Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- Theo reports that his Codex agent cannot learn its specific current model from the system prompt. This user report does not establish a uniform implementation across all clients.
- After asking about the need, Tibo heard examples involving model-aware Skills, ultrafast workflow guidance and PR attribution, then promised a fix at 06:50 UTC on October 15.
- UI model selection, context supplied to the agent and its generated self-description are different forms of evidence. A self-reported name alone should not determine tool permissions or attribution.
- There is no evidence of a released fix in this conversation. Use trusted host configuration and records where available, with an explicit fallback when model identity is unknown.
What Tibo has committed to so far
At 23:12 UTC on October 14, Theo reported that the Codex agent he was using could obtain only a generic GPT-6 identity from its system prompt, rather than its specific current model. At 06:27 UTC on October 15, Tibo pointed to an apparent model indication at the bottom of the screenshot and asked when the agent itself needed that information.
After Theo supplied concrete use cases, Tibo agreed to fix the issue at 06:50 UTC, or 14:50 in Beijing. The public exchange establishes that commitment, without a released fix, version, date or client scope. It does not show that an individual session has already changed.
Separate the UI selection from the context an agent can read
An interface can show the user a selected model without necessarily supplying that same information in the agent’s readable context. Theo’s complaint concerns the latter. Tibo’s screenshot response does not by itself establish that the agent already receives the specific name, and one user report cannot establish the internal design of every Codex client.
The agent’s generated self-description is another source again. It may lack enough context to answer, so a model name in a response proves neither the actual execution identity nor a backend switch. Keep the UI selection, execution information supplied by the host and the model’s own answer distinct when making a record.
The use cases: Skills, speed trade-offs and PR attribution
Theo wants Skills to choose different behavior according to the model. He also gives an example of guidance to avoid slow Git operations in ultrafast mode. These are requested use cases; the exchange announces no model-routing feature and provides no command timings or cost changes.
His other example is model attribution in GitHub pull requests. Theo explicitly says that this currently works in his setup because he injects the information on the T3 Code side. That describes his integration, without establishing native automatic attribution in Codex or the same execution information being available to every host.
Make the basis for each decision explicit
First, inspect the model selection, configuration and execution records your current client or host actually provides, and retain the available names with the relevant session. Keep a user’s selection separate from an actual execution identifier. If the host does not provide the latter, leave it unknown rather than presenting the selected name as a verified backend result.
Second, state the capabilities a Skill needs and the operations it may perform. For example, whether a lengthy repository scan is allowed or needs confirmation can be an explicit task constraint. Define a conservative fallback when model identity is unavailable instead of requiring the agent to guess its model.
Third, identify the trusted record behind any model attribution in a PR. Label a configuration choice as a configuration choice; use actual execution evidence when you have it. Neither a generated model name, coding style nor perceived speed proves model identity. These are recommendations for records and task constraints, rather than product functions announced in this exchange.
A hypothetical fallback when the model is unknown
Suppose a team has two code-review strategies and uses a model-specific strategy only when trusted execution information from its host identifies the corresponding model. When that information is missing, it uses a general review that does not depend on a model name and retains a human confirmation step. This is a fallback example, not an executed Skill or evidence of an available autodetection interface.
If a formal fix is announced, check which clients it covers, what information it supplies to the agent and whether it applies to existing sessions. This conversation provides none of those answers. It also makes no connection between the identity fix and Day10 resets, usage allowances or account entitlements.
Sources
Frequently Asked Questions
If the UI and the agent give different model names, has the backend switched models?
The mismatch alone cannot establish that. The UI selection and the context available to the agent may have different sources. Check the host’s existing configuration and execution records first; without actual execution evidence, retain the mismatch as unresolved rather than concluding that a backend switch occurred.
Can I put a model name in the prompt and use it as proof for PR attribution?
That does not prove the execution identity. Text you supply can express task requirements, but it is not an execution record. If attribution rests only on a configuration choice, label it accordingly. Without a trusted source, there is no need to guess a specific model.
Can an organization enable this fix for everyone now?
The exchange announces no released fix, organization toggle, activation procedure or administrator policy. Teams can standardize their recording and attribution rules now, then assess client coverage when actual release information appears.