Skip to main content
Smithers 0.32.0 is a big feature release. The Monitor opens from any browser and from another machine, oneshot and the authoring loop right-size the approach before doing work, review and eval verdicts stop conflating “blocked” with “rejected,” and a run whose workflow you edited can resume instead of forking. Underneath, the runtime moved to Effect 4.

​
Upgrading

Behavior changes are collected in the upgrade notes.

​
Open the Monitor from any browser, on any machine

Pasting a monitor URL into a browser used to hit a 401, because the browser had no way to present your bearer token. The gateway now serves a session handoff endpoint: GET /v1/auth/session?token=<bearer>&next=<path> exchanges a valid bearer for an HttpOnly session cookie and redirects to the page you asked for. Every URL the CLI prints goes through it, so monitor, UI, and GUI links open signed in. The launch commands also stop assuming loopback, so the Monitor works over Tailscale, a LAN, or an SSH tunnel:
The Smithers Monitor following a live release-canary run from running to completed

The Monitor following a live run end to end: the run appears with a running status pill, the health banner reads Healthy while the execution tree progresses and events stream in, and it ends on Completed with 4/4 tasks done.

  • A non-loopback bind mints a login token automatically, and a wildcard bind prints one dialable URL per interface, Tailscale first, then LAN.
  • The session cookie is SameSite=Lax, scoped to the gateway’s own port, and the next redirect is constrained to a same-origin path, so the cookie stays yours: a cross-origin page cannot reuse it, and two workspace gateways on one host keep separate sessions.
  • Review printed URLs before sharing them: they carry a login token.
  • When no gateway is reachable, the CLI error names the runtime state file it checked, the port it probed, and the exact smithers gateway --host/--port command to start one.

​
oneshot and authoring right-size the work first

smithers oneshot now routes a goal by its shape. A trivial edit is applied directly, a well-scoped goal runs as a single strong agent, and only genuinely multi-stage work expands into a full workflow, so a one-line change no longer turns into a multi-node run.
  • create-workflow’s clarify stage does the same before it builds anything: it routes a trivial edit directly, points a well-scoped goal at smithers oneshot with the exact command, and scaffolds a workflow only for work that needs ordered stages.
  • oneshot warns on a dirty working copy and has the agent triage the tree before starting the goal, so pre-existing edits are not swept into its commits.
  • smithers monitor <runId> is now steerable: a built-in oneshot opens its dedicated transcript, steer, and restart surface, while other runs open the all-runs Monitor.

​
Honest review and eval verdicts

Review verdicts distinguish blocked (required verification could not run for an environmental reason, such as a missing service) from approved: false (a real rejection), so infrastructure trouble no longer reads as rejected work. Eval runs get the same treatment. A case that fails on a known harness signature is graded INCONCLUSIVE instead of failed, and bunx smthrs eval exits 5 (instead of 1) when every red case is inconclusive, so CI can tell “the work regressed” from “the harness fell over.”

​
Resume a run whose workflow you edited

retry-task and up --resume accept --accept-workflow-change, so a run parked on RESUME_METADATA_MISMATCH after an in-flight workflow edit can continue instead of forcing a fork:
The mismatch error names the flag, so the fix is in the message.

​
Local services stay reachable in the network sandbox

allowNetwork: false used to cut off everything, including the local dev server or Postgres your task was testing against. It now means no egress: loopback traffic, loopback binds, and unix sockets stay allowed while remote endpoints stay denied, and a denial names the --allow-network escape hatch.

​
Approvals and the run UI

  • Approval decisions bind to durable proofs, so a granted or denied decision is recorded against the exact attempt it answered and survives replay.
  • A review panel’s seats are configurable, so you choose which models sit on a panel instead of taking a fixed roster.
  • The gateway approval panel gains accessibility fixes, loading skeletons, and richer decision feedback.

​
New workflows in the pack

Each ships with its own live UI:
  • pr-polish-panel and review-since-publish for reviewing and polishing work before it lands.
  • whole-foods-meal-planner as an MCP-tool example and smithers-repo-federation as a federation example.
  • api-ab-benchmark, which compares the Effect and JSX authoring surfaces on the same task.
  • Merge-train workflows (sol-issue-train, xcombo-fix-train, stacked-ship) that route issues through green-gated pushes.
  • First-class monitor workflows: drop a .smithers/monitor/<workflowId>.tsx and the CLI launches it as a sibling run that watches, and can heal, each run of that workflow. The background monitor also breaks silence on retry churn and otherwise-quiet runs.
Touring a completed release-canary run in the Smithers Monitor, including the debug XML view and frames scrubber

Touring a finished run in the Monitor: the execution tree with per-node ok pills and durations, the Timeline and Usage panels with rate-limit gauges, the Debug XML view of the workflow, the Frames scrubber, and the Notable/All/Activity event tabs.

​
Reliability fixes

  • Resume honors ownership precedence and validates its metadata before reactivating a run, and registered accounts survive a resume.
  • The run UI subscribes to live approval, run-event, and run-list feeds instead of stale snapshots, surfaces run-list transport failures, and decodes persisted UltraGrill and durable event frames correctly.
  • Monitor health no longer goes stale, and a run’s durable sidecar identity is preserved across reads.
  • bunx smthrs usage reports Kimi Code subscription quota alongside the other providers.
  • The VCS layer rejects mutable jj pointers so a revert or time-travel lands on a stable revision.
  • Studio toggles crons through idempotent upserts and hides cached runs after an unauthorized response.

​
Under the hood

  • The runtime moved to Effect 4.0.0-beta.102 across the engine, scheduler, server, db, vcs, time-travel, sandbox, memory, and observability packages. Workflow and engine behavior is unchanged; the migration is a dependency and internals update.
  • packages/ui-core extracts the runs domain (list filtering and sorting, tree building, event frames, diff parsing, health, view models) into one shared home, and packages/tui-ui holds the pure TUI leaves (run tree, event log, status glyphs) built on it.
  • A new coverWorkflow testing helper drives component-level coverage with schema mocks, and a new zmux-based real-PTY e2e harness drives the TUI runs list and the Monitor mirror end to end.
  • Internal scripts and plugins now execute the working-tree Smithers, enforced by a repo check, so a checkout never silently runs a published build.

​
Upgrade notes

  • Passing --host (or SMITHERS_GATEWAY_HOST) to monitor, ui, or gui mints a login token automatically and prints one URL per dialable interface instead of the old loopback-only warning. Review printed URLs before sharing them.
  • allowNetwork: false now permits loopback and unix-socket traffic. If you relied on it to isolate a task from local services, that isolation no longer holds; remote egress stays denied.
  • bunx smthrs eval exits 5 when every red case is inconclusive. CI scripts that treat any nonzero exit as regression should special-case 5.
  • Runs stuck on RESUME_METADATA_MISMATCH after an in-flight workflow edit can resume with --accept-workflow-change instead of forking.

​
The full changelog

This page is the tour, not the inventory. The complete commit-level list is in CHANGELOG.md on GitHub. Found a bug? bunx smthrs bug files it with your run context attached.