Preserve preview harness, editor, email and task improvements

Snapshot current maintainer-preview application changes and regression fixtures for integration into lab. Excludes local runtime data, evaluation outputs and source backups. Focused Python regression selection: 140 passed; full suite not certified.

src/agent_tools/__init__.py

@@ -20,6 +20,7 @@ logger = logging.getLogger(__name__)
 
 from .subprocess_tools import BashTool, HostShellTool, PythonTool
 from .web_tools import WebSearchTool, WebFetchTool, PdfExtractTool, PrivateBrowserTool, YouTubeTool
+from .weather_tools import WeatherTool
 from .media_tools import ExtractTextTool, InspectMediaTool, TranscribeMediaTool
 from .filesystem_tools import ReadFileTool, WriteFileTool, EditFileTool, ApplyPatchTool, LsTool, GlobTool, GrepTool, GetWorkspaceTool
 from .coding_tools import TodoWriteTool
@@ -39,6 +40,7 @@ TOOL_HANDLERS = {
     "host_shell": HostShellTool().execute,
     "python": PythonTool().execute,
     "web_search": WebSearchTool().execute,
+    "get_weather": WeatherTool().execute,
     "web_fetch": WebFetchTool().execute,
     "pdf_extract": PdfExtractTool().execute,
     "youtube_tool": YouTubeTool().execute,

src/agent_tools/document_tools.py

@@ -1,6 +1,7 @@
 from typing import Any, Dict, List, Optional
 import hashlib
 import html
+import difflib
 import logging
 import re
 from src.constants import MAX_READ_CHARS
@@ -295,10 +296,13 @@ def parse_edit_blocks(content: str) -> list:
     # preserve whitespace inside the actual find/replace text.
     pattern = (
         r'<<<FIND>>>[ \t]*(?:\r?\n)?(.*?)[ \t]*(?:\r?\n)?'
-        r'<<<REPLACE>>>[ \t]*(?:\r?\n)?(.*?)[ \t]*(?:\r?\n)?<<<END>>>'
+        r'<<<(REPLACE|REPLACE_ALL)>>>[ \t]*(?:\r?\n)?(.*?)[ \t]*(?:\r?\n)?<<<END>>>'
     )
     for m in re.finditer(pattern, content, re.DOTALL):
-        edits.append({"find": m.group(1), "replace": m.group(2)})
+        edit = {"find": m.group(1), "replace": m.group(3)}
+        if m.group(2) == 'REPLACE_ALL':
+            edit['replace_all'] = True
+        edits.append(edit)
     if not edits and "<<<FIND>>>" in content and "<<<REPLACE>>>" in content:
         # Some native callers stop generation immediately after the replace
         # body. Treat end-of-content as the terminal marker only in that
@@ -711,6 +715,74 @@ class UpdateDocumentTool:
         finally:
             db.close()
 
