Skip to main content
Smithers 0.30.0 is here! Across the v0.29.0..HEAD range that is 168 commits, 759 files changed, 96,014 insertions and 17,320 deletions, including 57 feature commits, 73 fix commits (3 of them security fixes), 2 refactors, 10 test commits, and 18 docs commits. The headline is a large new agentic UI component library and the gateway-ui widgets built on top of it (live per-node chat, a fleet dashboard), alongside a hardened browser viewer, a Monitor overview redesigned around triage, the new smithers oneshot command for single-agent asks, and a model registry update that moves orchestration and gating onto Claude and adds Kimi K3.

​
Upgrading

Behavior changes are collected in the upgrade notes at the bottom.

​
An agentic UI component library

Building a UI for an agent run used to mean hand-rolling message bubbles, tool-call panels, and diff viewers for every workflow. packages/ui now ships a component library built for exactly that surface:
  • Conversation: Message, MessageBranch, Bubble, MessageScroller.
  • Input: prompt and attachment components.
  • Agent internals: reasoning traces, tool-call panels, code blocks.
  • Work tracking: plans, tasks, queues, approvals, checkpoints.
  • Grounding: sources and citations.
  • Identity: agent identity and context components.
  • Coding artifacts: stack traces, test results, commits, secrets.
  • Sandbox previews: AgentSandbox, WebPreview, JSXPreview.
  • A workflow canvas built on @xyflow/react, for rendering a run’s graph directly in a workflow UI.
The library shipped through ten parallel worktree lanes behind a CAS-serialized merge queue, with a dual-reviewed integration pass, so a custom workflow UI can now compose these pieces instead of writing markup against raw run events. This release also lands a design-contract pass on top of the library: cancelled states read as neutral instead of alarming, colors and shadows route through shared styleguide tokens, spacing follows a 2px grid, and pressed-state feedback is consistent across components.

​
Live chat and a fleet dashboard in gateway-ui

gateway-ui gains the widgets a Monitor or custom workflow UI needs to watch more than one thing at a time:
  • NodeChatStream streams an agent’s live chat for a single node.
  • FleetTable and NodeStageStrip give a fan-out run (many nodes running the same task shape) a dashboard instead of a flat event list.
  • RunMeta and a MonitorButton add a one-click deep link from a workflow UI into the run’s Monitor page.
  • RunEventLog is now structured instead of a raw JSON dump: rows are severity-tinted, and repeated heartbeats coalesce instead of scrolling the feed.

​
A hardened browser viewer

Smithers can now drive a real browser from a workflow and let you watch it happen. The gateway adds a Playwright-driven browser session service: createBrowserSession, browserAct, browserContext, browserPick, closeBrowserSession, and listBrowserSessions, backed by a live screencast in the viewer. Because a browser session can touch arbitrary pages and arbitrary keystrokes, this shipped alongside a hardening pass:
  • Session creation is transactional, so a failed setup step cannot leave a half-registered session behind.
  • The pinned browser proxy is tested against DNS rebinding.
  • Single-keypress actions aimed at sensitive fields are now redacted. Only full type actions were redacted before; typing a password one key at a time through the viewer leaked each character into the journal and activity broadcast. Both action shapes are redacted now.

​
A triage-first Monitor

The Monitor overview used to answer one question: is it done yet? It is now organized around triage:
  • A “Needs you” band surfaces pending approvals, human-parked runs, and recent failures (under 24h) first.
  • An “Active now” band lists what is currently running.
  • Operational stats (throughput, counts) move to a demoted footer instead of competing for attention at the top.
  • A Debug fold tucks frames and the raw XML tree away until you ask for them.
  • A keyboard kit makes triage fast without a mouse: / to search, j/k to move, Enter to open, Esc to back out.
Every inspector block (Details, Prompt, Tool calls, Transcript, Output, Diff) is now a collapsible section that unmounts when closed, so a collapsed Transcript stops polling in the background. The inspector also gains a Prompt panel showing the task’s initial prompt. See the Monitor guide.

