🦋 New version release

[brentrager]

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

@smooai/smooth@0.4.0

Minor Changes

• e7e533e: MCP servers, CLI-wrapper plugins, and project-scoped config.

  • th mcp add/list/remove/test/path for stdio MCP servers (Playwright,
    GitHub, filesystem, etc.). Servers register as <server>.<tool>.
  • th plugin init/list/path/remove for file-based CLI-wrapper tools at
    .smooth/plugins/<name>/plugin.toml. Plugins register as
    plugin.<name> and run the configured shell command with
    {{placeholder}} substitution.
  • Both MCP and plugins resolve from ~/.smooth/ (global) and
    <repo>/.smooth/ (project); project entries shadow global on
    name collision. --project flags on add / init / remove /
    path scope to the repo.
  • No trust gate on loading configs — Narc screens every tool call
    (CliGuard, injection, secrets, LLM judge) and the microVM contains
    the blast radius. See SECURITY.md and docs/extending.md.

• 3657db9: Chat gets sessions. The smooth-web chat page now lists prior
  sessions in a sidebar, each persisting its own message history in
  the Dolt session_messages table. New /api/chat/sessions
  endpoints (create/list/get/delete + messages). The LLM receives the
  last 50 messages on every turn so multi-turn context is preserved.
  Session titles auto-rename to the first 60 chars of the opening
  prompt. Chat layout fixed so the input row sits flush to the
  viewport bottom (no stray scroll).

• 3657db9: Retire SQLite. Pearls, sessions, memories, config, and worker metadata
  all live in the Dolt store at ~/.smooth/dolt/ (home) or
  <repo>/.smooth/dolt/ (per-project). smooth.db is gone; the
  dashboard reads "Dolt store (pearls + config)" instead of
  "Database (SQLite)". th pearls migrate-from-sqlite removed —
  transitional tool, no longer needed.

• 3657db9: Run in sandbox — the agent does its work in a microVM, you review it live.

  • smooai/smooth-operator image (unified — agent installs toolchains at
    runtime via mise; covers node/python/rust/go/bun/deno/~140 more).
  • th run [pearl-id] [--keep-alive] [--image ...] [--memory-mb N] —
    dispatches via /api/tasks, streams SSE events, optionally keeps the
    VM alive for dev-server review.
  • th operators list / kill <id> — see and tear down running VMs.
  • th cache list / prune / path / clear — project-scoped sandbox
    cache at ~/.smooth/project-cache/<name>-<hash>/, bind-mounted
    at /opt/smooth/cache inside the VM. LRU prune by mtime.
  • Auto-forward common dev-server ports (3000, 3001, 4000, 4200, 5000,
    5173, 8000, 8080, 8888) on keep-alive runs; print reachable
    http://localhost:<host> URLs after the agent completes.
  • Per-run memory override threaded through
    TaskRequest → SandboxConfig.

• 4f91014: th service — background service wrapper for th up.

  User-level install by default on all three platforms:

  • macOS: writes ~/Library/LaunchAgents/com.smooai.smooth.plist,
    drives launchctl bootstrap gui/$UID.
  • Linux: writes ~/.config/systemd/user/smooth.service, drives
    systemctl --user enable --now. Prints a hint to run
    loginctl enable-linger so the service survives logout.
  • Windows: creates a logon-triggered Scheduled Task via schtasks.

  Commands: install [--system], uninstall, start, stop, restart,
  status, logs [-f]. --system prints the system-level artifact +
  install instructions to stdout instead of running under sudo.

  Logs stream to ~/.smooth/service.log and ~/.smooth/service.err.

Patch Changes

• 11f0c00: Fix 5 cast_integration tests that had been failing in CI since the
  Wonk bearer-token auth was added in f7676d8. The release workflow
  has been red for ~8 days, stranding 12 changesets and blocking every
  version bump.

  Root cause: ALLOW_EXAMPLE_POLICY has [auth] token = "test-token",
  so Wonk's require_operator_token middleware rejects any request
  without Authorization: Bearer test-token with a 401 (empty body).
  The tests built reqwest::Client::new() directly and called
  .post(...).json(...).send().await.unwrap().json().await.unwrap(),
  which panicked at the final .json() with
  reqwest::Error { kind: Decode, source: Error("EOF while parsing a value") }.

  Fix: introduce TEST_AUTH_TOKEN = "test-token" next to the policy
  fixture, attach .bearer_auth(TEST_AUTH_TOKEN) to every direct Wonk
  request, and switch spawn_goalie to WonkClient::with_auth so its
  /check/* calls carry the header too. The goalie_forwards_..._for_allowed_request
  test had surfaced as a 403 != 200 assertion for the same reason —
  Goalie was failing its auth to Wonk and correctly denying the request.

  Narc / Scribe / Archivist tests were never affected (those services
  do not require auth).

• 3b1a88a: Diagnostic logging on the sandboxed dispatch path so we can tell
  why th run / th code --headless fail when they do:

  • bill::exec_sandbox logs exec start + non-zero-exit with
    captured stdout/stderr tails (was silent before, making code=-1
    failures opaque).
  • Dispatch handler now runs a preflight /bin/sh probe against
    the sandbox before exec-ing the runner — surfaces whether
    bind-mounts landed, whether the runner binary is visible + executable,
    and what the guest's /opt actually contains.

  Pearl th-1ec3ce (P0) tracks the underlying issue: on plain alpine,
  microsandbox's bind-mounts aren't reaching the guest at all, so
  every sandboxed dispatch fails with exit=-1 stderr="". Fix requires
  digging into microsandbox's mount-arg plumbing; these changes just
  give us the visibility to do it.

• 72f7eef: Fix P0 (th-1ec3ce): sandbox bind-mounts not landing in the guest VM.

  The microsandbox guest agent does not mkdir -p mount targets before
  calling mount -t virtiofs — mounts to paths that don't pre-exist in
  the rootfs (/opt/smooth/bin, /opt/smooth/policy, /workspace)
  silently fail. We were falling back to plain alpine because our
  custom smooth-operator image was only in Docker Desktop's local
  store and microsandbox couldn't pull it; alpine has an empty /opt,
  so every mount missed.

  Fix: publish smooai/smooth-operator and smooai/boardroom images
  to GitHub Container Registry (public), and default to pulling from
  there. The Dockerfile.smooth-operator pre-creates /workspace,
  /opt/smooth/bin, and /opt/smooth/cache/mise — so every bind-mount
  target now exists before the guest agent tries to mount on top of it.

  • SandboxConfig default image: alpine → ghcr.io/smooai/smooth-operator:latest
  • th run default: smooai/smooth-operator:latest → ghcr.io/smooai/smooth-operator:latest
  • scripts/build-smooth-operator-image.sh + build-boardroom-image.sh:
    default SMOOTH_IMAGE_REPO to ghcr.io/smooai/..., add --push
    flag so one command builds + publishes.
  • Preflight probe now confirms mounts land: /opt/smooth/bin/smooth-operator-runner
    is executable inside the VM and the runner boots as expected.

  Users can override with SMOOTH_WORKER_IMAGE / SMOOTH_OPERATOR_IMAGE
  if they publish a fork to a different registry. Public pulls from
  ghcr.io/smooai/* require no auth.

• 3657db9: Buildable OCI images for the microVM cast:

  • docker/Dockerfile.smooth-operator + scripts/build-smooth-operator-image.sh
    → smooai/smooth-operator:<version> (alpine + mise + runner).
  • docker/Dockerfile.boardroom + scripts/build-boardroom-image.sh
    → smooai/boardroom:<version> (alpine + boardroom bin + smooth-dolt).
  • Both scripts delegate to the existing cross-compile flow
    (build-operator-runner.sh / build-boardroom.sh).
  • Fixed a latent package-name bug in build-boardroom.sh
    (-p smooth-bigsmooth → smooai-smooth-bigsmooth).

  Still pending: registry publish on release so microsandbox can
  pull without Docker on end-user machines.

• 3657db9: Pearl fixes:

  • /api/pearls + /api/projects/pearls default to unbounded
    (?limit=0). The dashboard was silently capped at 100 — a repo
    with 150+ pearls showed "100 closed, 0 open" when the open ones
    were past the limit. LLM tool callers still get a 100-row
    default via list_pearls().
  • PearlStore::close is now invoked on every error-path exit of
    the sandboxed dispatch handler (runner not found, workspace
    create failed, LLM provider missing, runner exited non-zero).
    Previously only exit-0 closed the pearl; leaked Task: ...
    pearls accumulated from E2E runs.
  • install:th now re-adhoc-signs a neighbor smooth-dolt binary
    in ~/.cargo/bin/ so scp'd copies work under launchd
    without a manual codesign.

• 3657db9: th doctor --init-home-repo scaffolds ~/.smooth/ as a git repo for
  backup / cross-machine sync. Writes a .gitignore that excludes
  secrets (providers.json), service logs, audit logs, the Dolt
  store (has its own push/pull), the project cache, and ephemeral
  debug captures. Idempotent. Optional --remote <url> adds origin.

• 3657db9: Wordmark + cast renames in user-facing surfaces:

  • th CLI + web dashboard now say "Big Smooth" (not "Leader") and
    "Smooth Operators" (not "Sandbox").
  • New horizontal logo (images/logo.png, crates/smooth-web/web/public/logo.svg)
    replaces the old mark.
  • th up / th status / th doctor banners render "Smooth" with
    the logo's gradient colors via ANSI 24-bit truecolor escapes
    ("Smoo" orange→pink, "th" teal→blue).
  • /health service field renamed smooth-leader → big-smooth.
  • SMOOTH_SANDBOX_MAX_CONCURRENCY env + th up --max-operators N
    flag expose the previously hardcoded pool cap of 3.

• 3b1a88a: Wordmark gradient ("Smoo" orange→pink, "th" teal→blue) live in
  th up / th status / th doctor banners via a new
  gradient module in smooth-cli. Helper smooth() stitches the
  two halves into the full word; smoo_ai() covers the Smoo AI
  brand. Pure ANSI 24-bit truecolor, no new deps. 4 unit tests.