fix(#322): the third reader of the same marker, and a freeze guard that outlived its claim

[fdaviddpt]

Builds on @jmossie82's #325 (merged). Their patch is unchanged — this is the sweep it opened up.

Part of #322

The third and fourth sites

10# landed in save-session.sh. scripts/post-tool-hook.sh reads the same tmp/last-save-ts at :335 and tmp/no-transcript-notice at :192, both case-guarded since #230, neither carrying 10#. So 08/09 clear the guard and are read as octal, the body is abandoned, and:

• :335 — IN_COOLDOWN stays false, the hook forks a save inside the window
• :192 — worse: the abandoned body swallows its own exit 0, so the line below the fi deletes the notice marker on the assumption a transcript was found

Both tests assert behaviour (save-session.pid absent / the notice marker surviving), not just the diagnostic. That mattered: the first draft asserted proc.stderr alone and passed against the broken code, because bash's line goes to hook-errors.log rather than stderr on this path.

SAVE_COOLDOWN was left alone, deliberately

I briefed this as having the same gap. It does not. LAST_TS is inside $(( )); SAVE_COOLDOWN is the right-hand operand of [ … -lt … ], the test builtin, which parses base 10 and does not evaluate:

$ [ 9 -lt 010 ] && echo decimal || echo octal      -> decimal
$ X='a[$(touch p3)]'; [ 5 -lt "$X" ]              -> integer expression expected, rc 2, nothing run

10# there would be inert and worse than inert — it would advertise a hazard at a site that has none, and its absence elsewhere would then read as a finding. The measurement is in the code comment so it does not get "fixed" later.

The judgment call that needs a second opinion

test_the_per_tool_call_path_is_not_touched (#298/#299) asserted post-tool-hook.sh is byte-identical to origin/main. That blocked a correct one-token fix and would block every future edit to the file. It is also self-cancelling: once the edit it objects to lands on main, it compares against the new main and passes. So what it actually asserts is "nobody has touched this file since main" — a different claim from the one its name makes, "the per-tool-call path is still cheap". A check that reads as evidence about the world when it is evidence about the diff.

Replaced the post-tool-hook.sh arm with the three properties #298/#299 care about — the divergence library is not sourced, case_divergence is not called, no git is spawned — each with its own message. Stronger than byte-equality after this merges, and cannot go vacuous. The lib-slug.sh / lib-memory-dir.sh arms keep their byte compare untouched.

If you would rather that were its own PR against #298, say so — but then the third site cannot land at all, because the guard forbids it.

Verification

Their tests against main's save-session.sh: 9 failed, 3 passed (the 3 are the two whitespace controls and the valid-marker case, by design). Their tests with their patch: 12 passed. My two, before my 10#: red with value too great for base at :325 and :186. Full suite 1519 passed, 43 skipped.

macOS + bash 3.2.57 only. bash 5 unverified — no bash 5 on this machine. CI is the authority.

[fdaviddpt]

Code review

No blocking issues. One non-blocking finding, filed as #330 rather than held against this PR.

The byte-identity pin on the per-tool-call path is gone, and the three substring checks that replace it can be walked around — GIT=git; "$GIT" -C ..., an inlined divergence check without the literal case_divergence, or any added spawn or file read that is not a git call. #299 cited that byte-identity as its evidence for the zero-cost claim. Some loosening was unavoidable, since the old assertion blocked this PR's own one-character fix, so the right answer is a cost-counting pin rather than keeping the byte compare. Scored 60. Details in #330.

Checked and clean: the two 10# sites are both already case-guarded to digits, so 10# never meets an empty string; SAVE_COOLDOWN is correctly left alone as test's right operand; the 50-git-backup.sh / 50-git-restore.sh deferral matches #327's ordering; bash 3.2.57 compatibility holds across the changed lines; no comment or docstring left contradicting the code.

Adjacent, already open: #326 names post-tool-hook.sh:316-326 — the same line this PR touches — for a range check rather than a syntax check. A future-dated marker sticks the throttle on with no self-heal. Out of scope here, still open there.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}