Smithers agent runtime
SkillAI & modelsLets your agent configure the Smithers agent loop, model seats, budgets, sandbox limits, and flow-call authorization.
Available today. Use it from your connected AI after setup.
No other account needed.
Add ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.
Then ask your AI: use the Smithers agent runtime skill
About this skill
Compose or change the Smithers agent loop, host services, model seats, budgets, sandbox limits, or flow-call authorization.
What this skill tells your AI
The instructions your AI receives, as published by smithersai/smithers in .agents/skills/smithers-agent-runtime/SKILL.md and read by ahel’s review.
Read the agent composition guide and relevant package tests before changing host wiring or agent boundaries. Agent.Service owns one cell loop. Cells reach capabilities through ctx.call to ordinary flows; AgentSession and AgentAction adapt that same loop. A foreign CLI agent implements Agent.Service rather than adding a second loop.
The host resolves credentialed seats through SeatResolver, supplies quota and budget policy, and declares bounded sandbox limits. The completion evaluator is a host decision: select it before opening resources and refuse missing configuration rather than silently using a scripted or unavailable judge in production.
Budget accounting is durable. An unreadable or unwritable ledger is unknown spend and fails closed. Budget.check is advisory; a custom provider boundary needs scoped reserve and record around the real call. Check authorization before opening a journaled activity so a later grant can unblock a parked action. Keep the flow's declared capabilities and effects consistent with the bound implementation; verify recovery and refusal behavior in tests, not only a fresh run.
For model changes, read the model guide: canonical requests exclude credentials, which resolve just before transport. Ambient provider keys do not enable live tests; SMITHERS_LIVE_MODEL_TESTS=1 opts into quota-consuming calls. For command tools, read the Std guide: Bash hermetic is a lexical check, not OS confinement, and truncated output must stay visibly truncated. For scheduled or webhook launches, read the triggers guide: retryable claims require a durable idempotencyKey dedupe at RunnerService.start. For scoring receipts, read the scorers guide: a completed batch can contain inconclusive observations, and store failures are warnings; verify persistence by readback before claiming a grade was saved.
For host capability composition, read the kernel guide. Replace the ordinary host service tags with their guarded implementations in place; a parallel protected service lets dependencies bypass the guard. Path is intentionally unchecked, so a path service alone is no confinement boundary.
For untrusted execution, read the sandbox provider guide and choose a provider whose documented boundary meets the threat model. The package itself adds no isolation; DirectorySandbox is a scratch directory with process lifecycle, not filesystem, user, or network confinement. Verify provider cancellation with its conformance suite.
Signals
- GitHub stars
- 422
- Forks
- 50
- Last commit
- Sep 2026
Advanced
- Catalog kind
- skill
- Key
smithers-agent-runtime- Source
- github.com/smithersai/smithers