@4meta5/chain

Declarative skill chaining for Claude Code workflows.

Overview

Chain enforces development workflows through capability-based skill resolution. Define what capabilities are required for a task (like test_written, test_green), and Chain automatically:

1. Matches prompts to profiles - Detects task type from your prompt
2. Resolves skill chains - Finds skills that provide required capabilities
3. Enforces tool gates - Blocks tools until capabilities are satisfied
4. Tracks progress - Persists session state across invocations

Quick Start

# Install
npm install @4meta5/chain

# Validate your configuration
chain validate

# Manually activate a profile
chain activate bug-fix

# Check current session status
chain status

# Use in Claude Code hooks (auto-activates based on prompt)
chain hook-pre-tool-use --tool '{"tool":"Write"}' --prompt "fix the login bug"

Configuration

Chain uses two YAML files in a chains/ directory:

chains/skills.yaml

Defines available skills and what they provide:

version: "1.0"

skills:
  - name: tdd
    skill_path: .claude/skills/tdd
    provides: [test_written, test_green, test_refactored]
    requires: []
    conflicts: []
    risk: low
    cost: low
    artifacts: ["**/*.test.ts"]
    tool_policy:
      deny_until:
        write_impl:
          until: test_written
          reason: "RED phase: Write failing test first"
        commit:
          until: test_green
          reason: "GREEN phase: Tests must pass"

chains/profiles.yaml

Defines workflow profiles and their requirements:

version: "1.0"

default_profile: permissive

profiles:
  - name: bug-fix
    description: TDD workflow for bug fixes
    match: [fix, bug, broken, error, failing test]
    capabilities_required: [test_written, test_green, test_refactored]
    strictness: strict
    priority: 10
    completion_requirements:
      - name: tests_pass
        type: command_success
        command: npm test

  - name: permissive
    description: No enforcement
    match: []
    capabilities_required: []
    strictness: permissive
    priority: 0

How It Works

Profile Matching

When you use --prompt, Chain matches your prompt to profiles using regex patterns:

# "fix the login bug" matches bug-fix profile (contains "fix" and "bug")
chain hook-pre-tool-use --tool '{"tool":"Write"}' --prompt "fix the login bug"

Higher scores (more pattern matches) win. On ties, higher priority wins.

Capability Resolution

Chain builds a DAG of skills based on capabilities:

Required: test_written, test_green
    ↓
Resolution:
  1. tdd provides [test_written, test_green] → selected
    ↓
Chain: [tdd]

Skills can have requirements (requires: [X]), which Chain resolves recursively.

Tool Gating

Skills define what tools are blocked until capabilities are satisfied:

tool_policy:
  deny_until:
    write:
      until: test_written
      reason: "Write failing test first (TDD RED)"

When the hook runs, it checks if the tool's intent is blocked:

$ chain hook-pre-tool-use --tool '{"tool":"Write"}'
CHAIN ENFORCEMENT: BLOCKED

Intent: write
Reason: Write failing test first (TDD RED)

Path-Aware Intents

Chain supports fine-grained blocking based on file paths. Instead of blocking all writes, you can target specific file categories:

TDD Example: Block implementation writes until tests exist, but always allow writing tests:

tool_policy:
  deny_until:
    write_impl:
      until: test_written
      reason: "TDD RED: Write a failing test first"
    commit:
      until: test_green
      reason: "TDD GREEN: Tests must pass"

With this config:

• Write to src/index.ts → blocked (write_impl)
• Write to src/index.test.ts → allowed (write_test)
• Write to README.md → allowed (write_docs)

Fallback behavior: If you block the base write intent, it blocks all file writes regardless of category. Path-aware intents provide finer control when needed.

Session Persistence

Chain stores session state in .chain-state.json:

{
  "session_id": "auto-1234567890",
  "profile_id": "bug-fix",
  "chain": ["tdd"],
  "capabilities_required": ["test_written", "test_green"],
  "capabilities_satisfied": [],
  "blocked_intents": {
    "write": "Write failing test first (TDD RED)"
  }
}

CLI Commands

Hook Options

chain hook-pre-tool-use \
  --tool '{"tool":"Write","input":{"path":"src/foo.ts"}}' \
  --prompt "fix the login bug" \  # Optional: enables auto-activation
  --no-auto                       # Optional: disable auto-activation
  --cwd /path/to/project          # Optional: working directory

Programmatic API

import {
  ChainActivator,
  createRouteDecision,
  PreToolUseHook,
  matchProfileToPrompt,
  resolve,
  loadConfig,
} from '@4meta5/chain';

// Load configuration
const config = await loadConfig('/path/to/project');

// === Router Integration ===
// Activate from a RouteDecision (for middleware integration)
const activator = new ChainActivator(cwd, config.skills, config.profiles);

const decision = createRouteDecision(
  'req-123',                    // request_id for idempotency
  'fix the login bug',          // query
  'immediate',                  // mode: immediate | suggestion | chat
  [{ name: 'bug-fix', score: 0.9 }]  // candidates
);

const activation = await activator.activate(decision);
// → { activated: true, session_id: '...', chain: ['tdd'], blocked_intents: {...} }

// Idempotent: same request_id returns existing session
const same = await activator.activate(decision);
// → { activated: true, idempotent: true, session_id: same as before }

// === Profile Matching ===
// Match prompt to profile
const profile = matchProfileToPrompt(
  'fix the login bug',
  config.profiles
);
// → { name: 'bug-fix', matchScore: 2, matchedPatterns: ['fix', 'bug'], ... }

// Resolve skill chain
const result = resolve(profile, config.skills);
// → { chain: ['tdd'], blocked_intents: { write_impl: '...' }, ... }

// === Hook Usage ===
// Use the hook
const hook = new PreToolUseHook(cwd, config.skills, config.profiles);
const result = await hook.check(
  { tool: 'Write', input: { path: 'src/foo.ts' } },
  { prompt: 'fix the login bug' }
);
// → { allowed: false, message: 'BLOCKED...', blockedIntents: [...] }

Integration with Claude Code

Add to your .claude/settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": ".*",
        "hooks": [
          {
            "type": "command",
            "command": "chain hook-pre-tool-use --tool '$TOOL_INPUT'"
          }
        ]
      }
    ],
    "Stop": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "chain hook-stop"
          }
        ]
      }
    ]
  }
}

Strictness Levels

Related Skills

Chain works with skills from 4meta5/skills:

• tdd - Test-Driven Development (RED → GREEN → REFACTOR)
• dogfood - Enforce dogfooding the tools you build
• repo-hygiene - Repository cleanup and documentation

License

MIT