Skip to content

feat: migrate shield-swap-sdk to the new shield_swap.aleo stack - #62

Open
iamalwaysuncomfortable wants to merge 24 commits into
masterfrom
feat/shield-swap-update
Open

feat: migrate shield-swap-sdk to the new shield_swap.aleo stack#62
iamalwaysuncomfortable wants to merge 24 commits into
masterfrom
feat/shield-swap-update

Conversation

@iamalwaysuncomfortable

Copy link
Copy Markdown
Member

Migrates shield-swap-sdk from the retired shield_swap_v3.aleo to the freshly deployed shield_swap.aleo stack, per the migration guide (amm-v3#62 target). Hard cutover — no state carries over.

Contract cutover

  • Wire layer regenerated from deployed testnet bytecode: PROGRAM_ID = shield_swap.aleo; U256 {hi, lo} prices, MerkleProof struct, PositionNFT.withdrawal, reshaped PoolState/Slot/Tick/Position/SwapOutput, new compliance records; claim_multi_hop_output and set_token_decimals gone.
  • Q128.128 math: new _MAGIC_X128 tick table + MIN/MAX_SQRT_RATIO_X128 pinned from the contract and verified against the reference implementation (f(±400000) matches exactly); u256_to_int / int_to_u256_plaintext helpers.
  • Raw native amounts end to end: no dust rule, no 9-decimal normalization; SlotView.price re-derived; collect_all requests exactly the chain-reported owed amounts.
  • Freezelist proofs: mint / claim_swap_output / collect carry [MerkleProof; 2] arrays — empty-tree defaults while the lists are empty, wrapper_proofs= override for later.
  • Immutable withdrawal: mint(withdrawal=…) (defaults to recipient) is stored on the NFT; collect always pays it — the recipient kwarg is removed.
  • Blinding: CLAIM_OR_SWAP_DOMAIN unchanged in the new bytecode; only DEFAULT_PROGRAM flips (v3 vectors kept as algorithm tests, new-program vectors pinned).

Automatic router dispatch

Every verb detects each token's wrapped/plain shape (from_wrapper_token_id chain probe, cached; transport errors propagate rather than misclassifying) and dispatches per the guide's §6 matrix:

  • swap → shield_swap_router.swap_from_wrapped when the input is wrapped (funded with underlying records; deposit + burn happen in-transaction; amount_out_min > 0 enforced at prepare time)
  • claim → 4-way dispatch on the finalized SwapOutput shapes (covers the plain-start/wrapped-claim asymmetry)
  • mint/increase/collect → the 11 LP-router transitions, wrapper proofs interleaved per the deployed input orders
  • decrease/burn → always direct

Record selection understands credits.aleo (microcredits) and skips recipient-bound wrapper records. Mirrored in AsyncShieldSwap.

Supporting changes

  • aleo.codegen (main SDK): fixed-length Array ABI type support (fmt_array, toposort/ref-check recursion into array elements).
  • API layer → staging: DEFAULT_API_URL and regen default now https://amm-api-staging.dev.provable.com (SHIELD_SWAP_API_URL override); models regenerated (wrapper_programamm_token_program, new underlying_program/underlying_token_id drive record-program resolution); redeem no longer rotates the session credential.
  • Devnode harness: deploys the new 5-program stack (core + multisig + freezelist + 2 test tokens), initializes the freezelist, registers plain tokens via the two-arg allow_token with idempotent bootstrap.

Verification

  • shield-swap-sdk unit suite: 171 passed
  • main SDK suite: 905 passed (includes codegen Array coverage)
  • live read + drift tier (real testnet + staging API): 11 passed — pinned ABI matches deployed bytecode, X128 invariants on live slots, local pool/tick key derivation parity, registry↔chain wrapped-detection agreement
  • devnode lifecycle (hermetic, full stack): 11 passed — pools → mint → increase → decrease → both-direction swaps + claims → owner≠withdrawal collect payout → burn
  • live write tiers (plain + wrapped swap roundtrips) are wired and skip cleanly pending DPS e2e credentials

