diff --git a/.gitignore b/.gitignore
index 75a7b67c..2bdceaad 100644
--- a/.gitignore
+++ b/.gitignore
@@ -39,3 +39,7 @@ Thumbs.db
 docs/*
 !docs/ui-ux/
 !docs/ui-ux/**
+
+# Local-only PR review harness: rendering drivers, sample bank, fixtures.
+# Used by Claude during deep reviews; never shared in the repo.
+.local-review/
diff --git a/CHANGELOG.md b/CHANGELOG.md
index b8c5a7ad..5deb1c43 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,24 @@
 
 ## [Unreleased]
 
+
+## v0.50.216 — 2026-04-26
+
+### Added
+- **Compression chain collapse** — `get_importable_agent_sessions()` now merges linear compression continuation chains into a single sidebar entry, showing the chain tip's activity time and model. The chain root's title and start time are preserved for display; the latest importable segment is used for import. Non-compression parent/child pairs are unchanged. (`api/agent_sessions.py`, `tests/test_gateway_sync.py`) Closes #1012 [#1012 @franksong2702]
+- **Comprehensive markdown renderer improvements** — blockquote grouping, strikethrough, task lists, CRLF normalisation, nested blockquotes, lists inside blockquotes. See details below. (`static/ui.js`) [#1073]
+
+### Fixed
+- **Blockquote rendering** — consecutive `> lines` now group into one `<blockquote>`, blank `>` continuation lines become `<br>`, bare `>` (no space) handled, `>>` nested blockquotes recurse correctly, lists inside blockquotes render `<ul>`, inline markdown (bold/italic/code) works inside quotes. (`static/ui.js`) [#1073]
+- **Strikethrough** — `~~text~~` now renders as `<del>text</del>` in all contexts (paragraphs, blockquotes, list items). (`static/ui.js`) [#1073]
+- **Task lists** — `- [x]` renders as ✅, `- [ ]` renders as ☐ in all unordered list contexts including inside blockquotes. (`static/ui.js`) [#1073]
+- **CRLF line endings** — Windows `\r\n` line endings are normalised at the start of `renderMd()` so `\r` never appears in rendered text. (`static/ui.js`) [#1073]
+- **HTML/HTM preview in workspace** — `.html` and `.htm` files now render correctly in the workspace preview iframe. Root cause: `MIME_MAP` was missing these extensions; the fallback `application/octet-stream` caused browsers to refuse to render in the iframe. (`api/config.py`) [#1070]
+- **Approval card obscured by queue flyout** — the approval card's "Allow once / Allow session / Always allow / Deny" buttons are no longer hidden behind the queue flyout when both are visible simultaneously. (`static/style.css` — one line: `z-index:3` on `.approval-card.visible`) [#1071]
+- **`/steer`, `/interrupt`, `/queue` not working while agent is busy** — typing these commands while the agent is running now executes them immediately instead of queuing the raw text. Root cause: `send()` returned early inside the busy block before reaching the slash-command dispatcher. Fix: intercept the three control commands at the top of the busy block. (`static/messages.js`) [#1072]
+- **Reasoning chip always visible** — the composer reasoning chip is now shown for all effort states. When effort is unset/default it shows a muted "Default" label; when explicitly set to `none` it shows "None". Previously both states hid the chip entirely, removing the affordance to inspect or change it. (`static/ui.js`, `static/style.css`) Closes #1068 [#1074 @franksong2702]
+- **Steer settings copy updated** — removed "falls back to interrupt" / "interrupt + send" language across all 6 locales; steer mode now correctly described as "mid-turn correction without interrupting". (`static/i18n.js`, `static/index.html`) [#1072]
+
 ## v0.50.215 — 2026-04-26
 
 ### Added
diff --git a/ROADMAP.md b/ROADMAP.md
index 523f73a1..0735b3a4 100644
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -3,7 +3,7 @@
 > Goal: Full 1:1 parity with the Hermes CLI experience via a clean dark web UI.
 > Everything you can do from the CLI terminal, you can do from this UI.
 >
-> Last updated: v0.50.215 (April 26, 2026) — 2319 tests collected
+> Last updated: v0.50.216 (April 26, 2026) — 2428 tests collected
 > Tests: 2107 collected (`pytest tests/ --collect-only -q`)
 > Source: <repo>/
 
diff --git a/TESTING.md b/TESTING.md
index cbb814ab..f513a021 100644
--- a/TESTING.md
+++ b/TESTING.md
@@ -8,7 +8,7 @@
 > Prerequisites: SSH tunnel is active on port 8787. Open http://localhost:8787 in browser.
 > Server health check: curl http://127.0.0.1:8787/health should return {"status":"ok"}.
 >
-> Automated coverage: 2319 tests collected via `pytest tests/ --collect-only -q`. Includes onboarding coverage for bootstrap/static wizard presence, real provider config persistence (`config.yaml` + `.env`), the `/api/onboarding/*` backend, the onboarding skip/existing-config guard, and CSS regression coverage for smooth thinking/tool card disclosure animation.
+> Automated coverage: 2428 tests collected via `pytest tests/ --collect-only -q`. Includes onboarding coverage for bootstrap/static wizard presence, real provider config persistence (`config.yaml` + `.env`), the `/api/onboarding/*` backend, the onboarding skip/existing-config guard, and CSS regression coverage for smooth thinking/tool card disclosure animation.
 > Run: `pytest tests/ -v --timeout=60`
 >
 > Local regression focus: verify that a previously closed workspace panel stays visually closed from first paint through boot completion on desktop refresh; there should be no brief open-then-close flash.
diff --git a/api/agent_sessions.py b/api/agent_sessions.py
index 00ab8b93..53f46132 100644
--- a/api/agent_sessions.py
+++ b/api/agent_sessions.py
@@ -6,13 +6,126 @@ from pathlib import Path
 logger = logging.getLogger(__name__)
 
 
+def _optional_col(name: str, columns: set[str], fallback: str = "NULL") -> str:
+    return f"s.{name}" if name in columns else f"{fallback} AS {name}"
+
+
+def _is_compression_continuation(parent: dict | None, child: dict) -> bool:
+    """Mirror Hermes Agent's compression-child guard.
+
+    A child is a continuation only when the parent ended because of compression
+    and the child started after that compression boundary. Plain parent/child
+    relationships are left alone for future subagent-tree work.
+    """
+    if not parent:
+        return False
+    if parent.get('end_reason') != 'compression':
+        return False
+    ended_at = parent.get('ended_at')
+    if ended_at is None:
+        return False
+    try:
+        return float(child.get('started_at') or 0) >= float(ended_at)
+    except (TypeError, ValueError):
+        return False
+
+
+def _project_agent_session_rows(rows: list[dict]) -> list[dict]:
+    """Collapse compression chains into one logical sidebar row.
+
+    The visible conversation should still look like the original chain head
+    (title and timestamps), while importing should use the latest importable
+    segment so the user continues from the current compressed state.
+    """
+    rows_by_id = {row['id']: row for row in rows}
+    children_by_parent: dict[str, list[dict]] = {}
+    continuation_child_ids = set()
+
+    for row in rows:
+        parent_id = row.get('parent_session_id')
+        if not parent_id:
+            continue
+        children_by_parent.setdefault(parent_id, []).append(row)
+        if _is_compression_continuation(rows_by_id.get(parent_id), row):
+            continuation_child_ids.add(row['id'])
+
+    for children in children_by_parent.values():
+        children.sort(key=lambda row: row.get('started_at') or 0, reverse=True)
+
+    def compression_tip(row: dict) -> tuple[dict | None, int]:
+        current = row
+        seen = {row['id']}
+        latest_importable = row if (row.get('actual_message_count') or 0) > 0 else None
+        segment_count = 1
+        for _ in range(len(rows_by_id) + 1):
+            candidates = [
+                child for child in children_by_parent.get(current['id'], [])
+                if child['id'] not in seen and _is_compression_continuation(current, child)
+            ]
+            if not candidates:
+                return latest_importable, segment_count
+            current = candidates[0]
+            seen.add(current['id'])
+            segment_count += 1
+            if (current.get('actual_message_count') or 0) > 0:
+                latest_importable = current
+        return latest_importable, segment_count
+
+    projected = []
+    for row in rows:
+        if row['id'] in continuation_child_ids:
+            continue
+
+        segment_count = 1
+        tip = row
+        if row.get('end_reason') == 'compression':
+            tip, segment_count = compression_tip(row)
+        if not tip or (tip.get('actual_message_count') or 0) <= 0:
+            continue
+
+        if tip is row:
+            projected.append(dict(row))
+            continue
+
+        merged = dict(row)
+        # Keep the chain head's visible identity (title, started_at), but
+        # point the row at the latest importable segment for navigation AND
+        # surface the tip's recency so an actively-used chain bubbles to the
+        # top of the sidebar by its true last activity. Without overriding
+        # last_activity, a long-lived chain whose tip is being edited NOW
+        # would sort by the root's old timestamp and fall below recently
+        # touched standalone sessions — exactly the inverse of what a user
+        # expects from "Show agent sessions" sorted by activity.
+        for key in (
+            'id', 'model', 'message_count', 'actual_message_count',
+            'ended_at', 'end_reason', 'last_activity',
+        ):
+            if key in tip:
+                merged[key] = tip[key]
+        if not merged.get('title'):
+            merged['title'] = tip.get('title')
+        if not merged.get('source'):
+            merged['source'] = tip.get('source')
+        merged['_lineage_root_id'] = row['id']
+        merged['_lineage_tip_id'] = tip['id']
+        merged['_compression_segment_count'] = segment_count
+        projected.append(merged)
+
+    projected.sort(
+        key=lambda row: row.get('last_activity') or row.get('started_at') or 0,
+        reverse=True,
+    )
+    return projected
+
+
 def read_importable_agent_session_rows(db_path: Path, limit: int = 200, log=None) -> list[dict]:
-    """Return non-WebUI agent sessions that have readable message rows.
+    """Return non-WebUI agent sessions projected as importable conversations.
 
     Hermes Agent can create rows in ``state.db.sessions`` before a session has
-    any messages. WebUI cannot import those rows, so both the regular
-    ``/api/sessions`` path and the gateway SSE watcher must filter them the
-    same way.
+    any messages, and long conversations can be split into compression-linked
+    rows. WebUI cannot import empty rows and should not show compression
+    segments as separate conversations, so both the regular ``/api/sessions``
+    path and the gateway SSE watcher use this shared projection.
     """
     db_path = Path(db_path)
     if not db_path.exists():
@@ -36,20 +149,27 @@ def read_importable_agent_session_rows(db_path: Path, limit: int = 200, log=None
             )
             return []
 
+        parent_expr = _optional_col('parent_session_id', session_cols)
+        ended_expr = _optional_col('ended_at', session_cols)
+        end_reason_expr = _optional_col('end_reason', session_cols)
+
         cur.execute(
-            """
+            f"""
             SELECT s.id, s.title, s.model, s.message_count,
                    s.started_at, s.source,
+                   {parent_expr},
+                   {ended_expr},
+                   {end_reason_expr},
                    COUNT(m.id) AS actual_message_count,
                    MAX(m.timestamp) AS last_activity
             FROM sessions s
             LEFT JOIN messages m ON m.session_id = s.id
             WHERE s.source IS NOT NULL AND s.source != 'webui'
             GROUP BY s.id
-            HAVING COUNT(m.id) > 0
             ORDER BY COALESCE(MAX(m.timestamp), s.started_at) DESC
-            LIMIT ?
             """,
-            (int(limit),),
         )
-        return [dict(row) for row in cur.fetchall()]
+        projected = _project_agent_session_rows([dict(row) for row in cur.fetchall()])
+        if limit is None:
+            return projected
+        return projected[:max(0, int(limit))]
diff --git a/api/config.py b/api/config.py
index 75909f31..1ad055ce 100644
--- a/api/config.py
+++ b/api/config.py
@@ -442,6 +442,8 @@ MIME_MAP = {
     ".bmp": "image/bmp",
     ".pdf": "application/pdf",
     ".json": "application/json",
+    ".html": "text/html",
+    ".htm": "text/html",
     ".xls": "application/vnd.ms-excel",
     ".xlsx": "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet",
     ".doc": "application/msword",
diff --git a/static/commands.js b/static/commands.js
index e7d7cfdf..aeb25d71 100644
--- a/static/commands.js
+++ b/static/commands.js
@@ -595,8 +595,8 @@ async function cmdSteer(args){
  * Shared implementation for /steer and the busy_input_mode='steer' path.
  *
  * Tries the real steer endpoint first. On any non-accept response (no cached
- * agent, agent lacks steer, stream dead, etc.) falls back to interrupt mode:
- * queue the message + cancel the stream so the existing drain re-sends.
+ * agent, agent lacks steer, stream dead, etc.) falls back to interrupt+queue:
+ * queues the message and cancels the stream so the drain re-sends it.
  *
  * @param {string} msg - The steer text.
  * @param {boolean} explicitSteer - True if the user explicitly invoked /steer
diff --git a/static/i18n.js b/static/i18n.js
index f2c8c3b8..58675e1b 100644
--- a/static/i18n.js
+++ b/static/i18n.js
@@ -101,23 +101,23 @@ const LOCALES = {
     no_active_session: 'No active session',
     cmd_queue: 'Queue a message for the next turn',
     cmd_interrupt: 'Cancel current turn and send a new message',
-    cmd_steer: 'Steer the agent with a correction (falls back to interrupt)',
+    cmd_steer: 'Inject a mid-turn correction without interrupting the agent',
     cmd_queue_no_msg: 'Usage: /queue <message>',
     cmd_queue_not_busy: 'No active task — just send normally',
     cmd_queue_confirm: 'Message queued',
     cmd_interrupt_no_msg: 'Usage: /interrupt <message>',
     cmd_interrupt_confirm: 'Interrupted — sending new message',
     cmd_steer_no_msg: 'Usage: /steer <message>',
-    cmd_steer_fallback: 'Steer unavailable — interrupted and queued instead',
+    cmd_steer_fallback: 'Steer unavailable — queued for next turn instead',
     cmd_steer_delivered: 'Steer delivered — agent will see it on its next tool result',
     steer_leftover_queued: 'Steer queued for next turn',
-    busy_steer_fallback: 'Steer not available — interrupted instead',
+    busy_steer_fallback: 'Steer unavailable — queued for next turn',
     busy_interrupt_confirm: 'Interrupted — sending new message',
     settings_label_busy_input_mode: 'Busy input mode',
-    settings_desc_busy_input_mode: 'Controls what happens when you send a message while the agent is running. Queue waits; Interrupt cancels and starts fresh; Steer sends a correction (currently falls back to interrupt).',
+    settings_desc_busy_input_mode: 'Controls what happens when you send a message while the agent is running. Queue waits; Interrupt cancels and starts fresh; Steer injects a correction mid-turn without interrupting (falls back to queue when agent or stream unavailable).',
     settings_busy_input_mode_queue: 'Queue follow-up',
     settings_busy_input_mode_interrupt: 'Interrupt current turn',
-    settings_busy_input_mode_steer: 'Steer (interrupt + send)',
+    settings_busy_input_mode_steer: 'Steer (mid-turn correction)',
 
   slash_skill_badge:'Skill',
   slash_skill_desc:'Invoke this skill',
@@ -742,7 +742,7 @@ const LOCALES = {
     busy_steer_fallback: 'Steer недоступен — прервано',
     busy_interrupt_confirm: 'Прервано — отправка нового сообщения',
     settings_label_busy_input_mode: 'Режим ввода при занятости',
-    settings_desc_busy_input_mode: 'Определяет поведение при отправке сообщения во время работы агента. Очередь ждёт; Прерывание отменяет и начинает заново; Steer отправляет исправление (сейчас как прерывание).',
+    settings_desc_busy_input_mode: 'Определяет поведение при отправке сообщения во время работы агента. Очередь ждёт; Прерывание отменяет и начинает заново; Steer внедряет коррекцию без прерывания.',
     settings_busy_input_mode_queue: 'Поставить в очередь',
     settings_busy_input_mode_interrupt: 'Прервать текущий оборот',
     settings_busy_input_mode_steer: 'Steer (прерывание + отправка)',
@@ -1340,23 +1340,23 @@ const LOCALES = {
     no_active_session: 'No hay ninguna sesión activa',
     cmd_queue: 'Poner mensaje en cola para el siguiente turno',
     cmd_interrupt: 'Cancelar turno actual y enviar nuevo mensaje',
-    cmd_steer: 'Redirigir al agente con una correcci\u00f3n (usa interrupci\u00f3n)',
+    cmd_steer: 'Inyectar una corrección a mitad del turno sin interrumpir al agente',
     cmd_queue_no_msg: 'Uso: /queue <mensaje>',
     cmd_queue_not_busy: 'Sin tarea activa \u2014 env\u00eda normalmente',
     cmd_queue_confirm: 'Mensaje en cola',
     cmd_interrupt_no_msg: 'Uso: /interrupt <mensaje>',
     cmd_interrupt_confirm: 'Interrumpido \u2014 enviando nuevo mensaje',
     cmd_steer_no_msg: 'Uso: /steer <mensaje>',
-    cmd_steer_fallback: 'Steer no disponible \u2014 interrumpido y encolado',
+    cmd_steer_fallback: 'Steer no disponible — en cola para el siguiente turno',
     cmd_steer_delivered: 'Steer entregado \u2014 el agente lo ver\u00e1 en su pr\u00f3ximo resultado de herramienta',
     steer_leftover_queued: 'Steer en cola para el pr\u00f3ximo turno',
-    busy_steer_fallback: 'Steer no disponible \u2014 interrumpido',
+    busy_steer_fallback: 'Steer no disponible — en cola para el siguiente turno',
     busy_interrupt_confirm: 'Interrumpido \u2014 enviando nuevo mensaje',
     settings_label_busy_input_mode: 'Modo de entrada ocupada',
-    settings_desc_busy_input_mode: 'Controla qu\u00e9 sucede al enviar un mensaje mientras el agente est\u00e1 activo. Cola espera; Interrumpir cancela y empieza de nuevo; Steer env\u00eda una correcci\u00f3n (actualmente usa interrupci\u00f3n).',
+    settings_desc_busy_input_mode: 'Controla qué sucede al enviar mensajes mientras el agente está activo. Cola espera; Interrumpir cancela y empieza de nuevo; Steer inyecta una corrección sin interrumpir (usa cola si el agente no está disponible).',
     settings_busy_input_mode_queue: 'Poner en cola',
     settings_busy_input_mode_interrupt: 'Interrumpir turno actual',
-    settings_busy_input_mode_steer: 'Steer (interrupci\u00f3n + env\u00edo)',
+    settings_busy_input_mode_steer: 'Steer (corrección a mitad de turno)',
     no_personalities: 'No se encontraron personalidades (añádelas a ~/.hermes/personalities/)',
     available_personalities: 'Personalidades disponibles:',
     personality_switch_hint: '\n\nUsa `/personality <name>` para cambiar, o `/personality none` para limpiar.',
@@ -1920,23 +1920,23 @@ const LOCALES = {
     no_active_session: 'Keine aktive Sitzung',
     cmd_queue: 'Nachricht f\u00fcr den n\u00e4chsten Durchgang einreihen',
     cmd_interrupt: 'Aktuellen Durchgang abbrechen und neue Nachricht senden',
-    cmd_steer: 'Agent mit Korrektur lenken (f\u00e4llt zur\u00fcck auf Unterbrechung)',
+    cmd_steer: 'Korrektursignal einf\u00fcgen ohne Unterbrechung',
     cmd_queue_no_msg: 'Verwendung: /queue <Nachricht>',
     cmd_queue_not_busy: 'Keine aktive Aufgabe \u2014 normal senden',
     cmd_queue_confirm: 'Nachricht eingereiht',
     cmd_interrupt_no_msg: 'Verwendung: /interrupt <Nachricht>',
     cmd_interrupt_confirm: 'Unterbrochen \u2014 neue Nachricht wird gesendet',
     cmd_steer_no_msg: 'Verwendung: /steer <Nachricht>',
-    cmd_steer_fallback: 'Steer nicht verf\u00fcgbar \u2014 unterbrochen und eingereiht',
+    cmd_steer_fallback: 'Steer nicht verf\u00fcgbar \u2014 f\u00fcr n\u00e4chsten Durchgang eingereiht',
     cmd_steer_delivered: 'Steer geliefert \u2014 der Agent sieht es bei seinem n\u00e4chsten Tool-Ergebnis',
     steer_leftover_queued: 'Steer f\u00fcr n\u00e4chsten Durchgang eingereiht',
-    busy_steer_fallback: 'Steer nicht verf\u00fcgbar \u2014 unterbrochen',
+    busy_steer_fallback: 'Steer nicht verf\u00fcgbar \u2014 f\u00fcr n\u00e4chsten Durchgang eingereiht',
     busy_interrupt_confirm: 'Unterbrochen \u2014 neue Nachricht wird gesendet',
     settings_label_busy_input_mode: 'Eingabemodus bei Besch\u00e4ftigung',
-    settings_desc_busy_input_mode: 'Steuert, was passiert, wenn Sie w\u00e4hrend der Agentenaktivit\u00e4t eine Nachricht senden. Warteschlange wartet; Unterbrechen bricht ab und startet neu; Steer sendet eine Korrektur (aktuell wie Unterbrechen).',
+    settings_desc_busy_input_mode: 'Steuert, was passiert, wenn Sie w\u00e4hrend der Agentenaktivit\u00e4t eine Nachricht senden. Warteschlange wartet; Unterbrechen bricht ab und startet neu; Steer f\u00fcgt eine Korrektur ein ohne zu unterbrechen.',
     settings_busy_input_mode_queue: 'In Warteschlange einreihen',
     settings_busy_input_mode_interrupt: 'Aktuellen Durchgang unterbrechen',
-    settings_busy_input_mode_steer: 'Steer (Unterbrechen + Senden)',
+    settings_busy_input_mode_steer: 'Steer (Korrektur ohne Unterbrechung)',
     no_personalities: 'Keine Persönlichkeiten gefunden (füge sie in ~/.hermes/personalities/ hinzu)',
     available_personalities: 'Verfügbare Persönlichkeiten:',
     personality_switch_hint: '\n\nNutze `/personality <name>` zum Wechseln, oder `/personality none` zum Löschen.',
@@ -2303,7 +2303,7 @@ const LOCALES = {
     busy_steer_fallback: 'Steer \u4e0d\u53ef\u7528 \u2014 \u5df2\u4e2d\u65ad',
     busy_interrupt_confirm: '\u5df2\u4e2d\u65ad \u2014 \u6b63\u5728\u53d1\u9001\u65b0\u6d88\u606f',
     settings_label_busy_input_mode: '\u5fd9\u788c\u8f93\u5165\u6a21\u5f0f',
-    settings_desc_busy_input_mode: '\u63a7\u5236\u5728\u4ee3\u7406\u8fd0\u884c\u65f6\u53d1\u9001\u6d88\u606f\u7684\u884c\u4e3a\u3002\u961f\u5217\u7b49\u5f85\uff1b\u4e2d\u65ad\u53d6\u6d88\u5e76\u91cd\u65b0\u5f00\u59cb\uff1bSteer \u53d1\u9001\u7ea0\u6b63\uff08\u76ee\u524d\u56de\u9000\u4e3a\u4e2d\u65ad\uff09\u3002',
+    settings_desc_busy_input_mode: '\u63a7\u5236\u5728\u4ee3\u7406\u8fd0\u884c\u65f6\u53d1\u9001\u6d88\u606f\u7684\u884c\u4e3a\u3002\u961f\u5217\u7b49\u5f85\uff1b\u4e2d\u65ad\u53d6\u6d88\u5e76\u91cd\u65b0\u5f00\u59cb\uff1bSteer\u4e2d\u9014\u6ce8\u5165\u7ea0\u6b63\uff0c\u4e0d\u4e2d\u65ad\u3002',
     settings_busy_input_mode_queue: '\u52a0\u5165\u961f\u5217',
     settings_busy_input_mode_interrupt: '\u4e2d\u65ad\u5f53\u524d\u56de\u5408',
     settings_busy_input_mode_steer: 'Steer\uff08\u4e2d\u65ad + \u53d1\u9001\uff09',
@@ -3215,23 +3215,23 @@ const LOCALES = {
     no_active_session: '\u7121\u6d3b\u8e8d\u6703\u8a71',
     cmd_queue: '\u5c07\u8a0a\u606f\u52a0\u5165\u4e0b\u4e00\u8f2a\u7684\u4f47\u5217',
     cmd_interrupt: '\u53d6\u6d88\u7576\u524d\u56de\u5408\u4e26\u767c\u9001\u65b0\u8a0a\u606f',
-    cmd_steer: '\u7528\u7d0a\u6b63\u8a0a\u606f\u5f15\u5c0e\u4ee3\u7406\uff08\u56de\u9000\u70ba\u4e2d\u65ad\uff09',
+    cmd_steer: '\u5728\u56de\u5408\u9032\u884c\u4e2d\u6ce8\u5165\u7d3a\u6b63\uff0c\u4e0d\u4e2d\u65b7\u4ee3\u7406',
     cmd_queue_no_msg: '\u7528\u6cd5\uff1a/queue <\u8a0a\u606f>',
     cmd_queue_not_busy: '\u6c92\u6709\u6d3b\u52d5\u4efb\u52d9 \u2014 \u76f4\u63a5\u767c\u9001\u5373\u53ef',
     cmd_queue_confirm: '\u8a0a\u606f\u5df2\u52a0\u5165\u4f47\u5217',
     cmd_interrupt_no_msg: '\u7528\u6cd5\uff1a/interrupt <\u8a0a\u606f>',
     cmd_interrupt_confirm: '\u5df2\u4e2d\u65ad \u2014 \u6b63\u5728\u767c\u9001\u65b0\u8a0a\u606f',
     cmd_steer_no_msg: '\u7528\u6cd5\uff1a/steer <\u8a0a\u606f>',
-    cmd_steer_fallback: 'Steer \u4e0d\u53ef\u7528 \u2014 \u5df2\u4e2d\u65ad\u4e26\u52a0\u5165\u4f47\u5217',
+    cmd_steer_fallback: 'Steer \u4e0d\u53ef\u7528 \u2014 \u5df2\u52a0\u5165\u4e0b\u4e00\u8f2a\u4f47\u5217',
     cmd_steer_delivered: 'Steer \u5df2\u9001\u9054 \u2014 \u4ee3\u7406\u5c07\u5728\u4e0b\u4e00\u500b\u5de5\u5177\u7d50\u679c\u4e2d\u770b\u5230',
     steer_leftover_queued: 'Steer \u5df2\u52a0\u5165\u4e0b\u4e00\u8f2a\u4f47\u5217',
-    busy_steer_fallback: 'Steer \u4e0d\u53ef\u7528 \u2014 \u5df2\u4e2d\u65ad',
+    busy_steer_fallback: 'Steer \u4e0d\u53ef\u7528 \u2014 \u5df2\u52a0\u5165\u4e0b\u4e00\u8f2a\u4f47\u5217',
     busy_interrupt_confirm: '\u5df2\u4e2d\u65ad \u2014 \u6b63\u5728\u767c\u9001\u65b0\u8a0a\u606f',
     settings_label_busy_input_mode: '\u5fd9\u788c\u8f38\u5165\u6a21\u5f0f',
-    settings_desc_busy_input_mode: '\u63a7\u5236\u5728\u4ee3\u7406\u904b\u884c\u6642\u767c\u9001\u8a0a\u606f\u7684\u884c\u70ba\u3002\u4f47\u5217\u7b49\u5f85\uff1b\u4e2d\u65ad\u53d6\u6d88\u4e26\u91cd\u65b0\u958b\u59cb\uff1bSteer \u767c\u9001\u7d0a\u6b63\uff08\u76ee\u524d\u56de\u9000\u70ba\u4e2d\u65ad\uff09\u3002',
+    settings_desc_busy_input_mode: '\u63a7\u5236\u5728\u4ee3\u7406\u904b\u884c\u6642\u767c\u9001\u8a0a\u606f\u7684\u884c\u70ba\u3002\u4f47\u5217\u7b49\u5f85\uff1b\u4e2d\u65b7\u53d6\u6d88\u4e26\u91cd\u65b0\u958b\u59cb\uff1bSteer\u4e2d\u9014\u6ce8\u5165\u7d3a\u6b63\uff0c\u4e0d\u4e2d\u65b7\u3002',
     settings_busy_input_mode_queue: '\u52a0\u5165\u4f47\u5217',
     settings_busy_input_mode_interrupt: '\u4e2d\u65ad\u7576\u524d\u56de\u5408',
-    settings_busy_input_mode_steer: 'Steer\uff08\u4e2d\u65ad + \u767c\u9001\uff09',
+    settings_busy_input_mode_steer: 'Steer\uff08\u4e2d\u9014\u7d3a\u6b63\uff09',
     no_active_task: '\u7121\u57f7\u884c\u4e2d\u7684\u4efb\u52d9\u53ef\u505c\u6b62\u3002',
     no_notes_yet: '\u5c1a\u7121\u5099\u8a3b\u3002',
     no_profile_yet: '\u5c1a\u7121\u8a2d\u5b9a\u6a94\u3002',
diff --git a/static/index.html b/static/index.html
index 1d63ca13..5bd43af5 100644
--- a/static/index.html
+++ b/static/index.html
@@ -667,7 +667,7 @@
               <select id="settingsBusyInputMode" style="width:100%;padding:8px;background:var(--code-bg);color:var(--text);border:1px solid var(--border2);border-radius:6px">
                 <option value="queue" data-i18n="settings_busy_input_mode_queue">Queue follow-up</option>
                 <option value="interrupt" data-i18n="settings_busy_input_mode_interrupt">Interrupt current turn</option>
-                <option value="steer" data-i18n="settings_busy_input_mode_steer">Steer (interrupt + send)</option>
+                <option value="steer" data-i18n="settings_busy_input_mode_steer">Steer (mid-turn correction)</option>
               </select>
               <div style="font-size:11px;color:var(--muted);margin-top:4px" data-i18n="settings_desc_busy_input_mode">Controls what happens when you send a message while the agent is running. Queue waits for the current task; Interrupt cancels and starts fresh; Steer injects a mid-turn correction without interrupting (falls back to interrupt when the agent is not yet cached or the stream has ended).</div>
             </div>
diff --git a/static/messages.js b/static/messages.js
index 64535066..4ee16886 100644
--- a/static/messages.js
+++ b/static/messages.js
@@ -14,6 +14,23 @@ async function send(){
   if(S.busy||compressionRunning){
     if(text){
       if(!S.session){await newSession();await renderSessionList();}
+      // Busy-control slash commands must be intercepted HERE, before the
+      // busyMode routing block, so the user can always type /steer, /interrupt,
+      // or /queue while the agent is running and have them execute immediately.
+      // Without this intercept they fall through to the queue and execute after
+      // the current turn ends — by which point there is no active stream and
+      // cmdSteer / cmdInterrupt say "No active task to stop."
+      if(text.startsWith('/')){
+        const _pc=typeof parseCommand==='function'&&parseCommand(text);
+        if(_pc&&['steer','interrupt','queue'].includes(_pc.name)){
+          const _bc=COMMANDS.find(c=>c.name===_pc.name);
+          if(_bc){
+            $('msg').value='';autoResize();
+            await _bc.fn(_pc.args);
+            return;
+          }
+        }
+      }
       const busyMode=window._busyInputMode||'queue';
       if(busyMode==='steer'&&S.activeStreamId&&typeof _trySteer==='function'){
         // Real steer: clear the input first so the user gets immediate
diff --git a/static/style.css b/static/style.css
index e08bb26c..44bfd67d 100644
--- a/static/style.css
+++ b/static/style.css
@@ -401,7 +401,7 @@
   .composer-flyout{position:relative;height:0;z-index:1;}
   /* ── Approval card ── */
   .approval-card{position:absolute;left:0;right:0;bottom:-24px;max-width:var(--msg-max);margin:0 auto;padding:0 20px;box-sizing:border-box;width:100%;overflow:hidden;pointer-events:none;}
-  .approval-card.visible{pointer-events:auto;}
+  .approval-card.visible{pointer-events:auto;z-index:3;}
   .approval-inner{background:var(--surface);backdrop-filter:blur(8px);border:1px solid var(--accent-bg-strong);border-radius:14px;padding:16px 18px 40px;transform:translateY(100%);opacity:0;transition:transform .4s cubic-bezier(.32,.72,.16,1),opacity .25s ease;}
   .approval-card.visible .approval-inner{transform:translateY(0);opacity:1;}
   .approval-header{display:flex;align-items:center;gap:8px;margin-bottom:10px;font-size:13px;font-weight:600;color:var(--error);}
@@ -655,6 +655,7 @@
   .composer-workspace-label{min-width:0;overflow:hidden;text-overflow:ellipsis;white-space:nowrap;}
   .composer-reasoning-wrap{position:relative;flex:0 1 auto;min-width:0;}
   .composer-reasoning-chip{display:inline-flex;align-items:center;gap:8px;max-width:180px;padding:8px 10px 8px 12px;border-radius:999px;border:1px solid transparent;background-color:transparent;color:var(--muted);font-weight:500;cursor:pointer;transition:color .15s,background-color .15s,border-color .15s;}
+  .composer-reasoning-chip.inactive{opacity:.78;}
   .composer-reasoning-chip:hover{color:var(--text);background-color:var(--hover-bg);}
   .composer-reasoning-chip.active{color:var(--text);background:var(--accent-bg);border-color:var(--accent-bg);}
   .composer-reasoning-icon,.composer-reasoning-chevron{display:inline-flex;align-items:center;justify-content:center;flex-shrink:0;line-height:1;}
diff --git a/static/ui.js b/static/ui.js
index 77d842c4..cbb8d42c 100644
--- a/static/ui.js
+++ b/static/ui.js
@@ -432,15 +432,31 @@ window.addEventListener('resize',()=>{
 // ── Reasoning effort chip ────────────────────────────────────────────────────
 let _currentReasoningEffort=null;
 
