Files
msd-core/.planning/ROADMAP.md
Lex Christopherson d994732a45 docs(05): create phase plan for subagent codebase analysis
Phase 05: Subagent Codebase Analysis
- 2 plans in 2 waves
- Plan 05-01: Create gsd-entity-generator subagent (Wave 1)
- Plan 05-02: Refactor analyze-codebase to use subagent (Wave 2)
- Ready for execution

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
2026-01-20 11:44:32 -06:00

4.5 KiB

Roadmap: v1.9.0 Codebase Intelligence System

Goal: Make GSD feel intelligent and automagical in how it navigates and understands both greenfield and brownfield projects.

Phases: 5 (4 complete)


Current Milestone: v1.9.0

Phase 1: Foundation & Learning ✓

Goal: Establish index schema and incremental learning via PostToolUse hook Status: Complete Plans: 2/2

Phase 2: Context Injection ✓

Goal: Inject codebase awareness into every session via SessionStart hook Status: Complete Plans: 2/2

Phase 3: Brownfield & Integration ✓

Goal: Deep analysis command for existing codebases, workflow integration Status: Complete Plans: 3/3

Phase 4: Semantic Intelligence & Scale ✓

Goal: Transform syntax-only indexing into semantic understanding with graph-based relationships Status: Complete Plans: 5/5

Plans:

  • 04-01-PLAN.md — SQLite graph layer with sql.js (Wave 1)
  • 04-02-PLAN.md — Graph-backed rich summary generation (Wave 2)
  • 04-03-PLAN.md — Semantic entity generation via Claude API (Wave 2)
  • 04-04-PLAN.md — CLI query interface for getDependents (Wave 3)
  • 04-05-PLAN.md — Wire plan-phase.md to inject intel into planner (Wave 3)

Wave Structure:

  • Wave 1: 04-01 (SQLite foundation)
  • Wave 2: 04-02, 04-03 (parallel - both depend only on 04-01)
  • Wave 3: 04-04, 04-05 (parallel - consumption layer)

Why this phase:

  • Current system provides "2-3 ls commands worth of information" (Claude's own assessment)
  • Missing: what files actually DO, who uses them, blast radius of changes
  • Senior engineers at top companies need real intelligence, not file counts

Delivers:

  • SQLite graph layer (sql.js - zero native deps) for relationship queries
  • Entity-based semantic documentation (Claude writes understanding, not just syntax)
  • Semantic /gsd:analyze-codebase that creates initial entities
  • Rich summary generation from accumulated semantic knowledge
  • CLI query interface for "what uses this file?" queries

Requirements:

  • INTEL-04: Entity files capture semantic understanding (purpose, what exports do)
  • INTEL-05: Relationships queryable ("what uses this file?", "blast radius")
  • INTEL-06: /gsd:analyze-codebase creates initial entity docs via Claude
  • INTEL-07: Summary reflects accumulated semantic knowledge

Success Criteria:

  1. Claude can answer "what uses src/lib/db.ts?" from SessionStart context
  2. Summary includes file purposes, not just file counts
  3. Transitive dependency queries work (blast radius)
  4. Works at scale (500+ file codebases)

Phase 5: Subagent Codebase Analysis

Goal: Prevent context exhaustion on large codebases by delegating analysis to subagents Depends on: Phase 4 Plans: 2 plans

Plans:

  • 05-01-PLAN.md — Create gsd-entity-generator subagent (Wave 1)
  • 05-02-PLAN.md — Refactor analyze-codebase to use subagent delegation (Wave 2)

Wave Structure:

  • Wave 1: 05-01 (agent definition)
  • Wave 2: 05-02 (command refactor + verification)

Why this phase:

  • Current entity generation loads file contents in orchestrator context
  • On large codebases (500+ files), orchestrator exhausts context during file selection and batching
  • Subagent delegation gives fresh 200k context for entity generation

Delivers:

  • gsd-entity-generator subagent following gsd-codebase-mapper pattern
  • Refactored /gsd:analyze-codebase that spawns subagent instead of inline batching
  • Preserved orchestrator context for large codebase analysis

Requirements:

  • INTEL-08: Entity generation delegated to subagent (not inline)
  • INTEL-09: Subagent writes entities directly, returns statistics only
  • INTEL-10: Orchestrator passes file paths, not file contents

Success Criteria:

  1. Entity generation works via subagent spawn
  2. Orchestrator context preserved (no file contents loaded)
  3. Entities correctly formatted and graph.db updated
  4. Works on 500+ file codebases without context exhaustion

Traceability

Requirement Phase Status
INTEL-01 Phase 1 ✓ Complete
INTEL-02 Phase 2 ✓ Complete
INTEL-03 Phase 3 ✓ Complete
INTEL-04 Phase 4 ✓ Complete (04-03)
INTEL-05 Phase 4 ✓ Complete (04-04)
INTEL-06 Phase 4 ✓ Complete (04-03)
INTEL-07 Phase 4 ✓ Complete (04-02)
INTEL-08 Phase 5 Planned (05-01, 05-02)
INTEL-09 Phase 5 Planned (05-01)
INTEL-10 Phase 5 Planned (05-02)

Created: 2026-01-19 Updated: 2026-01-20 — Phase 5 planned with 2 plans in 2 waves