arcgentic

Arcgentic

Arcgentic is a harness engineering layer for AI coding agents. It turns ad-hoc prompting into a gated engineering workflow.

中文文档 -> README.zh-CN.md

Arcgentic helps Codex and Claude Code run software work as a disciplined sequence: clarify the idea, plan the work, build it, self-audit it, optionally run a realistic user test, audit it independently, then close only when the evidence is good enough.

It is for people who already use AI coding tools, but do not want every session to depend on memory, vibes, or a lucky prompt.

Where it came from

Arcgentic started as the development discipline behind Moirai, a real agent project where AI coding had to survive 30+ strict development rounds, repeated NEEDS_FIX audits, planned handoffs, role boundaries, self-audit, external audit, and recoverable session state.

This plugin packages the patterns that survived that work so Codex and Claude Code users can apply them to their own complex projects. It is not a new coding agent, and it does not copy Moirai-specific phase numbers, fact shapes, or runtime internals. It is the workflow layer extracted from real agent development.

Harness engineering

Arcgentic sits in the harness layer around a coding agent.

Codex and Claude Code are the agents. Arcgentic is the engineering harness that gives those agents roles, handoffs, stop states, audit gates, and evidence. This is the same direction people describe as moving from vibe coding toward agentic engineering: the important work is not only prompting the model, but building the workflow around the model so its output can be checked, routed, and trusted.

Related reading:

• Martin Fowler / Thoughtworks: Harness engineering for coding agent users
• Andrej Karpathy's broader "vibe coding -> agentic engineering" framing

30-second version

Question	Answer
What is it?	A harness engineering layer for Codex / Claude Code: roles, handoffs, audits, stop states, and pass/fix gates around the coding agent.
What problem does it solve?	AI coding sessions drift: scope changes silently, context gets lost, tests are skipped, and "done" often means "the assistant said it is done."
Who should use it?	Heavy Codex / Claude Code users, agent builders, AI-native teams, and people doing complex multi-round engineering work.
What does it add?	A repeatable gated workflow with automated role dispatch: planning, dev self-audit, optional user testing, external audit, and closeout.
What does it not do?	It does not replace your judgment, your tests, or your review process. It makes those steps harder to skip.

Platform status

Platform	V2 status	Verification
Codex	Complete V2	Verified in a real Codex project workflow, including automatic Orchestrator thread setup and role-thread dispatch.
Claude Code	Complete V2	Verified in a real Claude Code session for single-session-subagent mode via a foreground Agent tier-0 broker dispatch — see tests/dogfood/gate-3-claude-code-native-broker/RESULT.md. That gate covered a single Planner role dispatch, not a full Planner→Developer→Auditor round loop. The SendMessage-based footer-correction retry path and ListAgents were not exercised by this gate (the footer was valid on the first attempt), and multi-session-subthread mode and the hook-backed fallback path remain unverified.

Codex is the best current experience. In the verified Codex path, the current project conversation becomes Orchestrator; Arcgentic then creates or reuses the role threads, names them, sends the right role prompt, waits for their return signal, and dispatches the next role without manual thread switching.

Claude Code support now has dogfood evidence for single-session-subagent mode via a foreground Agent tier-0 broker dispatch, so treat that specific path as a verified real workflow. The SendMessage-based footer-correction retry path and ListAgents were not exercised by that gate, and multi-session-subthread mode and the hook-backed fallback path have not yet been verified in a real Claude Code session — all of these should still be treated as unproven.

Distribution status

Channel	Status	Use it for
GitHub Release	Published v2.2.1	Release notes, source archive, and verification context.
PyPI	Published arcgentic==2.2.1	Python CLI: gates, V2 state helpers, Claude Code broker, and audit tooling.
npm	Published arcgentic@2.2.1	Plugin asset bundle and Codex local install helper.
Claude Code plugin marketplace	Manifest ready	Main Claude Code install path; single-session-subagent mode now has dogfood evidence via a foreground Agent tier-0 broker dispatch, but multi-session-subthread mode and the hook-backed fallback path remain unverified.

Install

Codex local install

git clone https://github.com/Arch1eSUN/Arcgentic.git arcgentic
cd arcgentic
bash scripts/install-codex-local.sh --plugin-root .

Then start in a saved project workspace and ask:

Use Arcgentic to build this idea: <your idea>

npm bundle install

Use this if you want the Arcgentic plugin assets through npm:

npm install -g arcgentic
arcgentic install-codex-local

The npm package is a zero-dependency plugin bundle and Codex local install helper. It includes the skills, agents, scripts, schemas, templates, and platform manifests. The Python CLI is still published separately on PyPI.

