dual-replay

Dual Execution Replay

Instruments the latest execution path to run every transaction twice — once with a base executor (built from an execution-layer cut taken at <base-sha>) and once with the tip executor at <tip-sha> — and diffs the two outputs. The base result is committed; the tip result is for comparison only.

Most of the work is handled by scripts/dual_replay/. This skill is the contract for invoking that script and the playbook for when its happy path fails.

Usage

/dual-replay <base-sha> <tip-sha> [diff-mode] [cut-name]

• diff-mode (optional, default strict): one of strict, status-only, gas-only, status-and-gas.
• cut-name (optional, default replay_cut).

Additional knobs (pass through if the user supplied them):

• --gas-tolerance <percent>: tolerated gas delta as a percentage (default 0, i.e. exact equality). Only meaningful for diff modes that compare gas.
• --output-dir <path>: where effects-diff artifacts go (default target/dual-replay/).
• --timings-file <path>: where per-tx execution timings are appended as CSV (default <output-dir>/timings.csv).

Timings output

Every transaction's base and tip execution is timed and appended as one CSV row to --timings-file (default <output-dir>/timings.csv). Columns: digest,base_ns,tip_ns,base_gas,tip_gas,status_match. Writes are buffered and flushed every 500 rows. The file is opened in append mode, so resumed runs accumulate into the same file — delete it between runs if a clean dataset is needed.

If either SHA is missing or does not resolve to a commit, ask the user for clarification. Do not prompt for other arguments — fall back to defaults.

Happy path

Run the script's run subcommand. It does everything end-to-end and creates a commit on the current branch:

python3 scripts/dual_replay.py run \
  --base <base-sha> --tip <tip-sha> \
  --cut-name <cut-name> \
  --diff-mode <diff-mode> \
  [--gas-tolerance <percent>] \
  [--output-dir <path>]

On success, report the commit SHA produced (git rev-parse HEAD) and stop. Do not open a PR.

The script never prompts. If you need to override the dirty-tree refusal because the user explicitly said so, pass --force-dirty.

Recovery playbook

If run exits non-zero, the script prints which subcommand failed. Drive the remaining steps individually — they read persisted state from .dual-replay.json so you don't need to re-pass flags:

python3 scripts/dual_replay.py cut    --base <sha> --tip <sha> --cut-name <name>
python3 scripts/dual_replay.py inject
python3 scripts/dual_replay.py build
python3 scripts/dual_replay.py commit

cut failed

Usually environmental: dirty tree, unresolvable SHA, or scripts/execution_layer.py cut itself errored. Read the script's stderr, surface it to the user, and stop. Do not attempt code edits — the working tree may be on an unexpected commit.

inject failed (anchor not found)

The script reports the file and the regex it tried. The tip has likely renamed or moved a symbol since the templates were written. Do not refresh the templates speculatively. Instead:

1. Read the named file and locate the equivalent symbol at tip.
2. Apply the dual-exec changes manually for this run, following the shape in the TPL_* template constants in scripts/dual_replay.py. Substitute ${cut_pkg} → the underscored cut name (e.g. replay_cut → sui_adapter_replay_cut).
3. Skip the inject step (state is preserved) and continue: python3 scripts/dual_replay.py build, then commit.
4. If the same anchor breaks repeatedly across invocations, the regex in scripts/dual_replay.py should be updated — but only after confirming the new symbol name is stable.

build failed

cargo check -p sui-execution -p sui-types failed. The script already handles the common scaffolding (Clone derives on the gas types it knows about, the serde_json dep on sui-execution). What's left is genuinely case-by-case. Read the captured stderr and apply the smallest fix that compiles:

• Generated cut references an API that no longer exists at tip. Apply the smallest compatibility patch. Prefer adapting code inside the cut (it's a snapshot, you're free to adjust it) over changing shared code at tip. Sometimes restoring a tiny helper in shared code is cleaner — judge per case.
• Cascading Clone derives. The script derives Clone on SuiGasStatus (gas.rs), gas_v2::SuiGasStatus, gas_v2::SuiCostTable, gas_v2::ComputationBucket, and tables::GasStatus. If a new field type in any of these (or a transitively-reached type) is not Clone, add #[derive(Clone)] to it — and if this becomes a recurring fix, add the type to GAS_CLONE_TARGETS in scripts/dual_replay.py so the script handles it directly.
• Missing gas-model setter. If the template calls set_gas_model_version on SuiGasStatus and the method doesn't exist, add a narrow one. Keep it minimal.
• Cut crates fail formatting/metadata checks because they think they're part of the nested Move workspace. Update external-crates/move/Cargo.toml so the generated cut crates are either members or explicitly excluded — keep the metadata consistent. Don't skip formatting.

After each fix, re-run python3 scripts/dual_replay.py build. When it passes, run commit.

commit failed

Almost always pre-commit hook failure. Read the hook output, fix the cited issue (often cargo fmt or cargo xclippy complaints), and re-run commit. Never use --no-verify.

Constraints

• Do not modify scope beyond what the dual-replay flow requires. If you notice unrelated issues, mention them but do not fix them in this commit.
• Do not reset or clean the working tree to "tidy up" after the cut. The generated cut crates must survive the checkout from base to tip; the script uses git checkout -m for exactly this reason.
• Do not double-run or diff non-normal execution paths (dev-inspect, genesis state update, verifier, layout resolver) unless the user explicitly asks for it. If a compatibility patch routes one of those paths through the cut just to compile, that's fine — note it.
• The final commit message should be: instrument dual execution replay from <base-short-sha> to <tip-short-sha>. The script produces this automatically.
• Do not push. Do not open a PR. The bot's caller does that separately.