aru-code 0.41.0__tar.gz → 0.42.0__tar.gz

{aru_code-0.41.0/aru_code.egg-info → aru_code-0.42.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: aru-code
-Version: 0.41.0
+Version: 0.42.0
 Summary: A Claude Code clone built with Agno agents
 Author-email: Estevao <estevaofon@gmail.com>
 License-Expression: MIT

aru_code-0.42.0/aru/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.42.0"

{aru_code-0.41.0 → aru_code-0.42.0}/aru/tui/app.py

@@ -1068,6 +1068,24 @@ class AruApp(App):
                 self.query_one(ContextPane).refresh_from_session()
             except Exception:
                 pass
+            # Layer 9 self-heal — re-assert Textual's mouse-tracking
+            # sequences at every turn boundary. If any content path
+            # emitted a rogue DEC private-mode escape during the turn
+            # (``\x1b[?1000l`` or similar) the terminal would have
+            # silently disabled wheel reporting for every scroll area
+            # in the app, with no way for us to detect it. Calling the
+            # driver's enable-mouse path writes four short SGR sequences
+            # (``?1000h`` / ``?1003h`` / ``?1015h`` / ``?1006h``) and
+            # restores wheel input no matter what corrupted the state.
+            # Idempotent when mouse tracking was never disabled. See the
+            # Layer 9 post-mortem at the top of
+            # ``aru/tui/widgets/chat.py`` for the full analysis.
+            try:
+                driver = self._driver
+                if driver is not None:
+                    driver._enable_mouse_support()
+            except Exception:
+                pass
 
