[fhammerschmidt]

Watch mode could panic or exit when a configuration or source-layout change made full rebuild initialization fail. Initialization errors now propagate through the build lock, are reported, and keep the watcher alive for recovery.

After a failed initialization, source edits retry full initialization until the current configuration can be loaded, preventing compilation with stale suffixes, dependencies, or options. Malformed dependency configurations retain their resolved directory so fixing a newly added workspace dependency can trigger recovery. Unsupported JSX versions return configuration errors, and a Drop guard releases the build lock on errors and unwinding.

Dependency failures preserve exit code 2 and package/root-path diagnostic context. The package-tree progress line remains available; the detailed error is reported once.

Both recovery tests use temporary projects. The invalid-config test covers invalid JSON and unsupported JSX, records emitted files across recovery, and uses the same assertions for Rust and OCaml. Disabling Rust's source-retry fix makes the test fail with stale output, without relying on debug-log wording.

Recovery limit: installing a missing dependency or replacing a non-local dependency directory does not itself trigger a retry in Rust. Edit a watched source file or touch rescript.json after installation or replacement. Automatic detection remains a follow-up.

Validation:

• Full CI passes on Linux x64 (OCaml 5.0 and 5.5), Linux ARM64, macOS ARM64, and Windows x64, including native watcher unit, parity, and focused integration tests.
• All 57 shared canonical fixtures pass with the OCaml CLI on every platform and with the Rust CLI on Linux ARM64, including both recovery scenarios.
• PR package publication, all eight npm/pnpm installation tests, coverage, changelog validation, and DCO pass.
• Locally, all 150 Rust unit tests, Clippy with warnings denied, formatting, shell syntax, and whitespace checks pass. The stale-output negative control and 28 dependency-error checks verify the recovery and diagnostic assertions.

[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 2026-10-09T14:22:30.923223Z	f83cd4f	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.

[pkg-pr-new]

Open in StackBlitz

rescript

npm i https://pkg.pr.new/rescript@8641

@rescript/belt

npm i https://pkg.pr.new/@rescript/belt@8641

@rescript/darwin-arm64

npm i https://pkg.pr.new/@rescript/darwin-arm64@8641

@rescript/linux-arm64

npm i https://pkg.pr.new/@rescript/linux-arm64@8641

@rescript/linux-x64

npm i https://pkg.pr.new/@rescript/linux-x64@8641

@rescript/runtime

npm i https://pkg.pr.new/@rescript/runtime@8641

@rescript/win32-x64

npm i https://pkg.pr.new/@rescript/win32-x64@8641

commit: 0e3814b

[fhammerschmidt]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3377441a3d

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".

[fhammerschmidt]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Keep them coming!

Reviewed commit: f83cd4ffd8

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".

[github-actions]

Developer playground preview: https://rescript-lang.github.io/rescript/dev-playground/?version=pr-8641

[cknitt]

🤖 This review was generated by Claude (Claude Code) on behalf of @cknitt.

Verdict: approve with nits. No correctness bugs found in the watch loop. Init errors propagate instead of panicking. initialization_failed forces a full re-init rather than an incremental build on stale state. Every fallible step in initialize_build runs before cleanup_previous_build, so a failed init leaves nothing half-cleaned. Both new tests fail without the fix.

• Exit code change: the two process::exit(2) calls in packages.rs are gone and main.rs's exit_code returns 1. rescript build/clean/format now exit 1 instead of 2 when a dependency is missing or has a malformed config. weird_deps/weird_devdeps only assert != 0, so CI doesn't catch it. Please acknowledge it in the PR/changelog, or keep 2.
• Missing dependency dir gets no recovery watch: only DependencyConfigError adds one. Example: a dependency added to rescript.json before npm install. Finishing the install produces no event, so watch stays failed until a .res file or rescript.json is touched. The same happens when a non-local dependency with a bad config is reinstalled, because the directory is replaced. Still far better than exiting; could be documented, or the root node_modules watched non-recursively while init is failing.
• Duplicate test number: 07-watch-recovers-from-invalid-config.sh sits next to the existing 07-changed-interface-after-failed-implementation.sh. Renumber the new ones to 08/09.
• Test 07 can pass without proving anything: if the mv of the invalid config spans two debounce windows, a second init failure can bump initial_errors before the Test.res edit is processed, and the "no marker in Test.mjs" check then passes trivially. This can't cause a flaky failure, only a false pass. Waiting for a marker tied to the edit would tighten it.
• Nits:
  • With show_progress, dependency errors are printed twice (inspect_err in packages.rs, then eprintln!("{error:#}")).
  • The rustfmt re-indent in watcher.rs (~509-540) and packages.rs makes the diff noisier.
  • Out of scope, but other panics can still kill watch mode after a config change, e.g. "JSX version {v} is unsupported" (config.rs:834) and .expect("Could not canonicalize") (packages.rs:562). Possible follow-up.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 81.13%. Comparing base (6a3eed1) to head (0e3814b).

