mirror of
https://github.com/openai/codex.git
synced 2026-11-15 21:50:46 +00:00
Resample Guardian denials with the original request (#59659)
## What changed When `features.guardian_denial_reconsideration` is enabled, retry the first terminal denial once using the unchanged inference request. Buffer provisional output so only the selected response enters reviewer history, while allowing tool calls to proceed normally. Preserve the original denial if the retry fails or times out, and continue to honor cancellation. Account for token usage from both attempts, including failed reviews. Remove `auto_review.experimental_denial_reconsideration_prompt` and warn when it remains in configuration. ## Testing Update the Guardian scenario and snapshot to verify identical retry inputs, retained tool evidence, and reuse of the selected assessment in a subsequent review, with asynchronous compaction enabled. GitOrigin-RevId: cc4e303d40a02acc824c8b4b954852325e64ad52
This commit is contained in:
@@ -2960,7 +2960,7 @@ async fn try_run_sampling_request(
|
||||
let mut response_id_for_assessment = None;
|
||||
let mut pending_assessment: Option<codex_api::GuardianAssessment> = None;
|
||||
let mut response_completed = false;
|
||||
let mut request_token_usage = None;
|
||||
let mut request_token_usage: Option<codex_protocol::protocol::TokenUsage> = None;
|
||||
let mut includes_reasoning = step_context.settings.model_info.use_responses_lite;
|
||||
// Catalog policies can enable adaptive review independently of the legacy scorer flag.
|
||||
// Capture evidence whenever Guardian is available; the reviewer resolves the policy.
|
||||
@@ -2973,21 +2973,25 @@ async fn try_run_sampling_request(
|
||||
} else {
|
||||
None
|
||||
};
|
||||
let stream = client_session
|
||||
.stream(
|
||||
&prompt,
|
||||
&step_context.settings.model_info,
|
||||
&step_context.session_telemetry,
|
||||
effort,
|
||||
step_context.settings.reasoning_summary,
|
||||
step_context.settings.service_tier.clone(),
|
||||
responses_metadata,
|
||||
&inference_trace,
|
||||
)
|
||||
.instrument(trace_span!("stream_request"))
|
||||
let mut resampling = crate::guardian::ResponseResampling::new(&sess);
|
||||
let mut stream_request = async |prompt: &Prompt| {
|
||||
client_session
|
||||
.stream(
|
||||
prompt,
|
||||
&step_context.settings.model_info,
|
||||
&step_context.session_telemetry,
|
||||
effort.clone(),
|
||||
step_context.settings.reasoning_summary,
|
||||
step_context.settings.service_tier.clone(),
|
||||
responses_metadata,
|
||||
&inference_trace,
|
||||
)
|
||||
.instrument(trace_span!("stream_request"))
|
||||
.await
|
||||
};
|
||||
let stream = stream_request(&prompt)
|
||||
.or_cancel(&cancellation_token)
|
||||
.await?;
|
||||
drop(prompt);
|
||||
let mut stream = match stream {
|
||||
Ok(stream) => stream,
|
||||
Err(error) => {
|
||||
@@ -3041,6 +3045,16 @@ async fn try_run_sampling_request(
|
||||
let mut active_item_is_streaming_to_client = false;
|
||||
let receiving_span = trace_span!("receiving_stream");
|
||||
let outcome: CodexResult<SamplingRequestResult> = loop {
|
||||
if response_completed && resampling.accepted_output.is_empty() {
|
||||
if resampling.update_usage_boundary {
|
||||
sess.state.lock().await.history.reset_last_usage_boundary();
|
||||
}
|
||||
break Ok(SamplingRequestResult {
|
||||
needs_follow_up,
|
||||
mailbox_preempted: *draining_for_mailbox,
|
||||
last_agent_message,
|
||||
});
|
||||
}
|
||||
let handle_responses = trace_span!(
|
||||
parent: &receiving_span,
|
||||
"handle_responses",
|
||||
@@ -3058,8 +3072,8 @@ async fn try_run_sampling_request(
|
||||
);
|
||||
|
||||
let receive_started = Instant::now();
|
||||
let event = stream
|
||||
.next()
|
||||
let event = resampling
|
||||
.next(&mut stream)
|
||||
.instrument(trace_span!(parent: &handle_responses, "receiving"))
|
||||
.or_cancel(&preempt)
|
||||
.or_cancel(&cancellation_token)
|
||||
@@ -3112,7 +3126,11 @@ async fn try_run_sampling_request(
|
||||
Ok(event) => event,
|
||||
Err(err) => {
|
||||
if let CodexErrorDetails::ResponseIncomplete(incomplete) = err.details() {
|
||||
request_token_usage = incomplete.token_usage.clone();
|
||||
if let Some(usage) = &incomplete.token_usage {
|
||||
request_token_usage
|
||||
.get_or_insert_default()
|
||||
.add_assign(usage);
|
||||
}
|
||||
sess.record_incomplete_response_usage(
|
||||
&turn_context,
|
||||
&step_context.settings,
|
||||
@@ -3134,10 +3152,13 @@ async fn try_run_sampling_request(
|
||||
}
|
||||
};
|
||||
|
||||
record_turn_ttft_metric(&turn_context, &event).await;
|
||||
let Some(event) = resampling.filter(event, &step_context.tool_router) else {
|
||||
continue;
|
||||
};
|
||||
sess.services
|
||||
.session_telemetry
|
||||
.record_responses(&handle_responses, &event);
|
||||
record_turn_ttft_metric(&turn_context, &event).await;
|
||||
|
||||
match event {
|
||||
ResponseEvent::Created { response_id } => {
|
||||
@@ -3490,7 +3511,11 @@ async fn try_run_sampling_request(
|
||||
.await;
|
||||
}
|
||||
response_completed = true;
|
||||
request_token_usage = token_usage.clone();
|
||||
if let Some(usage) = &token_usage {
|
||||
request_token_usage
|
||||
.get_or_insert_default()
|
||||
.add_assign(usage);
|
||||
}
|
||||
sess.services
|
||||
.analytics_events_client
|
||||
.track_code_mode_tool_call(
|
||||
@@ -3520,9 +3545,13 @@ async fn try_run_sampling_request(
|
||||
&turn_context,
|
||||
&step_context.settings,
|
||||
token_usage.as_ref(),
|
||||
super::TokenUsageContext::Sampling {
|
||||
includes_reasoning,
|
||||
input_item_count,
|
||||
if resampling.retry_pending {
|
||||
super::TokenUsageContext::SpendingOnly
|
||||
} else {
|
||||
super::TokenUsageContext::Sampling {
|
||||
includes_reasoning,
|
||||
input_item_count,
|
||||
}
|
||||
},
|
||||
)
|
||||
.await;
|
||||
@@ -3531,14 +3560,23 @@ async fn try_run_sampling_request(
|
||||
if let Err(err) = budget_result {
|
||||
break Err(err);
|
||||
}
|
||||
if resampling.retry_pending {
|
||||
response_completed = false;
|
||||
response_id_for_assessment = None;
|
||||
pending_assessment = None;
|
||||
includes_reasoning = step_context.settings.model_info.use_responses_lite;
|
||||
stream = match resampling
|
||||
.retry(stream_request(&prompt), &cancellation_token)
|
||||
.await
|
||||
{
|
||||
Ok(stream) => stream,
|
||||
Err(error) => break Err(error),
|
||||
};
|
||||
continue;
|
||||
}
|
||||
if let Some(false) = end_turn {
|
||||
needs_follow_up = true;
|
||||
}
|
||||
break Ok(SamplingRequestResult {
|
||||
needs_follow_up,
|
||||
mailbox_preempted: *draining_for_mailbox,
|
||||
last_agent_message,
|
||||
});
|
||||
}
|
||||
ResponseEvent::OutputTextDelta(delta) => {
|
||||
// In review child threads, suppress assistant text deltas; the
|
||||
|
||||
Reference in New Issue
Block a user