Grantkee Claude Extensions Tn Domain Epoch

tn-domain-epoch

This skill exists because epoch boundaries are the single most dangerous part of the protocol. Every validator must agree, byte-for-byte, on what state was active at the boundary. A reader that consults the wrong source — canonical tip vs. closing-epoch snapshot, governance-mutable contract vs. pinned epoch record — will produce a different block from its peers, and that's a chain split.

If you are about to modify code that:

• lives under crates/node/src/manager/** (EpochManager, RunEpochMode, epoch task spawning)
• touches GasAccumulator, WorkerFeeConfig, base-fee carryover, or per-epoch reward counters
• reads epoch_state, epoch_info, EpochState, or anything derived from concludeEpoch
• spawns or shuts down per-epoch tasks (vote collector, primary/worker network handles)
• handles consensus-output replay on startup, mode change, or epoch rollover

…load this skill before writing a single line.

Why epoch boundaries are different

Within an epoch, parameters are stable — committee, worker configs, stake versions, and fee-strategy choices are fixed. Mid-epoch governance calls may mutate the underlying contract storage, but those mutations have no effect on the running epoch. They take effect only after the closing block applies concludeEpoch(...).

Therefore, two distinct snapshots exist at every boundary:

1. Closing-epoch final state — the state root produced by the last block of epoch N. Contains the parameters all validators used during epoch N. This is the canonical reference for anything carried forward (base fees, leader counts, gas totals).
2. Opening-epoch initial state — the state root that begins epoch N+1. Contains the new committee and any governance updates applied at close.

Reading the current contract storage (canonical tip) instead of one of these two snapshots is the most common cause of validator divergence.

Invariants

These are hard rules. Violating any of them is a Critical-severity bug.

1. WorkerConfig at the boundary comes from the closing epoch's final block, never from canonical tip. Governance can rewrite the WorkerConfigs contract mid-epoch, but the values that applied during epoch N are pinned in the state at epoch_info.blockHeight + epoch_info.numBlocks - 1 (or equivalently: the last finalized block whose nonce decodes to epoch == N). Reading from canonical tip leaks future governance into past computation.

2. Execution state lags consensus. reth_env.finalized_header() may return None or an older block than the consensus DB knows about. Never assume a header at a specific number exists during catchup — it may not have been executed yet. If your code falls back to a default when the header is missing, that default must match what every other validator would compute, or you will fork.

3. Per-epoch state must be reset at RunEpochMode::NewEpoch, not at ModeChange. ModeChange keeps the epoch alive (e.g., switching observer ↔ active validator). Resetting things like GasAccumulator leader counts in ModeChange corrupts in-flight epoch totals.

4. The replay window is [last_executed_consensus_number + 1, last_committed_consensus_number]. Code that catches up from a stored consensus header must not re-execute headers below last_executed_consensus_number (double-execution) and must not skip headers above (state divergence). The replay path is the only legitimate way to advance execution past a saved-but-unexecuted consensus output.

5. Network handles, task managers, and watch channels are epoch-scoped — but the underlying networks are node-scoped. The ConsensusNetwork runs for the lifetime of the node; the handles into it (per-epoch PrimaryNetworkHandle) get rebuilt every epoch with the new committee. Don't store epoch-scoped handles in node-lifetime structs, and don't tear down node-lifetime networks on epoch transition.

Pre-write Checklist

Answer each question before writing the change. Write your answers in your scratch context — they don't need to ship in code, but skipping the exercise is how the regression bug got merged.

1. Which epoch's data am I reading? Name it: closing epoch N, opening epoch N+1, or "stable mid-epoch". If the answer is "the current value of the contract", you are probably wrong.

2. What is the canonical source for this value? Cross-reference references/canonical-sources.md. If the value isn't in the table, add it before writing the code.

3. Is the header I'm reading guaranteed to exist on every node at this point in the pipeline? During catchup_accumulator, replay_missed_consensus, or any startup path: probably not. What's the fallback, and would every other validator compute the same fallback?

4. Does my fallback diverge? If you write .unwrap_or(SOME_DEFAULT) for a value that participates in block production or state transition, the default must be deterministically correct or the code path must be unreachable. MIN_PROTOCOL_BASE_FEE as a fallback for an EIP-1559 base fee is the canonical example of a forking default.

5. Will governance ever change this value during the epoch? If yes, you must read it from a pinned epoch boundary, not canonical tip. WorkerConfig, StakeConfig, committee membership, and issuance schedule all fall into this bucket.

6. What state does this leave behind for the next epoch? GasAccumulator totals, base-fee carryover, leader counts — what does the next epoch need, and where am I writing it so the next-epoch startup can find it?

Canonical Sources

Quick lookup for "where do I read X from?". Full version with file pointers in references/canonical-sources.md.

Common Bug Patterns

Pattern 1: Tip-instead-of-snapshot

The agent reads a governance-mutable value from canonical state when it needs the boundary snapshot. This was the catchup_accumulator regression:

// WRONG — get_worker_fee_configs reads from canonical tip, which reflects
// post-epoch governance updates. Two validators with different sync states
// will see different configs and produce different blocks.
let configs = reth_env.get_worker_fee_configs(num_workers)?;
for (worker_id, config) in configs.iter().enumerate() {
    let base_fee = match config {
        WorkerFeeConfig::Eip1559 { .. } => {
            // WRONG — epoch_info.blockHeight is the START of the new epoch.
            // That block may not be executed yet, falling back to MIN_PROTOCOL_BASE_FEE,
            // which forks against caught-up nodes that read the real value.
            reth_env.header_by_number(epoch_state.epoch_info.blockHeight)?
                .and_then(|h| h.base_fee_per_gas)
                .unwrap_or(MIN_PROTOCOL_BASE_FEE)
        }
        ...
    };
}

The fix uses one snapshot for everything: read the closing-epoch's final block once, then derive both the worker configs and the base-fee carryover from that pinned state. The closing-epoch final block is the only state root that all nodes — caught-up or catching up — agree on after consensus has committed the epoch.

Pattern 2: Forking-default fallback

Any .unwrap_or(...) on a value that feeds block production or state transition is a forking risk unless every validator would compute the same fallback under the same conditions. Common offenders: unwrap_or(MIN_PROTOCOL_BASE_FEE), unwrap_or_default() on a U256 fee, unwrap_or(0) on a leader count.

The right move is almost never a fallback. It's either: (a) the data is guaranteed to exist at this point in the pipeline (prove it and expect() with a clear panic message), or (b) the data may not exist, in which case the surrounding operation should be deferred or skipped, not faked with a default.

Pattern 3: ModeChange resetting NewEpoch state

Code that resets per-epoch counters runs in the wrong RunEpochMode arm:

// WRONG — ModeChange means we're switching observer/active mid-epoch.
// The epoch is still alive. Resetting GasAccumulator destroys in-flight totals.
match run_epoch_mode {
    RunEpochMode::ModeChange | RunEpochMode::NewEpoch => {
        gas_accumulator.reset();
    }
    _ => {}
}

ModeChange and NewEpoch look symmetric but have opposite meanings for epoch-scoped state. If you're touching the RunEpochMode match, list every per-epoch resource and explicitly decide which arm owns its lifecycle.

Further Reading

• references/invariants.md — every invariant with a code-pointer to where it must hold
• references/bug-patterns.md — historical bugs caught by this skill, with diffs
• references/canonical-sources.md — full value-to-source lookup with file paths and types

Source: grantkee/claude-extensions — distributed by TomeVault.