# ADR 004: Coder Community trial

- HTML version: https://robbiepalmer.me/projects/agent-friendly-remote-development/adrs/004-coder-community
- Project: Agent-friendly Remote Development (https://robbiepalmer.me/projects/agent-friendly-remote-development.md)
- Status: Proposed
- Date: 2026-09-06
- Initiatives: Semi-autonomous Software Development (https://robbiepalmer.me/initiatives/semi-autonomous-software-development.md)

# Context

[ADR 003](/projects/agent-friendly-remote-development/adrs/003-portable-per-user-workspaces)
defines a portable per-user workspace. Direct K3s manifests can provide that
contract. The project would still need to build user accounts, workspace
ownership, provisioning controls, private application access, health reporting,
and lifecycle APIs.

[Coder Community](https://coder.com/pricing) is a self-hosted control plane for
developer workspaces. It can install on Kubernetes with Helm, and its official
Kubernetes template creates a pod with a persistent volume mounted at
`/home/coder`. A template can package the T3 Code image, resource limits,
private endpoint, and workspace volume.

# Decision

Run a one-day comparison between two implementations on the first host:

1. the smallest direct K3s definition that satisfies ADR 003; and
2. the same workspace managed by Coder Community.

Keep my workspace unchanged during the experiment. Select Coder only if one
template can create and remove my brother's workspace, preserve his home,
enforce access and resource limits, and reduce the custom lifecycle code and
operator work.

# Technical fit

Coder would own user and workspace records, Terraform-based create and delete
operations, start and stop controls, health reporting, and access through its
UI, CLI, and API. T3 Code would remain the agent interface inside the
workspace.

The official Kubernetes installation requires PostgreSQL, which adds a service
to operate and back up. Run a small PostgreSQL workload in K3s for the
experiment. A production database location can be decided after the control
plane itself earns a place.

Coder's tunnels may overlap with the project's Tailscale routing. The test must
identify which layer owns authentication and the private application URL. A
pilot user must not need broad operator membership to reach their workspace.

# Licence boundary

Coder Community is licensed under
[AGPL-3.0](https://github.com/coder/coder/blob/main/LICENSE). If this project
modifies Coder and lets users interact with that version over a network, it
must offer those users the corresponding source. That obligation is compatible
with this repository's AGPL-3.0 licence.

Code in Coder's `enterprise` directory uses a
[separate licence](https://github.com/coder/coder/blob/main/LICENSE.enterprise).
The experiment must use Community features and avoid a dependency on separately
licensed enterprise code.

# Acceptance test

Both implementations must complete the ADR 003 workspace test. Record for each:

* lines and concepts in the versioned workspace definition;
* manual steps for creation, login, update, recovery, and removal;
* idle memory and CPU use;
* new state and backup procedures;
* the user's authentication and private-routing flow; and
* failures that require direct cluster or database repair.

Reject Coder if it requires broad operator access, weakens per-user provider
identity, consumes too much of the 8 GB host at idle, or leaves more lifecycle
work than it removes.

# Consequences

The experiment gives the project an executable comparison with a mature open
source control plane. Selecting Coder could remove much of the account and
lifecycle work. It would also add Coder upgrades, PostgreSQL, template
maintenance, and another authentication surface.

Selecting direct K3s keeps the first implementation small. The project would
then own the missing user and workspace lifecycle features and should use the
experiment record to define that work.

---

Markdown index of this site: https://robbiepalmer.me/llms.txt