join-the-team

AI-Native Developer Experience Harness

A team-project AI harness bootstrap that gives humans and agents a shared operating contract from day one, moving AI leverage from an individual “IC superhero” advantage to a repeatable team capability on an equal playing field.

📦 Now listed on Claude Plugin Hub — install the harness as the join-the-team plugin for Claude Code, Codex, Kimi Code, and Google Antigravity. Jump to install.

This repository has gone through three real phases — and the progression tracks something happening across the field.

It started as a reference guide: DEVELOPER_EXPERIENCE.md and BOOTSTRAP.md, something to read, walk through, and adapt. A concrete baseline for teams getting serious about AI-native delivery — the operating model, the guardrails, spec-first discipline, the Adversarial Gate.

Then it became context for your agents: CLAUDE.md and GEMINI.md here, plus the thin AGENTS.md/CLAUDE.md/GEMINI.md adapters the bootstrap generates — drop them into a project and the agent picks up the operating contract at session start without a briefing. The harness shifted from something humans read to something agents use.

Now it ships as an installable plugin — join-the-team (current version in CHANGELOG.md). One install across Claude Code, Codex, Kimi, and Antigravity, and the full harness is live: 22 skills, session-start orientation, slash commands, drift-checked in CI. The discipline travels with the agent from day one, across every project, without anyone copying files or re-explaining the contract.

The point: this stopped being something you read and became something you install.

And what you install is the operating model — that is the value proposition here. Skill libraries are everywhere now; what teams are missing is the contract: who holds authority, how risk is tiered, what evidence binds a review, what "done" actually means when humans and agents ship together. The 22 skills are the capability layer that executes inside that contract — not the other way round.

The kernel stays model-, vendor-, and IDE-agnostic throughout. Thin platform adapters handle discovery and invocation; the authority, risk, evidence, review, and completion semantics are non-negotiable.

Speed is easy. Safe speed is engineered.

This is deliberately opinionated. Refine it for your team, technology, domain, authority model, and definition of done. Keep one coherent shared contract while doing so.

New project? Bootstrap in fifteen minutes. Existing agent setup? Two ways to adopt below — by hand, or as a plugin.

Install — two ways to adopt

By hand — clone the repo and share it with your agents. Copy skills/ into your project, keep the CLAUDE.md/GEMINI.md/AGENTS.md adapters at root, and your assistant picks up the contract at session start. BOOTSTRAP.md is the fifteen-minute walkthrough.

git clone https://github.com/jpantsjoha/ai-native-developer-experience

As a plugin — the harness ships as join-the-team: one canonical skill set, thin per-harness adapters, drift-checked in CI. It is listed on Claude Plugin Hub; install straight from this repository into whichever coding assistant you run:

Claude Code

/plugin marketplace add jpantsjoha/ai-native-developer-experience
/plugin install join-the-team@join-the-team-marketplace

Kimi Code

/plugins install https://github.com/jpantsjoha/ai-native-developer-experience

Antigravity (Gemini)

agy plugin install https://github.com/jpantsjoha/ai-native-developer-experience

Codex — no install command needed: Codex reads AGENTS.md as its always-on adapter and discovers skills from .agents/skills/. The by-hand clone above is the install; if you are vendoring rather than cloning, put the directory at .agents/skills/ (or symlink it there) so Codex finds it.

Full per-harness detail (session-start hooks, verification, update path): Claude Code · Kimi Code · Codex · Antigravity

Antigravity is a first-class surface: this harness was built and battle-tested on Google Cloud's agent stack, and ships cloud guardrails (cloud-expert, with GCP, AWS, Azure and Alibaba Cloud references) plus adk-expert and mcp-server-scaffold alongside the vendor-neutral contract.

The plugin composes with — never duplicates — companion skill plugins; see INTEGRATIONS.md.

Verify the install

After installing, confirm the harness is actually live rather than merely present:

which skills are available?

You should see using-the-harness plus the cloud and gate skills (cloud-expert, adversarial-gate, pr-reviewer). Then run the built-in validator against a bootstrapped repository:

/join-the-team:validate

Expect PASS operating-model validation. A seed profile may pass with warnings — that is expected before the profile is promoted to active.

Usage

The harness is not a set of commands you memorise. Skills trigger on intent; three slash commands cover the lifecycle.

Three commands, in order:

/join-the-team:bootstrap    # seed the operating model (dry-run first, never overwrites)
/join-the-team:init         # record the humans: roles, authority, escalation channels
/join-the-team:validate     # check adapter drift, adoption state, evidence binding

Typical sessions — say the intent, and the router picks the skill:

You say	What fires	What you get
"Design the ingestion service"	spec-first-delivery → the-architect	Spec and ADR before code
"Is this safe to ship?"	release-readiness	Go/no-go against failure modes and rollback
"Review this PR"	pr-reviewer	Structured verdict with findings
"How would this break?"	adversarial-gate	Red-team pass argued against your own approach
"We're on GCP, multi-tenant"	cloud-expert + governance-guardrail	IAM, residency and policy guardrails
"Where do things stand?"	sitrep	Status, blockers, next actions
"Get it done, I'll be back later"	get-it-done	Pursues the agreed goal to a merged outcome; banks decisions instead of blocking
"The Mac is slow / the tests left Chrome open"	system-performance-guardrail	Snapshot, owner attribution, bounded cleanup, before/after numbers
"Is everything live? Any errors this week?"	cloud-checkup	Read-only GCP checkup: one status table, one OVERALL: line, in session or on a schedule
"Is the page fast enough? Why is this route slow?"	frontend-performance-audit	Core Web Vitals and traces from the real authenticated page, not an estimate

Worked example — a risk-touching change:

You:   We need to let tenants export their own audit logs.
Agent: [spec-first-delivery] Spec before code. Who is the actor, what is out of scope?
       [governance-guardrail] Multi-tenant export is R2 — data-boundary check required.
       [the-architect] ADR-00N drafted: signed URLs vs streamed export, trade-offs recorded.
       [adversarial-gate] How would I break this? Tenant A requesting tenant B's logs.
       → Named human approves the R2 classification before implementation starts.

The point is the routing, not the vocabulary: you describe intent, the contract decides which gate applies.

Configuration

The harness works unconfigured. Three seams are worth knowing:

Seam	File	What it controls
Operating contract	CLAUDE.md / GEMINI.md / AGENTS.md	Always-on rules the agent loads each session
Project profile	PROJECT-OPERATING-PROFILE.md (generated by bootstrap)	Risk tiers, named authority, escalation, definition of done
Data seams	.agents/mcp_config.json	Governed MCP access — a template, not live config

Adjust the risk tiers and authority model to your team; keep one coherent contract while you do. The MCP file ships as an example: copy an entry, rename it, point it at a source you control. See the governed-data-seam pattern.

Troubleshooting

Symptom	Cause	Fix
Skills discovered but no orientation at session start	The session-start hook needs bash on PATH	Install bash; without it the plugin degrades to discovery without injection
Antigravity reports hooks: skipped (not found)	The root hooks.json is missing from the install	Reinstall; agy plugin install must report hooks: 1 processed
Codex-installed copy has no .agents/skills/	Codex's install cache flattens symlinks	Expected — all 22 skills are at skills/; use .agents/skills/ only in a repo checkout
/join-the-team:validate warns about unresolved fields	Profile is still a seed	Fine for R0/R1 work; resolve placeholders and promote to active before R2/R3
Slash commands missing after install	Client not restarted	Restart the client; Claude Code applies plugin updates on restart
Agent ignores the contract mid-session	Context drift after a long session	/clear, then let the session-start hook re-inject the orientation skill

Still stuck? Open an issue with the client, version (plugin.json), and the output of make spec-conformance.

---

Is this the same as BMAD?

No — and the difference is the point. BMAD is a methodology engine: it drives an idea through phased workflows — analysis, planning, solutioning, implementation — with AI personas playing Analyst, PM, Architect, Developer. It answers what to do next. join-the-team answers who is accountable, and what is allowed — the contract the work runs inside: risk tiers, named human authority, evidence bound to an exact commit, a definition of done that means the checks pass. They scale on different axes — BMAD scales ceremony by size, this harness scales gates by risk. So they compose rather than compete: run BMAD as the workflow engine, and let join-the-team keep a mixed human-and-agent team coherent while it does. Reach for it the moment "who decided this, and on what evidence?" stops being obvious. The workflow you can borrow anywhere — the contract is the part teams keep missing.

---

What it looks like

Install the plugin, and at session start the using-the-harness skill grounds the agent in the contract before any work. Ask it what the harness changes, and it answers in the terms that matter — where the system accelerates, and where it stops:

Coherence and auditability — not speed. The value is visible when the system knows where to accelerate and where to stop: a risk-touching decision triggers the