Score tests on the impact path and prefer generated OpenAPI clients

[Modsofthenation]

Why

Merge confidence, N+1, OpenAPI stitching, and Django boot were a little too eager: off-path tests could cover a sink, CharField accesses looked like relations, string fetch competed with generated clients, weak churn notes could knock HIGH down to MEDIUM, and django.setup() mutated the reviewer process.

What changed

• Confidence counts TESTED_BY only when both ends sit on the impact subgraph (including two-hop producers on that path).
• Schema-aware N+1: drop known non-relation fields (CharField, etc.), keep unknown names at medium, raise to high when the schema says relation. Findings stay on extra — no new walkable edges. Residuals use the same filter.
• django.setup() runs in a subprocess so it is not process-global. Failures and malformed payloads still surface as django.setup() skipped: ….
• Generated OpenAPI clients stitch at high confidence. Client nodes are keyed by file+URL so generated/invoices.ts is not overwritten by api.ts. String URL matches stay dashed; if a generated client already covers the same template they are marked superseded_by_generated (~0.4).
• INDEX_REVISION is part of the sidecar digest so identity changes force an incremental rebuild.
• Evolution only flips HIGH→MEDIUM for serious notes (hotspot, silo, cross-context, temporal coupling), not a 2-commit complexity blip.
• Residuals name serializer/Zod overlap fields and superseded string stitches.

Tests

Unit coverage for impact-only tests, schema-aware N+1 (findings + residuals), boot JSON framing/malformed payload, generated vs inferred client stitch, and file-scoped generated clients. Vertical slice still expects InvoicePage (not MePage), suggested_reviewers == ['billing-team'], celery + dramatiq, confidence medium/low.

Adversarial review

• 3f5467e: generated and handwritten clients shared one node id (client:{url}).
• ab99fa0: CharField N+1 still leaked into residuals; extractor identity did not invalidate the index; boot JSON KeyError could crash the parent.

  

Summary by CodeRabbit

• Improvements

  • Improved detection and confidence scoring for relationship-based N+1 query findings.
  • Django project analysis now runs more reliably with isolated bootstrapping and clearer error reporting.
  • Prevented collisions between API clients from different files.
  • Automatically refreshes stale indexes when analysis logic changes.
  • Improved generated API client detection and route stitching, including supersession details.
  • Refined review confidence and residual reporting for impacted code.

• Tests

  • Expanded coverage for Django bootstrapping, relationship analysis, confidence scoring, indexing, and generated API clients.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 6ceddfe8-80ba-41e2-b846-31d31f1fa55f

📥 Commits

Reviewing files that changed from the base of the PR and between a5f4a75 and ab99fa0.

📒 Files selected for processing (12)

• src/loadpath/architecture/rules.py
• src/loadpath/extractors/django_boot.py
• src/loadpath/extractors/react.py
• src/loadpath/index.py
• src/loadpath/review/confidence.py
• src/loadpath/review/engine.py
• src/loadpath/stitch/openapi.py
• tests/e2e/test_brokers_and_django.py
• tests/unit/test_architecture_rules.py
• tests/unit/test_confidence.py
• tests/unit/test_index_and_stitch.py
• tests/unit/test_react_extractors.py

---

📝 Walkthrough

Walkthrough

The PR isolates Django boot execution, improves graph identity and index invalidation, makes N+1 analysis relationship-aware, scopes confidence testing to impacted graphs, and distinguishes generated OpenAPI clients during stitching.

Changes

Graph extraction and review updates

Layer / File(s)	Summary
Isolated Django boot src/loadpath/extractors/django_boot.py, tests/e2e/test_brokers_and_django.py	Django boot runs in a subprocess by default. Marked JSON payloads are parsed and malformed or failed boots produce residual errors.
Graph identity and index invalidation src/loadpath/extractors/react.py, src/loadpath/index.py, tests/unit/test_react_extractors.py	React client identifiers include source files. Generated and inferred metadata is recorded. Index digests include INDEX_REVISION.
Relationship-aware review analysis src/loadpath/architecture/rules.py, src/loadpath/review/confidence.py, src/loadpath/review/engine.py, tests/unit/test_architecture_rules.py, tests/unit/test_confidence.py	N+1 findings resolve relationship fields. Confidence and residual collection use normalized, impact-scoped data and serious evolution notes.
Generated-client stitching src/loadpath/stitch/openapi.py, tests/unit/test_index_and_stitch.py	Generated-client detection is centralized. Stitching records generated coverage, confidence, and superseded inferred edges.

Estimated code review effort: 4 (Complex) | ~45 minutes

Sequence Diagram(s)

sequenceDiagram
  participant DjangoBoot
  participant BootSubprocess
  participant PayloadParser
  participant ExtractedGraph
  DjangoBoot->>BootSubprocess: execute isolated Django boot
  BootSubprocess->>PayloadParser: emit marked JSON payload
  PayloadParser->>ExtractedGraph: reconstruct nodes and edges
  PayloadParser-->>DjangoBoot: return graph or residual error

Loading

Possibly related PRs

• Modsofthenation/PR-Reviewer#1: Extends the same architecture, indexing, review, Django boot, React extraction, and OpenAPI stitching areas.
• Modsofthenation/PR-Reviewer#2: Shares the N+1 detection implementation and architecture-rule tests.

Suggested reviewers: cursoragent

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch cursor/review-accuracy-ccb4

---

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