graph-review

graph-review

Run scripts/graph-audit.mjs to surface code-health signals derived from .understand-anything/knowledge-graph.json. Default to the summary rollup; drill into individual audits only when a check is non-zero or the user asks for specifics.

This skill is deterministic: it reads a static JSON and emits counts/lists. No LLM call inside. Use the model's judgment only for interpreting the output and recommending follow-ups.

Preconditions

• .understand-anything/knowledge-graph.json must exist at the repo root. If absent, tell the user: "No knowledge graph found. Run /understand first." Do NOT try to regenerate it inside this skill — /understand is a separate multi-phase pipeline.
• node available on PATH (already required by the project).

Workflow

1. Rollup first — get one-line counts per check:

   node scripts/graph-audit.mjs summary
   

2. Decide drill-in scope based on rollup:
   • orphans > 0 → run orphans. After Phase 7 cohort-wiring this should always be 0; a non-zero means /understand was interrupted or the graph is stale.
   • dead-code > 0 is expected and noisy — only drill if user explicitly asks. Always preface with the under-capture caveat (see Caveats).
   • redline-r1 > 0 → run redline-r1. Every hit is a real R1 violation (sandbox bridges already excluded).
   • redline-r10 > 0 → run redline-r10. Means src/harness/ exceeded 12 files; check whether new file is justified per CLAUDE.md R10 "3 questions before adding code".
3. Drill in with the relevant subcommand and --limit=N to widen lists:

   node scripts/graph-audit.mjs redline-r1 --limit=50
   node scripts/graph-audit.mjs dead-code --json | jq '.items[] | select(.file | startswith("src/gateway/"))'
   

4. Cross-check before recommending fixes — especially for dead-code (see Caveats). Use rg, cargo machete, or IDE find-usages.

Commands

Command	Output	Notes
summary	One-line counts per audit	Default; safe and fast
orphans	Nodes with zero edges	Expected: 0. Non-zero = stale or broken graph
dead-code	function/class nodes with no inbound calls edge	CANDIDATES only — see Caveats
redline-r1	src/ files naming platform-native APIs (AppKit / Vision / win32 / objc / cocoa / etc.)	Violates R1 brain-limb separation. src/sandbox/ + src/bin/aleph-server/ excluded (legitimate bridges)
redline-r10	src/harness/ .rs file count vs. 12-file budget	R10 thin-harness invariant — see CLAUDE.md §R10

Flags: --json (machine output, pipeable to jq), --limit=N (default 30, widens lists in human mode), --graph=path (point at a non-default graph file — useful for subdomain graphs like frontend-knowledge-graph.json), --scope=prefix/ (restrict audit to nodes whose filePath starts with the prefix; trailing / auto-appended).

Module-scoped audits

Use --scope=<path-prefix> to confine any audit to a subtree. Filtering is on the graph node's filePath field, so it covers all 4 audits consistently. Trailing / is auto-appended when the prefix has no . (avoiding src/gateway accidentally matching src/gateway_extras/).

# Only audit src/gateway/
node scripts/graph-audit.mjs summary --scope=src/gateway/

# Only harness dead-code candidates
node scripts/graph-audit.mjs dead-code --scope=src/harness/

# R1 violations under a single feature subtree
node scripts/graph-audit.mjs redline-r1 --scope=src/builtin_tools/browser_tools/

Per-audit behaviour under scope:

Audit	Scoped behaviour
orphans	Only orphans whose filePath is within scope.
dead-code	Only function/class nodes within scope. Global call-edge stat in warning stays global (not scope-adjusted) — that's intentional, the under-capture is a project-wide signal.
redline-r1	Only src/ files within scope (still excludes bridge subtrees src/sandbox/ and src/bin/aleph-server/).
redline-r10	If scope doesn't overlap src/harness/, reports N/A and returns 0. Otherwise filters harness files by scope.

Natural-language triggers map to scope flags. When the user says:

• "审 src/gateway/" → --scope=src/gateway/
• "只看 harness 的死代码" → dead-code --scope=src/harness/
• "看看 builtin_tools 里的 R1" → redline-r1 --scope=src/builtin_tools/

Caveats

• dead-code is severely under-captured. The graph currently captures ~4.5% of true call edges (~159 calls edges across ~3518 function nodes). The dead-code report will show thousands of false positives. Treat output as candidates requiring manual verification, not a deletion list. Always say so when reporting. The fix lives upstream in /understand's file-analyzer subagent, not here.
• redline-r1 token list is conservative. Currently tokens like use objc::, use cocoa::, use core_graphics::, use core_foundation::, NSWorkspace, AVCaptureDevice, CGEvent, use windows::, use winapi::, CheckNetIsolationEnableLoopback, CreateProcessW. Add new tokens as new platform APIs appear in the codebase (edit the NATIVE_API_TOKENS array in scripts/graph-audit.mjs).
• redline-r10 file budget is wired to CLAUDE.md §R10. Current budget: 12 files. If CLAUDE.md is bumped, update HARNESS_BUDGET in scripts/graph-audit.mjs in the same commit — the two must stay in sync.
• The script writes nothing. Pure read. Safe to run repeatedly, including in CI.

Reporting back to the user

After running, summarise in 3-5 lines max:

• Which checks passed (one bullet, group them).
• Which need attention (one bullet per non-zero check, with the count).
• Top 1-2 candidates worth investigating, if any (with file path).

Do NOT dump full lists into the response. If listing is genuinely needed (e.g., user explicitly asks "show me all R1 violations"), prefer --limit=5 and offer --limit=N to widen.

For R1 / R10 hits, link the redline in CLAUDE.md:

• R1 — brain-limb separation: no platform-native API calls outside src/sandbox/ (OS isolation bridge) and src/bin/aleph-server/ (libc/process control).
• R10 — thin-harness budget: src/harness/ ≤ 12 files / ~4900 LOC. Adding a file requires answering the "3 questions before adding code" in CLAUDE.md.

When NOT to use this skill

• User asks "who calls function X?" or "what does file Y depend on?" — this skill doesn't do per-symbol drill-down. Suggest rg, IDE find-usages, or future MCP server.
• User wants the dashboard visualisation — invoke understand-anything:understand-dashboard instead.
• Graph is stale (commits since last /understand) — re-run /understand first; this skill audits whatever's on disk and won't warn you that the graph is old.