@tiens.nguyen/gu-cli 1.0.734 → 1.0.736

package/coding-choice-tags.mjs

@@ -95,7 +95,46 @@ export function ollamaWhere(kind, url) {
   return (local ? "127.0.0.1" : host) + port;
 }
 
-/** The coder's name as shown: "ollama.com/kimi-k3" for an Ollama coder, the bare name otherwise. */
+/**
+ * WHICH SERVER A CODER RUNS ON — for EVERY coder, not only Ollama ones (#263). Reported: "you
+ * should show the server before the model, correctly — later I can have another model on a
+ * different server". A bare "claude-opus-5" cannot say which provider (or which of two gateways)
+ * it goes to; "api.anthropic.com/claude-opus-5" can.
+ *   · Ollama keeps ollamaWhere's rules (ollama.com, 127.0.0.1, a non-default port).
+ *   · Anything else: its host, this machine as 127.0.0.1, and the port when it is not the
+ *     scheme's default — two local servers on two ports are two servers.
+ */
+export function coderWhere(kind, url) {
+  if (kind === "ollama") return ollamaWhere(kind, url);
+  let u;
+  try {
+    u = new URL(String(url ?? ""));
+  } catch {
+    return "";
+  }
+  const host = u.hostname.toLowerCase().replace(/\.$/, "").replace(/^\[|\]$/g, "");
+  if (!host) return "";
+  const local = host === "localhost" || host === "127.0.0.1" || host === "::1";
+  const port = u.port ? `:${u.port}` : ""; // URL() already drops :443 on https and :80 on http
+  return (local ? "127.0.0.1" : host) + port;
+}
+
+/**
+ * WHERE THE CODER IN USE RUNS — the host the bar and banner put in front of its name (#263).
+ * The folder's pick (a choice id, or a bare model name), else the account default. No list (an
+ * older API) → "", never a guess from raw payload fields (key-choice-regression.test.mjs).
+ * One rule for startup and for every path that changes the coder: computing it in only one of them
+ * is what left "ollama.com/" in front of a newly added Claude model.
+ */
+export function whereOfActiveCoder(choices, sessionModel) {
+  const list = Array.isArray(choices) ? choices : [];
+  const pick = String(sessionModel ?? "");
+  const active =
+    (pick && list.find((c) => c.id === pick || c.model === pick)) || list.find((c) => c.isDefault) || null;
+  return active ? coderWhere(active.kind, active.url) : "";
+}
+
+/** The coder's name as shown: "<server>/<model>" ("api.anthropic.com/claude-opus-5"), bare when no server is known. */
 export function coderDisplayName(name, where) {
   const n = String(name ?? "").trim();
   return where && n ? `${where}/${n}` : n;

package/coding-provider-setup.mjs

@@ -233,6 +233,24 @@ export function modelListRequest(url, key = "") {
   return { url: modelsUrlFor(url), headers: keyed ? { Authorization: `Bearer ${k}` } : {} };
 }
 
+/**
+ * Why a provider's model list could not be shown, and what to type instead (#266). Reported:
+ * "did not list its models without a key — type the id … what id?". The line now says WHY
+ * (no key saved / key refused / unreachable) and WHAT to type: the model's exact name as that
+ * provider calls it, with the current one as the example.
+ *   reason: "no-key" | "refused" | "failed"
+ */
+export function describeListFailure({ host = "", reason = "failed", status = 0, keyName = "", example = "" } = {}) {
+  const what = `type the model's name exactly as ${host} calls it${example ? ` (e.g. ${example})` : ""}`;
+  if (reason === "no-key") {
+    return [`no API key is saved for ${host}, so it cannot list its models —`, `${what}, or add a key with ctrl+k first.`];
+  }
+  if (reason === "refused") {
+    return [`${host} refused the key${keyName ? ` "${keyName}"` : ""}${status ? ` (HTTP ${status})` : ""} when listing its models —`, `${what}, or fix the key with ctrl+k.`];
+  }
+  return [`couldn't list ${host}'s models${status ? ` (HTTP ${status})` : " (no answer)"} —`, `${what}.`];
+}
+
 /**
  * The hosted providers offered by name when adding one (#261), before "another URL". Claude has
  * no embedding models (Anthropic points to Voyage AI), so it is not offered for the embedder.

package/gonext_agent_chat.py

@@ -1152,6 +1152,18 @@ def _anthropic_usage(u) -> "types.SimpleNamespace":
                                  cache_read_input_tokens=cr, cache_creation_input_tokens=cw)
 
 
+def _log_anthropic_usage(u) -> None:
+    """One content-free line per Claude request with Anthropic's OWN cache figures (task #264).
+    Without it the log could not say whether caching worked — the reported question — although
+    every response carried the answer."""
+    g = (lambda n: int(_usage_field(u, n) or 0))
+    read, write, fresh = g("cache_read_input_tokens"), g("cache_creation_input_tokens"), g("input_tokens")
+    total = read + write + fresh
+    pct = f"{100 * read // total}%" if total else "—"
+    _log(f"claude usage: prompt {total} = cache read {read} + cache write {write} + uncached {fresh} "
+         f"({pct} read from cache) · output {g('output_tokens')}")
+
+
 class _AnthropicChatAdapter:
     """client.chat.completions.create() backed by Anthropic's native Messages API (task #262)."""
 
@@ -1176,6 +1188,7 @@ class _AnthropicChatAdapter:
         if kw.get("stream"):
             return self._stream(req)
         msg = self._sdk.messages.create(**req)
+        _log_anthropic_usage(getattr(msg, "usage", None))
         import types
         text = "".join(getattr(b, "text", "") or "" for b in (msg.content or []) if getattr(b, "type", "") == "text")
         return types.SimpleNamespace(
@@ -1217,6 +1230,7 @@ class _AnthropicChatAdapter:
                         usage["output_tokens"] = int(out)
         if stop_reason == "refusal":
             _log("claude: the model declined this request (stop_reason=refusal)")
+        _log_anthropic_usage(usage)
         yield chunk(content="", finish=_ANTHROPIC_FINISH.get(stop_reason, "stop"))
         yield chunk(usage=_anthropic_usage(usage))
 
@@ -8243,12 +8257,18 @@ def run_agent_chat(cfg):
     # Variable text — the deploy target and the current date/time — lives in the
     # variable tail next to the task instead.
     tool_hint = (
+        # ONE batching rule, the system prompt's (task #265): this line used to say "ONE code block
+        # that calls a SINGLE tool", contradicting the scaffold's BATCH rule from nearer the task —
+        # measured as a 102-step turn in which every file read was its own full round trip.
         f"Solve the TASK step by step (up to {max_steps} steps). At EACH step write a "
-        "brief Thought, then ONE code block that calls a SINGLE tool. You will SEE that "
-        "tool's result (Observation) before the next step — use it to decide what to do "
-        "next (e.g. web_search to find a URL, THEN fetch_url to read it). When you have "
-        "enough to answer, call final_answer(<your answer>) in a code block. Prefer "
-        "FEWER steps, and never repeat a tool call that already succeeded.\n\n"
+        "brief Thought, then ONE code block. BATCH INDEPENDENT LOOKUPS into that block — "
+        "several reads, greps or listings whose arguments you already know, printing each "
+        "result under a header (up to about 8). Make ONE CHANGE per block (an edit, a file "
+        "write or a command): if a line fails, the lines after it never run. Take a separate "
+        "step only when a call depends on a result you have not seen yet (e.g. web_search to "
+        "find a URL, THEN fetch_url to read it). You SEE the Observation before the next step. "
+        "When you have enough to answer, call final_answer(<your answer>) in a code block. "
+        "Prefer FEWER steps, and never repeat a tool call that already succeeded.\n\n"
         "{{TOOL_LIST}}\n"  # filled in below, once from the REAL registered tool objects
         + _rag_tool_block + _ws_tool_block + _auto_test_block +
         "\n"

package/gu-repl.mjs

@@ -120,7 +120,7 @@ import {
 } from "./terminal-layout.mjs";
 import { completedFlush } from "./completed-flush.mjs";
 import { codingBannerLabel, isOllamaCloudUrl, SECRET_KEY_HINT } from "./coding-key-prompt.mjs";
-import { codingChoiceTags, coderDisplayName, ollamaWhere } from "./coding-choice-tags.mjs";
+import { codingChoiceTags, coderDisplayName, coderWhere, whereOfActiveCoder } from "./coding-choice-tags.mjs";
 import {
   isDoubleClick,
   clickRun,
@@ -5242,6 +5242,7 @@ async function reconfigureOllamaServer() {
   // Persist the CLEARING too. The stored override names a model on the OLD box; leaving it on
   // disk would resurrect it at the next launch, against a server that has never heard of it.
   await saveSessionCoding(resolve(process.cwd()));
+  await syncActiveCoderWhere(); // the bar now names the account default — with ITS host (#263)
   // The GLOBAL output-token total is keyed to the coding backend (task #125), and the backend
   // just moved — re-fetch it for the new one, exactly as /model does after a pick, or the
   // footer keeps counting against the box that is no longer in use.
@@ -5355,7 +5356,7 @@ async function chooseCodingModelFrom(ps, all, question, current) {
   let list = coders;
   let showingAll = filtered === 0;
   for (;;) {
-    const rows = [...list, "type a model id…"];
+    const rows = [...list, "type a model name…"];
     // The row that lifts the filter, offered only while it is on and only when it hides something.
     if (!showingAll) rows.push(`show all ${all.length} models…`);
     // Opens on the CURRENT model where there is one — a long list that opens at the top tells the
@@ -5368,7 +5369,7 @@ async function chooseCodingModelFrom(ps, all, question, current) {
     const idx = await pickFromList(rows, at >= 0 ? at : 0);
     if (idx < 0) return null;
     if (idx === list.length) {
-      const typed = (await ask("  model id: ")).trim();
+      const typed = (await ask("  model name: ")).trim();
       return typed || null;
     }
     if (idx === list.length + 1) {
@@ -5394,13 +5395,13 @@ async function chooseCodingModelFrom(ps, all, question, current) {
  * unlisted one may be exactly right.
  */
 async function chooseFromModelList(all, question, current = 0) {
-  const rows = [...all, "type a model id…"];
+  const rows = [...all, "type a model name…"];
   const at = typeof current === "string" ? all.indexOf(current) : Number(current) || 0;
   console.log(dim(`  ${question}  (${PICK_HINT})`));
   const idx = await pickFromList(rows, at >= 0 ? at : 0);
   if (idx < 0) return null;
   if (idx === all.length) {
-    const typed = (await ask("  model id: ")).trim();
+    const typed = (await ask("  model name: ")).trim();
     return typed || null;
   }
   return all[idx];
@@ -5458,15 +5459,22 @@ async function changeProviderModel(ps, rows, p) {
   const md = await import("./model-doctor.mjs");
   const { proxyFetch } = await import("./proxy-dispatcher.mjs");
 
