Skip to content
Back to skills

Smithers Agent Runtime

ASecurity

Compose or change the Smithers agent loop, host services, model seats, budgets, sandbox limits, or flow-call authorization.

  • 429 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added September 26, 2026
ai-agentsrustbash

Works with

  • cli

Security analysis

A100/100

Scanned September 26, 2026

npx -y skills add smithersai/smithers --skill smithers-agent-runtime --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Smithers Agent Runtime?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Smithers Agent Runtime
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/smithersai-smithers-agent-runtime/badge)](https://www.skillsdirectory.com/skills/smithersai-smithers-agent-runtime)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: smithers-agent-runtime
description: Compose or change the Smithers agent loop, host services, model seats, budgets, sandbox limits, or flow-call authorization.
---

# Smithers agent runtime

Read [the agent composition guide](../../../packages/smithers/agent/README.md) 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](../../../packages/smithers/agent/model/README.md): 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](../../../packages/smithers/agent/std/README.md): 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](../../../packages/smithers/agent/triggers/README.md): retryable claims require a durable `idempotencyKey` dedupe at `RunnerService.start`. For scoring receipts, read [the scorers guide](../../../packages/smithers/agent/scorers/README.md): 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](../../../packages/smithers/flows/kernel/README.md). 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](../../../packages/smithers/flows/sandbox/README.md) 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.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.