agentic-os

Turn your repo into a governed multi-agent setup in one interview — scoped agent contracts, enforcement hooks (not just prompts), and optional SDLC orchestration. Works in Claude Code and Cursor; the SDLC pipeline (agentic-sdlc) also runs in Codex.

In plain words

You install two plugins into your AI coding editor (Claude Code or Cursor). The first, agentic-os, sets up guardrails for AI agents in your project: what files they may touch, when they must stop and ask a human, and how their work gets independently reviewed before anything is committed. The second, agentic-sdlc, runs a software-delivery pipeline on top of those guardrails: it takes an idea or a ticket and carries it through requirements, a spec, a plan, tested code, and a review-ready pull request. You don't need to be a programmer to benefit — there are role setups for developers, QA, architects, DevOps, business analysts, and project/portfolio managers, and the non-coding roles work entirely through plain chat phrases. Unfamiliar terms (HITL, gate, preset…) are one-liners in the Glossary.

Start here

New to this repo? You do not need to read everything below. Pick a path:

Five-minute flow (any editor):

1. Install the plugin(s) for your editor (Claude Code or Cursor).

2. Open the git repo you want to improve (not this marketplace repo).

3. Run /agentic-init --defaults (or /agentic-init for the full interview).

4. Run /agentic-doctor — expect passed: true in .agentic/agentic-os/doctor.json.

5. git status — review scaffolded files; nothing is committed for you.

6. If you use the SDLC pipeline, run sdlc-doctor — expect passed: true in .agentic/agentic-sdlc/doctor.json.

Safe first try: use the throwaway repo walkthrough before touching a real project.

Full install checklist

Complete this before /agentic-init in the repo you want to equip:

After equip — which doctor?

Cursor users: this plugin is not in Browse Marketplace → All (curated public plugins only). You add a custom marketplace and install from the User tab — see Cursor install.

Related MCP server: CodeAudit MCP

Why this exists

Left alone, multi-agent coding fails in a few specific, recurring ways: an agent's write scope silently creeps into files it wasn't asked to touch; an agent reviewing its own work rationalizes away the gaps it should have caught; "autonomous mode" is a blunt on/off switch instead of a resolution strategy sized to each decision's actual risk; and stack facts get re-guessed from scratch every session, with nothing stopping a Postgres-flavored rule from leaking into a MongoDB agent's contract. agentic-os closes each of those with a mechanism that's enforced, not just requested in a prompt — see docs/PRINCIPLES.md for what each one does and why a plain agent session doesn't already have it.

What it is

agentic-os packages that governance layer as an installable system. One command interviews you (role, autonomy level, stack) and scaffolds a coherent system into your project: canonical agent contracts, enforcement hooks (blind pre-commit review, write-scope guard, instruction-quality spawn gate), a human-in-the-loop escalation ladder, and — where it fits your stack — generated stack-specific agents. It is stack-universal: six curated profiles (Next.js, Django, Spring, Rails, Go, Playwright) are recognized instantly, and anything else — a NestJS backend, a schemaless Mongo service, or no code stack at all — gets real, evidence-grounded persistence and API agents via live repo discovery instead of a degraded stub (proven live against both a migration-managed and a schemaless non-curated backend — see tests/universal/README.md). Frontend/UI generation on a non-curated stack uses the same discovery mechanism but is earlier in its own verification cycle. It is also role-agnostic (developer, QA, BA/PO, architect, delivery).

Today it's delivered as a Claude Code and Cursor plugin (the agentic-sdlc pipeline additionally ships Codex packaging) — that's the install mechanism, not the architectural boundary: the scaffolded contracts are harness-neutral by design (thin per-host pointer files over a canonical, host-independent body).

This repo is a marketplace (Claude Code and Cursor) hosting three plugins:

• agentic-os — the product: the /agentic-init, /agentic-doctor, and /agentic-upgrade skills plus the template library, generators, and role presets they scaffold from.

• agentic-sdlc — the SDLC pipeline and the decision-router that resolves every judgment gate (ask the human in HITL mode; deterministic → fast-path → stand-in subagent → escalate in autonomous mode) with a full decisions.jsonl audit trail. For an interactive visual map of the whole pipeline, open plugins/agentic-sdlc/sdlc.html in a browser.

