tryharness 0.7.18 → 0.7.23

package/lib/worktrees.mjs

@@ -142,17 +142,33 @@ export function createWorktrees({ store, root }) {
     return { ok: true };
   };
 
+  /**
+   * The thread's folder changed: its copy of the old repository goes when nothing in it is the person's yet, and the
+   * next turn makes whatever the new folder needs. Changes not yet merged or discarded keep it where it is.
+   */
+  const moved = (threadId) => {
+    const tree = get(threadId);
+    if (tree?.state === "open" && fs.existsSync(tree.path)) {
+      if (stat(tree.path, `${tree.base}..HEAD`).length) return { ok: false, error: `this thread changed files on ${tree.branch}; merge or discard them first` };
+      retire(tree);
+      git(tree.repo, ["branch", "-D", tree.branch]);
+    }
+    if (tree) save(threadId, null);
+    return { ok: true };
+  };
+
   /** A thread that carries on a session from a terminal works where that session did: agents find sessions by folder. */
   const inPlace = (threadId) => save(threadId, { state: "in_place" });
 
   /**
    * Where T3 Code runs a thread's agents (t3.mjs runInT3): the checkout T3 made for the thread's first agent, shared by
-   * the rest as its own worktree is, or the thread's own copy when it worked here before; else a new one on its own
-   * branch from what the person has checked out (as folderFor's), or T3's project folder when the thread's folder is in
-   * no repository.
+   * the rest as its own worktree is, or the thread's own copy when it worked here before; a task's room that works in
+   * place, T3's project checkout itself; else a new one on its own branch from what the person has checked out (as
+   * folderFor's), or T3's project folder when the thread's folder is in no repository.
    */
   const t3Workspace = (thread) => {
     const current = get(thread.id);
+    if (current?.state === "in_place" || current?.inPlace) return { type: "root" };
     if ((current?.state === "t3" || current?.state === "open") && current.path) return { type: "existing_worktree", worktreePath: current.path, ...(current.branch ? { branch: current.branch } : {}) };
     const repo = repositoryOf(thread.goal.workspace);
     if (!repo) return { type: "root" };
@@ -163,8 +179,8 @@ export function createWorktrees({ store, root }) {
   /** The thread works in T3 Code from now on; `where` and `branch` once T3 has made its checkout. */
   const inT3 = (threadId, where = null, branch = null) => {
     const current = get(threadId);
-    if (current?.state !== "t3" || (where && where !== current.path)) save(threadId, { ...current, state: "t3", path: where ?? current?.path ?? null, branch: branch ?? current?.branch ?? null });
+    if (current?.state !== "t3" || (where && where !== current.path)) save(threadId, { ...current, ...(current?.state === "in_place" ? { inPlace: true } : {}), state: "t3", path: where ?? current?.path ?? null, branch: branch ?? current?.branch ?? null });
   };
 
-  return { folderFor, commitTurn, diff, summary, merge, discard, inPlace, t3Workspace, inT3, get };
+  return { folderFor, commitTurn, diff, summary, merge, discard, moved, inPlace, t3Workspace, inT3, get };
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tryharness",
-  "version": "0.7.18",
+  "version": "0.7.23",
   "description": "Harness Relay: the daemon that turns the agents on this machine into your personal swarm. Memory, history, and work handed between any of them.",
   "bin": {
     "tryharness": "cli.mjs"
@@ -32,15 +32,17 @@
     "@clack/core": "0.5.0",
     "@clack/prompts": "0.11.0",
     "@huggingface/tokenizers": "^0.1.3",
+    "jpeg-js": "0.4.4",
     "node-sqlite3-wasm": "0.8.60",
-    "onnxruntime-web": "1.26.0-dev.20260416-b7804b056c",
+    "onnxruntime-node": "1.24.3",
     "picocolors": "^1.1.1",
+    "pngjs": "7.0.0",
     "qr": "0.4.2",
     "ws": "^8.18.3"
   },
   "optionalDependencies": {
-    "tryharness-desktop-darwin-arm64": "0.7.17",
-    "tryharness-desktop-darwin-x64": "0.7.17",
+    "tryharness-desktop-darwin-arm64": "0.7.23",
+    "tryharness-desktop-darwin-x64": "0.7.23",
     "tryharness-desktop-linux-x64": "0.7.4",
     "tryharness-desktop-win32-x64": "0.7.4"
   }

package/lib/embedder.mjs

@@ -1,156 +0,0 @@
-import fs from "node:fs";
-import path from "node:path";
-import { ensureDir } from "./files.mjs";
-
-// The daemon's own embedder for the context space: the same model, quantization, pooling
-// and normalization as the browser (src/shared/memory/embeddingService.ts), so the two
-// sides describe the same space under the same version key. It runs on onnxruntime-web's
-// Node build (wasm, no native addon) with the same tokenizer library transformers.js uses,
-// and mirrors transformers' feature-extraction pipeline step by step: tokenize with special
-// tokens, truncate to the model's 512, pad to the batch, take the [CLS] vector, L2-normalize.
-// Weights are downloaded into the Harness home during setup, never fetched from a browser.
-
-export const EMBEDDING_MODEL = "Xenova/bge-base-en-v1.5";
-export const EMBEDDING_DIMENSIONS = 768;
-export const EMBEDDING_VERSION = "bge-base-en-v1.5:q8:cls:l2:asymmetric:v1";
-const QUERY_PREFIX = "Represent this sentence for searching relevant passages: ";
-const MODEL_FILES = ["config.json", "tokenizer.json", "tokenizer_config.json", "onnx/model_quantized.onnx"];
-const REMOTE = `https://huggingface.co/${EMBEDDING_MODEL}/resolve/main/`;
-// One text per run: a q8 model quantizes per batch tensor, so the same text in different company
-// differs in the last decimals (transformers.js shows the same, cosine 0.992 alone versus batched).
-// Entries arrive one at a time and a vector must not depend on its neighbours, so no batching.
-const BATCH = 1;
-
-export function modelDirectory(modelsDir) {
-  return path.join(modelsDir, ...EMBEDDING_MODEL.split("/"));
-}
-
-/** True when every model file is present and non-empty. */
-export function modelPresent(modelsDir) {
-  const dir = modelDirectory(modelsDir);
-  return MODEL_FILES.every((file) => {
-    try { return fs.statSync(path.join(dir, file)).size > 0; } catch { return false; }
-  });
-}
-
-async function download(url, target, { fetchImpl, log }) {
-  const partial = `${target}.partial`;
-  ensureDir(path.dirname(target));
-  let have = 0;
-  try { have = fs.statSync(partial).size; } catch {}
-  const headers = have > 0 ? { range: `bytes=${have}-` } : {};
-  const response = await fetchImpl(url, { headers, redirect: "follow" });
-  if (response.status === 416) { fs.renameSync(partial, target); return; }
-  if (!response.ok) throw new Error(`download of ${path.basename(target)} failed: HTTP ${response.status}`);
-  const resumed = response.status === 206;
-  const total = Number(response.headers.get("content-length") || 0) + (resumed ? have : 0);
-  const out = fs.createWriteStream(partial, { flags: resumed ? "a" : "w" });
-  let done = resumed ? have : 0;
-  let shown = -1;
-  const name = path.basename(target);
-  for await (const chunk of response.body) {
-    if (!out.write(chunk)) await new Promise((resolve) => out.once("drain", resolve));
-    done += chunk.length;
-    const percent = total ? Math.floor((done / total) * 10) * 10 : -1;
-    if (percent !== shown && percent >= 0) { shown = percent; log?.(`embedder: ${name} ${percent}%`); }
-  }
-  await new Promise((resolve, reject) => { out.end(); out.on("finish", resolve); out.on("error", reject); });
-  if (total && done !== total) throw new Error(`download of ${name} stopped at ${done} of ${total} bytes; run setup again to resume`);
-  fs.renameSync(partial, target);
-}
-
-/** Downloads any missing model file into modelsDir. Idempotent; a partial file resumes. */
-export async function ensureModel({ modelsDir, log = () => {}, fetchImpl = fetch } = {}) {
-  const dir = modelDirectory(modelsDir);
-  const missing = MODEL_FILES.filter((file) => { try { return fs.statSync(path.join(dir, file)).size === 0; } catch { return true; } });
-  if (!missing.length) return { downloaded: [], present: true };
-  log(`embedder: downloading ${EMBEDDING_MODEL} (${missing.length} file${missing.length === 1 ? "" : "s"}, about 110 MB) into ${modelsDir}`);
-  for (const file of missing) await download(REMOTE + file, path.join(dir, file), { fetchImpl, log });
-  log("embedder: ready");
-  return { downloaded: missing, present: true };
-}
-
-function truncate(encoded, maxLength) {
-  for (const key of Object.keys(encoded)) encoded[key] = encoded[key].slice(0, maxLength);
-}
-
-function pad(encoded, maxLength, padId) {
-  for (const key of Object.keys(encoded)) {
-    const fill = key === "input_ids" ? padId : 0;
-    while (encoded[key].length < maxLength) encoded[key].push(fill);
-  }
-}
-
-/**
- * A lazily loaded embedder over local files only. Loading the session takes about half a
- * second and inference runs on the calling thread (onnxruntime-web's Node build has no
- * worker), so each short text holds the event loop for a few tens of milliseconds;
- * callers embed in the background, never inside a request the person is waiting on.
- */
-export function createEmbedder({ modelsDir }) {
-  const dir = modelDirectory(modelsDir);
-  let loading = null;
-  const load = () => {
-    if (loading) return loading;
-    loading = (async () => {
-      if (!modelPresent(modelsDir)) throw new Error(`the embedding model is not present in ${modelsDir}; run npx tryharness to download it`);
-      const [{ Tokenizer }, ort] = await Promise.all([import("@huggingface/tokenizers"), import("onnxruntime-web")]);
-      const tokenizerJson = JSON.parse(fs.readFileSync(path.join(dir, "tokenizer.json"), "utf8"));
-      const tokenizerConfig = JSON.parse(fs.readFileSync(path.join(dir, "tokenizer_config.json"), "utf8"));
-      const modelConfig = JSON.parse(fs.readFileSync(path.join(dir, "config.json"), "utf8"));
-      const tokenizer = new Tokenizer(tokenizerJson, tokenizerConfig);
-      ort.env.wasm.numThreads = 1;
-      const session = await ort.InferenceSession.create(path.join(dir, "onnx", "model_quantized.onnx"), { executionProviders: ["wasm"] });
-      return {
-        ort, tokenizer, session,
-        maxLength: Number(tokenizerConfig.model_max_length) || 512,
-        padId: Number(modelConfig.pad_token_id ?? 0),
-        wantsTypeIds: session.inputNames.includes("token_type_ids"),
-      };
-    })();
-    loading.catch(() => { loading = null; });
-    return loading;
-  };
-
-  const embedBatch = async (texts) => {
-    const { ort, tokenizer, session, maxLength, padId, wantsTypeIds } = await load();
-    const encoded = texts.map((text) => {
-      const { ids, attention_mask, token_type_ids } = tokenizer.encode(text, { add_special_tokens: true, return_token_type_ids: true });
-      return { input_ids: [...ids], attention_mask: [...attention_mask], ...(wantsTypeIds ? { token_type_ids: [...(token_type_ids || ids.map(() => 0))] } : {}) };
-    });
-    const width = Math.min(Math.max(...encoded.map((e) => e.input_ids.length)), maxLength);
-    for (const e of encoded) {
-      if (e.input_ids.length > width) truncate(e, width);
-      else if (e.input_ids.length < width) pad(e, width, padId);
-    }
-    const tensor = (key) => new ort.Tensor("int64", BigInt64Array.from(encoded.flatMap((e) => e[key].map((v) => BigInt(v)))), [texts.length, width]);
-    const feeds = { input_ids: tensor("input_ids"), attention_mask: tensor("attention_mask"), ...(wantsTypeIds ? { token_type_ids: tensor("token_type_ids") } : {}) };
-    const outputs = await session.run(feeds);
-    const hidden = outputs.last_hidden_state;
-    const [, seq, dims] = hidden.dims;
-    const data = hidden.data;
-    return texts.map((_, row) => {
-      const start = row * seq * dims;
-      const vector = new Float32Array(dims);
-      let norm = 0;
-      for (let i = 0; i < dims; i += 1) { vector[i] = data[start + i]; norm += vector[i] * vector[i]; }
-      norm = Math.sqrt(norm) || 1;
-      for (let i = 0; i < dims; i += 1) vector[i] /= norm;
-      return vector;
-    });
-  };
-
-  return {
-    version: EMBEDDING_VERSION,
-    dimensions: EMBEDDING_DIMENSIONS,
-    ready: () => load().then(() => true),
-    async embedDocuments(texts) {
-      const out = [];
-      for (let i = 0; i < texts.length; i += BATCH) out.push(...(await embedBatch(texts.slice(i, i + BATCH))));
-      return out;
-    },
-    async embedQuery(text) {
-      return (await embedBatch([QUERY_PREFIX + text]))[0];
-    },
-  };
-}