Files
codex/codex-rs/core/src/compact_remote.rs
T
Charley Cunningham bb0ac5be70 Fix compaction context reinjection and model baselines (#12252)
## Summary
- move regular-turn context diff/full-context persistence into
`run_turn` so pre-turn compaction runs before incoming context updates
are recorded
- after successful pre-turn compaction, rely on a cleared
`reference_context_item` to trigger full context reinjection on the
follow-up regular turn (manual `/compact` keeps replacement history
summary-only and also clears the baseline)
- preserve `<model_switch>` when full context is reinjected, and inject
it *before* the rest of the full-context items
- scope `reference_context_item` and `previous_model` to regular user
turns only so standalone tasks (`/compact`, shell, review, undo) cannot
suppress future reinjection or `<model_switch>` behavior
- make context-diff persistence + `reference_context_item` updates
explicit in the regular-turn path, with clearer docs/comments around the
invariant
- stop persisting local `/compact` `RolloutItem::TurnContext` snapshots
(only regular turns persist `TurnContextItem` now)
- simplify resume/fork previous-model/reference-baseline hydration by
looking up the last surviving turn context from rollout lifecycle
events, including rollback and compaction-crossing handling
- remove the legacy fallback that guessed from bare `TurnContext`
rollouts without lifecycle events
- update compaction/remote-compaction/model-visible snapshots and
compact test assertions (including remote compaction mock response
shape)

## Why
We were persisting incoming context items before spawning the regular
turn task, which let pre-turn compaction requests accidentally include
incoming context diffs without the new user message. Fixing that exposed
follow-on baseline issues around `/compact`, resume/fork, and standalone
tasks that could cause duplicate context injection or suppress
`<model_switch>` instructions.

This PR re-centers the invariants around regular turns:
- regular turns persist model-visible context diffs/full reinjection and
update the `reference_context_item`
- standalone tasks do not advance those regular-turn baselines
- compaction clears the baseline when replacement history may have
stripped the referenced context diffs

## Follow-ups (TODOs left in code)
- `TODO(ccunningham)`: fix rollback/backtracking baseline handling more
comprehensively
- `TODO(ccunningham)`: include pending incoming context items in
pre-turn compaction threshold estimation
- `TODO(ccunningham)`: inject updated personality spec alongside
`<model_switch>` so some model-switch paths can avoid forced full
reinjection
- `TODO(ccunningham)`: review task turn lifecycle
(`TurnStarted`/`TurnComplete`) behavior and emit task-start context
diffs for task types that should have them (excluding `/compact`)

## Validation
- `just fmt`
- CI should cover the updated compaction/resume/model-visible snapshot
expectations and rollout-hydration behavior
- I did **not** rerun the full local test suite after the latest
resume-lookup / rollout-persistence simplifications
2026-02-20 23:13:08 -08:00

296 lines
11 KiB
Rust

use std::sync::Arc;
use crate::Prompt;
use crate::codex::Session;
use crate::codex::TurnContext;
use crate::compact::InitialContextInjection;
use crate::compact::extract_trailing_model_switch_update_for_compaction_request;
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 crate::error::CodexErr;
use crate::error::Result as CodexResult;
use crate::protocol::CompactedItem;
use crate::protocol::EventMsg;
use crate::protocol::RolloutItem;
use crate::protocol::TurnStartedEvent;
use codex_protocol::items::ContextCompactionItem;
use codex_protocol::items::TurnItem;
use codex_protocol::models::BaseInstructions;
use codex_protocol::models::ResponseItem;
use futures::TryFutureExt;
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,
) -> CodexResult<()> {
run_remote_compact_task_inner(&sess, &turn_context, initial_context_injection).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(),
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).await
}
async fn run_remote_compact_task_inner(
sess: &Arc<Session>,
turn_context: &Arc<TurnContext>,
initial_context_injection: InitialContextInjection,
) -> CodexResult<()> {
if let Err(err) =
run_remote_compact_task_inner_impl(sess, turn_context, initial_context_injection).await
{
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;
// Keep compaction prompts in-distribution: if a model-switch update was injected at the
// tail of history (between turns), exclude it from the compaction request payload.
let stripped_model_switch_item =
extract_trailing_model_switch_update_for_compaction_request(&mut history);
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 = Prompt {
input: history.for_prompt(&turn_context.model_info.input_modalities),
tools: vec![],
parallel_tool_calls: false,
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.otel_manager,
)
.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;
// Reattach the stripped model-switch update only after successful compaction so the model
// still sees the switch instructions on the next real sampling request.
if let Some(model_switch_item) = stripped_model_switch_item {
new_history.push(model_switch_item);
}
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()),
};
sess.replace_history(new_history.clone(), reference_context_item)
.await;
sess.recompute_token_usage(turn_context).await;
let compacted_item = CompactedItem {
message: String::new(),
replacement_history: Some(new_history),
};
sess.persist_rollout_items(&[RolloutItem::Compacted(compacted_item)])
.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),
/// keeping only real user messages as parsed by `parse_turn_item`.
///
/// 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(_))
),
ResponseItem::Message { role, .. } if role == "assistant" => true,
ResponseItem::Message { .. } => false,
ResponseItem::Compaction { .. } => true,
ResponseItem::Reasoning { .. }
| ResponseItem::LocalShellCall { .. }
| ResponseItem::FunctionCall { .. }
| ResponseItem::FunctionCallOutput { .. }
| ResponseItem::CustomToolCall { .. }
| ResponseItem::CustomToolCallOutput { .. }
| ResponseItem::WebSearchCall { .. }
| 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
}