fix(scripts): row 211 R3 push guard tags, installer hooks dir, busy check errors (#1709)

This commit is contained in:
jason.woltje committed 2026-11-02 13:50:11 -06:00
1 parent 3a121885cf
commit 7d03f212db
9 files changed
+378 -25

No files matched your search

+173
View File
@@ -149,3 +149,176 @@ happened. The suite, the mutants and every other `scripts/test-*.sh` run in a CL
- The hook doesn't check tags differently from branches: a moved tag is refused like a rewritten branch.
- `slot-busy.sh` can call a process BUSY that isn't a run, for example an editor with a `test-*.sh` open.
It never reads files under a seat's directory, so it can't tell.
## R3, the GPT review (2026-11-02)
The lead's implementation sub-agent (Claude Opus 5.5, a child of Sage's T3 c5c1e8ba) wrote this round from
the brief `/mnt/storage/scratch/mosaic-lead/briefs/row211-r3-fix.md` and the review
`/mnt/storage/scratch/mosaic-lead/reviews/batch1-a-gpt.md` (F1 to F5 and the edge-case table). Base
09a847eda, branch `t3/row211-push-guard-r3-fix`: dca5f10cc and 3e39689c1 picked clean as 5734c1a80 and
c8f2cb70c, then one R3 commit. The suite goes from 61 cases to 85 and the mutants from 13 to 21. The
sections above stay as written; where R3 changes them, it says so here.
### F1: a failed read counted as a vanished process
What: `slot-busy.sh` skipped any process whose cmdline read failed, so a live process it couldn't read (a
mode-000 cmdline of a `node --test`, say) could end in CLEAR. Its own status read was unchecked too: it put
`$$` in the skip set first, and a failed read only ended the ancestor walk.
Fix: after a failed read the scan checks the pid directory. Gone means the process exited, and it's
skipped. Still there is an error: the scan names that cmdline, counts it and goes on, so BUSY lines still
print, then exits 2 and never prints CLEAR. Its own status must read and give a numeric PPid, or it exits 2
before the scan. A failed read further up the walk still only ends the walk, which can give a false BUSY but
never a false CLEAR.
Tests, section "slot-busy, read errors" (fake `/proc`): "a pid directory with no cmdline: exit 2", "it
names that cmdline and never says CLEAR", "a cmdline that can't be read (a directory): exit 2, no CLEAR", "a
read error beside a BUSY process: exit 2, the BUSY line kept, no CLEAR", "a node --test process whose
cmdline is unreadable (mode 000) is never CLEAR", "a failing read (tr exits 1) with every pid directory
still there: exit 2, no CLEAR", "a process that vanishes during its read isn't an error: exit 0 and CLEAR"
(a `tr` shim on PATH removes the pid directory, then fails) and "its own process unreadable (awk exits 1):
exit 2, no CLEAR" (an `awk` shim).
Mutants: `unreadable-read-clear` (every failed read skipped), `vanished-is-error` (no directory check, so
an exited process is an error), `read-errors-ignored` (no final error exit) and `self-read-unchecked` (its
own failed read ends the walk, as before).
### F2: the snapshot exclusion matched a substring
What: any command line holding `/shell-snapshots/snapshot-` was skipped, so a real `node --test` of a file
under a `shell-snapshots` directory read as CLEAR.
Fix: `RE_SNAPSHOT` matches only the wrapper's shape: a shell (`sh`, `bash` or `zsh`, with or without a
path), options that include `-c`, then `source` or `.` of a `shell-snapshots/snapshot-*.sh` file as the
first command. Any other command line that names such a path goes through the rules.
Tests: "busy on: node --test /seat/tests/shell-snapshots/snapshot-guard.test.mjs (node --test)" and "busy
on: bash /seat/shell-snapshots/snapshot-gate.sh (gate)".
Mutant: `snapshot-substring` (the old substring test put back).
### F3: an existing tag could be replaced or moved
What: tags went through the ancestor check, and `git merge-base --is-ancestor` peels a tag. A new annotated
tag object on the same commit, or a lightweight tag moved to a descendant, passed as a fast-forward.
Fix: after the delete and new-ref checks and before the object and ancestor checks, a `refs/tags/*` ref
whose remote object name isn't zero and differs from the local one is refused ("an existing tag can't be
moved or replaced"). A new tag still passes, and a tag delete is still refused as a delete. This corrects
the tag line in "Known limits" above: a moved tag was refused only when it left its line, not when
it moved to a descendant or was re-made on the same commit.
Tests, section "tags": "a new annotated tag passes", "push --force of a new tag object on the same commit
is refused", "the remote tag is still the first tag object", "a new lightweight tag passes", "push --force
of a lightweight tag moved to a descendant is refused", "the remote lightweight tag hasn't moved" and "a
tag delete is refused, and the tag stays".
Mutant: `tag-replace-passes`.
### F4: a symlinked hooks directory
What: the installer wrote through `.git/hooks` even when it was a symlink, so an install could set the hook
of every repository sharing that directory, or write outside the repository.
Fix: `own_hooks` refuses (exit 2) when `<common>/hooks` is a symlink, or when it exists and its physical
path isn't the common directory's physical path plus `/hooks`. It runs before the pre-push symlink check
and again after `mkdir -p`, so both checks come before the temporary file and any install write. The
existing refusals (`core.hooksPath`, a different hook without `--force-install`, a symlinked `pre-push`)
are unchanged.
Tests: "the install refuses a symlinked hooks directory (exit 2)", "it refuses it with --force-install too"
and "nothing was written in the directory the symlink names".
Mutant: `hooks-dir-symlink-followed` (`own_hooks` returns at once).
### F5: the snapshot fixture proved nothing
What: the fixture wrapper ran `eval 'node --test x.test.mjs'`. There `node` follows a quote, and the node
rule needs a space or `/` before it, so the fixture wasn't BUSY even without the exclusion and the
`snapshot-matched` mutant survived.
Fix: two wrappers whose text matches a rule without the exclusion. Pid 4194403 runs `node --test
x.test.mjs` straight after the `source ... &&`. Pid 4194405 has the harness's real shape: `source ...
2>/dev/null || true && eval 'bash scripts/test-x.sh > x.log' && pwd -P >| ...`. The paired BUSY case is the
same node command under `/bin/bash -c` with no snapshot source. The F2 cases are above.
Tests: "an agent's shell-snapshot wrapper isn't a match" (now neither 4194403 nor 4194405) and "busy on:
/bin/bash -c node --test x.test.mjs (node --test)".
Mutant: `snapshot-matched`, whose site is now the `RE_SNAPSHOT` test. It is killed (3 FAIL). The suite
alone can't fail before this fix, because the old substring exclusion skips these wrappers too; the killed
mutant is the before-and-after evidence.
### Object names of the wrong length (the review's table)
What: the hook took any hex string as an object name, so a remote field of `0` read as a new ref.
Fix: after the non-hex check, each field must be exactly 40 or 64 characters, or the ref is refused.
Tests: "a remote object name of 0 is refused, not read as a new ref", "a 39-character local object name is
refused" and the control "a 64-character object name passes the length check (64 zeros refuse as a
delete)".
Mutant: `short-sha-accepted`.
### What didn't change
The earlier cases all still pass: an empty push, a new branch and a fast-forward pass; a delete, `--force`,
`+refspec` and `--force-with-lease` rewrites are refused; a remote tip missing locally is refused with
"fetch first", and nothing is fetched.
### R3 runs
`TMPDIR` was a fresh `mktemp -d /mnt/storage/scratch/tmp/XXXXXX`. Every run was in the foreground, with
output to files there.
- RED, all 24 new cases in place and no fix: 67 passed, 18 failed, 2026-11-02 19:27:01Z to 19:27:06Z. The
six new cases that pass there are controls that must hold before and after: both new tags pass, the tag
delete is refused, the 64-character control, the bare `bash -c` node case and the vanished process.
- GREEN: 85 passed, 0 failed, 19:28:58Z to 19:29:04Z. The run before it gave 80 passed, 5 failed: the new
closing error said "not CLEAR", which the "no CLEAR" assertions caught. It now says "the host isn't
known to be idle".
- Mutants: 19:30:11Z to 19:33:22Z, base green (85/0), runner exit 0, 21 of 21 killed, none survived, no
NO MATCH. Every kill exited 1.
