> ## Documentation Index > Fetch the complete documentation index at: https://smithers.sh/llms.txt > Use this file to discover all available pages before exploring further. # 0.32.0 > Smithers 0.32.0 adds monitor workflows that watch and heal runs, task-shaped oneshot routing, release review automation, safer workflow recovery, and a broad child-run durability hardening pass. **Smithers 0.32.0 is a durability and operations release.** Most changes target the places durable runs break in production: resumed runs, child workflows, sandboxed networks, monitors, and time travel. It also adds task-shaped oneshot routing, a release-review workflow, monitor workflows that can heal a run, and a resume path that accepts an edited workflow instead of refusing to continue. ## Upgrading ```bash theme={"theme":{"light":"github-light","dark":"github-dark"}} bunx smithers-orchestrator update # upgrade the CLI bunx smithers-orchestrator upgrade # agent-assisted: reads the changelogs and applies what your project needs bunx smithers-orchestrator packs update # refresh installed workflow packs ``` Behavior changes are collected in the [upgrade notes](#upgrade-notes). ## Task-shaped oneshots and release review `smithers oneshot` now chooses its first available model from the goal shape. UI and frontend goals lead with Kimi K3, while general implementation goals lead with Claude Opus 5. Explicit `--agent` and `--model` choices still win, and `smithers oneshot --status` reports the resolved task type and chain. The built-in pack also gains `review-since-publish`, a three-model review panel that compares the working tree with the last published tag, merges verified findings, partitions fixes into file-disjoint lanes, and repeats until clean. ## Monitor workflows: a workflow that watches your workflow Long runs fail in boring ways: a node wedges, an agent stalls, and nobody is watching at 3am. Now you can ship a monitor alongside any workflow. Drop a file at `.smithers/monitor/.tsx` and Smithers runs it against every run of that workflow: it observes health, applies the healing actions you define for stalled or wedged-node conditions, and escalates to a durable `HumanTask` when it cannot fix things itself. * Healing actions are ordinary workflow code, so a monitor can retry a node, resume a parked run, or gather diagnostics before waking you. * Escalations are durable human requests, so the question survives restarts and shows up in `bunx smithers-orchestrator human` and the Monitor UI. * This is a first release of the surface; expect the condition vocabulary to grow in coming versions. ## Resume a run whose workflow you edited Editing a workflow file while one of its runs was parked used to be a dead end: resuming failed with `RESUME_METADATA_MISMATCH` and your only option was forking a fresh run. Now you decide: ```bash theme={"theme":{"light":"github-light","dark":"github-dark"}} bunx smithers-orchestrator retry-task RUN_ID TASK_ID --accept-workflow-change bunx smithers-orchestrator up --resume RUN_ID --accept-workflow-change ``` * The run carries the edited workflow forward from the resume point. * The mismatch error now names the flag, so the fix is in the message. ## The CLI watcher speaks up sooner `smithers claude monitor` used to alert only on discrete bad transitions, so a run that was quietly retrying the same node forever, or emitting nothing at all, never broke the silence. It now fires on retry churn (`node-retrying`) and on long silence from a live run (`run-progress`), so a wedged detached run pings you instead of wasting an afternoon. Oneshots got the same attention: `smithers monitor ` now serves a dedicated live UI for built-in oneshots where you can watch the transcript and steer the agent mid-run. The oneshot and hijack surfaces now share one `OneshotSurface` component, so steering works the same everywhere. ## Right-sized authoring and honest review verdicts Two authoring-loop fixes reduce wasted runs: * `create-workflow` now right-sizes the request before building anything. A trivial edit is routed directly, a well-scoped goal goes to `smithers oneshot`, and only genuinely multi-stage work gets a full workflow scaffold. * Review verdicts now distinguish `blocked` (the verification could not run for an environmental reason, such as a missing service) from `approved: false` (a real rejection). Harness faults no longer read as rejected work. `smithers oneshot` also warns when the working copy is dirty and has the agent triage the leftover changes before starting on your goal, instead of silently building on top of them. ## Durability hardening This is where much of the release's hardening work went. * **Async `deps` callbacks complete when they finish, not before.** A `` async function child could be claimed complete while its work, including a launched child run, was still pending, so downstream tasks read an empty payload. The engine now awaits the callback before claiming completion (#1415). * **Child runs behave like part of the run.** A large batch of subflow fixes landed: cancel and pause now propagate to child runs, approvals keep the right owner, signals are preserved across child waits (#1405), snapshot lineage stays isolated, failed child branches all reset on retry (#1410), subflow rewind leases are hardened (#1411), and child runs keep gateway attribution (#1389) and appear in run lists. * **Time travel restores real bytes.** jj attempt pointers recorded the change id, and since jj resolves a change id to its current commit, `revert` and `timetravel` could resolve to the very state you were trying to leave: a silent filesystem no-op. Pointers now pin the immutable commit id. * **The network sandbox keeps loopback open.** `allowNetwork: false` was cutting off local dev servers, local Postgres, and unix sockets along with the internet. Loopback now stays reachable while remote egress is still denied, and denial errors name the `--allow-network` escape hatch. * **Evals stop blaming your workflow for harness deaths.** A red case whose failure matches a harness signature (network, DNS, OOM, rate limit) is now graded INCONCLUSIVE instead of failed, and `bunx smithers-orchestrator eval` exits with code 5 when every red case is inconclusive, so CI can retry instead of reverting good work. * **Subflow timers wake durably.** Parent subflows now persist their child timer deadline, close stale waiting attempts, and resume through the gateway timer sweep instead of parking forever. * **Event waits reject stale matches.** A new wait cannot consume an older matching signal that predates the wait node. ## Other improvements * Gateway RPC, REST, SDK, and MCP surfaces gained typed queries for a run's child-run descendants and direct children, with regenerated contracts and docs (#1391). * New `packages/ui-core` and `packages/tui-ui` packages extract the shared runs domain (run list, tree, event frames, health, and view models) used by both the TUI and the web UI. * The approval panel gained accessibility improvements, loading skeletons, and richer decision feedback. * A new `eval-driven-development` skill documents red-classify-first, green-ratchet, and circuit-breaker discipline for automated repair loops. * Dependency ranges widened for effect, zod, and pglite, and mermaid is now an optional peer dependency. * New docs: capability pages, a monitor workflows guide, and a compatibility policy page. * Whole Foods ordering webhooks must match the server-configured HTTPS endpoint and resolve only to public addresses, closing input-driven SSRF and redirect paths. * The TUI run search now captures printable keys without triggering global shortcuts, and unknown backend statuses stay out of active-run counts. ## Upgrade notes * If a resumed run previously failed with `RESUME_METADATA_MISMATCH` after an in-flight workflow edit, re-run with `retry-task --accept-workflow-change` (or `up --resume --accept-workflow-change`) instead of forking a fresh run. * Workflows using `` with async function children should confirm downstream consumers were not silently reading an empty `{}` payload before this release; the compute bridge now awaits the callback before claiming completion. * Sandboxed tasks with `allowNetwork: false` can now reach loopback services. If you relied on loopback being blocked, that is no longer the case. * `bunx smithers-orchestrator eval` can now exit with code 5 (all red cases inconclusive). CI scripts that treat any nonzero exit as a regression should handle it as a retry signal. ## The full changelog This page is the tour, not the inventory. The complete commit-level list is in [CHANGELOG.md on GitHub](https://github.com/smithersai/smithers/blob/main/CHANGELOG.md). Found a bug? `bunx smithers-orchestrator bug` files it with run context attached.