v0.2.3

Documented

• The README is rewritten in Simplified Technical English. Metaphor verbs ("wires", "rides", "anchors", "comes along") became literal verbs ("lets", "uses", "opens at", "stays intact"). Long sentences split under the 20- and 25-word limits. Conditions moved before their commands. The intro theatre metaphor and tagline stay — STE deletes persuasion by design, so the brand voice is preserved. Code blocks, CLI flags, and file paths are untouched.
• Added the assets/greenroom-preview.png the README already renders.

---

Authored by Jesse Robbins (@jesserobbins)

v0.2.2-alpha — Greenroom

Changed

• Document the copy-only trust boundary of collect.
• Clarify that recovered content is read as bytes and never executed or interpreted.
• Bump the greenroom skill to 0.2.2-alpha.

v0.2.1-alpha — recognize skills-CLI installs

Fixed

• install.sh recognizes a standalone install made by npx skills add. The CLI
  does not copy into ~/.claude/skills/<name>: it installs to
  <root>/.agents/skills/<name> and symlinks .claude/skills/<name> at that.
  The check required a real directory, so a working CLI install read as "a
  symlink into somewhere else" — the run refused to link the commands, exited
  non-zero, and told the user to delete a perfectly good copy of greenroom. A
  symlink now counts when it is neither ours (which must refresh) nor another
  greenroom checkout's (which must repoint). Found by running the real CLI: every
  fixture in the suite had built a real directory, so the shape the CLI actually
  produces was never exercised.

v0.2.0-alpha — dual-channel distribution

Added

• A GitHub Actions workflow runs tests/smoke.sh on every pull request and on
  pushes to main, on Linux and macOS, with PyYAML installed so the frontmatter
  check takes the full parse rather than its degraded name-only fallback. The
  suite is the only guard on the distribution shape; leaving it to local runs
  left every guarantee resting on a developer remembering.
• greenroom installs as a standalone skill on any agent:
  npx skills add jesserobbins/greenroom. The skills/greenroom/ directory is
  now self-sufficient — it carries its own scripts/ and templates/ — because
  that directory, and nothing else, is what the skills CLI copies. A smoke test
  drives a full scaffold from an isolated copy of it to keep that guarantee
  honest.

Changed

• Breaking: the skill is renamed setup → greenroom-setup → greenroom,
  so the plugin invocation is now /greenroom:greenroom and a standalone install
  is /greenroom. The npx skills CLI derives a skill's install directory and
  @handle from the name: field, so a bare name: setup installed as
  ~/.claude/skills/setup/ and collided with any other ecosystem skill named
  setup; the name had to carry the greenroom identity. greenroom-setup was
  the first fix, but it covered only one of the skill's four subcommands, so the
  final name is the bare project name. Plugin users need no action; install.sh
  migrates a manual install in place.
• scripts/ and templates/ moved from the repo root into
  skills/greenroom/. greenroom.py resolves templates relative to its own
  location, so this needed no code change.
• Slash commands are now hollow triggers that hold no logic — all behaviour
  lives in the skill. npx skills never reads commands/, so anything encoded
  there is unreachable for standalone users.
• SKILL.md is a router: detail moved into skills/greenroom/references/, which
  costs nothing until read, and flag documentation defers to --help. The
  always-resident file drops from ~3,300 words (~22.8 KB) to ~1,400 (~8.8 KB) —
  roughly 60% — with no functionality removed; the detail now sits in references/
  and is paid for only when a task needs it. Smoke tests assert a word ceiling so
  it cannot silently regrow.
• install.sh no longer builds a ~/.claude/skills/greenroom/ script-root
  shim; the skill directory carries its own script. It migrates the old shim and
  the stale greenroom-setup link, both ownership-checked.
• Smoke tests now parse SKILL.md frontmatter as YAML rather than asserting a
  fixed description string. The skills CLI silently drops a skill whose
  frontmatter fails to parse, so the class of bug matters more than the instance.

