Reliability & Ops

These packages give you a circuit breaker designed specifically for agent-to-tool and agent-to-agent communication, with per-tool circuit isolation, confidence-aware tripping, and cost-based rate limiting. You would adopt them to prevent cascading failures when an LLM agent repeatedly calls a degraded tool, or to automatically route traffic away from expensive or unreliable model endpoints. The most distinctive thing is that circuits are identified by a logical `circuitId` (not a hostname), and the state machine uses a lazy auto-transition model with no timers, plus optional leader-elected persistence across restarts via Firestore, DynamoDB, or Redis.

These packages give you a complete session management layer for multi-turn AI agent conversations, handling conversation history windowing, token budget enforcement, context compression, and agent handoff. You'd adopt them to avoid rebuilding the same session lifecycle logic—create, update, end, delete, and list sessions with participants and messages—that every agent system needs. The most distinctive thing is the pluggable storage adapter interface with production backends for Firestore, DynamoDB, and Redis, each implementing optimistic concurrency with version-checked writes and deterministic message ordering, plus three compression strategies (sliding window, summarization, hybrid) with cached summaries so the summarizer isn't re-invoked on every fetch.

These packages give you a repair engine that takes a Zod schema and malformed LLM output, then runs it through six graduated strategies—stripping markdown fences, extracting JSON from prose, fixing syntax errors, coercing types, fuzzy-matching keys, and removing extra fields—to return valid, schema-conforming data instead of crashing. You'd adopt them to handle the common failure modes of LLM structured output generation: trailing commas, truncated streams, Python literals, hallucinated field names, and JSON buried in conversational wrappers. The most distinctive thing is the graduated pipeline approach—each strategy runs in sequence, the engine validates after each step, and it returns as soon as the data conforms, giving you detailed diagnostics including per-field errors and best-effort partial data when full repair fails.