Skip to main content
The repo layout:

​
Quick commands

cargo check passing means very little for the on-chain program. cargo build-sbf is the gate, and deployment to devnet is the only thing that proves an instruction works.

​
The toolchain trap

The program requires modern Agave platform-tools (v1.54 / rustc 1.89, installed by Agave CLI 4.1.1 from release.anza.xyz/stable/install). The historical anchor build failure (edition2024 parse errors) was caused entirely by old platform-tools v1.43. Two rules:
  1. Build with cargo build-sbf from the program directory, not anchor build.
  2. Any anchor CLI invocation (including anchor idl build) may silently re-install Solana 2.1.0 and repoint active_release back to the old tools. After any anchor command, repoint:
IDL note: anchor idl build chokes on the two ostatus PDAs whose seed is a function call on an instruction argument. Workaround: temporarily swap those seeds for a plain argument field to emit the IDL, then restore. The runtime .so keeps the real hash-based seeds.

​
Backend environment

Go DB tests need DATABASE_URL (auto-read from the repo-root .env, gitignored — copy .env.example). Each test run creates a scratch database and drops it afterwards. They’re slow against Neon (~300 ms per statement) and occasionally network-flaky — a package failing with dial error / no such host is the network, not the code; re-run it. Scope test runs while iterating: go test ./internal/matching/ is instant, ./internal/api/ takes ~3 minutes. pgx runs in the simple query protocol (Neon’s pooler breaks prepared statements), which imposes three conventions: explicit ::bigint casts in SQL arithmetic on parameters, JSONB parameters passed as Go string (never []byte), and uint64 salts scanned via int64.

​
Running the server

See backend/.env.example for the full environment surface.
Public devnet RPC rate-limits aggressive status polls — the crank and harness poll gently (1.5–2.5 s); a transaction that “times out” usually landed (check solana confirm).