+function _normalizeReasoningEffort(eff){
+  return String(eff||'').trim().toLowerCase();
+}
+
+function _formatReasoningEffortLabel(effort){
+  if(effort==='none') return 'None';
+  if(!effort) return 'Default';
+  return effort;
+}
+
 function _applyReasoningChip(eff){
-  _currentReasoningEffort=eff;
+  const effort=_normalizeReasoningEffort(eff);
+  _currentReasoningEffort=effort;
   const wrap=$('composerReasoningWrap');
   const label=$('composerReasoningLabel');
+  const chip=$('composerReasoningChip');
   if(!wrap||!label) return;
-  if(!eff||eff==='none'){wrap.style.display='none';return;}
   wrap.style.display='';
-  label.textContent=eff;
-  _highlightReasoningOption(eff);
+  label.textContent=_formatReasoningEffortLabel(effort);
+  if(chip){
+    const inactive=!effort||effort==='none';
+    chip.classList.toggle('inactive',inactive);
+    chip.title='Reasoning effort: '+_formatReasoningEffortLabel(effort);
+  }
+  _highlightReasoningOption(effort);
 }
 
 function fetchReasoningChip(){
@@ -664,7 +680,7 @@ function _sanitizeThinkingDisplayText(text){
 }
 
 function renderMd(raw){
-  let s=raw||'';
+  let s=(raw||'').replace(/\r\n/g,'\n').replace(/\r/g,'\n');
   // ── MEDIA: token stash (must run first, before any other processing) ───────
   // Detect MEDIA:<path-or-url> tokens emitted by the agent (e.g. screenshots,
   // generated images) and replace them with inline <img> or download links.
@@ -731,6 +747,8 @@ function renderMd(raw){
     t=t.replace(/\*\*\*(.+?)\*\*\*/g,(_,x)=>`<strong><em>${esc(x)}</em></strong>`);
     t=t.replace(/\*\*(.+?)\*\*/g,(_,x)=>`<strong>${esc(x)}</strong>`);
     t=t.replace(/\*([^*\n]+)\*/g,(_,x)=>`<em>${esc(x)}</em>`);
+    // Strikethrough: ~~text~~ → <del>text</del>
+    t=t.replace(/~~(.+?)~~/g,(_,x)=>`<del>${esc(x)}</del>`);
     // #487: Image pass — runs while code stash is active so ![x](url) inside
     // backticks stays protected as a \x00C token and is never rendered as <img>.
     // Must run before _code_stash restore and before _link_stash so the image
@@ -748,7 +766,7 @@ function renderMd(raw){
     t=t.replace(/\x00G(\d+)\x00/g,(_,i)=>_img_stash[+i]);
     // Escape any plain text that isn't already wrapped in a tag we produced
     // by escaping bare < > that are not part of our own tags
-    const SAFE_INLINE=/^<\/?(strong|em|code|a|img)([\s>]|$)/i;
+    const SAFE_INLINE=/^<\/?(strong|em|del|code|a|img)([\s>]|$)/i;
     t=t.replace(/<\/?[a-z][^>]*>/gi,tag=>SAFE_INLINE.test(tag)?tag:esc(tag));
     return t;
   }
@@ -759,10 +777,47 @@ function renderMd(raw){
   s=s.replace(/\*\*\*(.+?)\*\*\*/g,(_,t)=>`<strong><em>${esc(t)}</em></strong>`);
   s=s.replace(/\*\*(.+?)\*\*/g,(_,t)=>`<strong>${esc(t)}</strong>`);
   s=s.replace(/\*([^*\n]+)\*/g,(_,t)=>`<em>${esc(t)}</em>`);
+  s=s.replace(/~~(.+?)~~/g,(_,t)=>`<del>${esc(t)}</del>`);
   s=s.replace(/\x00O(\d+)\x00/g,(_,i)=>_ob_stash[+i]);
   s=s.replace(/^### (.+)$/gm,(_,t)=>`<h3>${inlineMd(t)}</h3>`).replace(/^## (.+)$/gm,(_,t)=>`<h2>${inlineMd(t)}</h2>`).replace(/^# (.+)$/gm,(_,t)=>`<h1>${inlineMd(t)}</h1>`);
   s=s.replace(/^---+$/gm,'<hr>');
-  s=s.replace(/^> (.+)$/gm,(_,t)=>`<blockquote>${inlineMd(t)}</blockquote>`);
+  // Group consecutive > lines into one <blockquote>.
+  // Handles: blank continuation lines (> alone), nested blockquotes (>>),
+  // lists inside blockquotes (> - item), and inline markdown in quoted text.
+  function _applyBlockquotes(src){
+    return src.replace(/((?:^>[^\n]*(?:\n|$))+)/gm,block=>{
+      const lines=block.split('\n');
+      // Drop trailing bare '>' artifact
+      while(lines.length&&(lines[lines.length-1].trim()==='>'||lines[lines.length-1]===''))
+        {if(lines[lines.length-1].trim()==='>'){lines.pop();break;}lines.pop();}
+      const stripped=lines.map(l=>l.replace(/^>[ \t]?/,''));
+      const innerRaw=stripped.join('\n');
+      let inner;
+      if(/^>/m.test(innerRaw)){
+        // Nested blockquote: recurse so >> → <blockquote><blockquote>
+        inner=_applyBlockquotes(innerRaw);
+      } else if(/(^(?:  )?[-*+] .+)/m.test(innerRaw)){
+        // List inside blockquote: run list pass on stripped inner content
+        inner=innerRaw.replace(/((?:^(?:  )?[-*+] .+\n?)+)/gm,lb=>{
+          const ll=lb.trimEnd().split('\n');let h='<ul>';
+          for(const li of ll){
+            const txt=li.replace(/^ {0,4}[-*+] /,'');
+            let ih;
+            if(/^\[x\] /i.test(txt)) ih='<span class="task-done">✅</span> '+inlineMd(txt.slice(4));
+            else if(/^\[ \] /.test(txt)) ih='<span class="task-todo">☐</span> '+inlineMd(txt.slice(4));
+            else ih=inlineMd(txt);
+            h+=`<li>${ih}</li>`;
+          }
+          return h+'</ul>';
+        });
+      } else {
+        // Plain lines: blank line → <br>, text → inlineMd
+        inner=stripped.map(l=>l.trim()===''?'<br>':inlineMd(l)).join('\n');
+      }
+      return `<blockquote>${inner}</blockquote>`;
+    });
+  }
+  s=_applyBlockquotes(s);
   // B8: improved list handling supporting up to 2 levels of indentation
   s=s.replace(/((?:^(?:  )?[-*+] .+\n?)+)/gm,block=>{
     const lines=block.trimEnd().split('\n');
@@ -770,8 +825,12 @@ function renderMd(raw){
     for(const l of lines){
       const indent=/^ {2,}/.test(l);
       const text=l.replace(/^ {0,4}[-*+] /,'');
-      if(indent) html+=`<li style="margin-left:16px">${inlineMd(text)}</li>`;
-      else html+=`<li>${inlineMd(text)}</li>`;
+      let _ih;
+      if(/^\[x\] /i.test(text)) _ih='<span class="task-done">✅</span> '+inlineMd(text.slice(4));
+      else if(/^\[ \] /.test(text)) _ih='<span class="task-todo">☐</span> '+inlineMd(text.slice(4));
+      else _ih=inlineMd(text);
+      if(indent) html+=`<li style="margin-left:16px">${_ih}</li>`;
+      else html+=`<li>${_ih}</li>`;
     }
     return html+'</ul>';
   });
@@ -820,7 +879,7 @@ function renderMd(raw){
   // Our pipeline only emits: <strong>,<em>,<code>,<pre>,<h1-6>,<ul>,<ol>,<li>,
   // <table>,<thead>,<tbody>,<tr>,<th>,<td>,<hr>,<blockquote>,<p>,<br>,<a>,
   // <div class="..."> (mermaid/pre-header). Everything else is untrusted input.
-  const SAFE_TAGS=/^<\/?(strong|em|code|pre|h[1-6]|ul|ol|li|table|thead|tbody|tr|th|td|hr|blockquote|p|br|a|img|div|span)([\s>]|$)/i;
+  const SAFE_TAGS=/^<\/?(strong|em|del|code|pre|h[1-6]|ul|ol|li|table|thead|tbody|tr|th|td|hr|blockquote|p|br|a|img|div|span)([\s>]|$)/i;
   s=s.replace(/<\/?[a-z][^>]*>/gi,tag=>SAFE_TAGS.test(tag)?tag:esc(tag));
   // Autolink: convert plain URLs to clickable links.
   // Stash <a>, <img> and <pre> blocks so autolink never runs inside them.
diff --git a/tests/test_1062_busy_input_modes.py b/tests/test_1062_busy_input_modes.py
index fc1fa735..7160d82b 100644
--- a/tests/test_1062_busy_input_modes.py
+++ b/tests/test_1062_busy_input_modes.py
@@ -176,6 +176,74 @@ class TestSendBusyBranchDispatch:
         )
 