• agentic-qe — a standalone, tool-agnostic catalog of Quality Engineering AI blueprints (28, organized by STLC stage) plus two skills: qe-blueprints scaffolds a ready-to-fill agent framework from a chosen blueprint, and eval-harness generates a two-layer evaluation framework for skills and agents. Independent of the governance flow — no /agentic-init required.

There's also mcp/ — a read-only MCP server that serves this same governance/SDLC/QE methodology to hosts that don't speak the Claude Code or Cursor plugin format (any MCP-capable client). It's a separate package with its own build, published on npm and the MCP Registry; see mcp/README.md.

Use it from any MCP client

Don't use Claude Code or Cursor? The same methodology is a published MCP server — your assistant can search the docs, list the role presets and QE blueprints, turn a role into an install plan, and audit an existing install, all read-only:

claude mcp add agentic-os -- npx -y agentic-os-mcp

Works in any MCP-capable assistant. See mcp/README.md for per-role examples and the full tool list.

Which plugin? For the governed platform, install agentic-os + agentic-sdlc plus superpowers: agentic-os is the installer/governance layer, agentic-sdlc is the SDLC orchestrator, and /agentic-init wires them together and registers missing dependencies in Phase 3 if any are absent. agentic-qe is optional and stands alone — install it if you want the QE blueprint catalog and scaffolder, with or without the other two.

Why separate plugins, not one? The governance layer and the pipeline have different lifecycles and different audiences: agentic-os scaffolds into your repo once and evolves with your codebase, while agentic-sdlc is a set of runtime skills you invoke per task — and some teams want the guardrails without the pipeline (or, on Codex, the pipeline alone). agentic-qe is separate again: a knowledge-and-scaffolding add-on with no dependency on the other two. Splitting them keeps each independently versioned, upgraded, and installable.

/agentic-init Phase 3 treats agentic-sdlc and superpowers as non-optional (plugins/agentic-os/manifest/dependencies.json). Install all three up front to avoid a pending-restart loop.

Prerequisites

• Claude Code or Cursor — see Install. This is a custom plugin marketplace; you add it once per editor, then equip projects with /agentic-init.

• python3 on your PATH — enforcement hooks are Python scripts.

• git — the target repo must be a git repository (/agentic-init can git init if it is not).

• node on your PATH — required when using agentic-sdlc; sdlc-doctor runs node --version and git --version (plugins/agentic-sdlc/skills/sdlc-doctor/SKILL.md).

• superpowers ≥ 6.1.0 and agentic-sdlc ≥ 0.4.4 — required by /agentic-init (plugins/agentic-os/manifest/dependencies.json).

• Optional: gh (GitHub CLI) for GitHub ticket/MR adapters.

Install

Install superpowers

/agentic-init and agentic-sdlc require the superpowers plugin ≥ 6.1.0 (plugins/agentic-os/manifest/dependencies.json). Install it before equipping a target repo.

Claude Code:

/plugin install superpowers@claude-plugins-official

If that marketplace is missing, add it first:

/plugin marketplace add anthropics/claude-plugins-official
/plugin install superpowers@claude-plugins-official

Fallback marketplace (same plugin family):

/plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers@superpowers-marketplace

Cursor:

1. Customize → Plugins → Browse Marketplace → All → install Superpowers (curated public plugin), or

2. Install superpowers in Claude Code and enable “Automatically import agent configs from other tools”, then reload Cursor.

Confirm version ≥ 6.1.0 on the plugin card or via sdlc-doctor after equip.

Claude Code

Install superpowers (above), then add this marketplace:

/plugin marketplace add Jarroslav/agentic-os
/plugin install agentic-os@agentic-os
/plugin install agentic-sdlc@agentic-os

Or try it straight from a local clone, no publish needed:

git clone https://github.com/Jarroslav/agentic-os
# in Claude Code:
/plugin marketplace add /absolute/path/to/agentic-os
/plugin install agentic-os@agentic-os
/plugin install agentic-sdlc@agentic-os

Restart the session so the plugins load (Claude Code activates plugins at session start).

Cursor

agentic-os is a custom marketplace, not a curated Cursor Store listing. You will not find it by searching Browse Marketplace → All.

Install (recommended)

1. Open Customize → Plugins (or Settings → Plugins; use Customize if you see the “Plugins are moving to Customize” banner).

2. Add marketplace and paste the Git clone URL (must end in .git):

   https://github.com/Jarroslav/agentic-os.git

   Or, from a local clone:

   /absolute/path/to/agentic-os

3. Open the User tab (not All).

4. Install both plugins from the agentic-os marketplace:

   • agentic-os — /agentic-init, /agentic-doctor, /agentic-upgrade

   • agentic-sdlc — SDLC skills + subagents (display name Agentic SDLC)

5. Install superpowers ≥ 6.1.0 (Install superpowers — curated Superpowers from Browse Marketplace → All works in Cursor).

6. Reload the window (Command Palette → “Developer: Reload Window”).

You should see each plugin as its own card (often tagged Imported), with bundled skills listed — similar to other third-party plugins. The page Rules, Skills, Subagents is a flat list of everything on your machine; that is not the plugin install view.

Alternative: import via Claude Code

If you already use Claude Code with “Automatically import agent configs from other tools” enabled in Cursor:

/plugin marketplace add Jarroslav/agentic-os
/plugin install agentic-os@agentic-os
/plugin install agentic-sdlc@agentic-os

Reload Cursor. Plugins registered in Claude appear as Imported in Cursor.

Equip your project

Open the target repo (your app, not the agentic-os clone), then:

/agentic-init            # full interview (six screens, pre-filled from stack discovery)
/agentic-init --defaults # accept detected defaults, no prompts

The interview never commits — it scaffolds files, shows a settings diff before merging, and leaves the working tree for you to review.

Then:

/agentic-doctor          # verify governance install → .agentic/agentic-os/doctor.json
/agentic-upgrade         # reconcile after a plugin version bump

If you use SDLC skills (/sdlc-start, /sdlc-autonomous, …), also run sdlc-doctor — it writes .agentic/agentic-sdlc/doctor.json and checks superpowers, node, and git.

Cursor troubleshooting

Cursor reads .cursor-plugin/marketplace.json from the repo root. Use a Git clone URL ending in .git, not the GitHub browser URL.

Try it in two minutes (throwaway repo)

Nothing here touches a real project — build a disposable repo and watch the whole install cycle end-to-end:

mkdir /tmp/try-agentic && cd /tmp/try-agentic && git init
printf '{"name":"try","dependencies":{"next":"15.0.0"}}' > package.json

Open Claude Code or Cursor in that directory (after the marketplace/plugin install steps and a session reload), then run:

/agentic-init --defaults

What happens, in order:

1. Preflight — detects this is a git repo, then runs stack discovery: a cheap marker check spots the next dependency and matches the nextjs-supabase profile, then a subagent confirms that match against the real repo (a non-matching repo gets a full from-scratch inspection instead, not a dead-end fallback).

2. Interview — with --defaults, all six screens (role preset, HITL dial, autonomy matrix, gates, stack confirm, ticket/MR adapter) are accepted at their detected defaults instead of prompted.

3. Dependency check — verifies every non-optional dependency (agentic-sdlc, superpowers) is registered; prints a pending-restart notice for any that aren't.

4. Scaffold — writes .agentic/agents/, .agentic/guides/, .claude/hooks/, .githooks/pre-commit; always writes CLAUDE.md as a marker-delimited block, and writes AGENTS.md whole on a fresh repo like this one (it only becomes a marker-delimited block when the file already exists). Nothing is committed — it's your working tree to review.

5. Generate — spawns per-slot subagents for the generated set (the union across every selected role preset): writer contracts, any applicable read-only gate like migration-validator, and stack guides — each independently audited against the instruction-quality rubric before being armed in the scorecard.

