@tiens.nguyen/gu-cli 1.0.704 → 1.0.705

package/gonext_agent_chat.py

@@ -1329,6 +1329,11 @@ def _render_numbered_tool_list(tools) -> str:
     return ("\n".join(lines) + "\n") if lines else ""
 
 
+# A step line only `gu --verbose` shows (#45): retry notices and context trimming are how the
+# agent keeps itself going, and the user can do nothing with them.
+_VERBOSE_PREFIX = "__VERBOSE__::"
+
+
 def _summarise_step(step_log):
     """Return a short human-readable description of an agent step."""
     tool_calls = getattr(step_log, "tool_calls", None) or []
@@ -1337,6 +1342,9 @@ def _summarise_step(step_log):
     step_num = getattr(step_log, "step_number", None)
 
     parts = []
+    # A retry the agent recovers from by itself is machinery, not progress: the terminal shows
+    # it only under `gu --verbose` (#45). The model still gets smolagents' exact error text.
+    verbose_only = False
     for tc in tool_calls:
         name = getattr(tc, "name", "")
         args = getattr(tc, "arguments", None)
@@ -1404,15 +1412,18 @@ def _summarise_step(step_log):
             # step loop), so the model keeps the precise wording it needs to self-correct.
             _msg = "the model wrote no code this step — asking again"
             parts.append(f"→ {_msg}" if parts else _msg)
+            verbose_only = True
         elif "code parsing failed" in err.lower() or "syntaxerror" in err.lower():
             # The sibling shape: there WAS a block, and its body would not compile.
             _msg = "the model's code wouldn't run — asking again"
             parts.append(f"→ {_msg}" if parts else _msg)
+            verbose_only = True
         else:
             parts.append(f"→ Error: {_clip(err, 160)}")
 
     # No numeric "Step N:" prefix — show only the semantic action.
-    return (" | ".join(parts) if parts else "thinking…")
+    text = " | ".join(parts) if parts else "thinking…"
+    return f"{_VERBOSE_PREFIX}{text}" if verbose_only else text
 
 
 # Keywords that strongly indicate the user wants to make an HTTP/network request,
@@ -8689,7 +8700,7 @@ def run_agent_chat(cfg):
             # once, and that answer is remembered. The line exists so the figure is not mistaken
             # for something the endpoint reported — see the "measured vs remembered" note above.
             _emit({"type": "step", "text": (
-                f"context window {_ctx_window // 1000}k for {coding_model_id} (published default — "
+                f"{_VERBOSE_PREFIX}context window {_ctx_window // 1000}k for {coding_model_id} (published default — "
                 f"this endpoint does not report one). Corrects itself only if it is too HIGH; "
                 f"/context-window to set it yourself")})
     else:
@@ -8705,7 +8716,7 @@ def run_agent_chat(cfg):
         # command immediately and cannot act on an env var without leaving, editing a file and
         # restarting — which is how this stayed unset long enough to matter.
         _emit({"type": "step", "text": (
-            f"context window unknown — compacting at ~{_compact_at // 1000}k tokens; "
+            f"{_VERBOSE_PREFIX}context window unknown — compacting at ~{_compact_at // 1000}k tokens; "
             "run /context-window <tokens> to use more")})
 
     # Task #90 Phase 3 — completion gate (code-enforced, complements the prompt). Fires at
@@ -8958,7 +8969,7 @@ def run_agent_chat(cfg):
                         if _cs and not _compact_state["told"]:
                             _compact_state["told"] = True
                             _emit({"type": "step", "text": (
-                                f"Context reached ~{_in_tok // 1000}k tokens — trimming older "
+                                f"{_VERBOSE_PREFIX}Context reached ~{_in_tok // 1000}k tokens — trimming older "
                                 "steps to stay fast (files are re-readable)")})
                         if _cs:
                             _log(f"context compact: saved ~{_cs} chars "
@@ -9477,9 +9488,9 @@ def run_agent_chat(cfg):
                 # to say. On empty-content there is genuinely nothing to show — the model spent
                 # the turn in its reasoning channel and wrote nothing — so a short line stays.
                 _emit({"type": "step", "text": (
-                    "The model thought but didn't act — asking it again…"
+                    f"{_VERBOSE_PREFIX}The model thought but didn't act — asking it again…"
                     if why == "empty-content"
-                    else "Asking the model to take the action it described…")})
+                    else f"{_VERBOSE_PREFIX}Asking the model to take the action it described…")})
                 retry_kwargs = dict(completion_kwargs)
                 retry_kwargs["messages"] = list(completion_kwargs.get("messages") or []) + [
                     {"role": "user", "content": self._CODE_NOW_DIRECTIVE}

package/gu-repl.mjs

@@ -137,6 +137,7 @@ import { REQUIRED_AGENT_MODEL, REQUIRED_AGENT_PREFIX, REQUIRED_EMBED_MODEL } fro
 // plain SGR code: terminals disagree about colour, so accent.mjs picks 24-bit, 256 or the
 // sixteen-colour magenta from what this one actually claims to support.
 import { accent } from "./accent.mjs";
+import { isNoCodeRetry, splitVerboseLine, stepHasCode } from "./verbose-lines.mjs";
 import {
   createTranscript,
   displayRows,
@@ -8466,6 +8467,16 @@ async function runAgentTurn(history, { attachJobId = "" } = {}) {
   // 400 chars — enough for a one-line label, not for the paragraphs we now print). Bounded so a
   // runaway generation cannot grow without limit; the prose we keep is far shorter anyway.
   let stepThought = "";
+  // #45: the prose of a step whose output had NO code, held back until the next line says what
+  // became of it. A no-code retry notice discards it (the model restates the plan on the retry);
+  // anything else prints it. Always null under --verbose, which shows everything as it comes.
+  let heldThought = null;
+  const releaseHeldThought = () => {
+    if (heldThought === null) return;
+    const raw = heldThought;
+    heldThought = null;
+    if (printThoughtBlock(raw, onContent)) lastWasBlank = false;
+  };
   let hoverRow = -1; // last mouse viewport row (from motion events) → highlight when it's on the thought line
   let thoughtRow = -1; // absolute viewport row of the thought line (line 1), learned via \x1b[6n
   let lastDsrAt = 0; // throttle the cursor-position probe
@@ -8894,7 +8905,7 @@ async function runAgentTurn(history, { attachJobId = "" } = {}) {
     carry += stripThinkTags(chunk);
     let nl;
     while ((nl = carry.indexOf("\n")) >= 0) {
-      const line = carry.slice(0, nl);
+      let line = carry.slice(0, nl);
       carry = carry.slice(nl + 1);
       const alreadyFlushed = carryFlushedLen;
       carryFlushedLen = 0; // reset — `carry` now starts a fresh pending line
@@ -8907,6 +8918,21 @@ async function runAgentTurn(history, { attachJobId = "" } = {}) {
         almostDone = /almost/i.test(line);
         continue;
       }
+      // #45: retry notices and context-window bookkeeping are the agent's own machinery. Shown
+      // under --verbose only; otherwise swallowed, together with the no-code false start the
+      // retry notice is about.
+      {
+        const v = splitVerboseLine(line);
+        if (v.verboseOnly) {
+          if (!REPL_VERBOSE) {
+            if (isNoCodeRetry(v.text)) heldThought = null;
+            lastContentAt = Date.now();
+            continue;
+          }
+          line = v.text;
+        }
+      }
+      if (heldThought !== null && line.trim() !== "") releaseHeldThought();
       // Task #136: a RAG index file was written. Swallow the line — it belongs in the
       // BOTTOM BAR, not in the action list — and remember the newest one. Kept out of the
       // bullet stream deliberately: indexing writes a shard per batch, so printing them
@@ -8939,7 +8965,11 @@ async function runAgentTurn(history, { attachJobId = "" } = {}) {
           // The step's model output is complete. Print its reasoning HERE, before the tools
           // report what they did, so the transcript reads in the order it happened: the model
           // says what it intends, then the actions it took appear under it.
-          if (printThoughtBlock(stepThought, onContent)) lastWasBlank = false;
+          if (!REPL_VERBOSE && stepThought.trim() && !stepHasCode(stepThought)) {
+            heldThought = stepThought; // #45: decided by the next line — see releaseHeldThought
+          } else if (printThoughtBlock(stepThought, onContent)) {
+            lastWasBlank = false;
+          }
           stepThought = "";
         }
         continue;
@@ -9328,6 +9358,7 @@ async function runAgentTurn(history, { attachJobId = "" } = {}) {
             lastWasBlank = false;
           }
         }
+        releaseHeldThought(); // #45: nothing followed it, so it was not a false start
         clearStatus();
         // Task #104: expose this turn's code-model INPUT + the CUMULATIVE output-token
         // total (all turns for this user + coding server) so the caller can print a

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gu-cli",
-  "version": "1.0.704",
+  "version": "1.0.705",
   "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",
@@ -60,6 +60,7 @@
     "job-pools.mjs",
     "rag-job-config.mjs",
     "completed-flush.mjs",
+    "verbose-lines.mjs",
     "agent-model-label.mjs",
     "coding-choice-tags.mjs",
     "coding-provider-setup.mjs",

package/verbose-lines.mjs

@@ -0,0 +1,33 @@
+/**
+ * Step lines only `gu --verbose` shows (#45).
+ *
+ * The agent keeps itself going with retries ("the model wrote no code this step — asking
+ * again") and context bookkeeping ("context window 128k … published default", "Context reached
+ * ~40k tokens — trimming older steps"). The user can do nothing with either, so the agent tags
+ * them with this prefix and the terminal drops them unless it runs with --verbose.
+ */
+export const VERBOSE_PREFIX = "__VERBOSE__::";
+
+/** Untag a line: `{ verboseOnly, text }`. Lines without the prefix come back unchanged. */
+export function splitVerboseLine(line) {
+  const s = String(line ?? "");
+  return s.startsWith(VERBOSE_PREFIX)
+    ? { verboseOnly: true, text: s.slice(VERBOSE_PREFIX.length) }
+    : { verboseOnly: false, text: s };
+}
+
+/**
+ * The notice that the model's LAST reply carried no code at all. That reply's prose is then a
+ * false start: the model is asked again and restates it, so printing both shows the same plan
+ * twice around a notice the user does not need.
+ */
+export function isNoCodeRetry(text) {
+  return /wrote no code this step|thought but didn't act|take the action it described/i.test(
+    String(text ?? "")
+  );
+}
+
+/** Did a step's raw model output contain a code block? */
+export function stepHasCode(raw) {
+  return /<code[\s>]/i.test(String(raw ?? ""));
+}