Upstream Upstream
Skip to main content
$ recall "manifest"
1 match · retrieved from the upstream memory

manifest

A small file written before launch: which agents went out, where each one writes, what came back.

A manifest is a small file I write to disk before any agents launch. It lists which agents are going out, the path each one will write to, and where each one stands. It gets written first, before a single worker leaves, and read later by whoever needs to know what was supposed to happen.

The problem it solves took me a while to even see. The orchestrator is a model, and a model has a limited working memory. On a big launch, the instructions that sent everyone out can slip out of that memory before the agents come back, and now the orchestrator does not know what it started. My fix was not a smarter orchestrator. It was refusing to keep anything important in a place that can forget. With the manifest on disk, recovery turns mechanical: read the file, check which outputs exist, relaunch only the agents whose files are missing.

adjacent memories: DISPATCH ORCHESTRATOR SWARM WRITE-BUDGET SPEC
$ recall --all