fix(local): wire OpenAI-compatible provider into local backend

[just-cameron]

Summary

• add a first-class local openai-compatible provider to the CLI, TUI, and provider selector
• discover arbitrary model IDs from the configured /v1/models endpoint and use the same published model for turns
• support keyed and keyless endpoints without sending the internal no-key sentinel as authorization

Before and after

Before, letta --backend local connect openai-compatible returned Unknown provider, so users had to configure arbitrary endpoints as LM Studio or llama.cpp.

After, the local provider has its own identity and stable openai-compatible/<model-id> handles. The base URL is required. The API key is optional.

Scope and risk

• local backend only; the existing cloud provider flow is unchanged
• model capabilities remain text-only and non-reasoning because the standard model-list response does not report them
• the provider keeps the last successful model list when a refresh fails
• unconfigured custom endpoints are not probed
• this adds one configured OpenAI-compatible connection, not multiple named profiles

Validation

• 93 focused provider, CLI, storage, selector, discovery, and turn tests
• command-level smoke using the source CLI and a mock OpenAI-compatible server
• keyed and keyless discovery and Chat Completions requests
• bun run check
• listener tests that failed in the full parallel unit run passed when rerun in isolation
• full integration suite could not create its required test agents in the current test environment; CI will run it independently

Overlord (agent-c2adbf5c-8419-4211-8cd8-3740db164974)

👾 Generated with Letta Code

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.