Agent systems are running into problems I recognise: exactly-once effects, ordering under retry, durable state across restarts. I don't think anyone has clean answers yet, mine included — the repos here are attempts at pieces of it.
- Copenhagen, Denmark
-
23:53
(UTC +02:00)
Highlights
Pinned Loading
-
forkrun
forkrun PublicRecord, replay and fork AI agent runs. Forkrun records every model call and tool call, then re-runs your agent's real code against the recording.
TypeScript
-
natsacl
natsacl PublicLeast-privilege NATS permissions compiled from your TypeScript: one broker user per service, a CI drift gate, and JetStream consumer filters verified against your streams.
TypeScript
-
punchin
punchin PublicFork a recorded voice-agent call at the turn it went wrong. Punchin replays the prefix for free and runs the rest live against a customer pinned to what the real one wanted, knew and asked for.
Python
-
ballast
ballast PublicVersion control for what a model has learned. Ballast stores weight deltas as content-addressed blocks, records merges as views over their inputs, and deletes with a proof.
Python
-
sublift
sublift PublicPython library for measuring what a subscription retention experiment was actually worth. Censored lifetime value over an explicit horizon, competing risks, and inference that stays valid when you …
Python
-
lashing
lashing PublicMCP server that lets AI agents book and track container shipments through the DCSA open standards, with every write validated, authorized and logged.
Python
If the problem persists, check the GitHub status page or contact support.