 
+    def test_slash_commands_intercepted_before_busymode_routing(self):
+        """The three busy-control slash commands (/steer /interrupt /queue) must be
+        intercepted at the TOP of the busy block — before the busyMode routing — so
+        they execute immediately while the agent is running.
+
+        Without this intercept, typing /steer while busy queues the text as a plain
+        message.  When it drains after the turn ends there is no active stream, so
+        cmdSteer says "No active task to stop." and the steer is lost entirely.
+        """
+        send_idx = MESSAGES_JS.find("async function send(")
+        assert send_idx >= 0, "send() not found"
+        # Look in the first 500 chars of the busy block for the intercept
+        busy_start = MESSAGES_JS.find("S.busy||compressionRunning", send_idx)
+        assert busy_start >= 0, "busy block not found"
+        # The intercept must appear BEFORE the busyMode assignment
+        intercept_idx = MESSAGES_JS.find("'steer','interrupt','queue'", busy_start)
+        busymode_idx = MESSAGES_JS.find("_busyInputMode||'queue'", busy_start)
+        assert intercept_idx >= 0, (
+            "send() must intercept /steer /interrupt /queue before the busyMode "
+            "routing block — otherwise they queue instead of executing immediately"
+        )
+        assert intercept_idx < busymode_idx, (
+            "The slash-command intercept must come BEFORE the busyMode routing "
+            "so /steer executes while the agent is running, not after the turn ends"
+        )
+
+    def test_steer_intercept_calls_handler_directly(self):
+        """The busy-intercept must dispatch via _bc.fn(_pc.args), not queue the text."""
+        send_idx = MESSAGES_JS.find("async function send(")
+        busy_start = MESSAGES_JS.find("S.busy||compressionRunning", send_idx)
+        intercept_idx = MESSAGES_JS.find("'steer','interrupt','queue'", busy_start)
+        assert intercept_idx >= 0
+        # Get the intercept block (up to the next busyMode assignment)
+        busymode_idx = MESSAGES_JS.find("_busyInputMode||'queue'", busy_start)
+        intercept_block = MESSAGES_JS[intercept_idx:busymode_idx]
+        assert "_bc.fn(_pc.args)" in intercept_block, (
+            "The intercept must call the command handler directly via _bc.fn(_pc.args)"
+        )
+        assert "return;" in intercept_block, (
+            "The intercept must return after dispatching so send() does not also queue"
+        )
+
+    def test_steer_intercept_clears_input_before_await(self):
+        """The intercept must clear $('msg').value BEFORE awaiting the handler.
+
+        Without the sync clear, the input field still shows '/steer foo' after
+        the steer fires. If the user presses Enter again (a common reflex while
+        waiting for the toast), send() re-runs and either re-fires the command
+        or — once the turn ended — drops a confusing 'No active task to stop.'
+        """
+        send_idx = MESSAGES_JS.find("async function send(")
+        busy_start = MESSAGES_JS.find("S.busy||compressionRunning", send_idx)
+        intercept_idx = MESSAGES_JS.find("'steer','interrupt','queue'", busy_start)
+        busymode_idx = MESSAGES_JS.find("_busyInputMode||'queue'", busy_start)
+        intercept_block = MESSAGES_JS[intercept_idx:busymode_idx]
+        clear_idx = intercept_block.find("$('msg').value=''")
+        await_idx = intercept_block.find("await _bc.fn")
+        assert clear_idx >= 0, (
+            "The intercept must clear $('msg').value (so the field doesn't keep "
+            "showing /steer foo after the command fires)"
+        )
+        assert await_idx >= 0, "await _bc.fn(...) must be present in the intercept"
+        assert clear_idx < await_idx, (
+            "$('msg').value='' must be cleared BEFORE awaiting the handler — "
+            "otherwise a reflexive Enter press during the await re-fires the command"
+        )
+
+
 # ── Boot init + settings panel wiring ───────────────────────────────────
 
 class TestBootAndPanelsWiring:
diff --git a/tests/test_779_html_preview.py b/tests/test_779_html_preview.py
index dec992c7..f3dcd876 100644
--- a/tests/test_779_html_preview.py
+++ b/tests/test_779_html_preview.py
@@ -62,6 +62,19 @@ def test_sandbox_allows_scripts_only():
         )
 
 
+def test_mime_map_includes_html_and_htm():
+    """MIME_MAP must map .html/.htm to text/html — without this, _handle_file_raw
+    falls back to application/octet-stream and browsers refuse to render the
+    response inside the preview iframe (issue #779 follow-up: PR #1070)."""
+    from api.config import MIME_MAP
+    assert MIME_MAP.get(".html") == "text/html", (
+        "MIME_MAP['.html'] must be 'text/html' for the workspace HTML preview iframe"
+    )
+    assert MIME_MAP.get(".htm") == "text/html", (
+        "MIME_MAP['.htm'] must be 'text/html' for the workspace HTML preview iframe"
+    )
+
+
 def test_inline_html_response_sets_csp_sandbox():
     """Defense-in-depth: ?inline=1 HTML responses must set Content-Security-Policy:
     sandbox so the same origin isolation applies even when the URL is opened
diff --git a/tests/test_approval_card_layering.py b/tests/test_approval_card_layering.py
new file mode 100644
index 00000000..3127b0fd
--- /dev/null
+++ b/tests/test_approval_card_layering.py
@@ -0,0 +1,40 @@
+"""
+Regression test for PR #1071: approval card must render above the queue flyout.
+
+Both `.approval-card` and `.queue-card` are siblings inside `.composer-flyout`
+and share the same absolute positioning slot just above the composer. When
+both are visible at the same time (queue flyout open + tool approval card
+sliding up) the approval card MUST win the stacking order so its security-
+relevant Allow / Deny buttons stay clickable.
+
+The old CSS had `.queue-card { z-index: 2 }` and no z-index on
+`.approval-card.visible`, so the queue card painted on top and blocked the
+approval buttons. The fix raises `.approval-card.visible` to z-index 3.
+
+This test pins the invariant: approval-card.visible z-index must be strictly
+greater than queue-card z-index.
+"""
+import re
+from pathlib import Path
+
+CSS = Path("static/style.css").read_text(encoding="utf-8")
+
+
+def _z_index_of(selector_regex: str) -> int | None:
+    m = re.search(selector_regex + r"\s*\{[^}]*z-index:(\d+)", CSS)
+    return int(m.group(1)) if m else None
+
+
+def test_approval_card_visible_outranks_queue_card():
+    queue_z = _z_index_of(r"\.queue-card")
+    approval_visible_z = _z_index_of(r"\.approval-card\.visible")
+    assert queue_z is not None, ".queue-card must declare a z-index"
+    assert approval_visible_z is not None, (
+        ".approval-card.visible must declare a z-index — without it, the approval "
+        "buttons get covered by the queue flyout (PR #1071)"
+    )
+    assert approval_visible_z > queue_z, (
+        f".approval-card.visible z-index ({approval_visible_z}) must be strictly "
+        f"greater than .queue-card z-index ({queue_z}) so approval buttons "
+        f"remain clickable when both flyouts are open."
+    )
diff --git a/tests/test_blockquote_rendering.py b/tests/test_blockquote_rendering.py
new file mode 100644
index 00000000..fad580ad
--- /dev/null
+++ b/tests/test_blockquote_rendering.py
@@ -0,0 +1,215 @@
+"""Regression tests for the blockquote rendering fix (fix/blockquote-rendering).
+
+Root cause: the old rule was `s.replace(/^> (.+)$/gm, ...)` which had three bugs:
+  1. `.+` required at least one character — bare `>` lines passed through as literal `>`
+  2. Each line became its own `<blockquote>` — no grouping, so 10-line quotes became
+     10 stacked `<blockquote>` elements
+  3. Fenced code blocks inside blockquotes left orphaned `>` literals after the
+     fence-stash pass had consumed the code content
+
+Fix: group consecutive `>` lines into a single `<blockquote>`, handle bare `>` lines
+as `<br>`, and strip the `>` prefix before passing each line to `inlineMd()`.
+"""
+import re
+import pathlib
+
+UI_JS = (pathlib.Path(__file__).parent.parent / "static" / "ui.js").read_text(encoding="utf-8")
+
+# ---------------------------------------------------------------------------
+# Python mirror of the new blockquote rule + inlineMd (for behavioural tests)
+# ---------------------------------------------------------------------------
+
+import html as _html
+
+
+def _esc(s):
+    return _html.escape(str(s), quote=True)
+
+
+def _inline_md(t):
+    """Minimal inlineMd mirror — bold, italic, inline-code only."""
+    t = re.sub(r"`([^`\n]+)`", lambda m: f"<code>{_esc(m.group(1))}</code>", t)
+    t = re.sub(r"\*\*\*(.+?)\*\*\*", lambda m: f"<strong><em>{_esc(m.group(1))}</em></strong>", t)
+    t = re.sub(r"\*\*(.+?)\*\*", lambda m: f"<strong>{_esc(m.group(1))}</strong>", t)
+    t = re.sub(r"\*([^*\n]+)\*", lambda m: f"<em>{_esc(m.group(1))}</em>", t)
+    return t
+
+
+def _apply_blockquote(s):
+    """Python mirror of the new group-based blockquote rule in ui.js."""
+    def replacer(m):
+        block = m.group(0)
+        lines = block.split("\n")
+        # Drop a lone trailing ">" artifact that the regex can leave
+        while lines and lines[-1].strip() in (">", ""):
+            if lines[-1].strip() == ">":
+                lines.pop()
+                break
+            lines.pop()
+        processed = []
+        for l in lines:
+            stripped = re.sub(r"^>[ \t]?", "", l)
+            if stripped.strip() == "":
+                processed.append("<br>")
+            else:
+                processed.append(_inline_md(stripped))
+        inner = "\n".join(processed)
+        return f"<blockquote>{inner}</blockquote>"
+
+    return re.sub(r"((?:^>[^\n]*(?:\n|$))+)", replacer, s, flags=re.MULTILINE)
+
+
+# ---------------------------------------------------------------------------
+# Source-level structural tests
+# ---------------------------------------------------------------------------
+
+class TestBlockquoteSourceStructure:
+    """The new rule must be present in ui.js and the old single-line rule must be gone."""
+
+    def test_old_single_line_rule_removed(self):
+        """The old `.+` pattern that skipped blank lines must be gone."""
+        assert "replace(/^> (.+)$/gm" not in UI_JS, (
+            "Old single-line blockquote rule still present — it misses blank '>'"
+            " lines and creates one <blockquote> per line"
+        )
+
+    def test_new_group_rule_present(self):
+        """The new grouping regex must be present."""
+        assert "(?:^>[^\\n]*(?:\\n|$))+" in UI_JS, (
+            "New group-based blockquote rule not found in ui.js"
+        )
+
+    def test_prefix_strip_present(self):
+        """The fix must strip the '> ' prefix from each line."""
+        assert "replace(/^>[" in UI_JS or "replace(/^>[ " in UI_JS, (
+            "Expected prefix-strip pattern not found in the blockquote block"
+        )
+
+
+# ---------------------------------------------------------------------------
+# Behavioural tests (using the Python mirror)
+# ---------------------------------------------------------------------------
+
+class TestMultiLineBlockquote:
+    """Consecutive > lines must become ONE <blockquote>, not many."""
+
+    def test_single_line_still_works(self):
+        out = _apply_blockquote("> Hello world")
+        assert out.count("<blockquote>") == 1
+        assert "Hello world" in out
+        assert ">" not in out.replace("<blockquote>", "").replace("</blockquote>", "")
+
+    def test_two_consecutive_lines_grouped(self):
+        src = "> Line one\n> Line two"
+        out = _apply_blockquote(src)
+        assert out.count("<blockquote>") == 1, (
+            f"Expected 1 <blockquote>, got {out.count('<blockquote>')}: {out!r}"
+        )
+
+    def test_ten_lines_one_blockquote(self):
+        src = "\n".join(f"> Line {i}" for i in range(10))
+        out = _apply_blockquote(src)
+        assert out.count("<blockquote>") == 1
+
+    def test_two_separate_quotes_stay_separate(self):
+        src = "> First quote\n\n> Second quote"
+        out = _apply_blockquote(src)
+        # Each quote is its own group (separated by a blank line)
+        assert out.count("<blockquote>") == 2
+
+
+class TestBlankContinuationLines:
+    """Bare '>' lines (blank continuation) must not appear as literal '>'."""
+
+    def test_bare_gt_line_no_literal(self):
+        src = "> Para one\n>\n> Para two"
+        out = _apply_blockquote(src)
+        assert out.count("<blockquote>") == 1, f"Expected 1 blockquote: {out!r}"
+        # No stray '>' outside of HTML tags
+        text_only = re.sub(r"<[^>]+>", "", out)
+        assert ">" not in text_only, f"Literal '>' in text: {text_only!r}"
+
+    def test_bare_gt_no_space_handled(self):
+        """'>' with no space at all should also be consumed, not rendered literally."""
+        src = ">no space after"
+        out = _apply_blockquote(src)
+        assert out.count("<blockquote>") == 1
+        text_only = re.sub(r"<[^>]+>", "", out)
+        assert ">" not in text_only
+
+    def test_blank_line_becomes_br(self):
+        src = "> First\n>\n> Second"
+        out = _apply_blockquote(src)
+        assert "<br>" in out, f"Expected <br> for blank > line: {out!r}"
+
+
+class TestInlineMarkdownInsideBlockquote:
+    """Bold, italic, and inline code must still render correctly inside a blockquote."""
+
+    def test_bold_inside_blockquote(self):
+        out = _apply_blockquote("> This is **important**")
+        assert "<strong>" in out
+        assert "<blockquote>" in out
+
+    def test_inline_code_inside_blockquote(self):
+        out = _apply_blockquote("> Run `git status` first")
+        assert "<code>" in out
+        assert "<blockquote>" in out
+
+    def test_italic_inside_blockquote(self):
+        out = _apply_blockquote("> *emphasis* here")
+        assert "<em>" in out
+        assert "<blockquote>" in out
+
+
+class TestNoPhantomTrailingBr:
+    """The fix must drop both empty trailing lines (from a trailing \\n in the
+    match) and bare '>' artifacts. Without this, the common case — a blockquote
+    followed by another paragraph — renders with a phantom <br> right before
+    </blockquote>, leaving a visible blank line at the bottom of the quote.
+    """
+
+    def test_input_ending_with_newline_no_trailing_br(self):
+        """`> Hello\\n` must NOT produce `<blockquote>Hello\\n<br></blockquote>`."""
+        out = _apply_blockquote("> Hello\n")
+        assert "<br></blockquote>" not in out, (
+            f"Trailing <br> leaked inside the blockquote (phantom blank line): {out!r}"
+        )
+
+    def test_blockquote_followed_by_paragraph_no_trailing_br(self):
+        """The common real-world shape: quote + blank line + paragraph."""
+        src = "> Quoted text\n\nNormal paragraph"
+        out = _apply_blockquote(src)
+        assert "<br></blockquote>" not in out, (
+            f"Trailing <br> leaked inside blockquote when followed by paragraph: {out!r}"
+        )
+
+    def test_multiline_quote_ending_with_newline_no_trailing_br(self):
+        out = _apply_blockquote("> Line one\n> Line two\n")
+        assert "<br></blockquote>" not in out, (
+            f"Multi-line quote ending with \\n must not leave a trailing <br>: {out!r}"
+        )
+
+    def test_quote_with_blank_continuation_then_newline(self):
+        """`> A\\n>\\n> B\\n` — the internal `<br>` for the blank line stays,
+        but the trailing newline must not add a second `<br>` at the end."""
+        out = _apply_blockquote("> A\n>\n> B\n")
+        # Internal <br> for the blank-line continuation is intentional
+        assert "<br>" in out
+        # But there must not be a <br> immediately before the closing tag
+        assert "<br></blockquote>" not in out, (
+            f"Trailing <br> leaked at end of blockquote: {out!r}"
+        )
+
+
+class TestBlockquoteFollowedByParagraph:
+    """A blockquote followed by a normal paragraph must not bleed into each other."""
+
+    def test_non_blockquote_paragraph_untouched(self):
+        src = "> Quoted text\n\nNormal paragraph"
+        out = _apply_blockquote(src)
+        assert out.count("<blockquote>") == 1
+        assert "Normal paragraph" in out
+        # Normal paragraph must be outside the blockquote
+        after_bq = out[out.index("</blockquote>"):]
+        assert "Normal paragraph" in after_bq
diff --git a/tests/test_gateway_sync.py b/tests/test_gateway_sync.py
index 1667042f..f9b04e63 100644
--- a/tests/test_gateway_sync.py
+++ b/tests/test_gateway_sync.py
@@ -86,6 +86,14 @@ def _ensure_state_db():
             timestamp REAL NOT NULL
         );
     """)
+    for column, ddl in (
+        ('parent_session_id', 'ALTER TABLE sessions ADD COLUMN parent_session_id TEXT'),
+        ('ended_at', 'ALTER TABLE sessions ADD COLUMN ended_at REAL'),
+        ('end_reason', 'ALTER TABLE sessions ADD COLUMN end_reason TEXT'),
+    ):
+        existing = {row[1] for row in conn.execute("PRAGMA table_info(sessions)").fetchall()}
+        if column not in existing:
+            conn.execute(ddl)
     conn.commit()
     return conn
 
@@ -113,6 +121,50 @@ def _insert_gateway_session(conn, session_id='20260401_120000_abcdefgh', source=
     conn.commit()
 
 
+def _insert_agent_session_row(
+    conn,
+    session_id,
+    source='weixin',
+    title='Agent Session',
+    model='openai/gpt-5',
+    started_at=None,
+    parent_session_id=None,
+    ended_at=None,
+    end_reason=None,
+    messages=1,
+):
+    """Insert an agent session row with optional compression lineage."""
+    started_at = started_at or time.time()
+    conn.execute(
+        "INSERT OR REPLACE INTO sessions "
+        "(id, source, title, model, started_at, message_count, parent_session_id, ended_at, end_reason) "
+        "VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)",
+        (
+            session_id,
+            source,
+            title,
+            model,
+            started_at,
+            messages,
+            parent_session_id,
+            ended_at,
+            end_reason,
+        ),
+    )
+    conn.execute("DELETE FROM messages WHERE session_id = ?", (session_id,))
+    for i in range(messages):
+        conn.execute(
+            "INSERT INTO messages (session_id, role, content, timestamp) VALUES (?, ?, ?, ?)",
+            (
+                session_id,
+                'user' if i % 2 == 0 else 'assistant',
+                f'{title} message {i + 1}',
+                started_at + i,
+            ),
+        )
+    conn.commit()
+
+
 def _remove_test_sessions(conn, *session_ids):
     """Remove specific test sessions from state.db (parallel-safe cleanup)."""
     for sid in session_ids:
@@ -229,6 +281,347 @@ def test_gateway_watcher_hides_sessions_without_messages(monkeypatch):
             pass
 
