[cknitt]

Windows-only CI failures (~18 jobs, 2026-10-20 to 10-24) were all in the OCaml rewatch test steps:

• Canonical suite (5×): rm: cannot remove 'lib/watch.lock': Device or resource busy, then "Watcher process N did not exit". The watcher re-reads the lock every 100 ms through Unix.openfile, which on Windows doesn't share delete unless O_SHARE_DELETE is passed. An rm that lands during a read fails, and the watcher is left orphaned. Rust's File::open shares delete by default, so this started with the OCaml port.
• Parity (1×): mv of a watched directory failed with Permission denied. Same mechanism: the watcher hashes a file inside it.
• Focused, retained-parse (6×): after B is edited, the log never shows A being parsed again within 20 s. The log dump from change 3 caught this on this PR's first Windows run: A was parsed, but its line was lost from REWATCH_COMPILER_CALL_LOG. Parallel compiler requests each reopen that log in append mode, and on Windows appending is seek-then-write, so one line can overwrite another. The initial build's compile line for A was missing from the same log. So the duplicate-event explanation in Fix watch tests that flake on duplicate change events #8902 was wrong.
• Parity (4×): "The hosted runner lost communication with the server", about 45 min into the step despite timeout-minutes: 20, and no log survived.

Changes:

1. File_util.open_for_reading passes O_SHARE_DELETE (a no-op on Unix). All read-only opens and the lock publisher use it.
2. exit_watcher in rewatch/tests/utils.sh retries the rm and kills a watcher that still doesn't exit, so it can't run into later tests.
3. The retained-parse test prints watch.log and calls.log when its wait fails.
4. A watchdog in the parity step (Windows only) reports the largest processes with their command lines and stops rescript/bsc processes when one exceeds 3 GB, more than 64 are running, or free memory falls below 1.5 GB. The step then fails with that report instead of the runner dying silently.
5. Writes to the compiler call log go through a mutex.

New tests: a unit test in file_util_tests.ml renames and deletes a file held open by open_for_reading. It only proves anything on the Windows unit-test step. I checked the watchdog wrapper locally with a stub powershell.exe. The PowerShell script itself first runs on Windows CI.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GtJW7McE7BZqNZGt5VULiK

[github-actions]

Changelog

🐛 Bug fix

• On Windows, let other programs delete or rename files while the build system reads them, so stopping a watcher by removing lib/watch.lock no longer fails intermittently.

Semantic Diff

4 commits · Full page

Open files with delete sharing on Windows

_{314e53f · cknitt + Claude · Oct 25, 2026, 05:42 UTC}

Unix.openfile opens files on Windows without FILE_SHARE_DELETE unless asked, so while the watcher read lib/watch.lock (every 100 ms) or hashed a source, other processes could not delete or rename it. Removing the lock then failed with "Device or resource busy" and the watcher kept running; Rust's File::open shares delete by default.

Retry removing watch.lock and stop a watcher that ignores it

_{a1da0c0 · cknitt + Claude · Oct 25, 2026, 05:43 UTC}

A single failed rm left the watcher running into later tests, where the runner later killed it as an orphan. Retry for five seconds, and kill the watcher when it still does not exit so the failure stays in its test.

Diagnose Windows rewatch CI failures that leave no evidence

_{94ef5ea · cknitt + Claude · Oct 25, 2026, 05:43 UTC}

The retained-parse test now prints watch.log and calls.log when the edit of B does not re-parse A. A watchdog in the parity step reports and stops runaway rescript/bsc processes: four Windows runners lost contact with GitHub during that step, past its timeout, without any log.

✨ Serialize writes to the compiler call log

_{c5694c7 · cknitt + Claude · Oct 25, 2026, 06:18 UTC · new}

Parallel domains appended to REWATCH_COMPILER_CALL_LOG through separate handles, and Windows emulates append as seek-then-write, so one domain's line could overwrite another's. The retained-parse test then missed A's parse line (and the initial build's implementation line for A) although A was parsed: the Windows "flake" in that test was a lost log line.

[cknitt]

@codex review

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-25T06:29:13.016866Z	c5694c7	Manual request

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c5694c7e44

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Wait for the watchdog before reading and deleting its report

On Windows, if the parity step finishes while the watchdog is sleeping or collecting process details, removing the sentinel does not stop its current iteration: windows_process_watchdog.ps1 checks the sentinel only before its five-second sleep. This cleanup immediately reads the report and deletes its directory without waiting for PowerShell, so an in-flight watchdog action can be missed and the step can return success despite stopping processes. The watchdog can also continue into subsequent CI steps. Save the background PID and wait for it to finish after removing the sentinel, before inspecting the report or deleting its directory. This also satisfies the scoped process-ownership requirement.

AGENTS.md reference: rewatch-ocaml/AGENTS.md:L29-L31

Useful? React with 👍 / 👎.

[pkg-pr-new]

Open in StackBlitz

rescript

npm i https://pkg.pr.new/rescript@9034

@rescript/belt

npm i https://pkg.pr.new/@rescript/belt@9034

@rescript/darwin-arm64

npm i https://pkg.pr.new/@rescript/darwin-arm64@9034

@rescript/linux-arm64

npm i https://pkg.pr.new/@rescript/linux-arm64@9034

@rescript/linux-x64

npm i https://pkg.pr.new/@rescript/linux-x64@9034

@rescript/runtime

npm i https://pkg.pr.new/@rescript/runtime@9034

@rescript/win32-x64

npm i https://pkg.pr.new/@rescript/win32-x64@9034

commit: c5694c7

[github-actions]

Developer playground preview: https://rescript-lang.github.io/rescript/dev-playground/?version=pr-9034