Skip to content

Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure

TypeScript

docs(crds): state the real contentAccess default (absent means content)

Andy Paviad90c76e37f59
diff --git a/charts/clerum-crds/crds/llmhook.yaml b/charts/clerum-crds/crds/llmhook.yaml
index 07570d3ba..fb4217a14 100644
--- a/charts/clerum-crds/crds/llmhook.yaml
+++ b/charts/clerum-crds/crds/llmhook.yaml
@@ -176,13 +176,18 @@ spec:
the projected non-system messages/response bodies; `metadata` =
model+params+usage only, no bodies. moderate/postCallSuccess/
onError are inherently content-bearing; only a preCall-only hook
- may be `metadata`. Omit to let the platform infer (content points
- -> content, preCall-only -> metadata). The install-time content/
- egress gate keys on this: a content hook with egress or a remote
- target requires trust_level high. NOTE: enforced by the runtime
- content projection (target model B, §12.4); until that lands the
- field is advisory and the gate treats every message-carrying
- point as content-bearing (interim fix A).
+ may be `metadata` (the CEL rule below rejects the rest, and
+ install-hook rejects them again with content_access_conflict).
+ OMITTING THIS FIELD MEANS `content`, not `metadata`: absent is
+ read as content-bearing by both the install-time gate and the
+ runtime projection, deliberately, so an unstated projection can
+ never under-declare what a third-party hook sees. A hook that
+ must not receive bodies has to say `metadata` explicitly. Two
+ consequences of the default: a preCall hook that omits this
+ receives every non-system message, and — because the content/
+ egress gate keys on the same value — it is admissible with
+ `egressBindings` or a `remote` target ONLY at trust_level high
+ (otherwise install fails with content_egress_requires_high_trust).
order:
type: integer
default: 100
@@ -241,8 +246,9 @@ spec:
- rule: "!(has(self.capabilities) && self.capabilities.exists(c, c == 'may_deny')) || self.failMode == 'closed'"
message: "a deny-authoritative hook (may_deny) must fail closed"
# contentAccess: metadata is contradictory for an inherently content-bearing
- # point. Absent field passes (platform infers), so this is non-breaking for
- # existing CRs; it only rejects an explicit metadata on moderate/post/onError.
+ # point. An absent field passes (absent is read as `content`), so this is
+ # non-breaking for existing CRs; it only rejects an explicit metadata on
+ # moderate/post/onError.
- rule: "!has(self.contentAccess) || self.contentAccess == 'content' || !self.lifecyclePoints.exists(p, p == 'moderate' || p == 'postCallSuccess' || p == 'onError')"
message: "contentAccess: metadata is not allowed for a hook using moderate/postCallSuccess/onError (those points are inherently content-bearing)"
status: