Herdr Architect

herdr-architect

A Claude Code skill that turns one session into an architect over two [Herdr](https://herdr.dev) tabs: an implementer and an advisor. The architect writes specs and verifies results but never types implementation code. Every deliverable is reviewed by a second model with a clean context before it is reported done.


Roles

Role	Model	Where
Architect	the session you talk to, any model	your pane: specs, verification, reporting
Implementer	GPT-5.6 Sol via Codex CLI	its own tab, sandboxed to the working tree and /tmp
Advisor	Fable 5.1 via Claude Code in plan mode	its own tab, read-only, cleared after every consult

Both tabs stay visible in Herdr, so you can watch Sol work or read the advisor's reasoning at any time.

Flow

1. setup.sh runs preflight and opens or reuses both tabs.
2. The architect writes a spec with six fixed parts and sends it to Sol. The call returns as soon as Sol starts working; the architect writes the next spec or the review consult while it does.
3. Sol writes a report file in a fixed format. The architect reads the diff and re-runs the verification itself.
4. The advisor is consulted before architectural commitments, after two failed attempts, and once at the end of every deliverable. It answers in under 300 words and ends with VERDICT: ship | fix-first | rethink.
5. fix-first goes back to Sol as a numbered list. The architect verifies the fix against that list.

Requirements

• Claude Code running inside a Herdr tab (HERDR_ENV=1)
• codex with access to gpt-5.6-sol
• claude with access to the fable model
• python3

Tested with Herdr 0.8.2, Codex CLI 0.152.1 and Claude Code 2.1.258.

Install

git clone https://github.com/kennethgeerts/herdr-architect ~/.claude/skills/herdr-architect

Use

In a Claude Code session inside