row 237 candidate on 816d48b03 (gate tree)

This commit is contained in:
sage committed 2026-10-20 20:27:37 -05:00
1 parent 816d48b03e
commit 23529aab2a
4 files changed
+123

No files matched your search

+75
View File
@@ -0,0 +1,75 @@
# Row 237: one decision publishes once (#1743)
Megavolt, 2026-10-21 (from `date -u`). Brief: `docs/plans/2026-10-20_agent-credentials.md`
§ "row 199 follow-up N1", blob dca67d310e1b. Source: Duckworth's row 199 round 1 note N1
(duckworth-0068). Base: `cb5fdb064` (row 199's landing). Reviewer Duckworth, gate owner Sage.
## The defect
`checkPublish` opens a ticket after checking the decision is unspent. The decision is spent only at
`recordPublish`, which runs after the bus process has pushed. Two `checkPublish` calls on one decision
therefore both get tickets, and both push. With different branches, the commit lands on two fork refs:
`refs/heads/<user>/feat` and `refs/heads/<user>/other`. The second `recordPublish` then refuses
`decision-consumed`, but by then its push has already reached the fork.
## The fix, and why this option
`checkPublish` refuses `publish-busy` when an open ticket in `#publishes` holds the same business and
decision (`packages/bus/src/broker.mjs:1038-1040`). It runs after the authority and fact checks, so a
spent decision still refuses `decision-consumed` first.
The brief's other option, rechecking `#spent` in `publishTarget`, doesn't close the race. Both tickets
reach `publishTarget` before either push ends, and nothing is spent until `recordPublish`. So both
rechecks pass and both pushes run. The only state that exists while the first push runs is the open
ticket, so the check reads that.
A refusal (`publishRefused`) drops the ticket and leaves the decision unspent, so a retry on the same
decision gets a new ticket. That is row 199's behaviour, unchanged. A within-role publish has no
decision and isn't affected.
## Tests
- `packages/bus/tests/publish.test.mjs:186`. A second ticket refuses `publish-busy`, on the same branch
and on another. After `publishRefused` the retry gets a ticket and a third refuses busy. After
`recordPublish` there is one `action.allowed` for git.push.remote and the next call refuses
`decision-consumed`. Another agent's decision on the same task isn't held.
- `packages/cli/tests/publish-socket.test.mjs:269`. End to end through the git socket, the bus runtime
and the TLS fake Gitea. While the first publish holds its ticket, a second publish on the same
decision to branch `other` gets `publish-busy`. The fake Gitea gets exactly one `git-receive-pack`,
the fork has only `refs/heads/coder/feat`, and there is one `action.allowed`.
Runs on this branch, single files:
- publish.test.mjs: 13/13.
- publish-socket.test.mjs: 8/8.
## Mutant
M1 removes the `publish-busy` line. Both tests fail:
- The bus test fails at the first busy assert.
- The cli test fails because the second publish answers `ok: true` with `ref: refs/heads/coder/other`.
That is the defect: two refs on the fork.
Logs are in `/mnt/storage/scratch/megavolt-r237/` (m1-bus.log, m1b-cli.log). `cmp` confirmed
broker.mjs matched the saved copy after the restore. The first cli mutant run (m1-cli.log) was against
a hook that recursed on every checkPublish. I fixed the hook to fire only for branch `feat` and reran
it (m1b-cli.log).
## Residuals, not fixed here
1. **A ticket that never ends holds its decision until the bus restarts.** `#publishes` is in memory.
That happens when the host dies between checkPublish and recordPublish or publishRefused. It fails
closed (`publish-busy`), and a restart clears it.
2. **publish-timeout (sage-1440).** The socket files `publishRefused` and drops the ticket, but the bus
process's push may still be running. A retry on the same decision can then open a ticket while that
push is live. This is the same residual the socket already logs ("the bus process's push may still
reach the fork"). Closing it means the runtime's `remotes` must keep the decision busy until its push
settles, which reaches into runtime.mjs. That is beyond the brief's two options.
3. **checkLand has the same shape.** `#landings` takes a ticket per call and the decision is spent at
`recordLand`. Whether two landings on one decision can both reach main isn't checked here. That's a
follow-up for Sage to row or dismiss.
## Gate
Per the brief: Duckworth approves the manifest, the bus suite and every `scripts/test-*.sh` are green
on Sage's gate, and M1 is killed. This row also changes a cli test file, so the cli suite (or that file
alone) should run on the gate too.