-  // PROBED WITHOUT A KEY, DELIBERATELY. The terminal does not hold this provider's secret, and
-  // the one key it could fetch is the ACCOUNT DEFAULT — which may belong to a different provider
-  // entirely. Sending it here to be helpful would be the exact leak this task spent a phase
-  // preventing. Many servers (Ollama, vLLM, LM Studio) list models unauthenticated; a hosted one
-  // will refuse, and then the id is typed, which always works.
+  // LISTED WITH THIS PROVIDER'S OWN KEY (#266). This used to probe with NO key, on purpose: the
+  // only key the terminal could fetch was the ACCOUNT DEFAULT, which may belong to another provider.
+  // Since #245 a key belongs to its server, so the key used is the one bound to THIS row — the same
+  // check addServer makes — and never another provider's. Keyed providers (Claude, OpenAI, Kimi,
+  // ollama.com) can finally list their models here instead of asking for an id.
+  const sk = await import("./server-keys.mjs");
+  const own = await withPendingStatus(`reading the key for ${p.host}`, async () => {
+    const keys = await fetchServerKeys(p.id);
+    const inUse = sk.keyInUse(keys ?? [], p.id, sessionCodingKeyId);
+    if (!inUse || inUse.backendId !== p.id) return { key: "", name: "" };
+    const r = await fetchCodingKey(p.id);
+    return r.keyId && r.keyId !== inUse.id ? { key: "", name: "" } : { key: r.key, name: String(inUse.name ?? "") };
+  });
   const probe = await withPendingStatus(`asking ${p.host} what it has`, async () => {
     try {
-      // Keyless, as above — but in the list's own dialect (the Claude API's headers, #261).
-      const lr = ps.modelListRequest(p.url);
+      const lr = ps.modelListRequest(p.url, own.key); // the list's own dialect too (Claude's headers, #261)
       const res = await proxyFetch(lr.url, { headers: lr.headers, signal: AbortSignal.timeout(10_000) });
       if (!res.ok) return { status: res.status };
       return { body: await res.json().catch(() => null) };
@@ -5486,7 +5494,11 @@ async function changeProviderModel(ps, rows, p) {
     const rn = ps.retiredNote(parsed.retired, p.host);
     if (rn) console.log(GUTTER + dim(rn));
   } else {
-    console.log(GUTTER + dim(`${p.host} did not list its models without a key — type the id.`));
+    const refused = probe.status === 401 || probe.status === 403;
+    const reason = refused ? (own.key ? "refused" : "no-key") : "failed";
+    for (const l of ps.describeListFailure({ host: p.host, reason, status: probe.status, keyName: own.name, example: p.model })) {
+      console.log(GUTTER + dim(l));
+    }
   }
 