Claude Code install

/plugin marketplace add Arch1eSUN/Arcgentic
/plugin install arcgentic@arc-studio

Then start inside your project:

Use Arcgentic to build this idea: <your idea>

For Claude Code V2 experimental workflow setup:

If your Claude Code session's own tool list already includes Agent, SendMessage, and ListAgents (native tooling / "tier 0"), no setup step is needed — the session broker uses those tools directly and this is the preferred, dogfood-verified path (see "Platform status" above). Only run install-hooks below as a fallback, for sessions that lack those three tools:

arcgentic claude-code-broker install-hooks \
  --settings .claude/settings.local.json \
  --state .agentic-rounds/state.yaml

CLI install

Use this if you only need the command-line helper:

pipx install arcgentic
arcgentic --help

Minimal example

Without Arcgentic:

User: Build a small expense splitter.
AI: writes code
AI: says it is done
User: later discovers missing edge cases, unclear scope, no audit trail

With Arcgentic:

User idea
-> current conversation becomes Orchestrator
-> Orchestrator creates or reuses Planner and sends the planning prompt
-> Planner returns the plan to Orchestrator
-> Orchestrator creates or reuses Developer and sends the dev prompt
-> Developer implements and returns a self-audit
-> Orchestrator dispatches optional Test only if realistic use needs it
-> Orchestrator creates or reuses Auditor and sends the audit prompt
-> Auditor returns PASS / NEEDS_FIX / AUDIT_INCOMPLETE
-> Orchestrator routes the next step

The important difference is not that the AI writes more text. The important difference is that each role has a job, each stage has a stop condition, and "done" is not accepted until the workflow can explain why.

Arcgentic recommends a mode first

When you start Arcgentic with a new idea, the current session becomes Orchestrator. Before it plans or builds, it should judge whether the idea is a small fast project or a larger project that needs stronger review. Then it recommends one project-level mode and asks you to confirm or override it:

Mode	Choose it when	Tradeoff
Single session, multiple agents	You want the fastest run and a smaller demo surface.	Faster completion, weaker audit isolation. Planner, Developer, Test, and Auditor run inside the current Orchestrator session as fixed named role agents and are reused across rounds.
Multiple sessions, multiple threads	You want stronger separation between planning, development, testing, and audit.	Slower completion, stronger audit discipline. Planner, Developer, Test, and Auditor use fixed project threads.

The choice is made once for the project. Arcgentic should not ask again every round unless you start a new project or intentionally reset the workflow.

What changes in real use

Before

• One long AI coding session tries to remember everything.
• The assistant mixes planning, coding, review, and closeout in one context.
• Fixes are sometimes treated as audit work.
• The next session has to reconstruct what happened.
• "Pass" often means the assistant felt confident.

After

• The current session is the Orchestrator.
• Planner, Developer, Test, and Auditor are separate roles.
• Developer owns building, repairs, and self-audit.
• Auditor owns stricter independent review.
• Test is used only when realistic user behavior needs separate verification.
• Closeout happens after the project/phase conditions are satisfied.

The V2 workflow

Arcgentic V2 follows this shape:

idea
-> brainstorm and planning
-> round handoff
-> development
-> developer self-audit
-> optional user-test
-> external audit
-> pass or fix
-> next round, next phase, or closeout

The roles are fixed:

Role	Owns
Orchestrator	Routing, role dispatch, waiting, and deciding which role acts next.
Planner	Brainstorming, project plan, phase/round structure, handoffs, and closeout decisions.
Developer	Building, fixes, local verification, and self-audit.
Test	Realistic user/session testing when the plan says it is needed.
Auditor	Independent evidence review and PASS / NEEDS_FIX / AUDIT_INCOMPLETE decisions.

Arcgentic does not create a new role identity every round. The role names stay fixed: Orchestrator, Planner, Developer, Test, and Auditor.

Custom role/state topology

The role/state routing that decides "who acts next" is no longer a hardcoded table. It is a Topology (toolkit/src/arcgentic/topology.py) that a project can override via project.arcgentic_v2.topology in state.yaml — routes, which states each role may act from, and conditional next-role selection based on the returning role's artifacts (for example, routing differently depending on an audit outcome field). Zero-config projects get the exact same 5-role sequence as before, byte for byte; custom topologies are validated at parse time (unknown role keys and routes with no matching next state are rejected before they can leave a project stuck mid-round). This is for projects that need a different role graph than the default plan/dev/audit loop — most projects should never need to touch it.

Codex V2

Codex V2 is the verified path.

In Codex, Arcgentic can run either V2 mode.

The verified automation is: