bide v0.1.0

Most agent frameworks recover a crashed run by re-running it, so your side effects have to be idempotent and a charge, an email, or a shipment can fire twice. bide is built the other way around. It is a Go library for agents that act in the real world (money, records, work done under audit) where "probably ran once" is not good enough.

One append-only journal underpins four guarantees no other framework pairs in a single library:

• Side effects fire at most once. A resumed run replays its decisions from the journal and never re-executes an effect it already committed. Not at-least-once that you make idempotent yourself: at most once.
• Thousands of durable runs in a single process. Crash-safe and highly available across node handoffs, with no cluster and no external workflow engine to run. You import a library, you do not operate infrastructure.
• An audit trail you can prove, not just read. Every run emits RFC 6962 Merkle proofs that anyone can verify offline, without trusting the vendor or the process that produced them.
• Shared state that provably converges. Governed multi-agent state carries machine-checked convergence, not eventual-consistency hope.

And it is a full framework, not only a durability layer: model adapters for OpenAI, Anthropic, and Gemini; tools; the plan typed flow builder; durable timers and typed human-in-the-loop; MCP tool sources; retrieval and memory; and OpenTelemetry in one line.

This release also ships bide-audit, a standalone verifier for proof bundles, prebuilt for Linux, macOS, and Windows on amd64 and arm64 (attached below with checksums), so an auditor can check a run with no Go toolchain.

Live-verified end to end, CI green on all three platforms, Go 1.27. This is a v0: the guarantees hold today, and the API may still move before v1.

Install:

go get github.com/bide-ai/bide@v0.1.0

Docs and guides: https://bide-ai.com