Fixed

• install.sh no longer replaces a symlink the user owns. It refreshed any
  symlink at the skill path unconditionally, so a link the migration correctly
  declined to touch was silently repointed one step later. Refreshing is now
  ownership-checked like migration, and both checks resolve relative symlink
  targets instead of only absolute ones.
• The old script-root shim is dismantled only when every entry in it is one of
  the two links we created — the decision is made before anything is removed, so
  a SKIP migration: ... leaving it untouched is now true. It previously tore
  out our links and then discovered rmdir could not empty the directory,
  leaving the user's old script fallback broken with no skill linked over it. OS
  noise (.DS_Store, Thumbs.db, .localized) does not count as a file the
  user left, and a directory that is empty or holds nothing but that noise is
  simply removed rather than called foreign — either one previously failed the
  whole install over something invisible. It is also
  recognized when those links merely dangle, which is what they do in the common
  upgrade: old clone deleted, greenroom re-cloned elsewhere, install.sh re-run.
• A dangling symlink at the skill path is replaced rather than skipped, and a
  greenroom-setup link is migrated away whether it dangles or points into
  another live checkout. Our own links become dangling the moment the clone is
  moved, skipping them made a re-run from the relocated clone silently install
  nothing, and a stale one left the retired skill name registered forever. The
  command names (new.md, add.md, sync.md) are generic enough that a user
  may have bound them to their own repo, so a dangling link there — an unmounted
  volume, a moved clone — is left alone and reported.
• A copied greenroom-setup install (from npx skills add ...@greenroom-setup)
  is reported with the command to remove it. It is not a link install.sh made,
  so it is not ours to delete, but left unmentioned the retired skill name keeps
  resolving forever. The migration covers every shape that path has taken —
  skills/setup (0.1.4–0.1.7), skills/greenroom-setup (0.1.8), and a hand
  repointed link at the renamed skills/greenroom; whose checkout the target
  belongs to is the test, not what it declares. A copied
  ~/.claude/skills/setup/ from before the first rename is reported too, with
  ownership proven rather than assumed, since setup is the generic name the
  whole rename chain was about.
• Symlink ownership no longer fails for a target sitting directly under /,
  where the resolved path picked up a doubled slash and matched nothing.
• install.sh is tracked executable. The README's manual-install block says
  ./install.sh, which failed with "permission denied"; every test invoked it as
  bash install.sh, so nothing caught it.
• Re-installing from a second clone repoints the links instead of hard-failing —
  including the ancient root symlink, which holds no SKILL.md to identify it
  and is now recognized by the plugin manifest at its target.
  Ownership was proven only against the current $REPO_DIR, so a link from an
  older clone that still existed on disk was neither ours nor dangling: the run
  skipped it and exited non-zero. A link into any directory that declares itself
  a greenroom payload is now recognized and repointed.
• The hollow slash commands are no longer registered when the skill they invoke
  did not install. They only say "invoke the greenroom skill", so linking them
  anyway just moved the failure from install time — where the remedy is printed
  — to use time.
• A checkout holding no skills/*/SKILL.md — a partial or corrupt clone — fails
  loudly instead of reporting Done. 0 skill(s) and registering the commands.
• The greenroom-setup migration waits until the replacement skill has actually
  linked. A 0.1.8 user has both artifacts, so a blocked shim path used to leave
  them with the old name removed and the new one never installed — strictly worse
  off than before the run.
• A standalone install already sitting at ~/.claude/skills/greenroom (from
  npx skills add -g, the install the README leads with) is reported and left
  alone. It is a real directory, so running install.sh on top of it used to
  fail the run and tell the user to remove a perfectly good copy of greenroom.
• An install that installs no skill now exits non-zero and prints the remedy,
  instead of reporting Done. 0 skill(s) and leaving the user's /greenroom
  quietly pointing at an older clone.
• SKILL.md documents how to find scripts/greenroom.py — a resolver covering
  the plugin, npx skills add, and manual-clone layouts — rather than naming an
  absolute path the agent had no way to derive. It walks up from $PWD for the
  project-local npx skills add layout the README leads with, so it still
  resolves once the agent has cd'd into a repo — skipping $HOME itself, which
  is an ancestor of nearly every cwd and would otherwise let a global install win
  the project tier, and stopping at the project boundary so an unrelated
  ancestor's .claude/ cannot outrank the plugin cache (a greenroom wrapper is
  the one boundary the walk crosses, since it sits above its repos); the
  plugin-cache tier sorts on
  the version directory alone, so neither a lexically-first version nor a second
  cached marketplace owner can outrank the newest; the
  script is invoked through python3 so a lost exec bit is not fatal, and an
  unresolved path fails loudly instead of running python3 "". A smoke test
  extracts the snippet from SKILL.md and runs it against each install shape.
• The skill's description no longer contains an unquoted colon-space (layout:
  → layout —). The npx skills YAML parser silently dropped the skill when the
  plain-scalar description value contained : , which made greenroom
  undiscoverable via npx skills add and unlistable on skills.sh.

v0.1.8-alpha

Added

• A .greenroom marker file at the wrapper root is now the editor-neutral
  signal that identifies a greenroom wrapper. new, retrofit, and sync
  all write it (write-if-absent), and sync adds it to older wrappers that
  predate it. Wrapper detection no longer depends on the VS Code workspace file.
• --workspace / --no-workspace flags on new, retrofit, and sync to
  force or suppress the <project>.code-workspace write regardless of detection.

Changed

• The <project>.code-workspace file is now written only when a VS Code-family
  editor is detected — code, cursor, codium/vscodium, or windsurf on
  PATH, or an existing .vscode/ or *.code-workspace in the wrapper. A terminal-only
  setup (any editor that reads AGENTS.md natively) no longer gets an unused
  workspace file. Existing wrappers are unaffected: a present *.code-workspace
  is itself a detection signal, so sync keeps refreshing it.

v0.1.7-alpha

Changed

• The plugin and marketplace author/owner email is now correct

v0.1.6-alpha

Fixed

• Every shell command greenroom prints for you to paste now quotes its path
  arguments, so they stay valid when a wrapper, repo, or private-dir path
  contains a space. Affects the stale-cwd cd <wrapper> hint, the
  gh repo create <owner>/<repo> --source=<path> repo-creation offer, and
  collect's git -C <private> status follow-up. Previously a path like
  ~/My Projects/foo produced a command that split on the space when pasted.

v0.1.5-alpha

Added

• add/retrofit now prints a note when it moves the repo out from under the
  shell you ran it in (running add from inside the repo). That shell keeps
  a stale handle to the moved directory, so ls/pwd there look wrong until
  you cd to the wrapper; the note says so and gives the exact cd. The
  layout on disk is correct — this is cosmetic, not data loss.

v0.1.4-alpha

Changed

• The skill now lives at skills/setup/SKILL.md instead of a root SKILL.md,
  so it invokes as /greenroom:setup instead of the stuttering
  /greenroom:greenroom. Auto-activation (by description) is unchanged. ([#3])
• install.sh now links each skill under a greenroom- prefix
  (~/.claude/skills/greenroom-setup), so a manual install invokes
  /greenroom-setup instead of a generic /setup that could collide with your
  own skills. For script resolution it builds ~/.claude/skills/greenroom/ as a
  plain directory holding scripts/ and templates/ symlinks, so the commands'
  script-path fallback and the collect examples still find greenroom.py
  without exposing the plugin manifest (which Claude Code would otherwise
  auto-load as a greenroom@skills-dir plugin).
• install.sh is safer around a path it does not own: if ~/.claude/skills/ greenroom is an unrelated symlink or a real file, it skips the script-root
  setup with a notice (rather than writing scripts//templates/ through the
  symlink into your directory, or aborting the whole install) and still links
  the skill and commands. The README's manual-install section now documents the
  greenroom-setup skill link and the greenroom/ script-root.

v0.1.3-alpha

Fixed

• sync no longer treats $HOME (or the filesystem root, or a standard system
  directory) as a greenroom wrapper, even when a stray non-greenroom
  .code-workspace is present or --wrapper points there explicitly. A single
  unrelated workspace file could previously make greenroom scaffold
  ~/CLAUDE.md and other wrapper files into the home directory, where
  CLAUDE.md is loaded into nearly every session. ([#4])
• retrofit now refuses when the resolved wrapper is a forbidden root. It
  previously checked only the parent, so retrofitting an already-wrapped repo
  whose wrapper is the boundary ($HOME or GREENROOM_ROOT) could still
  scaffold into it; the final target now clears the same guard sync uses.
• Wrapper detection reads .code-workspace files as UTF-8 (matching the write
  side) and skips an undecodable file instead of crashing classification.

Added

• A .code-workspace now only marks a directory as a greenroom wrapper if it
  carries the greenroom sentinel {"greenroom": {"wrapper": true}}; greenroom
  stamps this into every workspace it writes. A generic greenroom key alone
  does not qualify, and a pre-sentinel wrapper still classifies via its
  -private sibling.
• Optional GREENROOM_ROOT env var: a boundary greenroom never crosses upward.
  Set it to your projects' parent dir (e.g. export GREENROOM_ROOT="$HOME/GitHub").
  greenroom refuses it as a scaffold target but accepts it as a parent, so
  new/retrofit can create a project directly under your projects dir while
  sync still refuses to treat the boundary itself as a wrapper.
• greenroom now states its supported platforms (macOS and Linux; Windows via
  WSL2) and refuses to run on native Windows.
• The forbidden-root floor of standard $HOME subdirectories now includes the
  XDG dirs that differ on Linux (Videos, Templates) alongside the macOS
  ones, for parity across both supported platforms.