A core trio of AI coding CLIs forging production-grade code together. Claude Code leads a six-CLI builder pool — Antigravity CLI, Codex CLI, and an optional tier (OpenCode, Kimi, Cursor) — plus 19 specialized subagents, coordinated through a roster, per-task leases, portable skills, and parallel review swarms.
/setup now walks role assignment — keep the shipped posture, customize any role's CLI · model · effort, or restore the defaults — with validated writes, a content-level roster guard, and model overrides that reach every external-CLI lane. Built on the v3.0.0 builder-pool foundation below.
/setup role step — keep · customize · restore, validated writesCODEX_MODEL joins its siblingsagy) — Gemini 3.1 Pro by default, Flash opt-in/goal + ops/.sprint-complete sentinelops/roster.toml — role → CLI + model + fallback/setup onboarding + /cli-watch · /repo-watch--output-schema structured verdictsEvery model does what it does best. Claude leads and builds. Antigravity analyzes. Codex tests. Any roster member can build. Knowledge compounds across sessions.
Core trio + optional tier
Opus/Fable, max effort
Portable workflows
Parallel review swarm
Claude Code leads; Antigravity (1M-token analysis), Codex (sandboxed tests), and optional OpenCode/Kimi/Cursor are eligible builders. Each role is assigned from ops/roster.toml; every build is leased and cross-reviewed.
Security sentinel, performance oracle, architecture strategist, test-gap analyzer, findings synthesizer — each with restricted tools and focused expertise.
Model-agnostic workflow modules injected into any agent — TDD, systematic debugging, wave orchestration, shadow-path tracing, knowledge compounding.
Up to 7 reviewers analyze code simultaneously — Antigravity + Codex + 5 Claude specialist agents — then merged with confidence tiering.
Every non-trivial solution is documented in ops/solutions/. A learnings-researcher automatically searches these before planning new work.
All agents coordinate through shared markdown in ops/ — AGENTS.md, GOALS.md, TASKS.md, MEMORY.md, CHANGELOG.md, STATE.md. Auditable, git-friendly, no databases.
The single-writer rule is retired. Safety no longer comes from restricting who writes — it comes from leasing every task, isolating it in its own git worktree, and requiring a pinned non-author reviewer to approve before anything merges.
Every build is leased, isolated in a worktree, cross-reviewed by a pinned non-author, and squash-merged onto the sprint integration branch before a user-gated promote to main.
Every implementation task — lead-authored included — claims a per-task lease in ops/leases.toml before touching a single file.
The lease runs in an isolated git worktree, confined by a per-adapter environment allowlist. Builders never share the main tree.
A pinned non-author reviewer must approve. No agent self-merges — the builder can never be its own reviewer.
Each approved task lands as exactly one squash commit on the sprint integration branch, keeping history legible.
lease_promote gates the merge to main on a protected-path set plus a user-approval knob.
A lease moves claim → build → review → merge → release. A stale lease, or one that hits the same error three times, is reclaimed and reassigned to a fresh builder.
A single CLI-neutral file — ops/roster.toml — maps each role to a CLI, model, and effort. Fallback chains degrade gracefully and always terminate at the core trio, which can never be disabled.
Roles are the task types — builder, reviewer, tester, analyst, documenter — so choosing model and effort per task type is this file's native grain.
Lead agent and default builder — plans, builds, and coordinates every other CLI.
Analyst and documenter — 1M-token Phase 0 scans, architecture reviews, and docs.
Reviewer and tester — structured --output-schema verdicts and sandboxed TDD.
Eligible builder and reviewer once enrolled — skipped cleanly in every fallback chain when absent.
Eligible builder once enrolled — a decline persists as a flag flip, nothing else to restore.
Eligible builder once enrolled — the model is pinned explicitly, never the Auto router.
Run /ship for fully autonomous execution, or invoke each phase individually with dedicated commands.
Each runs in its own context window with restricted tools and focused expertise.
Validates task plans for completeness, assignments, dependencies
Merges review outputs with deduplication and confidence tiering
Runs build, tests, lint between waves
Searches solutions and decisions for relevant patterns
Orchestrates agent team workers with file ownership
Merges parallel research outputs into unified analysis
Per-task quality gate during team builds
SQL injection, XSS, auth bypass, OWASP Top 10
O(n²) loops, N+1 queries, memory leaks
Over-engineering, YAGNI violations, unnecessary abstraction
Naming, file organization, code style consistency
SOLID principles, coupling/cohesion, module boundaries
Untested code paths, missing edge cases, weak assertions
Industry patterns, anti-patterns, tradeoff analysis
Current docs for specific frameworks and libraries
Code evolution and architectural decisions via git
Validates bugs are reproducible before fixes begin
Post-deployment health checks and smoke tests
Reads GitHub PR comments and implements changes
Skills are markdown files that any agent can consume. Injected into Antigravity or Codex agent definitions, with prompt-prefix fallback via $(cat skills/SKILL.md).
| Skill | Consumer | What It Teaches |
|---|---|---|
codebase-mapping | Antigravity (Phase 0) | Full-repo analysis: structure, data flow, patterns, debt |
writing-plans | Claude (Phase 1) | Task decomposition with shadow paths, error maps, interface context |
shadow-path-tracing | Claude (Phase 1) | Enumerate every failure path alongside the happy path |
wave-orchestration | Claude (Phase 2) | Dependency-grouped parallel execution with integration checks |
test-driven-development | Codex (Phase 5) | RED-GREEN-REFACTOR: no production code without failing test |
systematic-debugging | Codex, Claude | Error taxonomy, assumption tracking, bisection, circuit breaker |
iterative-refinement | Claude (Phase 4) | Review-fix-review loops with convergence modes (max 3 cycles) |
review-synthesis | Claude (Phase 4) | Merge multi-reviewer findings with confidence tiering |
verification-before-completion | All agents | Evidence-based completion checklist — no "done" without proof |
knowledge-compounding | Claude (Phase 6) | Document solutions to ops/solutions/ for future sprints |
session-continuity | Claude | Save and resume via STATE.md across sessions |
scope-cutting | Claude | Systematically cut scope by unblocking value and risk |
watch-cycle | Claude (lead) | CLI/repo watch cycle: primary-source research, gap table, adopt/defer ADR |
Two external CLIs and five Claude specialist agents analyze the same code simultaneously through different lenses, then a findings-synthesizer merges everything.
HIGH — verified in codebase. MEDIUM — pattern match. LOW — heuristic only, can never be P1. Prevents wasting time on phantom issues.
Each reviewer has a "Do Not Flag" list — readability-aiding redundancy, documented thresholds, sufficient assertions, consistency-only style changes.
Max 3 review cycles. P1 fixed immediately, P2 this cycle, P3 logged for later. Escalate to user if not converged after 3 rounds.
Enforced at every stage of the pipeline. No shortcuts.
No build without a validated plan plan-checker — max 3 iterations
No production code without a failing test test-driven-development
No fix without diagnosis systematic-debugging
No "done" without proof verification-before-completion
No merge without review review-synthesis — max 3 cycles
3-attempt ceiling per issue, then escalation report systematic-debugging
Long sprints don't die to context limits or silent failures.
Native /goal checklist gates each session; completion is signaled only by the ops/.sprint-complete sentinel after the verification checklist passes
coordinate.sh spawns fresh sessions with clean context, notifies on completion via webhook or OS notification
Auto-checkpoints STATE.md before context compaction — prevents state loss during mid-sprint compaction
context-monitor.sh warns at 8+ consecutive reads without writes — breaks the reading loop
Tracks and warns on accumulated tool failures — 5 consecutive or 10 total triggers alert
Watchdog pattern on all external-CLI calls — SIGTERM after timeout, SIGKILL after 5s grace
Per-subagent risk accumulation — halt at >20% risk or 50+ file changes
Hooks, agents, skills, and commands register automatically. No manual configuration needed.
/setup gates the core trio live before any sprint runs. The optional tier enrolls the same way and is skipped cleanly in every fallback chain when absent.
| CLI | Binary | Tier | Minimum | Notes |
|---|---|---|---|---|
| Claude Code | claude | Core | 2.1.212 | Lead + builder; Fable 5 → Opus 4.8 → Sonnet 5 ladder |
| Antigravity | agy | Core | 1.1.3 | Tested 1.1.4; pins Gemini 3.1 Pro by default — still the latest Pro (Flash opt-in via roster, e.g. Gemini 3.6 Flash) |
| Codex | codex | Core | 0.144.0 | Reviewer + tester; --output-schema structured verdicts |
| OpenCode | opencode | Optional | 1.18 | OpenRouter provider connected (glm-5.2) |
| Kimi Code | kimi | Optional | 0.15 | OAuth device-code or API key (kimi-k3) |
| Cursor | cursor-agent | Optional | date-versioned | Pin grok-4.5, never the Auto router |
Python 3 is required for the hook handlers. Every CLI answers a headless READY probe — /setup runs them for you and writes the result to ops/roster.toml.
| Command | What It Does |
|---|---|
/ship <goal> full | Fully autonomous end-to-end sprint, gated to completion by /goal + sentinel |
/coordinate <goal> full | Full lifecycle with context-exhaustion recovery |
/plan <goal> phase | Analyze codebase, plan with shadow paths, validate |
/build phase | Wave orchestration build. --team for agent team mode |
/review phase | Parallel review + synthesis. --full for all 7 reviewers |
/test phase | Gap analysis + Codex TDD. --gaps-only to just identify gaps |
/wrap phase | Compound knowledge, archive reviews, write STATE.md, git trailers |
/quick <change> util | Changes touching < 3 files. Skips heavy machinery |
/debug <bug> util | Structured debugging: reproduce, diagnose, fix |
/deep-research <topic> util | 5 parallel research agents + research-synthesizer |
/status util | Sprint overview: phase, tasks, blockers |
/pause util | Quick checkpoint to STATE.md |
/resume util | Continue from STATE.md checkpoint |
/compound util | Document a solved problem or decision |
/analyze <url> util | Deep compatibility analysis of an external repo |
/resolve-pr <#> util | Read GitHub PR comments and implement changes |
/setup util | Guided roster onboarding — gate the core trio, enroll/decline each optional CLI |
/cli-watch util | Watch the six CLIs vs primary sources — gap report + adopt/defer ADR + re-probe. Schedulable monthly via /schedule |
/repo-watch util | Mine external repos for adoptable patterns — prioritized recommendations (recommends only). Schedulable monthly via /schedule |
ops/ directory with AGENTS.md, GOALS.md, MEMORY.md, and CHANGELOG.md templates./setup to get them live. The optional tier (OpenCode, Kimi, Cursor) is genuinely optional: enroll any subset via /setup, and an absent one is skipped cleanly in every roster fallback chain. Claude alone still runs the pipeline (degraded — you lose the multi-model review/test/analysis benefits)./quick for changes touching fewer than 3 files. It skips Phase 0, plan validation, and the full review swarm./goal gate holds completion until the checklist passes — signaled only by the ops/.sprint-complete sentinel after verification. Outside a session, coordinate.sh spawns fresh Claude processes with clean context windows, detecting completion via that sentinel. A PreCompact hook auto-checkpoints STATE.md before context compaction./compound saves it to ops/solutions/. Future /plan commands automatically search this directory before starting new work — so every sprint gets smarter. Each sprint should make the next sprint easier.