- /auth/verify: send challenge_id from the challenge payload
- sessions are httpOnly cookies + X-CSRF-Token (legacy body-JWT kept)
- is_authenticated covers both credential kinds
- salted retry on DPS consumer-username collisions
- lifecycle live test: invite codes are pasted, never generated; shed
  SHIELD_SWAP_PRIVATE_KEY so the fresh-profile premise holds
…nt under cookie auth

- _headers: prefer the live session (csrf + cookies) over Authorization;
  ss_ tokens don't cover the /access tier and the server reads the header first
- 401 while both credentials are loaded drops the expired session and
  retries as bearer (15-min sessions)
- _auth_done recognizes cookie sessions; CSRF never persisted as jwt;
  from_profile prefers the durable ss_ token over stale session creds
Slot-derived insert hints only validate for a pool's FIRST position —
finalize asserts the true linked-list predecessors, so every later mint
was rejected on populated pools. find_tick_predecessor walks the ticks
mapping from the MIN sentinel (fresh pools anchor at the sentinel;
initialized ticks return themselves since validation is skipped).
… 0.3.0

shield-swap-sdk now floors aleo-sdk at 0.3 (Array codegen runtime).
Docs updated: staging auth/session + referral-vs-access invites, live-
verified LP behaviors (tick-list hints, wrapped-side exact amounts,
slippage headroom), sdk-abi example re-pointed at the new stack.
* docs: voice.md docstrings for the undocumented public surface

Adds docstrings to the 206 public classes/methods in sdk/python/aleo and
shield-swap-sdk that had none, following .agents/voice.md: present-tense
verb lead, side effects named (network / fee / local-only), and each
argument, return, and raised error described by consequence.

Section syntax follows each file's local idiom rather than one global
rule — numpy `Parameters` inside facade/ (54 existing blocks, and
voice.md shipped in the same commit as those files), Google `Args:`
elsewhere. shield-swap keeps its terser prose voice and its
"see :meth:`ApiClient.X`" convention for async mirrors. Where a sync
counterpart was already documented the text is mirrored, with
async-specific facts corrected: wait_for_transaction_confirmation yields
to the event loop via asyncio.sleep, AsyncDexCall.simulate explains why
it is not awaited, and three cross-references now point at the async
classes instead of RecordScanner / AleoNetworkClient.

Three behaviours were checked against the code rather than assumed:
get_block_range's end-inclusivity varies by node build (the e2e test
says so) so the docstring warns against relying on the last element;
decrypt_enabled gates find_credits_record(s), not owned(), which
decrypts opportunistically; and codegen's main() exits via argparse on a
usage error rather than returning 1.

Also removes ApiClient.generate_access_codes and its async mirror.
Minting invite/access codes should not be part of the SDK surface.
redeem_access_code stays, so a code obtained out-of-band still works,
and human-pasted referral invites keep going through redeem_code. Note
this reduces SDK surface, not access: POST /access/generate is still
reachable directly, and the real gate is the server-side generate right.
test_minting_access_codes_is_not_exposed guards the removal.

shield-swap-sdk/AGENTS.md is generated from these docstrings, so both
copies are regenerated. Four TIER2 entries (api.get_pools,
api.get_tokens, derive_pool_key, derive_tick_key) rendered with blank
bodies before and now carry real text, which pushed the page past
test_gen_context.py's compactness budget — that threshold moves 22k to
24k, with the reason recorded alongside the prior 20k to 22k raise.

Verified: 890 passed (sdk, -m "not slow"), 174 passed (shield-swap),
pyright strict 0 errors on sdk, gen_context.py --check clean. The 15
pyright errors in shield-swap-sdk are pre-existing and unchanged
(confirmed by re-running against HEAD).

* fix(facade): exact credits/microcredits conversion (#66)

Both directions went through binary floating point and lost value.

`credits_to_microcredits` multiplied by 1_000_000 as a float and then
truncated toward zero with int(), so ordinary amounts silently underpaid:
1.005 credits became 1_004_999 microcredits, not 1_005_000. Sub-microcredit
input was dropped with no error at all (0.9999999 -> 999_999).

`microcredits_to_credits` returned a float. Microcredits are a u64, and past
2**53 a float cannot hold the integer — u64 max round-tripped off by one, and
2884 of the first 200_000 microcredit values failed
`micro -> credits -> micro` identity.

Both now compute in Decimal. Float input is routed through str(), which
recovers the shortest representation that round-trips — i.e. the literal the
caller wrote — which is what rescues 1.005; str and Decimal input are exact
already. Sub-microcredit precision now raises ValueError naming the value,
with allow_rounding=True to opt back into truncation, so lost value is an
error rather than a silent underpayment.

`from_microcredits` returns Decimal instead of float. It still compares equal
to the obvious float, so existing assertions hold unchanged, but mixing it
into float arithmetic now raises — convert with float() deliberately if you
want that, accepting the loss.

Scope is contained: these are user-facing convenience helpers only. No
internal fee or amount path consumes them (the SDK is integer microcredits
end to end), and shield-swap does not use them.

Verified: 896 passed (sdk, -m "not slow"; +6 new), pyright strict 0 errors,
shield-swap 174 passed and unaffected.

* docs: clarify OwnedFilter uuid default wording

* docs: drop the self-hosted-scanner recommendation from record docstrings

* docs: drop redundant 'Hits the network' notes and the block-range build caveat

* docs: make get_pools/get_tokens concrete, name methods instead of 'verbs'

* docs(voice): ban "reach for" and vague hedges

* docs: drop self-hosted-scanner advice everywhere, keep the view-key disclosure

* docs(voice): drop self-hosting from the privacy stance

* docs: drop self-hosted-scanner mentions from the READMEs

* docs(shield-swap): say "methods", not "verbs"

* docs: tighten get_pools, explain why token info can be None

* docs: simplify the token-info None explanation

* docs(voice): require plain verbs; reword get_tokens decimals note

* docs: reword get_swap lag note and get_public_balances

* docs: state get_ohlcv's real granularity values and unix-second bounds

* docs: explain what the Journal is on the class itself

* docs: state that a profile holds one Aleo address

* docs: add a Profile create/load usage example
* fix(shield-swap): resolve the DEX API host per network

The hard-coded staging host (amm-api-staging.dev.provable.com) now 404s —
the deployment moved and shield_swap.aleo is live on mainnet as well as
testnet. There are two API hosts now, one per network, so a single
module-level constant cannot be right for both.

Replaces DEFAULT_API_URL's baked value with SHIELD_SWAP_API_URLS keyed by
network and api_url_for(network) to resolve it. ShieldSwap and
AsyncShieldSwap now default api_url from their bound client's
network_name, so the off-chain indexer always matches the chain being
read — a testnet pool key means nothing to the mainnet indexer, and the
old default silently guaranteed one of the two was wrong.

SHIELD_SWAP_API_URL still overrides every network. An unknown network
raises rather than falling back, and the standalone-ApiClient default
points at testnet deliberately: an accidental default must not reach
mainnet.

Verified both hosts serve /tokens and /pools unauthenticated and gate
/access/status with 401. 179 passed (shield-swap; +5 new).

* fix(shield-swap): walk the tick list on increase_liquidity; OHLCV takes unix seconds

Two correctness bugs, both confirmed against the live testnet.

increase_liquidity derived its insert hints from pick_insert_hint, which
reads slot.next_init_below/above — those bracket the pool's CURRENT tick,
not the target. Any bound further out than one initialized tick therefore
got a hint above itself, which finalize rejects after the fee is spent.
mint already walked the on-chain list (aa71c33); increase_liquidity now
does the same via find_tick_predecessor.

pick_insert_hint is deleted rather than left in place. It was unused after
this change, unexported, untested, and its own docstring conceded it
returns hints the contract rejects — a known-wrong helper is a trap.

get_ohlcv typed from_ts/to_ts as str, but the API declares from/to as
int64 unix seconds (inclusive start, exclusive end). test_api_get_ohlcv
passed ISO-8601 strings and failed with 400 "query parameters do not match
the expected schema" — it had never run in CI, being live-marked. Both are
now int, and the test passes real unix seconds.

