Space Architect 🚀

Task-scoped workspaces that double as Architect Loop projects — for humans and their agents! ✨🛰️

space-architect is a Ruby toolkit for spaces — task-scoped project workspaces that hold repos, notes, and artifacts under one obvious filesystem root — fused with the Architect Loop: a structured judgment-and-build cycle for you and a fleet of headless AI builders. 🌌

The gem ships three composable binaries over clean library seams:

Each is a first-class executable. architect also forwards architect space … and architect src … to the other two, so a project can drive everything from one command when that's handier. 🎀

A space is just a regular directory with a tiny YAML identity file and room for everything a task needs:

~/architect/spaces/20260531-name-of-space/
  space.yaml        # identity: id, title, status, repos, notes, tags
  README.md
  repos/            # cloned (or copy-on-write'd) repositories
  notes/            # scratch, prompts, logs
  architecture/     # iteration files: I<NN>-<name>.md + ARCHITECT.md index + BRIEF.md
  build/            # lane worktrees + scratch (build/<id>-<lane>/)
  tmp/              # workspace-local temp — use this instead of /tmp

Run a command from anywhere inside a space and it just works — space and architect walk up from $PWD until they find the nearest space.yaml. No "current space" state to get out of sync; where you are is the space you mean. 🧭

Add it to your Gemfile:

gem "space-architect"

bundle install

Or grab it yourself:

gem install space-architect

The gem is space-architect; it installs three executables — space, architect, and src. 🎀

# Spaces
space init                                        # create XDG config + state files
space new "Name of Space"                         # blast off a new space 🚀
space new "Name of Space" -r org/repo -r org/lib  # …with repos cloned in (repeat -r)
space list                                        # see all your spaces
space show                                        # show the space you're standing in

# Evergreen checkouts (copy-on-write sources for fast provisioning)
src repo add github.com/example-org/example-app     # tend it 🌲
src sync                                            # one sync pass
src status                                          # per-repo evergreen status

# Architect Loop (run from inside a space)
architect install-skills                          # install agent skills (once per machine)
architect init                                    # scaffold ARCHITECT.md + architecture/
architect new my-feature                          # scaffold the next iteration file
architect freeze my-feature                       # lock the Acceptance Criteria ❄️
architect dispatch my-feature lane-a              # send a headless builder to work

# Containers (run from inside a space) — a portable, reproducible-by-SHA image of the space
space pack                                         # render build/oci/ (Dockerfile + entrypoint)
space build                                        # pack + build & tag <space-id>:<git-sha>
space run                                          # run it — login shell, auth from your env

space init
space new "Name of Space"
space new "Name of Space" -r org/repo -r example-tools/alpha -r example-tools/beta
space list                                   # alias: space ls
space show 20260531-name-of-space
space path 20260531-name-of-space
space current                                # based on $PWD
space show                                   # based on $PWD
space status                                 # report: metadata + loop status (based on $PWD)
space status 20260531-name-of-space          # report another space
space status done                            # set: active | paused | done | archived
space status 20260531-name-of-space done
space config set default_provider github.com
space config set default_organization example-org
space repo add example-app                   # github.com/example-org/example-app
space repo add example-tools/alpha example-tools/beta
space repo add gitlab.com/example-org/api
space repo resolve example-app example-tools/async
space repo ls                                # alias: space repos ls
space use 20260531-name-of-space             # records recent state, prints the path
space ls --color=always                      # auto | always | never (--colors also accepted)

space status is report-or-set: with no status keyword (bare, or with just a space id) it reports the space — its metadata plus a compact loop-status block (project status, current iteration, derived state) when the space runs an Architect project, quietly omitted otherwise; pass a status keyword (active, paused, done, archived) to set it instead. 🔁

Repos are passed with a repeatable -r flag (-r org/repo -r org/lib); the comma form (-r a,b) works too. Space ids are date-prefixed (20260531-name-of-space) so they sort naturally, and duplicate names on the same day get a counter (…-name-of-space-2). 📅

Everything space does is also reachable as architect space … from within a project.

