[spalen0]

Problem

yChad nonce 2299 was rated HIGH by our AI explainer (gist), while Govlens rated it MEDIUM (gist). The main driver: the governance Executor's execute(bytes) _script is a packed script (target(160) | size(96) header + calldata per call, per the verified Executor.vy), not ABI calldata. try_decode_inner_calldata can't read it, so it went to the prompt as raw hex. The model sliced it one hex digit off, reported a nonexistent selector 0x4759be10, and treated the six calls as "unresolved downstream effects". That was a main stated reason for HIGH.

In fact the script makes six accept_management() calls: the Executor accepts management of five reward distributors and the Funding Distributor.

Changes

• utils/calldata/wrappers.py: split_packed_script mirrors Executor.execute (requires exact consumption, a non-zero target, and ≤64 calls). execute(bytes) is registered as an executing wrapper, so proxy-upgrade detection and protocol context see the inner calls.
• utils/calldata/decoder.py: DecodedCall.script_calls holds each inner call, decoded against its own target's ABI when the chain is known. Without a chain id, only selectors resolvable offline are decoded.
• utils/llm/ai_explainer.py / utils/llm/report.py: the prompt and the gist Call Flow list each script call with its target. Script targets feed labels, address links, references and risk anchors.
• System prompt: never derive selectors or targets by slicing raw hex. Undecoded payloads are stated as undecoded.
• known_selectors.py: Executor execute / set_operator and set_management / accept_management.

Rendered prompt for the nonce 2299 script:

Call 1: execute(bytes)
    bytes _script: ↳ Executor script, 6 calls; each is made by the Executor, not the sender
      Script call 1 → 0x95547eDe56cF74B73dd78a37f547127dffDA6113: accept_management()
      ...
      Script call 6 → 0xbCc932e4750C3E465A7E54A06A34F9EdF8f6116b (Funding Distributor): accept_management()

Not in scope

• Govlens also used off-chain governance context: the on-chain proposal that bundled these management transfers failed (602 of 1,306 YFI in favour). We still don't fetch proposal status. That would be a follow-up in protocol_context.py.
• The simulation prompt still caps state changes and events at 10.

Testing

• New tests/test_executor_script.py (14 tests): split / malformed scripts, unwrap, decoder with ABI and offline paths, prompt and report rendering, label and reference collection. Uses the real nonce 2299 script.
• Full suite: 1939 passed, 4 skipped. ruff check / ruff format clean.
• Not run end-to-end against the live tx (no API keys in this environment).

🤖 Generated with Claude Code