Repository navigation
ci: scratch push to validate the #707 Windows fixes - #710
Merged
blafourcade merged 10 commits intoAug 23, 2026
Merged
Conversation
The journal's per-write ACL reset walked (`icacls /T`) into files it does not own - a checked-out `.gitkeep` among them - and left it with no usable ACE, so an ordinary `git add -A` right after got "Permission denied" opening it. Drops `/T`: a file this code writes already gets its own direct icacls pass, and `(OI)(CI)` alone makes the directory's own grant apply to anything created afterward, so nothing needed the recursion. A marker file also stops the directory reset from repeating on every session-start once it has already run. Closes the rest of the set the Windows measurement found: the doc/code-parity test for the figures location now checks the sink's real per-platform answer instead of a POSIX literal, folding the Windows pin into the same three subtests rather than a fourth; eleven CLI integration tests get the direction-correct fix each one actually needed - a native `join` used for a virtual, always-"/" relative path in one production translator (a real small defect, fixed), the same native-vs-virtual-path mismatch inverted across four test files, and three hardcoded POSIX literals; the golden baseline's own normalizer gains a drive-letter/backslash fold so it redacts a Windows path the same way it already redacts a POSIX one (a real gap, not a content difference - the settings.json hash mismatch traced to the same un-redacted path, not different bytes); the other golden file's two timeouts get a per-test bump, not the e2e project's global one. The windows-probe job also gets a step that echoes `icacls` for the runs directory and the written journal file, then runs `git add -A` against the real checkout - proof for both properties on the runner itself, not just inferred from a passing test. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VWNxk63AGKkqE8HRqHLjGp
This reverts commit db1310a.
The marker file from the first pass permanently disabled re-tightening the runs directory once written - exactly the "checked-out aidd_docs/ runs/ needs this chmod" case the reset exists for in the first place. Dropping /T alone already closes the git collision (it never touches a file already sitting in the directory), so the reset goes back to running on every write, unconditionally, matching what it did before this task beyond no longer recursing. Also names, rather than only routes around, the real defect the copilot/claude Mode A test fixes surfaced: resolveSourceForSettings re-resolves an already-absolute local source path against projectRoot unconditionally. Harmless for a real builtDir (always drive-qualified on a real Windows machine), but a genuinely drive-less absolute local path - e.g. a marketplaces.json committed on POSIX and read on Windows - would silently gain the current process's drive instead of failing loud. Not fixed here; named in both tests instead of left implied by a test-side literal-to-computed edit. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VWNxk63AGKkqE8HRqHLjGp
…#707 review) normalize()'s blanket backslash fold ran over the manifest after JSON.stringify - a settings.json path's separator is doubled by JSON's own escaping there, but so is nothing else, and the same fold would have flattened `\"` and `\n` right along with it, corrupting content JSON.parse would then throw on rather than merely miscompare. Walks the parsed manifest and normalizes each string value directly instead (no JSON escaping to fight at that point), and gives the one place that still reads raw JSON bytes off disk - recomputing a real file's hash - its own narrow un-escape: a doubled backslash back to one, scoped to .json files only, and only correct because every settings.json this matrix writes carries path/repo/plugin-name values and nothing else that JSON would need to escape. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VWNxk63AGKkqE8HRqHLjGp
…e test (#707) The remaining failure in this file was the same resolve() mismatch as the copilot/claude Mode A tests: FrameworkBuildUseCase.execute() calls resolve(options.outDir) before ever touching the filesystem through it, but this test's own colliding-file path was built from the un-resolved builtMarketplaceDir() output. Harmless for a real projectRoot; this test's drive-less "/proj" is what exposed it. One real, unfixed defect surfaced investigating the file's last failure and is now named at the assertion it breaks, not fixed: EnsureBuiltMarketplaceUseCase.nested() compares sourceDir/builtDir with a hardcoded "/" rather than path.sep, so on Windows it never recognises genuine nesting - the "dogfood" build (a marketplace whose source is the project being built) skips the temp-dir detour meant to keep it out of guardPaths()'s way. Compounded by resolve() only ever being applied to sourceDir (ensure-built-marketplace-use-case.ts:70), never to builtDir, so a separator fix alone would not close it - both sides need to compare normalized, equally-resolved paths. Out of this task's surgical scope; reported rather than patched blind. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VWNxk63AGKkqE8HRqHLjGp
builtMarketplaceDir() joins with the native separator, so a drive-less projectRoot yields a drive-less builtDir on Windows. That raw value was captured unresolved into FlatBuildStrategy's write target (absOut) at construction, while FrameworkBuildUseCase.execute() separately resolved its own outDir copy for the preBuild existence check only - letting writes and checks silently diverge on Windows. Resolve builtDir at the same point sourceDir already is.
…ls (#707) nested() compared sourceDir/builtDir with a hardcoded "/", never path.sep, so it silently missed real nesting on Windows - mirror guardPaths()'s own "/"-normalized comparison. Also update the four test-side builtDir literals left unresolved after the prior commit started resolving builtDir in production: they now seed/expect the same drive-qualified path the use-case actually reads and returns.
…#707) Same gap as the prior two commits: execute() now always resolves builtDir before capturedOutDirs sees it, but this test's cacheRoot comparison literal stayed unresolved.
…707) This repo has no .gitattributes, so a Windows checkout's core.autocrlf rewrites every text file's line endings to CRLF - hashDirectory() was hashing those raw bytes, diffing the frozen claude cell against the LF-committed stored baseline on line endings alone, not real content. Corrects the prior diagnosis: AC #1 was not only a timeout (fixed by the earlier 120s bump) but a second, separate content-hash mismatch the timeout fix then exposed.
blafourcade
merged commit Aug 23, 2026
716661c
into
claude/aidd-telemetry-layer-e403uf
14 of 15 checks passed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Scratch PR to run the windows-probe job against the fixes for #707. Base is the layer branch (not main) so it merges cleanly and pull_request triggers fire. Will be closed and the branch deleted once the run is read.