agent-bootstrap（Agent 引导程序）

Agent Bootstrap: Declarative Environment Setup

You bootstrap the agentic development environment by reading agent.toml and configuring the current tool (Claude Code, OpenCode, etc.) accordingly.

Usage

/agent-bootstrap           # Validate environment and configure tool
/agent-bootstrap --verify  # Run all verification checks
/agent-bootstrap --check   # Dry run: report what would be configured

Core Principle

Self-bootstrap without external dependencies. The manifest (agent.toml) + this skill = complete environment setup. No curl|bash, no package managers, no external tools.

Quick Reference

Use this skill when:

• Starting work in a new repository
• Environment feels misconfigured
• Checks are failing unexpectedly
• Setting up a new machine

What this skill does:

1. Reads agent.toml from repository root
2. Validates environment (required vars, commands, runtimes)
3. Configures plugins for the current tool
4. Generates MCP server config
5. Reports status (or runs verification checks with --verify)

The States

State AB0: No Manifest

Symptoms:

• No agent.toml in repository root
• User asks to bootstrap but no manifest exists

Key Questions:

• Should we create a starter agent.toml?
• What environment requirements does this project have?

Interventions:

• Offer to create a minimal agent.toml based on detected project type
• Check for common patterns (Cargo.toml → Rust, package.json → Node, etc.)
• Link to docs/agent-toml-spec.md if it exists

State AB1: Manifest Found, Validation Needed

Symptoms:

• agent.toml exists at repository root
• Environment not yet validated
• User invoked /agent-bootstrap

Key Questions:

• Are all required environment variables set?
• Are all required commands in PATH?
• Do runtime versions meet constraints?

Interventions:

• Parse agent.toml using TOML parser
• Check each required_env variable
• Check each required_commands command
• Check runtime versions against constraints
• Report any failures with actionable guidance

State AB2: Environment Valid, Configuration Needed

Symptoms:

• Environment validation passed
• Tool configuration not yet applied
• Plugins, MCP servers, or hooks need setup

Key Questions:

• Which plugins need to be enabled?
• What MCP servers should be configured?
• Which hooks should be applied?

Interventions:

• For Claude Code:

- Update .claude/settings.json with enabled plugins - Generate .mcp.json from MCP server definitions - Note: Hooks require user to copy to their settings (tool-specific)

• Report configuration changes made

State AB3: Configured, Verification Mode

Symptoms:

• User invoked /agent-bootstrap --verify
• Need to run verification checks

Key Questions:

• Which checks should run (based on trigger)?
• Did all checks pass?

Interventions:

• Run all [[verification.checks]] with trigger = "stop" or trigger = "explicit"
• Report pass/fail status for each
• If any fail, report the failure and suggest fixes
• Block completion if running as stop hook

Execution Process

Phase 1: Parse Manifest

# Read agent.toml from repository root

Parse the TOML file and extract:

• [manifest] - version check (1.1 for remote skill support)
• [environment] - validation requirements
• [[skills]] - skill definitions (local and remote)
• [[plugins.*]] - plugin configuration
• [[mcp.servers]] - MCP server definitions
• [[hooks.*]] - hook definitions
• [[verification.checks]] - verification checks

Phase 1.5: Install Remote Skills

For each skill with a source field (rather than path), install from remote:

Source Format Detection:

Installation Process:

1. Check if skill already installed: ls ~/.claude/skills/skill-name/SKILL.md
2. If not installed, determine source type and run appropriate command
3. Verify installation succeeded

Report Format:

Installing remote skills...

[1/34] gitea-workflow
  Source: jwynia/agent-skills@gitea-workflow
  ✓ Installed

[2/34] code-review
  Source: jwynia/agent-skills@code-review
  ✓ Already installed

[3/34] custom-skill
  Source: https://gitea.example.com/.../raw/.../custom-skill
  ✗ Failed: Connection refused

Remote skills: 32/34 installed, 1 already present, 1 failed

Raw URL Installation (for Gitea/GitLab):

When source starts with https:// and contains /raw/:

# Create skill directory
mkdir -p ~/.claude/skills/skill-name

# Fetch SKILL.md directly
curl -sSL "https://gitea.example.com/.../raw/.../skill-name/SKILL.md" \
  -o ~/.claude/skills/skill-name/SKILL.md

# Verify file was created
test -f ~/.claude/skills/skill-name/SKILL.md

Skills with version field:

When a skill has both source and version:

[[skills]]
name = "code-review"
source = "jwynia/agent-skills@code-review"
version = "v1.2.0"

Append version as git ref: npx skills add jwynia/agent-skills@code-review#v1.2.0 -g -y

Phase 2: Validate Environment

For each requirement, check and report:

Environment Variables:

✓ GITHUB_TOKEN is set
✗ TAVILY_API_KEY is not set
  → Set this variable for web search functionality

Commands:

✓ git (found at /usr/bin/git)
✓ cargo (found at /home/user/.cargo/bin/cargo)
✗ node not found in PATH
  → Install Node.js: https://nodejs.org/

Runtimes:

✓ rust 1.78.0 >= 1.75.0
✗ python 3.9.7 does not satisfy >=3.10
  → Upgrade Python to 3.10 or later

Phase 3: Configure Tool

Detect Current Tool:

• Look for .claude/ → Claude Code
• Look for .opencode/ → OpenCode
• Look for .cline/ → Cline
• Look for .cursor/ → Cursor

For Claude Code:

1. Plugins - Update .claude/settings.json: {"enabledPlugins": {"plugin-name@marketplace": true}}
2. MCP Servers - Generate .mcp.json: {"server-name": {"type": "http", "url": "https://...", "headers": {"Authorization": "Bearer ${ENV_VAR}"}}}
3. Hooks - Report hook configuration (requires manual setup): Hooks defined in agent.toml: - post_tool_use: rustfmt on *.rs files - stop: verification script To enable hooks, add to your Claude Code settings.

Phase 4: Report Status

Bootstrap complete for delft-core

Environment:
  ✓ All required commands found (git, cargo, rustfmt, npx)
  ✓ All runtime versions satisfied (rust 1.78.0, node 20.11.0)
  ⚠ 1 optional env var not set (TAVILY_API_KEY)

Remote Skills:
  ✓ 34 remote skills installed
  ⚠ 1 skill failed to install (see above)

Configuration:
  ✓ Enabled plugin: rust-analyzer-lsp
  ✓ Generated .mcp.json with 1 server
  ⚠ Hooks require manual configuration

Skills: 40 skills available (6 local, 34 remote)

Next steps:
  - Set TAVILY_API_KEY for enhanced web search
  - Review hooks configuration in agent.toml
  - Retry failed remote skill installs: npx skills add <source> -g -y

Phase 5: Verification (--verify mode)

Run each verification check:

Running verification checks...

[1/4] fmt
  Command: cd src && cargo fmt --check
  ✓ PASSED (0.8s)

[2/4] build
  Command: cd src && cargo build
  ✓ PASSED (12.3s)

[3/4] clippy
  Command: cd src && cargo clippy --workspace -- -D warnings
  ✗ FAILED (8.2s)

  error: unused variable `x`
   --> src/crates/delft-cli/src/main.rs:42:9

[4/4] tests
  Command: cd src && cargo test
  ⏭ SKIPPED (trigger = explicit, use --verify to run)

Verification: 2/3 passed, 1 failed

Environment Variable Checking

Check if variable is set (not its value):

# Bash approach
if [ -z "${VAR_NAME+x}" ]; then
  echo "VAR_NAME is not set"
fi

For Claude Code, use the Bash tool to check:

printenv | grep -q "^VAR_NAME=" && echo "set" || echo "not set"

Command Checking

command -v git >/dev/null 2>&1 && echo "found" || echo "not found"

Runtime Version Checking

# Rust
rustc --version | grep -oE '[0-9]+\.[0-9]+\.[0-9]+'

# Node
node --version | grep -oE '[0-9]+\.[0-9]+\.[0-9]+'

# Python
python3 --version | grep -oE '[0-9]+\.[0-9]+(\.[0-9]+)?'

File Operations

Read agent.toml

Use the Read tool to read agent.toml from repository root.

Update settings.json

Use the Edit tool to update .claude/settings.json with plugin configuration.

Generate.mcp.json

Use the Write tool to create .mcp.json with MCP server configuration.

Anti-Patterns

The Assumption Trap

Pattern: Assuming environment is correct without validation. Problem: Leads to confusing errors later. Fix: Always validate. Report clearly what's missing.

The Secret Leak

Pattern: Writing environment variable values to files. Problem: Secrets in version control. Fix: Only write ${VAR_NAME} references, never values.

The Implicit Bootstrap

Pattern: Auto-running bootstrap without user consent. Problem: Security risk; unexpected changes. Fix: Always require explicit /agent-bootstrap invocation.

The Over-Configuration

Pattern: Configuring everything even when not needed. Problem: Confusing state; hard to debug. Fix: Only configure what's declared in manifest.

Example Interactions

Example 1: Fresh Start

User: /agent-bootstrap

Your approach:

1. Read agent.toml from repository root
2. Validate environment requirements
3. Configure Claude Code settings
4. Generate MCP config if servers defined
5. Report status and next steps

Example 2: Verification

User: /agent-bootstrap --verify

Your approach:

1. Read verification checks from agent.toml
2. Run each check with trigger = "stop" or trigger = "explicit"
3. Report pass/fail for each
4. Summarize results

Example 3: Check Mode

User: /agent-bootstrap --check

Your approach:

1. Parse agent.toml
2. Report what would be validated and configured
3. Do not make any changes
4. List any issues found

Example 4: No Manifest

User: /agent-bootstrap (no agent.toml exists)

Your approach:

1. Report that no manifest found
2. Offer to create a starter manifest based on project detection
3. If user agrees, create minimal agent.toml
4. Link to specification documentation

Integration Graph

Inbound (From Other Skills)

Outbound (To Other Skills)

Complementary Skills

Manifest Schema Reference

See docs/agent-toml-spec.md for the complete schema specification.

Key sections:

• [manifest] - version (1.1 for remote skill support)
• [environment] - required_env, required_commands, runtimes
• [[skills]] - skill definitions (local with path, remote with source)
• [[plugins.{tool}]] - plugin configuration
• [[mcp.servers]] - MCP server definitions
• [[hooks.{tool}]] - hook configuration
• [[verification.checks]] - verification checks

Skills Schema (v1.1)

# Local skill
[[skills]]
name = "delft-authoring"
path = ".claude/skills/delft-authoring"

# Remote skill (GitHub)
[[skills]]
name = "code-review"
source = "jwynia/agent-skills@code-review"

# Remote skill with version pin
[[skills]]
name = "gitea-workflow"
source = "jwynia/agent-skills@gitea-workflow"
version = "v1.0.0"