✓ VERIFIED
MODULES

Survives Claude Code's auto-compaction without state loss. Detects post-compaction requests and re-injects the most relevant pieces of the most recent eviction archive into the system prompt.

Per-request, per-day, and per-month USD budget enforcement. Short-circuits requests that would breach a configured limit with a 429-shaped response.

Returns cached responses for byte-identical requests. Hash-keyed by a stable serialization of the canonical request — fastest cache type with zero false positives.

Content filtering at the gateway layer. PII redaction (email, SSN, credit-card), profanity blocking, and user-supplied regex patterns. v2 adds NIM and Anthropic Constitutional backends.

Compresses older messages instead of dropping them when a request exceeds the context window utilization target. Keeps long conversations going past context limits.

Prunes MCP tools to the ones actually relevant to the user's last message. Solves the 67k-tokens-of-MCP-tools problem by scoring each tool against the request via embeddings.

Injects relevant past patterns into the system prompt and forges new ones from successful resolutions. The Golden Loop — your gateway gets smarter over time.

Restructures requests to maximize Anthropic prompt-cache hits. Sorts tools deterministically and stamps cache-control markers on the longest stable prefix.

Pulls archived context back when the user references past turns. Companion to ipc — when the user says 'remember when…', searches eviction archives semantically and re-injects the most relevant turns.

Picks the best model for each request. Strategies: cheapest-first, fallback chain, or q-learning based on per-query historical success rate.

Returns cached responses for semantically similar (not just identical) requests. Embeds the request, vector-searches past requests, replays the cached response if similarity is above threshold.

Observes (and in v2 will cache) MCP tool-call results. Records tool_use → tool_result pairs and stamps observability metadata for repeated calls. Excludes side-effecting tools by default.