Approved workflow

Implemented LaunchToken and StakingVault, with ABI exports, vendored dependencies, tests, and deployment/frontend documentation.

• forge build, forge test, and forge fmt --check passed.
• 32 tests passed, including 8,192 invariant calls.
• Clean offline build and tests passed.
• Slither’s 10 findings were triaged; Aderyn was unavailable.

Rewards use SEVEN. Adding stake resets that wallet’s entire seven-day lock.

See README and verification details. Manifest generation and independent launch review remain separate stage responsibilities.

The launch.json manifest is written and verified. Only that file is new in the working tree.

What the manifest declares

• Token: LaunchToken named SevenDay, symbol SEVEN, 18 decimals, no constructor arguments. These match the values set in src/LaunchToken.sol.
• Contracts: one application contract, StakingVault, with constructorArgs of $token. Its ABI export confirms a single nonpayable address parameter and no owner argument, so no $owner reference is used.
• Pool: native ETH as the zero address, fee 3000, tickSpacing 60, and the standard sqrtPriceX96 initial price.
• Notes: a plain-language description of the vault's lock, funding and accounting behavior, stating that the factory supplies the LP, MerkleDistributor and PoolInitializationGuard, and that supply split, policy and attestation belong to services.

Checks run

• forge build compiles cleanly and forge test passes all 32 tests across 4 suites, including the invariant suite.
• A scratch script in test/scratch/ validated the manifest against the canonical JSON Schema and its refinements: exact key sets, identifier pattern, reserved names, backward-only references, references only in address-typed arguments, initialPrice below 2^256, notes under 4000 characters, and constructor argument counts cross-checked against the exported ABIs.
• git status shows launch.json as the only change. No configuration, library, or source file was touched.

Created .imd-findings.json with no substantiated defects and coverage for all eight entry points.

All 32 tests passed offline, including 8,192 invariant calls. Exported ABIs match.

Limitations: launch.json is absent; service-dependent protected tests were not run. Source and configuration remain unchanged.

Created .imd-findings.json with no substantiated control-flow defects and coverage for all 8 entry points.

• All 32 tests passed, including 8,192 invariant calls.
• Both ABI exports match compiled source.
• Nine dependency checksum mismatches reflect whitespace-only differences.
• Manifest binding remains unreviewed because launch.json is absent; service-injected protected tests were not run.

Source and configuration files are unchanged.

The review is complete. The findings file is written and parses, with two findings and eleven coverage rows. No source, test, docs, or config files were changed. The only new files are the findings JSON and two scratch tests under test/scratch/, which is gitignored.

What I found

• Medium: a 1-wei top-up bypasses the active-period guard on restartRewards. The restart function refuses to reschedule queued idle rewards while a period is running, but fundRewards calls the same scheduler with no such guard and folds the whole queue into the seconds left in the period. A staker who is alone near the end of a period calls fundRewards(1) and collects a week of idle emissions in one minute. The scratch proof fails on current code: the attacker claims 604,800 SEVEN after staking 1 SEVEN for 60 seconds, where the intended behaviour caps them at 60 SEVEN. The proof source is embedded in the finding.
• Low: a donation near period end is streamed over the remaining seconds with no minimum-duration or deadline guard. A whale that stakes in front of the donor's transaction one second before the period ends takes about 99.99 percent of the donation. The README documents this as an accepted trade-off, so I reported it as low with the front-running amplifier and a small fix.

Coverage

• All eight entry points have rows. The token's three functions and the vault's stake, unstake and claim hold: no owner, no roles, caller-only effects, checkpoint before weight changes, exact-pull and checked reserve subtraction.
• fundRewards and restartRewards carry finding 1.
• The constructor hands out no role, and the ABI exports in docs/abi match the compiled ABIs exactly.
• The manifest row is marked unreached because no launch.json exists in the tree yet. The README's declared constructor binding, $token with no owner argument, is consistent with the source.

Static-analysis leads: the weak-PRNG, strict-equality and timestamp lines are accounting remainders, intentional zero checks and the requested time-based lock, not defects. The reentrancy-balance lead is covered by the reentrancy guard on every mutating entry point and the fee-free launch token.

The review is complete and the findings file is written. No tracked files were changed; the only artifacts are .imd-findings.json and the proof under test/scratch/.

Outcome: one substantiated finding, medium severity, with a failing Foundry proof. The rest of the Math Precision, Boundary and Numerical Gap checks hold.

Finding 1 (medium), src/StakingVault.sol:170, _schedule. The reward scheduler folds the idle queue into every reschedule and, during an active period, divides it by the seconds remaining rather than by the seven-day duration. At the deadline boundary this collapses the documented seven-day stream into one second. Concrete state: a donor funds 604,800 tokens with nobody staked, so the emissions queue up. One second before the period ends, an attacker stakes 1 wei and calls fundRewards(1). The whole queue is emitted in that final second, and the attacker claims 604,800 tokens plus 1 wei at the deadline. The expected outcome is at most one second of the existing stream plus the attacker's own wei, with the queue held for a seven-day restart. The proof at test/scratch/QueueDump.t.sol fails on the current code and passes on a scratch copy where the queue only joins a fresh period.

What I covered and found sound:

• Index arithmetic never truncates to zero and loses under one billionth of a wei per checkpoint. Per-user mulmod remainder tracking is exact. I probed 1 wei total stake with maximal rate and 7e26 total stake with a 1 wei per second rate.
• No overflow in any intermediate. All mulDiv inputs stay under 1e63 and elapsed-times-rate stays under 1e27.
• The accrual subtraction cannot underflow because every writer keeps lastUpdateTime at or below the applicable time.
• A top-up exactly at periodFinish starts a clean period with no division by zero. The lock check admits exactly unlockTime.
• Reserve accounting adds exactly the funded amount to obligations per reschedule. The existing invariant suite and my probes confirm principal is never paid as rewards.
• The exported ABIs in docs/abi are byte-identical to the compiled artifacts.

Static analysis leads: the weak-PRNG, balance-reentrancy, and strict-equality lines are modulo remainders, a guarded exact-delta check, and intentional zero checks. None reproduced as a defect.

Not reached: no launch.json exists in the tree yet, so the manifest row is marked unreached. The constructor shape in the README matches the token-then-vault dependency order the manifest will need.

Added 21 tests covering failure paths, boundaries, fuzz properties, and two stateful invariants.

• forge build --offline: passed.
• forge test --offline: 53 passed, 0 failed, 0 skipped.
• New invariants: 65,536 random calls, no unexpected reverts.
• Slither: 10 alerts reviewed; no confirmed defect found.
• Aderyn: unavailable.

All deliverables are under test/. Verification details.

This content was flagged for possible cybersecurity risk. If this seems wrong, try rephrasing your request. If you’re doing authorized security work that requires more cyber permissive safeguards, apply for Daybreak access via https://platform.openai.com/settings/organization/status-and-access before retrying.

Findings file is written, both snippets match the source lines, and the working tree shows no tracked changes (the findings file is excluded by the repo's own git config and test/scratch/ is gitignored).

Result: two findings kept, one medium and one low, written to .imd-findings.json with a 13-row coverage record.

Finding 1, medium, merged from audit_permissions and audit_math. Both specialists reported the same root cause at src/StakingVault.sol:170. During an active period, _schedule folds queuedRewards into the budget and divides it by the seconds left until periodFinish. restartRewards refuses that reschedule with ActiveRewardPeriod, but fundRewards reaches the same code with no guard and accepts 1 wei. A wallet holding 2 wei can stake 1 wei one second before the finish, donate 1 wei, and collect a full week of idle emissions in that second. I ran both specialist proofs and they fail for the stated reason. I attached my own proof instead, written in neutral wording, which fails on this code with:

604800000000000000000001 > 1000000000000000001

Severity stays medium: only donated rewards are misdirected, principal and the reserve accounting are intact, and it requires a period with a large idle queue. The description notes that the fuzz test at test/StakingVaultBoundaries.t.sol:207 pins the current fold and must change with the fix.

Finding 2, low, from audit_permissions. A donor's own late top-up is spread over the remaining seconds, so a same-block staker captures it. Reproduced with the exact numbers the specialist gave. Kept separate because it has a different mechanism and fix, and kept low because the README and integration doc document it as an accepted trade-off.

My own pass found nothing further. I traced all eight entry points, the lastUpdateTime bound that keeps rewardPerToken from underflowing, the reserve conservation invariant, and the per-wallet remainder arithmetic. The checked-in ABI exports are item-for-item equal to the compiled ones. The manifest is schema-valid and consistent with the ownerless vault's single address argument. The slither leads are all false positives for this code. The full suite passes offline from a clean environment, with 53 tests green.

Fixed idle rewards being folded into last-second top-ups. Added optional minimum-duration protection for donors; the original funding method retains its documented behavior.

Updated tests, ABI, documentation, and .imd-responses.json for both findings.

Validation passed: forge build, all 39 tests, the supplied proof, and forge fmt --check.

The manifest is revised and verified. Nothing else is pending.

Outcome: launch.json was already structurally correct against the accepted tree. The only change this round is one sentence in the notes.

What I checked

• The accepted StakingVault changed after my previous manifest commit, but its constructor is still a single address argument, so the contracts entry stays ["$token"].
• LaunchToken still sets name SevenDay, symbol SEVEN, 18 decimals, no constructor arguments. The token block is unchanged.
• The pool block keeps native ETH, fee 3000, tickSpacing 60 and the standard sqrtPriceX96 initial price.
• I validated the file against the canonical schema and its refinements with a local script. It passes, and the notes are well under the 4000-character limit.
• forge build and forge test pass. All 39 tests in the accepted suite succeed.

What changed

• The notes now describe the accepted source's second fundRewards overload with its minimum-duration guard, and the queued-reward reserve that anyone can restart after a period ends. The earlier wording predated those additions.

Scope: git status shows launch.json as the only modified file. No source, test, ABI, or configuration file was touched.

Extended tests under test/ for guarded funding, failed-call rollback, and queued-reward conservation, including randomized call sequences.

forge build --offline and forge test --offline passed: 64 tests, zero failures or skips. The extended vault invariant completed 32,768 calls without unexpected reverts.

No confirmed contract defects found.

Both earlier findings are settled as fixed, and I found no new defect. The findings file is written with an empty list and a full coverage record. No tracked file was changed.

Round-1 medium, idle queue folded into an active top-up: fixed. The scheduler now adds the queued rewards only when it starts a fresh seven-day period, and an active top-up spreads only the new donation plus the remaining scheduled budget over the remaining seconds. I reran the proof I attached last round and both specialist proofs on the revised code. My proof and the math specialist's proof pass. The permissions specialist's proof fails only on its own stale assertion that the rate must be the old inflated value, and its real check (the sole staker earns at most one minute of stream) passes once that line is dropped. A scratch test I wrote also confirms exact conservation through an idle period, a dust top-up, a restart shared by two stakers, and full claim and exit.

Round-1 low, late top-up emitted over the last seconds: settled. The author added an opt-in overload that takes a minimum remaining duration and reverts atomically before any transfer. I verified the revert moves no donor funds and leaves the queue untouched. The single-argument overload keeps the documented fixed-end behavior, which the README and manifest notes now state plainly. That residual behavior is an accepted trade-off, not a defect.

Other checks that hold:

• Full delivered suite passes, 66 tests including fuzz and invariant suites.
• The exported ABI files in docs/abi match the compiled ABIs exactly.
• The manifest is schema-valid, binds the vault's only constructor argument to the token, and names no owner since the vault is ownerless.
• Every static-analysis lead was a false positive: modulo for dust accounting, a guarded exact-balance pull, and intended timestamp-based accrual.

Coverage: all nine entry points are marked as holding, plus rows for the reserve and balance invariants, the manifest, and the analyzer leads.

Implemented source, static export, deployment manifest, and documentation. Build, typecheck, manifest verification, and all 29 browser checks pass. No live transactions were broadcast.

Validation report · Design documentation

The checkout’s .git is read-only. Delivery commit a486ac3 is available in the verified 3.82 MB Git bundle.