@tiens.nguyen/gonext-local-worker 1.0.332 → 1.0.333

package/gonext-local-worker.mjs

@@ -1832,8 +1832,7 @@ function _isLocalCodingUrl(baseURL) {
 /** Resolve to "ollama" | "openai" | "local". `kind` is the user's Settings choice
  *  ("ollama" | "openai" | "" = Auto). Only Auto costs a probe. */
 async function resolveCodingBackend(kind, baseURL) {
-  if (kind === "ollama") return "ollama";
-  if (kind === "openai") return "openai";
+  if (kind === "ollama" || kind === "openai" || kind === "local") return kind;
   if (await _probeIsOllama(baseURL)) return "ollama";
   return _isLocalCodingUrl(baseURL) ? "local" : "openai";
 }
@@ -1872,8 +1871,9 @@ async function runAgentChatJob(job) {
   const cancelAc = new AbortController();
   let cancelled = false;
   // Task #108: the user's Coding backend setting ("ollama" | "openai" | "" = Auto).
-  const codingKind =
-    payload?.codingKind === "ollama" || payload?.codingKind === "openai" ? payload.codingKind : "";
+  const codingKind = ["ollama", "openai", "local"].includes(payload?.codingKind)
+    ? payload.codingKind
+    : "";
   const runRes = await workerFetch(`/api/worker/jobs/${jobId}`, {
     method: "PATCH",
     body: JSON.stringify({ jobStatus: "running" }),

package/gonext-repl.mjs

@@ -876,32 +876,52 @@ async function chooseModel() {
   }
   const allowed = Array.isArray(probe?.codingAllowed) ? probe.codingAllowed : [];
   const def = (probe?.codingDefault ?? "").trim();
-  if (allowed.length === 0) {
+  // Task #123: a choice is a (backend, model) PAIR — several backends (Ollama, an
+  // OpenAI-compatible endpoint, local MLX) can be enabled at once, and picking one moves
+  // the URL, the kind and the API key with it. `codingChoices` carries the backend of each
+  // entry so the list can say which is which; `codingAllowed` is the same ids, and stays
+  // the fallback for an API that predates this.
+  const choices = Array.isArray(probe?.codingChoices) && probe.codingChoices.length
+    ? probe.codingChoices
+    : allowed.map((id) => ({ id, model: id, kind: "" }));
+  if (choices.length === 0) {
     console.log(dim("  No coding model configured. Set one in the web app → Settings → Agent.\n"));
     return;
   }
-  if (allowed.length === 1) {
+  if (choices.length === 1) {
     console.log(
-      dim(`  Only one coding model is available: ${allowed[0]}.\n`) +
-        dim("  Add more in the web app → Settings → Agent → Load Models (checkboxes).\n")
+      dim(`  Only one coding model is available: ${choices[0].model}.\n`) +
+        dim("  Add more in the web app → Settings → Agent → Coding backends.\n")
     );
     return;
   }
   const active = sessionCodingModel || def;
-  // Arrow-key picker (↑/↓ + Enter) — pre-highlight the active model, tag the default.
-  const labels = allowed.map((m) => (m === def ? `${m}  (default)` : m));
-  const activeIdx = allowed.findIndex((m) => m === active);
+  // Arrow-key picker (↑/↓ + Enter) — pre-highlight the active model, tag the default, and
+  // name the backend when more than one is in play (the same model name can appear twice).
+  const KIND_LABEL = { ollama: "Ollama", openai: "OpenAI-compatible", local: "local MLX", "": "auto" };
+  const showKind = new Set(choices.map((c) => c.kind ?? "")).size > 1;
+  const labels = choices.map((c) => {
+    const kind = showKind ? `  ${dim(`· ${KIND_LABEL[c.kind ?? ""] ?? c.kind}`)}` : "";
+    return `${c.model}${kind}${c.id === def ? "  (default)" : ""}`;
+  });
+  const activeIdx = choices.findIndex((c) => c.id === active);
   console.log(dim("  Choose the coding model  (↑/↓ then Enter):"));
   const idx = await pickFromList(labels, activeIdx >= 0 ? activeIdx : 0);
   if (idx < 0) {
     console.log(dim("  (cancelled)\n"));
     return;
   }
-  const chosen = allowed[idx];
+  const chosen = choices[idx];
   // Empty override means "use the account default" — so don't send an override when the
   // user picks the default (keeps the payload clean and lets a later default change win).
-  sessionCodingModel = chosen === def ? "" : chosen;
-  console.log(green(`  ✓ coding model → ${chosen}${chosen === def ? " (default)" : ""}\n`));
+  sessionCodingModel = chosen.id === def ? "" : chosen.id;
+  // Remember the friendly name for the token footer, which should never show a raw
+  // "<kind>::<model>" wire id.
+  sessionCodingLabel = chosen.id === def ? "" : chosen.model;
+  const where = showKind ? ` (${KIND_LABEL[chosen.kind ?? ""] ?? chosen.kind})` : "";
+  console.log(
+    green(`  ✓ coding model → ${chosen.model}${where}${chosen.id === def ? " (default)" : ""}\n`)
+  );
 }
 
 // Probe whether SSH KEY auth already works for a server (so we can tell the user to run
@@ -1252,7 +1272,12 @@ let jobIdForApprovalRace = "";
 
 // Per-session coding-model override chosen via /model (empty = use the account default).
 // Sent to /agent-ask as codingModelOverride; the API validates it against the whitelist.
+// Task #123: with several backends configured this is a QUALIFIED id ("<kind>::<model>"),
+// which resolves to a whole backend — URL, kind and API key — not just a model name.
 let sessionCodingModel = "";
+// The friendly model name behind sessionCodingModel, for display only: the token footer
+// must never show a raw "<kind>::<model>" wire id. Empty = running the account default.
+let sessionCodingLabel = "";
 // The default coding model id (from the startup agent-payload probe) — shown in the
 // per-turn token footer next to the GLOBAL output total so it's clear which model that
 // lifetime count belongs to. The /model override (sessionCodingModel) wins when set.
@@ -2383,7 +2408,7 @@ async function main() {
         // name so it's clear which coder that running total belongs to. Guard on THIS turn's
         // activity so a plain-reply greeting that did no code-model work shows nothing (#112).
         if ((inputTokens ?? 0) > 0 || (turnOutputTokens ?? 0) > 0) {
-          const coder = sessionCodingModel || defaultCodingModel;
+          const coder = sessionCodingLabel || defaultCodingModel;
           console.log(
             dim("tokens · ↑ in ") +
               fmtTokens(inputTokens ?? 0) +

package/gonext_agent_chat.py

@@ -949,10 +949,8 @@ def _resolve_coding_backend(kind, base_url, probe=None):
     `probe` is injectable so the resolution table can be replayed offline without network.
     """
     k = (kind or "").strip().lower()
-    if k == "ollama":
-        return "ollama", "explicit setting"
-    if k == "openai":
-        return "openai", "explicit setting"
+    if k in ("ollama", "openai", "local"):
+        return k, "explicit setting"
     if (probe or _is_ollama_server)(base_url):
         return "ollama", "auto: server answers as Ollama"
     if not _is_local_url(base_url):
@@ -3620,7 +3618,7 @@ def run_agent_chat(cfg):
     # coding_api_key: the Bearer key for an OpenAI-compatible coder (empty otherwise). Used
     # for the code-model calls (and coding-model auto-detect), NOT the chat agent_api_key.
     coding_kind = (cfg.get("codingKind") or "").strip().lower()
-    if coding_kind not in ("ollama", "openai"):
+    if coding_kind not in ("ollama", "openai", "local"):
         coding_kind = ""
     coding_api_key = (cfg.get("codingApiKey") or "").strip()
     # Task #107: when on, emit each code-model response so the worker persists it to Mongo.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.332",
+  "version": "1.0.333",
   "description": "Polls GoNext cloud API for async local LLM jobs and runs them against Ollama/OpenAI-compatible servers on this Mac",
   "type": "module",
   "license": "MIT",