@jetrabbits/agentic

Agent Intelligence Configuration (agentic)

18 areas · 10 Software specs · 8 DevOps specs · 7 SDLC agents + 2 specialists · 105+ skills · 73+ workflows

A unified catalog of agentic specializations and the agentic CLI. Install orchestrator-ready rules, skills, workflows, and prompts into any project — and run a full SDLC agent team out of the box.

• Website
• Coverage scorecard
• CLI usage guide
• JEV native hooks and real CLI testing
• Claude Code integration
• Telegram notifications
• JEV research and opt-in CLI prototype
• Releases (prebuilt binaries)
• NPM package

Coverage snapshot

Coverage report for the current catalog across areas, specs, skills, and workflows.

Why agent-guides is different

Most agent prompt libraries give you a smarter single agent. agent-guides gives you a coordinated team:

@product-owner → @team-lead → @developer → @qa → (@devops-engineer) → post-task review

@pm (delivery planning) and @analyst (system analysis) join when the change needs them; @devops-engineer is an optional stage before the read-only post-task review by @instruction-reviewer and @memory-curator.

Each role has hard boundaries, explicit handoff criteria, and measurable success metrics. The Product Owner orchestrates the full SDLC; no other agent library ships this out of the box.

Works with Claude Code, opencode, Cursor, Codex, kilocode, antigravity and Gemini.

What this is

agent-guides is a structured knowledge base for AI coding agents. Instead of writing long system prompts, you install focused, composable guidance files that agents load on demand based on the task at hand.

Each area (e.g., devops/kubernetes) provides:

File type	Purpose	When loaded
AGENTS.md	Navigation index, load order, constraints	First — always
rules/*.md	Hard constraints the agent must follow	Always, for the active spec
skills/*/SKILL.md	Technical capabilities and patterns	On demand, matching "When to load"
workflows/*.md	Orchestrated step-by-step processes	On /command trigger
prompts/*.md	Human copy-paste templates (EN + RU)	Reference / paste

Repository structure

agent-guides/
├── areas/
│   ├── software/              # Application development
│   │   ├── general/           # Shared baseline — all software specs inherit from here
│   │   ├── backend/           # REST/GraphQL APIs, services, database access
│   │   ├── frontend/          # Components, accessibility, performance, CSS
│   │   ├── full-stack/        # End-to-end product features
│   │   ├── data-engineering/  # dbt, warehouses, pipelines, streaming
│   │   ├── mlops/             # Training, evaluation, serving, monitoring
│   │   ├── mobile/            # iOS, Android, React Native, Flutter
│   │   ├── platform/          # K8s manifests, IaC, CI/CD, secrets
│   │   ├── qa/                # Test strategy, coverage, flakiness, performance
│   │   └── security/          # Threat modeling, SAST/DAST, auth, compliance
│   └── devops/                # Platform and operations
│       ├── kubernetes/        # Cluster bootstrap, workload ops, RBAC, upgrades
│       ├── ci-cd/             # GitHub Actions, GitLab CI, quality gates, supply chain
│       ├── infrastructure/    # Terraform, Ansible, IaC standards, drift detection
│       ├── observability/     # Prometheus, Loki, Tempo, Grafana, SLOs
│       ├── sre/               # SLOs, error budgets, incidents, chaos engineering
│       ├── networking/        # Ingress, TLS, service mesh, DNS, VPC
│       ├── devsecops/         # Shift-left, SBOM, OPA/Kyverno, container hardening
│       └── database-ops/      # PostgreSQL, Redis, migrations, backup/restore
├── extensions/
│   ├── opencode/              # OpenCode agent definitions, commands, skills
│   │   └── agents/            # SDLC agents + post-task specialists for .opencode/agents/
│   ├── claude/                # Claude Code configs
│   │   └── agents/            # SDLC agents + post-task specialists for .claude/agents/
│   ├── antigravity/           # Antigravity platform configs
│   ├── codex/                 # Codex custom agents and override configs
│   │   └── agents/            # SDLC agents + post-task specialists for .codex/agents/
│   └── gemini/                # Gemini-specific configs
│   │   └── agents/            # SDLC agents + post-task specialists for .gemini/agents/
├── areas/template/            # Authoring templates — start here for new content
├── docs/                      # Setup and usage guides
├── AGENTS.md                  # Root agent guidance (loaded into every project)
└── src/                       # Rust CLI installer (self-contained binary)

Quick start

Requirements

agentic is a single self-contained binary — no runtime dependencies. The knowledge base (areas, extensions) is embedded into the binary.