| Mutant | FAIL | First failing case |
|---|---|---|
| ancestor-dropped | 8 | push --force of a rewritten branch is refused |
| unknown-object-passes | 2 | a push over a remote tip missing locally is refused, saying fetch first |
| delete-passes | 12 | the install into a clean clone exits 0 |
| install-over-hookspath | 2 | it says core.hooksPath is set |
| busy-exits-0 | 14 | a node --test process: exit 1 |
| differing-hook-replaced | 3 | the install refuses a different installed pre-push (exit 2) |
| replace-in-place | 1 | the replacement was a rename: the old file keeps its bytes |
| symlink-followed | 1 | the install refuses a symlinked pre-push, even with --force-install |
| ancestor-error-passes | 1 | a failed ancestor check (an unknown local object) is refused |
| snapshot-matched | 3 | no matching process: exit 0 and CLEAR |
| ancestors-matched | 1 | this suite, an ancestor, has no BUSY line |
| no-node-rule | 9 | a node --test process: exit 1 |
| no-compose-rule | 2 | busy on: docker compose -p x up -d (docker compose) |
| unreadable-read-clear | 6 | a pid directory with no cmdline: exit 2 |
| vanished-is-error | 1 | a process that vanishes during its read isn't an error: exit 0 and CLEAR |
| read-errors-ignored | 6 | a pid directory with no cmdline: exit 2 |
| self-read-unchecked | 1 | its own process unreadable (awk exits 1): exit 2, no CLEAR |
| snapshot-substring | 2 | busy on: node --test /seat/tests/shell-snapshots/snapshot-guard.test.mjs (node --test) |
| tag-replace-passes | 4 | push --force of a new tag object on the same commit is refused |
| hooks-dir-symlink-followed | 3 | the install refuses a symlinked hooks directory (exit 2) |
| short-sha-accepted | 2 | a remote object name of 0 is refused, not read as a new ref |
- Real host: `bash scripts/slot-busy.sh` on the real `/proc` exited 1 on other seats' gate and `node
--test` runs, with no error line. Both live shell-snapshot wrappers on the host matched `RE_SNAPSHOT`.
- Pins: `git grep -nE '[0-9a-f]{64}' -- 'packages/*/tests/*' 'scripts/test-*.sh'` gives 33 lines (the
positive control). None names a path this round changes. The only test that names one is
`scripts/test-git-guard.sh` itself, and no test reads `docs/TOOLS.md`.
### Known limits after R3
- The tag line in "Known limits" above is superseded by F3. The other two lines still hold.
- The resolved-path refusal in `own_hooks` has no test of its own. A hooks directory that isn't a symlink
but resolves elsewhere needs a bind mount, so root, and the symlink check catches the symlink case first.
`hooks-dir-symlink-followed` turns off both checks together.
- A wrapper that sources a test script into its own shell, with no child process, is still skipped, as
before R3. That script's children are still matched.
- The installer's temporary file is named by pid (`.pre-push.install-git-hooks.$$`). R3 leaves it alone.
- The "Runs" section above was written before any run. The R3 runs here are the ones on record for this
round.
+10 -1
View File
@@ -68,7 +68,16 @@ run differing-hook-replaced $I '[ "$FORCE" = 1 ] || die 2 "refused: a different'
run replace-in-place $I 'mv -f -- "$tmp" "$HOOK" ||' 'cat -- "$tmp" > "$HOOK" && rm -f -- "$tmp" ||'
run symlink-followed $I 'if [ -L "$HOOK" ]; then die 2' 'if false; then die 2'
run ancestor-error-passes $H ' *) refuse "$rref: the ancestor check failed' ' *) : "$rref: the ancestor check failed'
run snapshot-matched $S $' if [[ "$c" == *"/shell-snapshots/snapshot-"* ]]; then return 1; fi\n' ''
run snapshot-matched $S $' if [[ "$c" =~ $RE_SNAPSHOT ]]; then return 1; fi\n' ''
run ancestors-matched $S ' [ -z "${SKIP[$pid]:-}" ] || continue' ' :'
run no-node-rule $S 'if [[ "$c" =~ $RE_NODE ]]; then echo "node --test"; return 0; fi' ':'
run no-compose-rule $S 'if [[ "$c" =~ $RE_COMPOSE ]]; then echo "docker compose"; return 0; fi' ':'
# R3, the GPT review (F1 to F5 and the object-name length check).
run unreadable-read-clear $S ' [ -e "$d" ] || continue' ' continue'
run vanished-is-error $S $' [ -e "$d" ] || continue\n' ''
run read-errors-ignored $S '[ "$errors" = 0 ] || {' 'true || {'
run self-read-unchecked $S ' || { echo "slot-busy: error: cannot read its own process" >&2; exit 2; }' ' || p=""'
run snapshot-substring $S '[[ "$c" =~ $RE_SNAPSHOT ]]' '[[ "$c" == *"/shell-snapshots/snapshot-"* ]]'
run tag-replace-passes $H 'if [ "${rref#refs/tags/}" != "$rref" ] && [ "$lsha" != "$rsha" ]; then' 'if false; then'
run hooks-dir-symlink-followed $I 'own_hooks() {' 'own_hooks() { return 0;'
run short-sha-accepted $H ' if ! idlen "$lsha" || ! idlen "$rsha"; then' ' if false; then'