[Bug]: Amazon Bedrock Codex model IDs are misclassified as legacy and break text generation defaults #9659

Before submitting

• I searched existing issues and pull requests and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/server Codex provider/model manifest and text generation, with visible effects in the web and mobile model pickers.

Summary

Codex authenticated through Amazon Bedrock reports provider-qualified model IDs such as openai.gpt-5.6-sol, but T3 compares them against unqualified IDs such as gpt-5.6-sol. Exact string matching incorrectly classifies current Bedrock models as legacy and leaves T3’s generated-text defaults pointing at model IDs that are absent from the live catalog.

Steps to reproduce

1. Configure the Codex provider using a Codex CLI authenticated through Amazon Bedrock.
2. Refresh the provider model catalog.
3. Observe that the live catalog contains:
   • openai.gpt-5.6-sol
   • openai.gpt-5.6-terra
   • openai.gpt-5.6-luna
4. Open the model picker.
5. Leave textGenerationModelSelection at its server default and start a thread or request “Regenerate title.”

Expected behavior

• Bedrock’s qualified 5.6 model IDs should match the corresponding current model families in T3’s manifest.
• Current models should not appear under Legacy.
• Preferred/default model selection should resolve to an ID present in the live provider catalog.
• Automatic thread titles and other generated text should execute with the provider-qualified model ID required by Bedrock.

Actual behavior

All of the observed Bedrock models are stored with isLegacy: true, including the three 5.6 models that are explicitly listed as current in T3’s manifest under their unqualified names.

The relevant paths currently use exact model-string comparisons:

• isLegacyModel checks currentModels.includes(slug):
  https://github.com/pingdotgg/t3code/blob/main/apps/server/src/provider/ModelManifest.ts
• applyPreferredCodexDefaultModel compares live slugs directly with PREFERRED_DEFAULT_CODEX_MODELS:
  https://github.com/pingdotgg/t3code/blob/main/apps/server/src/provider/Layers/CodexProvider.ts

T3’s default text-generation model is the unqualified gpt-5.6-luna:

https://github.com/pingdotgg/t3code/blob/main/packages/contracts/src/model.ts

CodexTextGeneration then passes modelSelection.model directly to codex exec --model:

https://github.com/pingdotgg/t3code/blob/main/apps/server/src/textGeneration/CodexTextGeneration.ts

Consequently, automatic naming/regeneration targets an ID that is not present in the Bedrock-backed live catalog unless the user explicitly selects the qualified openai.gpt-5.6-luna.

Impact

• Current Codex models are hidden or de-prioritized as Legacy.
• Mobile rejects legacy models for implicit defaults.
• T3’s preferred Codex default ranking cannot match provider-qualified IDs.
• Automatic thread titles, title regeneration, generated branch names, and other generated text can fail when using the unqualified default.

Environment

• T3 Code Alpha 0.0.38
• macOS 26.6.2, arm64
• Codex CLI 0.153.0-alpha.5
• Authentication: Amazon Bedrock
• No ChatGPT authentication

Suggested fix

Introduce a provider-aware canonical model-family comparison:

openai.gpt-5.6-sol  ->  gpt-5.6-sol
openai.gpt-5.6-luna ->  gpt-5.6-luna

Use the canonical family only for manifest classification, aliases, and preference matching. Preserve the full provider-qualified ID for dispatch.

For stored/default model selections, resolve the generic default against the live catalog before invoking the CLI, selecting the qualified catalog entry when present.

Regression tests should cover:

• Bedrock-qualified current models are not legacy.
• Preferred defaults match qualified IDs.
• Text-generation defaults resolve to a live qualified ID.
• The original qualified ID is retained in codex exec --model.

Disclosure

I used AI assistance to organize and cross-check the local T3 cache and current source. I reviewed the report, and no prompts, credentials, local paths, or thread identifiers are included.