Then:

/agentic-doctor

runs all 8 checks (file manifest vs. install journal, hook compilation, canned-event dry-runs of four enforcement hooks, a 3-part HITL smoke test, settings registration, git hook + dependencies, scorecard coverage/thresholds, and agent-registry table integrity) and writes the result to .agentic/agentic-os/doctor.json.

Cursor note: same commands work in Cursor chat once agentic-os is installed from the custom marketplace. If /agentic-init is missing, reload the window and confirm the plugin card appears under Customize → Plugins.

git status         # inspect exactly what was scaffolded; nothing was committed

Role presets

Presets are additive — install several and their template sets union (strictest HITL wins).

The HITL dial

The install sets how much agents may do before a human must weigh in:

• strict — every step is user-invoked; agents recommend, they don't act (e.g. QA: agents author tests but never run them). Dispatcher orchestration.

• gated-autonomous — pipelines run, but judgment gates and the escalate_on risk flags (default: security, breaking-change, migration, spend) stop them for a human decision.

• autonomous — the agentic-sdlc decision-router resolves gates with deterministic checks, fast-paths, and stand-in reviewers, escalating to you only on low confidence, a matching risk flag, or malformed agent output.

Underneath, three mechanisms make HITL real, not advisory:

1. Policy files (.agentic/guides/policy/) — the autonomy matrix, size ceiling, env write boundaries, secret deny-lists, and the escalation ladder.

2. Resolver conventions — every agent ends its output with ## Summary / ## Why / ## Blocking / ## Non-blocking / ## Escalate to human, parsed fail-closed by a Stop/SubagentStop hook: a non-empty Blocking stops the parent (no silent retry); a non-empty Escalate to human forces an AskUserQuestion.

3. Hard gates — exit-2 hooks: blind pre-commit review (sha256 stamp of the staged diff), write-scope guard, instruction-quality spawn gate, human-gated command blocks.

What gets scaffolded

Into your target repo (harness-neutral canonical contracts, thin Claude pointers):

.agentic/agents/            canonical agent contracts (single source of truth)
.agentic/guides/            policy/, standards/, agent-registry.md, project.md
.agentic/agentic-sdlc/      config.json (decision-router wiring)
.claude/hooks/             the enforcement hooks
.claude/agents/, commands/ thin pointers + orchestration commands
.githooks/pre-commit       the review gate's git-level twin
CLAUDE.md / AGENTS.md      a managed block (your content outside it is untouched)
docs/audits/               the instruction-quality scorecard

Mature repos are handled non-destructively: managed marker blocks, deep-merged settings (shown as a diff first), skip-by-default name collisions, chained (never replaced) git hooks.

Using it, by role

After install, what you reach for depends on the preset(s) you chose:

• developer — describe a feature; the pipeline orchestrator runs the staged flow (generated schema/api/component agents → read-only gates → blind pre-commit review). Every git commit is blocked until the staged diff is reviewed. /agentic-doctor confirms the fleet is spawnable.

• qa — /dispatch routes each request to one owning agent (strict HITL, one step at a time). Generate test-case drafts from a story, automate approved work-item IDs (with real-ID + existing-coverage gates), triage a red test against the flaky protocol. Agents author tests but never run them — they hand you the exact command.

• ba-po — draft stories and requirements through the agentic-sdlc intake skills, wired to your ticket adapter.

• architect — the governance layer: AGENTS.md/PATTERNS.md/agent-registry, the instruction-auditor + scorecard spawn gate, generated architecture guides.

• pm-delivery — ticket/MR adapters, the PR pipeline gate, MR-monitoring.

• devops — the delivery-infrastructure layer: chained git hooks, the quality-gates registry, the PR pipeline gate, and MR-monitoring that fixes red CI and reviewer comments until the MR merges.

• portfolio — oversight without touching code: ask for run status, audit a repo's docs/agent-setup health, keep durable notes across sessions, and turn stakeholder asks into normalized requirements.

If you don't write code (pm-delivery, ba-po, portfolio): everything you need is a plain chat phrase, not a CLI command. Type things like "draft a story for password reset", "what's the status of the current run?", "watch MR !123", or "turn this stakeholder email into requirements" directly into the editor's chat. The one honest prerequisite: someone installs Claude Code or Cursor with these plugins for you first — after that, no terminal is required.

