AI Agent Integration¶
Eidos is agent-agnostic and model-agnostic. It provides a standardized plugin bundle conforming to agent-plugins.org v1.0.0 and 4 cohesive Agent Skills conforming to agentskills.io so any AI coding assistant—such as Google Antigravity, Claude Code, Cursor, GitHub Copilot, Aider, Cline, OpenCode, or OpenAI Codex—can understand, query, and mutate your living specifications safely through direct CLI execution.
┌─────────────────────────────────────────┐
│ AI Coding Assistant │
│ (Google Antigravity, Claude, Cursor, …) │
└────────────────────┬────────────────────┘
│
Direct CLI / Skills │ Rules & Guidelines
(eidos validate, …) │ (AGENTS.md, .mdc)
▼
┌─────────────────────────────────────────┐
│ Eidos Agent Plugin │
│ (agent-plugins.org v1) │
└────────────────────┬────────────────────┘
│
Inspect Architecture │ Scaffold & Apply
& Invariants │ SpecDeltas (DLT-*)
▼
┌─────────────────────────────────────────┐
│ Living Specifications │
│ (eidos/*.adoc in Git repo) │
└─────────────────────────────────────────┘
The CLI-First Architecture for AI Agents¶
Modern AI coding agents (Claude Code, Cursor Composer, Google Antigravity, Copilot, Cline, Aider) operate inside terminal-capable sandboxes, subshells, and Docker containers. Eidos is built CLI-first:
- Zero Daemon Flakiness: No long-running background stdio server processes to crash, disconnect, deadlock, or require manual port management.
- Universal Cloud & Sandbox Portability: Works out-of-the-box in GitHub Codespaces, remote SSH sessions, CI/CD runners, and ephemeral cloud devboxes where local stdio daemon connections are unreliable.
- Deterministic
--jsonContracts: Every inspection, validation, and drift check command supports--jsonoutput, providing deterministic schemas that agents can parse programmatically without regex or LLM hallucination. - Token-Efficient Progressive Disclosure: Through
eidos context, agents extract targeted architectural slices (~140 to ~1,500 tokens) rather than dumping thousands of lines of documentation into the context window. - Strict Exit Code Semantics: Commands return standard Unix exit codes (
0for success/clean,1for validation errors or drift), enabling automated tool loops and pre-commit hooks to halt instantly on failure.
One-Command Setup¶
Eidos automatically configures your project for all supported AI coding assistants in a single step:
# Default installation (configures project scope, Cursor, and Claude Code):
eidos plugin install
# Full installation across all supported AI assistants:
eidos plugin install all
# Targeted installation for a specific assistant:
eidos plugin install cursor
eidos plugin install claude
eidos plugin install copilot
Installation Targets¶
| Target Argument | Target Path / Config | Description |
|---|---|---|
all |
All environments below | Installs plugins and configs for Antigravity, Claude, Cursor, Aider, Cline, OpenCode, Codex, and GitHub Copilot |
project |
.agents/plugins/eidos/ |
Standard project-scope installation conforming strictly to agent-plugins.org v1.0.0 |
antigravity |
.agents/plugins/eidos/ |
Installs plugin bundle with rules, 4 skills, hooks, and extensions (agent-plugins.org v1.0.0) |
claude |
.claude/plugins/eidos/, .claude/skills/ |
Installs Claude Code plugin, 4 flat skills, slash commands (/validate, /drift), and configures .claude/settings.json |
cursor |
.cursor/skills/, .cursor/rules/eidos.mdc |
Installs 4 flat skills and Cursor rules (.mdc) for CLI execution |
copilot |
.github/plugin/, .github/skills/, .github/copilot/ |
Installs GitHub Copilot plugin (Agent Plugins 1.0), 4 skills, settings, and instructions |
aider |
.aider/skills/, .aider.conventions.md, .aider.conf.yml |
Installs 4 skills, conventions index, and auto-read configuration |
cline |
.cline/plugins/, .cline/skills/ |
Installs plugin bundle, 4 flat skills, and AGENTS.md conventions |
open-code |
.opencode/plugins/, .opencode/skills/ |
Installs plugin bundle, 4 flat skills, and AGENTS.md instructions |
codex |
.codex/plugin.json, .codex/skills/, .codex/rules.md |
Installs OpenAI Codex plugin bundle with 4 skills and rules |
Assistant Integration Guides¶
Select your AI coding assistant below for step-by-step setup instructions and configuration examples:
Google Antigravity natively discovers plugins adhering to the agent-plugins.org v1.0.0 standard inside .agents/plugins/.
