Pg patch review

pg-patch-review — multi-agent comprehensive PG patch review

The deep-review counterpart to the manual seven-phase review-checklist. The cf6402 validation run on 2026-06-02 proved the corpus + skills compose for a one-author review; this skill turns that loop into a repeatable multi-agent pipeline.

When to use this skill vs review-checklist

Situation	Use
CF entry you intend to mail a real review on	this skill
Quick "is this even sane?" pass on a patch	review-checklist (manual seven-phase)
Self-review of your own patch before mailing	patch-submission (which already invokes review-checklist)
Generic non-PG GitHub PR review	neither — this is PG-specific

The two are NOT redundant: review-checklist is the seven-phase scaffold each critic agent applies inside its assigned slice. This skill is the orchestration layer above that scaffold (project discovery + parallel critic fan-out + synthesis).

Companion skills (each critic loads what it needs)

• review-checklist — seven-phase scaffold each critic walks inside its slice
• wal-and-xlog — WAL records / redo / hint bits (used by breaking-change critic)
• locking — lock primitive choice + acquisition order (architecture critic)
• catalog-conventions — pg_proc / OID assignment (breaking-change critic)
• testing — regress vs isolation vs TAP vs module (test-coverage critic)
• coding-style — pgindent / include order / C99 subset (style critic)
• commit-message-style — upstream PG commit-message style (style critic, AND used by synthesizer for review-email tone)
• memory-contexts — palloc placement / context lifetimes (architecture critic when relevant)
• error-handling — ereport / SQLSTATE choices (style critic when relevant)

Inputs

• A patch reference (required), one of:
  • CF number: 6402, #6402, or CF 6402
  • GitHub PR number on postgres/postgres: pr 19234 or #19234
  • A local .patch file or directory of patches: /tmp/v3-0001-foo.patch
• Optional flags (from the slash command):
  • --skip-build — patch already applied + built in dev/, skip Phase 0
  • --no-flaky-isolation — skip the isolation suite (macOS sometimes flakes)
  • --subsystem=<name> — hint for which knowledge/subsystems/*.md to load first (e.g. --subsystem=access-nbtree for CF #6402)

Output

• A draft review email at sessions/<date>-cf<N>-review.md (or sessions/<date>-pr<N>-review.md) using the PG-hackers house style.
• A per-critic appendix at the bottom of the same file with the raw findings each sub-agent produced (blocking / warning / suggestion).
• A run log appended to that session file: which patches applied, which tests ran, which subsystem docs the critics consulted, wall time.
• A dev/ branch named cf<N>-review (or pr<N>-review) with the patch applied — disposable after the review is sent.

When NOT to invoke

• Patch already merged upstream — use git log + corpus walkthrough instead.
• Patch is your own work — use patch-submission (it invokes the same critics but on the self-review path).
• Patch is non-PG — wrong skill.

Method — five stages

Stage 0 — mechanical pre-amble (5-10 min)

Done by the /pg-review slash command. If invoked directly (without /pg-review), this skill does it inline before stage 1. See .claude/commands/pg-review.md for the exact recipe.

Inline Stage 0 (when /pg-review wasn't used) — minimum commands:

1. cd dev && git checkout master && git pull
2. git checkout -b cf<N>-review (or pr<N>-review)
3. Fetch the patch (CF: curl the v<N> from the CF entry; PR: gh pr checkout <N>; .patch file: copy in).
4. git am /path/to/v*.patch (apply all v<N> hunks in order).
5. ninja -C build-debug install — must be warning-clean.
6. meson test --no-rebuild regress/regress — record pass/fail.
7. meson test --no-rebuild --suite isolation — record pass/fail.
8. git diff --name-only HEAD~<N>..HEAD — capture the touched-files list.
9. Note any pre-existing flakes (e.g. macOS recovery/040_standby_failover_slots_sync).

The output of this stage is:

• dev/ on branch cf<N>-review (or pr<N>-review) with the patch applied.
• A built binary (ninja install clean, no new warnings).
• meson test --no-rebuild regress/regress result (pass/fail per test).
• meson test --no-rebuild --suite isolation result.
• A list of files the patch touches (git diff --name-only HEAD~N..HEAD).
• A note on any pre-existing flakes (e.g. macOS recovery/040_standby_failover_slots_sync).