 
+def test_compression_chain_collapses_to_latest_tip_in_sidebar():
+    """Show one logical agent conversation for a compression continuation chain."""
+    conn = _ensure_state_db()
+    ids_to_remove = ('chain_root_001', 'chain_empty_mid_001', 'chain_tip_001')
+    t0 = time.time() - 600
+    try:
+        _insert_agent_session_row(
+            conn,
+            'chain_root_001',
+            title='Magazine Style PPT Skill',
+            started_at=t0,
+            ended_at=t0 + 100,
+            end_reason='compression',
+            messages=3,
+        )
+        _insert_agent_session_row(
+            conn,
+            'chain_empty_mid_001',
+            title='Magazine Style PPT Skill #2',
+            started_at=t0 + 101,
+            parent_session_id='chain_root_001',
+            ended_at=t0 + 200,
+            end_reason='compression',
+            messages=0,
+        )
+        _insert_agent_session_row(
+            conn,
+            'chain_tip_001',
+            title='Magazine Style PPT Skill #3',
+            started_at=t0 + 201,
+            parent_session_id='chain_empty_mid_001',
+            messages=2,
+        )
+
+        post('/api/settings', {'show_cli_sessions': True})
+        data, status = get('/api/sessions')
+        assert status == 200
+        ids = {s.get('session_id') for s in data.get('sessions', [])}
+        tip = next((s for s in data.get('sessions', []) if s.get('session_id') == 'chain_tip_001'), None)
+
+        assert 'chain_tip_001' in ids
+        assert 'chain_root_001' not in ids
+        assert 'chain_empty_mid_001' not in ids
+        assert tip is not None
+        assert tip.get('title') == 'Magazine Style PPT Skill'
+        assert tip.get('message_count') == 2
+        # created_at = the chain head's started_at (preserves original conversation date)
+        assert abs(tip.get('created_at') - t0) < 0.01
+        # updated_at = the tip's last message timestamp so the sidebar entry
+        # bubbles to the top by true recency, not by the root's stale activity.
+        # tip messages are at t0+201 and t0+202, so last_activity = t0 + 202.
+        assert abs(tip.get('updated_at') - (t0 + 202)) < 0.01
+
+        from api.agent_sessions import read_importable_agent_session_rows
+
+        rows = read_importable_agent_session_rows(_get_state_db_path(), limit=None)
+        projected_tip = next((row for row in rows if row.get('id') == 'chain_tip_001'), None)
+        assert projected_tip is not None
+        assert projected_tip.get('title') == 'Magazine Style PPT Skill'
+        assert projected_tip.get('_lineage_root_id') == 'chain_root_001'
+        assert projected_tip.get('_lineage_tip_id') == 'chain_tip_001'
+        assert projected_tip.get('_compression_segment_count') == 3
+    finally:
+        try:
+            _remove_test_sessions(conn, *ids_to_remove)
+            conn.close()
+        except Exception:
+            pass
+        post('/api/settings', {'show_cli_sessions': False})
+
+
+def test_compression_chain_with_empty_latest_tip_falls_back_to_latest_importable_segment():
+    """Empty latest tips should not make the whole conversation disappear."""
+    conn = _ensure_state_db()
+    ids_to_remove = ('empty_tip_root_001', 'empty_tip_001')
+    t0 = time.time() - 500
+    try:
+        _insert_agent_session_row(
+            conn,
+            'empty_tip_root_001',
+            title='Long Conversation',
+            started_at=t0,
+            ended_at=t0 + 100,
+            end_reason='compression',
+            messages=2,
+        )
+        _insert_agent_session_row(
+            conn,
+            'empty_tip_001',
+            title='Long Conversation #2',
+            started_at=t0 + 101,
+            parent_session_id='empty_tip_root_001',
+            messages=0,
+        )
+
+        post('/api/settings', {'show_cli_sessions': True})
+        data, status = get('/api/sessions')
+        assert status == 200
+        ids = {s.get('session_id') for s in data.get('sessions', [])}
+
+        assert 'empty_tip_root_001' in ids
+        assert 'empty_tip_001' not in ids
+        root = next((s for s in data.get('sessions', []) if s.get('session_id') == 'empty_tip_root_001'), None)
+        assert root and root.get('title') == 'Long Conversation'
+    finally:
+        try:
+            _remove_test_sessions(conn, *ids_to_remove)
+            conn.close()
+        except Exception:
+            pass
+        post('/api/settings', {'show_cli_sessions': False})
+
+
+def test_compression_chain_with_all_empty_segments_is_hidden():
+    """A compression chain with no importable segment should not appear."""
+    conn = _ensure_state_db()
+    ids_to_remove = ('all_empty_root_001', 'all_empty_tip_001')
+    t0 = time.time() - 450
+    try:
+        _insert_agent_session_row(
+            conn,
+            'all_empty_root_001',
+            title='Empty Long Conversation',
+            started_at=t0,
+            ended_at=t0 + 100,
+            end_reason='compression',
+            messages=0,
+        )
+        _insert_agent_session_row(
+            conn,
+            'all_empty_tip_001',
+            title='Empty Long Conversation #2',
+            started_at=t0 + 101,
+            parent_session_id='all_empty_root_001',
+            messages=0,
+        )
+
+        post('/api/settings', {'show_cli_sessions': True})
+        data, status = get('/api/sessions')
+        assert status == 200
+        ids = {s.get('session_id') for s in data.get('sessions', [])}
+
+        assert 'all_empty_root_001' not in ids
+        assert 'all_empty_tip_001' not in ids
+    finally:
+        try:
+            _remove_test_sessions(conn, *ids_to_remove)
+            conn.close()
+        except Exception:
+            pass
+        post('/api/settings', {'show_cli_sessions': False})
+
+
+def test_non_compression_child_is_not_collapsed_into_parent():
+    """Parent/child relationships that are not compression continuations stay flat."""
+    conn = _ensure_state_db()
+    ids_to_remove = ('branch_parent_001', 'branch_child_001')
+    t0 = time.time() - 400
+    try:
+        _insert_agent_session_row(
+            conn,
+            'branch_parent_001',
+            title='Branch Parent',
+            started_at=t0,
+            ended_at=t0 + 100,
+            end_reason='branched',
+            messages=2,
+        )
+        _insert_agent_session_row(
+            conn,
+            'branch_child_001',
+            title='Branch Child',
+            started_at=t0 + 101,
+            parent_session_id='branch_parent_001',
+            messages=2,
+        )
+
+        from api.agent_sessions import read_importable_agent_session_rows
+
+        rows = read_importable_agent_session_rows(_get_state_db_path(), limit=None)
+        ids = {row.get('id') for row in rows}
+
+        assert 'branch_parent_001' in ids
+        assert 'branch_child_001' in ids
+    finally:
+        try:
+            _remove_test_sessions(conn, *ids_to_remove)
+            conn.close()
+        except Exception:
+            pass
+
+
+def test_agent_session_limit_applies_after_compression_projection():
+    """A long raw chain should count as one logical sidebar row before limiting."""
+    conn = _ensure_state_db()
+    chain_ids = [f'limit_chain_{i:03d}' for i in range(8)]
+    standalone_id = 'limit_standalone_001'
+    t0 = time.time() - 300
+    try:
+        for i, sid in enumerate(chain_ids):
+            _insert_agent_session_row(
+                conn,
+                sid,
+                title=f'Limit Chain #{i + 1}',
+                started_at=t0 + i,
+                parent_session_id=chain_ids[i - 1] if i else None,
+                ended_at=t0 + i + 0.5 if i < len(chain_ids) - 1 else None,
+                end_reason='compression' if i < len(chain_ids) - 1 else None,
+                messages=1,
+            )
+        _insert_agent_session_row(
+            conn,
+            standalone_id,
+            title='Limit Standalone',
+            started_at=t0 + 20,
+            messages=1,
+        )
+
+        from api.agent_sessions import read_importable_agent_session_rows
+
+        rows = read_importable_agent_session_rows(_get_state_db_path(), limit=2)
+        ids = [row.get('id') for row in rows]
+
+        assert len(rows) == 2
+        assert chain_ids[-1] in ids
+        assert standalone_id in ids
+        assert not any(sid in ids for sid in chain_ids[:-1])
+        chain = next(row for row in rows if row.get('id') == chain_ids[-1])
+        assert chain.get('title') == 'Limit Chain #1'
+        assert chain.get('_lineage_root_id') == chain_ids[0]
+        assert chain.get('_compression_segment_count') == len(chain_ids)
+    finally:
+        try:
+            _remove_test_sessions(conn, *(chain_ids + [standalone_id]))
+            conn.close()
+        except Exception:
+            pass
+
+
+def test_compression_chain_bubbles_to_top_by_tip_activity():
+    """An actively-used compression chain must surface in the sidebar by its
+    TIP's last activity, not by the (stale) root's last activity.
+
+    Without overriding ``last_activity`` from the tip, a long-running chain
+    whose tip is being actively edited NOW would sort by the root's old
+    timestamp and fall below recently touched standalone sessions — the
+    inverse of what users expect from "Show agent sessions" sorted by
+    recency. This regression test pins the override.
+    """
+    conn = _ensure_state_db()
+    ids_to_remove = ('bubble_root_001', 'bubble_tip_001', 'bubble_standalone_001')
+    now = time.time()
+    # Root started long ago; tip is being edited "now" (very recent message)
+    root_started = now - 30 * 86400
+    root_ended = now - 28 * 86400
+    tip_started = root_ended + 1
+    tip_latest_msg = now - 5  # 5 seconds ago — most recent activity in the DB
+    # A standalone session active 2 days ago — older than tip, much newer
+    # than the root. Without the fix, the chain row sorts by ROOT's age and
+    # standalone wins; with the fix, the chain wins.
+    standalone_msg = now - 2 * 86400
+    try:
+        _insert_agent_session_row(
+            conn,
+            'bubble_root_001',
+            title='Bubble Root',
+            started_at=root_started,
+            ended_at=root_ended,
+            end_reason='compression',
+            messages=2,
+        )
+        # Override message timestamps so root's last_activity is genuinely old.
+        conn.execute("DELETE FROM messages WHERE session_id = 'bubble_root_001'")
+        conn.execute(
+            "INSERT INTO messages (session_id, role, content, timestamp) VALUES (?, ?, ?, ?)",
+            ('bubble_root_001', 'user', 'old root msg', root_started + 60),
+        )
+        _insert_agent_session_row(
+            conn,
+            'bubble_tip_001',
+            title='Bubble Tip',
+            started_at=tip_started,
+            parent_session_id='bubble_root_001',
+            messages=1,
+        )
+        conn.execute("DELETE FROM messages WHERE session_id = 'bubble_tip_001'")
+        conn.execute(
+            "INSERT INTO messages (session_id, role, content, timestamp) VALUES (?, ?, ?, ?)",
+            ('bubble_tip_001', 'user', 'fresh tip msg', tip_latest_msg),
+        )
+        _insert_agent_session_row(
+            conn,
+            'bubble_standalone_001',
+            title='Bubble Standalone',
+            started_at=now - 2 * 86400 - 60,
+            messages=1,
+        )
+        conn.execute("DELETE FROM messages WHERE session_id = 'bubble_standalone_001'")
+        conn.execute(
+            "INSERT INTO messages (session_id, role, content, timestamp) VALUES (?, ?, ?, ?)",
+            ('bubble_standalone_001', 'user', 'standalone msg', standalone_msg),
+        )
+        conn.commit()
+
+        from api.agent_sessions import read_importable_agent_session_rows
+
+        rows = read_importable_agent_session_rows(_get_state_db_path(), limit=200)
+        ids = [row.get('id') for row in rows]
+        # Filter out unrelated rows from the shared DB
+        ids = [i for i in ids if i in ('bubble_root_001', 'bubble_tip_001', 'bubble_standalone_001')]
+
+        assert 'bubble_tip_001' in ids, (
+            f"Compression tip must appear in projected output. ids={ids}"
+        )
+        assert 'bubble_root_001' not in ids, (
+            "Compression root row must be hidden once the tip is the active row."
+        )
+
+        tip_pos = ids.index('bubble_tip_001')
+        standalone_pos = ids.index('bubble_standalone_001') if 'bubble_standalone_001' in ids else -1
+        assert standalone_pos == -1 or tip_pos < standalone_pos, (
+            f"Active compression tip (last msg 5s ago) must sort BEFORE standalone "
+            f"session (last msg 2d ago). Got order: {ids}. "
+            f"This indicates merged.last_activity is the root's stale value, "
+            f"not the tip's recent value."
+        )
+
+        tip_row = next(r for r in rows if r['id'] == 'bubble_tip_001')
+        assert abs(tip_row['last_activity'] - tip_latest_msg) < 0.01, (
+            f"Projected tip's last_activity must equal the tip's most recent "
+            f"message timestamp ({tip_latest_msg}), not the root's "
+            f"({root_started + 60}). Got: {tip_row['last_activity']}"
+        )
+    finally:
+        try:
+            _remove_test_sessions(conn, *ids_to_remove)
+            conn.close()
+        except Exception:
+            pass
+
+
 def test_gateway_sessions_excluded_when_disabled():
     """Gateway sessions are NOT returned when show_cli_sessions is off."""
     conn = _ensure_state_db()
diff --git a/tests/test_issue342.py b/tests/test_issue342.py
index be3b165d..e733d056 100644
--- a/tests/test_issue342.py
+++ b/tests/test_issue342.py
@@ -31,7 +31,7 @@ def test_autolink_regex_in_rendermd():
     rendermd_start = content.find('function renderMd(raw){')
     assert rendermd_start != -1, "renderMd function not found in ui.js"
     # Find the closing brace after renderMd (look for the autolink pattern within it)
-    rendermd_body = content[rendermd_start:rendermd_start + 5000]
+    rendermd_body = content[rendermd_start:rendermd_start + 10000]
     assert 'https?:\\/\\/' in rendermd_body, (
         "Autolink regex (https?:\\/\\/) not found inside renderMd() body."
     )
diff --git a/tests/test_reasoning_chip_btw_fixes.py b/tests/test_reasoning_chip_btw_fixes.py
index 3dbbb401..608d612d 100644
--- a/tests/test_reasoning_chip_btw_fixes.py
+++ b/tests/test_reasoning_chip_btw_fixes.py
@@ -28,6 +28,7 @@ INDEX = (REPO / "static" / "index.html").read_text(encoding="utf-8")
 UI_JS = (REPO / "static" / "ui.js").read_text(encoding="utf-8")
 COMMANDS_JS = (REPO / "static" / "commands.js").read_text(encoding="utf-8")
 MESSAGES_JS = (REPO / "static" / "messages.js").read_text(encoding="utf-8")
+STYLE_CSS = (REPO / "static" / "style.css").read_text(encoding="utf-8")
 
 
 # ── #1 dropdown escapes composer-left ─────────────────────────────────────────
@@ -115,6 +116,57 @@ class TestReasoningChipIcon:
         )
 
 
+# ── #1068 None/default reasoning chip stays visible ──────────────────────────
+
+
+class TestReasoningChipNoneState:
+    """Reasoning effort is a current setting like model selection.  Setting it
+    to None disables reasoning, but the chip must remain visible so users can
+    see and change the current level."""
+
+    def get_apply_reasoning_chip(self):
+        m = re.search(
+            r"function\s+_applyReasoningChip\b[\s\S]*?^}",
+            UI_JS,
+            re.MULTILINE,
+        )
+        assert m, "_applyReasoningChip not found in ui.js"
+        return m.group(0)
+
+    def test_none_and_default_do_not_hide_reasoning_chip(self):
+        fn = self.get_apply_reasoning_chip()
+        assert "wrap.style.display='';" in fn, (
+            "_applyReasoningChip must show the reasoning chip even for empty/"
+            "default or 'none' effort values"
+        )
+        assert "if(!eff" not in fn and "wrap.style.display='none'" not in fn, (
+            "_applyReasoningChip must not use a truthy guard that hides the "
+            "chip for the valid 'none' state"
+        )
+        assert "wrap.style.display='none'" not in fn, (
+            "the None/default reasoning state should be visible, not hidden"
+        )
+
+    def test_none_and_default_have_visible_labels(self):
+        assert "if(effort==='none') return 'None';" in UI_JS, (
+            "the disabled reasoning state must render a visible 'None' label"
+        )
+        assert "if(!effort) return 'Default';" in UI_JS, (
+            "the unset reasoning state must render a visible 'Default' label"
+        )
+
+    def test_none_and_default_are_visually_inactive_not_missing(self):
+        fn = self.get_apply_reasoning_chip()
+        assert "chip.classList.toggle('inactive',inactive)" in fn, (
+            "None/default should be shown with an inactive visual treatment "
+            "instead of removing the chip"
+        )
+        assert ".composer-reasoning-chip.inactive" in STYLE_CSS, (
+            "the inactive chip state needs a CSS rule so the visible None/"
+            "default state is intentionally muted"
+        )
+
+
 # ── #3 /reasoning immediately updates chip ────────────────────────────────────
 
 
diff --git a/tests/test_reasoning_chip_js_behaviour.py b/tests/test_reasoning_chip_js_behaviour.py
new file mode 100644
index 00000000..aeb56881
--- /dev/null
+++ b/tests/test_reasoning_chip_js_behaviour.py
@@ -0,0 +1,199 @@
+"""Behavioural tests that drive the actual `_applyReasoningChip()` from
+static/ui.js via node, not just a regex over the source.
+
+The static checks in test_reasoning_chip_btw_fixes.py confirm the *shape*
+of the function (no `display='none'`, the right toggle call exists, etc.)
+but they pass even if a runtime detail is wrong — e.g. if `inactive` were
+inverted, or `_normalizeReasoningEffort` mishandled whitespace, or the
+label fell through to a wrong value for an unknown input.
+
+This file pins the actual rendered output for every effort state so the
+chip's None/Default visibility cannot silently regress.
+"""
+import os
+import shutil
+import subprocess
+from pathlib import Path
+
+import pytest
+
+REPO_ROOT = Path(__file__).parent.parent.resolve()
+UI_JS_PATH = REPO_ROOT / "static" / "ui.js"
+
+NODE = shutil.which("node")
+pytestmark = pytest.mark.skipif(NODE is None, reason="node not on PATH")
+
+
+_DRIVER_SRC = r"""
+const fs = require('fs');
+const src = fs.readFileSync(process.argv[2], 'utf8');
+
+function makeEl() {
+  return {
+    style: {},
+    classList: {
+      _set: new Set(),
+      add(c){this._set.add(c)},
+      remove(c){this._set.delete(c)},
+      toggle(c, on){
+        const want = on === undefined ? !this._set.has(c) : Boolean(on);
+        if (want) this._set.add(c); else this._set.delete(c);
+      },
+      contains(c){return this._set.has(c)},
+    },
+    dataset: {},
+    title: '',
+    textContent: '',
+    querySelectorAll(){return []},
+  };
+}
+
+const els = {
+  composerReasoningWrap: makeEl(),
+  composerReasoningLabel: makeEl(),
+  composerReasoningChip: makeEl(),
+  composerReasoningDropdown: makeEl(),
+};
+els.composerReasoningWrap.style.display = 'none'; // mirrors the HTML default
+
+global.window = {};
+global.document = {
+  createElement: () => makeEl(),
+  addEventListener: () => {},
+  querySelectorAll: () => [],
+  querySelector: () => null,
+};
+global.$ = id => els[id] || null;
+global.api = () => ({ then: () => ({ catch: () => {} }), catch: () => {} });
+
+function extractFunc(name) {
+  const re = new RegExp('function\\s+' + name + '\\s*\\(');
+  const start = src.search(re);
+  if (start < 0) throw new Error(name + ' not found');
+  let i = src.indexOf('{', start);
+  let depth = 1; i++;
+  while (depth > 0 && i < src.length) {
+    if (src[i] === '{') depth++;
+    else if (src[i] === '}') depth--;
+    i++;
+  }
+  return src.slice(start, i);
+}
+eval(extractFunc('_normalizeReasoningEffort'));
+eval(extractFunc('_formatReasoningEffortLabel'));
+eval(extractFunc('_highlightReasoningOption'));
+eval(extractFunc('_applyReasoningChip'));
+
+const input = JSON.parse(process.argv[3]);
+_applyReasoningChip(input);
+const result = {
+  display: els.composerReasoningWrap.style.display,
+  label:   els.composerReasoningLabel.textContent,
+  inactive: els.composerReasoningChip.classList.contains('inactive'),
+  title:   els.composerReasoningChip.title,
+};
+process.stdout.write(JSON.stringify(result));
+"""
+
+
+@pytest.fixture(scope="module")
+def driver_path(tmp_path_factory):
+    p = tmp_path_factory.mktemp("reasoning_driver") / "driver.js"
+    p.write_text(_DRIVER_SRC, encoding="utf-8")
+    return str(p)
+
+
+def _apply(driver_path, value):
+    """Run _applyReasoningChip(value) against the actual ui.js."""
+    import json as _json
+    result = subprocess.run(
+        [NODE, driver_path, str(UI_JS_PATH), _json.dumps(value)],
+        capture_output=True, text=True, timeout=10,
+    )
+    if result.returncode != 0:
+        raise RuntimeError(f"node driver failed: {result.stderr}")
+    return _json.loads(result.stdout)
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# The chip MUST stay visible for every effort state (issue #1068).  This used
+# to be hidden for !eff and 'none', and the source-regex tests in
+# test_reasoning_chip_btw_fixes.py verify the literal `display='none'` is gone
+# — but only a behavioural check confirms the wrap actually receives `''`.
+# ─────────────────────────────────────────────────────────────────────────────
+
+
+class TestChipAlwaysVisible:
+
+    def test_empty_string_shows_chip_with_default_label(self, driver_path):
+        out = _apply(driver_path, "")
+        assert out["display"] == "", f"empty effort must show the chip: {out}"
+        assert out["label"] == "Default"
+        assert out["inactive"] is True
+
+    def test_null_shows_chip_with_default_label(self, driver_path):
+        out = _apply(driver_path, None)
+        assert out["display"] == ""
+        assert out["label"] == "Default"
+        assert out["inactive"] is True
+
+    def test_none_shows_chip_with_none_label(self, driver_path):
+        """The bug from #1068 — 'none' must NOT hide the chip."""
+        out = _apply(driver_path, "none")
+        assert out["display"] == "", (
+            f"'none' must show the chip (the regression that started #1068): {out}"
+        )
+        assert out["label"] == "None"
+        assert out["inactive"] is True
+
+    def test_low_shows_chip_active(self, driver_path):
+        out = _apply(driver_path, "low")
+        assert out["display"] == ""
+        assert out["label"] == "low"
+        assert out["inactive"] is False
+
+    def test_high_shows_chip_active(self, driver_path):
+        out = _apply(driver_path, "high")
+        assert out["display"] == ""
+        assert out["inactive"] is False
+
+
+class TestNormalizationEdgeCases:
+    """Pin the input-normalisation contract so it can't silently shift."""
+
+    def test_uppercase_normalises(self, driver_path):
+        # Even though the API and slash command use lowercase, defensive
+        # normalisation matters — copy/paste of an uppercase value or a
+        # mis-cased server response shouldn't break the chip.
+        out = _apply(driver_path, "NONE")
+        assert out["label"] == "None"
+        assert out["inactive"] is True
+
+    def test_whitespace_trimmed(self, driver_path):
+        out = _apply(driver_path, "  none  ")
+        assert out["label"] == "None"
+        assert out["inactive"] is True
+
+    def test_unknown_value_falls_through_visible(self, driver_path):
+        # Defensive: unknown effort still shows the chip rather than hiding.
+        out = _apply(driver_path, "banana")
+        assert out["display"] == ""
+        assert out["label"] == "banana"
+        assert out["inactive"] is False
+
+
+class TestTitleAttributeAccessibility:
+    """The chip's `title` is the hover tooltip and a screen-reader hint —
+    confirm it always carries the current state in human-readable form."""
+
+    def test_title_has_default_label_for_unset(self, driver_path):
+        out = _apply(driver_path, "")
+        assert out["title"] == "Reasoning effort: Default"
+
+    def test_title_has_none_label_for_none(self, driver_path):
+        out = _apply(driver_path, "none")
+        assert out["title"] == "Reasoning effort: None"
+
+    def test_title_has_active_label_for_high(self, driver_path):
+        out = _apply(driver_path, "high")
+        assert out["title"] == "Reasoning effort: high"
diff --git a/tests/test_renderer_comprehensive.py b/tests/test_renderer_comprehensive.py
new file mode 100644
index 00000000..74034051
--- /dev/null
+++ b/tests/test_renderer_comprehensive.py
@@ -0,0 +1,366 @@
+"""Comprehensive renderer audit tests for static/ui.js renderMd().
+
+This file covers the full suite of markdown constructs an LLM might produce,
+with a focus on edge cases and combinations. Tests are grouped by construct.
+
+Python mirrors the renderMd/inlineMd pipeline at the level needed for each
+test — either source-level assertions (checking the JS source directly) or
+behavioural assertions (checking rendered HTML via a Python mirror).
+"""
+import re
+import pathlib
+
+UI_JS = (pathlib.Path(__file__).parent.parent / "static" / "ui.js").read_text(encoding="utf-8")
+
+import html as _html
+
+
+def _esc(s):
+    return _html.escape(str(s), quote=True)
+
+
+def _inline_md(t):
+    """Mirror of inlineMd() in ui.js — processes one line of text."""
+    _code_stash = []
+    t = re.sub(r"`([^`\n]+)`",
+               lambda m: (_code_stash.append(f"<code>{_esc(m.group(1))}</code>")
+                          or f"\x00C{len(_code_stash)-1}\x00"), t)
+    t = re.sub(r"\*\*\*(.+?)\*\*\*", lambda m: f"<strong><em>{_esc(m.group(1))}</em></strong>", t)
+    t = re.sub(r"\*\*(.+?)\*\*",     lambda m: f"<strong>{_esc(m.group(1))}</strong>", t)
+    t = re.sub(r"\*([^*\n]+)\*",     lambda m: f"<em>{_esc(m.group(1))}</em>", t)
+    t = re.sub(r"~~(.+?)~~",         lambda m: f"<del>{_esc(m.group(1))}</del>", t)
+    t = re.sub(r"\x00C(\d+)\x00", lambda m: _code_stash[int(m.group(1))], t)
+    return t
+
+
+def _apply_blockquotes(src):
+    """Mirror of _applyBlockquotes() — handles nested + lists + blank lines."""
+    def replacer(m):
+        block = m.group(0)
+        lines = block.split("\n")
+        while lines and (lines[-1].strip() in (">", "")):
+            if lines[-1].strip() == ">":
+                lines.pop(); break
+            lines.pop()
+        stripped = [re.sub(r"^>[ \t]?", "", l) for l in lines]
+        inner_raw = "\n".join(stripped)
+        if re.search(r"^>", inner_raw, re.MULTILINE):
+            inner = _apply_blockquotes(inner_raw)
+        elif re.search(r"^(  )?[-*+] .+", inner_raw, re.MULTILINE):
+            def inner_list(lb):
+                ll = lb.strip().split("\n"); h = "<ul>"
+                for li in ll:
+                    txt = re.sub(r"^ {0,4}[-*+] ", "", li)
+                    if re.match(r"\[x\] ", txt, re.I): ih = f"✅ {_inline_md(txt[4:])}"
+                    elif txt.startswith("[ ] "): ih = f"☐ {_inline_md(txt[4:])}"
+                    else: ih = _inline_md(txt)
+                    h += f"<li>{ih}</li>"
+                return h + "</ul>"
+            inner = re.sub(r"((?:^(?:  )?[-*+] .+\n?)+)", lambda m2: inner_list(m2.group(0)),
+                           inner_raw, flags=re.MULTILINE)
+        else:
+            inner = "\n".join("<br>" if l.strip() == "" else _inline_md(l) for l in stripped)
+        return f"<blockquote>{inner}</blockquote>"
+    return re.sub(r"((?:^>[^\n]*(?:\n|$))+)", replacer, src, flags=re.MULTILINE)
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Source-level structural checks (JS must contain these patterns)
+# ─────────────────────────────────────────────────────────────────────────────
+
+class TestSourceStructure:
+    """Verify key patterns are present in ui.js."""
+
+    def test_crlf_normalisation_present(self):
+        assert ".replace(/\\r\\n/g,'\\n').replace(/\\r/g,'\\n')" in UI_JS, (
+            "renderMd must normalise \\r\\n and bare \\r to \\n at the start"
+        )
+
+    def test_strikethrough_in_inline_md(self):
+        assert "~~(.+?)~~" in UI_JS and "<del>" in UI_JS, (
+            "inlineMd must handle ~~strikethrough~~ → <del>"
+        )
+
+    def test_del_in_safe_tags(self):
+        assert "del" in UI_JS and "SAFE_TAGS" in UI_JS, (
+            "<del> must be in SAFE_TAGS so it is not HTML-escaped"
+        )
+
+    def test_del_in_safe_inline(self):
+        # SAFE_INLINE is used inside inlineMd
+        safe_inline_idx = UI_JS.find("SAFE_INLINE")
+        assert safe_inline_idx >= 0
+        window = UI_JS[safe_inline_idx: safe_inline_idx + 100]
+        assert "del" in window, "<del> must be in SAFE_INLINE"
+
+    def test_task_list_checked_handled(self):
+        assert "task-done" in UI_JS or "\\u2705" in UI_JS or "✅" in UI_JS, (
+            "Checked task list items [x] must produce a ✅ or task-done class"
+        )
+
+    def test_task_list_unchecked_handled(self):
+        assert "task-todo" in UI_JS or "\\u2610" in UI_JS or "☐" in UI_JS, (
+            "Unchecked task list items [ ] must produce ☐ or task-todo class"
+        )
+
+    def test_nested_blockquote_recurse(self):
+        assert "_applyBlockquotes" in UI_JS, (
+            "Blockquote handler must use a named function for recursive nesting"
+        )
+
+    def test_blockquote_handler_is_function(self):
+        assert "function _applyBlockquotes" in UI_JS, (
+            "Must define _applyBlockquotes as a named inner function for recursion"
+        )
+
+    def test_old_single_line_blockquote_removed(self):
+        assert "replace(/^> (.+)$/gm" not in UI_JS, (
+            "Old single-line blockquote rule must be removed"
+        )
+
+    def test_h1_h2_h3_handled(self):
+        for h in ("h1", "h2", "h3"):
+            assert f"<{h}>" in UI_JS or f"`<{h}>" in UI_JS
+
+    def test_ordered_list_value_attr(self):
+        assert 'value=' in UI_JS, "Ordered list items must use value= to preserve numbering"
+
+    def test_table_handler_present(self):
+        assert "<table>" in UI_JS and "<thead>" in UI_JS
+
+    def test_fenced_code_lang_header(self):
+        assert "pre-header" in UI_JS
+
+    def test_autolink_present(self):
+        # JS stores regex slashes as \/ — search for both forms
+        assert ("https?:\\/\\/" in UI_JS or "https?://" in UI_JS) and "target=\"_blank\"" in UI_JS
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Behavioural: inline formatting
+# ─────────────────────────────────────────────────────────────────────────────
+
+class TestInlineFormatting:
+
+    def test_bold(self):
+        assert _inline_md("**bold**") == "<strong>bold</strong>"
+
+    def test_italic(self):
+        assert _inline_md("*italic*") == "<em>italic</em>"
+
+    def test_bold_italic(self):
+        out = _inline_md("***bi***")
+        assert "<strong><em>" in out
+
+    def test_strikethrough(self):
+        out = _inline_md("~~deleted~~")
+        assert "<del>deleted</del>" == out
+
+    def test_strikethrough_inline(self):
+        out = _inline_md("keep ~~remove~~ keep")
+        assert "<del>remove</del>" in out
+        assert "keep" in out
+
+    def test_inline_code(self):
+        out = _inline_md("`git status`")
+        assert "<code>git status</code>" in out
+
+    def test_strikethrough_inside_code_not_processed(self):
+        out = _inline_md("`~~not deleted~~`")
+        assert "<del>" not in out
+        assert "~~not deleted~~" in out
+
+    def test_bold_with_inline_code(self):
+        # **`code`** → <strong><code>code</code></strong>
+        out = _inline_md("**`code`**")
+        # The code stash protects the backtick span from bold regex
+        assert "<code>" in out
+
+    def test_xss_in_bold(self):
+        out = _inline_md("**<script>alert(1)</script>**")
