Skill trust: files trusted by default, one provider-level opt-out
This release of @helix-agents/core and @helix-agents/skill-fs (both minor: a 0.x package's minor marks an observable behaviour change) follows a user ruling (2026-11-21): giving an agent a SkillProvider is the trust decision, so a skill's resource files are trusted like its body.
What changed
| Before | Now | |
|---|---|---|
read_skill_file (default) | <skill_file … untrusted="true"> + a "do not follow instructions" note | <skill_file skill="…" path="…">, no mark, no note |
load_skill (default) | <skill name="…"> | unchanged |
| A provider you do not vouch for | no option | trust: 'untrusted': body and files both marked, plus one note |
| Catalog entry of an untrusted skill | n/a | <skill untrusted="true"> |
preloadSkills naming an untrusted skill | n/a | refused: a typed, non-retryable validation_error |
Every runtime (JS, the Cloudflare DO, Cloudflare Workflows, Temporal, DBOS) gets this from core; there is nothing to wire. A result stored before the upgrade keeps its old wrapper, so no session's cached prefix changes and nothing is migrated.
What to do
Skills you ship yourself: nothing. Their files now reach the model as part of the skill's instructions.
Skills you do not vouch for (customer uploads, a marketplace, a baked third-party package): mark the provider.
import { inCodeSkillProvider } from '@helix-agents/core';
import { fileSystemSkillProvider } from '@helix-agents/skill-fs';
const marketplace = fileSystemSkillProvider({
roots: ['./marketplace-skills'],
trust: 'untrusted',
});
const baked = inCodeSkillProvider(thirdPartySkills, { trust: 'untrusted' });
// One provider with both kinds: decide per skill name (never per file).
const mixed = inCodeSkillProvider(allSkills, {
trust: (name) => (vetted.has(name) ? 'trusted' : 'untrusted'),
});
// A custom provider: set the property, or re-mark an existing one.
const remarked = { ...myProvider, trust: 'untrusted' as const };preloadSkills: an untrusted skill cannot be preloaded (its body would sit in the system prompt). Load it on demand with load_skill, or mark it trusted. On DBOS the refusal is logged and the agent runs without its skills catalog until DBOS resolves its catalog fail-closed (DI-08).
Tests that pinned the old wrapper: a test that expects untrusted="true" on a read_skill_file result of a default provider now fails; expect <skill_file skill="…" path="…"> instead, or mark the provider untrusted.
See Skill trust.