If stage 0 fails (patch doesn't apply, build breaks, regress fails) — stop. Report to the user. The patch is Waiting on Author mechanically; no point spending tokens on the critics.

Stage 1 — project discovery (orchestrator, ~5 min)

The main agent does this once before fanning out:

1. Touched files → touched subsystems. From the file list in stage 0, map each file to its knowledge/subsystems/<name>.md parent. Example: src/backend/access/nbtree/nbtpage.c → knowledge/subsystems/access-nbtree.md. Use the §2 "File map" section of each subsystem doc to confirm. If a touched file doesn't appear in any subsystem doc, NOTE THAT — the review must call out that an uncovered area was changed.

2. Load per-file docs. For each touched file, look up knowledge/files/<path>.md if present. These have INV-* invariants and per-function cites the critics will rely on.

3. Identify the "claims" the patch makes. Read the patch's commit message + cover letter (or the CF/PR description). Each claim that says "fixes X", "implements Y", "no behavior change", "doesn't touch on-disk format", "is purely a refactor" becomes a CHECK item the critics will validate.

4. Build the dispatch block. Produce a small reference block that every critic sub-agent receives in its prompt:

   Patch: <CF#|PR#|path>
   Branch: dev/cf<N>-review @ <short-sha>
   Files touched:
     - src/.../X.c (per-file doc: knowledge/files/.../X.c.md)
     - src/.../Y.h (no per-file doc — flag if material)
   Subsystems touched:
     - access-nbtree (knowledge/subsystems/access-nbtree.md)
     - storage-buffer (knowledge/subsystems/storage-buffer.md)
   Claims the patch makes (verbatim from commit msg):
     - "Replace duplicated metapage sanity checks..."
     - "No behavior change."
     - "Restores symmetry with _bt_getroot..."
   Stage-0 test result: regress 245/245 pass, iso 129/129 pass,
     1 unrelated TAP failure recovery/040_* (macOS flake)
   

5. Spot-check 3-5 file:line cites in the relevant subsystem doc against current source/ — if drift > 10% (cites stale by more than ~20 lines or naming since-removed symbols), STOP and tell the user the corpus needs an hf(corpus): refresh before this review.

Stage 2 — fan out 5 critic sub-agents IN PARALLEL

Launch all five (A-E; Critic E added 2026-06-12 from Phase C) in a single message with parallel tool calls. Each gets the dispatch block from stage 1 + its assigned slice. Each is read-only — sub-agents do NOT edit files or commit. Each returns a structured finding list.

Use the Agent tool with subagent_type: "general-purpose" for each. Estimate ~10-20 min wall time for all five to complete (in parallel).

Critic A — Architecture & invariants

Scope: does the patch fit the subsystem's existing invariants?

Loads: knowledge/subsystems/<each touched subsystem>.md + relevant per-file docs + review-checklist Phase 6 (Architecture).

Checks:

• Does the patch violate any INV-* invariant tagged in the subsystem doc? Cite the tag.
• Does the patch's locking match the subsystem's lock-order discipline? (E.g. for nbtree: buffer locks coupled in left-to-right order; for heap: buffer-pin-before-buffer-lock; for replication: never hold ProcArrayLock across...)
• Does the patch interact with parallel query / extensions / logical replication in any way the corpus warns about?
• Does the patch's claim of "no behavior change" hold up under inspection? (For refactors: every removed line must have an equivalent in the replacement.)
• Provenance check for any helper/struct the patch touches: run git -C source log -S '<symbol>' --oneline | head -5 to find when the symbol was introduced. A symbol that's existed for years + has multiple existing callers is a safer refactor target than a symbol introduced last release. Surfaces both "this is a long- overdue cleanup" and "this is racing an in-flight feature".

Output: structured findings list. Each item:

File: <path:line>
Severity: blocking | warning | suggestion
Invariant: INV-... (if applicable) or "no INV cited"
Description: what and why
Suggestion: proposed fix or question to the author

Critic B — Breaking-change scan

Scope: does the patch touch anything backwards-incompatible?

Loads: wal-and-xlog, catalog-conventions, review-checklist Phase 6 (Architecture — the ABI bullets), and the subsystem docs' §5 "Invariants and breaking-change surfaces" sections.

