges yourself before relying on the rest.";
   const file = handoff2 ? writeHandoff(handoff2.name, {
     source: `From sweep ${id}${manifest.config.name ? ` (${manifest.config.name})` : ""}, filter ${filter}`,
-    summary: viewText(handoff2.view),
+    summary: viewText2(handoff2.view),
     result: rows.length === 0 ? `no ${filter} chunks` : rows.map(rowLine).join("\n")
   }) : void 0;
   if (file) text3 += `
@@ -31590,7 +31834,7 @@ Name inputs with files (paths or globs under the server's working directory, the
 You own prompts, output format and verification: local answers are candidates, so check 3 to 5 (more where a mistake costs more) before relying on the rest, and treat low decision probabilities as unknown.
 Memory: a refusal names the programs holding memory; tell the user and never stop their processes, then retry once they free memory, or use a smaller installed model with the same capability.
 Approval: before ollama_pull, ask the user for one exact tag (name:tag) and its reported size, and pass that size as approvedSizeBytes; search results are not permission. Unload only a model you warmed (when its burst ends) or one the user named; delete only a tag the user names.
-Results: check isError first, then read the text; open structuredContent or read_response{file:response_file, section} only for a receipt field your next step needs. Pass page.next_cursor (models) and next_offset (read_response) back unchanged. A server restart loses scopes, sessions and jobs. CPU/GPU placement is operator-configured; executionPreference only chooses among eligible backends. Docs: freellama://docs/index.`;
+Results: check isError first, then read the text; open structuredContent or read_response{file:response_file, section} only for a receipt field your next step needs. A long answer (run_task, delegate_research, task_jobs get) comes back as a summary plus the output_file holding it; output:"inline" returns it whole, outputPath picks the file. Pass page.next_cursor (models) and next_offset (read_response) back unchanged. A server restart loses scopes, sessions and jobs. CPU/GPU placement is operator-configured; executionPreference only chooses among eligible backends. Docs: freellama://docs/index.`;
 var server = withToolResults(new McpServer(
   { name: "freellama", version: SERVER_VERSION },
   { instructions: INSTRUCTIONS, capabilities: { logging: {} } }