claude-ctrl
A structured workflow for agentic development implemented as skills and hooks, transforming Claude Code into a deterministic control plane. The core insight: "An instruction that lives only in model context is not a constraint." ClauDEX moves enforcement from prompts to event-based hooks with SQLite-backed policy evaluation, because LLMs are probabilistic systems that require deterministic enforcement boundaries.
Key Quotes
"An instruction that lives only in model context is not a constraint."
"LLMs are not deterministic systems with probabilistic quirks. They are probabilistic systems."
Key Themes
The workflow loop (Planner -> Guardian -> Implementer <-> Codex Critique -> Reviewer -> Guardian) is more structured than most agent workflows. The Guardian role is interesting -- it provisions worktrees and lands git changes, acting as the gatekeeper between planning and implementation. The read-only Codex/Gemini CLI providing independent critique is a form of cross-model review.
The SESAP concept (Self-Evaluating Self-Adaptive Programs) -- probabilistic systems constrained to deterministically produce desired outcome ranges -- is the theoretical foundation. It's a useful framing: don't try to make the LLM deterministic, just constrain its output to an acceptable range.
Connects to claude-code-config (Trail of Bits) (which also uses hooks for enforcement, but with a lighter touch), Pre-Commit Lint Checks (mechanical enforcement at a different layer), and Compound Engineering (building trustworthy systems rather than relying on model behavior).
Critical Analysis
At 1,566 commits this is one of the most actively developed projects in the Claude Code ecosystem, which suggests either deep conviction or endless yak-shaving. The "first-deny-wins" policy engine is a strong pattern -- it's the same approach used in firewalls and IAM systems, applied to agent behavior. The risk is over-engineering: if the enforcement infrastructure is heavier than the actual development work, you've traded one form of complexity for another. The 81.6% Python + 14.6% Shell composition suggests a lot of glue code, which is both necessary and fragile. For security-sensitive work (Trail of Bits' domain), this level of enforcement makes sense. For a personal project, it's probably overkill.
Sources: claude-ctrl Last updated: 2026-05-14