Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
49 changes: 9 additions & 40 deletions src/crates/assembly/core/src/agentic/coordination/coordinator.rs
Original file line number Diff line number Diff line change
Expand Up @@ -2185,7 +2185,7 @@ Update the persona files and delete BOOTSTRAP.md as soon as bootstrap is complet
fn build_manual_compaction_round_completed(
turn_id: &str,
outcome: &ContextCompactionOutcome,
context_window: usize,
context_window: Option<usize>,
) -> crate::service::session::ModelRoundData {
use crate::service::session::{ModelRoundData, ToolCallData, ToolItemData, ToolResultData};

Expand Down Expand Up @@ -2274,7 +2274,7 @@ Update the persona files and delete BOOTSTRAP.md as soon as bootstrap is complet
turn_id: &str,
compression_id: String,
error: &str,
context_window: usize,
context_window: Option<usize>,
) -> crate::service::session::ModelRoundData {
use crate::service::session::{ModelRoundData, ToolCallData, ToolItemData, ToolResultData};

Expand Down Expand Up @@ -6279,7 +6279,6 @@ Update the persona files and delete BOOTSTRAP.md as soon as bootstrap is complet
session_manager,
execution_engine,
event_queue,
session,
context_messages,
session_id_for_task,
turn_id_for_task,
Expand Down Expand Up @@ -6308,7 +6307,7 @@ Update the persona files and delete BOOTSTRAP.md as soon as bootstrap is complet
session_id: &str,
turn_id: &str,
outcome: &ContextCompactionOutcome,
context_window: usize,
context_window: Option<usize>,
) -> OpenBitFunResult<()> {
let model_round =
Self::build_manual_compaction_round_completed(turn_id, outcome, context_window);
Expand Down Expand Up @@ -6388,7 +6387,6 @@ Update the persona files and delete BOOTSTRAP.md as soon as bootstrap is complet
session_manager: Arc<SessionManager>,
execution_engine: Arc<ExecutionEngine>,
event_queue: Arc<EventQueue>,
session: Session,
context_messages: Vec<Message>,
session_id: String,
turn_id: String,
Expand All @@ -6401,17 +6399,17 @@ Update the persona files and delete BOOTSTRAP.md as soon as bootstrap is complet
commit_gate: Arc<ManualCompactionCommitGate>,
) -> OpenBitFunResult<()> {
let compression_id = format!("compression_{}", uuid::Uuid::new_v4());
let mut context_window = session.config.max_context_tokens;
let mut context_window = None;
// Preparation and compaction embed the full model/tool-catalog future.
// Keep that state on the heap instead of duplicating it through this
// task and its callers: even a pre-cancelled turn must enter the poll
// stack frame before it can observe cancellation.
let result = Box::pin(async {
let (manual_execution_context, resolved_context_window) =
let manual_execution_context =
prepare_compression_cancellable(&cancellation_token, async {
let manual_workspace_services =
Self::build_workspace_services(&manual_workspace).await?;
let manual_execution_context = ExecutionContext {
Ok(ExecutionContext {
session_id: session_id.clone(),
dialog_turn_id: turn_id.clone(),
turn_index,
Expand All @@ -6432,30 +6430,9 @@ Update the persona files and delete BOOTSTRAP.md as soon as bootstrap is complet
round_injection: None,
emit_lifecycle_events: false,
recover_partial_on_cancel: false,
};
let session_max_tokens = session.config.max_context_tokens;

// Unify context_window: min(model capability, session config)
let model_context_window =
match crate::infrastructure::ai::get_global_ai_client_factory().await {
Ok(factory) => {
let model_id =
session.config.model_id.as_deref().unwrap_or("default");
match factory.get_client_resolved(model_id).await {
Ok(client) => Some(client.config.context_window as usize),
Err(_) => None,
}
}
Err(_) => None,
};
let context_window = match model_context_window {
Some(mcw) => mcw.min(session_max_tokens),
None => session_max_tokens,
};
Ok((manual_execution_context, context_window))
})
})
.await?;
context_window = resolved_context_window;
execution_engine
.compact_session_context(
session_id.clone(),
Expand All @@ -6466,6 +6443,7 @@ Update the persona files and delete BOOTSTRAP.md as soon as bootstrap is complet
"manual",
cancellation_token,
commit_gate,
&mut context_window,
)
.await
})
Expand Down Expand Up @@ -7533,10 +7511,6 @@ Update the persona files and delete BOOTSTRAP.md as soon as bootstrap is complet

// Create execution context (pass full config and resource IDs)
let mut context_vars = std::collections::HashMap::new();
context_vars.insert(
"max_context_tokens".to_string(),
session.config.max_context_tokens.to_string(),
);
context_vars.insert(
"enable_tools".to_string(),
session.config.enable_tools.to_string(),
Expand Down Expand Up @@ -8116,10 +8090,6 @@ Update the persona files and delete BOOTSTRAP.md as soon as bootstrap is complet
.is_transient_session(&request.session_id),
);
let mut context_vars = HashMap::new();
context_vars.insert(
"max_context_tokens".to_string(),
session.config.max_context_tokens.to_string(),
);
context_vars.insert(
"enable_tools".to_string(),
session.config.enable_tools.to_string(),
Expand Down Expand Up @@ -11131,8 +11101,7 @@ Update the persona files and delete BOOTSTRAP.md as soon as bootstrap is complet
let session_id = session.session_id.clone();
let _execution_lease =
execution_lease.unwrap_or_else(|| self.register_session_execution(&session_id));
// Sync context window from AI config so subagents with large-context
// models are not prematurely capped at SessionConfig::default()'s 128128.
// Keep the legacy session window snapshot aligned with the subagent model.
if let Err(error) = self
.session_manager
.refresh_session_context_window(&session_id)
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -113,7 +113,7 @@ fn manual_compaction_transcript_restores_user_and_tool_payload() {
ConversationCoordinator::build_manual_compaction_round_completed(
&turn.turn_id,
&outcome,
128_000,
Some(128_000),
),
];
turn.status = TurnStatus::Completed;
Expand Down Expand Up @@ -177,7 +177,7 @@ fn manual_compaction_failure_round_preserves_runtime_identity_and_error() {
"compact-turn",
"compression-runtime".to_string(),
"summary request failed",
128_000,
Some(128_000),
);

assert_eq!(round.tool_items.len(), 1);
Expand Down Expand Up @@ -227,7 +227,6 @@ async fn manual_compaction_cancelled_before_setup_preserves_context_and_settles_
session_manager.clone(),
coordinator.execution_engine.clone(),
coordinator.event_queue.clone(),
session,
original,
session_id.clone(),
turn_id.clone(),
Expand Down Expand Up @@ -322,7 +321,7 @@ async fn applied_manual_compaction_emits_failed_terminal_when_turn_persistence_f
summary_source: "model".to_string(),
applied: true,
},
128_000,
Some(128_000),
)
.await;

Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -296,6 +296,56 @@ async fn current_context_estimate_remote_subagent_uses_cached_rules_without_loca
.await;
}

#[tokio::test]
async fn current_context_estimate_uses_the_model_window_with_legacy_session_limits() {
use crate::agentic::core::Session;
use openbitfun_core_types::UsageContextEstimateRequest;

let root = tempfile::tempdir().unwrap();
crate::service::workspace::legacy_compat::register_local_fixture_blocking(root.path());
let mut config = context_estimate_ai_config();
config.models[1].context_window = Some(1_048_576);
config.models[1].max_tokens = Some(384_000);
TEST_MODEL_RESOLUTION_AI_CONFIG
.scope(config, async {
let (coordinator, manager) = test_persistent_coordinator();
for (saved_window, model, expected_window) in [
(128_128, "child-model", 1_048_576),
(1_048_576, "parent-model", 64_000),
] {
let session = Session::new(
"Saved session".into(),
"Standard".into(),
SessionConfig {
workspace_path: Some(root.path().to_string_lossy().into_owned()),
model_id: Some(model.into()),
max_context_tokens: saved_window,
..Default::default()
},
);
manager
.persistence_manager()
.save_session(root.path(), &session)
.await
.unwrap();
let inputs = manager
.read_context_estimate_inputs(root.path(), &session.session_id)
.await
.unwrap();
assert_eq!(inputs.session.config.max_context_tokens, saved_window);
let estimate = coordinator
.estimate_session_context(inputs, &UsageContextEstimateRequest::default())
.await
.unwrap();
assert_eq!(
estimate.composition.unwrap().context_window,
Some(expected_window)
);
}
})
.await;
}

#[tokio::test]
async fn current_context_estimate_keeps_external_and_internal_worker_contexts_unsupported() {
use crate::agentic::tools::ToolRuntimeRestrictions;
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -59,7 +59,6 @@ impl ExecutionEngine {
trigger: &str,
runtime_messages: Vec<Message>,
before_pressure: TokenPressureSnapshot,
context_window: usize,
ai_client: Arc<crate::infrastructure::ai::AIClient>,
model_request_context: &ModelRequestContext,
tool_definitions: &Option<Vec<ToolDefinition>>,
Expand All @@ -77,6 +76,8 @@ impl ExecutionEngine {
OpenBitFunError::NotFound(format!("Session not found: {}", session_id))
})?;

let context_window = before_pressure.context_window;

// Record start time
let start_time = std::time::Instant::now();

Expand Down Expand Up @@ -276,11 +277,7 @@ impl ExecutionEngine {
let pressure = Self::estimate_auto_compression_pressure(
&preview,
tool_definitions.as_deref(),
context_window,
Self::compression_trigger_budget(
context_window,
ai_client.config.max_tokens,
),
before_pressure.trigger_budget(),
Self::prepended_reminder_tokens_for_pressure(
&prepended_prompt_reminders.ordered_reminders(),
),
Expand Down Expand Up @@ -375,12 +372,7 @@ impl ExecutionEngine {
let after_pressure = Self::estimate_auto_compression_pressure(
&new_messages,
tool_definitions.as_deref(),
context_window,
CompressionTriggerBudget {
input_limit: before_pressure.input_limit,
output_reserve_tokens: before_pressure.output_reserve_tokens,
safety_reserve_tokens: before_pressure.safety_reserve_tokens,
},
before_pressure.trigger_budget(),
prepended_reminder_tokens,
);
let compressed_tokens = after_pressure.total_tokens;
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -125,8 +125,7 @@ async fn apply_prefetch(
let pressure = ExecutionEngine::estimate_auto_compression_pressure(
messages,
None,
128_000,
ExecutionEngine::compression_trigger_budget(128_000, None),
CompressionTriggerBudget::for_model(128_000, None),
0,
);
engine
Expand All @@ -136,7 +135,6 @@ async fn apply_prefetch(
"auto",
messages.to_vec(),
pressure,
128_000,
server.client(),
&ModelRequestContext::default(),
&None,
Expand Down Expand Up @@ -579,8 +577,7 @@ async fn compression_failure_preserves_context_and_success_count() {
let pressure = ExecutionEngine::estimate_auto_compression_pressure(
&messages,
None,
128_000,
ExecutionEngine::compression_trigger_budget(128_000, None),
CompressionTriggerBudget::for_model(128_000, None),
0,
);
let error = engine
Expand All @@ -590,7 +587,6 @@ async fn compression_failure_preserves_context_and_success_count() {
"auto",
messages,
pressure,
128_000,
server.client(),
&ModelRequestContext::default(),
&None,
Expand Down
39 changes: 39 additions & 0 deletions src/crates/assembly/core/src/agentic/execution/context_budget.rs
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,7 @@ use openbitfun_core_types::model_pool::ModelPoolContextBudget;

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(super) struct CompressionTriggerBudget {
pub context_window: usize,
pub input_limit: usize,
pub output_reserve_tokens: usize,
pub safety_reserve_tokens: usize,
Expand All @@ -13,12 +14,25 @@ pub(super) struct CompressionTriggerBudget {
impl CompressionTriggerBudget {
const SAFETY_RESERVE_TOKENS: usize = 10_000;

/// The active model or pool owns the request budget. Persisted session
/// windows can describe a previous model or a legacy creation default.
pub(super) fn for_request(
model: &openbitfun_core_types::AIConfig,
pool: Option<ModelPoolContextBudget>,
) -> Self {
match pool {
Some(budget) => Self::for_pool(budget),
None => Self::for_model(model.context_window as usize, model.max_tokens),
}
}

pub(super) fn for_model(context_window: usize, configured_max_tokens: Option<u32>) -> Self {
let output_reserve_tokens = configured_max_tokens
.map(|value| value as usize)
.unwrap_or_else(|| automatic_max_output_tokens(context_window as u32) as usize);
let safety_reserve_tokens = Self::SAFETY_RESERVE_TOKENS;
Self {
context_window,
input_limit: context_window
.saturating_sub(output_reserve_tokens + safety_reserve_tokens),
output_reserve_tokens,
Expand All @@ -31,6 +45,7 @@ impl CompressionTriggerBudget {
let input_limit = budget.input_limit as usize;
let safety_reserve_tokens = Self::SAFETY_RESERVE_TOKENS.min(context_window);
Self {
context_window,
input_limit,
// This is an equivalent reserve for the common window, not a new
// max_tokens setting for any member's provider request.
Expand All @@ -40,3 +55,27 @@ impl CompressionTriggerBudget {
}
}
}

#[cfg(test)]
mod tests {
use super::*;
use crate::service::config::types::AIModelConfig;

#[test]
fn compression_budget_follows_each_selected_model() {
for (window, max_output, input_limit) in
[(1_048_576, 384_000, 654_576), (64_000, 16_000, 38_000)]
{
let model = openbitfun_core_types::AIConfig::try_from(AIModelConfig {
context_window: Some(window),
max_tokens: Some(max_output),
..Default::default()
})
.unwrap();
let budget = CompressionTriggerBudget::for_request(&model, None);
assert_eq!(budget.context_window, window as usize);
assert_eq!(budget.input_limit, input_limit);
assert_eq!(47_544 >= budget.input_limit, window == 64_000);
}
}
}
Loading
Loading