0.29.0

​

Upgrading

bunx smthrs update        # upgrade the CLI
bunx smthrs upgrade       # agent-assisted: reads the changelogs and applies what your project needs
bunx smthrs packs update  # refresh installed workflow packs

​

First-class memory

<Memory bank={`project-${ctx.input.repoId}`} tags={["branch:main"]} recall="auto" primers={["project-primer"]} retain="on-complete" tools>
  <Task id="implement" agent={agents.coder} output={outputs.impl}>…</Task>
</Memory>

• bank/banks, tags, recall ("auto", a fixed query, or false), budget, maxTokens, primers, retain, and tools: inherited by descendant tasks through context, overridable per task via the existing memory prop.
• Recall is scoped per bank, so distinct banks keep distinct recall scopes instead of bleeding into each other.
• Every injected block and remembered tag is capped by an explicit bound, so recall never blows past your token budget.
• Set HINDSIGHT_URL (plus optional HINDSIGHT_API_KEY) to activate the Hindsight-backed store. Without it, the existing local SQLite facts store and behavior are unchanged.

<Memory bank="repo-knowledge" recall="auto" retain="on-complete" tools>
  {/* inherits the bank, recall, retain, and tools from the provider */}
  <Task id="triage" agent={agents.triager} output={outputs.triage}>
    Classify the failure.
  </Task>

  {/* narrows recall to a fixed query and skips retaining a digest */}
  <Task
    id="repro"
    agent={agents.coder}
    output={outputs.repro}
    memory={{ recall: "known flaky tests in this suite", retain: "off" }}
  >
    Reproduce it.
  </Task>

  {/* opts out: no recall, no tools, no digest */}
  <Task
    id="format"
    agent={agents.fast}
    output={outputs.format}
    memory={{ recall: false, retain: "off", tools: false }}
  >
    Format the report.
  </Task>
</Memory>

export HINDSIGHT_URL="https://hindsight.example.com"
export HINDSIGHT_API_KEY="…"   # optional

​

XState as durable derived state

const state = useSmithersMachine(releaseMachine, {
  id: "release",
  input: ctx.input,
  events: [
    taskOutput(outputs.research, { nodeId: "research" }, () => ({ type: "RESEARCH_DONE" })),
    eventReceived("REVISE", reviseSchema, {}, (payload) => ({ type: "REVISE", payload })),
  ],
});

return state.matches("drafting")
  ? <Task id={`draft-r${state.context.revision}`} output={outputs.draft}>…</Task>
  : null;

• Event sources read typed durable rows and order them on the shared provenance sequence, with deterministic declaration-order tie-breaking.
• Signal sources support correlationId, so concurrent listeners using the same signal name do not consume one another’s history.
• A mount-time lint rejects actor-only and side-effecting machine features such as invoke, spawn, after, and custom actions, and detects constant task identities in re-enterable states before they can deadlock a run.
• A reducer-content hash and run-scoped, bounded prefix cache make incremental folds fast without serving a snapshot produced by another run, fork, or machine definition.
• xstate is an optional peer. Projects that never import the integration do not need it installed, and the main facade remains isolated from it.

​

A shared UI component library

• A dependency-free Markdown renderer that renders through React children, never innerHTML, so it is XSS-safe by construction. Links with an unsafe scheme render without an href and never reach your onLinkClick handler.
• A dependency-free DiffHunks component and diff domain, for when you do not need the full diff adapter.
• FileTree, CollapsiblePanel, and StageStrip components.

• Terminal over xterm.js, from smthrs/ui/adapters/terminal.
• MarkdownEditor over Milkdown Crepe, from smthrs/ui/adapters/markdown-editor.
• PierreDiffView over @pierre/diffs, from smthrs/ui/adapters/pierre-diff-view.

// dependency-free, safe from the barrel
import { Markdown, DiffHunks, FileTree, CollapsiblePanel, StageStrip } from "smthrs/ui";
// run-aware gateway components
import { WorkflowGraph, NodeOutputCard } from "smthrs/gateway-ui";
// heavy adapters: one subpath each, never the barrel
import { Terminal } from "smthrs/ui/adapters/terminal";
import { MarkdownEditor } from "smthrs/ui/adapters/markdown-editor";
import { PierreDiffView } from "smthrs/ui/adapters/pierre-diff-view";

​

Microsandbox: a first-class sandbox provider

import { Sandbox } from "smthrs";
import { createMicrosandboxSandboxProvider } from "smthrs/microsandbox";

const provider = createMicrosandboxSandboxProvider({
  image: "oven/bun:1",
  cpus: 2,
  memoryMib: 4096,
  security: "restricted",
});

export default parent.smithers((ctx) => (
  <parent.Workflow name="microsandbox-run">
    <Sandbox
      id="isolated-edit"
      provider={provider}
      workflow={childWorkflow}
      input={{ prompt: ctx.input.prompt }}
      allowNetwork
      output={parent.outputs.result}
    />
  </parent.Workflow>
));

​

Durable signal and output rows

// every durable output row for a node, in provenance order
const reviews = ctx.outputRows(outputs.review, { nodeId: "review" });
const rejected = reviews.filter((row) => !row.payload.approved);

// every delivered signal, whether or not a listener was mounted at the time
const revisions = ctx.signalRows("REVISE");
const latest = revisions.at(-1);

​

Security and durability hardening

• An empty bind={[]} provenance binding is parked as missing instead of dispatching an unproven task.
• Real-process test admission rejects impostor runner scripts instead of trusting an executable path alone.

• Concurrent resume paths (timer, signal, approval, webhook, operator) are deduplicated through a runId-keyed in-flight gate, and run-tracking cleanup no longer lets an older settling invocation wipe a newer invocation’s records.
• Cron fires are now claimed via compare-and-set before launching, so a scheduler cannot double-fire the same scheduled run.
• The driver strips the persisted runId column from table-shaped ctx.input, so it no longer trips an unrecognized-key error in a strict Zod input schema.
• The standalone Telegram client clamps retry_after to a documented cap (default 30s), races the retry wait against the call’s AbortSignal, and no longer treats an aborted body read as a retryable error that could trigger a duplicate request.
• Claude subscription accounts with organization-level overage purchasing disabled no longer have every agent invocation misclassified as a quota rejection.

​

Other improvements

• The driver gained an onError hook so external error reporters, Sentry for example, can capture run failures. It fires once per node or run failure occurrence, and the report names which one:
  await run(workflow, {
    input,
    onError: (report) => {
      Sentry.captureException(report.rawError, {
        tags: { runId: report.runId, phase: report.phase },
        extra: report.phase === "node"
          ? { nodeId: report.nodeId, iteration: report.iteration, attempt: report.attempt }
          : {},
      });
    },
  });
  
• Eval suites support LLM-judge assertions, for cases where correctness is about meaning rather than an exact JSON match:
  {"id":"deadline-summary","input":{"prompt":"Summarize the launch plan"},"expected":{"status":"finished"},"judge":{"instructions":"Output must be a polite summary mentioning the deadline","threshold":0.7}}
  
  judge.threshold defaults to 0.8. A case with both expected and judge passes only when every deterministic assertion passes and the judge score meets its threshold.
• The observability package supports OTLP authentication headers for authenticated trace export.
• Detached (up -d) run logs now default to <workspace>/.smithers/logs/<runId>.log with automatic retention-window and total-size-capped cleanup, instead of accumulating indefinitely in the workflow source directory. One workspace had 711 immortal log files totaling 1.4 GB before this fix.
• bunx smthrs ui no longer silently opens a different workflow’s UI when a long-lived gateway’s registry predates a newly authored workflow: the gateway now rescans on a registry miss, and the CLI fails loud with an actionable error instead of guessing.
• The CLI and gateway now tolerate leftover jj merge-conflict markers in the workspace package.json instead of crashing on JSON parse; the gateway boots read-only in that state.
• The db package now probes output tables via sqlite_master on sqlite-dialect external backends, fixing a table-detection bug.
• A new Workflow testing framework guide documents the expanded @smthrs/testing suites: cutpoint-conformance, replay-identity, and artifact-parity.
• A new example workflow, “Daily CEO Intel” (branded in its own docs as “The Smithers Signal”), demonstrates a durable daily-intelligence-brief pipeline with multi-provider SDK-agent fallback (Anthropic to OpenAI to Gemini) and Cloudflare KV and R2 publishing. It ships marked as an example in the workflow pack, not a supported product feature.

​

Upgrade notes

• npm upgrades directly from 0.28.0. Upgrading to 0.29.0 picks up the complete release described above, headlined by <Memory>.
• Memory bypasses the output cache when active. Tasks wrapped in an active bank-based <Memory> config bypass the durable output cache, because recalled context is a mutable input that is not part of the cache identity. Legacy namespace or object-form memory metadata remains inert and keeps its previous cache semantics.
• HindsightMemoryStore assumes a single writer per contract store. This constraint and its invariant tests are documented in the memory guides.
• Wait-for-event tagged results are an explicit opt-in. Existing waits keep receiving the raw signal payload unchanged. A timeout continuation now requires the tagged envelope instead of writing an ambiguous null.
• Detached run logs moved. up -d logs now default to <workspace>/.smithers/logs/<runId>.log instead of the workflow source directory. An explicit --log-dir still overrides, and SMITHERS_LOG_RETENTION_DAYS / SMITHERS_LOG_MAX_TOTAL_BYTES tune the automatic cleanup.
• Tool response bodies are capped by default. Generated OpenAPI tools and the built-in HTTP tool now stop reading a response body at 1 MiB. The transcription tool’s remote audio download keeps its 25 MiB default. Raise or lower either with maxResponseBodyBytes. A call that previously streamed a larger body now fails with an explicit limit error instead of buffering without bound.
• The Freestyle sandbox provider was removed. It shipped as an example provider and Microsandbox replaces it. Switch to Microsandbox or one of the other first-class providers (Vercel, AWS, GCP, Daytona).
• StatusClass gained a run member and statusClass() re-buckets. Running now maps to the brand class, and waiting, paused, queued and pending map to their new classes. An exhaustive switch or lookup map over StatusClass will fail to compile until you handle run.

​

The full changelog