+def _document_find_contexts(content, find):
+    """Give a failed caller exact contextual anchors instead of a blind retry."""
+    contexts = []
+    for index, match in enumerate(re.finditer(re.escape(find), content)):
+        if index >= 3:
+            break
+        start = content.rfind('\n', 0, match.start()) + 1
+        end = content.find('\n', match.end())
+        end = len(content) if end < 0 else end
+        # Rich-text paragraphs are often stored on one HTML line.
+        for tag in ('p', 'div'):
+            paragraph = content.rfind(f'<{tag}', 0, match.start())
+            close = content.find(f'</{tag}>', match.end())
+            if paragraph >= 0 and close >= 0 and close + len(tag) + 3 - paragraph <= 1200:
+                start, end = paragraph, close + len(tag) + 3
+                break
+        context = content[start:end]
+        if len(context) <= 1200 and content.count(context) == 1:
+            contexts.append(context)
+    return '\nExact unique anchors from the current document:\n' + '\n'.join(contexts) if contexts else ''
+
+
+def _document_find_repair_hint(content, find, count):
+    """Offer bounded exact source text for an unmatched or ambiguous edit."""
+    if isinstance(count, int) and count > 1:
+        return _document_find_contexts(content, find)[:900]
+    if not isinstance(count, int) or count != 0:
+        return ''
+    # Minified markup may be one enormous line. Parse tag boundaries instead
+    # of comparing a small FIND against that entire line. This is evidence for
+    # a corrected call, never permission to apply a fuzzy replacement.
+    if find.lstrip().startswith('<'):
+        from html.parser import HTMLParser
+
+        class SourceTags(HTMLParser):
+            def __init__(self):
+                super().__init__(convert_charrefs=False)
+                self.tags = []
+
+            def handle_starttag(self, tag, attrs):
+                raw = self.get_starttag_text()
+                if raw and len(raw) <= 1000:
+                    self.tags.append(raw)
+
+            def handle_startendtag(self, tag, attrs):
+                self.handle_starttag(tag, attrs)
+
+        parser = SourceTags()
+        parser.feed(content)
+        matches = difflib.get_close_matches(find, list(dict.fromkeys(parser.tags)), n=2, cutoff=0.7)
+        if matches:
+            return 'Copy an exact source fragment into FIND (including its spacing and quotes): ' + ' | '.join(
+                repr(match) for match in matches)
+    if re.fullmatch(r"[\w'-]{3,40}", find):
+        words = re.findall(r"[\w'-]{3,40}", content)
+        by_lower = {word.casefold(): word for word in words}
+        matches = difflib.get_close_matches(find.casefold(), by_lower, n=5, cutoff=0.6)
+        if matches:
+            return 'Closest words actually in the document: ' + ', '.join(
+                repr(by_lower[word]) for word in matches)
+    paragraphs = re.findall(r'<(?:p|div)\b[^>]*>.*?</(?:p|div)>', content, re.S | re.I)
+    if not paragraphs:
+        paragraphs = content.splitlines()
+    candidates = [p for p in paragraphs if len(p) <= 500]
+    matches = difflib.get_close_matches(find, candidates, n=2, cutoff=0.4)
+    return 'Closest exact passages in the document: ' + ' | '.join(matches) if matches else ''
+
+
 class EditDocumentTool:
     async def execute(self, content: str, ctx: dict) -> Dict:
         """Apply targeted FIND/REPLACE edits to an existing document."""
@@ -795,34 +867,66 @@ class EditDocumentTool:
                 return {"error": "No edits applied — FIND text cannot be blank"}
 
             updated_content = doc.current_content
-            applied = 0
-            skipped = 0
-            for edit in edits:
-                _find = edit["find"]
-                if _find == edit["replace"]:
-                    logger.warning("edit_document: skipping no-op FIND/REPLACE block")
+            applied, skipped, no_op_edits = 0, 0, 0
+            invalid_edits = []
+            # Validate against evolving content before the database write.
+            # Only exact unique matches may be saved; report every rejected
+            # entry explicitly so a partial batch cannot masquerade as complete.
+            prose = str(doc.language or '').lower() in {'text', 'markdown', 'richtext', 'email', ''}
+            for edit_number, edit in enumerate(edits, 1):
+                find = edit['find']
+                replacement = edit['replace']
+                if find == replacement:
                     skipped += 1
+                    no_op_edits += 1
                     continue
-                if _find in updated_content:
-                    updated_content = updated_content.replace(_find, edit["replace"], 1)
-                    applied += 1
-                else:
-                    # Defensive: the active-doc context shows a "N\t" line-number
-                    # gutter for reference. Weaker models sometimes copy that prefix
-                    # into FIND. If the exact match failed, retry with a leading
-                    # "<digits><tab>" stripped from each FIND line — but only use it
-                    # when that stripped form actually matches, so we never corrupt a
-                    # legitimately tab-prefixed document.
-                    _stripped = "\n".join(re.sub(r"^\d+\t", "", _l) for _l in _find.split("\n"))
-                    if _stripped != _find and _stripped in updated_content:
-                        updated_content = updated_content.replace(_stripped, edit["replace"], 1)
-                        applied += 1
-                        logger.info("edit_document: matched after stripping line-number gutter from FIND")
-                    else:
-                        logger.warning(f"edit_document: FIND text not found, skipping: {_find[:80]!r}")
-                        skipped += 1
+                if find not in updated_content:
+                    stripped = "\n".join(re.sub(r"^\d+\t", "", line) for line in find.split("\n"))
+                    if stripped != find and stripped in updated_content:
+                        find = stripped
+                count = updated_content.count(find) if find else 0
+                replace_all = edit.get('replace_all') is True
+                if count == 0 or (count != 1 and not replace_all):
+                    invalid_edits.append((edit_number, count, edit['find']))
+                    continue
+                position = updated_content.index(find)
+                positions = [m.start() for m in re.finditer(re.escape(find), updated_content)] if replace_all else [position]
+                if prose and re.fullmatch(r"[\w]+", find):
+                    for match_pos in positions:
+                        before = updated_content[match_pos - 1:match_pos] if match_pos else ''
+                        after = updated_content[match_pos + len(find):match_pos + len(find) + 1]
+                        if (before and (before.isalnum() or before == '_')) or (after and (after.isalnum() or after == '_')):
+                            invalid_edits.append((edit_number, 'part of a word', edit['find']))
+                            break
+                    if invalid_edits and invalid_edits[-1][0] == edit_number:
+                        continue
+                updated_content = updated_content.replace(find, replacement) if replace_all else updated_content[:position] + replacement + updated_content[position + len(find):]
+                applied += 1
+
+            partial_edits = bool(invalid_edits and applied)
+            if invalid_edits and not partial_edits:
+                details = '; '.join(
+                    f'#{number} ({reason} matches): {find[:100]!r}'
+                    if isinstance(reason, int) else f'#{number} ({reason}): {find[:100]!r}'
+                    for number, reason, find in invalid_edits[:8]
+                )
+                extra = f'; and {len(invalid_edits) - 8} more' if len(invalid_edits) > 8 else ''
+                return {
+                    'error': f'No edits applied. Invalid FIND entries: {details}{extra}. '
+                             'Do not repeat the unchanged call. Copy FIND exactly from the current source or the hints below, '
+                             'then retry the corrected entries. If no hint identifies the target, read the document first. '
+                             'Other entries were not saved. ' + ' '.join(
+                                 f'#{number}: {_document_find_repair_hint(doc.current_content, find, reason)}'
+                                 for number, reason, find in invalid_edits[:3]
+                                 if _document_find_repair_hint(doc.current_content, find, reason)
+                             ),
+                    'exit_code': 1, 'applied': 0,
+                    'invalid_edit_numbers': [number for number, _, _ in invalid_edits],
+                }
 
             if applied == 0:
+                if no_op_edits == len(edits):
+                    return {"error": "No edits applied: every FIND and REPLACE pair is identical. Write a changed replacement that fulfills the requested revision; keep FIND copied from the current document."}
                 return {"error": f"No edits applied — none of the FIND blocks matched the document content (skipped {skipped})"}
 
             missing_id = _missing_document_upload(owner, updated_content)
@@ -855,7 +959,7 @@ class EditDocumentTool:
             db.add(ver)
             db.commit()
 
-            return {
+            result = {
                 "action": "edit",
                 "doc_id": target_id,
                 "title": doc.title,
@@ -865,6 +969,17 @@ class EditDocumentTool:
                 "applied": applied,
                 "skipped": skipped,
             }
+            if partial_edits:
+                result.update({
+                    'partial': True,
+                    'rejected': len(invalid_edits),
+                    'invalid_edits': [
+                        {'number': number, 'matches': reason, 'find': find[:100],
+                         'hint': _document_find_repair_hint(updated_content, find, reason)}
+                        for number, reason, find in invalid_edits
+                    ],
+                })
+            return result
         except Exception as e:
             db.rollback()
             return {"error": f"Failed to edit document: {e}"}
@@ -897,8 +1012,8 @@ class SuggestDocumentTool:
                 return version_error
 
             # Validate that FIND text exists in document
-            valid = []
-            for s in suggestions:
+            valid, invalid = [], []
+            for number, s in enumerate(suggestions, 1):
                 find_text = s["find"]
                 # Browser selections from markdown, rich text, and email are
                 # rendered text, while the stored document may contain LF
@@ -906,22 +1021,44 @@ class SuggestDocumentTool:
                 # back to the exact source fragment used by the editor.
                 source_find = _visible_text_match_source(doc.current_content, find_text)
                 if source_find is not None:
+                    stored = doc.current_content or ''
+                    if stored.count(source_find) != 1:
+                        invalid.append({'number': number, 'find': find_text[:100],
+                                        'reason': 'ambiguous',
+                                        'hint': _document_find_contexts(stored, source_find)[:900]})
+                        continue
+                    if re.fullmatch(r"[\w]+", source_find):
+                        pos = stored.index(source_find)
+                        before = stored[pos - 1:pos] if pos else ''
+                        after = stored[pos + len(source_find):pos + len(source_find) + 1]
+                        if (before and before.isalnum()) or (after and after.isalnum()):
+                            invalid.append({'number': number, 'find': find_text[:100],
+                                            'reason': 'part of a word', 'hint': ''})
+                            continue
                     if source_find != find_text:
                         s = dict(s)
                         s["find"] = source_find
                     s["id"] = _stable_suggestion_id(target_id, s)
                     valid.append(s)
                 else:
-                    logger.warning(f"suggest_document: FIND text not found, skipping: {find_text[:80]!r}")
+                    invalid.append({'number': number, 'find': find_text[:100],
+                                    'reason': 'not found',
+                                    'hint': _document_find_repair_hint(doc.current_content or '', find_text, 0)[:900]})
 
             if not valid:
-                return {"error": "No suggestions matched the document content"}
+                details = '; '.join(f"#{item['number']} {item['reason']}: {item['find']!r} {item['hint']}"
+                                    for item in invalid[:5])
+                return {'error': 'No suggestions created: ' + details,
+                        'exit_code': 1, 'rejected': len(invalid)}
 
             return {
                 "action": "suggest",
                 "doc_id": target_id,
                 "suggestions": valid,
                 "count": len(valid),
+                "partial": bool(invalid),
+                "rejected": len(invalid),
+                "invalid_suggestions": invalid,
             }
         finally:
             db.close()

src/agent_tools/session_tools.py

@@ -156,7 +156,7 @@ async def list_sessions(content: str, session_id: Optional[str] = None, owner: O
             safe_name = (sess.name or "Untitled").replace("[", "\\[").replace("]", "\\]")
             msg_count = getattr(sess, "message_count", 0) or 0
             model = getattr(sess, "model", "unknown")
-            marker = " ← most recent" if i == 0 else ""
+            marker = " ← current chat" if sid == session_id else (" ← most recent" if i == 0 else "")
             lines.append(f"- **[{safe_name}](#session-{sid})** (id: `{sid}`, model: {model}, {msg_count} msgs, last active {_rel(ts)}){marker}")
 
         if not lines:
@@ -166,6 +166,7 @@ async def list_sessions(content: str, session_id: Optional[str] = None, owner: O
             "results": (
                 f"Found {len(rows)} session(s), sorted most-recent first:\n"
                 + "\n".join(lines)
+                + "\nFor the previous/last chat, exclude the row marked current chat. Use the exact returned ID, not an alias. If the target is ambiguous, ask using chat titles before changing anything."
                 + "\n\nAssistant: when replying to the user, preserve the chat-title markdown links exactly as shown, e.g. `[Chat](#session-id)`. Do not rewrite this as a plain, non-clickable table."
             )
         }

