Fix cat_tools dependency: install from pinned git SHA, not PGXN

[jnasbyupgrade]

What was broken

extension_drop.sql calls cat_tools.routine__parse_arg_types_text(), which only exists starting at cat_tools 0.3.0. The Makefile's cat_tools target and META.in.json's declared floor both said >=0.2.1 and used pgxn install.

PGXN's published cat_tools listing (https://api.pgxn.org/dist/cat_tools.json) is stuck at 0.2.1 from 2017 -- it's the old decibel/cat_tools listing and was never updated for the Postgres-Extensions/cat_tools fork's newer work. So pgxn install 'cat_tools>=0.2.1' in CI has actually always installed that ancient 0.2.1, which:

1. Doesn't have cat_tools.routine__parse_arg_types_text at all (added in 0.3.0), so CREATE EXTENSION extension_drop CASCADE has likely never actually succeeded.
2. Carries the pre-omit_column-fix catalog SELECT * bug that breaks with column "oid" specified more than once on PG12+.

Why PGXN can't be used right now

cat_tools's GitHub master has bumped to default_version = '0.3.0' and has the function extension_drop needs, but 0.3.0 has not been tagged or published to PGXN. Only 0.2.2/0.2.3 are tagged there, and even 0.2.3 predates the function. So there is currently no PGXN-published version of cat_tools that satisfies extension_drop's real requirement.

The fix

• Makefile: cat_tools target now clones Postgres-Extensions/cat_tools and builds/installs it from its master branch (tracked directly, not pinned to a commit SHA -- no tagged release has the function we need, so there's nothing more stable to pin to yet) instead of pgxn install. A comment at the fix site distinguishes the two claims: depending on cat_tools this way at all is temporary and will be reverted once 0.3.0 is tagged on PGXN, but building a dependency from a git ref is itself a supported technique in this Makefile -- worth reusing whenever a dependency's PGXN listing lags its real development, not something to design away.
• META.in.json: bumped the declared cat_tools build floor from 0.2.1 to 0.3.0 -- an honest reflection of what the code actually requires, independent of the install mechanism. Regenerated META.json via make META.json (never hand-edited).
• README.md: updated the stated cat_tools requirement/link to match, and noted the temporary git-source install.
• extension_drop.control's requires = 'cat_tools' is left as-is: PostgreSQL .control requires is a bare extension-name list (for CREATE EXTENSION ... CASCADE) with no version syntax, so no change was needed there.

Release safety

Cutting a real release while a dependency pin like this one is active would produce a citable, default-installable release whose declared dependency floor can't actually be resolved via pgxn install. Added a callout to RELEASE.md plus a small helper, bin/in_release (true when HISTORY.asc's top heading is a real version rather than the STABLE in-development placeholder), to check for this before releasing. Manual for now, not wired into CI -- doing that needs a generic way to detect "a pin exists" across whatever dependency needs one, not just cat_tools.

Related

Filed issue #11 to track a longer-term idea surfaced while working on this: a CI axis that tests the declared floor version of build/runtime dependencies (not just PG version), so a dependency floor like cat_tools 0.3.0 can't silently drift out of sync with what the code actually uses. Not implemented here -- cat_tools' floor isn't even testable until 0.3.0 is tagged.

Exit condition

Once cat_tools 0.3.0 is tagged and published to PGXN, revert the Makefile's cat_tools target to a plain pgxn install 'cat_tools>=0.3.0' --sudo (the comment at the fix site says so explicitly).

Verification

I confirmed via the CI job logs (using a temporary debug step I added and then reverted) that the git-clone/checkout/make install of the pinned cat_tools SHA succeeds cleanly on every PostgreSQL version in the matrix (9.3 through 17), installing cat_tools.control with default_version = '0.3.0' each time. That part of the fix is solid and verified end-to-end.

However, the schema/simple pgTAP tests still fail in CI, for a separate, pre-existing reason unrelated to cat_tools: CI's pg-build-test (from the pgxn/pgxn-tools image) invokes make installcheck directly, which does not go through pgxntool's own test target or its testdeps prerequisite. testdeps is what installs the test/extension/ support extension (extension_drop_test) that test/sql/schema.sql and test/sql/simple.sql depend on. Since pg-build-test skips that step, those tests fail with extension "extension_drop_test" is not available in CI, regardless of cat_tools. Locally, running pgxntool's actual make test (which does run testdeps first) installs extension_drop_test correctly, and with this fix, schema/simple then pass cleanly end-to-end (verified locally against PG17). This CI-invocation gap predates this branch and is out of scope for this PR; flagging it here so it isn't confused with a cat_tools regression. (There's also a related one-character pre-existing typo in this Makefile, line 4: $(DESTDIR)$datadir)/... is missing a ( before datadir -- harmless today since it just makes that target's dependency tracking always stale/rebuild, but worth a follow-up cleanup.)

Separately, zzz_build fails locally and in CI for an unrelated, already-merged reason: it plays sql/extension_drop.sql directly via \i rather than through CREATE EXTENSION, so it never benefited from CREATE EXTENSION's automatic client_min_messages raise-to-WARNING that PR #3 (224e89d, "Remove redundant client_min_messages handling from install script") relied on when it dropped the script's own explicit save/restore. Also pre-existing, also out of scope here.

None of this is gating today: pgxntool's .IGNORE: installcheck plus pg-build-test not gating on regression results means CI shows green regardless (as noted in the original bug report). This PR's all-checks-passed is green, and the cat_tools install itself, the actual subject of this PR, is verified working on every PG version.

Test plan

• Confirmed cat_tools git-pin install (clone + checkout + make install) succeeds on all 12 PG versions in CI (9.3-17)
• Locally (PG17), schema and simple, the tests that exercise cat_tools, pass cleanly once testdeps has installed extension_drop_test (i.e. via pgxntool's own make test)
• Follow-up needed (separate from this PR): make CI's pg-build-test invocation exercise pgxntool's testdeps/test target so schema/simple actually run to completion in CI
• Follow-up needed (separate from this PR, already pre-existing): zzz_build NOTICE regression from PR Remove redundant client_min_messages handling from install script #3

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 780ad270-4f1f-4ce9-91a5-467697975535

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}