CLI `build` missing post-processing steps that `build_or_update_graph_tool` MCP tool includes

[igagankalra]

Summary

code-review-graph build (CLI) and build_or_update_graph_tool(full_rebuild=True) (MCP tool) produce significantly different results. The CLI only runs Tree-sitter parsing, while the MCP tool also runs 4 post-processing steps. This means users who build via CLI get an incomplete graph with many features returning empty results.

This affects all CLI entry points — build, update, and watch — as well as the Claude Code PostToolUse hook which calls code-review-graph update.

Steps to Reproduce

1. Install: pip install code-review-graph
2. Build via CLI: code-review-graph build
3. Start MCP server: code-review-graph serve
4. Query communities → 0 results
5. Query flows → 0 results
6. Semantic search → 0 results (FTS not indexed)
7. Architecture overview → 0 communities

Workaround: calling build_or_update_graph_tool(full_rebuild=True) via the MCP server produces the correct results with all features populated.

Expected Behavior

CLI build should produce the same output as the MCP build_or_update_graph_tool(full_rebuild=True) — or at minimum, the CLI docs/README should clarify that additional steps are needed.

Root Cause

cli.py (build command) calls only:

result = full_build(repo_root, store)

tools/build.py (build_or_update_graph) calls full_build() plus 4 post-build steps:

1. Compute node signatures
2. Rebuild FTS index (search.rebuild_fts_index)
3. Trace execution flows (flows.trace_flows → flows.store_flows)
4. Detect communities (communities.detect_communities → communities.store_communities)

These post-build steps populate the flows, flow_memberships, communities tables and the nodes_fts virtual table. Without them, the corresponding MCP tools return empty results.

Feature Comparison

Post-processing step	CLI build	MCP build_or_update_graph_tool
Tree-sitter parse (nodes + edges)	✅	✅
Compute signatures	❌	✅
Rebuild FTS index	❌	✅
Trace execution flows	❌	✅
Detect communities	❌	✅

Affected Entry Points

The same gap exists in all non-MCP paths:

• code-review-graph build — CLI full build
• code-review-graph update — CLI incremental update
• code-review-graph watch — filesystem watcher (calls incremental_update internally)

The only path that works correctly is the MCP tool build_or_update_graph_tool().

Also Affects Claude Code Hooks

This isn't platform-specific — the bundled Claude Code hooks have the same gap:

• hooks/session-start.sh suggests running code-review-graph build (CLI) when no graph exists — this produces an incomplete graph
• hooks/hooks.json PostToolUse hook fires code-review-graph update (CLI) after every Write/Edit — incremental updates also skip post-processing

So even in Claude Code (the primary target platform), a user following the setup instructions gets an incomplete graph until an agent happens to call build_or_update_graph_tool() via the MCP server.

Suggestion

Extract the post-build pipeline from tools/build.py into a shared function and call it from both the MCP tool and the CLI commands. This would fix the CLI, the hooks, and the watch mode in one change.

Environment

• code-review-graph v2.1.0
• Python 3.14
• macOS (Apple Silicon)
• Tested with an agentic OpenCode workflow (multi-agent setup using the MCP server)

[tirth8205]

Fixed in v2.2.x — the CLI build and update commands now run the full post-processing pipeline (flows, communities, FTS, signatures). See the postprocess subcommand for manual runs.

[tirth8205]

Was this fixed? The issue was that CLI 'build' skipped post-processing steps that the MCP build tool ran.

[tirth8205]

Confirmed resolved: the CLI now has a 'postprocess' subcommand and build/update commands have --skip-postprocess/--skip-flows flags, indicating post-processing is now integrated into the CLI workflow. Issue correctly closed.