Ninety2UA/agent-triforge

Website ·
What's new ·
Install ·
Getting started ·
Documentation ·
Release history ·
FAQ
---
Why Agent Triforge
Agent Triforge puts several AI coding CLIs on one sprint. One CLI leads an eight-CLI builder pool, and a roster you can edit, ops/roster.toml, gives each role a CLI, a model and an effort. Any member can implement code; Devin builds only if you opt in.
- The lead is Claude Code or Codex. You choose it in at-setup, which records it in the roster's [lead] table; without the table, Claude Code leads. The lead plans, resolves the roster, dispatches builders and merges reviewed work. Claude's model ladder is Fable 5.1 → Opus 5.5 → Sonnet 5.5.
- The core trio is required: Claude, Antigravity (agy, Gemini 3.8 Flash (High) by default, 1M context) and Codex (gpt-6-astra, sandboxed).
- The optional tier is detected automatically: OpenCode (OpenRouter glm-5.3), Kimi Code (kimi-code/k3), Cursor (Grok 4.6) and Grok Build (grok-4.7; builder, reviewer or analyst). You enroll them through at-setup, and one you don't enroll is skipped cleanly. Devin (swe-1-6-slow, reviewer or analyst) joins only after you record your consent in at-setup.
- 19 specialist personas cover security, performance, architecture and more. One manifest sets each persona's tools, model tier and turn budget.
Every implementation task, the lead's own included, is built under a per-task lease in its own git worktree and merges only after a pinned reviewer, never its builder, has cross-reviewed it. Safety comes from that isolation and cross-review; any member may write code. Work is tracked in shared markdown files under ops/, reviews run in parallel, and what one sprint learns is written down for the next.
> *Each sprint should make the next sprint easier.*
When the lead solves a problem whose reasoning would otherwise be lost, it writes a note to ops/solutions/, or to ops/decisions/ for an architectural decision. The learnings-researcher persona searches both before new work is planned (the knowledge loop).
What's new (v4.1.1)
This release changes the documentation and the release tooling. Nothing changes in how you run Triforge.
- **The README follows agent-blueprint's order.** It starts with what Triforge is, then covers install for either lead, getting started, unattended runs, upgrading from 3.x and updating, and ends with the reference sections. It shows only the latest What's new; older releases are in the collapsed Release history.
- **The diagrams are light.** The 13 dark diagrams became 15 light ones that read on GitHub's light and dark themes, and on a phone a tap opens each one full size. Two are new: the install routes and the unattended-run loop. Checking every label against the code corrected a few stale facts: there are five quality gates, at-ship reviews in five lanes, the kill grace is 10 s, and test fix cycles stop at 3.
- **Each release has its own notes file.** A version's notes live in docs/releases/v .md, and the release workflow publishes that file as the GitHub release.
- **The maintainer gates are stricter.** validate-versions.sh fails a version bump that has no notes file and rejects Liquid delimiters under docs/, which break the GitHub Pages build. It also fails when a diagram source changed without a re-render, or when a diagram names a stale pin, a wrong count or a Codex invocation without its agent-triforge: prefix.
---
Install
Triforge installs as a plugin into the CLI that leads: one command adds the marketplace, one installs the plugin. The same repository serves both leads, because Codex also reads the .claude-plugin/ manifests (D-048).
If you are coming from Triforge 3.x, read Upgrading from 3.x first. The slash commands are now at- workflows, and a CLAUDE.md left from the 3.x template hides the project's AGENTS.md from Claude Code.
You need the core trio (Claude Code, Antigravity agy and Codex) installed and logged in, Python 3, and on macOS GNU coreutils. Then add the marketplace and install the plugin: