Skip to content

Getting Started

Jump to bottom
Ânderson Q edited this page Aug 14, 2026 · 7 revisions

Getting Started

Install

curl -fsSL https://tron.42labs.io/seed.sh | sh

Clones TRON into ~/.tron and symlinks the tron launcher into ~/.local/bin, so you can run tron from anywhere. It's idempotent (re-run any time to update), never uses sudo, and never edits your shell rc — if ~/.local/bin isn't on your PATH it prints the one line to add. Pin a version with … | TRON_REF=v0.4.1 sh; relocate with TRON_HOME / TRON_BIN.

Prefer to do it by hand? git clone https://github.com/4242labs/tron && cd tron && ./tron start.

Requirements

  • python3 3.11 or newer (the engine reads workflow.toml with tomllib, stdlib from 3.11 on) and git.
  • A background-capable agent runtime on PATH — it runs the worker agents TRON dispatches. TRON drives it; you never address it directly.
  • TRON ships a default set of worker personas (engine/prompts/persona_*.md — architect, reviewer, worker, aide): the agents it dispatches. They're replaceable artifacts — keep them, edit them, or bring your own to match how your team works. Nothing project-side is required to start.
  • A complete project scaffold: blocks with exact acceptance criteria and a declared, runnable test, and every dependency/asset in place before a block is dispatched. TRON gates on evidence, so an under-specified block or a missing asset walls rather than improvises.
  • Optional: a Telegram bot (TELEGRAM_BOT_TOKEN, TELEGRAM_CHAT_ID in a gitignored .env) so milestone notes and operator pages ride Telegram. Without it, everything degrades to the terminal.

Run

# Installed on PATH (via the one-liner above); from a clone, use ./tron instead.
tron start                   # run TRON on this repo: a short bootup, then it dispatches
                             #   the fleet and drives the pipeline to done
tron start <project>         # point it at a project path directly
tron --watch                 # long-running: idles, wakes on register work, STOP file exits
tron --selftest              # engine selftests — no agents, no tokens

python3 engine/tron.py runs the same engine and prompts for the project path. No demo project ships with a clone — point it at a project of your own (Authoring a Project), or run one of the bundled templates through the SIM harness (below).

The bootup runs a short fixed sequence: first AIDE's own model (blank keeps the default), then the scope (all, a phase, or a block range), the worker count, whether to pause for your OK before each merge to trunk, and finally the model for each role (architect, reviewer, worker — each with a recommended default you confirm or override). A non-interactive boot (e.g. the SIM harness) asks nothing and takes defaults.

TRON runs autonomously — not a chat REPL. It prints milestone notes as blocks land and pages you (terminal OPERATOR> prompt, or Telegram) only when it needs a decision. You reach TRON by dropping a file in the project root: parley.md (a question/instruction — the architect answers) or report-request.md (the architect writes a report). Under --watch, a STOP file ends the run.

Validate the engine

Every module ships runnable selftests (no agents, no tokens):

python3 engine/tron.py --selftest      # the engine
python3 engine/gate.py                 # the truth gate (real throwaway git repos)
python3 engine/tg.py                   # the Telegram formatting/wiring
python3 engine/glossary.py             # vocabulary doc-sync   (--write regenerates docs/GLOSSARY.md)
python3 engine/events.py               # event-vocabulary sync (--write regenerates docs/EVENTS.md)
python3 engine/workflow.py             # flow lint + doc-sync  (--write regenerates docs/WORKFLOW.md)
python3 engine/bpmn.py                 # BPMN doc-sync         (--write regenerates workflow/)
python3 evaluation/harness.py --selftest

The full suite also runs in CI on every push (.github/workflows/engine-ci.yml).

Simulate a run

The evaluation/ suite exercises the whole engine end-to-end against seeded projects — no live tokens needed to see the flow:

python3 evaluation/harness.py project-01 3   # 3 SIMs of PROJECT-01; writes a stats.md
#   --parallel N   seed a project-owned flow with [limits] max_parallel = N
#   --ablate ARM   run with ONE invariant disabled (truth_gate | judge_isolation | architect_first)

Templates live in evaluation/templates/ (project-01/02/03 small→large, project-04 scale rung, exp-* fixtures). Run outputs write under evaluation/ and are gitignored.

File layout

tron/
├── tron                # launcher
├── install.sh          # the `curl … | sh` installer (the pretty URL redirects here)
├── engine/             # the deterministic engine + workflow.toml + prompts/
├── docs/               # GENERATED reference (GLOSSARY · EVENTS · WORKFLOW) + voice.md
├── workflow/           # GENERATED interactive BPMN diagram
└── evaluation/         # the SIM validation suite (harness + templates)

A project you point TRON at brings its own committed core docs — context.md, principles.md, playbook.md, an optional own workflow.toml, policy.md, blocks/*.md, and the engine-written pipeline.md register. TRON reads the project's pipeline; it owns no pipeline and no agents of its own.

See also: Commands · Configuration · Operations.

TRON

Clone this wiki locally