mirror of
https://github.com/openai/codex.git
synced 2026-11-16 06:54:36 +00:00
Share Guardian user-message retention logic (#42031)
## What changed - Add a shared `select_user_messages` helper in `guardian-context` that always anchors the first user message and fills the remaining token budget from newest to oldest. - Use the helper for both core Guardian prompt rendering and the Guardian V2 async scorer so they apply the same retention policy. ## Testing - Cover empty and single-message transcripts, over-budget anchors, and newest-first selection across multiple token budgets. GitOrigin-RevId: 7db246013875263ceb5806d2a6a19b4e2a3db1c4
This commit is contained in:
1 parent
2b7c279735
commit
8209978616
5 files changed
+129
-60
No files matched your search
@@ -398,45 +398,27 @@ fn render_guardian_transcript_entries_with_offset(
|
||||
.collect::<Vec<_>>();
|
||||
|
||||
let mut included = vec![false; entries.len()];
|
||||
let mut message_tokens = 0usize;
|
||||
let mut tool_tokens = 0usize;
|
||||
let user_indices = entries
|
||||
let user_messages = entries
|
||||
.iter()
|
||||
.enumerate()
|
||||
.filter_map(|(index, entry)| {
|
||||
matches!(entry.kind, ConversationTranscriptEntryKind::User).then_some(index)
|
||||
matches!(entry.kind, ConversationTranscriptEntryKind::User).then_some(
|
||||
codex_guardian_context::UserMessageCost {
|
||||
index,
|
||||
tokens: rendered_entries[index].1,
|
||||
},
|
||||
)
|
||||
})
|
||||
.collect::<Vec<_>>();
|
||||
|
||||
if let Some(&first_user_index) = user_indices.first() {
|
||||
included[first_user_index] = true;
|
||||
message_tokens += rendered_entries[first_user_index].1;
|
||||
}
|
||||
|
||||
if let Some(&last_user_index) = user_indices.last()
|
||||
&& !included[last_user_index]
|
||||
&& message_tokens + rendered_entries[last_user_index].1
|
||||
<= GUARDIAN_TRANSCRIPT_RETENTION.max_message_transcript_tokens
|
||||
{
|
||||
included[last_user_index] = true;
|
||||
message_tokens += rendered_entries[last_user_index].1;
|
||||
}
|
||||
|
||||
for &index in user_indices.iter().rev() {
|
||||
if included[index] {
|
||||
continue;
|
||||
}
|
||||
|
||||
let token_count = rendered_entries[index].1;
|
||||
if message_tokens + token_count
|
||||
> GUARDIAN_TRANSCRIPT_RETENTION.max_message_transcript_tokens
|
||||
{
|
||||
continue;
|
||||
}
|
||||
|
||||
let selection = codex_guardian_context::select_user_messages(
|
||||
&user_messages,
|
||||
GUARDIAN_TRANSCRIPT_RETENTION.max_message_transcript_tokens,
|
||||
);
|
||||
for index in selection.indices {
|
||||
included[index] = true;
|
||||
message_tokens += token_count;
|
||||
}
|
||||
let mut message_tokens = selection.tokens;
|
||||
let mut tool_tokens = 0usize;
|
||||
|
||||
let mut retained_non_user_entries = 0usize;
|
||||
for index in (0..entries.len()).rev() {
|
||||
|
||||
Reference in new issue
Block a user