src/agent_tools/weather_tools.py

@@ -0,0 +1,78 @@
+"""No-key weather lookup backed by Open-Meteo."""
+
+import asyncio
+import json
+import re
+import urllib.parse
+import urllib.request
+
+
+def _get_json(url: str) -> dict:
+    request = urllib.request.Request(url, headers={"User-Agent": "Odysseus/1.0"})
+    with urllib.request.urlopen(request, timeout=8) as response:
+        return json.load(response)
+
+
+def weather_location_from_query(query: str) -> str | None:
+    """Extract a place only from straightforward weather lookup phrasing."""
+    text = re.sub(r"\s+", " ", query).strip(" ?.! ")
+    patterns = (
+        r"^(?:what(?:'s| is) the )?(?:current |today(?:'s)? |tomorrow(?:'s)? )?"
+        r"(?:weather|forecast)(?: like)? (?:in|for|at) (?P<place>.+)$",
+        r"^(?:weather|forecast) (?P<place>.+)$",
+        r"^(?P<place>.+?) (?:weather|forecast)\b.*$",
+    )
+    for pattern in patterns:
+        match = re.match(pattern, text, re.IGNORECASE)
+        if match:
+            place = re.sub(r"\b(?:today|tomorrow|now|current)\b.*$", "", match.group("place"), flags=re.IGNORECASE).strip(" ,")
+            if 1 <= len(place) <= 100:
+                return place
+    return None
+
+
+class WeatherTool:
+    async def execute(self, content: str, ctx: dict) -> dict:
+        try:
+            args = json.loads(content) if content.strip().startswith("{") else {"location": content}
+            if not isinstance(args, dict):
+                return {"error": "get_weather expects a location string or JSON object", "exit_code": 1}
+            location = str(args.get("location") or "").strip()
+            if not location or len(location) > 160:
+                return {"error": "get_weather requires a location (up to 160 characters)", "exit_code": 1}
+
+            geo_url = "https://geocoding-api.open-meteo.com/v1/search?" + urllib.parse.urlencode({
+                "name": location, "count": 1, "language": "en", "format": "json",
+            })
+            geo = await asyncio.to_thread(_get_json, geo_url)
+            places = geo.get("results") or []
+            if not places:
+                return {"error": f"No location found for {location!r}", "exit_code": 1}
+            place = places[0]
+            forecast_url = "https://api.open-meteo.com/v1/forecast?" + urllib.parse.urlencode({
+                "latitude": place["latitude"],
+                "longitude": place["longitude"],
+                "current": "temperature_2m,relative_humidity_2m,precipitation,weather_code,wind_speed_10m",
+                "daily": "temperature_2m_max,temperature_2m_min,precipitation_probability_max,weather_code",
+                "forecast_days": 3,
+                "timezone": place.get("timezone") or "auto",
+            })
+            forecast = await asyncio.to_thread(_get_json, forecast_url)
+            current = forecast.get("current") or {}
+            daily = forecast.get("daily") or {}
+            if not current.get("time") or not daily.get("time"):
+                return {"error": "Weather provider returned incomplete forecast data", "exit_code": 1}
+            place_parts = list(dict.fromkeys(filter(None, [place.get("name"), place.get("admin1"), place.get("country")])))
+            data = {
+                "location": ", ".join(place_parts),
+                "timezone": forecast.get("timezone"),
+                "current": current,
+                "current_units": forecast.get("current_units") or {},
+                "daily": daily,
+                "daily_units": forecast.get("daily_units") or {},
+                "source": forecast_url,
+                "provider": "Open-Meteo",
+            }
+            return {"output": json.dumps(data, ensure_ascii=False), "exit_code": 0, "evidence_status": "available"}
+        except (OSError, ValueError, KeyError, TypeError) as exc:
+            return {"error": f"Weather lookup failed: {exc}", "exit_code": 1}

src/agent_tools/web_tools.py

@@ -322,6 +322,37 @@ class WebSearchTool:
                 "exit_code": 1,
                 "untrusted_content": True,
             }
+        from .weather_tools import WeatherTool, weather_location_from_query
+        weather_location = weather_location_from_query(query)
+        if not sources and time_filter and weather_location:
+            # A forecast or current fact need not live on a newly published page.
+            try:
+                text, sources = await asyncio.wait_for(
+                    loop.run_in_executor(
+                        None,
+                        lambda: comprehensive_web_search(
+                            query, max_pages=max_pages, time_filter=None,
+                            return_sources=True,
+                        ),
+                    ),
+                    timeout=20,
+                )
+            except Exception:
+                pass
+        if not sources:
+            from src.turn_contract import active_turn_contract
+            contract = active_turn_contract()
+            policy = ctx.get("tool_policy") if isinstance(ctx, dict) else None
+            weather_allowed = (
+                weather_location
+                and "get_weather" not in (ctx.get("disabled_tools") or ())
+                and not (policy and policy.blocks("get_weather"))
+                and not (contract and not contract.permits("get_weather"))
+            )
+            if weather_allowed:
+                weather = await WeatherTool().execute(json.dumps({"location": weather_location}), ctx)
+                if weather.get("exit_code") == 0:
+                    return weather
         if progress_cb:
             await progress_cb({
                 "elapsed_s": 30,
@@ -669,7 +700,7 @@ class WebFetchTool:
         except Exception as e:
             return {"error": f"web_fetch: {url}: {e}", "exit_code": 1}
         err = result.get("error")
-        text = (result.get("content") or "").strip()
+        text = (result.get("linked_content") or result.get("content") or "").strip()
         title = result.get("title") or ""
 
         if not text:
@@ -711,7 +742,7 @@ class WebFetchTool:
                 "\n\n[...truncated; re-call web_fetch with query terms to retrieve matching passages]"
                 if not query else "\n\n[...truncated]"
             )
-        return {"output": output, "exit_code": 0}
+        return {"output": output, "exit_code": 0, "page_entries": result.get("page_entries") or []}
 
 
 class PdfExtractTool:
@@ -1840,7 +1871,6 @@ class YouTubeTool:
 
     async def execute(self, content: str, ctx: dict) -> dict:
         from services.youtube.youtube_handler import (
-            extract_youtube_id,
             extract_transcript_async,
             fetch_youtube_comments,
             init_youtube,
@@ -1878,11 +1908,29 @@ class YouTubeTool:
                 return await self._latest_channel_video(channel, max_results=max_results)
 
         url_or_id = str(args.get("url") or args.get("video_url") or args.get("video_id") or "").strip()
-        video_id = str(args.get("video_id") or "").strip()
-        if not video_id and url_or_id:
-            video_id = extract_youtube_id(url_or_id) or (url_or_id if _looks_like_youtube_video_id(url_or_id) else "")
-        if not video_id:
-            return {"error": f"youtube_tool {action}: provide a YouTube video URL or video_id", "exit_code": 1}
+        # The shared extractor accepts ID prefixes inside text. At the tool
+        # boundary require the complete target, not a truncated invented ID.
+        if url_or_id.startswith(('http://', 'https://')):
+            parsed = urllib.parse.urlparse(url_or_id)
+            host = (parsed.hostname or '').lower()
+            if host in {'youtube.com', 'www.youtube.com', 'm.youtube.com', 'music.youtube.com'}:
+                parts = parsed.path.strip('/').split('/')
+                candidate = (urllib.parse.parse_qs(parsed.query).get('v', [''])[0]
+                             if parsed.path == '/watch' else
+                             parts[1] if len(parts) == 2 and parts[0] in {'shorts', 'embed', 'live'} else '')
+            elif host in {'youtu.be', 'www.youtu.be'}:
+                candidate = parsed.path.strip('/')
+            else:
+                candidate = ''
+        else:
+            candidate = url_or_id
+        if (not re.fullmatch(r'[A-Za-z0-9_-]{11}', candidate)
+                or (args.get('video_id') and str(args['video_id']) != candidate)):
+            return {"error": f"youtube_tool {action}: invalid video target. Resolve the actual video URL "
+                    "from the user or an observed link. A title or channel page is not a video ID; "
+                    "open the referenced video in the browser or use latest_channel_video first.",
+                    "exit_code": 1, "failure_kind": "invalid_target"}
+        video_id = candidate
         url = url_or_id if url_or_id.startswith(("http://", "https://")) else f"https://www.youtube.com/watch?v={video_id}"
 
         if action == "comments":
@@ -1896,7 +1944,9 @@ class YouTubeTool:
                 joined = f"{fallback_error}"
                 if api_error:
                     joined = f"YouTube Data API unavailable: {api_error}; yt-dlp fallback failed: {fallback_error}"
-                return {"error": f"youtube_tool comments: {joined}", "exit_code": 1, "untrusted_content": True}
+                return {"error": f"youtube_tool comments: {joined}", "exit_code": 1,
+                        "failure_kind": "comments_unavailable", "video_url": url,
+                        "untrusted_content": True}
             return {"output": self._format_comments(comments_data, url), "exit_code": 0, "untrusted_content": True}
 
         if action == "transcript":
@@ -2664,14 +2714,15 @@ class PrivateBrowserTool:
             if err_text:
                 combined = f"[stderr]\n{err_text}\n\n{combined}".strip()
         fill_error = ""
+        empty_observation = (proc.returncode or 0) == 0 and self._empty_dom_observation(out)
         observe_state_change = action in {"open", "fill", "press"} and model_choice
-        failed_interaction = action in {"click", "fill"} and model_choice and (proc.returncode or 0) != 0
-        if failed_interaction or ((action == "click" or observe_state_change) and (proc.returncode or 0) == 0):
+        failed_interaction = action in {"click", "fill"} and (proc.returncode or 0) != 0
+        if empty_observation or failed_interaction or ((action == "click" or observe_state_change) and (proc.returncode or 0) == 0):
             # A click can navigate, replace the DOM, or open a modal. Return
             # the settled post-click DOM in the same tool result so callers do
             # not race navigation with a separate immediate read and so the
             # next conversational turn receives current element refs. A failed
-            # model-choice interaction also needs refs for a covering dialog
+            # interaction also needs refs for a covering dialog
             # or changed DOM. A successful fill may run input handlers that
             # open a modal or replace the field: CLI success is not proof that
             # the intended value survived. Observe only; never retry an action.
@@ -2718,7 +2769,8 @@ class PrivateBrowserTool:
             if page_errors:
                 combined = f"{combined}\n\n[page errors]\n{page_errors}".strip()
         if len(combined) > MAX_OUTPUT_CHARS:
-            combined = combined[:MAX_OUTPUT_CHARS] + "\n\n[...truncated]"
+            from src.browser_observation import compact_browser_observation
+            combined = compact_browser_observation(combined, budget=MAX_OUTPUT_CHARS)
         shopping_hint = self._shopping_landing_hint(combined)
         if shopping_hint:
             combined = f"{combined}\n\n[{shopping_hint}]"
@@ -2780,7 +2832,7 @@ class PrivateBrowserTool:
         deadline = loop.time() + min(timeout_s, 20)
         text, observation_note = "", ""
         first_rows, rows = [], []
-        for attempt in range(2 if model_choice else 1):
+        for attempt in range(2):
             proc = None
             try:
                 async with asyncio.timeout(max(0, deadline - loop.time())):
@@ -2821,9 +2873,14 @@ class PrivateBrowserTool:
             text, rows = observed, observed_rows
             if not attempt:
                 first_rows = rows
-            if not snapshots or any(snapshot.strip() != '(empty page)' for snapshot in snapshots):
+            if not snapshots or not self._empty_dom_observation(observed):
                 break
             commands = [["wait", "1000"], ["snapshot"]]
+        if not observation_note and self._empty_dom_observation(text):
+            observation_note = (
+                "Browser observation incomplete: the page still has no readable content after waiting. "
+                "Navigation success is not evidence that results loaded. Do not infer page results."
+            )
         fill_error = ""
         if verify_fill:
             fill_error = unverified
@@ -2845,7 +2902,30 @@ class PrivateBrowserTool:
             text = self._snapshot_observation(text)
         if observation_note:
             text += '\n' + observation_note
-        return text[:MAX_OUTPUT_CHARS], fill_error
+        if len(text) > MAX_OUTPUT_CHARS:
+            from src.browser_observation import compact_browser_observation
+            text = compact_browser_observation(text, budget=MAX_OUTPUT_CHARS)
+        return text, fill_error
+
+    @staticmethod
+    def _empty_dom_observation(text: str) -> bool:
+        """Recognize empty accessibility scaffolding, not an actual no-results message."""
+        try:
+            payload = json.loads(text)
+        except (ValueError, TypeError):
+            payload = None
+        if isinstance(payload, list):
+            snapshots = [row['result']['snapshot'] for row in payload
+                if isinstance(row, dict) and row.get('success') is True
+                and isinstance(row.get('result'), dict)
+                and isinstance(row['result'].get('snapshot'), str)]
+        else:
+            snapshots = [text] if isinstance(text, str) and text.strip() else []
+        scaffolding = {'- generic', '- main', '- none', '- presentation', '(empty page)'}
+        return bool(snapshots) and all(
+            all(line.strip() in scaffolding for line in snapshot.splitlines() if line.strip())
+            for snapshot in snapshots
+        )
 
     @staticmethod
     def _dialog_first_snapshot(snapshot: str) -> str: