Design Philosophy

This project uses a CLI ticket system for task management. Run tk help when you need to use it.

Tunacode

This project is tunacode, much like you! It's a TUI code agent that can be used to code and debug code or general agentic tasks.

src/tunacode/ui is the TUI interface that is used to interact with the user. src/tunacode/core is the core agent that is used to code and debug code or general agentic tasks. src/tunacode/tools is the tools that are used to code and debug code or general agentic tasks.

Tests are located in the tests/ directory covering tool decorators, tool conformance, compaction, and tool retry logic.

Design Philosophy

The TUI design is heavily inspired by the classic NeXTSTEP user interface. This choice reflects a commitment to "the next step of uniformity".

• Uniformity: The interface should provide a consistent and predictable experience across all interactions.
• User Informed: A core tenet is to keep the user constantly informed of the agent's state, actions, and reasoning. No "magic" should happen in the background without visual feedback.
• Aesthetic: The look should be professional, clean, and retro-modern, echoing the clarity and object-oriented nature of the NeXTSTEP environment.

UI Design Rule: Always call the neXTSTEP-ui skill for any UI changes.

Skill Location: .claude/skills/neXTSTEP-ui/

• SKILL.md - Design philosophy and guidelines
• NeXTSTEP_User_Interface_Guidelines_Release_3_Nov93.pdf - Original 198-page reference
• read_pdf.py - Chunked PDF reader (uv run python read_pdf.py --help)

Workflow Rules

• Never begin coding until the objective is explicitly defined. If unclear, ask questions or use best practices.
• Always use .venv and uv for package management.
• Small, focused diffs only. Commit frequently.

Code Style & Typing

• Enforce ruff check --fix . before PRs.
• Use explicit typing. cast(...) and assert ... are OK.
• # type: ignore only with strong justification.
• Mypy Status (2026-01-27): 50 errors in 17 files. Gate 2 dependency direction work complete (PR #316). Use git commit -n to bypass pre-commit hooks if blocked. Do NOT introduce new type errors.
• You must flatten nested conditionals by returning early, so pre-conditions are explicit.
• If it is never executed, remove it. You MUST make sure what we remove has been committed before in case we need to rollback.
• Normalize symmetries: you must make identical things look identical and different things look different for faster pattern-spotting.
• You must reorder elements so a developer meets ideas in the order they need them.
• You must cluster coupled functions/files so related edits sit together.
• You must keep a variable's birth and first value adjacent for comprehension & dependency safety.
• Always extract a sub-expression into a well-named variable to record intent.
• Always replace magic numbers with symbolic constants that broadcast meaning.
• Never use magic literals; symbolic constants are preferred.
• ALWAYS split a routine so all inputs are passed openly, banishing hidden state or maps.

Error Handling

• Fail fast, fail loud. No silent fallbacks. This is one of the most important rules to follow.
• Minimize branching: every if/try must be justified.

Dependencies

• Avoid new core dependencies. Tiny deps OK if widely reused.
• Run tests with: uv run pytest.

Testing

This project uses a tmux-driven system test suite for end-to-end tool verification. See tests/system/cli/test_tmux_tools.py for details.

The suite launches tunacode in a real tmux session, sends prompts, and verifies tool execution by checking the status bar shows last: <tool_name> and expected output appears in the captured pane.

Run with:

uv run pytest tests/system/cli/test_tmux_tools.py -v -m tmux

Requirements:

• tmux binary installed
• MINIMAX_API_KEY in environment or ~/.config/tunacode.json

Tested tools: bash, read_file, write_file, hashline_edit, discover, web_fetch

See docs/modules/tools/tools.md for tool documentation.

Scope & Maintenance

• No cutover shims, aliases, or dual-path compatibility layers. Do full replacement everywhere in the same change.
• Delete dead code (never guard it).
• Always run ruff ..
• Use git commit -n if pre-commit hooks block rollback.

Card Format

See docs/agent-docs/card-format.md for frontmatter spec and body sections when writing cards.

Quality Gates

See docs/agent-docs/quality-gates.md for Gates 0-6 (shims, coupling, dependency direction, contracts, docs, indirection, exception paths).

KB Directory

Maintain a .claude/ directory with:

• Do not use .claude/markdown; the seven folders below are the source of truth.

• metadata/ — dependency graphs, file classifications, error pattern database

• code_index/ — function call graphs, type relationships, interface mappings

• debug_history/ — error-solution pairs indexed by component/error type

• patterns/ — canonical implementation examples for this codebase

• cheatsheets/ — quick-reference guides per component with gotchas

• qa/ — solved problems database with reasoning

• delta/ — semantic changelogs explaining changes

Select the most semantically correct directory and create a card in it.

See .claude/debug_history/continuous-learning.md for bug/pattern/lesson entries.

We are currently in the middle of a large rewrite few test exist and documentation and that is okay. We will build the test and documentation as we go

The agent loop is https://github.com/alchemiststudios.ai/tinyAgent

pydantic-ai has been fully removed. tinyagent is the sole agent loop. pydantic (the validation library) remains. Do NOT reintroduce pydantic-ai.