P-58's deleted revision: GitHub still lists it, with its time, and a deletion moves nothing (#549)

Merged
DO
6d ago
+587-10

Closes #549.

The trial

P-58's leg that decision 0140 assumed and no trial had recorded: what GitHub answers for a description edit whose revision a person deleted from the history. Verdict: pass, on the condition written first. A second leg, on the newest revision, neither passed nor failed.

  • Where: canary pull request 101, opened and its description edited twice by the Maintainer's login. The Maintainer then deleted the revision of 20:47:01Z in the browser, by hand, at 23:02:37Z.
  • What was read: the query the facts port sends, adapters/edits.py's QUERY word for word, as the Maintainer's login: after the opening, after the two edits, before the deletion and after it. Each was answered 200 with no errors.
  • What GitHub said after the deletion: the deleted revision's node is still listed, in its place, with its editedAt, createdAt and editor as before, and its deletedAt set. Every other node is as it was. totalCount stayed 5, and lastEditedAt and editor did not move.
  • What that shows, and what it does not: the port is answered a deleted middle revision in the shape it reads, and refuses nothing. It reads 2026-11-10T20:48:14Z before the deletion and after it. But the revision deleted was a middle one with two later edits above it, so the port's time could not have moved back whatever GitHub did. Muse and the security pass both said so.
  • The newest revision's leg: its condition was written at 23:13:41Z, before it ran. The Maintainer's login edited the description once more, at 23:13:57Z, and a fifth read shows that edit as the newest. The Maintainer then looked for the deletion of that newest revision in the browser's history: GitHub offered none. That is recorded on his word; no exchange shows it.
  • What stands: seen, one middle revision deleted and still listed with its time; reported, no deletion offered for the newest. No rule of when GitHub lets a revision be deleted is drawn from the two. Nothing in the port changes.
  • Not asked: whether anything but the browser deletes a revision, the newest among them. The read through the reader App's token was not repeated: its key is the Maintainer's, and trials/P-58.reader.json found the reader answered as the Maintainer's login is.

What this holds

  • trials/P-58.deleted.json: the five exchanges, redacted.
  • sim/grounding.py: the two rules of the fake that P-58 asks name the new fixture beside the other two. The fake has no rule about a deletion, and none is added.
  • tests/test_adapters.py: one test feeds the reads before the deletion, after it and after the later edit to the port. The contract's two readings of those rules replay the new fixture and show each rule from it.
  • tools/mutants.py: one mutant, a revision deleted from the history is passed over, killed by the test decision 0140's row names.
  • The design's P-58 row and adapters/edits.py's docstring say what the leg showed, worded to a middle revision. No code changes.

No decision record: the issue asks for a rule only if a deleted node's editedAt were null, and it is not.

Removed, and undone

  • Removed from the fixture: the authorization header gh sent, on each of the five requests. The redaction also dropped 11 response headers a fixture may not keep from each capture.
  • The pass condition's wording: it was written at 2026-11-10T20:45:57Z, before anything ran. One phrase of its last sentence, "found that token answered", read to the redaction as the shape of a credential and was reworded to "found the reader answered" after the run. The condition itself is unchanged.
  • Undone in the canary: pull request 101 is closed and its branch, trial/p58-deleted-revision-549, deleted. Nothing else was made: no token, secret, label, workflow or environment.

Checks

  • python3 -m tools.check_trials: ok, before the fixture was added to git.
  • make checks: OK at d05f30c5, which is 59ab504f with main merged at 7f968f09 and the rewordings cubic asked for.
  • tests.test_adapters, 168 tests, and the fixture and grounding tests of tests.test_tools, 27 tests: OK.
  • Mutants, filtered to the edits port and the grounding: 20 of 20 killed.
  • CI at c31568e0 failed checks (0): naming the new fixture in the grounding left Contract.test_the_port_reads_each_grounded_rule_as_the_fake_assumes_it behind. Fixed in 59ab504f.

Reviews

Carbon, Muse and Astra, the Claude review and the security pass. Carbon, Muse and the two Claude reads read c31568e0; all their findings are fixed in 59ab504f. Astra read 59ab504f, after those fixes, and found nothing. Newer than Astra's read: the merge of main, and fb65053c and d05f30c5, which change words only, in the fixture's account, the design's row, a docstring and a comment. Codex: not run (quota). CodeRabbit: skipped, auto reviews are off on this repo.

cubic and CodeAnt, on 6bc4d7c4, and cubic again on fb65053c:

FindingWhereWhatDisposition
cubic, P2trials/P-58.deleted.json"A deletion is itself no edit" generalizes from a middle revision's deletionfixed in fb65053c: the account, the design's row and the test's comment say a middle revision's deletion moved no time, and nothing of the newest's
cubic, P3adapters/edits.py:10The browser's offering no deletion of the newest is given as seen, and is the Maintainer's wordfixed in fb65053c: the docstring and the design's row say he reported it, and that no exchange shows it
cubic, P2, second readtrials/P-58.deleted.json:5The closing sentence gave a rule of when the browser deletes a revision, from one deletion and a reportfixed in d05f30c5: it says what was seen, what was reported with no exchange showing it, and that no rule is drawn from the two
CodeAnt, nitpicktrials/P-58.deleted.json:5The pass condition accepts any time that is no earlier, so a time moved forward by the deletion would passdeclined: the condition was written before the run and is not rewritten after it. That lastEditedAt did not move is what the reads before and after recorded, which the condition lists as recorded, deciding nothing. CodeAnt's own verdict is low risk
ReviewerHeadUnread partsBlockingMajorMinorInfo
agyc31568enone0000
musec31568enone1100
astra59ab504none0000
claude-reviewc31568enone1001
claude-securityc31568enone0011
FindingSeverityCriterionWhereWhatDisposition
muse#1blockingPR-3sim/grounding.py:26Grounding tuples name P-58.deleted but the contract read does not, breaking test_adaptersfixed in 59ab504f: both readings replay P-58.deleted and show their rule from it
muse#2majorPR-1adapters/edits.py:9Deleted-revision trial deletes only a middle revision, leaving the post-approval newest-delete case ungroundedfixed in 59ab504f: the fixture, the design's row and the docstring say a middle revision was deleted and the time could not have moved back; a second leg, its condition written first, records that the browser offered no deletion of the newest; a new mutant holds that a deleted node counts
claude-review#1blockingPR-3sim/grounding.py:26The grounding names P-58.deleted for two rules, but the test that holds each grounded rule to its fixtures was not changed, and fails at this headfixed in 59ab504f: both readings replay P-58.deleted and show their rule from it
claude-review#2infononetrials/P-58.deleted.json:5The fixture's account times the read after the deletion at 23:02:55Z, and the exchange it records is dated 23:02:56fixed in 59ab504f: the read is timed by its answer's date, 23:02:56Z
claude-security#1minornonetrials/P-58.deleted.json:5The trial deleted a middle revision, so its 'time no earlier' leg could not fail, and the one deletion the hold of 0140 depends on, the newest edit's, is not recordedfixed in 59ab504f: the fixture, the design's row and the docstring say a middle revision was deleted and the time could not have moved back; a second leg, its condition written first, records that the browser offered no deletion of the newest; a new mutant holds that a deleted node counts
claude-security#2infononetrials/P-58.deleted.json:5The pass text times the read after the deletion at 23:02:55Z, and the exchange it describes is dated 23:02:56fixed in 59ab504f: the read is timed by its answer's date, 23:02:56Z

šŸ¤– Generated with Claude Code