@tiens.nguyen/gu-cli 1.0.812 → 1.0.813

package/gonext_agent_chat.py

@@ -4769,6 +4769,42 @@ def _collapse_inner_openers(text):
     return text[:body_from] + cleaned + "\n</code>"
 
 
+# DeepSeek DSML tool-call markup (issue #119). Some DeepSeek-family models (seen via
+# token-plan.maas.qwencloudapi.com) answer with their NATIVE tool-call text instead of our code
+# block: prose, then a DSML "calls" tag, a DSML invoke tag with name="code", then plain Python,
+# then closers that are often wrong (a DSML "parameter" close) or missing. The bars in the tags
+# are FULLWIDTH, so the patterns tolerate any short run of non-bracket characters around DSML.
+_DSML_INVOKE_CODE = re.compile(
+    r"<[^<>\n]{0,8}DSML[^<>\n]{0,8}invoke\s+name\s*=\s*[\x22\x27]code[\x22\x27]\s*>[ \t]*\r?\n?"
+)
+_DSML_ANY = re.compile(r"<[^<>\n]{0,8}DSML")
+_DSML_CLOSER = re.compile(r"\n?[ \t]*</[^<>\n]{0,8}DSML")
+
+
+def _dsml_to_code_block(text):
+    """Rewrite a DSML invoke name="code" body into a normal code block, or return the text
+    unchanged. Fires only when there is NO code tag already and the body COMPILES (the same
+    conservative contract as the other readings)."""
+    if not text or "DSML" not in text or re.search(r"<\s*/?\s*code\b", text, re.I):
+        return text
+    m = _DSML_INVOKE_CODE.search(text)
+    if not m:
+        return text
+    rest = text[m.end():]
+    c = _DSML_CLOSER.search(rest)
+    body = (rest[:c.start()] if c else rest).strip("\n")
+    if not body.strip():
+        return text
+    try:
+        compile(body, "<gu-code>", "exec")
+    except Exception:  # noqa: BLE001
+        return text
+    head = text[:m.start()]
+    first = _DSML_ANY.search(head)
+    prose = (head[:first.start()] if first else head).rstrip()
+    return (prose + "\n" if prose else "") + "<code>\n" + body + "\n</code>"
+
+
 def _code_block_readings(text):
     """Every plausible reading of where the code block is, CHEAPEST FIRST.
 
@@ -4829,6 +4865,10 @@ def _canonicalise_code_reply(text):
     # reproduced, inside the fix, the very shape of the bug being fixed: a strict cheap check
     # guarding a tolerant repair. If a pre-check cannot be as permissive as what it guards, it
     # does not belong in front of it.
+    _dsml = _dsml_to_code_block(text)
+    if _dsml != text:
+        _log("DSML tool-call markup converted to a code block (#119)")
+        return _dsml
     if not text or not _CODE_ANY_TAGGISH.search(text):
         return text
     fallback = None
@@ -11825,7 +11865,8 @@ def run_agent_chat(cfg):
                         if _still_bad:
                             self._parse_fail_streak += 1
                             _log("tool-call code still unparseable after repair "
-                                 f"(streak {self._parse_fail_streak}/3)")
+                                 f"(streak {self._parse_fail_streak}/3) model={getattr(self, 'model_id', '?')} "
+                                 f"reply[:300]={(msg.content or '')[:300]!r}")
                             if self._parse_fail_streak >= 3:
                                 _log("parse-fail streak hit limit → ending the turn honestly "
                                      "instead of looping on an unparseable tool call")

package/gu-repl.mjs

@@ -61,6 +61,7 @@ import {
   thoughtFull,
   thoughtBlockRows,
   ragSegmentLabel,
+  ragBarNote,
   searchSegmentLabel,
   runningDetailRows,
   formatThinkTime,
@@ -2276,9 +2277,13 @@ const ragSegmentText = () => {
   // RAG FAILED MID-TURN and searches were answered by the grep fallback: a WARNING (yellow),
   // always visible — the turn still got results, but not semantic ones.
   const liveFallback = b.ragState === "fallback" && !b.stale;
-  if (!REPL_VERBOSE && !liveDown && !liveFallback) return lexSeg;
+  // Task #34: a LIVE grep fallback (embedder down, or RAG failed mid-turn) is no longer on this
+  // row. In red with a "✗" it read as a serious fault while search kept working. It is now a soft
+  // yellow note on the bottom bar's mode row (ragBarNote → footerRows' modeNote).
+  if (liveDown || liveFallback) return lexSeg;
+  if (!REPL_VERBOSE) return lexSeg;
   if (b.stale) return dim(" · ") + dim(`${label}·`) + lexSeg;
-  return dim(" · ") + (liveDown ? red(label) : liveFallback ? yellow(label) : dim(label)) + lexSeg;
+  return dim(" · ") + dim(label) + lexSeg;
 };
 /**
  * IS ANYTHING NEWER? (task #181) — cached, never probed from a repaint.
@@ -2729,6 +2734,10 @@ const footerRows = () =>
     identity: identityLineText(),
     token: tokenLineText(),
     mode: modeLineText(),
+    // Task #34: "rag: embedder down → grep fallback" as a soft yellow note beside the mode, not an
+    // alarming red segment on the token row. yellow goes through paint(), so NO_COLOR turns it off.
+    modeNote: ragBarNote(lastPromptBreakdown),
+    notePaint: yellow,
     // How to paste here (issue #67): the platform's text key, plus gu's image key when the coder
     // can take an image.
     modeHint: pasteHint({ images: coderTakesImages() }),
@@ -3540,6 +3549,7 @@ function printThoughtBlock(raw, before, { gap = false } = {}) {
       collapsed: true,
       lines: THOUGHT_COLLAPSED_LINES,
     });
+    const rowsBeforeMerge = screenRows().length;
     const newEnd = transcript.replaceRange(prev.from, prev.to, grown.map((r) => GUTTER + dim(r)));
     if (newEnd !== null) {
       prev.prose = merged;
@@ -3548,7 +3558,7 @@ function printThoughtBlock(raw, before, { gap = false } = {}) {
       // move under the pointer, so the cached hover row must be dropped — same reason
       // toggleThoughtBlock does it after a resize of the block.
       hoverCtl.reset();
-      paintLiveTail();
+      repaintTailOrHoldView(rowsBeforeMerge); // #113/#35: a finished thought never pulls a reader down
       return true;
     }
     // Scrolled out of the buffer — fall through and print a new block rather than lose it.
@@ -3855,6 +3865,21 @@ function toggleNewestThought() {
   return toggleThoughtBlock(block);
 }
 
+/**
+ * #113 / #35 — A BLOCK AT THE TAIL CHANGED SIZE (a thought merged, a step group grew). At the
+ * live tail that is a paintLiveTail. Scrolled back, it is NOT: paintLiveTail sets scrollOffset
+ * to 0, which is exactly the "it finished thinking and threw me back to the bottom" report. The
+ * change happened below (or at the edge of) what the reader is looking at, so keep the same
+ * lines on screen — the offset moves by the rows added — and let the jump notice count up.
+ */
+function repaintTailOrHoldView(rowsBefore) {
+  if (barPinned() && scrolledBack()) {
+    scrollOffset = Math.max(1, scrollOffset + (screenRows().length - rowsBefore));
+    renderScrollback(); // same slice as before; ends in paintBar (the "N lines below" count)
+    return;
+  }
+  paintLiveTail();
+}
 /** Back to the live tail: repaint it from the transcript and hand the screen back. */
 function jumpToBottom() {
   if (!barPinned() || !scrollOffset) return; // already live — nothing to come back from
@@ -11876,13 +11901,16 @@ async function runAgentTurn(history, { attachJobId = "", ownJob = false } = {})
     if (stepGroupStillOpen()) {
       stepGroup.items.push(text);
       const painted = paintThoughtRows(stepGroup, hoverCtl.hovered === stepGroup);
+      const rowsBeforeStep = screenRows().length;
       const newEnd = transcript.replaceRange(stepGroup.from, stepGroup.to, painted);
       if (newEnd !== null) {
         const delta = newEnd - stepGroup.to;
         stepGroup.to = newEnd;
         if (delta === 0) {
           // Collapsed: one row before, one row after — rewrite that row and nothing else.
-          ui(thoughtBlockRepaintSeq(stepGroup));
+          // Scrolled back (#113/#35): those rows are live-tail coordinates, NOT the history the
+          // reader is looking at — painting them would draw over it. The transcript has it.
+          if (!scrolledBack()) ui(thoughtBlockRepaintSeq(stepGroup));
           paintBar();
         } else {
           // Open (the user clicked it mid-turn): it grew, and what is under it moved down.
@@ -11893,7 +11921,7 @@ async function runAgentTurn(history, { attachJobId = "", ownJob = false } = {})
             }
           }
           hoverCtl.reset();
-          paintLiveTail();
+          repaintTailOrHoldView(rowsBeforeStep); // #113/#35: never pull a reader down
         }
         return;
       }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gu-cli",
-  "version": "1.0.812",
+  "version": "1.0.813",
   "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",
@@ -40,6 +40,7 @@
     "test:cache-e2e": "GONEXT_OLLAMA1_LIVE=1 node --test --test-timeout=1800000 tests/prefix-cache-e2e.test.mjs",
     "test:tokens": "python3 -m unittest discover tests -p \"test_cached_tokens.py\" -v && node --test tests/prefix-cache.test.mjs tests/agent-turn-shape.test.mjs",
     "test:selection": "node --test tests/selection-scrollback.test.mjs tests/selection-while-thinking.test.mjs tests/terminal-selection.test.mjs tests/terminal-copy.test.mjs tests/terminal-viewport.test.mjs tests/hover-repaint.test.mjs",
+    "test:autoscroll": "node --test tests/no-autoscroll-hide-total.test.mjs tests/thought-spacing.test.mjs tests/terminal-viewport.test.mjs tests/selection-scrollback.test.mjs",
     "test:steps": "python3 scripts/py-tap.py tests/test_read_caps.py tests/test_trim_payback.py tests/test_explore.py tests/test_and_chains.py tests/test_edit_file_primary.py tests/test_bare_code_rescue.py tests/test_batching_rule.py tests/test_no_pdf_routing.py",
     "check:image": "node scripts/image-paste-check.mjs",
     "check:image-clipboard": "node scripts/image-paste-check.mjs --clipboard",

