neuralforge-labs/tlmforge
4 stars · Last commit 2026-06-03
README preview
# tlmforge **The missing process layer for Claude Code.** Claude Code writes code fast. tlmforge makes sure it's *correct* — by enforcing a spec audit before any code is touched, running 8 independent agents against your plan and your output, and blocking commits that haven't cleared an adversarial red-team review. [](LICENSE) [](.claude-plugin/plugin.json) [](https://claude.ai/code) [](https://python.org) --- ## The gap no other tool fills Two excellent tools already tried to solve this problem: **[obra/superpowers](https://github.com/obra/superpowers)** (174K stars) built a rigorous 7-phase TDD-first methodology. They also built a bounded multi-round plan review loop — and removed it in v5.0.6. Their [release notes](https://github.com/obra/superpowers/blob/main/RELEASE-NOTES.md) state: *"The subagent review loop doubled execution time (~25 min overhead) without measurably improving plan quality. Regression testing across 5 versions, 5 trials each, showed identical quality scores."* They replaced it with inline self-review checklists. **[garrytan/gstack](https://github.com/garrytan/gstack)** (89K stars) took a different approach: 35+ role-based slash commands (CEO, Staff Engineer, QA Lead, `/cso` for OWASP+STRIDE security audit, `/codex` for independent cross-model review via OpenAI Codex CLI — source: [codex/SKILL.md](https://github.com/garrytan/gstack/blob/main/codex/SKILL.md), [cso/SKILL.md](https://github.com/garrytan/gstack/blob/main/cso/SKILL.md)). Genuinely capable. The core limitation: every review skill is on-demand and user-invoked. Nothing automatically gates your commit. You can skip straight to implementation without a spec and ship without ever running `/cso`.