1. Installation¶
Run the installation command in your repository root:
This creates the following directory structure inside your project:
.agents/plugins/eidos/
├── plugin.json # agent-plugins.org v1.0.0 manifest
├── hooks.json # Document linting & drift gate hooks
└── skills/
├── eidos/ # Master orchestrator, context slicing & continuous drift detection
├── eidos-spec/ # Product, Arc42/C4 architecture, design systems & invariants (NO ADRs)
├── eidos-plan/ # Technical spikes (EXP-*), SpecDeltas (DLT-*) & Staff Architect review
└── eidos-ops/ # Greenfield init, brownfield extraction & Living Web Portal
2. How Antigravity Uses Eidos¶
- Master Engineering Skill: Antigravity indexes
skills/eidos/SKILL.mdto guide agent decision-making, strictly enforcing the 5 Eidos architectural invariants (Single Source of Truth, No ADRs, SpecDelta lifecycle, Invariant-Test traceability, and Spec validation) alongsideeidos contextandeidos drift. - Workflow-Aligned Agent Skills: Antigravity agents invoke
/eidos-specto author or inspect living specs and invariants,/eidos-planfor explorations, SpecDeltas, and Staff reviews, and/eidos-opsfor repository initialization, brownfield extraction, and Web Portal operations. - Quality Hooks: Quality gates intercept document edits to verify AsciiDoc schemas before changes are finalized.
Claude Code (Anthropic's agentic CLI) communicates with Eidos via native Agent Skills and direct CLI execution.
Claude Code (CLI)¶
Claude Code natively supports a standardized, file-based plugin architecture:
1. Manifest (plugin.json): Plugin namespace, versioning, and metadata.
2. Commands (commands/*.md): Native slash commands (/validate, /drift, /context).
3. Skills (skills/*/SKILL.md): Domain expertise triggerable by user tasks.
4. Subagents: Specialized roles dynamically forked by Claude.
5. Hooks (hooks.json): Lifecycle event handlers (pre-tool checks, lint gates).
1. Automatic Installation¶
Install the Eidos Claude Code plugin into your project:
This automatically:
- Installs the full plugin bundle into .claude/plugins/eidos/ (following Claude Code's Method 2: Project Root Injection) with plugin.json, .claude-plugin/plugin.json, and hooks.json.
- Generates Claude Code interactive slash commands in .claude/plugins/eidos/commands/ (/validate, /drift, /context).
- Installs all 4 cohesive Eidos skills into .claude/skills/<skill-name>/ so they are immediately discoverable by Claude Code's native /skills registry and direct command invocations (/eidos, /eidos-spec, /eidos-plan, /eidos-ops).
- Registers and enables the plugin in .claude/settings.json (plugins: [{"path": ".claude/plugins/eidos"}], enabledPlugins: {"eidos": true, "eidos@plugins-dir": true}).
2. Enforcing Rules in CLAUDE.md¶
To ensure Claude Code adheres to Eidos architectural governance, reference the Eidos rules in your project's CLAUDE.md:
<!-- CLAUDE.md -->
## Eidos Architectural Guidelines
- Adhere strictly to the guidelines in `.agents/plugins/eidos/skills/eidos/SKILL.md`.
- Treat `eidos/` living specifications as the single source of truth.
- Never create ADRs; all mutations must follow the SpecDelta lifecycle (`eidos delta create`).
- Validate specifications with `eidos validate --json` before completing tasks.
- Check for spec-to-code drift with `eidos drift --uncommitted --json`.
Cursor connects to Eidos via native Agent Skills (.cursor/skills/) and custom Cursor rules (.cursor/rules/eidos.mdc), invoking Eidos CLI commands.
1. Automatic Configuration¶
Generate Cursor rules and 4 native Agent Skills:
This automatically:
- Generates .cursor/rules/eidos.mdc with alwaysApply: true.
