CLI flags --provider and --model are silently ignored by the server

Bug

--provider and --model CLI flags are silently ignored when launching jcode (TUI) and jcode run. The server-side auto-detection overrides whatever the client requests, so the flags have no effect.

Expected behavior

jcode --provider zai --model glm-5-turbo

Should start a session on Z.AI / glm-5-turbo.

Actual behavior

The session starts on whatever provider the server auto-detects (e.g., Command Code / Claude Opus 5), completely ignoring --provider zai --model glm-5-turbo.

The TUI footer shows:

api-key:command code · Claude 5 Opus

instead of the requested Z.AI / glm-5-turbo.

Workaround

Use the /model slash command inside the TUI after launch:

/model zai:glm-5-turbo

This works correctly and switches to the requested provider/model.

Reproduction

# TUI mode — flags are ignored, defaults to auto-detected provider
jcode --provider zai --model glm-5-turbo

# run mode — same issue, picks a random provider instead
jcode run --provider zai --model glm-5-turbo "hello"
# → responds as "Claude Sonnet 4 via Anthropic API" instead of glm-5-turbo

# Both tested with multiple provider/model combos — always ignored

Environment

• jcode version: v0.64.3-dev (d6c7c36, dirty)
• Server: Forest v0.64.3-dev
• OS: Linux (x86_64)
• Providers configured: OpenAI (oauth), Gemini (oauth), Z.AI (api key), Command Code (api key), DeepSeek (api key)
• Auto-detected default: Command Code

Notes

• --provider zai is listed as a valid value in --provider's possible values
• glm-5-turbo is listed by jcode model list --provider zai
• The flags are accepted without error — they just have no effect
• The issue appears to be that the Forest server's provider auto-detection runs after client flags are parsed and overrides them

Start with the CLI handling for the --provider and --model flags in the TUI and run entry points, then trace how Forest server auto-detection selects a provider after those flags are parsed. Reproduce with the commands in the issue and compare against /model zai:glm-5-turbo. Done means both launch modes use the requested Z.AI provider and glm-5-turbo instead of the auto-detected model.