[krishhgg]

Phase 1 of the testing plan, item 3: lint the shell scripts the tests write. It was stacked on #85. #85 and #89 have merged, the base is now main, and 74a580d merges main in.

Why

Fakes of sudo, pmset, launchctl and the other tools decide what a test sees. A broken fake makes a test pass or fail for the wrong reason. A fake gh that used $10, which bash reads as ${1}0, faked 16 failures on #51. The ShellCheck job lints scripts/*.sh only. The fakes are strings inside Swift tests, and nothing linted them.

What changes

• One writer. TestScripts.write(_:to:executable:) (Tests/InsomniaTests/TestScripts.swift) writes a script and makes it executable, unless the test runs it through /bin/bash. The 30 call sites that write fakes, stand-in backstops, wrapper scripts and the fake swiftc's icon generator now go through it. Two kinds of script don't, because the ShellCheck job lints the originals: patched copies of the real scripts in scripts/, and test harnesses built from functions cut out of them.
• A copy of each script. When INSOMNIA_TEST_SCRIPT_DUMP names a folder, each script is also copied there, once per distinct text. UUIDs (the per-test temp folders) are zeroed first. That way the RecoveryScriptTests fixture's fakes, written for each of its 231 tests, land once. index.tsv names the test source line that first wrote each copy.
  • A script with no #! line is one a test runs with /bin/bash. Its copy (never the real file) starts with # shellcheck shell=bash, so ShellCheck lints it as bash instead of stopping at SC2148.
  • The index line is written before the copy. A copy can't exist without being linted, and an index line whose copy failed to write fails ShellCheck.
• The CI step. The Swift job sets the variable. A new last step downloads ShellCheck 0.11.0, checks its sha256 against the digest GitHub lists for the release asset, and lints the folder at ShellCheck's default severity, the same as the scripts job. It runs whenever the tests passed, so a lint failure doesn't hide the release build's result. An empty folder fails the step, so a renamed variable can't pass by checking nothing.
• Four deliberate constructs. ShellCheck flagged four things, and all are intended. Each gets a # shellcheck disable= line with the reason:
  • SC2053 in RecoveryScriptTests' mv fake: its mv.fail patterns are globs.
  • SC2254 in PackagingTests' mv fake: its mv-fails patterns are globs.
  • SC2015 in the sudo fake's visudo branch: A && B || reject means "reject unless both pass".
  • SC2034 in capture.sh: the test checks that the command substitution sees end of file, not what it captured.

Checks

• Local run at 2ca0534. The tree's diff hash matched before the run, after it, and at commit. The seven classes that write scripts ran with the variable set: RecoveryScriptTests, PackagingTests, LidSimulationGateScriptTests, PrivilegedCommandLockTests, CommandCancellationTests, DockerRuleEndpointTests and ReleaseWorkflowTests.
  • 288 tests, 1 failure. The failure is testTheSupervisorOutlivesItsRunAndGroupSignalsAndHoldsTheLockUntilItReapsTheCommand, the fake sudo race on main that Tests: let the fake sudo wait on SECONDS so one signal is logged once #89 fixes. This PR doesn't change that fake's text. The Mac did not sleep during the run.
  • The folder held 73 distinct scripts, up from 69 at b9c5ea5. The four new ones are the icon generator, the backstop stand-in, the sleep sentinel and the umask wrapper.
  • The CI step's script, run against that folder with the pinned ShellCheck, exited 0.
• Earlier local run at b9c5ea5. Clean tree before and after: 288 tests, 0 failures, 69 distinct scripts, step exit 0.
• The step fails when it should.
  • Before the annotations, the step failed on the five warnings and notes above.
  • With a $10 fake added to the folder, it failed with SC1037 ("Braces are required for positionals over 9").
  • With $10 planted in the umask wrapper's copy, one of the scripts without #!, it also failed with SC1037. The same copy without the shell=bash line fails with SC2148.
  • With an empty folder, it failed with "the tests wrote no scripts".
• CI. At 2ca0534 both jobs passed: macOS 26.6.2 (25G83) and macOS 27.0.1 (26A434), 1223 tests, 10 skipped, 0 failures each. On both, the ShellCheck step found 73 distinct scripts and passed. At 74a580d, which adds Tests: let the fake sudo wait on SECONDS so one signal is logged once #89's fake sudo text, both jobs passed again with the same totals.
• Reviews. Codex (gpt-6.1-sol, xhigh, read-only) reviewed a88d05c and found three gaps. The fake swiftc's icon generator sat in a heredoc that ShellCheck read as text. Four scripts without #! skipped the helper; Greptile's comment flags the same thing. A copy could exist without its index line. 2ca0534 fixes all three. Codex found nothing at 2ca0534, or in the merge at 74a580d, which adds exactly Tests: let the fake sudo wait on SECONDS so one signal is logged once #89's change.
• Not run locally. UIStatusTests, UIStartupTests and KeychainStoreTests (none write scripts). CI runs the whole suite on macOS 26 and 27.
• actionlint 1.7.12 passes locally.

Tests and CI only: no Mac report needed.

🤖 Generated with Claude Code

Confidence Score: 5/5

The PR appears safe to merge; no actionable new issues remain.

Summary

The PR sends test-written shell scripts through TestScripts.write and adds a CI step to lint their copies.

• CI now checks every distinct shell script that the tests write.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  Tests[Test script writers] --> Writer[TestScripts.write]
  Writer --> Original[Script used by the test]
  Writer --> Dump[Normalized lint copy and index.tsv]
  Dump --> Lint[CI ShellCheck step]

Loading

_{Reviews (3) · Last reviewed commit: "Merge main into ci/shellcheck-test-fakes" · Reviewed by Greptile}