Hunter Bown
b7bd02d814
Adds first-class DeepSeek V4 Pro and Flash support, updates the default model to deepseek-v4-pro, aligns legacy aliases with the current V4 1M context behavior, and fixes thinking-mode request handling. Key fixes: - Send DeepSeek's raw Chat Completions `thinking` parameter at the top level instead of SDK-only `extra_body`. - Preserve assistant `reasoning_content` for all prior thinking-mode tool-call turns so subsequent requests satisfy DeepSeek V4's replay requirement. - Fix npm wrapper concurrent first-run downloads by using per-process temporary download paths. - Add `.mailmap` so historical bot-attributed commits aggregate under Hunter Bown where mailmap is honored. Verified with the full local Rust gate, live DeepSeek V4 smoke, npm wrapper temp-install smoke, and green PR CI across Linux, macOS, and Windows.
1.8 KiB
1.8 KiB
Project Instructions
This file provides context for AI assistants working on this project.
Project Type: Rust
Commands
- Build:
cargo build - Test:
cargo test --workspace --all-features - Lint:
cargo clippy --workspace --all-targets --all-features - Format:
cargo fmt --all - Run:
cargo run -p deepseek-tui
Documentation
See README.md for project overview, docs/ARCHITECTURE.md for internals.
DeepSeek-Specific Notes
- Thinking Tokens: DeepSeek models output thinking blocks (
ContentBlock::Thinking) before final answers. The TUI streams and displays these with visual distinction. - Reasoning Models:
deepseek-v4-proanddeepseek-v4-flashare the documented V4 model IDs. Legacydeepseek-chatanddeepseek-reasonerare compatibility aliases fordeepseek-v4-flash. - Large Context Window: DeepSeek V4 models have 1M-token context windows. Use search tools to navigate efficiently.
- API: OpenAI-compatible Chat Completions (
/chat/completions) is the documented DeepSeek API path. Base URL configurable for global (api.deepseek.com) or China (api.deepseeki.com);/v1is accepted for OpenAI SDK compatibility, and/betais only needed for beta features such as strict tool mode, chat prefix completion, and FIM completion. - Thinking + Tool Calls: In V4 thinking mode, assistant messages that contain tool calls must replay their
reasoning_contentin all subsequent requests or the API returns HTTP 400.
Important Notes
- Token/cost tracking inaccuracies: Token counting and cost estimation may be inflated due to thinking token accounting bugs. Use
/compactto manage context, and treat cost estimates as approximate. - Modes: Three modes — Plan (read-only investigation), Agent (tool use with approval), YOLO (auto-approved). See
docs/MODES.mdfor details.