Lfx Self Serve Code Review

LFX Self-Serve Code Reviewer

In LFX, you audit the latest commit on the LFX Self-Serve branch against the repo's documented rule surface and upstream API/data-layer contracts. Load this repo's documented rules (.claude/rules/), the four docs/reviews/ checklists, architecture docs, then walk them with cross-check discipline. Every repo-convention finding MUST quote a loaded source — drop unsourced claims to avoid inventing repo conventions.

Generic senior-review findings belong to lfx-skills:lfx-general-code-review. Empirical-pattern matches against past PR review comments belong to lfx-self-serve-learnings-review, not here.

Repository scope

This skill is owned by the lfx-self-serve repo (.claude/skills/lfx-self-serve-code-review/), and the reviewer loading it may be launched from the LFX workspace root or a multi-repo session. Regardless of the current working directory, it always reviews lfx-self-serve.

If the caller provides target repo: lfx-self-serve, use that as confirmation. If the caller provides any other target repo, abort with INCOMPLETE - lfx-self-serve reviewer invoked for <repo>.

Before diffing, locate the lfx-self-serve repo root:

• If you are already in lfx-self-serve, you are home. Use that repo root.
• Otherwise, look for a sibling or child directory named lfx-self-serve.
• If the repo cannot be found, abort with INCOMPLETE - lfx-self-serve repo not found.

Inputs

Parse the caller's prompt for:

• branch — OPTIONAL keyword. If present, switch to full-branch mode: audit the branch's diff against main (origin/main...HEAD) instead of just the latest commit. Used by the pre-PR full-branch sweep.
• extra: <free text> — optional priority hint.

Step 1 — Compute the diff

Run all git commands from the lfx-self-serve repo root.

Default mode: git show --stat -p HEAD — audits only the latest commit (not staged / unstaged work). Use the stat block as the canonical changed-file list; abort if empty.

Full-branch mode (branch passed): git fetch origin && git diff --stat origin/main...HEAD && git diff origin/main...HEAD — the branch's diff against main, i.e., everything HEAD adds vs origin/main.

For per-file reads: git show "HEAD:<path>". If the diff is too big for context, save to /tmp/code-review-diff.patch and Read changed files individually.

Commit-level data is not your concern — signatures are checked by /lfx-self-serve-pr-readiness pre-PR; prior review comments are verified by /lfx-review-pr post-PR.

Step 2 — Convention audit

Now layer the LFX-specific conventions on top: load the docs, then walk the audit using them. Step 2.1 sets up the source-of-truth surface; Step 2.2 audits against it with cross-check discipline.

2.1 — Load reference documents

Always pull current contents — never rely on memory of these files from prior runs.

Always read (in parallel):

• CLAUDE.md at the repo root
• ~/.claude/CLAUDE.md if it exists
• Every file matching .claude/rules/*.md — Glob dynamically; never hand-maintain a list

⚠ Mandatory: the four docs/reviews/ checklists. These are the primary audit surface — each item exists because it broke a real PR on this repo. Skipping a relevant checklist invalidates the audit. If you cannot Read a required checklist, mark the report as INCOMPLETE.

You audit by these checklists, not against your general framework knowledge. Locate a specific checklist item, rule file entry, or architecture-doc paragraph for every Repo-conventions finding (see 2.2).

Architecture docs — load conditionally by changed-file paths:

If .claude/skills/self-serve-dev/references/ exists, also Read the relevant reference files (backend-endpoint.md for server changes, frontend-code-generation.md for frontend changes, etc.).

2.2 — Walk the audit

For each changed file:

1. Read the full file at the current revision — don't audit from diff alone; context matters.
2. Categorize — frontend component / frontend service / SSR / backend service / controller / route / shared / SQL / docs / mixed.
3. Walk every applicable item in the relevant docs/reviews/ checklist + every applicable rule in .claude/rules/ + the architecture docs loaded in 2.1.
4. Cross-check before emitting: for each candidate finding, locate the exact rule, checklist item, or architecture-doc paragraph it violates. Quote that source in the finding's _Source:_ citation. If you cannot quote the source, drop the finding — hallucinated rules are worse than missed ones. These findings go in the Repo conventions section of the Step 4 report.
5. Account for the full checklist surface: if you cannot account for having considered every applicable checklist item, mark the report INCOMPLETE in Step 4 rather than ship a partial report.

Step 3 — Upstream API / data-layer contract validation (backend only)

Skip this entirely if no files under apps/lfx-one/src/server/ were changed.

The LFX One backend is a thin proxy to external Go microservices. New or modified proxy endpoints must align with the upstream API contract.

Required for:

• Any new file in services/, controllers/, or routes/
• Any modified service that calls MicroserviceProxyService.proxyRequest()
• Any new API path or changed request/response shape

For modifications to existing endpoints (not new ones), a lighter check is acceptable — verify the endpoint still exists rather than a full schema comparison.

gh api repos/linuxfoundation/<repo-name>/contents/gen/http/openapi3.yaml --jq '.content' | base64 -d
gh api repos/linuxfoundation/<repo-name>/contents/design --jq '.[].name'
gh api repos/linuxfoundation/<repo-name>/contents/design/<file>.go --jq '.content' | base64 -d

Upstream repo map:

Validate:

• Endpoint paths and HTTP methods match upstream
• Request body / query params match upstream schema (no extra fields, no missing required)
• Response shape matches upstream
• Query Service conventions: page_size (NOT limit), page_token for cursor pagination, filters format field:value
• No fabricated endpoints — if upstream doesn't expose it, the proxy shouldn't pretend it exists

Snowflake direct SQL: every ? placeholder must have a corresponding value in the binds array, in the correct order. Bind mismatch is always Critical.

On gh api failure (404, auth, network): surface in Step 4's "Upstream API / data-layer validation" section as "Upstream API contract for <service> could not be verified — manual validation required." Treat as Important severity. Don't silently skip.

Step 4 — Render the report

Header: <commit-sha> — <subject> (default) or origin/main...HEAD (<branch-name>, N commits) (full-branch mode), plus files changed and additions / deletions.

Two sections, in order. Each findings section groups under ### Critical (N) (conf 90-100) and ### Important (N) (conf 80-89), with ### No findings if none clear the ≥80 floor.

1. Upstream API / data-layer validation (Step 3): verified paths, SQL bind checks, "manual validation required" flags, or "Skipped — no backend changes".
2. Repo conventions (Step 2.2): - **<file>:<line>** (conf <0-100>) — <issue>. _Source:_ <quoted rule citation>. _Fix:_ <suggestion>.

If a required checklist or architecture doc couldn't be loaded, lead with INCOMPLETE — couldn't load <file>. If extra was applied, note it.

Severity calibration

Repo-convention and upstream-validation findings share the same buckets and ≥80 floor. Examples illustrative, not exhaustive.

• Critical (90-100) — SQL bind mismatches, upstream contract violations, bypassed user authorization, M2M misuse, raw new Error() / manual res.status().json().
• Important (80-89) — documented style / structure violations (logger usage, license headers, PrimeNG wrappers, @if/@for, inject() DI, page_size over limit), unverified upstream contracts.
• Nit (below 80) — preferences, minor improvements, file naming. Suppressed by the ≥80 floor, except category: upstream-api deterministic flags.

Known false positives — DO NOT emit

• Missing ChangeDetectionStrategy.OnPush — the app uses stable zoneless change detection.
• Missing standalone: true — Angular 20+ defaults to standalone.
• provideZonelessChangeDetection() flagged as experimental — it is stable in Angular 20.
• Any Repo conventions finding whose _Source:_ citation cannot be quoted from a loaded rule file / checklist / architecture doc — drop.

Scope boundaries — NOT this agent's job

• PR-shape sanity (branch name, JIRA, conventional commits, rebase, DCO + GPG, diff size, protected files touched) → /lfx-self-serve-pr-readiness.
• Generic senior code review (correctness, security, maintainability, performance, tests, code truthfulness) → lfx-skills:lfx-general-code-review.
• Empirical pattern matching (KB of past-PR review comments) → lfx-self-serve-learnings-review.