0.27.0

Smithers 0.27.0 is one of the largest releases yet. It rebuilds the control plane around a single per-workspace Gateway daemon (autostarted by the CLI, idle-aware, hardened against DNS rebinding, and tested under 50 concurrent runs with no mocks) and replaces the bespoke gateway-WS sync stack with a TanStack DB collection engine that runs over REST + SSE locally and ElectricSQL for multiplayer. Around that core: a workflow-UI toolkit with a create-workflow builder that turns plain-English descriptions into runnable workflows, a full-screen TUI monitor behind up --interactive, Claude Code and Codex plugins that mirror every run into the chat, self-healing runs with an automatic post-failure autopsy, docs-driven development, serverless execution on Cloudflare, Telegram approvals, an interactive init wizard, smithers update, HTML run reports, and a batch of security and correctness fixes.

---

​

Breaking changes

• The bespoke gateway-WS collection stack is retired. gateway-client / gateway-react now sit on TanStack DB collections via createSmithersCollections with per-mode providers. The documented workflow-UI hook surface is preserved on top of collections, so workflow UIs keep working unchanged; direct consumers of the old sync primitives need to migrate.
• AI SDK 7 upgrade. onStepFinish is renamed to onStepEnd, and Smithers now requires Node >= 22.

See the migration notes at the bottom.

​

One Gateway daemon per workspace

The Gateway is now a singleton: CLI commands autostart one shared daemon per workspace (with identity checks so you never attach to the wrong one), and every run in that workspace flows through it over a single shared SQLite store.

• Manage it explicitly with bunx smithers-orchestrator gateway status and bunx smithers-orchestrator gateway stop.
• Idle spin-down: autostarted daemons exit after ~5 minutes with no WS clients, in-flight runs, crons, or timers. SMITHERS_GATEWAY_IDLE_MS tunes the window; an explicit bunx smithers-orchestrator gateway never idle-exits.
• Escape hatch for CI, sandboxes, and containers: set SMITHERS_NO_DAEMON=1 (or pass --no-daemon) to opt out of autostart. Commands that genuinely need a daemon fail loud and name the hatch instead of silently spawning one.
• DNS-rebinding defense: unauthenticated requests with a non-loopback Host are rejected (including correct handling of a bare ::1), and the local UI gateway proxy blocks cross-origin requests. Remote binds that legitimately serve non-loopback Hosts use auth as the gate, or the explicit SMITHERS_GATEWAY_TRUST_ANY_HOST=1 opt-out.
• Tested under load: a new no-mocks test runs 50 concurrent workflows through one gateway owner: all 50 finish, every run is attributed exactly once, and per-run event sequences stay contiguous.
• Run lists filter by workflow across the gateway API and UI surfaces.

​

TanStack DB sync engine

The UI data layer moved to TanStack DB collections with per-mode providers behind one API. The sync guide covers the architecture.

• Local mode: a QueryCollection provider over new /v1/api REST routes with SSE invalidation.
• Multiplayer mode: an ElectricSQL provider over Postgres, through a hardened Electric sync proxy.
• The hooks, collections, and mutation API are identical in both modes, and a 3-lens adversarial audit of the sync engine resolved 10 findings before release.

​

Workflow UI toolkit

Workflow UIs are now a first-class authoring surface:

• A reusable gateway-ui component package and a workflow UI styleguide provide the primitives custom workflow UIs are built from.
• A create-workflow builder UI with a code editor and live graph preview backs bunx smithers-orchestrator make-workflow: describe a workflow in plain English and watch it get designed, scaffolded, and graph-verified.
• Experimental: bunx smithers-orchestrator ui --app serves a local UI app preview with a concierge chat that can create and launch workflows (ui.smithers.sh). It requires an LLM API key and is not yet part of the supported product surface.

​

Full-screen TUI monitor

bunx smithers-orchestrator up --interactive now opens a full-screen OpenTUI + React monitor with five modes (tree, graph, logs, timeline, and hijack), fed by the same live gateway-react hooks as the web UI, with inline approvals. For watching an already-detached run, smithers monitor is the cockpit equivalent.

​

Claude Code and Codex plugins: durable runs, inside the chat

Both plugins are covered in depth at plugins.smithers.sh.

• Claude Code (integration guide): the plugin now mirrors every Smithers run into a live /workflows view by default, with a background monitor that surfaces approvals, human requests, failures, and stalls. It is driven by new versioned protocol commands (bunx smithers-orchestrator claude tick|node-wait|monitor), and the graph layer derives Claude-native phase plans, including from persisted frames, so a restarted session picks the view back up.
• Codex gets its own plugin whose skill drives the agent to run its work through Smithers and open a live bunx smithers-orchestrator ui view per workflow.

​

Self-healing runs

When a run fails, Smithers now investigates itself:

• Post-failure autopsy: the CLI auto-launches an autopsy workflow on run failure that diagnoses what went wrong from the event history and reports findings (guide).
• Fleet medic: the monitor-smithers sweep gains an auto-remediation medic stage that proposes and applies fixes for stuck or degraded runs.
• bunx smithers-orchestrator supervise continues to watch for stale running runs and auto-resume them.

​

Serverless and sandboxes

• Cloudflare: serverless resume and cron ticks with run-lease claims let a Worker own a durable run without a long-lived process, alongside a Cloudflare sandbox provider (integration guide).
• Run on plue: a plue sandbox provider and a run-on-plue workflow run any Smithers workflow on a real remote Freestyle VM, verified end to end (example).

​

Integrations: Telegram, GitHub, and Linear

• A new integrations package delivers external webhook events (GitHub, Linear, Telegram) into workflows, with durable persisted integration event state.
• Telegram becomes an approval surface: a Bot API helper package, inline approval buttons, a mini-app approval page with initData verification, and a daily-digest summary bot; see telegram.smithers.sh and the Telegram guide.

​

Docs-driven development

A new docs-driven development family of workflows treats your docs as the contract and drives agents to meet it: ddd-bug-scan finds where the code disagrees with the docs, ddd-generate-docs writes the missing contract, and ddd-improve closes the gap. The family also includes DDD quality-loop and test-coverage workflows, a spec pack, a dedicated workflow UI, and a feature-triage library, all covered by real-run e2e tests.

​

CLI

• Interactive init wizard (init.smithers.sh) with à-la-carte workflow/skill/agent-doc selection and persisted skill deselections; --yes / --non-interactive for CI. Non-interactive re-init now runs as a durable, replayable system workflow, falling back to the imperative path if the workflow cannot load.
• bunx smithers-orchestrator make-workflow "<task>" dispatches the create-workflow builder from a plain-English description, and init [prompt] chains straight into it: install the pack and describe your first workflow in one command. A seeded make-workflow tutorial teaches the authoring loop.
• bunx smithers-orchestrator update detects how Smithers was installed (bunx/npx, global bun/npm/pnpm/yarn, project-local) and upgrades it or prints the exact command, plus a passive once-a-day update notice (SMITHERS_NO_UPDATE_CHECK opts out).
• HTML run reports: bunx smithers-orchestrator workflow run --report emits an HTML report with open-in-browser.
• bunx smithers-orchestrator bug files bug reports with scrubbed run context (how we handle them).
• Agent availability probing: the CLI now probes whether agent CLIs can actually run instead of trusting auth files on disk.
• Version-pinned LLM docs: bunx smithers-orchestrator docs / docs-full serve llms.txt / llms-full.txt pinned to the installed version, with a git-tag fallback for older --docs-version requests.
• Project-local CLI resolution: a globally-installed smithers resolves and delegates to the nearest project-local install, the way tsx and bunx do.
• waiting-quota is handled consistently across the CLI, TUI, gateway, and UI: paused exit code, waiting badge/tone, and active-run treatment instead of reading as failed or terminal.

​

Agents

• OpenClaw is now a supported agent (openclaw.smithers.sh): a new adapter, a native OpenClaw plugin with an orchestrate skill, and working session resume (--session-id emitted, resume metadata persisted).
• elizaOS ships as the opt-in @smithers-orchestrator/agent-eliza package, wrapping the elizaOS AgentRuntime with an elizaOS-conventions authoring/loading layer; @elizaos/core stays out of default installs.
• A review agent tier joins the generated agents.ts, so plan/implement/review pools are first-class in fresh installs.
• SSRF fixes in agent tools: the http tool no longer attaches configured secret headers to hosts outside the allowlist, and the transcription tool rejects loopback/private/link-local audio URLs unless explicitly opted in.
• MCP stdio deadlock fixed: the transport now drains the server’s stderr pipe, so a verbose MCP server can no longer stall the session once the OS pipe buffer filled.

​

Workflow pack

• Every user-facing seeded workflow now ships a custom UI: the init pack seeds per-workflow browser UIs built on the new gateway-ui package, opened with bunx smithers-orchestrator ui RUN_ID.
• Fresh installs default the implement pool to Claude Sonnet 5 in the generated agents.ts.
• Re-init detects drifted seeded pack files and offers an interactive update, so upgrading the pack no longer means diffing by hand.
• Plan and review now run as model-diverse panels whose moderator synthesizes a single verdict, with resilient failover when one CLI cannot authenticate. The Panel component’s moderator is gated on actual panelist outputs instead of synthesizing a verdict blind.
• System workflows: internal durable processes (like re-init) are marked system: true and hidden from default listings while staying explicitly runnable, so frontends show only your workflows.
• Cron and event triggers turn a push into a pipeline: see automate.smithers.sh.

​

Code review service

The open-code-review service that backs /code-review grew verification teeth: finding-verification helpers, impact quizzes with verified walkthroughs, sticky PR status comments, cache-token-aware usage accounting, and a monthly spend cap with claim-slot quotas.

​

Engine and scheduler fixes

• Agent final-output JSON extraction (extractLastBalancedJson) is now a single O(n) pass; a 200KB prose payload previously blocked the event loop about 1.7 seconds.
• Timer durations like 1d now park instead of busy-looping: the scheduler’s duration parser matches the engine grammar (d unit, case-insensitive), and duration-timer deadlines stay anchored across decide passes.
• Subflows with arbitrarily-named outputs work as documented via an opt-in smithers(build, { output }) resolution.
• The jj repo probe is bounded with a 1.5s timeout so a hung jj no longer blocks durability startup; skipped workspace snapshots and failed durable-resume restores log structured warnings instead of silent nulls.
• The agent trace buffer and DevTools run store are now bounded (head+tail retention, configurable run/event caps).
• Task gains depsOptional for tolerating failed upstream dependencies, and human task kinds survive the snapshot pipeline.

​

Security hardening

Beyond the gateway host defense and agent-tool SSRF fixes: canonical 3-segment JWTs with timing-safe signature comparison, OIDC PR-claim verification in the review server, a hardened Electric sync proxy, run idempotency enforcement, and a structured TOOL_NETWORK_ISOLATION_UNENFORCED warning when allowNetwork:false cannot be enforced at the OS level (Linux).

​

Windows CI

The full typecheck and test suites now run on Windows in CI (sharded), with LF line-ending normalization via .gitattributes and numerous Windows-specific test and portability fixes.

​

Migration notes

• Consumers importing the old gateway-WS sync primitives (createGatewayCollection, SyncProvider-era APIs) should move to createSmithersCollections / SmithersCollectionsProvider; workflow UIs using the documented gateway-react hooks keep working unchanged.
• Ensure Node >= 22 before upgrading; SDK-agent callers using onStepFinish must rename to onStepEnd.
• CI, containers, and sandboxes that must not spawn a background gateway should set SMITHERS_NO_DAEMON=1 (or pass --no-daemon); commands that genuinely need a daemon then fail loud instead of autostarting.
• Remote gateway binds that legitimately serve non-loopback Hosts need auth (the token is the gate) or the explicit SMITHERS_GATEWAY_TRUST_ANY_HOST=1 opt-out.
• http-tool configurations that rely on defaultHeaders reaching hosts other than the baseUrl host must now list those hosts in allowedHosts; transcription of private/loopback audio URLs requires allowedAudioHosts or allowPrivateAudioUrl.