​
smithers oneshot

Not every ask needs a full workflow. smithers oneshot routes a single clear task through an implement (and optional review) pipeline, with its own live-status dashboard:
  • The oneshot UI shows a live status narrator: a cheap agent watches the implement and review agents’ recorded output and posts one-line “what is it doing” updates into a pulsing status card, so you are not staring at a raw transcript to tell whether it is stuck.
  • The kimi seat now runs on Kimi K3 (1M-token context) instead of K2.7.

​
Model registry: Claude orchestrates, Kimi K3 joins

The default SOTA model registry moves orchestration and gating decisions onto Claude: Opus 4.8 is the default orchestrator at medium reasoning effort, Fable 5 holds planning, Codex Sol keeps review, and substantial implementation work moves from Luna to Terra. Kimi K3 (1M-token context) is newly registered, with usable ids for the Kimi CLI, OpenCode, and the Cloudflare AI Gateway. If you pin agents by name in your own agents.ts, this changes nothing; if you rely on the shipped defaults, your orchestration and gating calls now route to Claude.

​
Other improvements

  • Detached runs no longer report success before they exist. up -d, workflow run -d, and detached resume now preflight-render the workflow graph before spawning: a parse, import, or graph error prints file:line:col and exits non-zero with no run created. Once the graph is valid, the CLI waits up to 10s for proof the run was admitted before printing a success run id.
  • The gateway listRuns RPC gains an offset filter for server-side pagination over its newest-first run ordering. Zero is a valid offset; a negative or non-integer value is rejected as INVALID_REQUEST.
  • A new OmpAgent adapter and omp engine let Smithers drive the OMP harness, with the full RPC lifecycle handled.
  • gateway-client now classifies an unreachable or non-envelope gateway response as GATEWAY_UNAVAILABLE. Query collections keep their last-known rows and surface one notice per session instead of spamming console.error on every query.
  • A new Recharts-based chart adapter (ChartContainer, ChartTooltip, ChartLegend) ships at smithers-orchestrator/ui/adapters/chart with a CVD-validated categorical color palette, for any workflow UI that needs to plot numbers instead of just listing them.
  • The create-ui workflow’s compliance loop is now hard-gated behind a deterministic scorer (design-system usage, plus mandatory live agent chat for agent workflows), closing prior export bypasses in the structural UI verifier.

​
Hardening

73 fix commits, 3 of them security fixes, landed alongside the features above, most visibly the browser-viewer redaction fix described earlier. A new 44-case deliverable-behavior eval suite now checks whether a doc-grounded agent picks the right deliverable shape (a self-contained HTML report versus a shipped UI component versus hand-rolled markup), and a release-notes-visual eval gates the auto-generated release report on containing real charts with honest practice: no dual axis, a visible legend, shipped chart components instead of hand-rolled canvas. 257 documentation pages and the Smithers skill were rewritten for concision, with mechanical checks (frozen frontmatter, headings, and code fences) and an idea-preservation audit pass, so none of the trims should have dropped a fact.

​
Upgrade notes

There are no breaking changes in 0.30.0.
  • Default orchestration and gating now route to Claude if you use the shipped SOTA registry defaults rather than a custom agents.ts. See the model registry section above.
  • The kimi oneshot seat now runs Kimi K3, not K2.7.
  • listRuns accepts a new offset parameter. Existing callers that omit it are unaffected; a caller that already passes an invalid offset-shaped value will now get INVALID_REQUEST instead of it being ignored.
  • gateway-client no longer logs every failed RPC to console.error when the gateway is unreachable; it logs once per session and classifies the failure as GATEWAY_UNAVAILABLE instead.
  • A detached run can now fail fast before it starts. If your tooling parses up -d output expecting an immediate run id on the happy path, a bad workflow file now exits non-zero with a file:line:col error instead of silently succeeding with no run behind it.

​
The full changelog

The complete commit-level history for this release is in CHANGELOG.md on GitHub. Found a bug? Run bunx smithers-orchestrator bug to file it.