   let model = "";
@@ -5498,7 +5510,7 @@ async function changeProviderModel(ps, rows, p) {
     }
   } else {
     console.log(GUTTER + dim(`Enter keeps ${p.model || "it unset"}`));
-    model = (await ask("  model id: ")).trim();
+    model = (await ask("  model name: ")).trim();
   }
   if (!model) {
     console.log(GUTTER + dim("(nothing changed)\n"));
@@ -5533,6 +5545,7 @@ async function changeProviderModel(ps, rows, p) {
     sessionCodingLabel = "";
     sessionCodingUrl = "";
     await saveSessionCoding(resolve(process.cwd()));
+    await syncActiveCoderWhere();
     console.log(GUTTER + dim("this folder's pick named the old model — cleared it."));
   }
   console.log(accent(`  ✓ ${p.host} → ${model}`) + dim("  ·  pick it in /model\n"));
@@ -5625,6 +5638,7 @@ async function removeCodingProvider(ps, rows, p) {
     sessionCodingLabel = "";
     sessionCodingUrl = "";
     await saveSessionCoding(resolve(process.cwd()));
+    await syncActiveCoderWhere();
     console.log(GUTTER + dim("this folder was using it — cleared the pick.\n"));
   }
   console.log(
@@ -5834,7 +5848,7 @@ async function addServer({ role = "coding", dialect = "openai", url: presetUrl =
         return;
       }
     } else {
-      model = String((await ask("  model id: ")) ?? "").trim();
+      model = String((await ask("  model name: ")) ?? "").trim();
     }
     if (!model) {
       console.log(GUTTER + dim("(cancelled — nothing was saved)\n"));
@@ -5979,6 +5993,9 @@ async function useCoderInThisFolder({ ps, rows, beforeChoices, url, host, locali
   sessionCodingModel = offered.id;
   sessionCodingLabel = model;
   sessionCodingUrl = url;
+  // …and the host the bar puts in front of it (#263): left alone, the PREVIOUS coder's
+  // "ollama.com/" stayed in front of a newly added Claude model until the next /model pick.
+  activeCoderWhere = whereOfActiveCoder(after?.codingChoices, sessionCodingModel);
   // The folder follows the SERVER's key (task #245) — its default, which a first key already is.
   sessionCodingKeyId = "";
   sessionApiKeyTail = key ? key.slice(-4) : "";
@@ -6072,9 +6089,9 @@ async function chooseModel() {
   // OpenAI-compatible providers are both "OpenAI-compatible", so the kind names neither.
   const { tags, show: showTag } = codingChoiceTags(choices);
   const labels = choices.map((c, i) => {
-    // An Ollama row says WHERE it runs in front of the name ("ollama.com/kimi-k3"), so the kind tag
-    // beside it would only repeat "Ollama". Other rows keep the tag rule.
-    const where = ollamaWhere(c.kind, c.url);
+    // Every row says WHERE it runs in front of the name ("ollama.com/kimi-k3",
+    // "api.anthropic.com/claude-opus-5" — #263), so the kind tag is only for a row with no server.
+    const where = coderWhere(c.kind, c.url);
     const tag = showTag && !where ? `  ${dim(`· ${tags[i]}`)}` : "";
     return `${coderDisplayName(c.model, where)}${tag}${c.id === def ? "  (default)" : ""}`;
   });
@@ -6119,10 +6136,12 @@ async function chooseModel() {
   // Task #125: remember the chosen backend's coding URL so the GLOBAL output-token total is
   // keyed to it. Default pick → "" so activeCodingUrl() falls back to defaultCodingUrl.
   sessionCodingUrl = chosen.id === def ? "" : (chosen.url ?? "");
+  // The server goes WITH the model, before any await (#263): set after the save below, the bar
+  // could repaint in between and show the NEW model under the OLD server for a moment.
+  activeCoderWhere = coderWhere(chosen.kind, chosen.url);
   // WRITTEN NOW, not at the end of the next turn: `/model` then `/exit` is how the choice is
   // made and then tested, and that path never reaches saveSession.
   await saveSessionCoding(resolve(process.cwd()));
-  activeCoderWhere = ollamaWhere(chosen.kind, chosen.url);
   const where = showTag && !activeCoderWhere ? ` (${tags[idx]})` : "";
   console.log(
     accent(
@@ -8644,6 +8663,16 @@ let defaultCodingModel = "";
 let sessionCodingUrl = "";
 /** "ollama.com" / "127.0.0.1" / a host — where the ACTIVE coder runs, when it is an Ollama one. */
 let activeCoderWhere = "";
+/**
+ * Re-read where the coder in use runs (#263) — for the paths that change it without holding the
+ * API's choice list (they clear the folder's pick back to the account default). Best-effort: an
+ * unreachable API leaves no prefix rather than a stale one.
+ */
+async function syncActiveCoderWhere() {
+  const payload = await fetchAgentPayload([{ role: "user", content: "ping" }]).catch(() => null);
+  activeCoderWhere = whereOfActiveCoder(payload?.codingChoices, sessionCodingModel);
+  if (barPinned()) paintBar();
+}
 // The account-default backend's URL (from the startup probe's codingChoices), used when no
 // /model override is active. Set once at startup alongside defaultCodingModel.
 let defaultCodingUrl = "";
@@ -10527,14 +10556,9 @@ async function main() {
     }
     {
       // WHERE THE CODER RUNS, for the banner and the bar: the folder's pick, else the default.
-      const choices = Array.isArray(probe?.codingChoices) ? probe.codingChoices : [];
-      const active =
-        (sessionCodingModel && choices.find((c) => c.id === sessionCodingModel || c.model === sessionCodingModel)) ||
-        choices.find((c) => c.isDefault) ||
-        null;
       // No choice list (an older API): no prefix, rather than guess from the ACCOUNT default's raw
       // payload fields, which key-choice-regression.test.mjs forbids for exactly that reason.
-      activeCoderWhere = active ? ollamaWhere(active.kind, active.url) : "";
+      activeCoderWhere = whereOfActiveCoder(probe?.codingChoices, sessionCodingModel);
     }
     // Task #133: no RAG storage question any more — the terminal ALWAYS keeps the index on
     // this machine. `/rag-cloud` opts into additionally mirroring it to the user's S3 bucket.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gu-cli",
-  "version": "1.0.734",
+  "version": "1.0.736",
   "description": "Gu CLI — the gu terminal plus the local worker that runs agent / OCR / PDF / embedding jobs on your Mac (Ollama / MLX / OpenAI-compatible).",
   "type": "module",
   "license": "MIT",