Files
stack/agents/darkwing/work/queue-55/FOLLOWUP-transport-code.md
T
jason.woltjeandClaude Opus 5.5 fa6d889cb7 queue-55: round 2 candidate packet (darkwing)
Answers Filbert's comment 27251 and Dewey's comment 27253: R1 type-ahead
window, R2 terminal cell widths, B1 cut inside the question, 30 mutants
killed, N1-N8. Candidate manifest sha256 3f2d8f8e. Gate on c506c6ec:
cli 106/106 and every scripts/test-*.sh exit 0. The feature stays
uncommitted.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
2026-10-11 13:39:05 -05:00

63 lines
2.7 KiB
Markdown

# Follow-up: the transport's "did not finish" error has no code
Found while building row 55 (#1544). It is not part of row 55. Sage will
queue it with the 2026-10-12 batch. The code was already like this at
`fe5c68fa`.
## What happens
`humanTransport` in `packages/cli/src/transport.mjs` (line 60 at
`fe5c68fa`) throws this when the bus CLI child errors or times out (30 s)
and no exit status comes back:
```js
throw new CliError("outcome-unknown: the bus transport did not finish", 1);
```
Every other error it throws sets `error.code`. This one doesn't.
`resolveDecision` in `packages/cli/src/cli.mjs` checks
`e.code === "outcome-unknown"` to choose its sentence:
```
outcome unknown: the broker may have recorded it. Check mosaic inbox or mosaic trail <id> before trying again
```
That check misses the transport error. So if `decision.resolve` times out,
`mosaic decide` prints `mosaic: outcome-unknown: the bus transport did not
finish` and exits 1. The exit code is right. The text is wrong: it doesn't
tell the user the decision may already be recorded, or where to look. A
retry could resolve a decision that already went through. The bus should
refuse a second resolve, but the user hasn't been told to check first.
The row 55 app is not affected. It matches the message text
(`/^outcome[ -]unknown/`) and shows its outcome-unknown overlay, which has
no retry.
No other check in `packages/cli/src` compares `.code` against
`outcome-unknown`. The launcher and notifier errors have their own codes.
## Suggested fix
1. Set `error.code = "outcome-unknown"` on the transport's did-not-finish
error, the same way the other two throws do.
2. Add a test in `packages/cli/tests/cli.test.mjs`: a transport whose
`decision.resolve` throws that error makes `mosaic decide` print the
"may have recorded it" sentence, exit 1.
3. Extend the `hang` case in `packages/cli/tests/transport.test.mjs` to
check `e.code === "outcome-unknown"`.
Files: `packages/cli/src/transport.mjs`, `packages/cli/tests/cli.test.mjs`,
`packages/cli/tests/transport.test.mjs`.
## A related case: a timed-out inbox read (Filbert N5, row 55 round 1)
`resolveDecision` reads the inbox before it sends anything. If that read
hits the same did-not-finish error, the app shows its outcome-unknown
overlay ("the broker may have recorded it"), though no choice was sent.
Once the transport sets `error.code = "outcome-unknown"`, `mosaic decide`
would print the same sentence for that read. The fix should tell the two
apart: only an error from `decision.resolve` is outcome unknown. A failed
inbox read before it is a plain failure, exit 1, with nothing sent. A test
for each, in `packages/cli/tests/cli.test.mjs` and
`packages/cli/tests/app.test.mjs`.