bootstrap（Bootstrap 初始化）

Bootstrap

Value: Communication -- establish shared understanding of environment, capabilities, and workflow before any work begins.

Purpose

Single entry point for configuring the SDLC workflow on any harness. Detects the project environment and harness capabilities, configures TDD mode (guided or automated), generates harness-appropriate instruction files, and recommends skills. Never silently installs or modifies anything.

Practices

Step 1: Detect the Environment

Gather project context silently before asking questions:

!test -f package.json && echo "js" || true
!test -f Cargo.toml && echo "rust" || true
!test -f pyproject.toml && echo "python" || true
!test -f go.mod && echo "go" || true
!test -f mix.exs && echo "elixir" || true
!git rev-parse --is-inside-work-tree 2>/dev/null && echo "git" || true
!ls skills/*/SKILL.md 2>/dev/null | sed 's|skills/||;s|/SKILL.md||' || true

Record: languages detected, git available, skills already installed. If no language files are found, note "none" and ask the user in Step 5.

Step 2: Detect Harness Capabilities

Probe for delegation primitives (see references/capability-detection.md): check for Agent tool and skill chaining (always available).

Step 3: Detect Harness Type

Identify the harness to generate the correct instruction files:

Step 4: Configure TDD Mode

Match the strategy to detected capabilities:

Chaining is always available and runs all TDD phases sequentially in a single context. Guided mode (/tdd red, /tdd green, etc.) is always available as a user-controlled alternative. Let the user override.

If Claude Code is detected and the user wants maximum enforcement, offer to install optional hook templates from skills/tdd/references/hooks/.

Step 4b: Detect Factory Mode

Check if the pipeline skill is installed (skills/pipeline/SKILL.md exists). If detected, factory pipeline mode is available and Step 5 includes an additional question.

Step 5: Ask the User

Question 1: What are you trying to do?

• "Start a new project" -- recommend Understand + Decide + Build phases
• "Add a feature or fix a bug" -- recommend Build + Ship phases
• "Set up team workflow" -- recommend all phases plus ensemble team

Question 2: How much process structure?

• "Minimal" -- recommend tdd, domain-modeling
• "Standard" -- recommend core + ship skills
• "Full" -- recommend all skills (include factory pipeline skills when pipeline is detected)

Question 3 (only when pipeline detected): What autonomy level? Conservative / Standard / Full. Generate .factory/config.yaml with the chosen level. See references/skill-recommendations.md for skill lists.

Step 6: Task Tracking Convention

Include in generated instruction files that agents should use task lists (TaskCreate, TaskList, or equivalent harness tools) to track work — even for tasks that seem simple. Simple tasks often reveal subtasks, and a visible task list keeps progress transparent and prevents steps from being forgotten.

This convention goes into the AGENTS.md "Conventions" section (see references/agents-md.md). If the task-management skill is installed, reference it; otherwise, include the basic convention directly.

Step 7: Collect Per-Skill AGENTS.md Contributions

Some installed skills ship a references/agents-md-setup.md file that specifies content to add to AGENTS.md. Scan all installed skills for this file:

!for skill in skills/*/; do
  test -f "${skill}references/agents-md-setup.md" && echo "${skill}"
done

For each skill that has one, read its agents-md-setup.md and incorporate the specified content into AGENTS.md using managed markers for that skill (<!-- BEGIN MANAGED: skill-name -->). The setup file documents what section to add or append to and provides the content template.

Size budget: AGENTS.md must stay under 32 KiB total. Per-skill contributions should be concise — a few bullet points or a short section. If the combined contributions would exceed the budget, summarize and point to the skill for details rather than including full content.

Step 8: Generate Instruction Files

Generate harness-appropriate files. For Claude Code, CLAUDE.md uses @AGENTS.md references instead of symlinks or content embedding to keep a single source of truth and avoid duplication. See references/agents-md.md for AGENTS.md best practices (small routing document, progressive disclosure, managed markers) and references/harness-files.md for harness-specific generation rules.

Step 9: Optional Ensemble Team

If the user selected "Set up team workflow" or "Full", offer the ensemble-team skill with its three presets (solo-plus, lean, full).

Step 10: Commit and Display

Stage generated files, commit with a descriptive message, and display:

• What was configured (harness, TDD mode, skills recommended)
• Next steps (/tdd to start a TDD cycle, or phase-specific commands)
• If ensemble team was configured, note the preset and member count

Enforcement Note

Advisory in all modes. Bootstrap detects and recommends; it never silently modifies. No skill or setting is installed without explicit user confirmation [H].

Constraints

• "Never silently installs" [H]: "Override" means the user explicitly says "yes, install X" — not inferred permission from context.
• AGENTS.md 32 KiB budget: Hard cap, not target. Summarize to fit.

Verification

• Environment was detected before asking questions
• Harness capabilities were probed (not assumed)
• No more than 2-3 questions were asked
• TDD strategy matched detected capabilities (see Step 4 table)
• Generated files use managed markers for safe re-runs
• Task tracking convention is included in generated AGENTS.md
• Per-skill AGENTS.md contributions were collected and included
• Nothing was installed without user confirmation

Dependencies

This skill works standalone. It recommends but does not require other skills. It generates configuration that references the tdd skill for TDD workflow.