Skip to content

What's New in 4.4.0

Docs for Briefcase AI v4.4.0 · see what’s new

briefcase-ai v4.4.0 makes append-only integrity chains safe for concurrent writers. It builds on two larger releases: 4.3 introduced canonical JSON, hash chains, and Ed25519 signing; 4.2 added invocation controls, TypeScript packages, auto-instrumentation, and expanded data adapters.

4.4.0: concurrent integrity chains

  • ChainConflictError makes stale-tail writes explicit.
  • HashChainAppender rereads the tail and retries with jitter, bounded by max_attempts or maxAttempts, default 32.
  • JsonlHashChainStore holds POSIX flock across tail reads and appends and incorporates lines written by other processes.
  • Windows JSONL coordination remains thread-only.

4.3.0: integrity and signing

  • briefcase.integrity adds strict canonical JSON, a compatibility profile, spec-versioned hash chains, in-memory and hardened JSONL stores, and chain segment verification.
  • The integrity extra adds local Ed25519 signing, verification, RFC 8037 JWKs, and JWK thumbprints.
  • Bitemporal and examiner-bundle hashes route through the compatibility profile without changing stored bytes.

4.2.0: controls and cross-language packages

  • briefcase.controls and @briefcase-ai/controls add gateway outcomes, quota ports, cooldowns, retries, provider fallback, suggestion pipelines, exporters, and shared semantic conventions.
  • @briefcase-ai/native exposes Rust-backed cost estimation and normalized Bedrock model IDs, including Amazon Nova and embedding prices.
  • @capture adds capture_content="full" | "hash" | "none", bounded full-mode rendering, one shared export worker, and a zero-work unconfigured path.
  • briefcase.auto.auto() and undo() instrument LangChain, CrewAI, LlamaIndex, AutoGen, AG2, OpenAI Agents, and PageIndex.
  • New integrations include OPA, RBAC/ABAC, GCP and OTel exporters, Kafka and webhook transports, Chroma/Pinecone/Weaviate, seven VCS adapters, Glue Iceberg, kdb+, and KMS-signed examiner bundles.

4.1.0: behavior that now executes

  • Replay uses with_executor() to run current code and evaluate policy rules.
  • DecisionSnapshot.content_hash() covers outputs and other decision content.
  • BufferedBackend batches writes and exposes flush(), pending(), reads, and context-manager cleanup.
  • TimeoutWrapper stops waiting at its deadline and returns the configured fallback effect.

Upgrade checklist

  1. Change content-sensitive comparisons from fingerprint() to content_hash().
  2. Put every model call behind Gateway when caps or quotas must be enforced.
  3. Choose capture_content="hash" or "none" when raw content must not persist.
  4. Anchor the head or prior hash outside a chain before treating it as evidence.
  5. Pin caller-trusted key IDs when verifying signed examiner bundles.