Wallerand Delevacq
FrançaisFR

Leg 03 / Public repository / 2026

A measured AI agent harness

My workstation and the AI coding agents I use every day, managed from a single public source, with encrypted secrets and an audit that checks whether agents follow each instruction.

Flight plan

Period
August to October 2026
Role
Sole author
Team
Solo project
Stack
Go, chezmoi, age, GPG, GitHub Actions
692
real sessions analysed to see which skills get used
24
ADRs on a solo project
2
agent hosts, Claude Code and Codex, configured from a single source

From repository to agent

One source of instructions, projected to each host, then measured on real sessions.

  1. WP01Canonical sourcesTechnical rules, voice, preferences
  2. WP02AdaptersClaude Code on two profiles, Codex
  3. WP03Skills and subagentReview, ADRs, obsolescence, handoff
  4. WP04MeasurementContext cost, triggering, instruction adherence

Context

I work on macOS and Linux, with two profiles (personal and work) and two AI agent hosts. The repository is public and holds encrypted private fragments.

What I built

  • A layered instruction harness, projected to Claude Code and Codex, with 18 skills and a subagent that reviews the changes automated checks cannot decide.
  • A harness audit: context cost, deployment lag, activation of each skill from transcripts, and mutation tests on rules.
  • Secrets encrypted with age and a GPG key on a YubiKey; CI refuses any command that could leak decrypted content into public logs.
  • A 14-command Go CLI and a single verification gate, replayed in CI on macOS and Linux for both profiles, with an idempotency check.

What it shows

I treat AI agents like a production system: every instruction has a cost and a measurement. And I apply to a solo project the governance I expect from a team.

One decision

Measure before adding a rule

Context
Every always-loaded rule costs context in every session, and nothing proves an agent follows it.
Decision
A rule or skill only enters the harness with three measurements: what it costs in context, whether agents actually trigger it, and whether the mistake it targets becomes rarer once it is added.
Consequence
Across 692 real sessions, I know which skills get used and which never trigger. When data is missing, the audit answers "unknown" instead of concluding that all is well.