Skip to content

iannelsondev/adc

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

73 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

ADC — Autonomous Development Configuration

A portable, provider-agnostic configuration system for AI coding agents — 12 agents, 159 skills, 79 rules, and 35 hooks that install into any supported CLI tool.

Structure

.
├── agents/          # 12 agent definitions (10 core + 2 AutoResearch)
│   ├── cards/       # A2A Agent Card JSON files (10 cards)
│   └── MIGRATION.md # Legacy agent → core agent mapping
├── hooks/           # Shell scripts for lifecycle hooks
├── plugins/         # Plugin install manifest
├── rules/           # Global coding convention rules
├── skills/          # 139 skills (63 workflow + 71 domain)
├── providers/       # Provider mapping files and routing profiles
├── specs/           # Spec-driven development (SPEC-001 through SPEC-093)
├── AGENTS.md        # Cross-provider agent behavior contract
├── SPEC.md          # Project manifest
├── CHANGELOG.md     # Version history
├── settings.json    # Main settings (permissions, hooks, env, plugins)
├── settings.local.json  # Machine-local settings overrides
├── mcp.json         # MCP server manifest (8 core + 4 authenticated + 4 optional)
├── install-mcp.sh   # MCP server installer
└── install.sh       # Symlink installer

Components

Settings

  • settings.json — Permissions (auto-allow all tools + MCP), hook wiring, env vars (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS), and plugins (rust-analyzer LSP). Includes $schema for editor autocomplete.
  • settings.local.json — Machine-specific overrides (currently allows the browse skill).
  • plugins/installed_plugins.json — Plugin install manifest for reproducibility. The installer prints claude plugin install commands for each plugin.

MCP Servers (mcp.json)

Pre-configured Model Context Protocol servers organized in three tiers:

Core (no auth, install and go):

Server What it does
context7 Up-to-date library docs and code examples (prevents hallucinated APIs)
sequential-thinking Structured reasoning for complex multi-step problems
fetch Fetch URLs and convert HTML to Markdown
memory Persistent knowledge graph across sessions
playwright Browser automation (navigate, interact, screenshot)
git Git operations (status, diff, log, commit, branch, blame)
kubernetes Kubernetes cluster management
docker Docker container management

Authenticated (require API keys):

Server What it does Auth
github GitHub API (repos, issues, PRs, code search) GITHUB_TOKEN
gitlab GitLab API (projects, MRs, issues, pipelines) GITLAB_TOKEN
brave-search Web search via Brave (2,000 free queries/month) BRAVE_API_KEY
grafana Dashboards, Prometheus, Loki, alerting GRAFANA_URL + GRAFANA_TOKEN

Optional (project-specific):

Server What it does
postgres PostgreSQL schema inspection and queries
surrealdb SurrealDB schema management and SurrealQL
semgrep Static analysis security scanning (OWASP Top 10)
sqlite SQLite database queries

Rules (rules/)

63 coding convention rules, automatically applied based on file path patterns:

Languages:

Rule Scope
go.md Go — structured logging, error wrapping, table-driven tests, context-first
rust.md Rust — ownership, error handling (thiserror/anyhow), tracing, unsafe auditing
python.md Python — type annotations, Pydantic v2, asyncio, pytest, ruff/mypy
typescript.md TypeScript — strict mode, discriminated unions, Zod, vitest
java.md Java 17-21+ — records, sealed classes, Spring Boot 3.x, virtual threads
svelte.md SvelteKit 5 — runes, props, Tailwind, TypeScript
react.md React 18/19 — Server Components, concurrent rendering, TanStack Query

Domains:

