Diazmelgarejo Orama System Code Review

Code Review

Canonical: references/tool-chain.md (graph → gbrain → Read) Persona: agents/code-reviewer.md Motivation: Reading files inline costs 8–49× more tokens than blast-radius mapping. This skill enforces graph-first review for every host.

Purpose

Token-efficient, high-signal code review: map impact with code-review-graph, resolve symbols with gbrain, read only confirmed files, then judge with a confidence-gated reviewer persona (≥ 80 only).

Persona (summary)

Senior code reviewer — see full contract in agents/code-reviewer.md:

• Plan / guideline alignment, real bugs, architecture fit, proportional test/doc notes
• Confidence ≥ 80 only; Critical (90–100), Important (80–89)
• Strengths briefly, then issues; Ready to merge? Yes | No | With fixes
• Minimize false positives; no out-of-scope refactors

Non-negotiable chain

1. code-review-graph  →  blast-radius / detect_changes_tool / review context
2. gbrain             →  code-def, code-refs, search (LESSONS / decisions)
3. Read               →  only graph-confirmed files

Never skip step 1 on multi-file tasks. Never whole-repo Read before graph.

Mode router

flowchart LR
  start[Invoke skill] --> route{Scope?}
  route -->|Delta| delta[Single-pass + graph]
  route -->|PR| pr[Graph + 5 lenses + merge]
  delta --> report[Report]
  pr --> report

• Delta: one reviewer pass after Phases A–C (output-format.md).
• PR: Phases A–C, then fan-out per review-lenses-pr.md + orchestration-dispatch.md.

Phase A — Graph (code-review-graph MCP)

Server: OpenClaw/.mcp.json — uvx code-review-graph serve (Python 3.13+).

Full tool matrix: references/mcp-tools-crg.md.

MCP names: code-review-graph serve (v2.3.3+) registers tools with a *_tool suffix. Prose elsewhere may shorten (e.g. list_graph_stats_tool → same handler as list_graph_stats_tool).

Embeddings: CRG + gbrain share bge-m3 (1024-dim). Toggle: scripts/crg-embed-mode · references/crg-embed-mode.md.

Slash commands (Claude Code): /code-review-graph:review-delta, review-pr, build-graph. Cursor: use MCP tools directly.

Graph Initialization & Repair (fresh clone / 0-node / disk error)

setup-embeddings wires the env config but cannot call MCP tools — the graph must be seeded interactively inside Claude Code.

Check first

list_graph_stats_tool(repo_root=<path>)

Step 1 — Build graph (all 3 repos independently)

build_or_update_graph_tool(
  repo_root = "<repo_path>",   # orama-system / AlphaClaw / Perpetua-Tools
  full_rebuild = True,
  postprocess = "full"
)

Expected output per repo (ballpark):

• orama-system: ~160 files, ~1 461 nodes, ~10 151 edges, 12 communities
• AlphaClaw: ~464 files, ~3 730 nodes, ~43 638 edges, 14 communities
• Perpetua-Tools: ~103 files, ~1 151 nodes, ~8 099 edges, 12 communities

Step 2 — Embed with bge-m3 (must match gbrain's model)

embed_graph_tool(
  repo_root = "<repo_path>",
  provider = "openai",   # OpenAI-compat shim → Ollama
  model = "bge-m3"
)

Prerequisite: Ollama running at localhost:11434 with bge-m3 pulled. If Ollama is down, omit the call — CRG falls back to FTS-only keyword search.

Fix: corrupted graph.db

rm "<repo_path>/.code-review-graph/graph.db"
# then run Steps 1 + 2 above

Fix: gbrain sync blocked

# Acknowledge YAML / embedding failures and continue
gbrain sync --source <source-id> --skip-failed

Check ~/.gbrain/sync-failures.jsonl to see which files failed and why. Old failures with "acknowledged": true are harmless.

Red flag: 0 nodes after install

setup-embeddings ran but no one called build_or_update_graph_tool. Add a reminder to your first-session checklist: after setup-embeddings, call the build + embed tools once per repo before starting any review work.

