feat(opencode-workspaces-plugin): add Daytona workspace adaptor plugin for OpenCode

[jamesmurdza]

Description

Adds apps/opencode-workspaces-plugin: an OpenCode plugin that provisions Daytona sandboxes as remote workspaces via OpenCode's experimental workspace-adaptor API.

This PR was originally ported from daytonaio/daytona#4504, with Nx build config replaced by the standalone npm layout used elsewhere in this repo, and fixing several issues in the original plugin. It is not published and does not change packages/opencode-plugin.

This app has extensive test coverage, including end-to-end tests that run a full user flow in the actual OpenCode TUI with the plugin installed.

Upstream bugs in OpenCode

To make this PR work, two PRs were raised in OpenCode:

• fix(server): don't forward host directory to remote workspace anomalyco/opencode#40136 — the actual fix: strips the directory query param before proxying, so the remote falls back to its own project root. Once this ships in a release and the plugin's pinned OPENCODE_VERSION is bumped to include it, the symlink workaround can be removed.
• fix(server): log upstream 5xx bodies from proxied workspace requests anomalyco/opencode#40135 — logs upstream 5xx bodies on the host proxy, making this class of failure (previously an opaque 500 buried in the sandbox) diagnosable from the host log.

These PRs have both been merged and shipped.

Related issue(s)

Scope

• This PR touches no more than one package/app (release routing is by changed file path).
• The PR title follows Conventional Commits
  (e.g. feat(pi-extension): ..., fix(adk-plugin): ...) — it becomes the squash-commit
  message that release-please reads.

Checks

• I ran the affected package's lint/build/test locally and they pass.
• I updated documentation where relevant.

Legal

• Every commit is signed off for the
  DCO (git commit -s); the sign-off matches the
  commit author.
  1 of 15 commits signed off — needs git rebase --signoff main.
• I agree to the
  Contributor License Agreement.
  On my first PR, the CLA assistant will comment and I will reply to sign (once).

[github-actions]

All contributors have signed the CLA. ✅ Thank you!
_{Posted by the CLA Assistant Lite bot.}

[jamesmurdza]

I have read the CLA Document and I hereby sign the CLA

[cubic-dev-ai]

All reported issues were addressed across 13 files

<sub>Tip: cubic can generate docs of your entire codebase and keep them up to date. Try it here.

Re-trigger cubic</sub>

[mislavivanda]

@jamesmurdza Reviewed the full source and live-tested everything against real Daytona sandboxes (5
workspaces created/destroyed over the session, zero leaks — the partial-create cleanup
demonstrably deletes its sandbox on failure). Typecheck, plugin, and integration suites are
green; the e2e TUI suite passes too, with one environment caveat documented below. Solid
work overall — the create() error path, the bounded health polls, and the log-based test
sweeps all held up under deliberate abuse.

A few requests before merge, all small:

1. Normalize "adaptor" → "adapter"

The upstream API spells it adapter everywhere (WorkspaceAdapter,
/experimental/workspace/adapter, workspace.adapter.list()); the PR mixes both variants,
sometimes in the same file (README's "Adaptor registration" section curls
/workspace/adapter three lines later). Occurrences: daytona/index.ts lines 7, 192
(variable), 360–361; plugin/index.ts line 8; plugin README lines 95, 97, 171, 183;
integration.test.ts line 131; root README table row. Verified the rename typechecks and
tests still pass.

2. README: "Running the e2e on any machine" section

The e2e drives the real TUI, so it inherits the developer's shell and global opencode
config — and fails on machines that differ from a vanilla setup. Both failure modes are
fixable purely via environment, no test changes needed. Verified on a machine where the
test fails as written and passes with:

SHELL=/bin/sh XDG_CONFIG_HOME=$(mktemp -d) npm test -- test/e2e-tui.test.ts

Please add a short README subsection under Running Tests documenting this, with the why:

• SHELL=/bin/sh — the test types cd … && opencode into a fresh tmux pane immediately
  after new-session; heavier shell inits (zsh + plugins) eat part of that line, so the
  TUI silently starts in the wrong directory (plugin log then shows worktree=<repo root>
  instead of the temp project). A plain sh pane has no init to race.
• XDG_CONFIG_HOME=$(mktemp -d) — isolates ~/.config/opencode. With a custom default
  agent configured globally (e.g. an oh-my-opencode setup), the warped session's first
  prompt dies silently on the remote: session created, zero assistant output, zero tokens,
  $0.00, no error surfaced. Isolated A/B on the same machine: vanilla config → reply in
  ~1s; custom default agent → dead session. Auth lives under XDG_DATA, so provider keys
  still forward.
• OPENCODE_SERVER_PASSWORD must be unset — all three suites poll /global/health
  unauthenticated and fail with opaque 401s when it's exported.

The custom-agent behavior itself looks like upstream workspace-API territory (remote can't
resolve host-only agents and swallows the prompt without an error) — worth filing as
feedback to the opencode team alongside the two PRs already linked in the description, and
worth a one-line "known limitation" note in this README.

3. Stale comment: the TUI-render "sync gap"

The e2e header (test/e2e-tui.test.ts:16–20) says OpenCode's global sync delivers state
counts but not message content, so the reply never renders in the host TUI and is only
logged informationally. On 1.18.15 the reply DOES render (the test itself prints "TUI
rendered the reply" now). Suggest updating the comment, and optionally promoting the
informational check to a hard assertion.

4. Minor / documentation

• create() unconditionally overwrites the sandbox repo's opencode.json (index.ts:292–300)
  and strips committed .opencode/ from the upload (index.ts:263) — repos carrying their own
  opencode config lose it remotely. Fine as a v1 trade-off, but worth a README note (or
  merging instructions into an existing config later).
• Only the last commit is uploaded (clone from worktree, --depth 1) — uncommitted changes
  don't travel. Worth one README sentence.
• Provider credentials are forwarded into sandbox envVars (that's how the remote can call
  models) and are visible in the Daytona dashboard — inherent to the design, deserves a
  one-line note.
• PR body is stale: it says "1 of 15 commits signed off", but all commits now carry matching
  Signed-off-by trailers.

On the cubic comments

• Temp-dir Date.now() collision — agreed, mkdtemp is a strict improvement; please take it.
• Unbounded create fetch — already bounded by bun's --timeout 180000 and the afterAll
  log-sweep exists precisely for hung creates; optional, and if added the bound should be
  generous (creates legitimately took 16–50s in live runs).
• Shared plugin log between test files — the path is the plugin's fixed log which tests
  only read (append-only + per-file offsets, files run sequentially), so the described
  orphan scenario can't occur; making the plugin's log path env-configurable would be nice
  future polish but isn't needed for merge.

[jamesmurdza]

@mislavivanda Thank you for the feedback. I've implemented all the changes except instead of documenting the second one (e2e test inherits the developer's shell and global opencode config) I updated the test so this doesn't happen.

Summary of changes:

1. adaptor → adapter — normalized the spelling everywhere to match the upstream API
2. Unique scratch dir per create() — replaced the Date.now()-based temp path with mkdtemp
3. e2e reproducibility — rather than documenting a manual env incantation, the e2e now pins its own environment at the tmux spawn (SHELL=/bin/sh, a temp XDG_CONFIG_HOME, and OPENCODE_SERVER_PASSWORD unset).
4. README Limitations note — documented three trade-offs: 1) only the last commit is uploaded (shallow clone), 2) a repo's own .opencode//opencode.json isn't carried over, and 3) provider credentials are forwarded into the sandbox env.

And (of course) I've run the test suite after these changes to make sure everything works.