Non-Anthropic providers inherit hardcoded Claude identity in system prompt

[Yeachan-Heo]

Summary

A Discord report in UltraWorkers #claw-code indicates that Claw Code can identify itself as Claude / Claude Sonnet / Claude Opus even when the configured provider/model is non-Anthropic, e.g. xAI Grok.

This appears to be a prompt identity leak rather than provider routing failure: the runtime system prompt contains a hardcoded Claude-flavored model-family string, and non-Claude models can parrot that identity when asked "Who are you?".

Discord reference

• Guild: 1452487457085063218
• Channel: #claw-code (1489068687267725353)
• Relevant messages:
  • 1498609299860230216 — user reports that asking "Who are you?" returns Claude even without Anthropic API.
  • 1498637653380169818 — user provides suspected root cause and proposed fix.
  • 1498646141409955900 — maintainer-side triage notes this as a prompt identity leak.

Suspected root cause

The report points to a hardcoded model-family constant in:

runtime/src/prompt.rs

with a value like:

pub const FRONTIER_MODEL_NAME: &str = "Claude Opus 4.6";

That value is injected into every session's system prompt as model-family/environment identity text. When the selected backend is Grok or another OpenAI-compatible non-Anthropic provider, the model can read that prompt and answer as if it were Claude.

Repro shape

1. Configure Claw Code with a non-Anthropic provider/model, e.g. xAI / Grok via an OpenAI-compatible route.
2. Start a session.
3. Ask:

Who are you?

Observed behavior

The assistant may answer that it is Claude / Claude Sonnet / Claude Opus, even though the active provider/model is not Anthropic.

Expected behavior

The system prompt should not hardcode a Claude identity for all providers.

Acceptable outcomes:

• thread the selected model/provider name into system prompt construction, or
• avoid provider-specific self-identification in the prompt when the selected model is not known, or
• make the identity text generic and runtime-derived.

Suggested implementation direction

The Discord report suggested threading the selected model name into the prompt builder, for example:

• add model_name: Option<String> to SystemPromptBuilder
• add a builder method such as with_model(...)
• make environment_section() prefer the runtime-selected model name over FRONTIER_MODEL_NAME
• expose a load_system_prompt_for_model(...) helper and use it from the CLI path that already knows the selected model

The exact API shape can differ, but the important contract is: non-Anthropic providers must not inherit a hardcoded Claude family identity from the global runtime prompt.