From 7257319004f26bbef3a9ef08ec300ca41d1912e1 Mon Sep 17 00:00:00 2001
From: pewdiepie-archdaemon <pewdiepie-archdaemon@users.noreply.github.com>
Date: Thu, 17 Sep 2026 20:17:03 +0000
Subject: [PATCH] Preserve every fetched search source within observation
 budget

---
 src/clean_agent_preview.py              | 35 +++++++++++++++++++++++
 tests/test_search_observation_budget.py | 37 +++++++++++++++++++++++++
 2 files changed, 72 insertions(+)
 create mode 100644 tests/test_search_observation_budget.py

diff --git a/src/clean_agent_preview.py b/src/clean_agent_preview.py
index d44f0f385..3bbe92a6b 100644
--- a/src/clean_agent_preview.py
+++ b/src/clean_agent_preview.py
@@ -162,6 +162,39 @@ SAFE_ACTIONS = {
 }
 
 
+def bounded_search_observation(output, budget=8000):
+    """Share the observation budget across fetched sources, not prefix order.
+
+    Search's full report repeats bodies in summaries/quotes/statistics. Retain
+    source attribution and an excerpt of every fetched page before truncating.
+    This is evidence selection, never answer generation.
+    """
+    if len(output) <= budget:
+        return output
+    pattern = re.compile(r'\n(\[CONTENT(?: \d+)?\] From: [^\n]+\nTitle: [^\n]*\n-+\n)')
+    matches = list(pattern.finditer(output))
+    if not matches:
+        return output
+    source_match = re.search(r'```sources\n.*?```', output, re.DOTALL)
+    query_match = re.search(r'^Query: .*$', output, re.MULTILINE)
+    prefix = '\n'.join(match.group(0) for match in (source_match, query_match) if match)
+    suffix = '\n[Excerpts shortened across sources; use web_fetch on a source URL for full details.]'
+    headers = [match.group(1) for match in matches]
+    room = budget - len(prefix) - len(suffix) - sum(len(h) + 2 for h in headers)
+    if room < 100 * len(matches):
+        return output  # Unusually large metadata: preserve existing hard cap.
+    per_page = room // len(matches)
+    blocks = []
+    for index, match in enumerate(matches):
+        end = matches[index + 1].start() if index + 1 < len(matches) else len(output)
+        body = output[match.end():end]
+        body = re.split(r'\n(?:Key Points:|TL;DR:|Important Quotes:|Data / Statistics:|={20,}|<!-- SOURCES:)', body, maxsplit=1)[0].strip()
+        if len(body) > per_page:
+            body = body[:per_page - 15].rstrip() + '\n[...excerpt]'
+        blocks.append(headers[index] + body)
+    return prefix + '\n\n' + '\n\n'.join(blocks) + suffix
+
+
 def preview_tool_result_text(result, tool, args):
     """Preserve failure evidence before applying the observation budget."""
     output = result.get('output') or result.get('error') or result
@@ -185,6 +218,8 @@ def preview_tool_result_text(result, tool, args):
         # leaving neither the model nor canonical renderer usable evidence.
         output = result['results']
     output = output if isinstance(output, str) else json.dumps(output, ensure_ascii=False)
+    if canonical(tool) == 'web_search' and not result.get('error') and result.get('exit_code') in (None, 0):
+        output = bounded_search_observation(output)
     if len(output) > 8000:
         output = output[:8000] + '\n[Tool result truncated at 8000 characters.]'
     return output
diff --git a/tests/test_search_observation_budget.py b/tests/test_search_observation_budget.py
new file mode 100644
index 000000000..451693dcd
--- /dev/null
+++ b/tests/test_search_observation_budget.py
@@ -0,0 +1,37 @@
+from src.clean_agent_preview import preview_tool_result_text
+
+
+def test_all_fetched_sources_survive_observation_budget():
+    sources = '```sources\n' + '\n'.join(
+        f'[{i}] Page {i}\nhttps://example.org/{i}' for i in range(1, 6)
+    ) + '\n```\nQuery: example research\n'
+    report = sources
+    for i in range(1, 6):
+        report += (f'\n[CONTENT {i}] From: https://example.org/{i}\n'
+                   f'Title: Page {i}\n------------------------------\n'
+                   + f'Evidence from page {i}. ' * 200
+                   + '\nTL;DR:\n' + 'Repeated summary. ' * 200)
+    output = preview_tool_result_text({'output': report, 'exit_code': 0}, 'web_search', {})
+    assert len(output) <= 8000
+    for i in range(1, 6):
+        assert f'[CONTENT {i}] From: https://example.org/{i}' in output
+        assert f'Evidence from page {i}.' in output
+    assert 'Repeated summary.' not in output
+    assert 'full details' in output
+
+
+def test_short_search_results_are_unchanged():
+    text = 'No matching sources were found.'
+    assert preview_tool_result_text({'output': text}, 'web_search', {}) == text
+
+
+def test_failure_status_is_not_lost_to_search_compaction():
+    result = {'output': 'Partial evidence. ' * 1000, 'error': 'fetch failed', 'exit_code': 1}
+    output = preview_tool_result_text(result, 'web_search', {})
+    assert 'fetch failed' in output[:100]
+
+
+def test_other_tool_observations_keep_existing_budget():
+    output = preview_tool_result_text({'output': 'x' * 9000}, 'bash', {})
+    assert output.startswith('x' * 8000)
+    assert 'truncated at 8000' in output