Skip to content

CLAUDE.md

Latest commit

 

History

History
103 lines (76 loc) · 5.01 KB

CLAUDE.md

File metadata and controls

103 lines (76 loc) · 5.01 KB
Raw

dotcontext

AI context toolkit for coding agents (Claude Code, Codex, opencode, Gemini, Copilot, Cursor) — commands, decisions, skills, and feature planning for your codebase.

Decision Compliance

IMPORTANT: Before implementing any change, check .context/decisions/ for related ADRs.

If a requested change conflicts with an existing decision:

  1. Stop and inform the user which ADR(s) would be affected
  2. Ask explicitly if they want to:
    • Proceed and update the decision
    • Modify the approach to comply with existing decision
    • Cancel the change
  3. If updating a decision, create a new version:
    • Change status to Superseded by ADR-XXX
    • Create new ADR with updated decision
    • Reference the previous ADR

Stack

  • Bash 3.2+ (POSIX-compatible shell scripting)
  • Single executable distribution (no framework)
  • No database (file-based templates)
  • Dependencies: curl/wget, sed, grep, jq (optional)

Commands

CLI

# Initialize context structure in a project (harness-selective; auto-runs /setup-context when Claude is selected)
dotcontext init [--name "Project Name"] [--agents claude,codex,...] [--yes] [--no-setup]

# Update CLI + templates
dotcontext update                     # Update CLI + templates (if in project)
dotcontext update --cli               # Only update CLI
dotcontext update --templates         # Only update templates

# Validate project setup
dotcontext doctor

# Shell tab completion
dotcontext completion [bash|zsh]

# Help / version (--features and --json expose the capability handshake)
dotcontext --help
dotcontext --version [--features|--json]

Claude Code (interactive with questions)

/setup-context             # Analyze codebase and populate context
/spec-dc [feature]         # Write a behavior spec — the WHAT (step 1)
/plan-dc [spec-path]       # Turn a spec into a plan (ADR review) — the HOW (step 2)
/execute-dc [plan-path]    # Implement a plan in parallel waves — the DO (step 3)
/code-review               # Review code changes
/commit [--amend]          # Smart commit with style-aware messages
/deep-context [query]      # Structured 4-step codebase exploration
/fix-bug [description]     # Test-driven bug fixing with parallel agents
/create-pr                 # Create PR with auto-generated description
/pr-comment                # Add comment to existing PR
/release [patch|minor|major] # Create release with version bump
/add-decision              # Add ADR interactively
/add-skill                 # Add skill guide interactively
/add-command               # Create custom command interactively

Critical Rules

  1. Always ask before assuming - When there is ambiguity, multiple valid approaches, or decisions to be made, use the AskUserQuestion tool to clarify before proceeding. Never assume user intent.
  2. POSIX compatibility - Use Bash 3.2+ compatible syntax (macOS ships with 3.2). Avoid bashisms that require newer versions.
  3. Safe defaults - Never overwrite existing user content: seed files are create-only, and managed files show a diff and prompt (--yes to skip) before updating. Always prompt before destructive operations.
  4. Template downloads - Templates are fetched from GitHub main branch. Validate downloads before using.
  5. Cross-platform support - Test on macOS, Linux, and Windows/WSL. Platform-specific code must have proper detection and fallbacks.

Architecture

Single Executable Pattern

The CLI is distributed as a single dotcontext bash script. Source code lives in src/ modules (core/, commands/, setup/) and is bundled into the single executable via make build. All commands are implemented as functions, routed via a case statement. This keeps development modular while simplifying distribution.

Template-Based Init

Templates live in templates/ directory on GitHub and are downloaded during dotcontext init. init is harness-selective — only the chosen agents' files are emitted:

  • templates/AGENTS.md → project AGENTS.md (the canonical instructions, always emitted)
  • templates/CLAUDE.md / templates/GEMINI.md → thin @AGENTS.md import stubs (emitted only for Claude / Gemini)
  • templates/.context/ → project .context/ directory
  • .claude/ (commands, agents, statusline, hooks) is emitted only when Claude is selected

Multi-Agent Integration

Six harnesses are supported (Claude Code, Codex, opencode, Gemini CLI, Copilot, Cursor) via an adapter registry (src/setup/agents.sh). Workflows reach each harness in its native form: Claude .claude/commands/, opencode .opencode/command/, Copilot .github/prompts/, and a ## Workflows section in AGENTS.md for Gemini/Cursor/Codex. Skills (SKILL.md with name/description frontmatter) go to .claude/skills/ and/or .agents/skills/. See ADR-016/017/018.

Additional Context

  • Domain and architecture → .context/CONTEXT.md
  • Architectural decisions → .context/decisions/
  • Task-specific skills → .claude/skills/
  • Bug reproduction guide → .claude/skills/bug-reproduction/SKILL.md