Checks:

• On-disk page format change? (pd_* fields, opaque-area layout.)
• WAL record change? (New record, new info byte, existing record extended.) If yes, is XLOG_PAGE_MAGIC bumped?
• Catalog change? (pg_proc.dat, pg_type.dat, etc.) If yes, is CATALOG_VERSION_NO bumped? Are new OIDs assigned?
• Public API / extension ABI? (Anything in src/include/.) Inline functions / macros there count. If touching back-branchable code: new struct members must go at the end; no signature changes on exported functions.
• Replication protocol change? (libpq wire format, walsender output plugins, logical decoding output formats.)
• pg_dump impact? (Any new schema object.)

Output: same structured list. For each blocking break, name the upgrade/backpatch story the author needs to provide.

Critic C — Test coverage

Scope: is the patch tested adequately for what it claims?

Loads: testing skill + the touched subsystems' "test surface" sections + src/test/ for the existing coverage of the touched code.

Checks:

• Does the diff include src/test/ changes? If not, is the claim "pure refactor, no new behavior, existing tests cover" defensible?
• For a refactor: does at least one existing test exercise the code path being refactored? (Find by git grep for the function name in src/test/.) If not, the "existing tests cover" claim is weak.
• For new behavior: does the new test ACTUALLY fail without the code change? (The classic "test passes both with and without the patch" bug.) Sub-agent can't easily verify this without re-running tests twice — instead it flags this as a question for the author or for manual follow-up.
• Corner cases: NULL, empty input, max-length, encoding edges, concurrent calls, parallel-worker visibility, replication catchup. Sub-agent enumerates which apply to this patch's surface.
• Isolation tests needed? Concurrent-modification scenarios?
• TAP tests needed? Multi-node, recovery, replication, crash-recovery scenarios?

Output: same structured list. The "blocking" bar here is whether the patch's correctness claim is mechanically testable from what's in the diff.

Critic D — Style & commit-message

Scope: would a committer have to fix the format before applying?

Loads: commit-message-style + coding-style + review-checklist Phase 5 (Coding review) + Phase 7 (Committer-readiness).

Checks:

• Patch filename: vN-NNNN-<title>.patch?
• Commit message: imperative title, no period, ~76-col wrap, no emoji, no Co-Authored-By (forbidden upstream), Author/Reviewed-by trailers if relevant, Discussion: link if relevant.
• Code style: matches surrounding module (camelCase vs snake_case), no leftover debug elog, no commented-out code, no new compiler warnings flagged in stage 0.
• Error messages follow the message style guide (capitalization, no period on errmsg, separated errdetail/errhint).
• git diff --check clean? (Trailing whitespace, broken tab/space mix.)
• pgindent clean? (May not be runnable locally — note pg_bsd_indent install state; CI will catch.)

Output: same structured list. Most items here are suggestion or warning; only fundamentally broken style is blocking.

Critic E — Reviewer-reflex probes (added 2026-06-12 from Phase C)

Scope: does the patch trigger any of the persona-driven reflexes the corpus has documented but the generic critics A-D don't encode?

Loads: knowledge/calibration/gap-catalog.md (the 11-item catalog) + knowledge/personas/<name>.md for each persona named in items 4-11 that triggers + knowledge/personas/committer-map.md + knowledge/personas/domain-ownership.md (item 11 cross-reference).

Checks (each maps 1:1 to a catalog item):

• Cleanup-on-early-return tracing (catalog #4). Scan the diff for a new return statement added inside a function whose entry block owns a resource handle (z_stream, BufFile, FileFd, MemoryContext, Relation, LWLock). If found, surface "trace cleanup path under the new error return — does <resource>_destroy() / _close() / _release() run on this branch?". Driver: daniel-gustafsson.md errorhandling discipline.

• Multibyte/encoding interaction (catalog #5). Scan the diff for byte-walking patterns (*p++, *input++, manual for loops over varlena/text/cstring) OR size caps on text-processing primitives. If found, surface "enumerate worst-case per encoding (UTF-8 documented, GB18030, EUC_JP, EUC_KR, EUC_CN, EUC_TW); cite the Unicode TR / SpecialCasing.txt entry for any UTF-8-specific bound". Driver: noah-misch.md §4 + jeff-davis.md Unicode standard fidelity.

• Subsystem-local cap discoverability (catalog #6). Scan the diff for a new #define in a contrib/*/*.c file (not header). If found, surface "move to <subsystem>.h if a public-style cap; cite the precedent constant in the same area (e.g. LQUERY_MAX_LEVELS for ltree_io.c)". Driver: peter-eisentraut.md style reflex.

• "Third state" cross-check for binary-format changes (catalog #7). Scan the diff for changes in how a flag-bit, version-bit, or layout-bit is interpreted. If found, surface "enumerate the third state: bit set but structure invalid, OR bit unset but structure looks valid — what handles each?". Driver: heikki-linnakangas.md binary-format reflex.

• injection_points reproducer for DoS / scratch-allocation / race claims (catalog #8). Scan the commit-message + COVER for phrases like "prevents N MB scratch", "N GB allocation", "fixes a race", "OOB read", "amplification". If found AND the patch has no src/test/modules/injection_points/ change, surface "include an injection_points measurement at the allocation / race-windowed boundary; the structural argument is not enough on a security claim". Driver: noah-misch.md §5.

• Hot-path branch-prediction / micro-benchmark (catalog #9). When the patch touches a function in src/backend/utils/adt/*, src/backend/access/{heap,nbtree}/, src/backend/optimizer/, or similar query-evaluator path AND adds a new guard check, surface "include micro-benchmark numbers confirming the guard is in the unlikely branch and adds <1% overhead on typical inputs". Driver: thomas-munro.md + heikki-linnakangas.md performance reflex on hot paths.

• Symmetric-check refactor for N-entry-point guards (catalog #10). Diff scan for 3+ near-identical added blocks (heuristic: same if/ereport/ereturn pattern at 3+ places). If found, surface "consider a shared inline helper <module>_check_<thing>() to keep entry points symmetric". Driver: peter-eisentraut.md symmetric-primitives reflex.

• Persona-aware backpatch routing (catalog #11). When COVER claims back-patching AND the predicted top committer for the touched subsystem (from domain-ownership.md top-committer column) has a 24mo backpatch rate < 5% (compute from committer-map.md or /usr/bin/git log --author=<name> --since= '2yr' --pretty=%s | grep -ciE 'back.?patch' ratio), surface "X doesn't backpatch in 24mo; the realistic v16/v17/v18 landing committer is Y (from domain-ownership.md reviewer column — pick the highest-ranked committer who backpatches at ≥10%). CC them on the thread.". Driver: Peter Eisentraut row in committer-map.md.

Severity rules for Critic E:

• Catalog #1-#3 are NOT this critic's job — they live in review-checklist Phase 0 (gates that block before the patch enters the critic fan-out).
• Catalog #4, #5, #7, #8 are warning (sometimes blocking if the COVER doesn't even acknowledge the question).
• Catalog #6, #9, #10, #11 are suggestion by default — they improve the patch but don't block.
• Catalog #5 escalates to blocking if the patch caps a text primitive AND no per-encoding analysis is in the COVER — that's a real correctness gap (e.g. SP2 had this; the 3× UTF-8 bound may not hold for GB18030).

REJECT-track escalation (M4). When Critic E surfaces 3+ blocking-severity findings from the catalog AND the context-awareness signal (engagement class contested OR a documented INV-* invariant is foreclosed), the critic's output should explicitly recommend a REJECT-A Stage-3 verdict rather than "Waiting on Author". The Stage-3 orchestrator then decides between REJECT-A (the grade above), REJECT-B (acknowledge that the critic may have missed a concern), or downgrade to non-REJECT if the findings don't actually compose to a design-level NACK. Critic E recommends; Stage 3 decides.

Critic E severity matrix at a glance:

Catalog #	Probe	Default	Escalates to blocking when
#4	Cleanup-on-early-return	warning	COVER doesn't acknowledge the cleanup question
#5	Multibyte / encoding	warning	text-primitive cap added with no per-encoding analysis in COVER
#6	Subsystem-local cap discoverability	suggestion	—
#7	"Third state" binary-format	warning	COVER doesn't enumerate bit-set-but-invalid AND bit-unset-but-looks-valid cases
#8	injection_points reproducer	warning	structural argument on a security claim with no injection_points test
#9	Hot-path micro-benchmark	suggestion	—
#10	Symmetric-check refactor	suggestion	—
#11	Persona-aware backpatch routing	suggestion	—

REJECT-track escalation: 3+ blocking from this table AND context- awareness signal (engagement class contested OR foreclosed INV-*) → recommend REJECT-A to Stage 3.

Output: same structured-finding list as critics A-D, plus an optional recommend_verdict: REJECT-A | REJECT-B field when the escalation rule above triggers.

Stage 3 — orchestrator consolidates (10 min)

Critic-E recommendation vs orchestrator verdict. Critic E may emit recommend_verdict: REJECT-A | REJECT-B when its catalog-item threshold (3+ blocking findings + context-awareness signal) fires. The orchestrator at Stage 3 decides; Critic E recommends. The orchestrator may downgrade the recommendation to "Waiting on Author" if the findings, in aggregate, do NOT compose to a design-level NACK. Critic E's recommendation is one input, not the verdict.

The main agent gathers all four critics' outputs and:

1. Deduplicates. Two critics may flag the same issue from different angles — merge into one finding with both rationales.

2. Resolves conflicts. If critic A says "this is fine" but critic B says "this breaks ABI", the orchestrator re-reads both and picks the stronger argument. Cite both in the merged finding.

3. Severity prioritization. Group findings into:

   • Blocking (must fix before commit; flip CF to "Waiting on Author")
   • Warning (should fix or justify)
   • Nits / suggestions (take or leave)
   • Open questions (need author input)

4. Verdict. Decide one of:

   • Ready for Committer
   • Waiting on Author (blocking issues)
   • Needs more info from author (open questions dominate)
   • REJECT-A — design fundamentally wrong, all critical problems identified, alternative proposed. The right deliverable is a thread reply explaining the rejection with cites; saves community cycles. Use this when the patch is in contested engagement class or the Context-awareness probe (from pg-feature-plan) flagged it.
   • REJECT-B — design wrong, but you missed at least one major concern that a critic from the community will raise. Solid but incomplete; send the reply, acknowledge gaps.
   • REJECT-C — rejected for the wrong reasons OR rejected when the proposal is actually sound. STOP — escalate to user before posting. Likely you need to re-run with looser priors or load more corpus.

   M4 origin: knowledge/shadow-implementations/money-fx-exchange/skill-gaps.md. The REJECT-A/B/C grades parallel the A-F grade rubric on non-REJECT outcomes — they're not lesser verdicts, just the right shape for proposals that shouldn't proceed.

Stage 4 — synthesize the review email

Use the commit-message-style skill's tone rules — imperative, plain text, no HTML, no emoji, ~76 col wrap. The review email lives at sessions/<date>-cf<N>-review.md (or pr<N>) and has this shape:

To: [email protected]
Cc: <author> <author@email>
Subject: Re: [PATCH v<N>] <patch subject>

<one-line summary of where the patch stands>

<one or two paragraphs of the high-level read — what the patch does, why
the corpus thinks it's coherent (or not). Cite specific anchors where
relevant: nbtpage.c:407, knowledge/subsystems/access-nbtree.md §4.>

Blocking issues:
  1. <one-line summary>
     <2-4 lines of context + concrete ask>
  2. ...

Warnings / consider:
  1. ...

Nits, take or leave:
  1. ...

Open questions:
  1. ...

Testing performed:
  - git am: <clean | rejected hunk in X>
  - ninja install: <clean | warnings: ...>
  - meson test regress/regress: <NNN subtests, all pass | failed: ...>
  - meson test --suite isolation: <NNN subtests, all pass | failed: ...>
  - Patch base: <upstream-master short-sha>

<closing line: "I think this is ready for a committer" / "Marking
Waiting on Author pending the items above" / etc.>

Regards,
[Reviewer]

Below the email in the SAME session file, append:

---

## Per-critic raw findings

### Critic A — Architecture & invariants
<paste the sub-agent's structured list>

### Critic B — Breaking-change scan
<paste>

### Critic C — Test coverage
<paste>

### Critic D — Style & commit-message
<paste>

## Stage 0 mechanical log
- Patch source: <URL or path>
- Base ref: <upstream-master sha>
- Apply: <git am output summary>
- Build: <ninja install summary>
- regress: <pass/fail counts + duration>
- isolation: <pass/fail counts + duration>
- Targeted suites: <if any>
- Pre-existing flakes encountered: <list, with dismissal rationale>

## Wall time
- Stage 0: <min>
- Stage 1: <min>
- Stage 2 (4 critics in parallel): <max of the four, plus orchestration overhead>
- Stage 3: <min>
- Stage 4: <min>
- Total: <min>

Boundaries vs other skills

• review-checklist (the eight-phase scaffold — Phase 0 added 2026-06-12 for reviewer-reflex gates): each critic walks the relevant phase of it. This skill orchestrates five critics doing that in parallel (A-E; E added 2026-06-12 from Phase C) and synthesizes. Don't bypass review-checklist's phase definitions — extend them.
• patch-submission: the self-review counterpart. If you're reviewing YOUR OWN patch before mailing, use that — it invokes the same critics but on the pre-submission path.
• commit-message-style (upstream PG): used by the synthesizer for the review-email tone AND by critic D for judging the patch's commit message.
• meta-commit-style (postgres-claude): does NOT apply to the review email (which goes to pgsql-hackers, not into postgres-claude). It WOULD apply to the session-log commit (if any) and to the STATE.md update.

What to escalate to the user mid-review

• Stage-0 fail (patch doesn't apply, build breaks, regress fails): stop, report, ask whether to send a "rebase needed" reply or to skip.
• Corpus drift detected in stage 1 (cites stale > 10%): stop, ask the user whether to (a) refresh the corpus first via a separate hf(corpus): commit (per Rule R9 of .claude/rules/pg-implement-discipline.md — corpus fixes are their own commits in the meta-repo), or (b) proceed with a "best-effort against possibly-stale docs" caveat noted in the review email's "Testing performed" block.
• Touched file not in any subsystem doc: don't stop; note in the review email's "Testing performed" block that this area is uncovered by the corpus. After the review, file a followup to document that subsystem.
• Two critics genuinely disagree after orchestrator consolidation: ask the user to break the tie before drafting the email.

Style notes

• The review email is the deliverable; everything else is working notes. Make the email scannable — bullets, no walls of prose.
• Cite specific file:line anchors in the email when relevant. The validation run proved this is what makes a review feel grounded vs generic.
• Distinguish blocking from nits in EVERY review. "Needs more tests" is not blocking unless the patch's correctness claim depends on the missing test.
• For performance-impacting patches: ask for pgbench numbers with exact recipe (hardware, build flags, run count, master baseline). Numerical claims without a recipe get bounced.
• If invoking via the /pg-review slash command, the command already did stage 0 — skip ahead to stage 1.

Where the artifacts live

• Review email + appendices: sessions/<date>-cf<N>-review.md in postgres-claude/ (this repo).
• Patch branch: dev/cf<N>-review (the mutable PG clone). Disposable after review is sent.
• No knowledge/ writes by this skill — if the review surfaces a corpus gap, file a follow-up hf(corpus): commit separately (per R10 of .claude/rules/pg-implement-discipline.md).

Validation reference

The 2026-06-02 v0 review of CF #6402 [unverified: session log not preserved in sessions/ at the time of this writing] is the calibration target — re-running THIS skill against that patch should reproduce a review of comparable quality (same draft conclusion, same blocking-vs-nit split) in less wall time than the v0 manual walk. A future preserved-and-named calibration session can replace this paragraph.

Cross-references

• .claude/skills/review-checklist/SKILL.md — the eight-phase scaffold each critic walks; Phase 0 hosts the REJECT-A/B/C grade rubric this skill's Stage 3 verdict consumes.
• .claude/skills/patch-submission/SKILL.md — invokes this skill in --self mode for the self-review path.
• .claude/skills/pg-feature-plan/SKILL.md — supplies the Context-awareness probe + Thread-engagement classification that drive Critic E's REJECT-track escalation.
• .claude/skills/commit-message-style/SKILL.md — Critic D + synthesizer use this for upstream PG commit-message format.
• .claude/skills/coding-style/SKILL.md — Critic D style check.
• .claude/skills/testing/SKILL.md — Critic C test-coverage check.
• .claude/skills/wal-and-xlog/SKILL.md, .claude/skills/catalog-conventions/SKILL.md — Critic B breaking-change scan.
• .claude/skills/locking/SKILL.md, .claude/skills/memory-contexts/SKILL.md, .claude/skills/error-handling/SKILL.md — Critic A architecture check.
• knowledge/calibration/gap-catalog.md — items 4-11 source Critic E's eight reflex probes.
• knowledge/personas/*.md — Critic E loads relevant persona docs per probe.
• knowledge/shadow-implementations/money-fx-exchange/skill-gaps.md — M4 origin (REJECT-A/B/C verdict).
• .claude/commands/pg-review.md — slash-command wrapper that runs Stage 0 inline.