Persist and prewarm agent tasks per thread by adrian-openai · Pull Request #17978 · openai/codex

adrian-openai · 2026-04-15T19:03:01Z

Summary

persist registered agent tasks in the session state update stream so the thread can reuse them
prewarm task registration once identity registration succeeds, while keeping startup failures best-effort
isolate the session-side task lifecycle into a dedicated module so AgentIdentityManager and RegisteredAgentTask do not leak across as many core layers

Testing

cargo test -p codex-core startup_agent_task_prewarm
cargo test -p codex-core cached_agent_task_for_current_identity_clears_stale_task
cargo test -p codex-core record_initial_history_

adrian-openai · 2026-04-16T16:43:33Z

Stack navigation for this slice:

PR1: Add use_agent_identity feature flag #17385 - Add use_agent_identity feature flag
PR2: Register agent identities behind use_agent_identity #17386 - Register agent identities behind use_agent_identity
PR3: Register agent tasks behind use_agent_identity #17387 - Register agent tasks behind use_agent_identity
PR3.1: Persist and prewarm agent tasks per thread #17978 - Persist and prewarm agent tasks per thread (this PR)
PR4: [codex] Use AgentAssertion downstream behind use_agent_identity #17980 - Use AgentAssertion downstream behind use_agent_identity
PR4.1: [codex] Use background agent task auth for backend calls #18094 - Use background agent task auth for backend calls

## Summary Stack PR3 for feature-gated agent identity support. This PR adds per-thread agent task registration behind `features.use_agent_identity`. Tasks are minted on the first real user turn and cached in thread runtime state for later turns. ## Stack - PR1: #17385 - add `features.use_agent_identity` - PR2: #17386 - register agent identities when enabled - PR3: #17387 - this PR, original task registration slice - PR3.1: #17978 - persist and prewarm registered tasks per thread - PR4: #17980 - use `AgentAssertion` downstream when enabled ## Validation Covered as part of the local stack validation pass: - `just fmt` - `cargo test -p codex-core --lib agent_identity` - `cargo test -p codex-core --lib agent_assertion` - `cargo test -p codex-core --lib websocket_agent_task` - `cargo test -p codex-api api_bridge` - `cargo build -p codex-cli --bin codex` ## Notes The full local app-server E2E path is still being debugged after PR creation. The current branch stack is directionally ready for review while that follow-up continues.

shijie-oai · 2026-04-17T14:38:52Z

+// These tests start full app-server processes; keep headroom for concurrent debug startup.
+const DEFAULT_READ_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(30);


Are we doing this only for debug? Would like to understand better what/if the code change here is causing longer run loop or something.

Good catch! This has been updated.

nicksteele-oai · 2026-04-17T16:25:20Z

@codex review

chatgpt-codex-connector · 2026-04-17T16:32:55Z

Codex Review: Didn't find any major issues. Chef's kiss.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

shijie-oai · 2026-04-17T19:32:45Z

    }

-    pub(crate) async fn task_matches_current_binding(&self, task: &RegisteredAgentTask) -> bool {
+    pub(crate) async fn task_matches_current_identity(&self, task: &RegisteredAgentTask) -> bool {


This is going to be a catch 22 situation - probably will be answered in a later PR and I will come back to comment on this. What happens if there is a mismatch on the agent runtime id between the session vs the task? Agent identity has ttl that can expire mid session and be reacquired?

Okay I am back - my concern still stands and will leave a PR level comment.

Discussed directly - generally this should not happen in the current implementation.

shijie-oai · 2026-04-17T19:46:29Z

+    fn latest_persisted_agent_task(
+        rollout_items: &[RolloutItem],
+    ) -> Option<Option<SessionAgentTask>> {
+        rollout_items.iter().rev().find_map(|item| match item {
+            RolloutItem::SessionState(update) => Some(update.agent_task.clone()),
+            _ => None,
+        })
+    }


I am not 100% sure if we want to actually persist agent task as a part of the rollout? Any downside for this to be in memory and limited the usage to the current session and on future resume we generate a new one or conceptually it is incorrect?

Discussed in person - we want the task to be persisted so that we can keep permissions the same across the entirety of the rollout, and not start from scratch every so often.

shijie-oai · 2026-04-17T19:48:04Z

-                RolloutItem::EventMsg(_) | RolloutItem::SessionMeta(_) => {}
+                RolloutItem::EventMsg(_)
+                | RolloutItem::SessionMeta(_)
+                | RolloutItem::SessionState(_) => {}


Ties to if we want to store sessionState at all.

Still think we do :)

shijie-oai · 2026-04-17T19:57:13Z


        // record_initial_history can emit events. We record only after the SessionConfiguredEvent is emitted.
        sess.record_initial_history(initial_history).await;
+        sess.start_agent_identity_registration();


DO we want to await on this?

[codex] We intentionally do not await this at startup. record_initial_history(...) needs to run first so resumed SessionState can restore/clear the persisted task before prewarm can cache a new one; after that, startup registration is best-effort prewarm. The actual user-turn path awaits ensure_agent_task_registered() in session/turn.rs, emits an error for that thread if registration fails, and retries on the next turn rather than blocking session startup or shutting down other threads.

shijie-oai

Main question is that do we need task agent stored within the rollout? I do not see the downside (and to be fair in most situation) to recreate the task identity when we resume (i.e. huge window span).

efrazer-oai · 2026-04-17T20:16:01Z

Main question is that do we need task agent stored within the rollout? I do not see the downside (and to be fair in most situation) we would need to recreate the task identity when we resume (i.e. huge window span).

I think I pushed for this originally, but I think you'd have better intuition here @shijie-oai. Core idea was to keep the task lifecycle very close to a session lifecycle (e.g. we can have multiple tasks per rollout now), but i'm g either way.

shijie-oai

I am not seeing a clear failure handling if an agent or task validation is outdated when checking against backend.

If you are agent identity only (for enterprise use case), it is okay to hard fail as we need to strictly enforce defined TTL. But if you are chatgpt auth and we assign agent identity on the fly, I wonder what is the default TTL for that agent identity is and how we should best recover. Maybe that agent identity has the same life cycle as chatgpt refresh token?

adrian-openai mentioned this pull request Apr 15, 2026

Register agent tasks behind use_agent_identity #17387

Merged

efrazer-oai reviewed Apr 15, 2026

View reviewed changes

Comment thread codex-rs/app-server/tests/suite/v2/client_metadata.rs Outdated

efrazer-oai reviewed Apr 15, 2026

View reviewed changes

Comment thread codex-rs/app-server/tests/suite/v2/client_metadata.rs Outdated

adrian-openai mentioned this pull request Apr 15, 2026

[codex] Use AgentAssertion downstream behind use_agent_identity #17980

Merged

github-actions Bot mentioned this pull request Apr 16, 2026

📊 AI CLI 工具社区动态日报 2026-04-16 gsscsd/big_model_radar#193

Open

adrian-openai mentioned this pull request Apr 16, 2026

[codex] Use background agent task auth for backend calls #18094

Merged

efrazer-oai reviewed Apr 16, 2026

View reviewed changes

Comment thread codex-rs/protocol/src/protocol.rs

adrian-openai force-pushed the dev/adrian/codex/agent-identity-register-task branch from b65f7d4 to dfc5e05 Compare April 16, 2026 18:11

Base automatically changed from dev/adrian/codex/agent-identity-register-task to main April 16, 2026 21:30

adrian-openai force-pushed the dev/adrian/codex/agent-task-state-prewarm branch 2 times, most recently from ea5ec0a to 56f4b38 Compare April 17, 2026 03:35

adrian-openai mentioned this pull request Apr 17, 2026

[codex] Use background task auth for additional backend calls #18260

Merged

adrian-openai force-pushed the dev/adrian/codex/agent-task-state-prewarm branch from 56f4b38 to 4c0010f Compare April 17, 2026 04:01

adrian-openai requested review from efrazer-oai, nicksteele-oai and shijie-oai April 17, 2026 04:02

shijie-oai reviewed Apr 17, 2026

View reviewed changes

adrian-openai force-pushed the dev/adrian/codex/agent-task-state-prewarm branch from 3dcdb8d to 67cff31 Compare April 17, 2026 17:20

shijie-oai reviewed Apr 17, 2026

View reviewed changes

Persist and prewarm agent tasks per thread

cff6870

adrian-openai added 8 commits April 17, 2026 20:02

Fix agent task prewarm resume ordering

eb5ca52

Validate restored agent task identity

51a9152

Fix agent task auth test fixture

c166d20

Clarify persisted agent task identity invariant

7a13688

Restore agent task state during resume

5f1a60b

Clarify initialize test startup timeout

25b8520

Clean up rebased session test imports

e107456

Stabilize OTEL SSE failure tests

0000662

adrian-openai force-pushed the dev/adrian/codex/agent-task-state-prewarm branch from e6f7ce3 to 0000662 Compare April 18, 2026 03:16

shijie-oai approved these changes Apr 19, 2026

View reviewed changes

adrian-openai merged commit e5b52a3 into main Apr 19, 2026
35 of 36 checks passed

adrian-openai deleted the dev/adrian/codex/agent-task-state-prewarm branch April 19, 2026 22:45

github-actions Bot locked and limited conversation to collaborators Apr 19, 2026

		// These tests start full app-server processes; keep headroom for concurrent debug startup.
		const DEFAULT_READ_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(30);

Conversation

adrian-openai commented Apr 15, 2026

Summary

Testing

Uh oh!

Uh oh!

Uh oh!

Uh oh!

adrian-openai commented Apr 16, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

nicksteele-oai commented Apr 17, 2026

Uh oh!

chatgpt-codex-connector Bot commented Apr 17, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

shijie-oai left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

efrazer-oai commented Apr 17, 2026

Uh oh!

shijie-oai left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

shijie-oai left a comment •

edited

Loading