Verified: 14/14 live reads pass against api.testnet.swap.shield.fi (the
OHLCV test was the only red one), 179 shield-swap, 873 sdk, 11 devnode.

* chore(shield-swap): regen OpenAPI per-network; pick up UsdcUsdQuote + pool valuation

* feat(shield-swap): owned-position views with the contract's view math

get_owned_positions(pool_key=?) and get_owned_position(token_id) answer
"what do I hold and what is it worth right now" without a transaction.

A position spans two sources that neither side can answer alone: the
private PositionNFT record carries identity (pool, range, withdrawal) and
no amounts; the public positions/slots/ticks mappings carry amounts and no
identity. Callers previously had to persist token ids externally and
reimplement two pieces of contract math to display a position.

position_math.py mirrors the amm-v3 view helpers bit-exactly —
amounts_for_liquidity (view_amounts_for_liquidity), fee_growth_inside
(get_fee_growth_inside), fee_owed, and u256_wrapping_sub (u256::u256_sub).
Fee growth is 256-bit and modular by design: an outside counter may exceed
the global one and the difference wraps at 2^256, so every subtraction goes
through the wrapping helper — a plain - would raise where the contract
wraps.

The 17 math vectors are transcribed from the contract's own
tests/test_amm_helpers.leo, including the wrap-negative fee_growth_inside
cases, so a divergence in either implementation fails the suite rather than
silently producing wrong balances.

state is None while a mint finalizes (record spendable, mapping not written)
and when a boundary tick is uninitialized — the identity stays usable in
both cases. Burned positions cannot appear, since burn consumes the record.

196 passed (+27: 17 math vectors, 10 join/filter/lag paths).

* feat(shield-swap): swap reserves its blinding counter and journals the handle

swap() derived its blinded identity through next_blinded_identity, which
scans for the first counter the chain does not carry. Correct in sequence,
unsafe in parallel: two swaps starting together read identical chain state,
reach the same counter, and the second reverts at finalize once the first
consumes it. Nothing surfaces locally — at proving time the address
genuinely was unused, because the check and the use are not atomic.

swap_many already avoided this by reserving from the journal; single swap()
did not. It now takes the same path: reserve one counter under the journal's
file lock, derive the identity at it, and record the resulting handle once
the broadcast is accepted. Two concurrent swaps can no longer collide.

Journaling the handle is the other half. The blinding factor is the only
thing that can claim a swap — lose it and the output is unclaimable by
anyone, which is the point of blinding it. Recording at accept time (not
confirmation) means a crash mid-flight leaves a claimable handle for
collect_all(). A journal write that fails after the swap lands raises
rather than being swallowed: the swap is already spent, so silently
dropping its claim secret is the worse outcome.

track=False opts out, an explicit identity= still wins, and without a
journal the on-chain probe remains the only option — documented as racing.

206 passed (+5 covering reservation, distinct counters, opt-out, explicit
identity, and the journal-less path).

* fix(shield-swap): keep pyright clean on the reserved-counter narrowing

* fix(shield-swap): the airdrop stage is testnet-only, say so on mainnet

The mainnet API publishes neither /airdrop nor /airdrop/{job_id} — verified
against both OpenAPI specs — so requesting one there returned 404 and blew
up onboard() with a DexApiError. The remedy is the caller funding the
account, not a retry, so the stage now raises NotFundedError naming the
network and the address. Mainnet onboarding is authenticate -> redeem ->
credentials, then the caller funds, then the funded stage passes.

* feat(shield-swap): from_profile takes network and endpoint

A profile-bound client could only ever be testnet: from_profile called
Profile.load_or_create with no arguments, and that defaults network to
testnet. With shield_swap.aleo now deployed on mainnet, mainnet was
unreachable through the documented entry point.

Both apply only when the profile is created — an existing one keeps what it
was created with, since its derived pool keys and blinded identities are
network-scoped and would not transfer. Give each network its own home.

Verified against both live deployments: 5 pools/8 tokens on testnet and
4 pools/8 tokens on mainnet, with locally derived pool keys matching each
indexer on both.