Everything obeys the HITL dial you set: an agent that hits a ## Blocking item stops and surfaces it; an ## Escalate to human item forces a question before anything proceeds.

Why not just prompt the agent yourself?

You can, and for a one-off change you should. The difference shows up when the work is big enough to span several agents, several sessions, or a stack the agent has to learn rather than be told about:

The claim isn't that agents can't do this work, or that a plain session has no guardrails — it has good ones. It's that those guardrails are per session and per tool, and the failures above are per agent and per decision. Nothing in a plain session stops a specific agent from quietly doing a specific thing wrong. See docs/PRINCIPLES.md for the reasoning behind each row.

Why hooks, not prompts? A prompt is a request the model may weigh against everything else in context; a hook is a program the harness runs outside the model, whose non-zero exit physically blocks the tool call. An agent under context pressure can rationalize its way past "please don't commit unreviewed code" — it cannot rationalize its way past a PreToolUse hook that exits 2. That's why every load-bearing rule here (blind review before commit, write scope, human-gated commands, instruction freshness) is enforced by a hook, with prompts reserved for guidance (docs/PRINCIPLES.md).

FAQ

Does it ever commit or push on its own? Not unless you ask it to. /agentic-init scaffolds files and shows you a settings diff before merging it — it never runs git add or git commit, so the working tree is yours to review. The SDLC pipeline stops at a review-ready branch and never opens a PR by itself. Two bundled skills do write to git, and only when you invoke them by name: mr-creator (commits, pushes, opens the PR) and mr-watch (pushes review fix-ups with --force-with-lease).

What if my stack isn't one of the six curated profiles? Then discovery inspects the repo from scratch instead of matching a profile, and you still get real, evidence-grounded agents — not a degraded stub. This is verified against non-curated fixtures spanning both persistence paradigms and both UI paradigms; see tests/universal/README.md.

What if I have no code stack at all? That's a first-class path by design, not a degradation: a pm-delivery or ba-po install declares generated: [] and wires ticket/MR adapters plus the governance layer instead of code agents. Honest caveat — the preset combinatorics are covered deterministically in CI, but this zero-capability path hasn't yet been driven through a live end-to-end /agentic-init run (tracked in ROADMAP.md).

Will it fight my existing CI, hooks, or CLAUDE.md? No. Mature repos are handled non-destructively: CLAUDE.md/AGENTS.md get a marker-delimited managed block (your content outside it is never touched), settings are deep-merged after showing you a diff, name collisions skip by default, and a pre-existing git hook is chained, not replaced.

Does this work outside Claude Code? The canonical agent contracts are harness-neutral by design — the host-specific files are thin pointers over them. Install via Claude Code or Cursor; on Codex, the agentic-sdlc pipeline installs and runs, but the agentic-os governance installer has no Codex packaging yet (tracked in ROADMAP.md; see INSTALL.md). Claude Code is the best-supported enforcement host today; Cursor packaging reuses the same skills and scaffolds the same .agentic/ layer.

What does /agentic-doctor actually check? Eight things: file manifest vs. install journal, hook compilation, canned-event dry-runs of four enforcement hooks, a 3-part HITL smoke test, settings registration, git hook + dependencies, scorecard coverage/thresholds, and agent-registry table integrity (that the routing matrix the orchestrator reads is a real, contiguous table with a row per generated agent). It writes the result to .agentic/agentic-os/doctor.json.

Can I uninstall it? Nothing is committed for you, so before your first commit git status shows exactly what to delete. After that, the install journal (.agentic/agentic-os/install.json) lists every file it wrote, with ownership.

Glossary

One-liners for the jargon used above, no forward references required:

• Agent — an AI worker with its own instructions and tool access, spawned to do one job (write a migration, review a diff, triage a test failure).

• Subagent — an agent spawned by another agent rather than by you.

• Agent contract — the written instructions an agent runs under: its job, the files it may write, and when it must escalate.