Rule Scope
api.md API — REST, GraphQL, gRPC, event-driven, middleware order
frontend.md Frontend — layout, semantic HTML, accessibility, performance
security.md Security — OWASP Top 10, auth, crypto, secrets, web security
testing.md Testing — table-driven, AAA pattern, coverage targets, E2E
database.md Database — queries, schema, migrations, PostgreSQL, SurrealDB
documentation.md Documentation — Diataxis framework, ADRs, API docs, code examples
observability.md Observability — structured logging, metrics, traces, SLOs, dashboards
infrastructure.md Infrastructure — containers, Kubernetes, Helm, CI/CD, secrets
k8s.md Kubernetes/Flux — GPU strategy, image tags, GitOps, storage
compliance.md Compliance — data classification, PII, NIST, GDPR, HIPAA, SOC 2

Systems & Process:

Rule Scope
linux.md Linux — systemd units, udev, packaging, desktop integration
desktop-gui.md Desktop GUI — Qt6/QML, GTK4/libadwaita, accessibility, Wayland
shell.md Shell scripts — safety (set -euo pipefail), variables, error handling
quality.md Code quality — function design, complexity, naming, review checklist
git.md Git — commit messages, branching, PR conventions, history

Hooks (hooks/)

Shell scripts wired into Claude Code lifecycle events:

Hook Event Purpose
guard-bash.sh PreToolUse (Bash) Blocks catastrophic commands (rm -rf /, force push main, DROP DATABASE, etc.)
guard-files.sh PreToolUse (Edit/Write) Blocks edits to sensitive files (.env, credentials, keys, kubeconfig)
autoformat.sh PostToolUse (Edit/Write) Auto-formats files by extension (gofmt, prettier, black, rustfmt)
guard-commit.sh PostToolUse (Bash) Warns about --no-verify and missing issue references in commits
notify.sh Notification Desktop notification via notify-send when Claude needs attention
post-compact.sh SessionStart (compact) Re-injects engineering discipline reminders after context compaction
track-tokens.sh SessionStart Logs session starts for usage tracking

Agents (agents/)

12 agent definitions following a core-agent-plus-skills architecture. Each core agent covers a broad role and loads domain-specific skills on demand. See AGENTS.md for the cross-provider behavior contract and agents/cards/ for machine-readable A2A Agent Cards.

Agent Tier Domains
developer sonnet Code implementation across all languages/frameworks, UI/UX, databases, debugging
reviewer sonnet Code review, doc review, accessibility audit, performance review
architect opus System, infrastructure, data, network, agent, and swarm architecture
tester sonnet Unit, integration, e2e, and performance test authoring
security-analyst opus Threat modeling, pentesting, AppSec, supply chain, LLM/agent security
operations sonnet Deployment, SRE, observability, containers, incident response, Linux
compliance sonnet RMF, STIG, FedRAMP, privacy, control mapping, evidence, SBOM
researcher sonnet Web research, market/competitive/customer/product analysis
writer sonnet Technical docs, API docs, architecture docs, diagrams
planner opus Project planning, requirements, migration, release, GTM strategy
research-orchestrator opus AutoResearch improvement loop (SPEC-078)
eval-runner sonnet Mechanical evaluation harness for AutoResearch (SPEC-078)

Skills (skills/)

139 skills in two categories:

63 workflow skills — slash-command invokable with /<skill-name>:

/accessibility-audit /agent-design /agent-eval /agent-security-audit /agent-security-design /api /api-docs /architecture-design /architecture-docs /brand /browse /bug-fix /changelog /ci-cd /cleanup /code-review /competitive-analysis /customer-research /database-migration /dependency-audit /deployment /discipline /documentation-review /e2e-test /evaluate-model /feature-development /finetune /gtm-plan /inline-comments /integration-test /issues /kickstart /llm-security-audit /monitoring-setup /mutation-testing /ops /performance-review /pricing-strategy /product-review /project /provenance-audit /quantize /readme-generation /red-team /release /requirements-analysis /risk-assessment /rmf /route /sdlc /secret-detection /security-review /seo-audit /spec /structure /swarm-design /task-breakdown /test-coverage-analysis /threat-modeling /timeline-estimation /ui-component /unit-test-writing /vulnerability-scan

71 domain skills — loaded by core agents when task context matches:

Domain skills provide focused expertise (Go conventions, Kubernetes security, STRIDE threat modeling, etc.) without procedural steps. Core agents select them via the skill routing convention in rules/skill-routing.md.

Installation

git clone https://github.com/iannelsondev/adc.git
cd adc

Pick your CLI tool and install. You can install into one or all of them — each gets its own translated agent library.

Multiple CLI tools

# Install into all your CLI tools at once
./install.sh --multi-provider claude,codex,copilot,gemini

# Install MCP servers
./install-mcp.sh

This installs ADC's agents, rules, hooks, and skills into each CLI's config directory and creates a route table for multi-provider routing.

Other orchestrators like opencode can also consume the provider mappings and routing profiles in the same manner.

Single CLI installation

Claude Code

./install.sh
./install-mcp.sh

Symlinks agents, rules, hooks, skills, and settings into ~/.claude/. Updates are automatic — git pull and changes are live.

  • Auth: Log in with claude (Anthropic subscription — Max, Team, or Enterprise)
  • Models: Opus, Sonnet, Haiku
  • Agents: Full support — 12 core agents with native tool names (Read, Write, Edit, Bash, Glob, Grep, Agent)
  • Skills: Full support — 139 skills (63 workflow + 71 domain)
  • Hooks: Full support — 7 lifecycle hooks (guard, autoformat, notifications)
  • Test: claude -p "hello" --model haiku

Codex CLI

./install.sh --provider codex

Translates agents into Codex format and installs into ~/.codex/. Generates config.toml with permissions.

  • Auth: Log in with codex login (ChatGPT Plus/Pro/Team account)
  • Models: GPT-5.4, GPT-5.1-codex, GPT-5-mini
  • Agents: Translated — tool names mapped (read_file, write_file, edit_file, bash), model tiers mapped to GPT equivalents
  • Skills: Not directly supported — use agent prompts instead
  • Hooks: Not supported (Codex has its own sandbox model)
  • Test: codex exec -m gpt-5.4 "hello" --full-auto

Copilot CLI

./install.sh --provider copilot

Translates agents into Copilot format and installs into ~/.copilot/.

  • Auth: Log in with copilot or gh auth login (GitHub Copilot subscription)
  • Models: Claude Opus 4.6, Claude Sonnet 4.6, Claude Haiku 4.5, GPT-5.2, Gemini 3 Pro (Copilot routes internally)
  • Agents: Translated — same interface as Claude Code (-p, --agent, --model)
  • Skills: Not directly supported — use agent prompts instead
  • Hooks: Not supported
  • Test: copilot -p "hello" --allow-all-tools

Gemini CLI

./install.sh --provider gemini

Translates agents into Gemini format and installs into ~/.gemini/.

  • Auth: Set GEMINI_API_KEY or configure ~/.gemini/settings.json (Google AI Studio or Vertex AI)
  • Models: Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 2.0 Flash Lite
  • Agents: Translated — tool names mapped (read_file, write_file, edit_file, shell), model tiers mapped to Gemini equivalents
  • Skills: Not directly supported — use agent prompts instead
  • Hooks: Not supported
  • Context window: 1M tokens (largest of any provider)
  • Test: gemini -p "hello"

What the installer does

The installer reads the provider's mapping file (providers/<provider>.json) and translates ADC's portable agent definitions — tool names, model tiers, hook events — into the format each CLI expects:

Provider Config directory Tool mapping Model mapping
claude ~/.claude/ Read, Write, Edit, Bash, Glob, Grep opus, sonnet, haiku
codex ~/.codex/ read_file, write_file, edit_file, bash gpt-5.4, gpt-5.1-codex, gpt-5-mini
copilot ~/.copilot/ read_file, write_file, edit_file, bash claude-opus-4.6, claude-sonnet-4.6, claude-haiku-4.5
gemini ~/.gemini/ read_file, write_file, edit_file, shell gemini-2.5-pro, gemini-2.5-flash, gemini-2.0-flash-lite

MCP servers

# Core MCP servers (no auth required)
./install-mcp.sh

# Everything including authenticated servers
export GITHUB_TOKEN=ghp_...
export GITLAB_TOKEN=glpat-...
export BRAVE_API_KEY=BSA...
./install-mcp.sh --all

Prerequisites

  • jq — required by hook scripts and the installer for JSON parsing
  • notify-send — optional, for desktop notifications (Linux only)
  • Language formatters (optional, for auto-format hook): gofmt, goimports, prettier, black, rustfmt

Updating

cd /path/to/adc
git pull
# Claude: symlinks mean changes are already active — no reinstall needed
# Codex/Gemini/Copilot: re-run ./install.sh --provider <name> to pick up changes
# Multi-provider: re-run ./install.sh --multi-provider <list> to pick up changes

Self-Improvement (/autoresearch)

ADC improves itself based on real-world usage. When agents give wrong guidance, users correct them. Those corrections accumulate as feedback memories. The AutoResearch system transforms those corrections into configuration improvements — automatically.

How it works

User corrects agent  →  Feedback memory stored  →  Corrections aggregate
         ↓                                                    ↓
   (tagged with artifact path,               (3+ corrections on same artifact
    correction type, content)                  triggers improvement proposal)
                                                              ↓
                                              /autoresearch plan --from-feedback
                                                              ↓
                                              Assertions generated from corrections
                                                              ↓
                                              /autoresearch run (on branch)
                                                              ↓
                                              Convergence chain + PR for review

Two improvement loops

Project-local: Corrections from your project improve your local configuration. Overrides go to .claude/skills/ and .claude/rules/, taking precedence over global ADC config for your project only.

ADC-global: When the same correction appears across multiple projects (3+ corrections, 2+ projects), it signals a systemic issue. The ADC maintainer runs /autoresearch run --global to improve the canonical configuration. Changes land on an autoresearch/ branch for human review — no autonomous merge.

Safety rails

  • Mechanical evaluation only — binary assertions, no LLM-as-judge
  • Oscillation detection — strategy categories exhausted after 3 consecutive failures
  • Guard command — structural validation before every commit (mandatory for all targets)
  • Convergence chain — auditable record linking each improvement to its metric delta and feedback sources
  • No autonomous merge — improvement branches require human review
  • Test case immutability — the loop cannot modify its own evaluation criteria
  • Provenance — every change attributed to the model that produced it via Co-Authored-By

Quick start

# Plan an improvement from user feedback (requires surrealmemory)
/autoresearch plan --from-feedback skills/go/SKILL.md

# Or plan manually with custom assertions
/autoresearch plan

# Run the loop
/autoresearch run

# Check progress
/autoresearch status

# Generate improvement report (for PR description)
/autoresearch report

See SPEC-077 (types), SPEC-079 (skill), SPEC-094 (enforcement), and SPEC-095 (feedback pipeline) for details.

Optional: Graph Backend (surrealmemory)

ADC specs can be indexed into a SurrealDB graph for queryable relationships, coherence scoring, and Mermaid diagram export via surrealmemory (sm).

This is optional — the base spec system works entirely with flat Markdown files. The graph backend adds:

  • Spec dependency traversal and coherence scoring
  • Memory graph with GraphRAG context assembly
  • MCP server exposing 16 tools to Claude Code agents
  • Work path DAG for causal reasoning
# Install sm (Arch Linux)
paru -S surrealmemory  # or download from GitHub releases

# Initialize and sync ADC specs into the graph
sm init
sm sync --specs-dir specs/
sm spec coherence

See SPEC-019 (Spec Graph) and SPEC-024 (Graph Memory) for details.

About

Portable AI coding agent configuration — 12 agents, 159 skills, 79 rules, 35 hooks. Self-improving from real-world usage via AutoResearch feedback pipeline. Installs into Claude Code, Codex, Copilot, or Gemini CLI.

Topics

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors

Languages