feat: DeepSeek V4 support with reasoning-effort control (0.4.0)

[Hmbown]

Summary

• Adds first-class support for deepseek-v4-pro and deepseek-v4-flash with 1M context windows and current DeepSeek alias handling.
• Fixes DeepSeek V4 thinking-mode requests to send top-level thinking in raw Chat Completions JSON, not SDK-only extra_body.
• Preserves prior assistant reasoning_content for all thinking-mode tool-call turns so multi-turn tool conversations do not hit DeepSeek's HTTP 400 replay error.
• Fixes the npm wrapper first-run download race and keeps the 0.4.0 release path aligned with GitHub release assets and npm Trusted Publishing.

Model & pricing

Model	Context	Cache hit in	Cache miss in	Output
deepseek-v4-pro	1M	$0.145 / 1M	$1.74 / 1M	$3.48 / 1M
deepseek-v4-flash	1M	$0.028 / 1M	$0.14 / 1M	$0.28 / 1M

Legacy deepseek-chat, deepseek-reasoner, deepseek-r1, deepseek-v3, and deepseek-v3.2 remain accepted local aliases and now budget against the current V4 1M context unless an explicit context suffix is provided.

Reasoning effort

Config field reasoning_effort in ~/.deepseek/config.toml:

• off -> thinking.type = "disabled"
• low / medium / high -> reasoning_effort: "high", thinking.type = "enabled"
• max / xhigh -> reasoning_effort: "max", thinking.type = "enabled"

Test plan

• cargo fmt --all -- --check
• cargo check --workspace --all-targets --locked
• cargo clippy --workspace --all-targets --all-features --locked -- -D warnings
• cargo test --workspace --all-features --locked
• cargo build --release --locked -p deepseek-tui-cli -p deepseek-tui
• target/release/deepseek-tui doctor
• Live DeepSeek V4 smoke: DEEPSEEK_MODEL=deepseek-v4-pro target/release/deepseek-tui --no-alt-screen --skip-onboarding -p "Reply with exactly OK."
• npm local asset mirror: npm run release:check, npm pack, temp-install npx deepseek / npx deepseek-tui smoke

[gemini-code-assist]

Code Review

This pull request updates the workspace to version 0.4.0, introducing comprehensive support for DeepSeek V4 models, including deepseek-v4-pro and deepseek-v4-flash. Key enhancements include the implementation of reasoning-effort tiers (off, high, max) with TUI integration, updated pricing calculations for context caching, and a refactored CLI that provides better command delegation to the TUI. The default model is transitioned to deepseek-v4-pro. Review feedback highlights several locations in the CLI where hardcoded legacy model IDs and test assertions need to be updated to maintain consistency with the new V4 defaults.

[gemini-code-assist]

The fallback model here is still hardcoded to deepseek-reasoner, which is now a legacy alias for deepseek-v4-flash. Since the flagship model has been updated to deepseek-v4-pro across the rest of the workspace (e.g., in crates/config/src/lib.rs and crates/tui/src/config.rs), this should be updated to maintain consistency for new logins.

Suggested change

	.unwrap_or_else(|| "deepseek-reasoner".to_string()),
	.unwrap_or_else(|| "deepseek-v4-pro".to_string()),

[gemini-code-assist]

This test assertion expects the legacy deepseek-reasoner model. It should be updated to reflect the new default model deepseek-v4-pro to match the intended behavior of the PR.

Suggested change

	Some("deepseek-reasoner")
	Some("deepseek-v4-pro")

[gemini-code-assist]

This test assertion should be updated to check for the new default model deepseek-v4-pro.

Suggested change

	assert!(saved.contains("default_text_model = \"deepseek-reasoner\""));
	assert!(saved.contains("default_text_model = \"deepseek-v4-pro\""));