@jetrabbits/agentic

agentic — shared rules and a quality gate for every coding agent

One set of project rules for Claude Code, Codex and OpenCode. A quality gate runs your checks and records acceptance. The lean default adds no role team; helper agents, workflows and model routing are optional.

Website · Releases · npm · Usage guide · Reference

The problem it solves

Teams now use several coding agents at once. Each agent gets its own prompts, its own habits and its own idea of "done". One agent runs the tests, another skips them. One agent asks before it touches production, another does not.

agentic gives every agent the same base:

• 📏 Rules are the base. 64 rules for 16 specializations (8 software, 8 DevOps). Each rule lists items that a reviewer can check in a diff. A rule loads only for the files it applies to (paths:) or on request.
• ✅ A gate decides "done". When the agent stops, the quality gate runs make test and make lint (or your own commands) and asks for a CHANGELOG line. It records the result in .agentic/gate/last.json.
• 🪶 Lean by default. The native main session does the work, with one read-only explorer helper. The always-on context is at most 2,110 tokens on any host. --profile team installs the optional role team.
• 🔁 Install once, keep in sync. agentic install writes the native files of each agent. agentic upgrade keeps them current and keeps your own edits.

14 skills and 10 workflows help the agent meet the rules. Numbers come from the tree (areas/*/*/) and from python3 scripts/context_budget.py --all-selections.

Runs on the agents you already use

	Agent	Status	Evidence
✅	Codex	Real runs	codex-cli 0.164 · benchmark round 6, 2026-10-27
✅	Claude Code	Real runs	sonnet · benchmark round 6, 2026-10-27
✅	OpenCode	Real runs	2.0.29 · benchmark round 4, 2026-10-24
☑️	Antigravity	Configs ship	Agents, rules and workflows are installed; no recorded run yet
☑️	Gemini	Configs ship	Agents and workflows are installed; no recorded run yet
☑️	Cursor · Kilocode	Configs ship	Rules and MCP settings are installed; no recorded run yet

"Real runs" means that the agent's own CLI ran the installed instructions and a separate grader checked the result. Details per feature: platform matrix.

See agentic in action

A real agentic tui session, recorded key by key. It selects Codex, Claude Code and OpenCode with the lean profile, Context7, DevOps infrastructure plus backend, and a company rules directory. Then the same request runs in the main session of all three agents inside the new project.

Quick start

npx @jetrabbits/agentic@latest          # interactive installer

Or non-interactive, for all three verified agents:

npx @jetrabbits/agentic@latest install --project-dir . \
  --agent-os codex,claude,opencode --areas software \
  --specializations software.backend,software.general

Keep it current:

agentic upgrade            # new binary + refreshed project instructions, your edits preserved

Other install options (curl script, release binary, build from source): usage guide.

Guardrails

The quality gate (.agent/gate/gate.py) runs from native hooks in Claude Code and Codex and from a plugin in OpenCode. Details: quality gate.

• Acceptance record. .agentic/gate/last.json stores accepted, reason_code, head, pack_version and profile for the last stop.
• Production guard. kubectl, helm or oc on a context that matches prod asks you first. You can add database hosts and your own patterns in .agent/gate.json.
• Database guard. Write or maintenance SQL on a non-local database is blocked. A connection with credentials from .env asks first. In host plan mode, every non-local database connection is blocked.
• Secret guard. The gate blocks access to .env* files and private keys, git push --force, and recursive deletes outside the project.
• Stale checkout. Install, upgrade and the session start warn when the checkout is older than the default branch on origin.

Company rules

Keep internal standards outside this public pack. --company-rules-path <dir> or --company-rules-git <url>[#ref][:subdir] (or the TUI step "Company rules") installs them for every host, and agentic upgrade replays them. A company rule wins over a pack rule on the same topic. replaces: or --company-rules-exclude drops the pack rule, and a duplicate report lists the overlaps. See company rules.

Standards for the pack itself

• Artifact standard: the acceptance contract for rules, skills, workflows and company rules (ART-* ids). make check-artifacts checks every automatic requirement.
• Review regulation: the Definition of Done for code review and design review, and the docs coverage rule. make check-docs-coverage BASE=<ref> fails when changed code has no changed docs page.

Your AI team

The default lean profile installs one helper: explorer, a read-only agent for code search. The main session runs the workflow steps itself.

--profile team is optional. It installs the role team: product-owner, pm, team-lead, developer, qa, analyst, devops-engineer, instruction-reviewer and memory-curator. A workflow then sends each step to the role that fits it. See role profiles.

Honest numbers

We measure instead of promising. Round 6 (2026-10-27) ran Claude Code with sonnet and Codex with gpt-6.1-terra (codex-cli 0.164 or later). Codex cost is an API-equivalent estimate.

Check	Result
Devops task T11 (Kubernetes rules), Claude Code	plain agent 0/3 · lean 3/3
Core tasks T1–T7 and T10, Claude Code lean, before and after the content cleanup	8/8 and 8/8
T11 lean after the content cleanup	uncached tokens × 0.72 · cost × 0.85
Lean helpers with JEV routing, Codex T6–T8	9/9 both · every helper on the session model · cost × 1.09 of lean
reviewer helper, T8v2 (hidden defect)	Codex 9/9 with and without it · it never ran · removed from lean

• The rules carry the devops result: every plain T11 run missed the Kubernetes security items.
• The cleanup kept quality, but full tokens fell less than the 10% target, because cached session history dominates them.
• JEV stays opt-in. It gives no measured quality gain.

All numbers and earlier rounds: round 6 results · comparison.

Learn more

• Usage guide and installed CLI lifecycle
• Reference: layout, platform matrix, install options
• Docs site: rules, skills and workflows per specialization
• Artifact standard · Review regulation · Company rules
• Claude Code integration · Telegram notifications
• Review pipeline · JEV hooks and real CLI testing
• Smart upgrade

Contributing

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

License

See LICENSE if present, or contact the maintainers.