Skip to content

kona-node: verify delegated derivation parity against op-node #23912

Open

Description

@einar-oplabs

Part of ethereum-optimism/protocol-team#545.

The tracker's completion criteria include verification work that no fix PR covers. This issue tracks it so the parity claim rests on exercised behavior, not only on unit tests in the three core fixes (#23675, #23676, #23677).

Scope:

Activity

  1. self-assigned this
    on Nov 4, 2026
  2. einar-oplabs commented on Nov 5, 2026

    @einar-oplabs
    ContributorAuthor

    Progress: #23947 (draft) makes the three follow_l2 acceptance tests run against kona-node and closes #22272, #22273 and #22274. The sequencer and deriving verifier stay on op-node, and the follow-source verifier takes the env-selected CL.

    Local results: op-node 3/3 pass, kona-node develop 3/3 pass, kona-node develop + #23899/#23900/#23908 3/3 pass, with no MissingUnsafeL2Block, BlockNotSafe or critical errors in either kona run.

    develop passes too, so these tests do not exercise the paths the fixes change. Still open here:

    • source-ahead EL catch-up;
    • SYNCING then VALID to an unchanged target;
    • stale or conflicting finalized references;
    • resets below a pending finalized target;
    • hash and ancestry checks with unsafe production paused.

    Those need EL fault injection or engine-API control that the devstack does not have yet.

  3. einar-oplabs commented on Nov 9, 2026

    @einar-oplabs
    ContributorAuthor

    Draft: #24007, stacked on #24006. It covers the items left open after #23947.

    • End-to-end scenarios (delegated/scenarios.rs): the real delegate and engine actors run against a scripted delegate and a strict Engine API mock that checks every forkchoice update by hash. One scenario per refactor[contracts]: Port OVM_ProxyEOA to use ovm-solc #545 completion criterion:
      • safe and finalized progress by hash, with unsafe production paused
      • a delegated safe head on another branch reorging the EL
      • source-ahead catch-up, retrying an unchanged target through SYNCING until VALID
      • stale or conflicting finalized references: finalized never moves back and a conflicting hash is never accepted
      • a reset to genesis while finalized is missing
    • Acceptance test: TestFollowL2_LabelsMatchELByHashWithSequencerPaused compares kona-node's labels with its EL and with the sequencer's chain by hash. It passed locally under both kona-node and op-node.
    • Bug found: the delegate recorded the finalized block it forwarded, not the one the engine applied. It is fixed in kona-engine: prevent delegated finalized-head regression #23899, and a scenario covers it.

    Remaining gaps:

    • A real op-reth answering SYNCING or holding conflicting finalized hashes. The devstack has no Engine API fault injection, so these are covered only against the mock.
    • The reset race between the safe update and the finalize from one poll. The two actors' ordering can't be scripted from outside; finalization_target_survives_resets_below_it covers it at engine level.
  4. einar-oplabs commented on Nov 11, 2026

    @einar-oplabs
    ContributorAuthor

    Claude: Status of #24007 at head 6e50e73d8f (2026-11-11)

    Implemented

    • End-to-end scenarios run the real delegate and engine actors against a scripted delegate and a strict Engine API mock. They check every forkchoice update and lookup by hash, poll cadence, the published derivation origin, and quiet windows longer than the engine's longest retry.
    • The follow_l2 acceptance test pauses the sequencer once kona-node follows its source. It uses progress-bounded waits on both layers and compares labels by hash.

    Checks

    • 776 unit tests pass.
    • memory-all-kona-op-reth-fusaka failed once on TestConductorRejectsStaleMembershipVersion (conductor, RPC EOF). The stack does not touch the conductor, so this is unrelated; the job was rerun. The follow_l2 tests pass.

    Open findings (Fable 5.1 xhigh, round 5). All are test quality; there are no production bugs.

    • One origin assertion is tautological in two of three stale-finality variants.
    • The 4s poll cadence is no longer pinned by a literal.
    • The Go finalized wait's 4-minute stall timeout can fire before the 8-minute bound.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions