mirror of
https://github.com/samvallad33/vestige.git
synced 2026-07-24 23:41:01 +02:00
Add the firewall causal slice so a poisoned/injected memory is screened, quarantined, and proven held-out on the receipt (influence_allowed=false): - screen_write() detector (neuroscience/microglial_firewall.rs): deterministic prompt-injection / exfiltration / poisoning screen with NFKC + homoglyph folding and word-boundary matching to avoid false quarantines. - Receipt gains quarantined[] / influence_allowed / engram_phases (additive, serde-default, back-compatible with old receipts). - New trace events memory.quarantine + episode.boundary (mirror the TS union). - gate_writes() now screens every write and, on a quarantine verdict, suppresses the node (held out of retrieval) + records the quarantine receipt + events, so influence_allowed=false is ENFORCED, not cosmetic. Content capped at 16KB. - VestigeEvent::MemoryQuarantined live broadcast. Verified: cargo test vestige-core/vestige-mcp green, clippy -D warnings clean, integration test proves a screened injection is absent from the write's retrieval. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| config.toml | ||