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #8641   +/-   ##
=======================================
  Coverage   81.13%   81.13%           
=======================================
  Files         456      456           
  Lines       63750    63750           
=======================================
  Hits        51724    51724           
  Misses      12026    12026           

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[fhammerschmidt]

The follow-ups document the exit-code change and dependency-install recovery limit, renumber the tests, and remove duplicate diagnostics. The recovery regression now waits for the source event and a subsequent initialization failure; disabling recovery makes it fail and reproduces stale output. Automatic detection of newly installed or replaced dependencies remains a follow-up; editing a watched source or touching rescript.json triggers recovery. CI, including the package-installation checks, coverage, and changelog validation now all pass.

[cknitt]

Checked against current master (applies cleanly, bugs still reproduce: invalid root config panics at watcher.rs:619, missing/malformed dependency exits 2, and lib/build.lock + lib/watch.lock are left behind). The OCaml implementation is not affected. The core watcher fix works for me: no stale output while the config is broken, and a correct rebuild once it's fixed.

A few things before merging:

1. Same class of panic remains. Changing the config to "jsx": {"version": 3} during watch still panics in Config::get_jsx_args (panic!("JSX version {version} is unsupported")) and again leaves both lock files behind. Turn that into a config error. A Drop guard in with_build_lock that releases the lock would also make this robust.
2. Exit code 2 → 1 breaks parity. Rust and OCaml currently both exit 2 for dependency errors (OCaml deliberately, rescript_ocaml.ml maps Package_error to 2). I'd keep 2 in Rust (e.g. downcast the dependency error in exit_code) instead of documenting a new intentional difference. If we do want 1, change both implementations together, and it's a user-visible change rather than a .fix.
3. Error messages lost context. The malformed-dependency error no longer names the package (Could not build package tree for 'bad' at path … → Could not initialize build: Could not read dependency config at …), and the missing-dependency error loses at path '<root>'. Restoring those would also avoid loosening require_both_errors_contain to per-implementation fragments.
4. Tests. Test 09 skips the source-retry check for OCaml, so its "no stale output" assertion passes trivially there. The Rust branch depends on the wording of a debug log line. Test 10 edits the shared testrepo/packages/standalone config, so a temporary fixture like test 07 would be safer.
5. Minor: dropping show_progress from packages::make also removes the [1/2] … Error building package tree stdout line, which is outside this PR's scope.

Optional follow-up: OCaml recovers automatically when a missing dependency gets installed (it watches resolution candidates); Rust still needs a manual touch, as the description notes.

[github-actions]

Semantic diff: https://cdn.rescript-lang.org/semantic-diff/pr-8641/index.html?v=0e3814bf5

[cknitt]

Rechecked at 0e3814b. All points from my previous comment are addressed:

1. "jsx": {"version": 3} is now a config error instead of a panic, and the build lock is released by a Drop guard.
2. Dependency errors exit 2 again, matching OCaml.
3. Errors name the package and root path again; the parity check uses one shared fragment.
4. Recovery tests use temporary projects, detect stale output via js-post-build instead of log wording, and assert the same for Rust and OCaml.
5. The [1/2] Error building package tree. line is back.

Verified locally with a build of this head: cargo test passes. In watch mode, invalid root JSON, an unsupported JSX version, a missing dependency and a malformed dependency all keep the watcher alive; a source edit while the config is broken emits nothing, and fixing the config rebuilds with the new suffix. build/clean/format exit 2 for dependency errors and 1 for root config errors and leave no locks behind.

Not blocking: automatic recovery on dependency installation remains a follow-up, as described. Separately, stopping watch with SIGTERM leaves lib/watch.lock behind, but master does the same, so that's not from this PR.

LGTM.