Add code-mode output truncation protocol contracts (#57272)

## What changed

Add session options for client- or server-owned output truncation and optional `max_output_tokens` budgets for wait and terminate observations. Introduce `TerminateRequest` and update session implementations and callers to use it.

Extend the stdio and gRPC schemas with optional truncation settings and budgets, and define the `bounded-output-v1` capability. Existing sessions retain client-side truncation; providers reject server mode by default. This change defines the contracts without enabling server-side truncation.

## Testing

Extend protocol serialization coverage for `outputTruncation: "server"` while preserving existing wire shapes when optional fields are absent.

GitOrigin-RevId: 22145897ea16d385c5ba9e7549109f7a669fceb4
This commit is contained in:
Sean Huang
2026-10-26 21:36:35 +00:00
committed by copyberry
parent 7077b1edb9
commit e963cab11a
38 changed files with 329 additions and 55 deletions
+1
View File
@@ -234,6 +234,7 @@ impl GrpcCodeModeHost {
let request = WaitRequest {
cell_id: CellId::new(request.cell_id),
yield_time_ms: request.yield_time_ms,
max_output_tokens: None,
};
let outcome = tokio::select! {
biased;
@@ -255,6 +255,7 @@ async fn session_closure_cancels_pending_termination() {
let termination = host.terminate(Request::new(proto::TerminateRequest {
session_id,
cell_id,
max_output_tokens: None,
}));
tokio::pin!(termination);
assert!(termination.as_mut().now_or_never().is_none());
@@ -14,6 +14,7 @@ pub(super) async fn open_session(
let mut stream = host
.open_session(Request::new(proto::OpenSessionRequest {
cell_execution_limits: None,
output_truncation: None,
}))
.await
.expect("open code-mode session")
@@ -120,6 +121,7 @@ async fn execute_stream_starts_immediately_and_wait_preserves_missing_cells() {
cell_id: cell_id.clone(),
wait_id: "wait-1".to_string(),
yield_time_ms: 60_000,
max_output_tokens: None,
}))
.await
.expect("wait for completion")
@@ -163,6 +165,7 @@ async fn execute_stream_starts_immediately_and_wait_preserves_missing_cells() {
cell_id: "missing-cell".to_string(),
wait_id: "wait-missing".to_string(),
yield_time_ms: 1,
max_output_tokens: None,
}))
.await
.expect("missing cells remain successful wait outcomes")
@@ -355,6 +358,7 @@ async fn cancellation_before_wait_admission_is_preserved() {
cell_id: cell_id.clone(),
wait_id: "pre-cancelled".to_string(),
yield_time_ms: 60_000,
max_output_tokens: None,
}))
.await
.unwrap_err();
@@ -363,6 +367,7 @@ async fn cancellation_before_wait_admission_is_preserved() {
host.terminate(Request::new(proto::TerminateRequest {
session_id,
cell_id,
max_output_tokens: None,
}))
.await
.unwrap();
+2 -1
View File
@@ -124,6 +124,7 @@ impl GrpcHostState {
.try_send(Ok(proto::SessionEvent {
event: Some(proto::session_event::Event::Opened(proto::SessionOpened {
session_id: id.to_string(),
output_truncation: None,
})),
}))
.map_err(|_| Status::internal("failed to publish the opened code-mode session"))?;
@@ -263,7 +264,7 @@ impl GrpcSession {
_ = self.closed.cancelled() => {
Err(Status::cancelled("code-mode session is closed"))
}
result = self.runtime.terminate(cell_id) => {
result = self.runtime.terminate(cell_id.into()) => {
result.map_err(Status::failed_precondition)
}
}
+6 -1
View File
@@ -127,6 +127,7 @@ async fn handshake_and_multiple_session_lifecycles_are_ordered() {
request: HostRequest::OpenSession {
session_id: session_id(id),
cell_execution_limits: None,
output_truncation: None,
},
})
.await
@@ -198,6 +199,7 @@ async fn disconnect_cancels_a_backpressured_host_writer() {
request: HostRequest::OpenSession {
session_id: session_id("backpressured-session"),
cell_execution_limits: None,
output_truncation: None,
},
})
.await
@@ -339,6 +341,7 @@ async fn incompatible_or_invalid_handshake_is_rejected() {
request: HostRequest::OpenSession {
session_id: session_id("session-1"),
cell_execution_limits: None,
output_truncation: None,
},
})
.await
@@ -405,6 +408,7 @@ async fn session_id_cannot_be_reused_after_shutdown() {
HostRequest::OpenSession {
session_id: id.clone(),
cell_execution_limits: None,
output_truncation: None,
},
),
(
@@ -433,6 +437,7 @@ async fn session_id_cannot_be_reused_after_shutdown() {
request: HostRequest::OpenSession {
session_id: id,
cell_execution_limits: None,
output_truncation: None,
},
})
.await
@@ -556,7 +561,7 @@ async fn execute_request_id_remains_active_until_initial_response() {
state
.session(&session_id)
.expect("session")
.terminate(cell_id.into())
.terminate(codex_code_mode_protocol::CellId::from(cell_id).into())
.await
.expect("terminate cell");
state.disconnect().await;
+14 -7
View File
@@ -382,6 +382,7 @@ impl HostState {
HostRequest::OpenSession {
session_id,
cell_execution_limits,
..
} => {
let result = CodeModeSessionCellExecutionLimits::try_from(
cell_execution_limits.unwrap_or_default(),
@@ -452,6 +453,7 @@ impl HostState {
HostRequest::Wait {
session_id,
request,
..
} => {
let result = match self.session(&session_id) {
Ok(session) => {
@@ -476,15 +478,20 @@ impl HostState {
HostRequest::Terminate {
session_id,
cell_id,
..
} => {
let result = match self.session(&session_id) {
Ok(session) => session.terminate(cell_id.into()).await.and_then(|outcome| {
let outcome = outcome.with_code_mode_host_duration(received_at.elapsed());
Ok(HostResponse::WaitCompleted {
outcome: WireWaitOutcome::try_from(outcome)
.map_err(|error| error.to_string())?,
})
}),
Ok(session) => session
.terminate(codex_code_mode_protocol::CellId::from(cell_id).into())
.await
.and_then(|outcome| {
let outcome =
outcome.with_code_mode_host_duration(received_at.elapsed());
Ok(HostResponse::WaitCompleted {
outcome: WireWaitOutcome::try_from(outcome)
.map_err(|error| error.to_string())?,
})
}),
Err(err) => Err(err),
};
self.respond(request_id, result);
+26 -7
View File
@@ -327,6 +327,7 @@ async fn grpc_early_yield_preserves_cell_and_later_waits() -> Result<()> {
WaitRequest {
cell_id: cell_id.clone(),
yield_time_ms: 60_000,
max_output_tokens: None,
},
Some(wait_signal.clone()),
);
@@ -347,6 +348,7 @@ async fn grpc_early_yield_preserves_cell_and_later_waits() -> Result<()> {
WaitRequest {
cell_id: cell_id.clone(),
yield_time_ms: 60_000,
max_output_tokens: None,
},
/*preempt*/ None,
),
@@ -397,13 +399,17 @@ async fn shutdown_immediately_rejects_new_operations() -> Result<()> {
WaitRequest {
cell_id: cell_id("missing"),
yield_time_ms: 1,
max_output_tokens: None,
},
/*preempt*/ None,
)
.await,
Err(expected.clone())
);
assert_eq!(session.terminate(cell_id("missing")).await, Err(expected));
assert_eq!(
session.terminate(cell_id("missing").into()).await,
Err(expected)
);
shutdown.await.map_err(anyhow::Error::msg)?;
Ok(())
@@ -452,6 +458,7 @@ async fn cancelling_execution_before_admission_keeps_the_session_usable() -> Res
WaitRequest {
cell_id: abandoned_cell.clone(),
yield_time_ms: 1,
max_output_tokens: None,
},
/*preempt*/ None,
)
@@ -512,6 +519,7 @@ async fn dropping_a_started_cell_off_runtime_terminates_its_buffered_remote_exec
WaitRequest {
cell_id: abandoned_cell.clone(),
yield_time_ms: 1,
max_output_tokens: None,
},
/*preempt*/ None,
)
@@ -540,6 +548,7 @@ async fn dropping_a_started_cell_off_runtime_terminates_its_buffered_remote_exec
WaitRequest {
cell_id: abandoned_cell.clone(),
yield_time_ms: 1,
max_output_tokens: None,
},
/*preempt*/ None,
)
@@ -751,6 +760,7 @@ async fn concurrent_wait_rejects_without_displacing_the_active_observer() -> Res
WaitRequest {
cell_id: running_cell.clone(),
yield_time_ms: 100,
max_output_tokens: None,
},
)
.await?;
@@ -762,6 +772,7 @@ async fn concurrent_wait_rejects_without_displacing_the_active_observer() -> Res
WaitRequest {
cell_id: running_cell.clone(),
yield_time_ms: 60_000,
max_output_tokens: None,
},
/*preempt*/ None,
),
@@ -785,7 +796,7 @@ async fn concurrent_wait_rejects_without_displacing_the_active_observer() -> Res
})
);
let actual = session
.terminate(running_cell.clone())
.terminate(running_cell.clone().into())
.await
.map_err(anyhow::Error::msg)?;
assert_eq!(
@@ -838,6 +849,7 @@ async fn dropping_a_wait_retires_its_observer_before_the_next_wait() -> Result<(
WaitRequest {
cell_id: running_cell.clone(),
yield_time_ms: 60_000,
max_output_tokens: None,
},
)
.await?;
@@ -850,6 +862,7 @@ async fn dropping_a_wait_retires_its_observer_before_the_next_wait() -> Result<(
WaitRequest {
cell_id: running_cell.clone(),
yield_time_ms: 1,
max_output_tokens: None,
},
/*preempt*/ None,
),
@@ -866,7 +879,7 @@ async fn dropping_a_wait_retires_its_observer_before_the_next_wait() -> Result<(
})
);
let actual = session
.terminate(running_cell.clone())
.terminate(running_cell.clone().into())
.await
.map_err(anyhow::Error::msg)?;
assert_eq!(
@@ -1091,6 +1104,7 @@ async fn sessions_enforce_independent_yield_limits() -> Result<()> {
WaitRequest {
cell_id: cell_id("1"),
yield_time_ms: 60_000,
max_output_tokens: None,
},
/*preempt*/ None,
),
@@ -1117,7 +1131,7 @@ async fn sessions_enforce_independent_yield_limits() -> Result<()> {
text_response("1", "isolated", actual.code_mode_host_duration())
);
let actual = limited
.terminate(cell_id("1"))
.terminate(cell_id("1").into())
.await
.map_err(anyhow::Error::msg)?;
assert_eq!(
@@ -1156,6 +1170,7 @@ async fn dropping_a_grpc_lease_retires_its_server_session() -> Result<()> {
let mut lease = client
.open_session(grpc::OpenSessionRequest {
cell_execution_limits: None,
output_truncation: None,
})
.await
.context("open raw gRPC session")?
@@ -1236,6 +1251,7 @@ async fn cached_session_recovers_after_a_remote_host_restarts() -> Result<()> {
WaitRequest {
cell_id: old_cell_id.clone(),
yield_time_ms: 60_000,
max_output_tokens: None,
},
)
.await?;
@@ -1280,7 +1296,7 @@ async fn cached_session_recovers_after_a_remote_host_restarts() -> Result<()> {
.context("host loss did not retire the original generation's cell")?;
let _replacement = HostHarness::start(&listen_url).await?;
assert!(session.terminate(old_cell_id.clone()).await.is_err());
assert!(session.terminate(old_cell_id.clone().into()).await.is_err());
let mut rejected = request(r#"await tools.echo({must_not_run: true});"#);
rejected.enabled_tools = vec![tool("echo")];
@@ -1291,6 +1307,7 @@ async fn cached_session_recovers_after_a_remote_host_restarts() -> Result<()> {
WaitRequest {
cell_id: old_cell_id.clone(),
yield_time_ms: 1,
max_output_tokens: None,
},
/*preempt*/ None,
);
@@ -1406,6 +1423,7 @@ async fn cached_session_recovers_after_a_remote_host_restarts() -> Result<()> {