     # ── Bus wiring — ToolsPane + StatusPane subscribe to plugin events ──
 

{aru_code-0.41.0 → aru_code-0.42.0}/aru/tui/widgets/chat.py

@@ -128,6 +128,91 @@ cost and layout cost on the Aru side. This bug is about the
 terminal's own private-mode state being corrupted from outside —
 entirely unrelated to how fast we render, and invisible to any
 latency benchmark. Treat as a seventh layer: output-hygiene.
+
+----
+
+Post-mortem — "mouse wheel dead during heavy streaming" (2026-04-23,
+``fix/scroll-refinement``)
+---------------------------------------------------------------------
+**Symptom:** mouse wheel over the ChatPane does nothing while the
+agent is actively streaming / running tool batches. TAB to focus
+the pane + arrow keys / PgUp / PgDn works fine. Asymmetric enough
+that it felt like "the mouse lost focus" — not a freeze.
+
+Reported against session ``final-fantasy-battle/.aru/sessions/
+e9397dc3.json``.
+
+**Original (incorrect) theory:** interaction between
+``self.anchor()`` (layer 4) and ``_scroll_up_for_pointer``. This
+was based on a misreading of Textual's source — see the Layer 9
+correction below. ``_scroll_up_for_pointer`` does *not* pass
+``release_anchor=False``; it defaults to ``True`` (``widget.py:3378``
+→ ``widget.py:2730``), so wheel-up already releases the anchor via
+the framework. The ``on_mouse_scroll_up`` handler we added
+(``ChatPane.on_mouse_scroll_up``) is therefore redundant with the
+framework's own behaviour — a no-op on the happy path. It is kept
+as defensive redundancy because removing it is the same shape of
+change as keeping it, but it should not be credited for "fixing"
+anything.
+
+**What the bug probably was:** the same Layer-7 class of issue
+that the next session surfaced again — a rogue DEC private-mode
+escape reaching the terminal and disabling X10 mouse reporting.
+See Layer 9 for the real signature and the robust fix.
+
+----
+
+Post-mortem — "wheel globally dead at end of stream" (2026-04-24,
+``fix/scroll-refinement`` continued)
+---------------------------------------------------------------------
+**Symptom:** immediately after a long streaming turn concluded,
+mouse wheel stopped working on *every* scrollable surface in the
+app — ChatPane, sidebars, modals — simultaneously. TAB to walk
+focus into a scrollbar and arrow-key scrolling from there worked.
+Classic Layer-7 fingerprint: terminal-level mouse reporting got
+turned off.
+
+Reported against session ``final-fantasy-battle3/.aru/sessions/
+7e9e4549.json``: one mega-turn with 120 tool calls interleaved
+with 66 text blocks, 31 plan-panel mounts via
+``add_renderable(scrollable=True)``, ~245 widgets in the pane.
+
+**What we could prove:** a byte-level scan of the saved session
+for C0 control chars turned up zero ``\\x1b`` bytes. The leak is
+either (a) from a path that isn't persisted to ``session.json``
+(tool ``stdout``/``stderr`` never reaches the chat directly but
+transient UI strings, skill output, or reasoning tokens might),
+or (b) a Windows ConPTY quirk during high-volume redraw where the
+driver's mouse-enable state drops without us emitting anything
+hostile. Chasing the exact source is caça ao fantasma; the
+mitigation is structural.
+
+**Two-prong fix:**
+
+1. **Close the last unsanitised content path —
+   ``_SanitizedRenderable``.** ``ChatMessageWidget`` already
+   sanitises everything that goes through its ``buffer``. Arbitrary
+   Rich renderables handed to ``add_renderable`` (plan panels, task
+   lists, diff previews, the logo) bypass that path and mount as
+   ``Static(renderable)``. The wrapper sits between the renderable
+   and Rich's console, filtering C0 bytes out of every segment's
+   ``.text`` before it reaches Textual's compositor. Matches the
+   Layer 7 sanitisation boundary for the unchecked route.
+
+2. **Self-heal at turn boundary —
+   ``AruApp._run_turn`` finally clause.** Call the driver's
+   ``_enable_mouse_support()`` after each turn finishes. That
+   re-emits Textual's own four mouse-enable sequences (``?1000h``,
+   ``?1003h``, ``?1015h``, ``?1006h`` — see
+   ``textual/drivers/windows_driver.py:56``). Cost is four short
+   writes; benefit is full recovery of wheel input regardless of
+   what corrupted the terminal state mid-turn. Idempotent: a no-op
+   when mouse tracking was never disabled.
+
+Treat as a ninth layer: defence-in-depth against terminal-state
+corruption. Prong 1 plugs the last known-possible leak inside our
+code; prong 2 recovers even if something outside our reach drops
+the state anyway.
 """
 
 from __future__ import annotations
@@ -182,6 +267,39 @@ def _sanitize_for_terminal(raw: str) -> str:
     return raw.translate(_CTRL_CHAR_TRANSLATION)
 
 
+class _SanitizedRenderable:
+    """Wraps a Rich renderable so its output segments are stripped of C0 bytes.
+
+    ``_sanitize_for_terminal`` covers every string passing through
+    ``ChatMessageWidget.buffer``. Arbitrary Rich renderables handed to
+    ``ChatPane.add_renderable`` (plan panels, task lists, diff previews, the
+    startup logo) skip that widget and mount as a plain ``Static(renderable)``
+    — so a rogue ``\\x1b[?1000l`` inside a task description, a panel title, or
+    subprocess output echoed into a panel would flow straight through Rich
+    segments to Textual's compositor and onto the terminal, disabling mouse
+    tracking globally (Layer 7 signature).
+
+    This wrapper closes that gap: ``console.render`` yields segments from the
+    inner renderable, we strip C0 bytes from any segment whose ``.text``
+    contains them, and re-emit the cleaned stream. Rich's ``Segment`` is a
+    ``NamedTuple`` so ``seg._replace(text=...)`` is a cheap immutable swap.
+    Unchanged segments are re-emitted unmodified — the hot path is a single
+    ``str.translate`` on segment text which typically no-ops.
+    """
+
+    def __init__(self, inner: Any) -> None:
+        self._inner = inner
+
+    def __rich_console__(self, console: Any, options: Any) -> Any:
+        for seg in console.render(self._inner, options):
+            if seg.text:
+                clean = _sanitize_for_terminal(seg.text)
+                if clean != seg.text:
+                    yield seg._replace(text=clean)
+                    continue
+            yield seg
+
+
 def _scan_fences(text: str) -> tuple[int, int]:
     """One-pass fence scanner. Returns ``(last_stable_split, open_fence_start)``.
 
@@ -875,16 +993,36 @@ class ChatPane(VerticalScroll):
         # us enqueuing a ``scroll_end`` after every delta / tool event.
         # (a) kills the ``call_after_refresh`` backlog that piled up when
         # the UI thread was busy rendering markdown; (b) releases the anchor
-        # when the user manually scrolls up, so they can read history
-        # mid-stream without the viewport snapping back every 50 ms; and
-        # (c) re-engages automatically when they return to the bottom via
-        # ``_check_anchor``. Matches Textual's own "streaming Markdown"
+        # when the user manually scrolls — wheel, keyboard, or drag all go
+        # through ``_scroll_to`` which releases by default (widget.py:2730);
+        # and (c) re-engages automatically when they return to the bottom
+        # via ``_check_anchor``. Matches Textual's own "streaming Markdown"
         # recipe (see ``Markdown.get_stream`` docstring).
         self.anchor()
         # Periodic flush; cheap because the reactive watcher already
         # debounces repaints when buffer doesn't actually change.
         self.set_interval(self.DEBOUNCE_SEC, self._flush_pending_delta)
 
+    def on_mouse_scroll_up(self, event) -> None:
+        """Defensive redundancy — explicitly release the anchor on wheel-up.
+
+        Originally added under a misreading of Textual's source (see the
+        Layer 8 correction in the module post-mortem). The framework's
+        ``_scroll_up_for_pointer`` calls ``_scroll_to`` *without*
+        ``release_anchor``, which defaults to ``True`` in
+        ``widget.py:2730`` — so Textual already releases the anchor on
+        wheel-up. This handler does the same thing one beat earlier and
+        is effectively a no-op on the normal path.
+
+        Kept because (a) removing it has the same shape of change as
+        keeping it and (b) if some future Textual refactor ever flips
+        the default, this keeps wheel-up behaving the way ChatPane
+        needs. No ``event.stop()`` — the framework handler still runs
+        after this and does the actual scroll.
+        """
+        if self._anchored and not self._anchor_released:
+            self.release_anchor()
+
     # ── API used by TextualBusSink and the App ────────────────────────
 
     def add_user_message(self, text: str) -> None:
@@ -924,7 +1062,11 @@ class ChatPane(VerticalScroll):
         """
         from textual.widgets import Static
         self._close_active_assistant()
-        widget = Static(renderable)
+        # Sanitise the renderable's segment stream — see
+        # ``_SanitizedRenderable`` docstring. This is the only content path
+        # into the ChatPane that doesn't go through ``ChatMessageWidget``,
+        # so it needs its own Layer-7 barrier.
+        widget = Static(_SanitizedRenderable(renderable))
         if scrollable:
             from textual.containers import VerticalScroll
             wrapper = VerticalScroll()

{aru_code-0.41.0 → aru_code-0.42.0/aru_code.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: aru-code
-Version: 0.41.0
+Version: 0.42.0
 Summary: A Claude Code clone built with Agno agents
 Author-email: Estevao <estevaofon@gmail.com>
 License-Expression: MIT

{aru_code-0.41.0 → aru_code-0.42.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "aru-code"
-version = "0.41.0"
+version = "0.42.0"
 description = "A Claude Code clone built with Agno agents"
 readme = "README.md"
 license = "MIT"

aru_code-0.41.0/aru/__init__.py

@@ -1 +0,0 @@
-__version__ = "0.41.0"