Files
codex/codex-rs/core/src/compact_remote.rs
rhan-oai 5779be314a [codex-analytics] add compaction analytics event (#17155)
- event for compaction analytics
- introduces thread-connection and thread metadata caches for data
denormalization, expected to be useful for denormalization onto core
emitted events in general
- threads analytics event client into core (mirrors approved
implementation in #16640)
- denormalizes key thread metadata: thread_source, subagent_source,
parent_thread_id, as well as app-server client and runtime metadata)
- compaction strategy defaults to memento, forward compatible with
expected prefill_compaction strategy

1. Manual standalone compact, local
`INFO | 2026-04-09 17:35:50 | codex_backend.routers.analytics_events |
analytics_events.track_analytics_events:526 | Tracked
codex_compaction_event event params={'thread_id':
'019d74d0-5cfb-70c0-bef9-165c3bf9b2df', 'turn_id':
'019d74d0-d7f6-7c81-acc6-aae2030243d6', 'product_surface': 'codex',
'app_server_client': {'product_client_id': 'CODEX_CLI', 'client_name':
'codex-tui', 'client_version': '0.0.0', 'rpc_transport': 'in_process',
'experimental_api_enabled': True}, 'runtime': {'codex_rs_version':
'0.0.0', 'runtime_os': 'macos', 'runtime_os_version': '26.4.0',
'runtime_arch': 'aarch64'}, 'trigger': 'manual', 'reason':
'user_requested', 'implementation': 'responses', 'phase':
'standalone_turn', 'strategy': 'memento', 'status': 'completed',
'active_context_tokens_before': 20170, 'active_context_tokens_after':
4830, 'started_at': 1775781337, 'completed_at': 1775781350,
'thread_source': 'user', 'subagent_source': None, 'parent_thread_id':
None, 'error': None, 'duration_ms': 13524} | `

2. Auto pre-turn compact, local
`INFO | 2026-04-09 17:37:30 | codex_backend.routers.analytics_events |
analytics_events.track_analytics_events:526 | Tracked
codex_compaction_event event params={'thread_id':
'019d74d2-45ef-71d1-9c93-23cc0c13d988', 'turn_id':
'019d74d2-7b42-7372-9f0e-c0da3f352328', 'product_surface': 'codex',
'app_server_client': {'product_client_id': 'CODEX_CLI', 'client_name':
'codex-tui', 'client_version': '0.0.0', 'rpc_transport': 'in_process',
'experimental_api_enabled': True}, 'runtime': {'codex_rs_version':
'0.0.0', 'runtime_os': 'macos', 'runtime_os_version': '26.4.0',
'runtime_arch': 'aarch64'}, 'trigger': 'auto', 'reason':
'context_limit', 'implementation': 'responses', 'phase': 'pre_turn',
'strategy': 'memento', 'status': 'completed',
'active_context_tokens_before': 20063, 'active_context_tokens_after':
4822, 'started_at': 1775781444, 'completed_at': 1775781449,
'thread_source': 'user', 'subagent_source': None, 'parent_thread_id':
None, 'error': None, 'duration_ms': 5497} | `

3. Auto mid-turn compact, local
`INFO | 2026-04-09 17:38:28 | codex_backend.routers.analytics_events |
analytics_events.track_analytics_events:526 | Tracked
codex_compaction_event event params={'thread_id':
'019d74d3-212f-7a20-8c0a-4816a978675e', 'turn_id':
'019d74d3-3ee1-7462-89f6-2ffbeefcd5e3', 'product_surface': 'codex',
'app_server_client': {'product_client_id': 'CODEX_CLI', 'client_name':
'codex-tui', 'client_version': '0.0.0', 'rpc_transport': 'in_process',
'experimental_api_enabled': True}, 'runtime': {'codex_rs_version':
'0.0.0', 'runtime_os': 'macos', 'runtime_os_version': '26.4.0',
'runtime_arch': 'aarch64'}, 'trigger': 'auto', 'reason':
'context_limit', 'implementation': 'responses', 'phase': 'mid_turn',
'strategy': 'memento', 'status': 'completed',
'active_context_tokens_before': 20325, 'active_context_tokens_after':
14641, 'started_at': 1775781500, 'completed_at': 1775781508,
'thread_source': 'user', 'subagent_source': None, 'parent_thread_id':
None, 'error': None, 'duration_ms': 7507} | `

4. Remote /responses/compact, manual standalone
`INFO | 2026-04-09 17:40:20 | codex_backend.routers.analytics_events |
analytics_events.track_analytics_events:526 | Tracked
codex_compaction_event event params={'thread_id':
'019d74d4-7a11-78a1-89f7-0535a1149416', 'turn_id':
'019d74d4-e087-7183-9c20-b1e40b7578c0', 'product_surface': 'codex',
'app_server_client': {'product_client_id': 'CODEX_CLI', 'client_name':
'codex-tui', 'client_version': '0.0.0', 'rpc_transport': 'in_process',
'experimental_api_enabled': True}, 'runtime': {'codex_rs_version':
'0.0.0', 'runtime_os': 'macos', 'runtime_os_version': '26.4.0',
'runtime_arch': 'aarch64'}, 'trigger': 'manual', 'reason':
'user_requested', 'implementation': 'responses_compact', 'phase':
'standalone_turn', 'strategy': 'memento', 'status': 'completed',
'active_context_tokens_before': 23461, 'active_context_tokens_after':
6171, 'started_at': 1775781601, 'completed_at': 1775781620,
'thread_source': 'user', 'subagent_source': None, 'parent_thread_id':
None, 'error': None, 'duration_ms': 18971} | `
2026-04-10 13:03:54 -07:00

345 lines
12 KiB
Rust

use std::collections::HashSet;
use std::sync::Arc;
use crate::Prompt;
use crate::codex::Session;
use crate::codex::TurnContext;
use crate::codex::built_tools;
use crate::compact::CompactionAnalyticsAttempt;
use crate::compact::InitialContextInjection;
use crate::compact::compaction_status_from_result;
use crate::compact::insert_initial_context_before_last_real_user_or_summary;
use crate::context_manager::ContextManager;
use crate::context_manager::TotalTokenUsageBreakdown;
use crate::context_manager::estimate_response_item_model_visible_bytes;
use crate::context_manager::is_codex_generated_item;
use codex_analytics::CompactionImplementation;
use codex_analytics::CompactionPhase;
use codex_analytics::CompactionReason;
use codex_analytics::CompactionTrigger;
use codex_protocol::error::CodexErr;
use codex_protocol::error::Result as CodexResult;
use codex_protocol::items::ContextCompactionItem;
use codex_protocol::items::TurnItem;
use codex_protocol::models::BaseInstructions;
use codex_protocol::models::ResponseItem;
use codex_protocol::protocol::CompactedItem;
use codex_protocol::protocol::EventMsg;
use codex_protocol::protocol::TurnStartedEvent;
use futures::TryFutureExt;
use tokio_util::sync::CancellationToken;
use tracing::error;
use tracing::info;
pub(crate) async fn run_inline_remote_auto_compact_task(
sess: Arc<Session>,
turn_context: Arc<TurnContext>,
initial_context_injection: InitialContextInjection,
reason: CompactionReason,
phase: CompactionPhase,
) -> CodexResult<()> {
run_remote_compact_task_inner(
&sess,
&turn_context,
initial_context_injection,
CompactionTrigger::Auto,
reason,
phase,
)
.await?;
Ok(())
}
pub(crate) async fn run_remote_compact_task(
sess: Arc<Session>,
turn_context: Arc<TurnContext>,
) -> CodexResult<()> {
let start_event = EventMsg::TurnStarted(TurnStartedEvent {
turn_id: turn_context.sub_id.clone(),
started_at: turn_context.turn_timing_state.started_at_unix_secs().await,
model_context_window: turn_context.model_context_window(),
collaboration_mode_kind: turn_context.collaboration_mode.mode,
});
sess.send_event(&turn_context, start_event).await;
run_remote_compact_task_inner(
&sess,
&turn_context,
InitialContextInjection::DoNotInject,
CompactionTrigger::Manual,
CompactionReason::UserRequested,
CompactionPhase::StandaloneTurn,
)
.await
}
async fn run_remote_compact_task_inner(
sess: &Arc<Session>,
turn_context: &Arc<TurnContext>,
initial_context_injection: InitialContextInjection,
trigger: CompactionTrigger,
reason: CompactionReason,
phase: CompactionPhase,
) -> CodexResult<()> {
let attempt = CompactionAnalyticsAttempt::begin(
sess.as_ref(),
turn_context.as_ref(),
trigger,
reason,
CompactionImplementation::ResponsesCompact,
phase,
)
.await;
let result =
run_remote_compact_task_inner_impl(sess, turn_context, initial_context_injection).await;
attempt
.track(
sess.as_ref(),
compaction_status_from_result(&result),
result.as_ref().err().map(ToString::to_string),
)
.await;
if let Err(err) = result {
let event = EventMsg::Error(
err.to_error_event(Some("Error running remote compact task".to_string())),
);
sess.send_event(turn_context, event).await;
return Err(err);
}
Ok(())
}
async fn run_remote_compact_task_inner_impl(
sess: &Arc<Session>,
turn_context: &Arc<TurnContext>,
initial_context_injection: InitialContextInjection,
) -> CodexResult<()> {
let compaction_item = TurnItem::ContextCompaction(ContextCompactionItem::new());
sess.emit_turn_item_started(turn_context, &compaction_item)
.await;
let mut history = sess.clone_history().await;
let base_instructions = sess.get_base_instructions().await;
let deleted_items = trim_function_call_history_to_fit_context_window(
&mut history,
turn_context.as_ref(),
&base_instructions,
);
if deleted_items > 0 {
info!(
turn_id = %turn_context.sub_id,
deleted_items,
"trimmed history items before remote compaction"
);
}
// Required to keep `/undo` available after compaction
let ghost_snapshots: Vec<ResponseItem> = history
.raw_items()
.iter()
.filter(|item| matches!(item, ResponseItem::GhostSnapshot { .. }))
.cloned()
.collect();
let prompt_input = history.for_prompt(&turn_context.model_info.input_modalities);
let tool_router = built_tools(
sess.as_ref(),
turn_context.as_ref(),
&prompt_input,
&HashSet::new(),
/*skills_outcome*/ None,
&CancellationToken::new(),
)
.await?;
let prompt = Prompt {
input: prompt_input,
tools: tool_router.model_visible_specs(),
parallel_tool_calls: turn_context.model_info.supports_parallel_tool_calls,
base_instructions,
personality: turn_context.personality,
output_schema: None,
};
let mut new_history = sess
.services
.model_client
.compact_conversation_history(
&prompt,
&turn_context.model_info,
turn_context.reasoning_effort,
turn_context.reasoning_summary,
&turn_context.session_telemetry,
)
.or_else(|err| async {
let total_usage_breakdown = sess.get_total_token_usage_breakdown().await;
let compact_request_log_data =
build_compact_request_log_data(&prompt.input, &prompt.base_instructions.text);
log_remote_compact_failure(
turn_context,
&compact_request_log_data,
total_usage_breakdown,
&err,
);
Err(err)
})
.await?;
new_history = process_compacted_history(
sess.as_ref(),
turn_context.as_ref(),
new_history,
initial_context_injection,
)
.await;
if !ghost_snapshots.is_empty() {
new_history.extend(ghost_snapshots);
}
let reference_context_item = match initial_context_injection {
InitialContextInjection::DoNotInject => None,
InitialContextInjection::BeforeLastUserMessage => Some(turn_context.to_turn_context_item()),
};
let compacted_item = CompactedItem {
message: String::new(),
replacement_history: Some(new_history.clone()),
};
sess.replace_compacted_history(new_history, reference_context_item, compacted_item)
.await;
sess.recompute_token_usage(turn_context).await;
sess.emit_turn_item_completed(turn_context, compaction_item)
.await;
Ok(())
}
pub(crate) async fn process_compacted_history(
sess: &Session,
turn_context: &TurnContext,
mut compacted_history: Vec<ResponseItem>,
initial_context_injection: InitialContextInjection,
) -> Vec<ResponseItem> {
// Mid-turn compaction is the only path that must inject initial context above the last user
// message in the replacement history. Pre-turn compaction instead injects context after the
// compaction item, but mid-turn compaction keeps the compaction item last for model training.
let initial_context = if matches!(
initial_context_injection,
InitialContextInjection::BeforeLastUserMessage
) {
sess.build_initial_context(turn_context).await
} else {
Vec::new()
};
compacted_history.retain(should_keep_compacted_history_item);
insert_initial_context_before_last_real_user_or_summary(compacted_history, initial_context)
}
/// Returns whether an item from remote compaction output should be preserved.
///
/// Called while processing the model-provided compacted transcript, before we
/// append fresh canonical context from the current session.
///
/// We drop:
/// - `developer` messages because remote output can include stale/duplicated
/// instruction content.
/// - non-user-content `user` messages (session prefix/instruction wrappers),
/// while preserving real user messages and persisted hook prompts.
///
/// This intentionally keeps:
/// - `assistant` messages (future remote compaction models may emit them)
/// - `user`-role warnings and compaction-generated summary messages because
/// they parse as `TurnItem::UserMessage`.
fn should_keep_compacted_history_item(item: &ResponseItem) -> bool {
match item {
ResponseItem::Message { role, .. } if role == "developer" => false,
ResponseItem::Message { role, .. } if role == "user" => {
matches!(
crate::event_mapping::parse_turn_item(item),
Some(TurnItem::UserMessage(_) | TurnItem::HookPrompt(_))
)
}
ResponseItem::Message { role, .. } if role == "assistant" => true,
ResponseItem::Message { .. } => false,
ResponseItem::Compaction { .. } => true,
ResponseItem::Reasoning { .. }
| ResponseItem::LocalShellCall { .. }
| ResponseItem::FunctionCall { .. }
| ResponseItem::ToolSearchCall { .. }
| ResponseItem::FunctionCallOutput { .. }
| ResponseItem::ToolSearchOutput { .. }
| ResponseItem::CustomToolCall { .. }
| ResponseItem::CustomToolCallOutput { .. }
| ResponseItem::WebSearchCall { .. }
| ResponseItem::ImageGenerationCall { .. }
| ResponseItem::GhostSnapshot { .. }
| ResponseItem::Other => false,
}
}
#[derive(Debug)]
struct CompactRequestLogData {
failing_compaction_request_model_visible_bytes: i64,
}
fn build_compact_request_log_data(
input: &[ResponseItem],
instructions: &str,
) -> CompactRequestLogData {
let failing_compaction_request_model_visible_bytes = input
.iter()
.map(estimate_response_item_model_visible_bytes)
.fold(
i64::try_from(instructions.len()).unwrap_or(i64::MAX),
i64::saturating_add,
);
CompactRequestLogData {
failing_compaction_request_model_visible_bytes,
}
}
fn log_remote_compact_failure(
turn_context: &TurnContext,
log_data: &CompactRequestLogData,
total_usage_breakdown: TotalTokenUsageBreakdown,
err: &CodexErr,
) {
error!(
turn_id = %turn_context.sub_id,
last_api_response_total_tokens = total_usage_breakdown.last_api_response_total_tokens,
all_history_items_model_visible_bytes = total_usage_breakdown.all_history_items_model_visible_bytes,
estimated_tokens_of_items_added_since_last_successful_api_response = total_usage_breakdown.estimated_tokens_of_items_added_since_last_successful_api_response,
estimated_bytes_of_items_added_since_last_successful_api_response = total_usage_breakdown.estimated_bytes_of_items_added_since_last_successful_api_response,
model_context_window_tokens = ?turn_context.model_context_window(),
failing_compaction_request_model_visible_bytes = log_data.failing_compaction_request_model_visible_bytes,
compact_error = %err,
"remote compaction failed"
);
}
fn trim_function_call_history_to_fit_context_window(
history: &mut ContextManager,
turn_context: &TurnContext,
base_instructions: &BaseInstructions,
) -> usize {
let mut deleted_items = 0usize;
let Some(context_window) = turn_context.model_context_window() else {
return deleted_items;
};
while history
.estimate_token_count_with_base_instructions(base_instructions)
.is_some_and(|estimated_tokens| estimated_tokens > context_window)
{
let Some(last_item) = history.raw_items().last() else {
break;
};
if !is_codex_generated_item(last_item) {
break;
}
if !history.remove_last_item() {
break;
}
deleted_items += 1;
}
deleted_items
}