src/lib/rag-retrieval-variants.ts (1)

169-172: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

New admission/community branch overlaps with the existing comparison branch (Lines 184-187).

Both blocks add "admission community patients"-style variants for overlapping query shapes (single-branch admission+community vs. comparison admission+discharge). The seen set dedupes, so this is not a correctness bug, but the near-duplicate logic adds maintenance surface.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/lib/rag-retrieval-variants.ts` around lines 169 - 172, The
admission/community variant block in rag-retrieval-variants overlaps with the
existing admission/discharge comparison branch and duplicates similar addVariant
logic. Consolidate the shared “admission community patients”-style variant
generation into one helper or shared conditional path used by both branches, and
keep the unique query-shape checks in the surrounding logic so the variants are
added from a single source of truth.

.github/workflows/live-drift.yml (1)

46-50: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

Consider bumping setup-node to the latest major.

actions/setup-node@v5 works, but v6.x is current (node24-based). Not urgent for a weekly/manual workflow.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/live-drift.yml around lines 46 - 50, The workflow is still
pinned to an older major of actions/setup-node; update the setup-node usage in
the live-drift workflow to the current major release. Locate the existing
actions/setup-node reference and bump it to the latest supported major while
keeping the current node-version-file, cache, and cache-dependency-path settings
unchanged.

src/lib/registry-corpus.ts (2)

431-447: 🗄️ Data Integrity & Integration | 🔵 Trivial | 💤 Low value

bestEffortReembedRegistryRecordAfterEdit reports zero counts even after partial writes.

When reembedRegistryRecordAfterEdit throws mid-batch (see the batching loop above), earlier batches may have already written documents/chunks, but the caught "failed" result here always reports documentCount: 0, chunkCount: 0, masking that partial writes occurred. Combined with the batching rollback gap, this makes it hard for downstream callers to detect and reconcile partial index state.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/lib/registry-corpus.ts` around lines 431 - 447,
bestEffortReembedRegistryRecordAfterEdit currently returns a failed result with
zero counts even when reembedRegistryRecordAfterEdit may have already written
some documents/chunks before throwing. Update the error path in
bestEffortReembedRegistryRecordAfterEdit to preserve/report the actual partial
documentCount and chunkCount from the attempted re-embed instead of hardcoding
zeros, using the existing reembedRegistryRecordAfterEdit flow and
RegistryCorpusEmbedResult shape so downstream callers can detect partial writes.

---

360-406: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Near-identical slug lookup logic duplicated three times.

reembedClinicalRegistryRecordBySlug, reembedMedicationRecordBySlug, and reembedDifferentialRecordBySlug repeat the same disabled-check → query-by-owner/slug(/kind) → maybeSingle() → not-found/error pattern. Consider extracting a shared helper (table name + optional kind filter + row caster) to reduce triplication and ease adding future corpus kinds.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/lib/registry-corpus.ts` around lines 360 - 406, The slug-based re-embed
flow is duplicated across reembedClinicalRegistryRecordBySlug,
reembedMedicationRecordBySlug, and reembedDifferentialRecordBySlug. Extract the
shared disabled-check, owner/slug lookup, optional kind filter, maybeSingle
handling, and not_found/error logic into a common helper, then have each
function pass its table name and row-specific embed/cast behavior to that
helper. Keep the existing public function names but remove the repeated query
boilerplate.

src/lib/rag.ts (2)

2622-2624: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

No direct test coverage found for the new admission/community gate.

Neither tests/rag-answer-fallback.test.ts nor tests/rag-routing.test.ts (the test files included in this cohort) exercise admission_community_requires_title_rescue or missing_admission_community_title_support. If coverage exists elsewhere (e.g. an existing rag.test.ts) this can be disregarded; otherwise consider adding a regression test for a query matching admission+community intent with/without matching title support.

Also applies to: 2666-2685, 2954-2963

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/lib/rag.ts` around lines 2622 - 2624, Add regression coverage for the new
admission/community routing gate in rag.ts, especially the branch in the
fast-path decision that returns admission_community_requires_title_rescue and
the missing_admission_community_title_support path. Update an existing relevant
test suite such as rag-answer-fallback.test.ts, rag-routing.test.ts, or
rag.test.ts to exercise a query that matches admission+community intent both
with and without title support, and assert the routing reason and fast-path
behavior in the affected helpers like hasAdmissionCommunityLookupIntent and
hasAdmissionCommunityTitleSupport.

---

2666-2685: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick win

Fallback title-match branch is narrower than the titleHit branch.

When match_explanation.titleHit/labelHit is set, support is granted via a general \badmission\b + \bcommunity\b check on the title (order-independent). When it isn't set, the fallback only matches three hardcoded literal phrases ("admission of community patient", "admission community pt", "admission to discharge for community"). A genuinely on-topic title with different word order or phrasing (e.g. "Community Patient Admission Pathway") would fail the fallback and incorrectly report missing title support, even though the same general regex check (used in the titleHit branch) would have matched it. This only degrades to the slower/structured path rather than causing a wrong answer, but the asymmetry looks unintentional.

♻️ Suggested simplification

 function hasAdmissionCommunityTitleSupport(results: SearchResult[]) {
   return results.slice(0, 5).some((result) => {
-    if (result.match_explanation?.titleHit || result.match_explanation?.labelHit) {
-      const title = normalizeDocumentAliasText(`${result.title} ${result.file_name}`);
-      return /\badmission\b/.test(title) && /\bcommunity\b/.test(title);
-    }
-    const title = normalizeDocumentAliasText(`${result.title} ${result.file_name}`);
-    return (
-      title.includes("admission of community patient") ||
-      title.includes("admission community pt") ||
-      title.includes("admission to discharge for community")
-    );
+    const title = normalizeDocumentAliasText(`${result.title} ${result.file_name}`);
+    return /\badmission\b/.test(title) && /\bcommunity\b/.test(title);
   });
 }

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/lib/rag.ts` around lines 2666 - 2685, The fallback branch in
hasAdmissionCommunityTitleSupport is too narrow compared with the
titleHit/labelHit branch, so align both paths by using the same general
normalizeDocumentAliasText plus /\badmission\b/ and /\bcommunity\b/ title check
for all results instead of only matching a few hardcoded phrases. Keep the
existing result slicing and title/file_name concatenation, but remove the
asymmetric literal-only fallback so Community Patient Admission Pathway-style
titles are recognized consistently.

✅ Autofix completed

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".