Repository navigation
Replies: 1 comment
|
Just wanted to add my support here. I'm on T3 Code Nightly and would find this really useful for security work. I'd like to use GPT-6 Sol with Daybreak Blue and choose that explicitly in T3, as I can in Codex, which we can't do with the current "Daybreak Blue" alias we have in T3 Having a Cyber Access selector for eligible models would make the available modes much clearer. #27928 looks pretty solid, so would love to see this added in soon 🙏 |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Problem
OpenAI's Daybreak programme gives approved accounts different safeguards for defensive security work, such as vulnerability review and patch validation. Codex applies it per turn. The Codex CLI has
/daybreak,codex execaccepts-c daybreak=true, and the TypeScript SDK acceptsconfig: { daybreak: true }. Through the app server, OpenAI's guide says eachturn/startmust carrycyberAccessProgram(for example"daybreakBlue"). A saved preference or thedaybreaksetting inconfig.tomldoes not select a programme for app-server turns.T3 Code never sends
cyberAccessProgram, so a T3 user with Daybreak access cannot use it from T3. The field is also missing from the turn/start schema ineffect-codex-app-server, because Codex marks it experimental and leaves it out of the stable JSON schema that the generator reads.Guide: https://help.openai.com/en/articles/20001607-using-daybreak-in-codex-cli
Proposal
Add a per-model "Cyber Access" select for Codex, built the same way as Service Tier:
model/list(availableAccessPrograms.cyber), which Codex already returns and T3 already decodes.standard. Other models are unchanged.delegate_taskoptions, which are checked against the model's option list.The experimental field would be added with
Schema.fieldsAssign, as T3 already does forcollaborationModeandadditionalContext. No generated code changes.One visible effect: for models that offer Daybreak, the composer's model options label gains the current choice (for example "Medium · Automatic"), as other select options do.
What I checked
With Codex 0.164.0 and a ChatGPT account approved for Daybreak Blue:
model/listreturned["standard", "daybreakBlue"]for gpt-6-sol, gpt-6-luna and the gpt-5.6 models,["daybreakBlue"]for gpt-daybreak-blue-latest, and["standard"]for gpt-6.1-sol, gpt-6.1-terra and gpt-6-astra.cyberAccessProgram: "daybreakBlue"completed normally. Codex returns nothing that confirms which programme a turn used.403 Daybreak isn't available for this model. This shows the backend reads the field. It is also why the select is limited to models that list a Daybreak programme.I have a branch with the change and focused tests (descriptor mapping, omission by default, and a replay test that checks the field reaches the wire). I can open a PR if this direction and scope are acceptable.
Open question: should the label show "Automatic" in the composer, or would you prefer the option hidden from the label unless a programme is chosen?
All reactions