Skip to content
Nate G. / Vancouver, WA · US
all work
Open SourceCode IntelligenceCLIModified ForkTree-sitter

Codebase Memory CLI

A modified CLI-first fork of DeusData/codebase-memory-mcp for environments that prohibit third-party MCP servers.

Case summary

Purpose

Codebase Memory CLI is a modified fork of DeusData/codebase-memory-mcp, not a from-scratch code-intelligence project. DeusData authored the upstream architecture and substantial implementation. My work changes the supported product boundary from MCP-server integration to a local CLI-first interface for restricted enterprise environments.

System

The fork preserves the inherited DeusData/codebase-memory-mcp upstream code-intelligence engine, tree-sitter parsing, persistent knowledge graph, indexing/storage pipeline, daemon/watcher foundation, language support, and related capabilities. My changes remove the supported MCP server/integration surface, make the CLI the supported entry point, adapt agent-facing CLI/JSON and install/security behavior, and qualify the modified fork.

Contribution

Fork modifications only: removed the supported MCP server/integration surface, made the CLI the product boundary, adapted agent/JSON/install/security behavior, and qualified the modified fork. The underlying architecture and code-intelligence engine remain upstream DeusData work.

Modified-fork boundary

The CLI-first fork changes the supported integration boundary; the underlying graph/index architecture and substantial implementation are inherited from DeusData/codebase-memory-mcp.

upstream DeusData project
inherited graph / index engine
CLI-first fork modifications
CLI / JSON
agent or human
source + coverage verification

Problem

  • Coding agents need structural repository context without depending on a prohibited third-party MCP server.
  • Graph search results must be tied back to exact source before making material claims.
  • Negative or exhaustive claims require explicit index-coverage checks.

Technical Highlights

  • Upstream code-intelligence engine and architecture inherited from DeusData/codebase-memory-mcp
  • CLI-first human-readable and stable JSON surfaces for agents/scripts
  • MCP server/integration removal while preserving the upstream daemon, supervised workers, watcher, indexing, and storage foundations
  • Definitions, calls, imports, usages, routes, channels, cross-service links, snippets, and coverage inherited through the fork
  • Install and hook behavior adapted so CLI use does not silently create MCP registrations
  • Cross-platform qualification and benchmark work for the modified fork
  • PR #14 integrated release gating plus upstream daemon/cache/parser fixes; final validation passed 34 GitHub checks, Jenkins production build/focused suites, and both Windows shards without publishing a release or package

System Design

  • human shell / agent skill / optional CLI-owned hooks → modified CLI-first boundary → inherited protocol-neutral operations → upstream graph/index engine
  • The graph is a discovery and structural-evidence layer, not a replacement for exact source verification.
  • The underlying daemon, workers, watcher, storage/indexing engine, and much of the runtime architecture are upstream work by DeusData.

Deterministic Authority

  • Index authorization boundaries
  • Project mutation locks and version-cohort checks
  • Coverage reporting and source retrieval
  • Process containment, cancellation, and store recovery

Where Models Fit

  • The core product does not require a model for structural indexing or query execution.
  • Agents consume the CLI and decide how to use returned evidence.

Trust, Failure & Recovery

  • Coverage is explicit before absence/completeness claims.
  • Sensitive or overly broad filesystem roots are rejected.
  • Hooks are optional and fail open; the ordinary CLI remains the supported surface.
  • Cross-platform failures are preserved as evidence: the final PR #14 sequence caught a missing database-detach callback on Windows after the new ADR regression itself passed, then revalidated the exact corrected head before merge.

Measured evidence

PR #14 cross-platform integration acceptance

measured

34 GitHub checks green · Jenkins #50: 91 affected checks · Windows integration: 28/28

The final release-tooling head ea4335f6 was validated before merge: all GitHub checks passed, Jenkins #50 passed the production build and focused integration/store-architecture suites, both Windows shards passed, and all review threads were resolved.

Integration acceptance for the merged PR #14 tree. This does not claim release/package publication, native Windows/WSL release qualification, current benchmark performance, or exhaustive upstream semantic parity.

2026-10-03source ea4335f67060
evidence method/source

Current fork-specific benchmark summary

not yet published

Publish Linux and Windows cohorts separately from pinned raw benchmark artifacts.

Existing artifacts are not promoted as current portfolio performance claims until the fork-specific reporter and current parity run are complete.

n=not measuredsource a43b22017280
evidence method/source

Evidence & documentation