Skip to content

v1.1.0 — Warden role + /goal contract

Latest
Latest

Choose a tag to compare

View all tags
@webdevtodayjason webdevtodayjason released this 19 May 15:28
· 2 commits to main since this release
v1.1.0
2ca447b

First public release of claude-layers — a layered behavioral configuration system for Claude Code.

This release ships the full v1.1.0 stack: the universal Layer 1 base, four role overlays (Forge/Quill/Scout/Warden), three example project overlays, the orchestrator overlay (Argent), and the verifiable /goal contract loop.

What's in the box

  • Layer 1 (CLAUDE.md) — eight numbered behavioral guidelines drawn from Karpathy's late-2025 Claude Code notes (HN discussion). Read before write, surface confusion, hold position under challenge, surgical edits.
  • Layer 2 — four role overlays:
    • FORGE.md (builder) — implements specs, refuses to architect
    • QUILL.md (writer) — writes specs/docs, refuses to write code
    • SCOUT.md (investigator) — traces and reports, refuses to decide the fix
    • WARDEN.md (verifier) — runs the /goal manifest, signs off or kicks back
  • Layer 3 — project overlays: PROJECT_TEMPLATE.md plus three reference examples (PROJECT_WEBAPP, PROJECT_API, PROJECT_CLI)
  • Orchestrator (ARGENT.md) — composes context across layers; v1.1.0 adds a binding dispatch contract that refuses to spawn "general-purpose" workers (every dispatch must name a role overlay, a project overlay, and a /goal manifest)
  • Slash commands — /goal writes the manifest, /goal-verify runs Warden
  • /goal protocol — manifest schema, six check types (test, command, file_exists, file_contains, diff_constraint, human_review), three sign-off states (true, false, "pending"), full Stop-hook contract documented in .claude/docs/GOAL_PROTOCOL.md
  • Reference Stop-hook config — .claude/settings.example.json blocks worker termination until signed_off === true
  • Bash installer — --minimal, --yes, --help flags; idempotent; backed by CI install test

The closed loop

/goal → manifest written → Argent dispatches role-typed worker →
worker delivers → /goal-verify → Warden → proof.json →
Stop hook reads proof → releases iff signed_off === true

Install

curl -fsSL https://raw.githubusercontent.com/frontier-infra/claude-layers/main/install.sh | bash

Or clone and run ./install.sh /path/to/your/project.

Failure modes addressed

  • Wrong assumptions run with silently → Layer 1 §2 (surface confusion)
  • Sycophantic capitulation → Layer 1 §3 (hold position)
  • Bloated abstractions → Layer 1 §4 (simplicity default)
  • Scope creep into adjacent code → Layer 1 §5 (surgical edits)
  • Builder making architecture decisions → Forge "what you do not own"
  • Worker self-assessing "done" without verification → Warden + /goal manifest + Stop hook
  • Orchestrator dispatching generic workers → Argent dispatch contract (role + project + manifest required)

Acknowledgments

Base discipline draws on Andrej Karpathy's late-2025 Claude Code notes. Layered architecture is informed by years of building multi-agent systems where flat configurations consistently failed at scale.

🤖 Generated with Claude Code

Assets 2
Loading