* fix(shield-swap): the write tier never passed its own credentials

conftest gated the write tier on ALEO_E2E_API_KEY / ALEO_E2E_CONSUMER_ID but
built its provider without them, so the hosted record scanner answered
Unauthorized and every write test failed on the first record read — before
reaching the chain. It also never registered the account, which scanning
requires.

Confirmed by contrast: the same account reads private balances fine when the
credentials are wired in (57 credits + test ETH + USDCx on testnet, 0.4
credits + USDCx on mainnet), and fails with exactly this Unauthorized when
they are not.

The provider now receives both, and the account is registered with the
scanner once a key exists.

* fix(shield-swap): authenticate the write-tier clients before auth-gated reads

get_route is auth-gated; the bespoke clients in test_swap_lifecycle never
established a session, so it answered 401 before anything was proved.

* fix(shield-swap): quote failures stay errors; expanduser; pyright to zero

Four fixes, each a way a wrong value used to reach the chain or the disk.

_quote_expected_out swallowed NotAuthenticatedError/NotRedeemedError into a
None return, which resolve_swap_params then replaced with a spot estimate.
Spot ignores the pool fee, so amount_out_min came out above what the pool
can pay: the caller paid for a proof the finalize rejected. "Could not ask"
is not "no route" — auth failures now propagate, and swap_many refuses
outright when it has no quote and slippage_bps < 10000 rather than proving
and broadcasting N swaps engineered to revert. slippage_bps=10000 ("accept
any output") still proceeds without one.

Profile never called expanduser, so load_or_create("~/x") and
SHIELD_SWAP_HOME=~/x each created a literal ~ directory in the cwd and
wrote the private key there — where no later run would look for it. Both
paths now expand.

pyright: 15 errors -> 0. Twelve were real. _lp_programs was annotated str
while its own docstring said "None when an explicit record made resolution
unnecessary"; the annotation was simply wrong. select_token_record's callers
relied on a short-circuit pyright cannot see, so mint and increase now share
_fund_side, which resolves record and program together and removes the
duplication. authenticate returned self._csrf (str | None) as str. Two
dict comprehensions produced list[str | None] despite filtering.

The last three were unresolved mcp imports — a declared optional extra, now
installed so pyright checks the MCP server rather than skipping it. That
surfaced a latent bug in sdk-abi's own stub: _aleo_abi.pyi declared
generate_abi with three parameters while the pyo3 signature and the runtime
both take four (imports=None). aleo.abi passes four and was correct; the
stub was wrong and is now generated from the Rust signature.

Bumps all three packages 0.3.1 -> 0.4.0: this branch removes
generate_access_codes, replaces DEFAULT_API_URL, and retypes get_ohlcv's
timestamps, none of which are patch-compatible.

220 shield-swap, 884 sdk, pyright clean in both packages.

* fix(shield-swap): quote the route in canonical amounts, not base units

test_private_swap_roundtrip proved, broadcast, and had the network reject
at1xze86e… — fee consumed. Diagnosed from the rejected transition's inputs:
amount_out_min was 1_844_890_080 while sqrt_price_limit sat at exactly
MIN_SQRT_RATIO_X128, the default extreme, so the price bound was not the
constraint. The minimum was simply unpayable.

Cause, measured against the live API on the ETH/ALEO pool:

  get_route(amount_in=10000000000000000)  -> 1863.544605   (raw base units)
  get_route(amount_in=0.01)               -> 1058.294112   (canonical)

/route takes a CANONICAL decimal amount. The test passed raw base units, so
the API quoted a trade of 10_000_000 ETH rather than 0.01, returned a price
from deep in the book, and the test scaled that into a minimum 76% above
what the pool would actually pay.

_quote_expected_out was correct throughout — it divides by 10**dec_in before
asking and returns 1058294112, matching the canonical quote exactly. The
test reimplemented the conversion and got it wrong, so it now calls the
helper instead.

test_reads_live had the same confusion, passing 10**decimals as amount_in.
It never failed because it asserts only shape, but it documented the wrong
convention; it now passes "1" with the units spelled out.

The write tier now passes for the first time: 2 passed, both roundtrips
proved, broadcast, confirmed and claimed against real testnet.

* fix(shield-swap): address self-review — async parity, build-time cost, dead code

Six findings from reviewing the branch.

swap() reserves its counter at BUILD time, not at the terminal method: the
blinded address is a transition input, so the identity must exist before
anything can be assembled. That means discarding a prepared call — or only
calling simulate() — still spends a counter, which contradicts the
"nothing happens until a terminal method" contract the README states. The
reservation cannot be deferred, so it is documented instead of hidden, with
track=False offered as the side-effect-free build. Three tests pin the
behaviour rather than leaving it as prose.

Async parity: get_owned_positions/get_owned_position now exist on
AsyncShieldSwap. sdk/AGENTS.md requires sync+async pairs with shared pure
logic, so the PositionNFT record shape moved to _core as
POSITION_RECORD_FIELDS + decode_position_record rather than being copied
into the second client. test_async_parity fails on any future read method
added to one client and not the other, and also fails when its own SYNC_ONLY
allowlist goes stale. The async swap docstring pointed at
"journal-reserved counters" the async client cannot reserve; it now states
that it probes, that the probe races, and that concurrent callers must pass
identity explicitly.

get_owned_positions read the slot once per position; ten positions in one
pool cost ten identical reads. A per-call cache keyed on pool brings that to
one, asserted by counting the calls.

Record detection keyed on "tick_lower" alone, so any future record type
carrying that field would be misread as a position. It now requires the whole
field set, with a test using an impostor record that shares one field.

amounts_for_liquidity ended in an unreachable `return 0, 0` — its three arms
are exhaustive by construction. A silent (0, 0) would read as "holds
nothing", so it raises AssertionError instead. amount0_delta divides by both
bounds and now documents ZeroDivisionError.

The swap docstring renders into AGENTS.md and pushed the page past its
compactness budget; it is written tighter rather than raising the cap a third
time (23866, 134 spare).

229 shield-swap (+9), 884 sdk, 14 live reads, pyright clean in both.

* fix(shield-swap): await the async record scan; both Copilot findings

AsyncShieldSwap.get_owned_positions did not await record_provider.find.
AsyncRecordsModule.find is `async def`, so `records` was a coroutine and
iterating it raised "TypeError: 'coroutine' object is not iterable" — the
method could never have worked. Reproduced before fixing.

My own parity test missed it because it asserted hasattr, not behaviour: a
method that exists and always raises satisfies a presence check. So this adds
test_owned_positions_async, which drives the async views against a fake whose
find() really is a coroutine — the join, the finalize lag, record filtering,
the pool filter, lookup by id, and the empty case. Verified as a real guard by
removing the await again: all six fail, and pass once it is restored.

Also: test_airdrop_stage_refuses_on_mainnet had its docstring after an
assignment, making it a no-op string expression rather than a docstring
(ast.get_docstring returned None). Moved to the first statement.

Both found by Copilot review; both were real.

235 shield-swap (+6), pyright clean, AGENTS.md current.

* fix(shield-swap): swap_many takes expected_out; share the position decoder

Two findings from a second review pass.

swap_many's refusal message told callers to "quote it yourself and pass
expected_out" — and swap_many had no expected_out parameter, so the advice
was impossible to follow. It now accepts one, which both makes the message
true and gives callers with their own price source (or an unreachable API) a
way through.

find_position_plaintext matched any dict whose `pool` field equalled the
target, so a record of another type carrying that field would be returned as
a position and then spent as one. That is the same defect fixed in
_owned_from_record last commit; its sibling was left behind. Both now go
through decode_position_record, with a test using a lookalike record.

The agent page passed its compactness budget again. It had been squeezed to
134 chars of headroom, so the cap moves 24k → 26k deliberately rather than a
third squeeze, with the reason recorded alongside the earlier raises. The new
paragraph is also written tighter; the page sits at 24130 with 1870 spare.

237 shield-swap (+2), pyright clean.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant