Add safety check notification and error handling by etraut-openai · Pull Request #19055 · openai/codex

etraut-openai · 2026-04-22T23:27:47Z

Adds a new app-server notification that fires when a user account has been flagged for potential safety reasons.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3ef5031e9c

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".

pakrym-oai · 2026-04-23T00:24:18Z

    #[serde(rename = "type")]
    pub(crate) kind: String,
+    #[serde(rename = "openai_verification_recommendation")]
+    openai_model_verification: Option<Value>,


considering we fire an event already, why do we also need an extra field per ResponsesStreamEvent?

Yep, good point. I removed the field from ResponsesStreamEvent.

pakrym-oai · 2026-04-23T00:25:30Z

+    message: Option<String>,
+}
+
+fn cyber_policy_message_from_body(body: &str) -> Option<String> {


Don't we have a helper for generic error shape parsing somwehere?

I couldn't find any that would work for this error shape. Let me know if you think I missed something.

yeah, we parse inline for other errors in this file.

pakrym-oai · 2026-04-23T00:30:52Z

+        )
+        .await;
+
+        if let Some(warning_message) = warning_message {


Why do we emit both Warning and ModelVerification? We should pick one

[codex] The two events serve different consumers: ModelVerification is the typed, app-server-visible signal that GUI clients can localize or render however they want, while Warning is the existing user-visible path the TUI renders today. The warning is intentionally not recorded into model-visible history, so it is only a client/UI notice. If you prefer, we can make the TUI render from the structured event and drop the paired warning, but that would mean adding client-side rendering behavior instead of reusing the existing warning path.

TUI/app can render whatever message it needs based on the event

Yep. I changed this so core only emits ModelVerification and then changed the TUI to handle it.

This dual-emit pattern exists all over the place currently. We should be able to remove a bunch of this redundancy once I do a few more turns of the crank on cleaning up the TUI.

pakrym-oai · 2026-04-23T00:33:49Z

            &explicitly_enabled_connectors,
            skills_outcome,
            &mut server_model_warning_emitted_for_turn,
+            &mut model_verification_emitted_for_turn,


is there something we can do to avoid carrying this state through the agent loop only to deduplicate warnings?

Yep. This was based on the way that the existing server_model_warning_emitted_for_turn was handled, but I agree that there are cleaner ways to handle this. I've removed the new model_verification_emitted_for_turn flag and the existing flag.

…ation

pakrym-oai · 2026-04-23T04:20:10Z

+                        continue;
+                    }
+                };
+                let model_verifications = model_verifications_from_event_value(&event_value);


should we add metadata to ResponsesStreamEvent and conition on event.kind() == "response.metadata" ?

pakrym-oai · 2026-04-23T04:20:36Z

        .map(ToString::to_string);
+    let model_verifications = stream_response
+        .headers
+        .get(OPENAI_MODEL_VERIFICATION_HEADER)


I don't think we care about the header, we always stream.

Yep, good point. That simplifies things. Done.

pakrym-oai · 2026-04-23T04:21:30Z

+    })
+}
+
+pub(crate) fn model_verifications_from_event_value(


this method is very permissive

Good point. Tightened it up.

pakrym-oai · 2026-04-23T04:22:01Z

+    }
+}
+
+fn json_value_as_model_verifications(value: &Value) -> Vec<ModelVerification> {


in general this parsing can be simplified I think considering we know the exact format.

pakrym-oai · 2026-04-23T04:22:33Z

        trace!("SSE event: {}", &sse.data);

-        let event: ResponsesStreamEvent = match serde_json::from_str(&sse.data) {
+        let event_value: Value = match serde_json::from_str(&sse.data) {


probably should reuse ResponsesStreamEvent below.

github-actions · 2026-04-23T04:28:16Z

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

chatgpt-codex-connector

💡 Codex Review

codex/codex-rs/codex-api/src/sse/responses.rs

Lines 71 to 79 in 5b565f9

    
           let server_model = stream_response 
        
               .headers 
        
               .get(OPENAI_MODEL_HEADER) 
        
               .and_then(|v| v.to_str().ok()) 
        
               .map(ToString::to_string); 
        
           let reasoning_included = stream_response 
        
               .headers 
        
               .get(X_REASONING_INCLUDED_HEADER) 
        
               .is_some();

Parse verification headers in spawn_response_stream

spawn_response_stream extracts OpenAI-Model and other response headers but never reads OpenAI-Verification-Recommendation. On HTTP/SSE transport (including websocket fallback), that header can be the only verification signal, so ResponseEvent::ModelVerifications is not emitted and downstream model/verification notifications are silently skipped.

codex/codex-rs/codex-api/src/endpoint/responses_websocket.rs

Lines 389 to 399 in 5b565f9

    
           let reasoning_included = response.headers().contains_key(X_REASONING_INCLUDED_HEADER); 
        
           let models_etag = response 
        
               .headers() 
        
               .get(X_MODELS_ETAG_HEADER) 
        
               .and_then(|value| value.to_str().ok()) 
        
               .map(ToString::to_string); 
        
           let server_model = response 
        
               .headers() 
        
               .get(OPENAI_MODEL_HEADER) 
        
               .and_then(|value| value.to_str().ok()) 
        
               .map(ToString::to_string);

Emit verification from websocket upgrade headers

connect_websocket() caches X-Reasoning-Included, X-Models-Etag, and OpenAI-Model from the handshake response, but never reads OpenAI-Verification-Recommendation. stream_request() only emits cached handshake-derived signals, so if verification is supplied only via upgrade headers, ResponseEvent::ModelVerifications is never produced.

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".

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5873bd93c0

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".

chatgpt-codex-connector · 2026-04-23T05:07:22Z

+    pub(crate) fn model_verifications(&self) -> Option<Vec<ModelVerification>> {
+        if self.kind() != "response.metadata" {
+            return None;
+        }
+
+        self.metadata
+            .as_ref()
+            .and_then(|metadata| metadata.get("openai_verification_recommendation"))
+            .and_then(model_verifications_from_json_value)


Parse verification recommendations from header metadata

model_verifications() only reads metadata.openai_verification_recommendation and requires type == "response.metadata". It never inspects response.headers/top-level headers (unlike response_model()), so header-delivered verification signals are dropped and ResponseEvent::ModelVerifications is never emitted. Fresh evidence: spawn_response_stream_ignores_model_verification_header codifies this loss path.

Useful? React with 👍 / 👎.

This is intentional for streaming Responses traffic. The agreed carrier for HTTP streaming and websocket is the response.metadata event; OpenAI-Verification-Recommendation headers are not used on these paths. The test is codifying that streaming headers are ignored so we don’t accidentally support the wrong carrier.

Add model verification notification

3ef5031

etraut-openai requested a review from a team as a code owner April 22, 2026 23:27

chatgpt-codex-connector Bot reviewed Apr 22, 2026

View reviewed changes

Comment thread codex-rs/codex-api/src/sse/responses.rs Outdated

etraut-openai added 2 commits April 22, 2026 17:02

Add typed cyber policy error handling

9be58fc

codex: address PR review feedback (#19055)

f1b7dda

etraut-openai changed the title ~~Add model verification notification to app-server~~ Add safety check notification and error handling Apr 23, 2026

pakrym-oai reviewed Apr 23, 2026

View reviewed changes

codex: address PR review feedback (#19055)

581db45

pakrym-oai reviewed Apr 23, 2026

View reviewed changes

Comment thread codex-rs/codex-api/src/api_bridge.rs

pakrym-oai reviewed Apr 23, 2026

View reviewed changes

codex: address PR review feedback (#19055)

2d783ee

pakrym-oai reviewed Apr 23, 2026

View reviewed changes

codex: remove redundant review comment (#19055)

78b2b89

This comment was marked as resolved.

Sign in to view

etraut-openai added 6 commits April 22, 2026 17:56

codex: parse verification metadata before stream events (#19055)

6144dab

codex: render model verification in clients (#19055)

fadfa2e

codex: keep turn dedupe state in context (#19055)

ea4c801

Merge remote-tracking branch 'origin/main' into etraut/safety-notific…

3d70d03

…ation

Fix model verification warning test assertions

b4502bb

Support response metadata verification events

e2b9edd

pakrym-oai reviewed Apr 23, 2026

View reviewed changes

Ignore streaming verification headers

dff1fed

pakrym-oai approved these changes Apr 23, 2026

View reviewed changes

marcmontanari added a commit that referenced this pull request Apr 23, 2026

codex: address PR review feedback (#19055)

ff44be9

marcmontanari added a commit that referenced this pull request Apr 23, 2026

codex: address PR review feedback (#19055)

1103252

etraut-openai added 4 commits April 22, 2026 21:29

Narrow verification metadata parsing

b0e5a46

Read verification metadata from stream events

5fe8acf

Require array verification metadata

b8d1618

Inline cyber policy error parsing

5b565f9

etraut-openai force-pushed the etraut/safety-notification branch from 1103252 to 5b565f9 Compare April 23, 2026 04:38

chatgpt-codex-connector Bot reviewed Apr 23, 2026

View reviewed changes

Comment thread codex-rs/codex-api/src/sse/responses.rs

etraut-openai added 2 commits April 22, 2026 21:56

Use metadata event in verification tests

3275dfa

Merge branch 'main' into etraut/safety-notification

5873bd9

chatgpt-codex-connector Bot reviewed Apr 23, 2026

View reviewed changes

etraut-openai merged commit bbff4ee into main Apr 23, 2026
25 checks passed

etraut-openai deleted the etraut/safety-notification branch April 23, 2026 05:24

github-actions Bot locked and limited conversation to collaborators Apr 23, 2026

	let server_model = stream_response
	.headers
	.get(OPENAI_MODEL_HEADER)
	.and_then(\|v\| v.to_str().ok())
	.map(ToString::to_string);
	let reasoning_included = stream_response
	.headers
	.get(X_REASONING_INCLUDED_HEADER)
	.is_some();

	let reasoning_included = response.headers().contains_key(X_REASONING_INCLUDED_HEADER);
	let models_etag = response
	.headers()
	.get(X_MODELS_ETAG_HEADER)
	.and_then(\|value\| value.to_str().ok())
	.map(ToString::to_string);
	let server_model = response
	.headers()
	.get(OPENAI_MODEL_HEADER)
	.and_then(\|value\| value.to_str().ok())
	.map(ToString::to_string);

Conversation

etraut-openai commented Apr 22, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

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!

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!

This comment was marked as resolved.

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!

Choose a reason for hiding this comment

Uh oh!

github-actions Bot commented Apr 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Apr 23, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

github-actions Bot commented Apr 23, 2026 •

edited

Loading