• required: nothing (the binary works standalone)
• optional: python3 + pip (only for the MemPalace integration), agent binaries such as codex, opencode, claude, gemini (for doctor smoke checks and actual agent work)

Install

Via npm (thin launcher that fetches the platform binary on first run):

npx @jetrabbits/agentic@latest

Or bootstrap the latest release binary (linux x86_64/aarch64, macOS Intel/ARM):

curl -fsSL https://raw.githubusercontent.com/sawrus/agent-guides/main/install | bash

Or download a binary for your platform from GitHub Releases and run:

./agentic self-install --force

Build from source (Rust toolchain required):

make release-build && make install

Run

agentic

Upgrade

agentic upgrade
agentic upgrade --project-dir /path/to/project
agentic upgrade --force --dry-run
agentic upgrade --force

upgrade downloads the newest release binary from GitHub Releases and atomically replaces the installed one, then synchronizes project instructions using the new executable. Recognizable legacy artifacts are adopted; non-overlapping local edits are merged. Conflicts are replaced after backup, and fully user-owned instructions are preserved. --force rebuilds Agentic artifacts without prompts, using manifest settings or default + software.general when settings are unavailable. Unrelated project files and global configuration are preserved. Backups live in .agentic-backups/ and may contain credentials; exclude them from source control. See smart upgrade behavior for details.

Full instructions

• CLI usage guide
• Installed CLI lifecycle

See agentic in action

Interactive walkthrough of the agentic CLI flow: choose a target project, pick agent platforms, and install the guidance bundle.

SDLC Agent team

The same 7-agent SDLC team works across Claude Code, OpenCode, Codex, and any tool that supports agent or subagent files. Agentic also ships post-task review specialists for instruction quality and docs/memory hygiene, automatic for hooked workflows and optional elsewhere.

Agent	Role	Invoke when
product-owner	Scope, acceptance criteria, SDLC orchestration	Start of any feature; final acceptance
pm	Planning, milestones, risk register	Scope is defined, execution needs tracking
team-lead	Architecture, code review, quality gates	Planning and pre-release sign-off
developer	Implementation, tests, delivery	Implementation plan is approved
qa	Verification, defect classification, go/no-go	Developer hands off an increment
analyst	Requirements, scenarios, API/data contracts, UX states	Planning; requirement and contract review
devops-engineer	CI/CD, IaC, platform reliability	Anything touching infra, pipelines, or deploy

Each agent has an Identity, Communication Style, Success Metrics, and explicit Boundaries (OpenCode agents also carry a one-line vibe), so roles never overlap and handoffs are always documented. In Claude Code the product-owner runs as the main session because subagents cannot start other subagents.

Post-task specialists run outside the mandatory SDLC role matrix. The six delivery workflows listed in the Review Pipeline invoke both automatically after successful top-level delivery:

Agent	Role	Invoke when
instruction-reviewer	Post-task instruction effectiveness review	Successful delivery workflow with a review hook
memory-curator	Post-task docs and memory hygiene recommendations	Successful delivery workflow with a review hook

See Review Pipeline for the guidance-mode pipeline and .reviews/<task-id>/ output convention. The installed REVIEW_PIPELINE.md is loaded only at review time. Reports are compact, evidence-backed, and advisory; no instructions, docs, or memory are changed automatically.

Platform	Agent path	Format	Guide
Claude Code	project/.claude/agents/*.md	Markdown with YAML frontmatter	Claude Code subagents · setup
OpenCode	project/.opencode/agents/*.md	Markdown with frontmatter	OpenCode agents · repo setup note
Codex	project/.codex/agents/*.toml	TOML custom agents	Codex subagents
Gemini	project/.gemini/agents/*.md	Markdown with YAML frontmatter	Gemini subagents

JEV: native decision hooks

JEV is an optional, paid decision service (TypeSafe AI) that runs inside each host's native hooks. It does not add an MCP server or a workflow runner.

• Workflow selection. On the first prompt it picks one of the installed .agent/workflows/*.md and tells the host to execute it as the declared initiator.
• Model routing. Before a subagent starts, it picks the cheapest sufficient model from the role defaults: Codex model ids, OpenCode catalog models, Claude opus/sonnet/haiku aliases. Explicit pins are preserved.
• Permission decisions. allow / deny / ask, with automatic approval limited to bounded project paths.
• Review packet. A task-scoped .agentic/jev/review-*.json packet goes to instruction-reviewer.
• Safety and audit. Errors, timeouts (8 s) and confidence below 0.9 keep the native behavior. Cleansed audit records are kept 30 days in .agentic/jev/.

agentic install --project-dir <dir> --agent-os claude,codex,opencode --areas software \
  --specializations software.backend,software.general --enable-jev   # TYPESAFE_API_KEY or JEV_API_KEY in env/.env
make blackbox-claude   # paid real-CLI check; also blackbox-codex, blackbox-opencode, blackbox (all)

Feature support by platform

✅ supported and verified in a real CLI blackbox run · ☑️ supported, covered by offline tests only · ➖ not supported

Feature	OpenCode	Codex	Claude Code	Gemini	Cursor · Kilocode · Antigravity
SDLC role subagents (9 roles incl. analyst, reviewers)	✅	✅	✅	☑️	➖
Rules (auto-loaded)	☑️	☑️ ¹	✅	☑️ ¹	☑️
Skills	☑️	☑️ ¹	✅	☑️ ¹	☑️
Workflows as commands	✅	✅ ¹	✅	☑️ ¹	☑️ ¹
MCP servers (Context7, MemPalace, Playwright, …)	☑️	☑️	☑️	☑️	☑️ (Cursor, Kilocode)
Memory guidance (MEMORY.md, MemPalace)	☑️	☑️	✅	☑️	☑️
Post-task review (instruction-reviewer + memory-curator)	✅	✅	✅	☑️	➖
Model profiles per role	☑️ (openai, githubcopilot, user)	☑️ (TOML defaults)	☑️ (anthropic)	➖	➖
Telegram: task finished	✅ (plugin) ²	✅ (hooks)	✅ (hooks)	➖	➖
Telegram: waiting for input	✅ (plugin) ²	☑️ (hooks) ²	☑️ (hooks) ²	➖	➖
JEV · workflow selection	✅	✅	✅	➖	➖
JEV · subagent model routing	✅	✅	✅	➖	➖
JEV · permission decisions	☑️	☑️	☑️	➖	➖
JEV · reviewer evidence packet	☑️	✅	✅	➖	➖
Real-CLI blackbox target	make blackbox-opencode	make blackbox-codex	make blackbox-claude	➖	➖

¹ Through the shared .agent/ directory that AGENTS.md tells the agent to read, not a native discovery path. ² Real CLI, real Telegram bot, outside the blackbox target (2026-10-16). OpenCode V2 delivered both events; the Claude/Codex waiting messages were sent through the same script with recorded hook payloads.

Real runs:

• Claude Code 2.1.301, Agentic 1.4.0, 2026-10-16: all ten checks passed (.agentic-blackbox/20261016-165654). JEV permission evaluations fell back below confidence 0.9 in that run.
• Claude Code 2.1.302, Agentic 1.4.0, 2026-10-17, opus session: all ten checks passed (.agentic-blackbox/20261017-232709) after the hand-back and profile-fallback fixes. JEV routed exploration to haiku; the other spawns fell back to the Claude profile role models (developer/qa → sonnet, team-lead → opus, reviewers → haiku) instead of inheriting opus. See verification.
• OpenCode 2.0.29 (openai/gpt-6-luna), Agentic 1.4.0, 2026-10-16 (.agentic-blackbox/20261016-233233): JEV selected the workflow, and the agent ran PO → pm → team-lead → developer → qa → instruction-reviewer + memory-curator with the new role names. An independent probe verified the calculator. JEV answered no_decision for all ten spawns, so subagent model routing was not re-confirmed (it was in the 1.3.0 run).
• Codex 0.163.0, Agentic 1.4.0, 2026-10-17 (.agentic-blackbox/20261016-232317): all six checks passed. JEV selected the workflow and applied gpt-6-luna to 8 of 9 spawns (PO, pm, team-lead, developer, qa, both reviewers); the calculator was verified independently, and the Telegram Stop hook delivered. The run used BLACKBOX_CODEX_SANDBOX=danger-full-access, because this host (kernel.apparmor_restrict_unprivileged_userns=1) blocks the bubblewrap workspace-write sandbox.
• Codex 0.161.0 and OpenCode V2, Agentic 1.3.0, 2026-10-14: model routing and calculator checks passed.

Details: JEV verification.

What gets installed where

agentic installs shared guidance into .agent/ and selected IDE files into directories like .claude/, .opencode/, and .codex/. Generated runtime guidance goes to AGENTS.md; OpenCode also gets .opencode/AGENTS.md.

project/.agent/
├── rules/       ← constraints active for every task in this project
├── skills/      ← capabilities loaded on demand
├── workflows/   ← processes triggered by slash commands
└── prompts/     ← human-paste templates

project/.claude/                 ← with --agent-os claude
├── CLAUDE.md    ← @AGENTS.md / @MEMORY.md imports
├── agents/      ← 9 role subagents (tools + model per profile)
├── rules/       ← auto-loaded rules
├── skills/      ← native skills
├── commands/    ← one /<workflow> wrapper per .agent/workflows/*.md
└── settings.json← agent: product-owner, MCP approvals, JEV/Telegram hooks (ownership-tracked)

Agent IDE options (interactive install)

MCP

• opencode-docs (safe): adds OpenCode documentation through a local MCP server.
• playwright (sensitive): adds browser automation through Playwright MCP.
• kubernetes (dangerous): adds Kubernetes cluster access through Kubernetes MCP.
• youtube-transcript (safe): adds YouTube transcript lookup through a local MCP server.
• docker-mcp (dangerous): adds Docker MCP Gateway as the generated docker MCP server.
• context7 (safe): adds a remote MCP server for up-to-date framework, library, SDK, and API documentation.
• mempalace (sensitive): adds a local memory MCP server for project context discovery and reuse.
• anydb (dangerous): adds database access through AnyDB MCP.

Telegram notifications (OpenCode, Claude Code, Codex)

• Every host sends through the shared .agent/notify/telegram.py: OpenCode through its V2 plugin, Claude Code and Codex through native hooks. Messages arrive when the agent finishes and when it waits for input (permission or question).
• Agent Markdown is rendered as Telegram formatting: bold, code, links, lists, tables as monospace blocks. Long answers arrive as a .md attachment with a preview.
• Enable with --enable-telegram or the interactive prompt; credentials stay in ~/.config/agentic/config.json. See Telegram notifications.

OpenCode Plugins

• Agent Model Mapping: maps .opencode/agents/*.md roles to main and fallback OpenCode models during interactive agentic install/agentic tui. Stored internally as agent-model-mapper; OpenCode startup never prompts or writes project files.

OpenCode Model Profiles

• OpenAI Model Profile: applies the bundled OpenAI model mapping from extensions/opencode/profiles/openai/.
• GitHub Copilot Model Profile: applies the bundled GitHub Copilot model mapping from extensions/opencode/profiles/githubcopilot/.
• User profiles: place opencode.json files under $HOME/.config/agentic/opencode/profiles/<profile-id>/. For example, $HOME/.config/agentic/opencode/profiles/DT/opencode.json appears in the optional OpenCode plugin menu as DT profile, and $HOME/.config/agentic/opencode/profiles/GH/opencode.json appears as GH profile.
• none: applies no model profile and does not copy the baseline extensions/opencode/opencode.json just for profile selection. OpenCode MCPs, Telegram notifications, and model mapping may still create or update .opencode/opencode.json when those options are selected.

Claude Code Model Profiles

• anthropic (default): opus for product-owner and team-lead, sonnet for developer, qa, devops-engineer and analyst, haiku for pm and both reviewers. Written as model: in .claude/agents/*.md.
• none (--claude-profile none): every role inherits the session model.
• With JEV enabled, models stay open so JEV hooks route each Agent call to an alias.

Contributing

See CONTRIBUTING.md for authoring standards, templates, and the pull request process.

Quick checklist before opening a PR:

• Used the appropriate template from areas/template/.
• No {{PLACEHOLDER}} values remain.
• Spec map in AGENTS.md updated.
• Constraints in rules use imperative language.
• Prompt examples include both EN and RU blocks.

License

See LICENSE if present, or contact the maintainers.