Skip to main content
Smithers 0.26.0 is a feature release. It adds first-class support for the Hermes agent runtime, makes every built-in workflow print something useful when it finishes, and fixes a few rough edges in the CLI. This release also rolls up the output-task and run-output work that was prepared as 0.25.5 and 0.25.6 but never published to npm, so npm goes straight from 0.25.4 to 0.26.0.

​
Agent integrations

Smithers now drops into the Hermes agent runtime, in both directions: the agent operates durable Smithers workflows, and a Smithers workflow can delegate a task back to the agent.
  • Hermes (Nous Research) gets a native Smithers plugin, which is richer than a bare MCP entry. It adds slash commands (/smithers run|ps|inspect|approve|deny|watch) usable in the Hermes CLI and every gateway it speaks to, a pre_llm_call hook that injects live run status and pending approval gates into Hermes on every turn, the smithers_run / smithers_ps / smithers_inspect / smithers_approve / smithers_deny / smithers_output / smithers_human_answer tools, and a bundled smithers:orchestrate skill. A new HermesCliAgent also lets a workflow <Task> delegate to the Hermes agent itself.
See Hermes for setup.

​
Every workflow run prints a useful result

A finished run used to print output: null even when the work produced something. Now it does not.
  • The workflow run / up summary prints the run’s output cleanly. It strips the internal run_id / node_id / iteration columns, unwraps a single row, and omits the field entirely when there is nothing to show, so a successful run never prints a noisy output: null.
  • Every built-in init-pack workflow, both the seeded ones (hello, create-workflow, route-task, and the rest) and the legacy-inline ones (plan, implement, review, debug, audit, and the rest), now ends with a deterministic output task that aggregates its genuinely useful terminal result: files changed, the plan steps, the chosen route, reviewer verdicts, the greeting, and so on. Add a final output task to your own workflows to control what their run prints.

​
CLI

bunx smithers-orchestrator memory list now works with no namespace. It lists every fact across all namespaces, grouped by namespace, instead of failing with a raw validation error. Passing a namespace still scopes to that one. The memory store gains a listAllFacts() method (and its listAllFactsEffect twin) that backs it.

​
Migration notes

No breaking changes are identified in 0.26.0. The new agent integrations and output tasks are additive; workflows that already ended with their own output task are unaffected. Because 0.25.5 and 0.25.6 never reached npm, installing 0.26.0 from 0.25.4 picks up all of their changes at once.