Skip to content

test(escrow-v2): document actual challenging vs refund race behaviour (bounty #1 race challenge/refund) - #9

Open
mjneuroart wants to merge 1 commit into
odiseo159-beep:mainfrom
mjneuroart:hermes/bounty-race
Open

mjneuroart wants to merge 1 commit into
odiseo159-beep:mainfrom
mjneuroart:hermes/bounty-race

Conversation

@mjneuroart

Copy link
Copy Markdown

Bounty #1 — Race between challenge and refund

Adds Hardhat tests reproducing the two signals documented in the challenge (bounty tests language: Spanish kept per conventions, titles/PR in English):

  • RACE_REAL: simultaneous challenge + refund in the same block -> double-ish effect (either refund twice or funds/reward split at same height).
  • GUARD_HOLDS: reentrancy guard correctly blocks the nested callback path, so the race is isolated to the challenge/refund ordering.

Run: npx hardhat test contracts/test/MergitEscrowV2.test.js
Both tests pass locally against the deployed MergitEscrowV2.

odiseo159-beep added a commit that referenced this pull request Oct 28, 2026
MergitEscrow.refund existed and nothing called it. When #9 and #10 expired
with four pull requests waiting, there was no way to close them: they would
have sat Open past their deadline forever, which reads as abandoned on the
explorer and on every builder record.

Dry-run by default. It checks the state, the caller and the deadline before
spending gas, so a refusal arrives as a sentence instead of a revert.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
odiseo159-beep added a commit that referenced this pull request Oct 28, 2026
Six of the first eight people who sent work could not be paid, and none of
them found out until after the merge. The preflight comments on every pull
request with the two things that decide it: a Bounty line pointing at a live
bounty this agent verifies, and an address the agent can resolve. One sticky
comment, rewritten on each edit.

It runs on pull_request_target because a fork's token cannot comment, and
because that is the one event that fires before a maintainer approves the
run, which is exactly when somebody needs telling. So it never checks out
the pull request's branch, takes the body through env instead of
interpolating it into a shell, gets no signing key, and holds only
pull-requests: write.

The watchdog is why #9 and #10 died. Nothing was wrong with the product; no
one watched the clock, and the contract will not settle a second past the
deadline. It now reports daily on bounties closing within 72 hours that
somebody is already working on, and on runs stuck waiting for approval. It
comments only when something is about to cost money, so the mention still
means something. On its first run it found ten runs waiting, the oldest
since 17 October.

Neither workflow holds a key or touches funds.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
odiseo159-beep added a commit that referenced this pull request Oct 28, 2026
A pull request named a bounty by its number, and reposting a bounty changes
that number. That one fact caused all of last week: four of eight wrote the
issue number because the issue is what they were looking at, and the four
who got it right went stale the moment #9 and #10 were refunded.

The number is an escrow detail. The issue is not. The agent now falls back
from the Bounty line to the issue the pull request closes, so PR #5 resolves
again as soon as the bounty behind issue 1 exists, without its author
touching anything.

Funding is open, so anyone can post a one-wei bounty pointing at our issue
with our agent as verifier. If that were enough to pick, it would be enough
to divert a payment. So when more than one live bounty fits an issue the
agent refuses and says which ones, and a repository can narrow it with a
funders list in mergit.json. There is no "take the first".

The explicit line still wins when it names a bounty that is still live and
still ours.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
odiseo159-beep added a commit that referenced this pull request Oct 28, 2026
On 25 October bounties #9 and #10 expired with four pull requests ready to
be paid. Nothing in the protocol failed; a GitHub button went unpressed. And
because settle refuses a second past the deadline, the only way out was
refund and repost, which changes the bounty number, which is the number
every one of those pull requests already carried. Eight people worked for
free over a date.

V3 is V2 plus extend(). The deadline only ever moves forward: letting a
funder shorten it would let them abandon somebody mid-work, so that is
refused for the funder too.

Two ways to move it. The funder, with no ceiling, because it is their money.
The verifier, up to the maxDeadline the funder set when posting. The second
is the one that matters: a watchdog can extend with the agent key that is
already in CI, and the funder key never has to go anywhere near a secret
store.

The cost of that, said plainly: a stolen verifier key can stall the funder's
refund up to the cap. It cannot take the money or redirect it, and the cap
is the funder's own choice. Setting maxDeadline equal to deadline turns the
whole path off.

15 tests, nearly all of them on what extend refuses to do. 58 passing.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
odiseo159-beep added a commit that referenced this pull request Nov 3, 2026
… that went missing

On 3 November the agent paid bounties #11 and #12 and crashed before the
receipts: `SITE` was declared at the bottom of on-event.mjs, and the loop that
pays runs at top level before that line. The money moved; the receipt did not.
It had been broken since a27d1c2 and nothing had been paid since.

Fixing that exposed a worse one that v1.2 introduced. The agent now only
accepts live bounties, so the second run every merge gets (the push, then CI
on main) found the bounty already closed and commented "Mergit did not pay
this pull request" on a pull request that had just been paid.

So a merged pull request is first checked against bounties that are already
settled, and one counts as its own only when the evidence hash recorded in the
settlement event is identical to the one the agent recomputes for this pull
request. A pull request that lost the race cannot match. When it matches and
there is no receipt, the agent writes it; when the receipt is there, nothing.

Also:
- a refusal on a merged pull request tells the maintainer how to retry once
  the bounty is funded, and is not posted twice;
- settle.mjs returns the status it reports, which was printing "undefined";
- verify.mjs and the record page leave `mergit/payable` out of CI and out of
  the evidence. That status is about to be written by the preflight for the
  maintainer, and it is not CI;
- the record page finds the paid pull request through the issue as well as
  the `Bounty: #N` line, and lets the evidence pick. Bounties #11 and #12 were
  paid to pull requests whose line still said #9 and #10.

Checked: the evidence recomputed with the new verify.mjs is identical to the
hash on chain for both payments, and the record page reproduces both, plus the
old bounty #3 through the line.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant