Temporal Parameter Staleness

TEMPORAL_PARAMETER_STALENESS Skill (Soroban)

Trigger Pattern: interval|period|duration|delay|cooldown|lock_period|timelock|unbonding|claim_delay|withdraw_delay|maturity|ledger_sequence|timestamp Inject Into: Breadth agents, depth-state-trace Finding prefix: [TPS-N] Rules referenced: R2, R8, R10, R13, R14

Cached parameters in multi-step operations become stale when authority changes them between steps. On Soroban, timing has unique properties: env.ledger().timestamp() provides Unix seconds (estimated, can lag wall clock), env.ledger().sequence_number() provides the exact ledger number (strictly monotonic), SCP consensus closes ledgers approximately every 5 seconds, and there is no concept of slots, epochs, or a clock sysvar. TTL (time-to-live) is measured in ledger numbers and can itself act as a timing mechanism.

Step 1: Enumerate Multi-Step Operations

Find all operations that span multiple transactions (ledgers):

For each multi-step operation:

• What parameters are read/cached at Step 1 (stored in Persistent/Instance storage)?
• What parameters are re-read at Step N?
• What parameters are used but NOT re-read at Step N?
• Which clock source is used? (env.ledger().timestamp() vs env.ledger().sequence_number())

Soroban Clock Semantics

Critical property: Within a single transaction (a single invocation tree, including sub-invocations via invoke_contract), ALL calls see the same env.ledger() values — there is no intra-transaction timing variation. Multi-step timing attacks require separate ledgers (separate transactions).

Timestamp vs sequence tradeoff:

• timestamp is human-intuitive (seconds) but can lag real time during validator slowdowns
• sequence_number is exact but requires converting to human time (~5s per ledger, subject to change if Stellar network parameters change)
• Hardcoded sequence_number durations break if Stellar's ledger close time changes (currently ~5s, configurable by validators)

Step 2: Identify Cached Parameters

For each parameter used across steps:

Soroban caching patterns:

• Instance storage cache: User's position struct stores a snapshot of config params at initiation time (e.g., position.fee_rate = config.fee_rate at deposit time)
• Persistent storage cache: Per-user Persistent entry stores the start ledger or timestamp (e.g., user_position.start_ledger = env.ledger().sequence_number())
• Re-read pattern: Step N re-reads from Instance storage directly — no staleness for that param
• Allowance TTL as cached permission: An allowance stored in Temporary storage expires by expiration_ledger — if the operation takes longer than expected, the allowance expires and Step N fails

Red flags: Parameter is cached in user's Persistent entry at Step 1 AND admin can change the source config AND Step N does NOT re-read from source.

Step 3: Model Staleness Impact

For each cached parameter that can become stale:

Scenario A: Parameter INCREASES between steps
1. User initiates at Step 1 — user Persistent entry stores param = X
2. Admin updates Instance config: param = X + delta
3. User completes at Step N — uses cached X from Persistent entry
4. Impact: {what happens with stale X when current config is X + delta}

Scenario B: Parameter DECREASES between steps
1. User initiates at Step 1 — user Persistent entry stores param = X
2. Admin updates Instance config: param = X - delta
3. User completes at Step N — uses cached X
4. Impact: {what happens with stale X when current config is X - delta}

BOTH directions are mandatory — increase and decrease often have different impacts.

Soroban-Specific Staleness Vectors

Step 3b: Update Source Audit

For each parameter updated from an external source (oracle contract, price feed):

• Is the source the correct representation of what this parameter tracks?
• Is the source contract address validated? Can a fake oracle be substituted via a user-supplied address?
• Should this parameter be fixed for a period (e.g., per epoch cycle) rather than continuously refreshed?
• Which functions update it? Which SHOULD? Any mismatch?
• Is there a refresh or crank function? Who calls it and when? What is the protocol state if it is never called?

Step 4: Retroactive Application Analysis

For fee/rate parameters that apply to existing state:

Soroban retroactive patterns:

• Instance config update: Admin changes fee_bps in Instance storage. All pending claims calculated at completion time using new rate — retroactively changes expected returns for users who initiated under old rate.
• Cooldown ledger update: Admin changes cooldown_ledgers in config. Users who initiated cooldown under old value may now need to wait longer or shorter — retroactive effect on in-flight positions.
• Rate-at-close pattern: If the protocol reads the current config rate at Step N (not the cached rate from Step 1), any admin change between steps applies retroactively to all pending operations.

Rule 2 direction check: Can the admin's parameter change make a user-facing function behave unexpectedly? (e.g., setting cooldown_ledgers = 0 removes withdrawal protection entirely; setting max_withdrawal = 0 blocks all withdrawals). Does the change retroactively affect users in active positions?