+        assert "<script>" not in out
+
+    def test_xss_in_strikethrough(self):
+        out = _inline_md("~~<img onerror=alert(1)>~~")
+        assert "onerror" not in out.lower() or "&lt;" in out
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Behavioural: blockquotes
+# ─────────────────────────────────────────────────────────────────────────────
+
+class TestBlockquotes:
+
+    def test_single_line(self):
+        out = _apply_blockquotes("> Hello")
+        assert out.count("<blockquote>") == 1
+        assert "Hello" in out
+
+    def test_multi_line_grouped(self):
+        out = _apply_blockquotes("> Line one\n> Line two\n> Line three")
+        assert out.count("<blockquote>") == 1
+
+    def test_blank_continuation_no_literal_gt(self):
+        out = _apply_blockquotes("> Para one\n>\n> Para two")
+        assert out.count("<blockquote>") == 1
+        text = re.sub(r"<[^>]+>", "", out)
+        assert ">" not in text, f"Literal > in output: {text!r}"
+
+    def test_blank_continuation_becomes_br(self):
+        out = _apply_blockquotes("> Para one\n>\n> Para two")
+        assert "<br>" in out
+
+    def test_bare_gt_no_space(self):
+        out = _apply_blockquotes(">no space after")
+        assert out.count("<blockquote>") == 1
+        assert "no space after" in out
+
+    def test_two_separate_blockquotes(self):
+        out = _apply_blockquotes("> First\n\n> Second")
+        assert out.count("<blockquote>") == 2
+
+    def test_inline_markdown_in_blockquote(self):
+        out = _apply_blockquotes("> **bold** and *italic*")
+        assert "<strong>" in out and "<em>" in out and "<blockquote>" in out
+
+    def test_inline_code_in_blockquote(self):
+        out = _apply_blockquotes("> run `git status` first")
+        assert "<code>" in out and "<blockquote>" in out
+
+    def test_strikethrough_in_blockquote(self):
+        out = _apply_blockquotes("> ~~old~~ new")
+        assert "<del>" in out and "<blockquote>" in out
+
+    def test_nested_blockquote_double(self):
+        out = _apply_blockquotes(">> deeply nested")
+        assert out.count("<blockquote>") == 2
+
+    def test_nested_blockquote_outer_and_inner(self):
+        out = _apply_blockquotes("> outer\n>> inner line")
+        assert out.count("<blockquote>") == 2
+
+    def test_list_inside_blockquote(self):
+        out = _apply_blockquotes("> - item one\n> - item two")
+        assert "<ul>" in out and "<li>" in out and "<blockquote>" in out
+
+    def test_task_list_inside_blockquote(self):
+        out = _apply_blockquotes("> - [x] done\n> - [ ] todo")
+        assert "✅" in out or "task-done" in out
+        assert "☐" in out or "task-todo" in out
+        assert "<blockquote>" in out
+
+    def test_blockquote_followed_by_paragraph(self):
+        out = _apply_blockquotes("> Quoted\n\nNormal text")
+        assert out.count("<blockquote>") == 1
+        after = out[out.index("</blockquote>"):]
+        assert "Normal text" in after
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Behavioural: task lists
+# ─────────────────────────────────────────────────────────────────────────────
+
+class TestTaskLists:
+
+    def _apply_list(self, block):
+        lines = block.strip().split("\n")
+        html = "<ul>"
+        for l in lines:
+            text = re.sub(r"^ {0,4}[-*+] ", "", l)
+            if re.match(r"\[x\] ", text, re.I):
+                html += f"<li>✅ {_inline_md(text[4:])}</li>"
+            elif text.startswith("[ ] "):
+                html += f"<li>☐ {_inline_md(text[4:])}</li>"
+            else:
+                html += f"<li>{_inline_md(text)}</li>"
+        return html + "</ul>"
+
+    def test_checked_item(self):
+        out = self._apply_list("- [x] done task")
+        assert "✅" in out and "done task" in out
+
+    def test_checked_uppercase_X(self):
+        out = self._apply_list("- [X] also done")
+        assert "✅" in out
+
+    def test_unchecked_item(self):
+        out = self._apply_list("- [ ] pending task")
+        assert "☐" in out and "pending task" in out
+
+    def test_mixed_task_and_normal(self):
+        out = self._apply_list("- [x] done\n- [ ] todo\n- normal")
+        assert "✅" in out and "☐" in out
+        assert "<li>" in out
+
+    def test_task_item_with_bold(self):
+        out = self._apply_list("- [x] **important** task")
+        assert "✅" in out and "<strong>" in out
+
+    def test_non_task_list_unaffected(self):
+        out = self._apply_list("- regular item\n- another item")
+        assert "✅" not in out and "☐" not in out
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Behavioural: strikethrough edge cases
+# ─────────────────────────────────────────────────────────────────────────────
+
+class TestStrikethrough:
+
+    def test_basic(self):
+        assert _inline_md("~~text~~") == "<del>text</del>"
+
+    def test_multiword(self):
+        out = _inline_md("~~multiple words here~~")
+        assert "<del>multiple words here</del>" == out
+
+    def test_inside_bold(self):
+        # **~~text~~** — outer bold picks up the raw ~~ which inlineMd then handles
+        # In practice bold runs first in the JS, then ~~ — let's verify the pattern exists
+        out = _inline_md("~~inside strikethrough~~")
+        assert "<del>" in out
+
+    def test_xss_escaped(self):
+        out = _inline_md("~~<b>bad</b>~~")
+        assert "<b>" not in out or "&lt;b&gt;" in out
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Edge-case combinations
+# ─────────────────────────────────────────────────────────────────────────────
+
+class TestEdgeCases:
+
+    def test_empty_string(self):
+        out = _apply_blockquotes("")
+        assert out == ""
+
+    def test_no_blockquote(self):
+        s = "just normal text"
+        assert _apply_blockquotes(s) == s
+
+    def test_crlf_in_blockquote(self):
+        # \r\n should not produce literal \r in output
+        src = "> line one\r\n> line two"
+        # First normalise \r\n (as renderMd does)
+        src = src.replace("\r\n", "\n")
+        out = _apply_blockquotes(src)
+        assert "\r" not in out
+        assert out.count("<blockquote>") == 1
+
+    def test_blockquote_with_code_and_nested(self):
+        src = "> `code`\n>> nested"
+        out = _apply_blockquotes(src)
+        # Outer blockquote wraps everything
+        assert out.count("<blockquote>") >= 2
+
+    def test_deeply_nested_blockquote(self):
+        src = ">>> triple nested"
+        out = _apply_blockquotes(src)
+        assert out.count("<blockquote>") == 3
+
+    def test_task_list_normal_list_mixed(self):
+        src = "> - [x] done\n> - normal item\n> - [ ] todo"
+        out = _apply_blockquotes(src)
+        assert "<blockquote>" in out
+        assert "<ul>" in out
diff --git a/tests/test_renderer_js_behaviour.py b/tests/test_renderer_js_behaviour.py
new file mode 100644
index 00000000..dd95c098
--- /dev/null
+++ b/tests/test_renderer_js_behaviour.py
@@ -0,0 +1,191 @@
+"""Behavioural tests that drive the ACTUAL renderMd() in static/ui.js via node.
+
+The Python mirrors in test_blockquote_rendering.py and
+test_renderer_comprehensive.py validate intent, but they can drift from the
+JS.  Twice now (PR #1073 commit 94d63d0 — phantom <br>; PR #1073 commit
+04e7b53 — leading-space-in-blockquote prefix-strip regex) the Python mirror
+was correct while the JS was not, so the static-mirror tests passed even
+though the live UI was broken.
+
+This file closes that gap by spawning ``node`` on the real ui.js and
+asserting the rendered HTML for the most common LLM-output shapes.
+Add a case here whenever the renderer fix targets a class of input the
+Python mirror cannot exercise faithfully.
+"""
+import os
+import shutil
+import subprocess
+import sys
+from pathlib import Path
+
+import pytest
+
+REPO_ROOT = Path(__file__).parent.parent.resolve()
+UI_JS_PATH = REPO_ROOT / "static" / "ui.js"
+
+NODE = shutil.which("node")
+
+pytestmark = pytest.mark.skipif(NODE is None, reason="node not on PATH")
+
+
+_DRIVER_SRC = r"""
+const fs = require('fs');
+const src = fs.readFileSync(process.argv[2], 'utf8');
+global.window = {};
+global.document = { createElement: () => ({ innerHTML: '', textContent: '' }) };
+const esc = s => String(s ?? '').replace(/[&<>"']/g, c => (
+  {'&':'&amp;','<':'&lt;','>':'&gt;','"':'&quot;',"'":'&#39;'}[c]));
+
+function extractFunc(name) {
+  const re = new RegExp('function\\s+' + name + '\\s*\\(');
+  const start = src.search(re);
+  if (start < 0) throw new Error(name + ' not found');
+  let i = src.indexOf('{', start);
+  let depth = 1; i++;
+  while (depth > 0 && i < src.length) {
+    if (src[i] === '{') depth++;
+    else if (src[i] === '}') depth--;
+    i++;
+  }
+  return src.slice(start, i);
+}
+eval(extractFunc('renderMd'));
+
+let buf = '';
+process.stdin.on('data', c => { buf += c; });
+process.stdin.on('end', () => { process.stdout.write(renderMd(buf)); });
+"""
+
+
+@pytest.fixture(scope="module")
+def driver_path(tmp_path_factory):
+    """Write the node driver to a tmp file (works around `node -e` arg quirks)."""
+    p = tmp_path_factory.mktemp("renderer_driver") / "driver.js"
+    p.write_text(_DRIVER_SRC, encoding="utf-8")
+    return str(p)
+
+
+def _render(driver_path, markdown: str) -> str:
+    """Run renderMd against the actual ui.js and return the rendered HTML."""
+    result = subprocess.run(
+        [NODE, driver_path, str(UI_JS_PATH)],
+        input=markdown,
+        capture_output=True,
+        text=True,
+        timeout=10,
+    )
+    if result.returncode != 0:
+        raise RuntimeError(f"node driver failed: {result.stderr}")
+    return result.stdout
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Blockquote prefix strip — the bug commit 04e7b53 introduced was a one-char
+# regex regression where `^>[\t]?` (only tab) replaced `^>[ \t]?` (space or
+# tab), producing leading-space artifacts and breaking lists-in-quotes
+# because the list-detection regex `^(  )?[-*+]` couldn't match the
+# space-prefixed lines.  These tests exercise the actual JS so the regex
+# can't silently regress to tab-only again.
+# ─────────────────────────────────────────────────────────────────────────────
+
+
+class TestBlockquotePrefixStrip:
+    """Drive the actual renderMd to confirm `> ` is fully stripped."""
+
+    def test_single_line_blockquote_no_leading_space(self, driver_path):
+        out = _render(driver_path, "> Hello world").strip()
+        assert "<blockquote>Hello world</blockquote>" in out, (
+            f"`> Hello world` must render as <blockquote>Hello world</blockquote> "
+            f"with no leading space.  Got: {out!r}.  Likely cause: prefix-strip "
+            f"regex consumes only \\t, not space."
+        )
+
+    def test_multiline_blockquote_no_leading_space(self, driver_path):
+        out = _render(driver_path, "> Line one\n> Line two").strip()
+        assert ">Line one\nLine two<" in out, (
+            f"Multi-line blockquote must strip the space after each `>`.  "
+            f"Got: {out!r}"
+        )
+        # Belt-and-braces: there must be no space-after-newline-in-content
+        assert "\n " not in out.replace("</blockquote>", ""), (
+            f"Inner content of blockquote should not contain leading-space "
+            f"lines.  Got: {out!r}"
+        )
+
+    def test_list_inside_blockquote_renders_as_ul(self, driver_path):
+        """The PR explicitly added 'lists inside blockquotes' as a feature.
+        With the prefix-strip bug, the list-detection regex can't match the
+        space-prefixed lines, so the list never renders.  This pins it."""
+        out = _render(driver_path, "> Steps:\n> - one\n> - two")
+        assert "<ul>" in out, (
+            f"`> - item` lines inside a blockquote must render as a <ul>.  "
+            f"Got: {out!r}.  Likely cause: prefix-strip leaves a leading "
+            f"space, list regex `^(  )?[-*+] ` can't match one-space prefix."
+        )
+        assert "<li>one</li>" in out
+        assert "<li>two</li>" in out
+
+    def test_task_list_inside_blockquote(self, driver_path):
+        """Task lists inside blockquotes render checkbox spans, not literal [x]."""
+        out = _render(driver_path, "> - [x] done\n> - [ ] todo")
+        assert 'class="task-done"' in out, (
+            f"`- [x]` inside a blockquote must produce a task-done span.  "
+            f"Got: {out!r}"
+        )
+        assert 'class="task-todo"' in out
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Common LLM output shapes — sanity-check the most frequent constructs render
+# the way a user would expect.
+# ─────────────────────────────────────────────────────────────────────────────
+
+
+class TestCommonLLMShapes:
+
+    def test_strikethrough_outside_quote(self, driver_path):
+        out = _render(driver_path, "This was ~~outdated~~ but is now fine.")
+        assert "<del>outdated</del>" in out
+
+    def test_strikethrough_inside_blockquote(self, driver_path):
+        out = _render(driver_path, "> This is ~~wrong~~ actually")
+        assert "<blockquote>" in out and "<del>wrong</del>" in out
+
+    def test_top_level_task_list(self, driver_path):
+        out = _render(driver_path, "- [x] done\n- [ ] todo\n- regular item")
+        assert 'class="task-done"' in out
+        assert 'class="task-todo"' in out
+        assert "regular item" in out
+
+    def test_nested_blockquote_recurses(self, driver_path):
+        out = _render(driver_path, ">>> deeply nested")
+        assert out.count("<blockquote>") == 3
+        assert out.count("</blockquote>") == 3
+
+    def test_quote_then_heading(self, driver_path):
+        out = _render(driver_path, "> Note this.\n\n## Heading")
+        assert "<blockquote>Note this.</blockquote>" in out
+        assert "<h2>Heading</h2>" in out
+
+    def test_crlf_does_not_leak_carriage_return(self, driver_path):
+        out = _render(driver_path, "Line1\r\nLine2\r\nLine3")
+        assert "\r" not in out, f"CRLF must be normalised; got {out!r}"
+
+    def test_llm_multiparagraph_quote_with_list(self, driver_path):
+        """The shape an LLM emits when summarising decisions inside a quote."""
+        src = (
+            "> Here are the key points:\n"
+            ">\n"
+            "> - Point one\n"
+            "> - Point two\n"
+            ">\n"
+            "> And a closing remark."
+        )
+        out = _render(driver_path, src)
+        assert "<blockquote>" in out
+        assert "<ul>" in out
+        assert "<li>Point one</li>" in out
+        assert "<li>Point two</li>" in out
+        assert "And a closing remark." in out
+        # No leading-space artifacts in the quoted text
+        assert "\n " not in out.replace("</blockquote>", "")