WaitRequest {
cell_id: replacement_cell_id.clone(),
yield_time_ms: 1,
max_output_tokens: None,
},
/*preempt*/ None,
)
@@ -1420,7 +1438,7 @@ async fn cached_session_recovers_after_a_remote_host_restarts() -> Result<()> {
})
);
let actual = session
.terminate(replacement_cell_id.clone())
.terminate(replacement_cell_id.clone().into())
.await
.map_err(anyhow::Error::msg)?;
assert_eq!(
@@ -1437,13 +1455,14 @@ async fn cached_session_recovers_after_a_remote_host_restarts() -> Result<()> {
WaitRequest {
cell_id: old_cell_id.clone(),
yield_time_ms: 1,
max_output_tokens: None,
},
/*preempt*/ None,
)
.await
.unwrap_err();
assert!(stale_wait.contains("stale code-mode host generation"));
let stale_termination = session.terminate(old_cell_id).await.unwrap_err();
let stale_termination = session.terminate(old_cell_id.into()).await.unwrap_err();
assert!(stale_termination.contains("stale code-mode host generation"));
session.shutdown().await.map_err(anyhow::Error::msg)?;
Ok(())
@@ -228,6 +228,7 @@ async fn routing_metadata_is_scoped_to_each_session_on_a_shared_provider() -> Re
WaitRequest {
cell_id: id.clone(),
yield_time_ms: 1,
max_output_tokens: None,
},
/*preempt*/ None,
)
@@ -237,7 +238,10 @@ async fn routing_metadata_is_scoped_to_each_session_on_a_shared_provider() -> Re
outcome,
WaitOutcome::LiveCell(RuntimeResponse::Yielded { .. })
));
session.terminate(id).await.map_err(anyhow::Error::msg)?;
session
.terminate(id.into())
.await
.map_err(anyhow::Error::msg)?;
}
for session in [&parent, &child] {
execute(
@@ -57,7 +57,7 @@ try { image(value); } catch { text(calls); }
.await
.context("getter did not signal startup")?;
let result = timeout(Duration::from_secs(1), session.terminate(id.clone()))
let result = timeout(Duration::from_secs(1), session.terminate(id.clone().into()))
.await
.context("workload did not stop within the cancellation deadline")?
.map_err(anyhow::Error::msg)?;
@@ -225,6 +225,7 @@ async fn completed_waits_drain_pending_notifications_before_returning() -> Resul
WaitRequest {
cell_id: cell_id("1"),
yield_time_ms: 5_000,
max_output_tokens: None,
},
/*preempt*/ None,
)
@@ -292,7 +293,7 @@ async fn termination_cancels_pending_notifications() -> Result<()> {
}
);
let actual = session
.terminate(cell_id("1"))
.terminate(cell_id("1").into())
.await
.map_err(anyhow::Error::msg)?;
assert_eq!(
@@ -150,6 +150,7 @@ async fn tcp_listener_opens_a_grpc_session() -> Result<()> {
let response = client
.open_session(grpc::OpenSessionRequest {
cell_execution_limits: None,
output_truncation: None,
})
.await
.context("failed to open gRPC code-mode session")?;
+4 -1
View File
@@ -40,6 +40,7 @@ async fn execution_timing_includes_javascript_but_excludes_delayed_reads() -> Re
let mut events = client
.open_session(grpc::OpenSessionRequest {
cell_execution_limits: None,
output_truncation: None,
})
.await?
.into_inner();
@@ -229,6 +230,7 @@ async fn observation_timing_excludes_previous_requests_and_background_time() ->
WaitRequest {
cell_id: cell_id.clone(),
yield_time_ms: 50,
max_output_tokens: None,
},
/*preempt*/ None,
)
@@ -254,7 +256,7 @@ async fn observation_timing_excludes_previous_requests_and_background_time() ->
let observed = Instant::now();
let terminated = session
.terminate(cell_id.clone())
.terminate(cell_id.clone().into())
.await
.map_err(anyhow::Error::msg)?;
let duration = terminated
@@ -276,6 +278,7 @@ async fn observation_timing_excludes_previous_requests_and_background_time() ->
WaitRequest {
cell_id: cell_id.clone(),
yield_time_ms: 50,
max_output_tokens: None,
},
/*preempt*/ None,
)
+2
View File
@@ -109,6 +109,7 @@ async fn assert_partial_results(provider: &dyn CodeModeSessionProvider) {
WaitRequest {
cell_id: cell_id.clone(),
yield_time_ms: 10,
max_output_tokens: None,
},
/*preempt*/ None,
)
@@ -133,6 +134,7 @@ async fn assert_partial_results(provider: &dyn CodeModeSessionProvider) {
WaitRequest {
cell_id: cell_id.clone(),
yield_time_ms: 5_000,
max_output_tokens: None,
},
/*preempt*/ None,
)
+14 -3
View File
@@ -206,6 +206,7 @@ async fn execute_to_terminal(
WaitRequest {
cell_id,
yield_time_ms: 60_000,
max_output_tokens: None,
},
/*preempt*/ None,
)
@@ -270,6 +271,7 @@ async fn interrupt_yields_observations_without_stopping_the_cell() {
let request = WaitRequest {
cell_id: cell_id.clone(),
yield_time_ms: 60_000,
max_output_tokens: None,
};
let response =
tokio::time::timeout(Duration::from_secs(5), session.wait(request, Some(signal)))
@@ -288,6 +290,7 @@ async fn interrupt_yields_observations_without_stopping_the_cell() {
WaitRequest {
cell_id,
yield_time_ms: 60_000,
max_output_tokens: None,
},
/*preempt*/ None,
),
@@ -350,6 +353,7 @@ async fn session_execution_limits_are_isolated_on_a_shared_process_host() {
WaitRequest {
cell_id: cell_id("1"),
yield_time_ms: 60_000,
max_output_tokens: None,
},
/*preempt*/ None,
),
@@ -367,7 +371,7 @@ async fn session_execution_limits_are_isolated_on_a_shared_process_host() {
);
limited
.terminate(cell_id("1"))
.terminate(cell_id("1").into())
.await
.expect("terminate cell");
limited.shutdown().await.expect("shutdown limited session");
@@ -468,6 +472,7 @@ text(result.value);
WaitRequest {
cell_id: cell_id("3"),
yield_time_ms: 1,
max_output_tokens: None,
},
/*preempt*/ None,
)
@@ -482,7 +487,7 @@ text(result.value);
})
);
let actual = session
.terminate(cell_id("3"))
.terminate(cell_id("3").into())
.await
.expect("terminate cell");
assert_eq!(
@@ -546,6 +551,7 @@ async fn dropping_long_wait_releases_observer_before_next_wait() {
WaitRequest {
cell_id: wait_cell_id,
yield_time_ms: 60_000,
max_output_tokens: None,
},
/*preempt*/ None,
)
@@ -561,6 +567,7 @@ async fn dropping_long_wait_releases_observer_before_next_wait() {
WaitRequest {
cell_id: running_cell_id.clone(),
yield_time_ms: 1,
max_output_tokens: None,
},
/*preempt*/ None,
),
@@ -577,7 +584,7 @@ async fn dropping_long_wait_releases_observer_before_next_wait() {
})
);
session
.terminate(running_cell_id)
.terminate(running_cell_id.into())
.await
.expect("terminate cell");
session.shutdown().await.expect("shutdown remote session");
@@ -652,6 +659,7 @@ return;
WaitRequest {
cell_id: wait_cell_id,
yield_time_ms: 60_000,
max_output_tokens: None,
},
/*preempt*/ None,
)
@@ -995,6 +1003,7 @@ async fn child_process_loss_cleans_up_and_rebuilds_the_shared_host() {
WaitRequest {
cell_id: wait_a_cell,
yield_time_ms: 60_000,
max_output_tokens: None,
},
/*preempt*/ None,
)
@@ -1008,6 +1017,7 @@ async fn child_process_loss_cleans_up_and_rebuilds_the_shared_host() {
WaitRequest {
cell_id: wait_b_cell,
yield_time_ms: 60_000,
max_output_tokens: None,
},
/*preempt*/ None,
)
@@ -1092,6 +1102,7 @@ async fn child_process_loss_cleans_up_and_rebuilds_the_shared_host() {
WaitRequest {
cell_id: cell_b.clone(),
yield_time_ms: 1,
max_output_tokens: None,
},
/*preempt*/ None,
)