Introduce modules/amm — the AMM business logic as a universal core Logos module,
consumed identically by the QML UI (via modules().amm_module) and headlessly
(logoscore call amm_module ...). The module is a thin transport adapter: the
domain math lives in the Rust amm_client crate (the transport-independent JSON
FFI), and the module sequences those pure ops with chain I/O delegated to the
logos_execution_zone wallet module. It reaches the same shared wallet instance
the UI opened (Basecamp loads core modules as singletons; standalone the
LogosAPI client cache dedups the connection), so it never opens a second wallet.
The module owns the full AMM surface — not just swaps:
- resolvePool / swapExactInput / tokenList (the swap path)
- newPositionContext / quoteNewPosition / submitNewPosition (add-liquidity)
apps/amm: delete the app-side orchestration (SwapRuntime, NewPositionRuntime,
AmmClient/BundledAmmClient) and the amm_client link. AmmUiBackend now owns only
wallet-session lifecycle and forwards every AMM slot to modules().amm_module.
The one wallet-keyset mutation add-liquidity needs — creating a fresh LP holding
— stays in the backend (via its wallet provider, keeping the account model and
on-disk storage coherent): the module returns "requires_fresh_lp" without
submitting, the backend creates the account and resubmits with its id.
flake.nix / CMakeLists / metadata: the module links the amm_client crate; the UI
links no external lib and depends on amm_module (injected into the UI builder's
flakeInputs so the dependency resolves).
- amounts/deadline declared nlohmann::json so the generated dispatch accepts a
JSON number (bare small ints on the CLI) or a string (exact u128 from the UI,
or a quote-wrapped big value on the CLI); JSON floats are rejected rather than
submit a silently-rounded amount.
- AMM_DEBUG-gated tracing for the swap path.
- Drop tests/cpp/NewPositionRuntimeTest.cpp with the class it covered
(module-level tests to follow).
- modules/amm/README.md: architecture, headless prerequisites, logoscore recipe.
The AMM host FFI was split across two crates with two ABI styles: the
typed-C `amm_client_ffi` (swap primitives, under programs/) and the
JSON/wire `amm_client` (new-position flow, under apps/). Fold both into a
single `amm_client` crate exposing one JSON C ABI, and delete
`programs/amm/client-ffi`.
Rust:
- Re-express the swap path as `api/swap.rs` operations on the existing
`call::<T>` dispatch — swap_pair, resolve_pool, swap_plan, program_id —
reusing `pair::derive_pair` (no more duplicated PDA derivation) and
`risc0_zkvm::serde` for the SwapExactInput words (the same encoding the
guest decodes). The account list and signer flags stay byte-identical
to the old typed path.
- Generate a single header (`include/amm_client.h`) covering all ops via
cbindgen; bump cbindgen 0.27 -> 0.28 for `#[unsafe(no_mangle)]` support.
C++:
- Extend the `AmmClient` wrapper with the four swap ops.
- Add `SwapRuntime` (mirrors `NewPositionRuntime`): reads accounts through
the wallet, drives the swap ops, submits the transaction.
- `AmmUiBackend` swap methods now delegate to `SwapRuntime`, dropping ~390
lines of typed-FFI and byte-twiddling. `program_id` becomes a JSON op,
and the swap clock is derived via `derive_pair` (clock_core::CLOCK_01)
instead of a hardcoded base58 literal — same account, verified.
networkSnapshot() rebuilt the new-position network context on every call —
deriving ammProgramId from $AMM_PROGRAM_BIN and resolving the $TOKENS_CONFIG
token ids, which run tokenList()'s remote account_id_from_base58 conversions.
It is called on the quote hot path (every keystroke) and, critically, from
inside runtime reply callbacks: after a create-pool submit, pool activation
runs refreshContext() from within a quoteNewPosition reply, so the nested
synchronous remote calls reentered the module connection and hung the reply.
refreshNewPositionContext never completed, contextLoading never cleared, and
the token selectors (gated on !contextLoading) stayed disabled — the view
became unusable after creating a pool.
$AMM_PROGRAM_BIN and $TOKENS_CONFIG are fixed for the process lifetime, so
resolve ammProgramId and the token ids once and cache them; networkSnapshot()
now returns the cached values with no per-call remote work. (Still gated to
"loading" until wallet state resolves, so the one-time resolve happens at
startup, not inside a callback.)
LogosWalletProvider::submitPublicTransaction passed the RISC0 instruction
words to send_generic_public_transaction as a QVariantList<u32>. That param
is a byte string (bstr): the module's QtRO glue mangles a list-of-u32 crossing
the module process boundary, so the guest deserialized a corrupted Instruction
variant and panicked, e.g.
Guest panicked: called `Result::unwrap()` on an `Err` value:
Custom("invalid value: integer `53765`, expected variant index 0 <= i < 10")
(53765 = 0x0000D205 — the guest read a 4-byte word where 0x05, the AddLiquidity
variant, is a single byte.) This broke every submit through the shared provider,
including the new-position / add-liquidity flow.
Send the little-endian bytes of the u32 words as a QByteArray instead — the same
encoding the AMM swap path already uses (it calls the module directly and never
went through this provider). signingRequirements stays a QVariantList<bool>;
only the instruction needed the bstr encoding.
The create-pool / new-position flow carried its own network layer:
AMM_UI_NETWORK + AMM_UI_DEVNET_FILE (a devnet.json) or a bundled
config/networks.json supplied the AMM program id and token set, and a
JSON-RPC channel/checkpoint "identity probe" gated the flow to a verified
network. Main already exposes all of this the way the Swap view consumes it,
so collapse onto those sources instead of a parallel system:
- ammProgramId <- $AMM_PROGRAM_BIN (derived like swapExactInput's program id;
doubles as the quote's network fingerprint so a quote can't
be replayed against a different deployment)
- tokenIds <- $TOKENS_CONFIG (amm-tokens.json), same as the Swap picker
- sequencer <- the wallet config (already surfaced via syncWalletState)
networkSnapshot() builds the ActiveNetworkSnapshot from those; status is
"ready"/"config_missing", gated to "loading" until wallet state resolves so no
module reads happen during construction. The channel probe is gone — submit
needs no channelId (the wallet module supplies the channel via
submitPublicTransaction), so the whole verification apparatus was overhead.
Removes: AMM_UI_NETWORK / AMM_UI_DEVNET_FILE, devnet.json / networks.json, the
ActiveNetwork class (+ its test) and its QNetwork channel probe, and Qt6Network.
ActiveNetwork.h keeps only the ActiveNetworkSnapshot struct. Run command drops
the AMM_UI_* vars:
```
LEE_WALLET_HOME_DIR=… AMM_PROGRAM_BIN=… TOKENS_CONFIG=… nix run .#amm-ui
```
Keep wallet assets as token choices without automatically selecting a pair. Keep pool-probe amounts internal so active-pool quotes do not populate the form before user input.
Add a "Create Pool" flow to the AMM app that lets a user open a new
liquidity position — seeding a pool's initial liquidity — from token
selection and amount entry, through a confirmation dialog, to on-chain
submission.
- client crate (apps/amm/client): pure, testable protocol logic — account
decoding, pair/position modelling, and quote/plan computation — exposed to
the app over a C ABI (config/networks.json drives network selection).
- C++ runtime + backend: AmmClient, ActiveNetwork, and NewPositionRuntime,
wired into AmmUiBackend (new resolve/quote/submit slots).
- QML flow: NewPositionForm, NewPositionFlow state, NewPositionConfirmation-
Dialog, TokenSelectorModal, and reusable Amm* presentational components
(theme, surfaces, buttons) + AmountMath.js.
- tests: C++ (NewPositionRuntimeTest, ActiveNetworkTest) and QML
(tst_NewPositionForm, tst_LiquidityPage, tst_TokenAmountInput, …).
The Logos.Wallet QML plugin was installed with an ELF "$ORIGIN" rpath.
macOS dyld does not expand "$ORIGIN", so loading the plugin failed with
Cannot load library liblogos_wallet_qmlplugin.dylib:
Library not loaded: @rpath/liblogos_wallet_qml.dylib
even though liblogos_wallet_qml.dylib sits in the same directory. Select
the rpath per platform: @loader_path on Apple, $ORIGIN elsewhere, so
@rpath/liblogos_wallet_qml.dylib resolves to the sibling library.
Turns the dummy-data AMM UI into a real client of the on-chain LEZ wallet.
Adds a hand-written ui_qml C++ backend (src/AmmUi*) over the core
logos_execution_zone module: create/open a local wallet, create and list
public/private accounts, and a navbar Connect / Connected + account-selector
+ Disconnect flow. Onboarding is password-only (no path picking) with a
per-app wallet at ~/.lee/amm-wallet (override: AMM_WALLET_HOME_DIR);
standalone gets its own wallet, Basecamp shares accounts via adopt-on-start.
Requires Nix with flakes; macOS also needs `sandbox = false` (the default).
The logos_execution_zone input is pinned to a module rev whose LEZ (lssa)
already includes the macOS Metal-build fix, so no `--override-input` is
needed — plain `nix run .` works:
cd apps/amm
nix run .
- create_new now returns the new wallet's BIP39 mnemonic (not an int status);
the app currently discards it, so the wallet can't yet be recovered. Surfacing
it in onboarding (+ restore_storage) is a follow-up.
- The wallet password is currently a no-op upstream (storage.rs: "TODO: use
password for storage encryption"); storage.json is plaintext. So Disconnect
is a UI-level lock and reconnect does not (cannot yet) re-prompt for it.
- wallet-ffi requires explicit config/storage paths; a *_default() FFI would
let the app drop its path handling.
- Bundled network config: connects to whatever WalletConfig::default() points
at; real testnet endpoints still TBD.
r4bbit