Remove reg* pseudotype columns; add object info functions

[jnasbyupgrade]

Drop the reg* pseudotype columns (regclass/regconfig/regdictionary/regnamespace/regoperator/regprocedure/regtype) from _object_reference._object_oid in favor of a single plain oid column (object_oid) and a plain oid classid, removing the now-unnecessary count_nulls-backed null_count trigger that used to arbitrate between them.

• Drop the count_nulls dependency entirely (the search_path workaround DO block, object_reference.control's requires, the Makefile install target, and test setup) — the trigger above was its only reason for being there.
• Add object_reference.object__describe() / object__identity() (thin wrappers around pg_describe_object()/pg_identify_object()), and object__cleanup() (best-effort delete, ignoring foreign_key_violation), wired to a new AFTER DELETE trigger on object_group__object so removing an object from its last group auto-attempts cleanup.
• object__getsert's core (_object_v__for_update) now refuses to track objects living in a pg_temp*/pg_toast_temp* schema.
• Test updates: object_group.sql's scratch tables switch from TEMP to regular tables (temp objects are now rejected) plus new coverage for the cleanup trigger; base.sql swaps the now-moot count_nulls-relocation test for coverage of object_oid, object__describe()/object__identity(), and temp-object rejection.
• _object_reference._object_oid.object_oid was later dropped too — it always equaled objid and never diverged.

sql/object_reference--0.1.0.sql (the frozen historical release) is untouched — frozen versioned SQL files are never hand-edited after release. default_version stays 'stable'; sql/object_reference--stable.sql is regenerated from sql/object_reference.sql via make.

Also adds README.md (installation, roles/security, key concepts, and the full public API reference) and fixes packaging metadata in META.in.json/META.json: adds cat_tools to runtime.requires (a real, current dependency per object_reference.control), updates decibel/* org URLs to Postgres-Extensions/*, and bumps the stated PostgreSQL build requirement from 9.5.0 to 12.0.0, matching the floor this repo's CI matrix already exercises (derived from cat_tools's own build requirement).

Fixes #35

Co-Authored-By: Claude noreply@anthropic.com

[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: 945490be-ba58-428a-a961-a2e3c360048a

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.}

[github-actions]

Claude finished @jnasbyupgrade's task in 8m 2s —— View job

---

Claude's Review

• Gather context (diff, CLAUDE.md, CI status)
• Review SQL changes (sql/object_reference.sql, control file, Makefile)
• Review test changes (test/sql/*.sql, expected outputs)
• Post review feedback

CI is green on this PR (make lint/make test both pass per the run linked below), and the diff matches what the PR description promises. Went through sql/object_reference.sql, the build/control changes, and both modified test files line-by-line against their expected/*.out counterparts.

What looks solid:

• reg* pseudotype columns, null_count trigger, and their unique indexes are cleanly removed; classid/object_oid collapse correctly into plain oid, and _object_oid__add's EXECUTE-built dynamic insert is replaced with a plain static INSERT — good simplification.
• count_nulls is fully excised (Makefile, control requires, test/deps.sql, test/load.sql, test/sql/zzz_build.sql) — no dangling references left anywhere except the frozen sql/object_reference--0.1.0.sql (expected, per repo convention) and the vendored pgxntool/README.*.
• object__describe/object__identity are thin, correctly-wired wrappers; object__cleanup's WHEN foreign_key_violation THEN NULL pattern is a clean way to make cleanup best-effort, and the new AFTER DELETE trigger on object_group__object is placed appropriately.
• The temp-schema guard in _object_reference._object_v__for_update (checking pg_identify_object(...).schema LIKE 'pg_temp%'/'pg_toast_temp%') runs before any row is written, and correctly no-ops when schema IS NULL (e.g. objects without a schema).
• Manually recomputed the pgTAP plan() math against the actual assertion count for both test/sql/base.sql (10) and test/sql/object_group.sql (29) — both match the committed expected/*.out files exactly, including the __remove off-by-one that commit 3486c06 already self-corrected (nice catch, and good that it was verified against real CI output rather than just the stale committed .out file).
• test/expected/zzz_build.out diff is just shifted line numbers / cascading NOTICEs from the shrunk _object_oid DDL — machine-regenerated noise, not hand-edited drift.

One non-blocking simplification worth considering: in sql/object_reference.sql, _object_oid.object_oid (~line 194) is now always set equal to objid — _object_oid__add's insert (~line 302) does VALUES (object_id, classid, objid, objsubid, objid). That makes the objid_must_match CHECK constraint (objid IS NOT DISTINCT FROM object_oid, ~line 190) tautological, and object_oid itself redundant storage of objid. Since this is a private (_object_reference-schema) table, it'd be safe to drop the column and reference objid directly in the views/object__* functions and in test/sql/base.sql's "Verify object_oid field is correct" check — unless a later PR in this 3-PR stack intends object_oid to diverge from objid for some object type, in which case it'd be worth a short comment noting why it's kept separate now.
Fix this →

No blocking issues found. Nice, focused reconciliation PR — the count_nulls removal and reg* collapse are handled consistently across every file that touched them.
· Branch: reconcile-object-functions

[github-actions]

Claude finished @jnasbyupgrade's task in 2m 8s —— View job

---

Claude's Review

• Gather context (diff, CLAUDE.md, CI status)
• Review SQL changes (sql/object_reference.sql, control file, Makefile)
• Review test changes (test/sql/*.sql, expected outputs)
• Post review feedback

CI is green (CI workflow run 31435600198 succeeded). This update adds two commits on top of what was previously reviewed:

9d909ce — Suppress NOTICEs during CREATE EXTENSION via SET LOCAL client_min_messages
Adds SET LOCAL client_min_messages = WARNING; near the top of sql/object_reference.sql (line 11) to silence the %TYPE resolution NOTICEs the script's many column-type references otherwise spam during install. SET LOCAL (not SET) is the correct choice here since it scopes to CREATE EXTENSION's implicit transaction and won't leak into the installing session. test/expected/zzz_build.out diff is exactly the now-suppressed NOTICE lines disappearing — matches the commit's intent.

a9d4a67 — Drop _object_oid.object_oid
This directly implements the non-blocking suggestion from the prior review round: since _object_oid__add's insert always set object_oid := objid and nothing else wrote to the table, the column and its objid_must_match CHECK were tautological. The commit cleanly removes both, updates the two views (_object_v, _object_v__for_update) to stop projecting object_oid, and repoints test/sql/base.sql's assertion at objid directly. Since _object_oid is a private _object_reference-schema table, there's no external API concern. test/expected/base.out/zzz_build.out diffs are consistent with a straight make results regeneration, not hand-edited.

Both commits are small, self-contained, and consistent with the rest of the stack (frozen sql/object_reference--0.1.0.sql and META files remain untouched, per repo convention). No new issues found — this addresses the only open item from the previous review pass.
· Branch: reconcile-object-functions

[github-actions]

Claude finished @jnasbyupgrade's task in 1m 36s —— View job

---

Claude finished @jnasbyupgrade's task in 3m 12s ——

Reviewing latest changes

• Gather context (diff since last review, CI status)
• Review new commit(s)
• Post review feedback

CI is green (CI run 31436646037 succeeded). Since the last review pass, one new commit was added:

3c06f4d — Fix multi-line comment to use block format, not consecutive -- lines

Converts the SET LOCAL client_min_messages = WARNING; explanation at the top of sql/object_reference.sql from a run of -- line comments to a proper /* ... */ block, per CODE_STYLE.md's rule against multi-line -- comments. The test/expected/zzz_build.out diff is exactly the resulting 2-line shift in the \echo-derived psql:test/temp_load.not_sql:NNN line numbers — consistent with a straight make results regeneration, not hand-edited drift.

This is a pure style/formatting fix with no behavioral change. No issues found — this addresses everything raised in prior review rounds (the object_oid redundancy from the first pass was already fixed in a9d4a67, confirmed in the second pass).

· Branch: reconcile-object-functions

[github-actions]

Claude finished @jnasbyupgrade's task in 1m 48s —— View job

---

Reviewing latest changes

• Gather context (diff since last review, CI status)
• Review new commit(s)
• Post review feedback

CI is green (CI run 31439158676 succeeded, including the PostgreSQL 12 leg). Since the last review pass, one new commit was added:

683cb57 — Add README.md and fix META packaging metadata (fixes #35)