• HITL (human-in-the-loop) — the rule that certain decisions stop and wait for a person instead of proceeding automatically.

• HITL dial — the install-time setting choosing how much runs unattended: strict (agents only recommend), gated-autonomous (pipelines run, risky decisions stop for you), autonomous (only low-confidence or flagged-risk decisions reach you).

• Gate — a checkpoint that work cannot pass until a condition holds (tests green, review approved, human said yes).

• Hook — a small program the editor runs automatically around an agent's actions; a failing hook physically blocks the action, unlike a prompt.

• Preset — a role-shaped bundle (developer, qa, ba-po, architect, pm-delivery, devops, portfolio) choosing which agents, hooks, and skills get installed.

• Scaffold — the set of files /agentic-init writes into your repo (contracts, hooks, policies, guides); nothing is committed for you.

• Doctor — a read-only verifier (/agentic-doctor, sdlc-doctor) that checks an install actually works and writes a pass/fail report.

• Adapter — a small config declaring which ticket/MR tool you use (GitHub, GitLab, Jira, Azure DevOps…), so no vendor is hardcoded.

• MR / PR — merge request (GitLab) / pull request (GitHub): a proposed change someone reviews before it lands.

• TDD (test-driven development) — writing the failing test first, then the code that makes it pass.

• Dispatcher vs pipeline — the two orchestration styles: dispatcher routes each request to one owning agent, one step at a time; pipeline runs the staged multi-agent flow.

• Skill — a packaged, invocable capability of a plugin (e.g. /agentic-init, sdlc-start, mr-watch).

Testing & development

bash tests/t0/run.sh                 # 61 hook unit tests
bash tests/t0/run-output-contract.sh # 12 output-contract parser checks
bash tests/run-matrix.sh             # T1–T8 acceptance (38 checks; re-runs the output-contract suite as T7)
cd mcp && npm run build && npm test  # mcp/ server: contract, content, and read-only tests

What CI proves, deterministically, on every PR: 99 checks — 61 hook unit tests (tests/t0/run.sh) plus the 38-check T1–T8 acceptance matrix — and JSON manifest/preset validation. The matrix executes the installer's deterministic phases against fresh and mature fixture repos; it is a skill-executability proof, not a mock. It covers non-destructive mature-repo handling, idempotent re-runs, upgrade classification, preset/ID resolution, dependency registration, the output-contract parser, and template rendering under quote-bearing answers. See tests/README.md.

What CI structurally cannot prove, and how it's proven instead: the generation loop is model-driven — you cannot spawn real subagents from a bash script, and mocking them would prove nothing about the actual claim. So it is verified by hand, repeatably, against fixtures on stacks the plugin has never seen, and every generation run below is recorded with its score: discovery → capability-driven generation → independent audit by a separate instruction-auditor subagent that re-checks every claim in the generated contract against the fixture's real code. Recorded results (tests/universal/README.md):

The 95s are the interesting ones. On FastAPI + Alembic the contract carried zero Postgres/Supabase RLS vocabulary from the exemplar it was shown; on Mongoose, zero migration vocabulary — and the rubric's evidence check caught a real generation mistake live, before it could reach a user's repo. The unverified claims are named in the log rather than rounded away.

That split is deliberate: the deterministic half is gated in CI, and the model-driven half is never quietly claimed as CI-covered.

Contributing

main is protected — no direct pushes; every change lands through a CI-passing pull request, and contributor PRs require a code-owner review. See CONTRIBUTING.md for the branch → PR flow.

Docs

• Engineering principles — the ideas behind the enforcement mechanisms, and why a plain agent session doesn't have them: docs/PRINCIPLES.md

• Upgrade flow: plugins/agentic-os/docs/UPGRADING.md

• Changelog: plugins/agentic-os/CHANGELOG.md

• Preset composition rules: plugins/agentic-os/presets/README.md

• MCP server (non-plugin hosts): mcp/README.md

• MCP server changelog: mcp/CHANGELOG.md

• Roadmap: ROADMAP.md

• Security policy: SECURITY.md

• Code of conduct: CODE_OF_CONDUCT.md

License: Apache-2.0.