British Army veteran | Founder, brAInwav | Codex-first toolmaker
Codex writes the code. I lead, inspect, and make the work accountable. The value is using both strengths properly.
Now (Aug 22, 2026): building synAIpse, Skills SDK, deterministic agent loops, and evidence tooling for teams that want AI coding to ship without losing trust.
By harness, I mean the operating layer around Codex and other coding agents: CLI entrypoints, repo-local guardrails, runtime evidence, review policy, memory, and handoff artifacts that make AI-assisted engineering repeatable.
Last updated: 2026-08-22
I build the layer that lets humans use coding agents with more confidence:
- Agent-ready repos with clear entrypoints, preflight checks, validation gates, and rollback-aware workflows
- Runtime evidence that separates local test truth from PR state, CI, review threads, tracker status, and merge readiness
- Capability systems for skills, plugins, prompts, hooks, and review agents that can be validated instead of merely trusted
- CLI products that make research, architecture, knowledge, and repo intelligence available to humans and agents
- Local-first memory and narrative tools that preserve intent, decisions, receipts, and context across long-running work
Codex, OpenAI, MCP, TypeScript, Node.js, React, Tauri, Swift, SwiftUI, Python, Bash, macOS, GitHub Actions, CircleCI, CodeRabbit.
Problem: AI coding is fast, but speed is not enough. Teams still need current context, bounded autonomy, repeatable validation, review evidence, and a clean handoff back to humans.
Solution: I build pragmatic Codex-first harnesses: CLIs, instruction systems, skills, evals, review gates, runtime cards, and workflow evidence that turn experiments into dependable engineering operations.
Why it helps: Shorter review loops, fewer vague agent claims, clearer operational defaults, and repos that are easier for both people and agents to pick up safely.
| Operating problem | What I built | Repo proof |
|---|---|---|
| Agents need a safe next step, not a wall of docs | Cockpit-style commands, runtime cards, repo-local gates, and evidence-backed handoff | synAIpse / coding-harness private work |
| Skills and plugins need lifecycle control | SDK-style authoring, routing, validation, evals, packaging, sync, and command-surface projections | Skills SDK |
| Long-running agent work needs deterministic state | Fresh session loops with file memory, receipts, context snapshots, gates, and review exits | ralph-gold |
| AI-assisted code needs recoverable context | Local-first session-to-commit narrative, timelines, and search across the why behind changes | trace-narrative |
| Reviewers and agents need architecture evidence | PR impact reports, repo orientation packs, policy validation, and agent handoff artifacts | diagram-cli |
| Research and knowledge tools need agent-safe UX | Scriptable CLIs with structured output, explicit policy gates, diagnostics, and safe defaults | rSearch, wSearch |
| Project | Why it matters | Signal |
|---|---|---|
| coding-harness | synAIpse AI Delivery Harness: a CLI control plane for agent-ready repos, runtime evidence, review gates, and safer PR handoff. | active |
| Agent-Skills | Skills SDK for authoring, routing, validating, packaging, and syncing Codex skills and plugins. | 8 stars |
| ralph-gold | Deterministic fresh-agent loop with file-based memory, gates, receipts, context snapshots, and review exit rules. | 2 stars |
| trace-narrative | Local-first app that links AI sessions, intent, commits, and timelines so teams can recover the why behind code changes. | active |
| diagram-cli | Architecture evidence CLI for PR review, repo orientation, agent handoff, policy validation, and Mermaid diagrams. | active |
| rSearch | Search, fetch, and download arXiv papers from the terminal. CLI plus TypeScript client. | active |
| wSearch | Script-friendly Wikidata REST, SPARQL, and Action API queries from the terminal. | active |
| mKit | MCP server boilerplate for Cloudflare Workers. | 1 star |
# ralph-gold
gh repo clone jscraik/ralph-gold
cd ralph-gold
uv tool install -e .
ralph --help# rSearch
npm i -g @brainwav/rsearch
rsearch --help# wSearch
npm i -g @brainwav/wsearch-cli
wsearch --help- skillsbar
- Design-System - Cross-platform UI workbench and component system for ChatGPT widgets and React apps.
- unfinished-cemetery - A ritualised archive of abandoned projects — post-mortems for software that died so we could learn what lives.
- evals - Shared local eval runner for artifact integrity, schema validity, evidence-backed claims, and deterministic scorer verdicts.
All published under @brainwav on npm:
| CLI | What it does | Install |
|---|---|---|
| rSearch | arXiv paper search, fetch, download | npm i -g @brainwav/rsearch |
| wSearch | Wikidata REST/SPARQL queries | npm i -g @brainwav/wsearch-cli |
- Shipping synAIpse / coding-harness - a portable AI delivery harness for agent-ready repos, review gates, runtime evidence, and safer PR handoff
- Building Skills SDK - a governed SDK for Codex skills, plugins, evals, review closeout, and runtime projections
- Building deterministic agent loops - using ralph-gold to keep task selection, gates, receipts, and exit rules explicit
- Making context durable - connecting AI sessions, commits, project memory, architecture evidence, and review artifacts
- Publishing practical CLIs - research, Wikidata, architecture, and repo-intelligence tools with structured output and agent-friendly diagnostics
AI delivery harnesses - make a repo safer for Codex and other coding agents with entrypoints, gates, evidence, and handoff contracts
Agentic developer workflows - Codex, MCP, review loops, PR automation, runtime cards, and validation policy
Developer tooling and CLIs - research, knowledge, architecture evidence, repo automation, diagnostics, and machine-readable UX
AI governance that actually runs - instructions, drift control, evals, skill lifecycle, review gates, and repeatable workflows that keep human intent visible
Founder/operator advisory - turn messy prototypes into dependable AI-assisted product and engineering systems without losing the point of the work
I keep an archive of retired experiments at unfinished-cemetery: short post-mortems for software that taught something useful before it was retired.
- system-design.md - A format specification for describing a visual identity to coding agents. DESIGN.md gives agents a persistent, structured understanding of a design system.
- GitHub: jscraik/system-design.md
- code-archaeology-kit - Privacy-aware git-history intelligence for churn, temporal coupling, abandoned structures, and cleanup targets.
- GitHub: jscraik/code-archaeology-kit
- xKit - Unofficial X Web API Toolkit
- GitHub: jscraik/xKit
- zSearch - Z.AI capabilities CLI and MCP server for agents and automation
- GitHub: jscraik/zSearch
- data-dashboard
- GitHub: jscraik/data-dashboard
- jules-companion
- GitHub: jscraik/jules-companion
