Smithers · v0.22.0

Author and run
agent workflows.

0.22.0 ships a batch for the people writing workflows: a session-forking primitive, one-command setup for 20+ coding agents, an outcome-first starter gallery, automatic prompt tuning, and live per-workflow UIs.
$ smithers starters
$ smithers optimize
$ smithers ui
<Task fork="plan">
smithers.sh
0.22.0
New · smithers starters

Start from the outcome you want.

idea-to-prdturn a rough idea into a written PRD
ship-a-changeplan, implement, and review a code change
customer-incidenttriage an incident and draft the writeup
launch-checklistrun the pre-launch checklist end to end
$ smithers starters  # plain outcomes, copy-paste commands
smithers.sh
10 starters map onto the seeded workflows
New · smithers optimize

Tune the prompts, keep only the wins.

baseline
0.61
optimized
0.84
Runs your eval suite twice: once with current prompts, once with GEPA-generated patches. It writes the new prompt artifact only when the optimized score clears --min-improvement. Structure, schemas, retries, and approvals stay untouched.
smithers.sh
cerebras · openai · anthropic · gemini · kimi · opencode
Under the hood · Hardening

A security and correctness sweep.

Security closed

✓ DOM XSS in agent markdown links
✓ local-RCE via cross-origin PTY upgrade
✓ path-traversal in account labels
✓ request-body DoS in the workspace API
✓ auth-bypass via prototype-key tokens

Correctness hardened

✓ engine reconciler and graph extraction
✓ time-travel snapshot fork and replay
✓ DB migrations and node-diff cache
✓ token accounting and judge scorers
✓ observability dashboards and event types
smithers.sh
plus memory-leak fixes across the gateway