- Installs all 4 skills into .cursor/skills/<skill-name>/ enabling Cursor's native Agent Skills progressive disclosure engine.
2. Cursor Rules (.cursor/rules/eidos.mdc)¶
Cursor Composer and Chat agents follow .cursor/rules/eidos.mdc:
---
description: "Eidos living specifications and architectural governance"
alwaysApply: true
---
# Eidos Living Specifications
This repository is governed by Eidos living specifications under `eidos/`.
## Core Principles
- Consult living specifications before modifying code or proposing architectural changes.
- Strictly enforce living specifications (NO ADRs). All architectural mutations follow the SpecDelta lifecycle (`eidos/deltas/`).
- Preserve bidirectional invariant-test traceability (`POL-002`).
- Verify specification schemas with `eidos validate eidos` (or `eidos validate --json`).
- Verify synchronization with `eidos drift --uncommitted` (or `eidos drift --uncommitted --json`) before finalizing tasks.
- Use Eidos CLI commands (`eidos context`, `eidos index search`, `eidos drift`) for real-time inspection.
- Consult installed Agent Skills in `.cursor/skills/` for specialized architectural workflows.
GitHub Copilot integrates with Eidos using native GitHub Copilot Plugins adhering to Agent Plugins 1.0 (agent-plugins.org v1.0.0), direct Agent Skills (.github/skills/), and repository settings (.github/copilot/settings.json).
1. Installation¶
This automatically:
- Installs the native Copilot plugin bundle into .github/plugin/ with plugin.json (declaring com.github.copilot extension).
- Installs all 4 Eidos skills flat into .github/skills/<skill-name>/SKILL.md where GitHub Copilot's Agent Skills engine natively scans and indexes them.
- Configures .github/copilot/settings.json enabling the eidos plugin (enabledPlugins: {"eidos": true}).
- Configures .github/copilot-instructions.md with Eidos living specification rules, safely demarcated with comment markers (<!-- EIDOS_COPILOT_START --> ... <!-- EIDOS_COPILOT_END -->) to preserve any custom user instructions.
2. Using with Copilot¶
GitHub Copilot automatically discovers the installed skills under .github/skills/, activates the plugin declared in .github/copilot/settings.json, and loads .github/copilot-instructions.md in Copilot Chat and Copilot Agent mode—enforcing the single source of truth under eidos/, zero spec-to-code drift, and bidirectional invariant-test traceability.
Aider pairs with Eidos living specifications by loading persistent conventions via .aider.conf.yml and .aider.conventions.md.
1. Installation¶
This automatically:
- Installs all 4 individual Eidos skills into .aider/skills/<skill-name>/SKILL.md.
- Generates .aider.conventions.md in the project root containing core Eidos guidelines (Single Source of Truth, No ADRs, Invariant traceability, spec validation, and drift gate) alongside an interactive index mapping tasks to /read .aider/skills/<skill>/SKILL.md.
- Configures .aider.conf.yml to automatically include .aider.conventions.md under read:.
2. Using with Aider¶
Run Aider as usual:
Aider automatically ingests.aider.conventions.md as read-only context on launch. You can verify specifications from within Aider using /run uv run eidos validate --json and check drift using /run uv run eidos drift --uncommitted --json. When working on specific architectural areas, prompt Aider to load the corresponding skill with /read .aider/skills/<skill-name>/SKILL.md.
Cline connects to Eidos through native Agent Skills (.cline/skills/) and plugin bundle (.cline/plugins/eidos). It natively discovers the repository's root AGENTS.md for architectural rules and conventions.
1. Installation¶
This automatically:
- Installs the complete plugin bundle into .cline/plugins/eidos/.
- Installs all 4 individual skills flat into .cline/skills/<skill-name>/SKILL.md where Cline's skill discovery engine indexes them for progressive disclosure.
- Leverages the root AGENTS.md open standard without requiring redundant rule files.
OpenCode integrates with Eidos via local plugin discovery under .opencode/plugins/eidos and flat skills under .opencode/skills/. OpenCode natively reads the root AGENTS.md for project instructions.
1. Installation¶
This automatically:
- Installs the Eidos plugin bundle into .opencode/plugins/eidos.
