Skip to content

fix(steer): the GENERATING-lane warning named a mode it could never report - #122

Merged
Dygreens merged 1 commit into
mainfrom
fix/generating-lane-warn-names-the-actual-mode
Aug 18, 2026
Merged

fix(steer): the GENERATING-lane warning named a mode it could never report#122
Dygreens merged 1 commit into
mainfrom
fix/generating-lane-warn-names-the-actual-mode

Conversation

@Dygreens

Copy link
Copy Markdown
Collaborator

A single warn! whose text asserted mode=type here is a regression while the structured field carried the mode that actually happened.

25 of 25 records, all-time, are mode="paste" — the mode the same sentence calls correct.

And that isn't luck. must_paste is:

generating || chars > 400 || picker_shaped

so inside if generating { … } the paste branch is taken for every input. The condition the warning described was unreachable by construction — stronger than never-observed.

So the line fired only on the healthy path, and a genuine mode=type regression — if the invariant ever broke — would have arrived as one line among identical healthy ones, indistinguishable. Same shape as the scheduler's "nothing was delivered" (#107) two days ago, in the same file: a loud probe whose output doesn't discriminate.

The fix follows the author's own stated intent

The surrounding comment already says it: make every mid-turn delivery self-announcing so a sweep can count them, and mode="type" "should be structurally impossible". That's instrumentation, not an alarm.

So the record stays for every delivery, at INFO, and only the impossible case is a WARN.

The unreachable branch is now pinned by an assertion, not a comment

That leaves a branch that can't currently fire, which is acceptable only while the invariant holds. So:

a_generating_lane_always_pastes_so_the_type_warning_stays_unreachable

sweeps chars 0..100_000 across both picker shapes. If anyone widens must_paste so a generating lane can be typed into, that test fails loudly and the warning becomes live and correct in the same moment.

What I did not do

No behavioural test for the log split itself, and I'd rather say so than imply otherwise: it changes no delivery behaviour, and this repo has no tracing capture to assert on a log line. The discriminating evidence is the measurement above plus the pinned invariant.

cargo clippy --workspace --all-targets -- -D warnings exit 0; cargo test --workspace 39 result lines, 0 failed.

🤖 Generated with Claude Code

https://claude.ai/code/session_01B4vuEScunv4K6RMwQoT9xx

…eport (AEAB-25)

A single `warn!` whose TEXT asserted "mode=type here is a regression" while the
structured field carried the mode that actually happened.

25 of 25 records, all-time, are mode="paste" — the mode the same sentence calls
correct. And this is not luck: `must_paste` is `generating || chars > 400 ||
picker_shaped`, so inside `if generating { … }` the paste branch is taken for
EVERY input. The condition the warning described was unreachable BY
CONSTRUCTION, which is stronger than never-observed.

So the line fired only on the healthy path, and a genuine mode=type regression —
if the invariant ever broke — would have arrived as one line among identical
healthy ones, indistinguishable. That is the same shape as the scheduler's
"nothing was delivered" (LR-24) fixed two days ago, in the same file: a loud probe
whose output does not discriminate.

The surrounding comment already states the intent — make every mid-turn delivery
self-announcing so a sweep can count them, and `mode="type"` "should be
structurally impossible". That is instrumentation, not an alarm. So the record
stays for every delivery, at INFO, and only the impossible case is a WARN.

That leaves a branch that cannot currently fire, which is acceptable only while
the invariant holds — so the invariant is now an assertion rather than a comment.
`a_generating_lane_always_pastes_so_the_type_warning_stays_unreachable` sweeps
chars 0..100_000 across both picker shapes. If anyone widens `must_paste` so a
generating lane can be typed into, that test fails loudly AND the warning becomes
live and correct in the same moment.

No behavioural test for the log split itself, and I would rather say so than imply
otherwise: it changes no delivery behaviour, and this repo has no tracing capture
to assert on a log line. The discriminating evidence is the measurement above plus
the pinned invariant.

CI's exact commands: clippy --workspace --all-targets -D warnings exit 0;
test --workspace 39 result lines, 0 failed.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B4vuEScunv4K6RMwQoT9xx

Amux-Session: amux-errors-and-bugs
@Dygreens
Dygreens merged commit a7a3238 into main Aug 18, 2026
3 checks passed
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