Work around the Anchor 2.0.0-rc.1 IDL enum bug, and fix the one failure that was ours

[mikemaccana]

Nine projects fail anchor build. Eight are one upstream bug, already fixed upstream but unreleased; the ninth was ours and is fixed here.

The upstream bug

anchor-derive-accounts 2.0.0-rc.1 emits enum type definitions with the JSON key "fields" where anchor-lang-idl-spec requires "variants", so anchor build cannot deserialize the IDL it just generated:

Error: missing field `variants` at line 1 column N

One shared helper causes it. type_def_header_parts in the derive crate's src/idl.rs hardcodes the key:

type_def_obj.insert("type".into(), json!({ "kind": kind_name, "fields": [FIELD_MARKER] }));

Both emitters call it: build_struct_type_def_emission with kind_name = "struct", which is correct, and build_enum_type_def_emission with kind_name = "enum", which is not. The spec is IdlTypeDefTy::Enum { variants: Vec<IdlEnumVariant> }, with no serde(default), so deserialization hard-fails.

Reproduced end to end with a twenty-line program containing #[derive(IdlType)] pub enum Side { Bid, Ask }. The emitted JSON is:

"types":[{"name":"Side","type":{"fields":[{"name":"Bid"},{"name":"Ask"}],"kind":"enum"}}]

The variant objects are correct; only the container key is wrong. Unit variants are enough, tuple variants fail identically, and the failure is invisible to cargo check, cargo clippy and cargo test because idl-build is a feature only anchor idl build turns on.

Fixed upstream in otter-sec/anchor#4947 (commit 51323eb, 19 Aug 2026), which adds exactly the missing branch:

let entries_key = if kind_name == "enum" { "variants" } else { "fields" };

That is not in the v2.0.0-rc.1 tag (12 Aug) and is not on crates.io.

Why the workaround is --no-idl and not a dependency patch

Patching only the derive crate to the fixed commit was tried and does not resolve: anchor-derive-accounts on anchor-next requires solana-address 2.7.0, and every program here pins >=2.6, <2.7 because anchor-lang 2.0.0-rc.1 is built against wincode 0.5 while solana-address 2.7 moved to 0.6. Cargo fails with a version conflict on solana-address. The comment in anchor.yml records this so nobody retries it.

Skipping IDL generation costs nothing today. None of the eight consumes a generated IDL: no idls/ directory, no declare_program! in any Rust source. The program is still compiled to a .so and its tests still run, which is what the workflow is actually checking.

The workflow block is marked TEMPORARY WORKAROUND, says to delete it once a release after 2.0.0-rc.1 ships, and links the upstream PR.

The one that was ours

finance/vault-strategy was failing for an unrelated reason:

error[E0277]: `mock_swap_router::AssetRate` has no IDL type information

It has BorshAccount<AssetRate> fields whose type is declared in mock-swap-router, and #[account] only generates the IDL glue when the declaring crate has its own idl-build feature enabled. vault-strategy's idl-build was ["anchor-lang/idl-build"] and never forwarded. It now forwards to mock-swap-router/idl-build, with a comment saying why it is required rather than optional.

Verified locally with anchor idl build: that error is gone. It is not in the skip list, so CI will exercise its IDL generation for real.

Still outstanding, not addressed here

finance/order-book carries impl anchor_lang::IdlBuild for OrderTreeRoot {} at ordertree.rs:37 for a trait v2 removed. It is behind #[cfg(feature = "idl-build")], so nothing compiles it while that project skips IDL generation, but it is dead code that will need the v2 equivalent when the skip is lifted.

basics/pyth is in the skip list rather than fixed because fixing it properly runs into the same upstream bug: its vendored PriceUpdateV2 needs #[derive(IdlType)], which requires its verification_level: VerificationLevel field's enum to be an IDL type too.

---

Generated by Claude Code