Skip to content

README.md

Latest commit

 

History

History
48 lines (37 loc) · 1.87 KB

README.md

File metadata and controls

48 lines (37 loc) · 1.87 KB
Raw

Chidori run store on Cloudflare Durable Objects

A small Worker that gives every Chidori run actor-grade storage: one Durable Object per run holds the run's journal and durable artifacts in SQLite-backed storage, so every write Chidori acknowledges has been replicated across data centers (Cloudflare's Storage Relay Service: a write is confirmed only after a majority of geographically distributed followers hold it, then batched into object storage) and is covered by 30-day point-in-time recovery.

Chidori's replay model is untouched — durability is still the deterministic effect journal. This backend changes where the journal sleeps at night: local disk stays the fast primary; the Durable Object mirror is the copy that survives losing the machine.

Deploy

cd integrations/cloudflare-durable-objects
npx wrangler deploy
npx wrangler secret put CHIDORI_RUN_STORE_TOKEN   # optional but recommended

Point Chidori at it

export CHIDORI_RUN_STORE="https://chidori-run-store.<account>.workers.dev"
export CHIDORI_RUN_STORE_TOKEN="<the same token>"
chidori serve agent.ts            # or chidori run agent.ts

Every run now tees its journal, snapshot manifest, signal inbox, and detached-agent registry to the Worker. Reads stay local; after machine loss, chidori resume <file> <run_id> (or the server loading a session) hydrates the run directory back out of the mirror automatically.

Optionally:

export CHIDORI_DURABILITY=strict   # a failed mirror write poisons the run
                                   # instead of being logged and tolerated

Protocol

The Worker speaks the HttpRunStore REST protocol defined in crates/chidori/src/runtime/store.rs (records + blobs per run, a run index, and the detached-agent registry). Any server implementing that protocol works as a mirror — the Durable Object deployment is the reference implementation.