- Installs all 4 individual skills flat into .opencode/skills/<skill-name>/SKILL.md.
- Leverages the root AGENTS.md open standard without requiring redundant rule files.
OpenAI Codex utilizes portable agent plugin packages containing plugin.json, skills, and rules.
1. Installation¶
This automatically:
- Scaffolds .codex/plugin.json declaring com.openai agent plugin extension and metadata.
- Installs all 4 individual skills flat into .codex/skills/<skill-name>/SKILL.md.
- Generates .codex/rules.md governing living spec workflows.
Any terminal-capable AI coding assistant or agent can operate against Eidos living specifications by reading the root AGENTS.md and invoking Eidos CLI commands directly:
CLI Automation Commands for AI Agents¶
AI agents execute Eidos CLI commands directly from terminal / subshell tools. All core commands support machine-readable --json output:
| CLI Command | Flags / Options | Description |
|---|---|---|
eidos context |
-f <file>, -c <id>, -t <task>, --tier <tier>, --json |
Slices a compact architectural context subgraph (anchor, guardrails, deep). |
eidos index repo-map |
--max-tokens <n>, --json |
Generates an AST-compressed structural repository map (~98% token reduction). |
eidos index find-symbol |
<query>, --limit <n>, --json |
Resolves classes, functions, and methods to exact file/line coordinates and owning specs. |
eidos index search |
"<query>", --limit <n>, --json |
Performs two-tier column-weighted BM25 full-text search across living spec titles, rules, and invariants. |
eidos index blast-radius |
<symbol>, --depth 2, --json |
Computes transitive dependency and governance blast radius for a symbol or spec path. |
eidos index status |
--json |
Reports index health, cache size, AST telemetry, entity/edge counts, and auto-sync freshness. |
eidos index sync |
--force |
Incrementally synchronizes or force-rebuilds the architectural knowledge graph. |
eidos drift |
--uncommitted, --base <ref>, --json |
Inspects spec-to-code drift, boundary-aware spec anchoring, and polyglot INV-* test bindings. |
eidos validate |
[PATH], --json |
Validates all living specifications against their bundled v1 JSON Schemas, x-eidos-ref links, and bidirectional contract parity. |
eidos doctor |
--path <path>, --json |
Runs workspace health diagnostics across CLI binary, living specs, index, and agent plugin hooks. |
eidos info |
Returns current engine version, loaded v1 schemas, and supported platforms. | |
eidos spec list |
--type <type>, --status <status>, --json |
Lists living specifications (including inline C4 Level 3 components CMP-*). |
eidos spec create |
<type> <id> --title <title> [--path <path>] |
Scaffolds a minimal draft blueprint across all 15 living specification types. |
eidos spec delete |
<id_or_path> [--force] |
Safely deletes a living specification with cross-reference safety checks. |
eidos spec component |
[promote\|inline] <id> |
Promotes inline C4 Level 3 components to eidos/architecture/components/ or inlines them back. |
eidos delta create |
[id] --title <title> [--kind <feature\|enabler\|bug>] [--slug] |
Scaffolds a new collision-free SpecDelta mutation document (DLT-*) for Planned Mode. |
eidos delta verify |
<id>, --json |
Verifies that all declared mutations and closed-schema expect: post-conditions hold in the spec catalog. |
eidos delta audit |
<id>, --strict, --json |
Audits removed/deprecated entities for dangling references vs. approved tombstones. |
eidos delta apply |
<id>, --verify/--no-verify, --json |
Verifies declared mutations by default (--verify), validates the SpecDelta, and reconciles it in-place (status: applied). |
eidos delta changelog |
--stdout, --output <file> |
Compiles Keep a Changelog release notes from Git tags and applied SpecDeltas. |
eidos init |
[PATH], --lite |
Bootstraps a repository with the full catalog or progressive 3-file Eidos Lite profile. |
Machine-Readable JSON Output Reference¶
When agents pass --json, Eidos outputs structured JSON payloads designed for programmatic consumption:
1. Specification Validation (eidos validate --json)¶
{
"is_clean": true,
"total_docs": 62,
"valid_docs": 62,
"schema_errors": {},
"broken_references": []
}
0 when all documents are valid (is_clean: true), 1 when any schema or cross-reference error is detected.
2. Drift Detection (eidos drift --uncommitted --json)¶
{
"is_clean": true,
"drift_detected": false,
"workflow_mode": "planned",
"code_files": ["src/eidos/cli/main.py"],
"spec_files": ["eidos/architecture/containers/cli.adoc"],
"delta_files": ["eidos/deltas/dlt-2026-016.adoc"],
"test_files": ["tests/test_cli_json.py"],
"touched_components": ["CNT-EIDOS-CLI"],
"unmapped_code_files": [],
"valid_specs": 62,
"total_specs": 62,
"validation_errors": [],
"uncovered_invariants": [],
"orphan_invariants": [],
"remediation_suggestions": [],
"trivially_covered_files": []
}
0 when clean, 1 when unmanaged code changes, trivial comment-only spec edits, or broken test bindings exist.
3. Context Slicing (eidos context -f src/eidos/cli/main.py --json)¶
{
"tier": "guardrails",
"components": [
{
"id": "CNT-EIDOS-CLI",
"title": "Eidos Command-Line Interface",
"type": "tech/container",
"source_path": "src/eidos/cli"
}
],
"contracts": [
{
"id": "CTR-EIDOS-CLI",
"protocol": "cli_args",
"container": "CNT-EIDOS-CLI"
}
],
"capabilities": [
{
"id": "CAP-AGENT-001",
"title": "Agent Adapter & Context Engine",
"domain": "DOM-AGENT-INTELLIGENCE"
}
],
"invariants": [
{
"id": "INV-AGENT-GRAPH-INDEX-FRESHNESS",
"rule": "Every CLI query lazily checks dirty status via SHA-256 and mtime hashing.",
"has_test_coverage": true,
"test_binding": "tests/test_graph_index.py::test_spike_b_incremental_dirty_check_and_lazy_sync"
}
],
"policies": [
{
"id": "POL-001-GIT-SOURCE-OF-TRUTH",
"title": "Git Repository as the Single Source of Truth"
}
]
}
The 7 Constitutional Guardrails¶
Eidos embeds seven non-negotiable architectural guardrails in its master skill skills/eidos/SKILL.md. When working in an Eidos repository, every AI assistant is instructed to obey them:
1. Single Source of Truth (POL-001)¶
All product domains, capabilities, personas, C4 architecture containers/components, interface contracts, design systems, and policies live inside eidos/ as structured AsciiDoc (.adoc) files with YAML frontmatter.
2. Zero ADRs (POL-001)¶
Standalone Architecture Decision Records (docs/adr/, eidos/adr/, or ADR-*.md) are strictly prohibited. Current architectural state lives in eidos/architecture/, durable design rationale and rejected alternatives are recorded in-place under the decisions: (DEC-*) frontmatter block of the owning specification, and transitions are governed by SpecDeltas (eidos/deltas/dlt-*.adoc).
3. Dual Workflow Mode Selection (POL-003)¶
Every task operates in either Direct Mode (fast, in-place co-evolution of code, tests, and owning specs for localized edits) or Planned Mode (formal SpecDelta DLT-* lifecycle for cross-container or breaking changes). Agents classify the workflow mode before writing code.
4. Mandatory Invariant-Test Traceability (POL-002)¶
Every capability invariant (INV-*) must declare a concrete test_binding (path/to/test_file::test_function or path/to/test_file::TestClass::test_method), and the bound test function must explicitly reference the INV-* ID.
5. Deterministic Engine Only (POL-004)¶
Eidos CLI commands are 100% deterministic local operations with zero external AI/LLM runtime calls. Agents always invoke Eidos CLI commands rather than guessing schemas or manually wiring symlinks.
6. Zero Unreconciled Drift¶
No task may conclude with failing specification schemas (eidos validate eidos) or unmanaged code-to-spec drift (eidos drift --uncommitted).
7. Mandatory Autonomous Adversarial Gap Self-Review¶
Before presenting work or concluding a task, the agent executes an autonomous adversarial review over its own git diff—auditing requirements completeness, edge cases, contract SemVer consistency (CTR-*), genuine test assertions (INV-*), and embedded decision capture (DEC-*)—and fixes any gaps immediately.
Stateless Stop Lifecycle Hook (eidos hook)¶
In agent harnesses that support lifecycle hooks (Google Antigravity, Claude Code, Cursor, Cline, OpenCode, and Codex), Eidos registers a single stateless Stop hook in hooks.json:
eidos-drift-gate(eidos hook):- Stateless Unified Check: When an agent finishes its task (
Stopevent),eidos hookrunsdetect_drift(uncommitted=True), which verifies living specification schemas and cross-references (SpecValidator), invariant test bindings (INV-*), and spec-to-code synchronization in a single pass. - Non-Blocking Advisory Warning: Returns
{"decision": "allow"}when clean, or{"decision": "allow", "systemMessage": "⚠️ ..."}with exact remediation steps when validation errors, broken invariant bindings, or code drift are detected—requiring zero state files, zero per-edit subprocesses, and never trapping agents in infinite retry loops.
Bundled Skills for AI Agents¶
When using an agent client that supports skill bundles (such as Google Antigravity, Claude Code, Cursor, Cline, OpenCode, Codex, Aider, or GitHub Copilot), 4 cohesive workflow skills are available:
eidos: Master workflow orchestrator (Direct vs. Planned Mode), token-efficient architectural context slicing (eidos context), and continuous spec-to-code drift detection (eidos drift).eidos-spec: Unified authoring reference for Product domains (DOM-*), capabilities (CAP-*), personas (PER-*), Arc42/C4 architecture (CNT-*, inline and promotedCMP-*,CTR-*, embeddeddecisions:), Multi-Design Systems (DS-*, W3C Design Tokens,UIC-*,CLT-*,SCR-*,FLOW-*), and system policies/invariants (POL-*,INV-*with polyglotfile::functest bindings).eidos-plan: Technical feasibility spikes (EXP-*), the SpecDelta mutation lifecycle (DLT-*,create,verify,audit,apply,changelog), and 5-dimension Staff Systems Architect reviews.eidos-ops: Greenfield repository initialization (eidos init,eidos init --lite), systematic reverse-engineering of brownfield codebases into living specifications, and Living Web Portal operations (eidos portal serve,eidos portal build).
Plugin Lifecycle & Maintenance¶
Eidos provides a complete suite of commands to manage, inspect, and update the installed agent plugin:
Inspect Plugin Manifest¶
Display full metadata and configuration for the Eidos Agent Plugin:
Output:
╭──────────────────────── Eidos Agent Plugin Manifest ─────────────────────────╮
│ Plugin Identifier: eidos │
│ Version: 0.8.0 │
│ Description: Git-native, agent-agnostic development AI ecosystem and │
│ lifecycle orchestration framework. │
│ License: MIT │
│ Specification: agent-plugins.org v1.0.0 │
│ Skills: eidos, eidos-ops, eidos-plan, eidos-spec │
│ MCP Servers: None │
│ Lifecycle Hooks: eidos-drift-gate │
│ Extensions: com.google.antigravity │
╰──────────────────────────────────────────────────────────────────────────────╯
Inspect Version Status¶
Compare the currently installed plugin against the version bundled with the Eidos engine:
Upgrade the Installed Plugin¶
When upgrading eidos-ai, update the project's installed plugin files in-place:
Verify Conformance¶
Ensure that your plugin bundle strictly complies with the agent-plugins.org v1.0.0 specification:
Output:
Agent Plugin Conformance (agent-plugins.org v1.0.0)
┏━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Property ┃ Value ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ Plugin Name │ eidos │
│ Plugin Version │ 0.8.0 │
│ Overall Status │ CONFORMANT │
│ Discovered Skills │ eidos, eidos-ops, eidos-plan, eidos-spec │
│ Configured MCP Servers │ None │
│ Lifecycle Hooks │ eidos-drift-gate │
│ Client Extensions │ com.google.antigravity │
└────────────────────────┴─────────────────────────────────────────────────────┘
[!NOTE] Under
agent-plugins.org v1.0.0, MCP servers are strictly optional. Eidos plugins decommissioned local stdio MCP servers in favor of direct CLI execution with machine-readable--jsonoutput, providing superior reliability in cloud, sandbox, and container environments.
Uninstall / Clean Up¶
To cleanly remove plugin files and agent configurations: