// Implement a new EIP for a hardfork under development in Erigon. Use when the user asks to implement, port, or wire up an EIP — covers spec lookup, dep analysis, prior-work check, implementation, lint, tests, and a wrap-up saved to `agentspecs/`.
Run the full Erigon test suite locally using GOGC=80 make test-all. Use this before marking a PR ready for review. Equivalent to the "All tests" CI workflow.
Run Erigon tests with Go race detector to find data races and concurrency bugs. Use this for concurrency-sensitive changes (parallel executor, p2p, txpool). Takes 30-60 minutes.
Run Erigon CI checks locally and/or trigger them remotely on a branch via GitHub Actions workflow_dispatch. Use this when you need to verify a branch passes all CI before or after pushing — especially for branches like bal-devnet-2 that don't auto-trigger on push/PR events.
Run benchmarkoor performance benchmarks against a locally-built Erigon binary and produce per-test MGas/s comparison tables. Covers image build, dataset reset, run invocation, result parsing, and before/after comparisons.
Run an ephemeral Erigon instance with a temporary datadir. Use this whenever the user wants to spin up a temporary, throwaway, or sandboxed Erigon node for quick testing, launch a second Erigon instance alongside an existing one, clone a datadir into a temp copy for safe experimentation, or find and clean up leftover ephemeral datadirs and processes from previous sessions. Handles port conflict detection and automatic port offsetting. Trigger on any mention of temporary/throwaway/ephemeral/disposable Erigon instances, running erigon briefly for testing or debugging, starting a second/additional erigon node, or cleaning up old temp erigon data.
Reference for all Erigon network ports. Use this when running multiple Erigon instances to avoid port conflicts. Lists every CLI flag that binds a port, its default value, and the protocol used.
| name | erigon-implement-eip |
| description | Implement a new EIP for a hardfork under development in Erigon. Use when the user asks to implement, port, or wire up an EIP — covers spec lookup, dep analysis, prior-work check, implementation, lint, tests, and a wrap-up saved to `agentspecs/`. |
| argument-hint | <fork> <eip-number> [devnet] |
| allowed-tools | Bash, Read, Write, Edit, Glob, Grep, WebFetch, Skill |
Fetch the specification from:
https://eips.ethereum.org/EIPS/eip-$1
Read the spec end-to-end and make sure you understand it. If there are any ambiguities or things that do not make sense in the EIP, bring these up and ask for clarifications. Once understood, map the EIP to the codebase — identify which packages, files, and existing constructs the EIP will touch.
Identify any other EIPs that EIP-$1 depends on or references. Each such EIP can be fetched using the same URL pattern as in Step 1, substituting its number:
https://eips.ethereum.org/EIPS/eip-<number>
Try to understand these additional EIPs and map them to the code. If there are any ambiguities or contradictions with the new EIP, make sure to raise these and ask questions.
Identify which hardforks contain the referenced EIPs, as well as any other EIPs already in the codebase that interplay with the new one. This provides more context and knowledge when analysing the code.
A list of forks and the EIPs they contain can be found at:
https://eips.ethereum.org/meta
They are organised as "meta EIPs" and usually start with Hardfork Meta followed by the name of the hardfork.
More recent forks use portmanteau naming to combine the execution-layer (EL) fork name and the consensus-layer (CL) fork name. EL forks are named after cities; CL forks are named after stars. For example, Glamsterdam is a portmanteau of Gloas (CL) and Amsterdam (EL).
Some EIPs only affect the EL, others only the CL, and some affect both. Erigon has both an EL implementation (under the execution package) and a CL implementation (under the cl package), so all EIPs apply to one or both layers in Erigon.
In some cases fork meta EIPs list which EIPs got included; in other cases they provide a summary of the changes instead of an EIP list.
The meta EIP for the current fork under development usually has these lists:
If a devnet name was provided as the third argument ($2), fetch its specification from:
https://notes.ethereum.org/@ethpandaops/$2
Devnets are short-lived test networks where changes for a hardfork under development are validated. Core developers iterate through a series of devnets, progressively adding EIPs and refining their specs, until all hardfork changes are ready for release.
A devnet spec typically lists:
Use the spec to pinpoint which changes for EIP-$1 are in scope for this devnet and how they differ from prior devnet iterations. This is especially valuable when an earlier version of the EIP was already implemented for a previous devnet — the delta between devnets tells you exactly what needs updating (and ties directly into Step 6 below).
The devnet argument is optional. If none was provided, skip this step — this is expected when an EIP is being implemented before any devnet spec exists for it.
Reference the Ethereum Yellow Paper for deeper information about the Ethereum protocol. It can be found at:
https://ethereum.github.io/yellowpaper/paper.pdf
Note however that the current version of it covers up to and including the Shanghai hard fork. Changes to the protocol after Shanghai can be inferred via the meta EIPs for the subsequent forks after Shanghai.
The yellow paper provides deep base knowledge that can be used to confirm correctness up to Shanghai, and later for parts that haven't been changed by subsequent hard forks. For parts that have been changed in subsequent hard forks, the corresponding EIP specs from those hardforks — read in chronological order — become the main source of truth.
Check whether any previous work for this EIP has already been done in the codebase. This may happen if we've previously implemented this EIP but against an older / outdated spec for a previous devnet.
In this case, analyse the current code and compare it to the latest specs. Implement the necessary changes to address any discrepancies that you find.
Implement the EIP changes in the code, based on the understanding gathered in Steps 1–6 and the codebase mapping you produced. Touch the packages, files, and existing constructs identified during that mapping.
If during implementation you discover something that is unclear or appears to contradict the spec, stop and ask for clarification rather than guessing.
After making code changes, run make lint and fix any reported issues. The linter is non-deterministic — run it repeatedly until it is clean.
Run local tests using the /erigon-test-all skill. Analyse and fix any failures in the implementation (see also "Question the tests — do not silently fix them" below). Keep iterating until all tests are fixed.
The most important tests when implementing a new EIP for the EL are the EEST spec test shards, exercised by the cmd/evm runners (statetest, blocktest, enginextest) via the Makefile targets:
make eest-spec-statetests-stable / …-devnet — state-tests against the stable/devnet EEST fixturesmake eest-spec-blocktests-stable-sequential / …-devnet — blockchain-tests against the stable/devnet EEST fixtures. The devnet shard always runs under ERIGON_EXEC3_PARALLEL=true (the in-development hardfork requires it); the …-sequential shard pins ERIGON_EXEC3_PARALLEL=false.make eest-spec-blocktests-stable-parallel — same fixtures as …-stable-sequential but with ERIGON_EXEC3_PARALLEL=true; useful for catching parallel-only regressions on stable fixtures.make eest-spec-enginextests-stable-sequential — engine-x tests against the stable EEST fixtures with ERIGON_EXEC3_PARALLEL=false. No devnet variant: the devnet tarball doesn't yet ship blockchain_tests_engine_x/.make eest-spec-enginextests-stable-parallel — same fixtures as …-stable-sequential but with ERIGON_EXEC3_PARALLEL=true; useful for catching parallel-only regressions on engine-x stable fixtures.make eest-spec-enginextests-benchmark-{1m,5m,10m,30m,60m,100m,150m}-{sequential,parallel} — engine-x tests against the per-gas-target benchmark fixtures, with --time per-test stats. Each gas target has a -sequential (ERIGON_EXEC3_PARALLEL=false) and -parallel (ERIGON_EXEC3_PARALLEL=true) variant.make eest-spec-blocktests-stable-race-{pre-cancun,cancun,prague,osaka}-{sequential,parallel} and make eest-spec-blocktests-devnet-race-amsterdam — race-detector variants split by fork. Each stable-race sub-shard has a -sequential / -parallel pair; the -parallel siblings exercise parallel exec3 under the race detector. The blocktests-devnet-race-amsterdam shard is always parallel (matches the non-race devnet behaviour).The shard list / failure budgets / exec3-parallel flags are defined in tools/eest-spec-shards.yml (single source of truth shared with the CI workflow and the local runner script). See EEST_SPEC_SHARDS / EEST_SPEC_RACE_SHARDS in the root Makefile for the partition into non-race vs race targets.
Pitfall: stale evm / evm.race binary. When iterating on an EIP implementation, always invoke shards via make eest-spec-<shard> rather than bash tools/run-eest-spec-test.sh <shard> — the make target lists evm (or evm.race) as a prereq and go build is cache-aware, so a fresh binary is built before each run. The script invoked directly bypasses that rebuild, so the runners exercise whatever build/bin/evm{,.race} happens to be on disk against current fixtures — silently inflating failures (e.g. devnet shards "regressing" by thousands of tests) or hiding regressions when comparing budgets before/after a change.
For an EIP on a hardfork under development, the -devnet shards are the primary signal; the -stable shards (and -benchmark-* shards, where applicable) are regression checks against prior hardforks.
These protocol tests are defined by JSON test fixture files inside release tarballs (eest_stable.tar.gz, eest_devnet.tar.gz, eest_benchmark.tar.gz) pinned by test-fixtures.json and lazy-downloaded + extracted by make test-fixtures-eest (which the eest-spec-* Makefile targets depend on) into test-fixtures-cache/<key>/. The cmd/evm runners walk the extracted JSON files directly.
These test fixture files are generated by the Ethereum Python spec reference implementation, which can be found at:
https://github.com/ethereum/execution-specs
The code version which generates the test fixtures for a given hard fork lives on a branch following the forks/<fork_name> naming convention. For example, if the fork name is osaka then the branch name will be forks/osaka. The branch for the user-provided hard fork under development is forks/$0.
Sometimes new EIPs under development may not be in the forks/$0 branch yet — they may be on a separate feature branch. To find it, look for a branch name that contains the provided EIP number $1. Once you find a branch (or several branches), ask the user to confirm which one to use.
The point of identifying the correct branch is to read and analyse the right reference code — i.e. the version of the Python spec implementation that corresponds to the fixtures we are testing against. Nothing in the local repo needs to be updated; this is purely about which code version the agent downloads / reads for cross-referencing while debugging or validating behaviour.
It is VERY IMPORTANT to question the tests and validate they are doing the correct thing as per the EIP specifications. This applies to both EL spec tests (driven by the tarballs pinned in test-fixtures.json) and CL spec tests (driven by cl/spectest — see "Running CL spec tests" below).
Sometimes some of the protocol spec tests may have bugs or mistakes in them due to inaccurate implementation in the upstream reference implementations (the Python execution-specs for the EL, the consensus-spec-tests for the CL).
Or some of our other local tests (outside of the ones driven by the EEST tarballs pinned in test-fixtures.json, the execution/tests/legacy-tests submodule, or the cl/spectest fixtures) may have gaps in them, test for the wrong thing, or test for logic from before the new EIP — in which case the test should be amended to be explicitly for the correct corresponding previous hardfork, and a new test should be written for the new hardfork which this EIP will go into, covering both the old logic (prior to this EIP) and the new logic (after it).
When you run into such situations do not attempt to fix those tests. Instead write a summary of your findings and ask for them to be reviewed and new guidance to be given.
If the EIP affects the CL (or has any CL-side touchpoints), also run the consensus-spec tests under cl/spectest. These mirror what the Consensus spec GitHub workflow (.github/workflows/test-integration-caplin.yml) runs.
The flow is:
cd cl/spectest
make tests # downloads the consensus-spec-tests fixture tarball (via tools/test-fixtures.sh) and unpacks it
make mainnet # runs the full mainnet CL spec test suite
For faster iteration when only one fork is relevant, the cl/spectest/Makefile exposes per-fork targets such as make capella, make deneb, make electra, make fulu, etc. Use the one matching the fork the EIP belongs to. If no per-fork target exists yet for the new fork, add one or run CGO_CFLAGS=-D__BLST_PORTABLE__ go test -run=/mainnet/<fork>/ -v --timeout 30m from cl/spectest (the CGO_CFLAGS flag mirrors the Makefile and is required for the BLS portable build).
If make tests fails (e.g. fixture download issue), the CI workflow treats it as a soft skip — but locally you should investigate and fix it before relying on the CL test results.
Once implementation is complete and tests are passing, produce a wrap-up summary of all the work done. The summary should cover:
eest-spec-* shards now pass (and CL cl/spectest results, if applicable), and any that were skipped or remain openInclude the summary in your final reply, and also save it as a markdown file at:
agentspecs/eip-$1-implementation/summary.md
Create the agentspecs/eip-$1-implementation/ directory if it does not already exist. The agentspecs/ directory is gitignored, so these notes stay local.