CI: gate .bo/.ba format changes on their header-tag bumps#1041
Conversation
The .bo and .ba readers decide compatibility solely by the tag strings in GenBin.hs and GenABin.hs; a format change that forgets the bump makes same-tag readers misparse stale files and crash with an internalError instead of the intended recompile error. Several recent PRs made this slip independently. The gate: a PR touching GenBin.hs must change the .bo tag, GenABin.hs the .ba tag, and BinData.hs (instances shared by both formats) both -- with an auditable escape hatch for provably byte-neutral changes, a commit trailer 'Bin-format: bo-unchanged' / 'ba-unchanged'. Validated against history: the ranges of PRs that forgot the bump fail, ranges that bumped correctly pass, and a byte-neutral BinData change fails until it carries the trailer. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019cqQdWKp7jr73GHQYnvhiq
The BinData/GenBin changes in this PR thread a Position transform through the encoder without altering any serialized bytes when the flag is unused -- verified by byte-comparison against the previous compiler (see the test's no-flag check). The .ba tag was bumped for the real format change (new Flags field); the .bo format is unchanged. Bin-format: bo-unchanged Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019cqQdWKp7jr73GHQYnvhiq
quark17
left a comment
quark17
left a comment
There was a problem hiding this comment.
This is good to check! Just a few comments.
(I also wouldn't mind a check for Error.getErrorText that there are no tag clashes and that the clauses are in tag order.)
| # | ||
| # Not every change to these files alters serialized bytes (comments, | ||
| # the hash plumbing, dump/trace code). For those, add a commit | ||
| # trailer to any commit in the range: |
There was a problem hiding this comment.
It seems dangerous for one commit to be able to exempt other commits from the check. I thought maybe you were doing this to allow someone to just push a commit to a PR, rather than force-push a change to an earlier commit -- but they couldn't push an empty commit just for its message, so I think that's OK.
| # - a change to src/comp/BinData.hs (Bin instances shared by both | ||
| # formats) requires BOTH tags to change | ||
| # | ||
| # Not every change to these files alters serialized bytes (comments, |
There was a problem hiding this comment.
These are also not the only files that would require a bump, right? I think some types have their format generated using generics, and so merely changing the ISyntax or ASyntax could require a bump? But this at least catches a lot of differences.
| - 'src/comp/GenBin.hs' | ||
| - 'src/comp/GenABin.hs' | ||
|
|
||
| jobs: |
There was a problem hiding this comment.
My preference would be to include this in ci.yml alongside the cleanliness checks, so that all of the checks are on one page and don't require looking elsewhere (assuming you know to even look for another page) (and it's particularly bad on the GitHub phone app).
2 participants
The
.boand.bareaders decide file compatibility solely by the header tag strings insrc/comp/GenBin.hsandsrc/comp/GenABin.hs. When a PR changes the serialized format but forgets the tag bump, a same-tag reader misparses stale files and dies with aninternalError("please report") instead of the intendedEBinFileVerMismatch"recompile" error. Recent history shows this slip is easy to make — several open PRs changedBininstances with the tag untouched, and I made the same mistake myself last week.This adds a cheap, build-free CI job (only triggered when the three files change):
src/comp/GenBin.hsmust change the.botag;src/comp/GenABin.hsmust change the.batag;src/comp/BinData.hs(theBininstances shared by both formats) must change both.Not every change to these files alters serialized bytes (comments, hash plumbing, dump code), so there is an auditable escape hatch: a commit trailer
Bin-format: bo-unchanged/Bin-format: ba-unchangedasserting the change is byte-neutral. The trailer lives in history, so a reviewer can hold the commit to its claim.Validated against real ranges before opening this PR:
GenABin.hswithout a bump → fails with the actionable message;BinData.hschange (empirically verified bit-identical output) with a bumped.babut unchanged.botag → fails on.bountil the trailer is added — the intended behavior for that case.The gate deliberately does not touch the gated files itself. Note for existing open PRs: a few will trip this on their next push until they bump tags (correctly) or add trailers — that is the gate doing its job.