Repository navigation
ci: run the unit and smoke suites on macOS and Windows - #144
Conversation
Adds a platforms job: unit suites on macOS and Windows, and the smoke suite on the installed tarball on Linux, macOS and Windows with Node.js 24 and 26. The smoke harness starts npm and npx through tinyexec's Windows shim handling, runs the installed bin file where npm links .cmd shims, sets USERPROFILE next to HOME, and takes the CLI's Node.js from ESTELA_SMOKE_NODE. Text files check out with LF everywhere. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Windows has no stand-in for DevFrame's cmd /c start, so the smoke tests pass --no-open there and check the printed URL only; node-pty rejects a signal name there. The session index tests compare platform paths and check POSIX modes on POSIX only. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… index rebuild The kit joins and splits session paths on / only, so a path typed with backslashes lost its Claude Code subagents and its Codex or Grok detection, and got a different trace id than the same file found by discovery. The CLI now hands the kit / separators on Windows, and trace ids hash that form. A Windows rebuild whose delete hits a file another process still has open now opens again instead of giving up. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…only Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n Windows loadSession and the catalog's summary pass now apply toKitPath, so every caller (the CLI, the server's tests, a future plugin) gets the kit's path form; the CLI keeps plain resolve(). The conformance tests build their paths in that form, the demo generator writes / on every system, and the platforms job runs every package's unit suite with --no-sort. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Windows' CREATE_NEW follows a dangling symbolic link, so the default export name allocation now treats any existing entry as taken. Also from review: the port-fallback smoke test passes --no-open, the CLI's Node.js directory goes on PATH so npx runs on it, a cross-platform disabled-index test, and the tests expect the kit path form and this system's absolute paths. Comments and the changeset state the Windows trace id change and the cost of the rebuild retry. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…atalog case everywhere From the follow-up review: the session view tests hand loadSession and summarizeRow this system's paths, the catalog's no-index case uses a file in place of the cache directory instead of chmod, the Playwright harness starts the CLI on ctx.node, and the export comment names the window left between the existence check and the open. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
@coderabbitai review |
Action performed
|
|
No actionable comments were generated in the recent review. Walkthrough
Merge Risk:
|
Summary
Closes #92. CI now runs on macOS and Windows as well as Linux, and the smoke suite runs the CLI on Node.js 24, the oldest release its
enginesallows. Running the suites on Windows turned up three product bugs, fixed here..github/workflows/ci.yml): a newplatformsmatrix job runs besideteston the same events. It runs every package's unit suite on macOS and Windows (pnpm -r --no-bail --no-sort test, so a failed package does not skip its dependents). It also runs the smoke suite on the installed tarball on Linux, macOS, and Windows, with the CLI on Node.js 24 and 26. Linux on 26 stays intest. Node.js 24 runs only the installed CLI, throughESTELA_SMOKE_NODE; the workspace toolchain stays on^26.13.1. The budgets and the Playwright suite stay intest, because their numbers are set on Linux runners..gitattributes):* text=auto eol=lf. Windows runners check out withautocrlf, and fixtures are read back by byte offset.packages/server): the kit joins and splits session paths on/only (Claude Code subagent directories, Codex rollout names, Grok session directories). A path with\therefore lost its subagents and its Codex or Grok detection. It also got a different trace id than the same file found by discovery, which yields the home directory's\followed by the kit's/.toKitPath(/on Windows, unchanged elsewhere) is now applied at the server's two entries to the kit,loadSessionandsummarizeRow, so every caller is covered: the CLI, MCP, and a later DSH plugin.traceIdOfhashes that form. POSIX ids are unchanged; the pinned ids insession-target.test.tsstill hold. The kit's redaction already matches both separator spellings of a Windows home.apps/cli/src/export/write-export.ts): Windows'wx(CREATE_NEW) follows a dangling symbolic link and creates its target.export-atifandexport-otlpwithout--outnow treat any existing entry, a link included, as a taken name. On POSIXwxalready refused it.packages/runtime-node): Windows refuses to delete a file another process has open. When several processes opened a broken index at once, the rebuilding process gotEPERMand ran without the index. On Windows that error is now retried like the busy and I/O errors, until the open's 5 s deadline.apps/e2e):.cmdshims.run()uses tinyexec'snormalizeSpawnCommandbut notx(), becausex()mergesprocess.envback into the child environment and the smoke environment removes variables on purpose..cmd,.ps1, and shell shims there.USERPROFILEis set next toHOME. The CLI's Node.js directory goes onPATHafter the fake opener, sonpxruns on it too.cmd /c start, so the tests that check the opened URL pass--no-openthere and check the printed URL only; the port-fallback test passes--no-openeverywhere.TERMfromnameon POSIX only.toKitPathform, and the CLI tests use this system's absolute paths./on every system.CONTRIBUTING.md: theplatformsjob.docs/development/setup.md: the Windows opener,ESTELA_SMOKE_NODE, and the node-pty pin, which had gone stale.@rivus/estelapatch for the three Windows fixes.Follow-up for the kit, recorded in #123: read Windows paths natively, by splitting on both separators or handing out
/paths from the Node platform, so thattoKitPathcan go. Trace ids keep hashing the/form either way.Validation
PR CI on the head
3074c60, 44047604319, passed every job:test(Linux, with the Playwright suite and the budgets), andplatformson ubuntu/Node 24, macos/Node 24, macos/Node 26, windows/Node 24, and windows/Node 26. The run before it, 44046926140, failed once on windows/Node 26: a catalog test removed its temp home before its lock-holding child had exited. The test now waits for that exit, as the session index test already did. Theworkflow_dispatchruns that showed the failures fixed here: 44035983739, 44036923678, 44042396654, 44043009782, and 44044193732.Windows smoke log lines from the last run show the CLI on 24 and the tests on 26:
node: v24.22.2 (C:\hostedtoolcache\windows\node\24.22.2\x64\node.exe), tests on v26.13.1.Locally on macOS (Node.js 26.13.1), all passed:
gh repo-checks repositorynode scripts/check-doc-translations.mjs,node scripts/check-conventions.mjs,node scripts/check-layers.mjs, plusnode scripts/check-release-intent.mjs,node scripts/format.mjs --check, andnode scripts/lint-scripts.mjsgh repo-checks pr-title,gh repo-checks pr-body.changeset/windows-paths-and-index-rebuild.md(patch)pnpm lint,pnpm test,pnpm typecheck,pnpm buildnode scripts/check-browser-bundle.mjsnode scripts/check-budgets.mjs. All within budget: tarball 661.5 kB of 790 kB,dist/cli.mjs1022.2 kB of 1230 kB, large-session peak RSS 476.7 MB of 1000 MB.cd apps/e2e && pnpm run test:smoke: 19 passed.The Playwright suite (
test:e2e) was not run locally; it runs in thetestjob, which passed. The local runs above predate the last two commits, which change only tests and comments; CI covers those commits.Risks
#/s/<id>link from an earlier version no longer opens its session, and the index computes those summaries again. Windows was untested before; the changeset says so. Letter case still counts in a Windows trace id:c:\usersandC:\Usersgive two ids.platformsjob is not a required check, becausemainhas no branch protection yet. Once protection is set, add its five contexts next totest./(Could not load C:\p\s.jsonl: claude-code: C:/p/s.jsonl:1: unknown format generation). Nothing parses across the two parts.--outpath already had the same window there, becauseO_NOFOLLOWis 0 on Windows.