jcode run hangs forever with OAuth providers when default_provider is an API-key provider

Version: v0.56.0 (9aa8c5ad), macOS (aarch64 machine, x86_64 build under Rosetta)

Summary

With [provider].default_provider = "deepseek" (an API-key / openai-compatible provider), jcode run --provider claude or --provider openai (OAuth) hangs indefinitely — no timeout, no output, no error, only killable via signal. The interactive TUI with the same OAuth provider works fine, and run with an API-key provider works fine. It affects both OAuth providers (Claude subscription and OpenAI/ChatGPT subscription).

Root cause (observed)

In the non-interactive run path, --provider / --model are effectively ignored and provider/model resolution becomes inconsistent:

• Session journal records the impossible pair "provider_key":"deepseek","model":"anthropic/claude-sonnet-4".
• --ndjson emits only {"type":"start","provider":"OpenRouter","model":"anthropic/claude-sonnet-4"} and then nothing — it never reaches the connection phase (connecting / waiting for response are never emitted).
• No API stream attempt … log line is ever written (API-key providers do write it right before sending).
• Stack sample of the hung process: main thread parked in run_main → tokio block_on on a condvar, all worker threads idle (waiting on a future that never wakes). The only open socket is the memory/embedding agent's connection to api.anthropic.com; the actual chat request never opens a socket.

Trigger isolation

• Keep default_provider = "deepseek", change only default_model to a Claude model → still hangs.
• Set default_provider = "claude" (or "openai") with a matching default_model → run works, and after that --provider / --model flags also work for per-run switching (verified switching to openai and back to deepseek).

So the trigger is default_provider being an API-key / openai-compatible provider while requesting an OAuth provider via run.

Reproduction

1. ~/.jcode/config.toml: default_provider = "deepseek", default_model = "deepseek-v4-pro". Log in to claude and openai via OAuth (jcode login). jcode auth status shows both available.
2. jcode run --provider claude --model claude-fable-5 --tool-profile none --quiet "OK" → hangs forever. (Same with --provider openai --model gpt-5.6-sol.)
3. Set default_provider = "claude", default_model = "claude-opus-4-8" → the exact same command completes in ~1s.

Expected

jcode run --provider <X> should use provider <X> (honoring --provider / --model over the config default), or fail fast with a clear error if the provider/model is unusable — never deadlock silently.

Workaround

Set [provider].default_provider (and default_model) to an OAuth provider; flags then work for per-run switching.

Start by reproducing the hang with jcode run --provider claude while default_provider is set to deepseek, then trace the non-interactive jcode run entry point where command-line provider and model values are resolved against config.toml. Compare that path with the working TUI and API-key runs. Done means the requested provider and model are honored, or an unusable combination fails clearly without hanging.