package/terminal-layout.mjs

@@ -146,8 +146,9 @@ export function ragSegmentLabel(b) {
   const rag = Number.isFinite(b.rag) ? b.rag : 0;
   if (rag > 0) return null; // …the caller renders the number + percent instead
   switch (b.ragState) {
+    // Task #34: no "✗" — search still works through grep, so this is a notice, not an error.
     case "unreachable":
-      return "rag ✗ embedder down → grep fallback";
+      return "rag: embedder down → grep fallback";
     case "no-index":
       return "rag — not indexed";
     case "unavailable":
@@ -167,7 +168,7 @@ export function ragSegmentLabel(b) {
     // RAG FAILED MID-TURN (indexing or query embedding) and the search was answered by the
     // keyword fallback instead. A warning, not an error: the turn still got results.
     case "fallback":
-      return "rag ⚠ unavailable → grep fallback";
+      return "rag: unavailable → grep fallback";
     // Nothing has been measured yet (the state before any turn has run). "0" is a claim we
     // cannot support; a dash says "no figure", which is the truth.
     case "unknown":
@@ -182,6 +183,34 @@ export function ragSegmentLabel(b) {
   }
 }
 
+/**
+ * The RAG notice for the BOTTOM BAR's mode row (task #34), or "" when there is none.
+ *
+ * Only THIS session's live verdict that search fell back to grep: "unreachable" (the embedder is
+ * down) or "fallback" (RAG failed mid-turn). It used to sit in red with a "✗" in the middle of the
+ * token row and read as a serious fault, while every search was still answered by grep. It is
+ * now a soft note beside "accept edits on". A stale verdict reloaded from a saved session is
+ * not a live fault, so it gets no notice.
+ */
+export function ragBarNote(b) {
+  if (!b || typeof b !== "object" || b.stale) return "";
+  if (b.ragState !== "unreachable" && b.ragState !== "fallback") return "";
+  return ragSegmentLabel(b) || "";
+}
+
+/** Cut plain text to `room` visible columns, ending with "…" when it was cut. */
+function fitPlain(s, room) {
+  s = String(s ?? "");
+  if (visibleLen(s) <= room) return s;
+  if (room <= 1) return room === 1 ? "…" : "";
+  let out = "";
+  for (const ch of s) {
+    if (visibleLen(out + ch) > room - 1) break;
+    out += ch;
+  }
+  return out + "…";
+}
+
 // The six origin buckets. They are scaled to sum to the server's prompt_tokens for ONE
 // request, so their sum IS the size of the newest context — the number to compare against
 // the model's window. Deliberately excludes ragFresh/ragCarried (shares OF rag, not extra)
@@ -544,11 +573,25 @@ export function buildFooterRows({
   // Right-aligned on the mode row (issue #67), dim. DROPPED WHOLE when it does not fit: every
   // footer row must stay one physical row (#109/#132/#138), and half a key name is worse than none.
   modeHint = "",
+  // Task #34: a soft notice (PLAIN text) after the mode text, painted with notePaint. It takes
+  // priority over the paste hint and is cut with "…" on a narrow window, never wrapped.
+  modeNote = "",
+  notePaint = dim,
   notice = "",
   missing = "",
 }) {
   const w = Math.max(20, width || 80);
-  const modeLeft = gutter + mode;
+  let modeLeft = gutter + mode;
+  if (modeNote) {
+    const NOTE_GAP = "  ";
+    const hintRoom = modeHint ? HINT_MIN_GAP + visibleLen(modeHint) : 0;
+    let room = w - 1 - visibleLen(modeLeft) - NOTE_GAP.length;
+    // Keep the paste hint only when the whole notice still fits beside it.
+    if (modeHint && visibleLen(modeNote) <= room - hintRoom) room -= hintRoom;
+    else modeHint = "";
+    // Too little room to say anything readable: leave the notice out.
+    if (room >= 8) modeLeft += NOTE_GAP + notePaint(fitPlain(modeNote, room));
+  }
   const hintGap = modeHint ? w - 1 - visibleLen(modeLeft) - visibleLen(modeHint) : -1;
   const modeRow =
     hintGap >= HINT_MIN_GAP