[scuttletoby]

Fixes #120
Fixes #121
Fixes #122
Fixes #123
Fixes #124

Follow-ups from the quick-check review of #113 (#113 (comment)), all in fork-ci/actions/ci-demand/.

What changes

• ci-demand: an unwritable cache is only a stderr warning and resume advice is wrong #120 A cache that cannot be written (read-only folder, full disk) was only a stderr warning. Now --json has cache_write_failed (the error, or null), and the first line of the text output says cache not written (...). When the run also stopped early, it no longer says "run again with the same --cache" (nothing was saved, so a rerun would fetch the same pages and a big range could never finish); it says "so a run again asks for the same pages: fix --cache first".
• ci-demand: four lines no test catches (private-cache refusal, fallback reserve, two UTF-8 settings) #121 Four lines that no test caught now have one test each: the "cannot be made private (0700)" refusal (chmod that has no effect), the 1000 reserve when X-RateLimit-Limit is missing, and UTF-8 for gh output and for the config file. The UTF-8 test runs the script itself in the C locale with UTF-8 mode off (ASCII), with a non-ASCII workflow file name and config text. The fake gh now writes its body as UTF-8 bytes like the real gh, and can leave a header out.
• ci-demand: a config file with a UTF-8 BOM is refused #122 The config file is read as utf-8-sig, so a UTF-8 BOM (some Windows editors) is accepted.
• ci-demand: a NaN cache date keeps a run-list page forever #123 A cache date that is NaN or -Infinity (json.load reads both) counts as damaged and the page is fetched again (math.isfinite).
• ci-demand README: cache folder must support chmod (FAT, WSL /mnt/c are refused) #124 README: on Linux or macOS the cache folder must allow private permissions; a folder on FAT/exFAT, WSL /mnt/c or some network mounts is refused (exit 2); pick a folder on a normal disk. Also documents the new --json field and that UTF-8 (with or without BOM) is read whatever the locale.

Tests

python3 -I -B fork-ci/actions/ci-demand/ci_demand.test.py (GitHub faked, no token, empty gh config): 41 tests, all pass (36 before). Red first: before the fix, the new or extended tests for #120, #122 and #123 failed (2 failures, 2 errors).

Mutation check (each change applied alone, then the suite run): all 8 are caught. These are: no private refusal, reserve fallback 0, gh without encoding='utf-8', the config without an encoding, the config as plain utf-8 (BOM), no isfinite, no cache_write_failed in the JSON, and the "run again" hint shown even after a failed write.

The reviewer's own round-2 test file: the cache-write and BOM tests pass. Its NaN test expects 2 requests, but it puts NaN into every cache file, including the 4 jobs pages. With the agreed fix, all 6 count as damaged, so 6 requests is right; the folded test checks that.

State question

"What happens when this runs again, stops half-way, or meets existing/old/empty/corrupt state?"

• Runs again: pages saved by a working cache are read back as before. When the cache could not be written, the result now says so instead of promising a resume.
• Stops half-way: unchanged (exit 3, INCOMPLETE). The resume hint is shown only when something could have been saved.
• Existing/old state: cache files of earlier runs of this version stay valid (same format, no FORMAT bump). A file with a NaN or -Infinity date, which only hand damage can cause, is fetched again once and then saved correctly. An existing config with a BOM now works. A config without one reads as before.
• Empty/corrupt: an empty cache folder works as before. Corrupt files are fetched again, never a crash. A folder that cannot be made private is still refused before any request (now tested).

General vs project

All general: the tool, its tests and its README name no project. Project facts stay where they were: the repository (--repo or ciDemand.repo in the project's own config) and the cache folder the caller picks.

🤖 Generated with Claude Code