A space is self-describing enough to become a container. space pack renders a portable OCI build context from the space; space build packs and builds it into a reproducible-by-SHA image; space run runs that image with your auth injected at runtime and stateful paths bind-mounted back to the host. 🐳

space pack                    # render build/oci/ (Dockerfile + entrypoint + ignore file)
space pack -o /tmp/ctx        # …to a different output directory
space build                   # pack, then build & tag <space-id>:<git-sha> and :latest
space run                     # run <space-id>:latest — login shell, auth from your env
space run architect status    # …or run a one-off command instead of the login shell
space run --tty               # force an interactive TTY (default: auto-detect)

What lands in the image. The context copies the whole space tree (filtered by a generated Dockerfile.dockerignore) onto a ruby:4.0.5 base with git, the Claude Code CLI, and the space-architect gem — installed from the in-space repos/space-architect checkout when present (a pinned build), else from RubyGems. Secrets never enter the layers: .env, *.key, *.pem, ssh keys, build/, and tmp/ are all excluded by the generated ignore file. 🔒

Reproducible by SHA. space build tags the image <space-id>:<sha>, where <sha> is the space repo's 12-char HEAD (suffixed -dirty when the tree has uncommitted changes), plus a moving :latest. Same commit → same tag → same image. It drives the container CLI, but the output is an ordinary OCI/Docker build context, so docker build -f build/oci/Dockerfile . (from the space root) works just as well.

Auth stays out of the layers. space run injects only the auth environment variables that are actually set — ANTHROPIC_API_KEY, CLAUDE_CODE_OAUTH_TOKEN, ANTHROPIC_BASE_URL — with -e at run time, so credentials live in your shell, never in the image. 🗝️

Forwarding payload credentials. Beyond the always-on auth trio, a space can declare run.env: in space.yaml — a list of host env var names forwarded into the container at run time. You can also pass space run --env VAR (repeatable) for ad hoc additions. All forwarding is bare -e VAR passthrough: values never appear in argv, ps, or the image. A requested-but-unset var warns on stderr instead of silently failing inside the guest.

run:
  env:                         # run-time: host var names, forwarded as bare -e VAR
    - FIREWORKS_API_KEY
    - OPENAI_API_KEY

Declaring provisioning & persistence — two optional keys in space.yaml:

pack:
  provision:                 # build-time: relative scripts, baked in by `space build`
    - scripts/setup-toolchain.sh
  persist:                   # run-time: absolute guest paths, bind-mounted per run
    - /root/.claude
    - /root/.local/state

provision scripts must live under the space root and be executable. Each script is copied into the image individually (COPY <script> /space/<script>) and then invoked (RUN /space/<script>) before the full space tree lands and before the gem is installed. This ordering is the cache-hygiene guarantee: editing any other space file leaves the provision and gem-install layers cached, so a rebuild completes in seconds instead of minutes. Scripts must therefore be self-contained — they run with only the base system layers and any earlier provision scripts' outputs; they cannot read other space files or call architect/space. Scripts must have the executable bit set; COPY preserves the mode from the build context.

persist paths must be absolute; space run bind-mounts each one from <space>/.state<path> on the host (created on first run) so a container's mutable state survives across runs. Both are validated at pack time.

Like everything else, these are reachable as architect space pack|build|run from inside a project.

src keeps local clones under <src_dir>/<host>/<owner>/<repo> (default ~/architect/src/…) clean, on their default branch, and freshly fetched — so spaces can provision repos by copy-on-write instead of cloning over the network.

src repo add github.com/example-org/example-app     # track + tend a repo
src org add github.com/example-org                  # track a whole org
src sync                                            # run one sync pass (--repo to scope)
src status                                          # per-repo evergreen status table
src clone example-app                               # APFS copy-on-write into $PWD
src daemon install                                  # per-user launchd agent for background sync

When space repo add finds a matching evergreen checkout, it copy-on-writes from it instead of hitting the network — instant provisioning. ⚡ The two surfaces share one layout by design, so they line up with zero configuration:

src repo add github.com/example-org/example-app   # tend it: keep it evergreen 🌲
space repo add example-app                          # copies it instantly ⚡

src has its own --plain / --json output modes for scripting, and its own fish integration (src shell fish install).

The Architect Loop is a structured build cycle for you and headless AI builders. Each loop lives inside a space as a project.

Roles:

• Architect — the judgment role: a strong reasoning model (or you), run interactively. Arbitrates disagreements, writes and freezes iteration files, calls kill/continue, merges builder output. Never writes implementation code.
• Builder — the execution role: a cheaper model run headless via architect dispatch, one per lane in its own git worktree. Reads the iteration's Builder Prompt, does the work, writes raw evidence to build/<id>-<lane>/report.md. Never grades its own work; never edits architecture/.

The loop is model-agnostic — which models fill the two roles is your choice (e.g. a strong Claude model judging a cheaper one on the same plan, or a cross-vendor pairing for more independent review). Set it per dispatch with architect dispatch --model …, or run several pairings head-to-head as a variant set (architect variant add). See docs/DESIGN.md §1–§2 for the reasoning.

Filesystem layout:

architecture/
  ARCHITECT.md              # cross-iteration index; project-wide state
  BRIEF.md                  # durable §-numbered project contract (optional)
  I01-<iteration>.md        # one self-contained file per iteration
build/
  I01-<iteration>-<lane>/   # lane worktree + scratch per dispatch
    run.jsonl               # streamed builder output
    report.md               # builder report (transcribed into the iteration file verbatim)

Iteration file anatomy — one file, grown section by section. You author the content in fresh scratch files; the CLI owns the persistence (each command writes the section, commits it, and prints back what changed). Every committing command takes -m/--message and --message-from <file> — your first line completes the subject after a short canonical prefix (I01 spec: <subject>), the rest becomes the body. The space's git log is the loop's durable memory: write detailed messages.

The freeze ❄️ — architect freeze <iteration> commits the frozen region (Grounds / Specification / Acceptance Criteria), records the freeze_sha, and prints the frozen Acceptance Criteria back. Any change to those sections afterward is an automatic iteration FAIL. The builder never edits the iteration file.

Re-grounding 🧭 — architect init also scaffolds a SessionStart hook that runs architect ground (emitting ARCHITECT.md, BRIEF.md, and the in-flight iteration) so every fresh session starts oriented — the loop leans on fresh-session judgment, and this is what makes picking up cold cheap. Builders inside a lane worktree are never grounded.

Command surface:

architect init                              # scaffold ARCHITECT.md + the space.yaml project: block + SessionStart hook
architect brief new --from <f>              # write the durable project BRIEF.md (authored in a scratch file)
architect new <iteration>                   # scaffold architecture/I<NN>-<iteration>.md
architect section <it> <section> --from <f> # write + commit a section (add -m/--message-from for a detailed commit)
architect freeze <iteration>                # freeze the Acceptance Criteria ❄️
architect worktree add <repo> <it> <lane>   # isolated worktree per lane (2–4 lanes)
architect dispatch <it> <lane> --prompt <f> # copy the lane prompt in + dispatch a builder (--detach to survive long runs)
architect verify <iteration>                # post-flight mechanical checks (reports only)
architect evidence <it> --lane <lane>       # transcribe the builder's report verbatim
architect gate <iteration>                  # run the frozen gate commands, stream raw output
architect merge <it> <lane>                 # integrate ONE judged-passing lane (--no-ff)
architect integrate <it> --lanes a,b        # integrate a set of passing lanes, in order
# landing (PR body + push/PR command) is the architect's end-of-project procedure (see skill)
architect status                            # project state (read-only)
architect variant add|compare|promote …     # competing (harness, model) lanes over one frozen spec
architect research dispatch|status|wait …   # parallel read-only research lanes (see below)

architect help (or bare architect, architect --help) lists these grouped by loop phase — Spec · Build · Judge · Land · Project · Groups, in canonical order — and, inside an architect space, ends with a compact loop-status block (project status + current iteration) so you can see where you are in the loop. 🧭

A typical session:

architect init                                   # first time
architect new my-feature                         # scaffold I01-my-feature.md
architect section my-feature specification --from spec.md -m "pull-based dispatcher seam"
architect freeze my-feature                      # lock it ❄️
architect dispatch my-feature lane-a --prompt tmp/prompts/I01-lane-a.md --detach   # send a builder; poll the report
architect verify my-feature                      # mechanical post-flight checks
architect evidence my-feature --lane lane-a      # transcribe raw evidence
architect gate my-feature                        # run the frozen gates yourself
# … read the diff against the spec, then write the Verdict …
architect integrate my-feature --lanes lane-a    # merge passing lanes → project/<slug>
# landing (PR body + push/PR command) is the architect's end-of-project procedure (see skill)

architect dispatch can push the builder's stream-json to an ingest server for live viewing:

# Push to an already-created run (you supply the full ingest URL):
architect dispatch my-feature lane-a \
  --push-url   $HOST/runs/<id>/ingest \
  --push-token $INGEST_TOKEN

# Create a run and push in one step (requires --push-token = server's INGEST_TOKEN):
architect dispatch my-feature lane-a \
  --push-host  $HOST \
  --push-token $INGEST_TOKEN

--push-host POSTs to <HOST>/runs, parses the new run id from the 201 response, derives <HOST>/runs/<id>/ingest, and streams there; the created run id and ingest URL are printed after dispatch starts. --push-url and --push-host are mutually exclusive, both require --push-token, and neither can be combined with --detach (the push tees the live pipe in-process).

When an iteration needs facts the repo doesn't already have, fan out parallel read-only research lanes — detached claude -p researchers (no Edit/Write/Bash) that you supervise:

architect research dispatch 01-official-api.prompt.md 02-changelog.prompt.md
architect research wait        # tails each lane's run.jsonl; --level 1-4, --quiet, --thinking
architect research status      # status of dispatched runs

Researchers gather; the architect verifies the load-bearing claims against sources and writes the iteration's Grounds section.

architect install-skills installs the bundled architect, architect-research, and architect-vocabulary skills for your harness:

architect install-skills                         # default: claude (~/.claude/skills/)
architect install-skills --provider opencode     # or codex | pi
architect install-skills --project               # into ./… instead of globally
architect install-skills --dry-run               # show what would change

architect-vocabulary loads the system's terms and a short orientation when you're in a space but don't want to run the loop.

Shells can't let a child process change their working directory, so space and src each ship a small fish wrapper function plus completions (commands, subcommands, spaces, statuses, config keys, repo refs). Install into fish's autoloaded directories:

space shell fish install
src shell fish install
exec fish

After restarting fish (or exec fish), space new "…" and space use <id> will cd into the selected space once the command succeeds; every other command keeps normal CLI behavior. 🚪 The functions and completions are written under ~/.config/fish/, so there's no need to edit config.fish. For one-off testing without installing:

space shell init fish | source

Config lives at ~/.config/space-architect/config.yml (XDG-aware) and defaults to:

version: 1
base_dir: ~/architect            # spaces_dir + src_dir hang off this by default
default_provider: github.com
default_organization:
git_clone_protocol: ssh          # ssh | https

Derived defaults: spaces_dir → <base_dir>/spaces, src_dir (evergreen checkout root) → <base_dir>/src. Override either explicitly. View values with space config show; set one with space config set KEY VALUE. Editable keys: base_dir, spaces_dir, src_dir, default_provider, default_organization, git_clone_protocol.

Repos are added to the current space under repos/ and tracked in space.yaml. When an up-to-date evergreen checkout exists at <src_dir>/<host>/<owner>/<name> (e.g. ~/architect/src/github.com/example-org/example-app), space copies it into the space instead of cloning over the network — a copy-on-write clone on APFS. ⚡ Set src_dir empty to always clone:

space config set src_dir ""

Clone URLs default to SSH (git@github.com:example-org/example-app.git); switch with space config set git_clone_protocol https. Multiple repos passed to space repo add are fetched concurrently, up to five at a time, on fibers — no threads, all cooperative. 🧵 After each repo lands, space runs mise trust in it. Each space also gets a workspace-local tmp/ — use it instead of /tmp.

The library is split into three namespaces you can require independently:

• Space::Core — the foundation: config, state, XDG, terminal, git/mise clients, the space store. The space CLI runs on this alone.
• Space::Architect — project state, the builder harness, dispatch, and the research supervisor.
• Space::Src — the evergreen engine (tracking, sync, copy-on-write clone).

require "space_core"       # just spaces
require "space_architect"  # the full loop (pulls in core + src)
require "space_src"        # just the evergreen engine

• Command Reference — every command, flag, and behavior
• Design — the source-backed rationale: the twelve invariant rules (R1–R12), the failure-mode → mitigation table, and why the loop is shaped this way
• Changelog — release history

bundle install
bundle exec rake test       # the full minitest suite
bundle exec rake build      # build the gem into pkg/
bundle exec rake install    # build + install into your user gem home

Bug reports and pull requests are welcome on GitHub at https://github.com/jetpks/space-architect!

Available as open source under the terms of the MIT License.

Made with 💖 and fibers 🧵 by Eric