Concepts

The opinionated middle of Synthex — what the system believes, how the lifecycle flows, what enforces correctness, and how the parallel-execution model stays safe at the agent layer.

• Philosophy

  Why Synthex models software delivery as a team of agents rather than a single coder. Excerpt + canonical link to AJ Brown's essay.

• The lifecycle

  Five phases — Discover → Plan → Build → Ship → Learn — and how each one hands off traceably to the next.

• Quality gates

  Typed acceptance criteria ([T] / [H] / [O]), mandatory code + security review, severity policy, and the bounded review loop.

• Reviewer pipeline

  Fan-out + consolidation: how multiple reviewer agents run in parallel and how their findings get merged into one ranked list.

• Parallel execution

  Worktree isolation, the concurrent_tasks budget, and how Synthex runs many tasks at once without them stomping on each other.

• Native looping

  Iterate any Synthex command (or arbitrary prompt) with `--loop`. Per-session state, cross-session resume, and coexistence with the official ralph-loop plugin.

• Multi-model review

  Add reviewers from other model families, such as Codex, Gemini, and local Ollama models, and get one merged, attributed findings list.

• Standing review pools

  Keep reviewers warm between reviews so review-code and performance-audit start without a fresh spawn. Lifecycle, routing, and config. Claude Code only.

• Workflow engines

  Opt-in Claude Code engines for code review and --loop verdicts: what they change and when to turn them on.

• The marketplace

  How Synthex is distributed via LumenAI, the plugin file shape, and what runs locally versus what comes from upstream.