Separate portable setup from runtime state
Selective links and layered configuration let shared practices travel without turning authentication or session history into configuration artifacts.
Selected work
Case study 05 / Workbench
A workbench for coding agents.
I built versioned setup and reusable workflows for coding agents, then connected them to browser session ownership, evidence for human review and a protocol for reviewing delegated work.
A conflicting edit. A safe stop.Watch the refusal and adoption.The actual manager in two fictional setups, preserving local settings.PythonShellChrome extensionAgent hooksSkills
metaclaude + metacodex + HelmA personal engineering project
01The starting point
Agent setup is part of the work.
A useful agent setup spans instructions, skills, hooks, settings and browser context. Copying those pieces by hand makes them drift between sessions and computers. Copying the whole runtime also carries credentials and state that belong to one machine.
metaclaude gives portable Claude Code setup a versioned home. metacodex applies the same boundary to Codex, with layered configuration and checks for local consistency. Both keep repeatable workflows close to the agent that needs them.
The workbench also needs a way to see and review work. Browser tabs have session ownership, observer images give the human a view after supported actions, and captioned walkthroughs turn a flow into a document. Helm gives the lead agent a shared protocol for reviewing worker evidence across two runtimes.
An illustrated workflow
Shared practice. Local boundaries.
A portable workbench brings the useful setup with it. The things that belong to one machine stay local.
- 01
Shared source
Instructions, skills and hooks live in versioned repositories.
metaclaude / metacodex - 02
Local choices
Machine settings and private credentials supply the local configuration.
Machine + secret overlays - 03
A working session
The selected runtime reads the assembled configuration and its own tools.
Claude Code or Codex
An illustrated workflow
Keep the evidence with the lesson.
A useful finding can improve the next session. Review is the step between one observation and a standing instruction.
- 01
Observe
A worker records what happened and the evidence behind its finding.
Result + supporting evidence - 02
Review
The lead agent checks the result and decides whether the lesson should generalise.
Helm captain protocol - 03
Carry forward
Accepted guidance becomes part of the workbench a later session can read.
Reviewed rules + handoff
A map of the working practice. It does not read or change a machine’s configuration.
02What the workflow looks like
A useful path through the product.
- 01
Install the portable setup
Manage shared instructions, hooks and skills through versioned configuration. Credentials, sessions and local runtime state retain their own machine-local boundary.
- 02
Check the local assembly
metacodex combines a portable base with machine and secret overlays. Doctor and trust tools inspect the setup, while drift adoption preserves legitimate local choices.
- 03
Return browser work to its owner
The tab bridge identifies a selected page’s session and can deliver it with an instruction. Observer screenshots and captioned flow PDFs provide separate forms of human-reviewable evidence.
- 04
Review delegated results
Helm keeps evidence, reports, handoffs and guidance proposals consistent while Claude/Bosun and native Codex use their own coordination tools. The captain checks a worker’s proposal before adopting it as a rule.
03Design + engineering
The decisions underneath.
Make browser identity specific
Two independent tabs can share a URL. Session ownership and tab identity preserve the intended destination when an agent inspects or receives a page.
Put review before shared guidance
A September Helm review caught conflicting report paths and a rule that assumed every screenshot tool had the same limits. The lead agent corrected both; an independent source check confirmed the changes.
Explore the individual features
The smaller pieces of this story.
metaclaudeGive reusable agent setup a versioned home.metacodexShare Codex setup while keeping local settings.Model policy guardCheck what an agent launch would use.Browser handoffSend the page to the session that owns it.Observer screenshotsGive the human a view of the browser action.Flow screenshots PDFMake a browser walkthrough someone can review.HelmReview delegated work across agent runtimes.Reusable skillsPackage a repeated task with a clear contract.Workbench syncUpdate the shared setup without copying local state.Browser evidenceKeep a browser result connected to its task.
04Selected release history
It got here one decision at a time.
These are historical changes recorded in repository tags. The capture and current-state notes below describe what is being shown here.
A shared captain protocol across runtimes
Helm established distinct runtime directions for Claude/Bosun and native Codex, with common reports, handoffs and reviewed instruction feedback.
Helmv1.0.0A repeatable browser walkthrough
metacodex added the adapted flow-screenshots-pdf skill and printable template, alongside model-routing guidance.
metacodexv1.0.6
About the material shown here
Product views.
Release notes.
Selected screenshots and edited recordings accompany changes checked against repository history through October 2026.
Current product and capture
The portable configuration, browser bridge and named skills are present in inspected source. The Workbench Lab replays eight actual manager commands from committed metacodex source in two fictional homes: a refused conflict, deliberate adoption, a shared update and preserved local-file hashes. Its synthetic settings and empty hooks are disclosed; it does not launch agents, authenticate or synchronise the fleet. Other workbench installations and observer behavior have not been audited as one live deployment inventory.
Source and historical evidence
The configuration story begins with metaclaude in December 2025. September records and Helm’s checked v1.0.0 tag establish its cross-runtime protocol. Commit 666911e contains the two artifact-rule corrections, supported by the archived review and follow-up source check; those checks did not exercise a screenshot tool. October bridge, observer and PDF records support the browser workflows; metacodex’s checked v1.0.6 tag contains the PDF skill. Private configuration, session details and machine identities are excluded.
About the demonstrations
Product imagery uses selected material prepared for this portfolio. Recordings show the interface in use. Architecture graphics describe system relationships and do not show live service availability.