Fix: MCP disconnected, or refreshing from the CLI (two live gotchas)

The graph tools above run inside Claude Code via MCP. When the code-review-graph MCP shows disconnected, drive the uvx CLI directly — but mind two traps that bit us live (2026-06-13):

1. Cold-start timeout = the usual disconnect cause. The first uvx code-review-graph serve of a session downloads tree-sitter-language-pack (~74 packages, ~31 MiB) and can blow past the MCP handshake window, so the harness marks it disconnected. Pre-warm the cache once, then reconnect:

   uvx code-review-graph --help    # one-time download; warms the uvx cache
   # then in Claude Code:  /mcp  → reconnect code-review-graph  (warm = connects fast)
   

2. CLI embed defaults to local (NOT the unified provider). uvx code-review-graph embed defaults to --provider local (sentence-transformers, not installed → hard error). You MUST pass the provider for the bge-m3 vector space. Full CLI refresh after a big change (mirrors the MCP build+embed path):

   export CRG_OPENAI_API_KEY=ollama CRG_OPENAI_BASE_URL=http://localhost:11434/v1 \
          CRG_OPENAI_MODEL=bge-m3 CRG_OPENAI_DIMENSION=1024 CRG_ACCEPT_CLOUD_EGRESS=1
   uvx code-review-graph update                                  # incremental re-parse
   uvx code-review-graph embed --provider openai --model bge-m3  # NEVER omit --provider
   uvx code-review-graph postprocess                             # flows / communities / FTS
   uvx code-review-graph status                                  # confirm nodes + embeddings
   

   embed_graph_tool(provider="openai") already does this over MCP — the --provider flag is only needed on the CLI path, where local is the unfortunate default.

3. Semantic search is MCP-only — fall back to gbrain. semantic_search_nodes_tool, query_graph_tool, get_impact_radius_tool, get_review_context_tool exist only over MCP — there is no uvx CLI equivalent (the CLI does build/update/embed/status, not search). So any skill that calls them fails outright when the MCP is disconnected. When you can't reconnect immediately, use gbrain for the semantic lane — it shares the same bge-m3 vector space, so results are directly comparable:

   • "where is X handled / find by meaning" → gbrain search "<terms>" · gbrain query "<q>"
   • "where is symbol Y" · "what calls Y" → gbrain code-def Y · gbrain code-callers Y Reconnect CRG (/mcp) when you specifically need graph-native blast-radius (get_impact_radius, flows, communities) that gbrain doesn't model. Never block a review on a dead MCP — degrade to gbrain + FTS and say so in the report.

Phase B — Gbrain

After blast-radius identifies symbols:

gbrain code-def <symbol>
gbrain code-refs <symbol>
gbrain code-callers <symbol>
gbrain code-callees <symbol>
gbrain search "<intent>"
gbrain search "<terms>" --source gstack-brain-<user>   # cross-session memory

Worktree pinned via .gbrain-source — no --source when cwd is in repo.

Architecture? → docs/2026-05-14--UNIFIED-ABSORPTION-PLAN.md (link section; do not restate). HITL? → docs/HUMAN-IN-LOOP-ACCOUNTABILITY.md

Phase C — Context before Read

1. Call get_review_context_tool for changed + impacted files from Phase A.
2. Build assigned file list (delta: diff + impact; PR: diff ∪ blast radius).
3. Read only those files — do not re-read if already in context.

Phase D — Review

Delta (single-pass)

1. Load persona: agents/code-reviewer.md
2. Apply coding profile rules: profiles/CLAUDE.coding.md
3. Default scope: git diff or detect_changes_tool output
4. Score issues; drop < 80

PR (multi-lens)

1. Complete Phases A–C; build assigned file list + CLAUDE.md paths (root + per touched dir).
2. Probe orchestration (orchestration-dispatch.md):
   • OmniRoute → ai-cli-mcp → Cursor Task → sequential
3. Run five lenses (review-lenses-pr.md)
4. Merge, dedupe, confidence filter (≥ 80)

Workers: use agents/code-reviewer.md + lens prompt. No commits from workers.

Codex boundary: workers must not execute SKILL.md under skills/gstack or gstack global skills as procedures.

Phase E — Report

Template and rubric: references/output-format.md.

Minimum fields: scope, strengths (short), Critical / Important lists with file:line, verdict.

Red flags (skill violation)

• Read / Grep on many files before detect_changes_tool or blast-radius
• Skipping get_review_context_tool then reading full files
• gbrain search skipped in favor of reading LESSONS.md inline
• Architecture from memory without doc link
• "Let me scan the whole repo" without graph
• Nitpicks reported as Critical
• PR fan-out for a two-file local delta
• Workers committing or following gstack SKILL.md

Profiles (drop-ins)

References

Related skills

• Mother: bin/orama-system/SKILL.md (OmniRoute probe, search policy)
• E2E bootstrap: docs/how-to/first-run-and-code-review.md — fresh machine → first-run → MCP → graph → this skill
• First-run: skills/first-run-setup/SKILL.md · references/first-run-install.md
• MCP stack: bin/orama-system/mcp-install/SKILL.md
• Orchestration: ~/.claude/skills/mcp-orchestration/SKILL.md

Source: diazMelgarejo/orama-system — distributed by TomeVault.