• README.md (new, 333 lines): installation, roles/security, key concepts, and full public API reference. Spot-checked every documented routine (object__getsert/object__getsert_w_group_id, object__describe/object__identity/object__cleanup, object_group__*, object__dependency__add/object_group__dependency__add, capture__start/stop/get_current/get_all, post_restore, unsupported/untested (+_srf variants)) against sql/object_reference.sql — all exist and match the documented signatures, so the "verified against current schema" claim in the commit message holds up.
• META.in.json/META.json: adds cat_tools: 0 to runtime.requires, matching object_reference.control's requires = 'cat_tools' (line 5) — this was a real gap, since object_reference genuinely can't build/run without it. decibel/* → Postgres-Extensions/* URL fixes are consistent — confirmed no other decibel org references remain anywhere in the tree outside the vendored pgxntool/ directory. The PostgreSQL build-requirement bump from 9.5.0 → 12.0.0 matches the actual floor the CI matrix exercises (jobs run PG 12–18).
• Both META files were edited in lockstep (META.in.json is the hand-maintained source, META.json the generated artifact) — no drift between them.

One thing worth a mention purely for context: this commit touches META.in.json/META.json, whereas earlier review rounds on this PR noted those files were untouched — that's expected here since this commit is fixing pre-existing packaging metadata issues (tracked as #35) rather than being part of the reg*-column/object-info-function reconciliation work; it doesn't conflict with the "frozen sql/object_reference--0.1.0.sql" convention, which this commit correctly leaves alone.

No issues found. This is a clean, well-scoped documentation and metadata fix.
· Branch: reconcile-object-functions

[github-actions]

Claude finished @jnasbyupgrade's task in 1m 24s —— View job

---

Reviewing latest changes

• Gather context (diff since last review, CI status)
• Review new commit(s)
• Post review feedback

CI is green (CI run 31440232562 succeeded). The branch was rebased onto master since the last review (commit hashes for previously-reviewed commits changed, e.g. 683cb57 → 7f58f0e), and one new commit was added on top:

f4bb263 — Restore EXCLUDED CODE marker lost during the linter-merge rebase conflict

During the rebase onto master (which now includes #16's SQL style linter), a conflict in the commented-out cat_tools schema-restriction DO $$ ... $$ block (sql/object_reference.sql line 15) was resolved by keeping a plain /* instead of master's /* EXCLUDED CODE: ... */ marker convention. This one-line fix restores the marker:

-/*
+/* EXCLUDED CODE: schema-restriction check below not currently enforced

I checked this against the linter's existing usage elsewhere in the repo (test/deps.sql:5, test/sql/object_group.sql:87,112, test/helpers/object_table.sql:165, test/sql/capture.sql:137) — the /* EXCLUDED CODE[: reason] format matches the established convention exactly, and this is the only such block in sql/object_reference.sql. Without the marker the linter would demand " * "-prefixed lines throughout the commented-out block, so this is a correct, minimal fix (1 file, +1/-1) with no behavioral change to the extension itself.

Confirmed the overall PR diff against master (14 files, +536/-161) still matches exactly what was described across all prior review rounds — this rebase only replayed existing commits plus this fix, it didn't introduce scope changes.

No issues found.
· Branch: reconcile-object-functions

[github-actions]

Claude finished @jnasbyupgrade's task in 1m 11s —— View job

---

Reviewing latest changes

• Gather context (diff since last review, CI status)
• Review new commit(s)
• Post review feedback

CI is green (CI run 31441782137 succeeded). Since the last review pass, one new commit was added:

fad854d — Move client_min_messages suppression from the extension script to the test

• Removes the SET LOCAL client_min_messages = WARNING; block from sql/object_reference.sql (added two review-rounds ago), on the reasoning that an extension's install script shouldn't unilaterally change a session setting for its caller, even scoped via SET LOCAL.
• Moves the suppression into test/sql/zzz_build.sql instead, right before the raw \i test/temp_load.not_sql load — this is the one test path that bypasses CREATE EXTENSION (loading the script directly via \i), so it's the only place these %TYPE-resolution NOTICEs actually surfaced. I confirmed test/pgxntool/setup.sql:2 opens with BEGIN;, so the SET LOCAL here correctly reverts at that transaction's end rather than leaking — the comment's claim checks out.
• test/expected/zzz_build.out diff is exactly the line-number shifts you'd expect from the extension script shrinking by 10 lines (e.g. temp_load.not_sql:186→176, 435→425, etc.) — consistent with a straight make results regeneration, not hand-edited drift.
• Full branch diff vs. origin/master still matches the PR description exactly (14 files, +538/-164).

This is a good call — it correctly narrows the workaround's blast radius to the one test that needed it, rather than having the shipped extension mutate a session-level setting on every install. No issues found.
· Branch: reconcile-object-functions