Module context_trust 

Context safety: provenance, trust, and pollution detection.

Everything the tiers/map/selection assemble eventually reaches a model. That makes context a trust boundary: a tool output, a fetched file, or a tampered data file can carry prompt-injection (“ignore previous instructions”), role-switching, hidden-unicode payloads, or exfiltration hints. This module tags each source with a provenance and trust level and scans its content for those patterns, so a human (and the assistant flow) can see what is in context, where it came from, and whether it looks poisoned before it is ever sent.

Structs

InjectionFinding

A suspicious pattern found in context content.

TrustReport

A full trust assessment of one context source.

Enums

SourceKind

Where a piece of context came from — the root of traceability.

TrustLevel

How much the content may be trusted to be inert data rather than instructions.

Functions

analyze_source

Assemble a full trust report for one source.

scan_injection

Scan content for injection / pollution patterns. is_code relaxes the instruction-shaped rules (expected in prose) but keeps the data-file rules.

trust_level

Derive a trust level from provenance and content classification. Workspace Rust source is trusted; data/vendored files are only semi-trusted (they can be tampered and should never carry instructions); anything external or model-generated is untrusted.