Add Guardian context cost and request token telemetry (#44164)

## What changed

- Record per-section text bytes, estimated text tokens, image bytes, and image counts for synchronous reviews and asynchronous scoring without logging evidence payloads.
- Emit estimated request tokens through `codex.guardian.context.request_tokens`. Synchronous estimates include assembled history, instructions, tool definitions, and output format, and measure the full logical request before WebSocket delta generation. Asynchronous estimates cover the assembled input.
- Add shared context budgeting helpers, including conservative image token reservations independent of encoded payload size and model-aware input limit calculation.

## Testing

Add coverage for separate text and image accounting, image estimates independent of encoded size, and section estimates that bound delivered messages. Extend asynchronous scorer and app-server tests to verify cost metric emission.

GitOrigin-RevId: aed45ecd9c23706f88caa51f2a4f2c77872d3bdb
This commit is contained in:
felixxia-oai authored and copyberry committed 2026-09-09 14:48:08 +00:00
1 parent 73a1148c9c
commit d3ffbbed5a
13 files changed
+472 -9

No files matched your search

+2
View File
@@ -9,6 +9,8 @@ mod decision;
mod feedback;
mod metrics;
mod prompt;
mod request_budget;
pub(crate) use request_budget::observe as observe_guardian_request;
mod review;
mod review_session;
mod reviewer_config;
+15 -3
View File
@@ -210,9 +210,21 @@ pub(crate) async fn build_guardian_prompt_items_with_parent_turn(
if transcript_entries.is_empty() {
transcript.items.push(placeholder.to_owned());
}
let items = sections
.compose(presentation, transcript)?
.into_user_inputs()?;
let context = sections.compose(presentation, transcript)?;
for (section, cost) in context.section_costs() {
for (measurement, value) in cost.measurements() {
session.services.session_telemetry.histogram(
codex_guardian_context::SECTION_COST_METRIC,
i64::try_from(value).unwrap_or(i64::MAX),
&[
("target", "sync"),
("section", section),
("measurement", measurement),
],
);
}
}
let items = context.into_user_inputs()?;
Ok(GuardianPromptItems {
items,
transcript_cursor,
@@ -0,0 +1,33 @@
//! Measures the complete synchronous request after image admission and wire
//! prefix assembly. Includes reused history, tool definitions and output format.
use codex_guardian_context::REQUEST_TOKENS_METRIC;
use codex_protocol::protocol::TruncationPolicy;
use crate::context_manager::estimate_item_token_count;
use codex_api::ResponsesApiRequest;
use codex_otel::SessionTelemetry;
pub(crate) fn observe(telemetry: &SessionTelemetry, request: &ResponsesApiRequest) -> usize {
let input = request
.input
.iter()
.map(estimate_item_token_count)
.fold(0i64, i64::saturating_add);
let instructions = TruncationPolicy::Bytes(request.instructions.len()).token_budget();
let metadata = serde_json::to_vec(&(&request.tools, &request.text))
.map(|bytes| TruncationPolicy::Bytes(bytes.len()).token_budget())
.unwrap_or(usize::MAX);
let total = usize::try_from(input)
.unwrap_or(usize::MAX)
.saturating_add(instructions)
.saturating_add(metadata);
// The assembled input already includes inherited history and the current
// review. Do not report a guessed old/new split after context injection.
telemetry.histogram(
REQUEST_TOKENS_METRIC,
i64::try_from(total).unwrap_or(i64::MAX),
&[("target", "sync"), ("component", "total")],
);
total
}