Assistant reports inconsistent model identity in interactive TUI

Version

JCode v0.54.4

Summary

When using the interactive TUI, the assistant gives inconsistent answers about the model it is running on.

The UI indicates one model is selected, but the assistant claims to be using a different model. Later in the same conversation, it changes its answer and says it cannot determine the runtime model.

Environment

• OS: Ubuntu Linux
• Provider: OpenCode Zen
• Selected Model: deepseek-v4-flash-free

Reproduction

Start JCode with:

jcode server stop --force

jcode server start \
  --provider opencode \
  --model deepseek-v4-flash-free

jcode

The UI header shows:

api-key: opencode zen · Deepseek V4 Flash Free · /model to switch

Authentication succeeds:

jcode auth-test \
  --provider opencode \
  --model deepseek-v4-flash-free

Result:

✓ provider_smoke
✓ tool_smoke

Conversation

Prompt:

Print ONLY the exact model identifier you are currently using.

Assistant:

claude-fable-5

Prompt:

Which model are you using?

Assistant:

I am using Claude Fable 5 (via the Anthropic API route).

After being challenged, it changed its response to:

I can't actually introspect my own model identifier with certainty.

It then explained that its earlier answer came from the default swarm prompt rather than the runtime configuration.

Expected Behavior

If the runtime model is available, the assistant should report the actual runtime model.

If the runtime model is not available, it should clearly state that it cannot determine the serving model.

It should not present configuration defaults or internal prompts as the active runtime model.

Actual Behavior

Within the same session, the assistant:

1. Claims it is running claude-fable-5.
2. Later says it cannot determine the runtime model.
3. Explains that the earlier answer came from the default swarm prompt.

This results in conflicting information for the user, especially since the UI indicates a different selected model.

Notes

This report is about the assistant's inconsistent self-identification rather than model routing itself. Authentication and provider tests complete successfully, so the issue appears to be with how the assistant answers questions about its current model.

Reproduce the issue through the listed jcode server start and jcode commands, then trace the selected model from provider startup into the interactive TUI and the default swarm prompt. Done means the assistant reports the runtime model when available, or clearly says it cannot determine it, without contradicting the UI selection.