Aller au contenu principal

ADR-030: AI-Native Context Management Architecture

Metadata​

FieldValue
Status✅ Accepted
Date2026-02-05
AuthorChristophe Aboulicam + Claude Code

Context​

STOA Platform is built exclusively with AI (Claude Code) in an "AI Factory" model where an architect supervises autonomous AI workers. Effective context management is critical: each Claude Code session starts fresh and must quickly load all necessary context to produce correct, consistent code.

Current problems​

  1. Duplication: Two CLAUDE.md files (root and .stoa-ai/) contain conflicting information. The .stoa-ai/CLAUDE.md references backend/ (old directory name, now control-plane-api/) and states coverage target is 80% (actual: 70%).
  2. Two memory systems: Root memory.md is actively maintained; .stoa-ai/memory.md is frozen at the CAB-1068 setup session.
  3. Monolithic context: Root CLAUDE.md is 11KB — all rules (Python style, TypeScript style, Rust, testing, security, deployment, git conventions, AI workflow) in a single file, consuming context window unnecessarily.
  4. No component context: 6 of 8 components lack any CLAUDE.md. When Claude Code works deep in control-plane-api/src/routers/, it has no component-specific guidance.
  5. Underutilized native features: Only 1 Claude Code skill (e2e-test) exists. Zero hooks configured. Zero .claude/rules/ files. The .stoa-ai/templates/PROMPT-ENTRY.md defines "Mode Audit", "Mode Fix", "Mode Refactor" as copy-paste text instead of native invocable skills.
  6. Permission sprawl: settings.local.json has 280 allow rules accumulated from ad-hoc sessions, including one-off commands that should not persist.
  7. No cross-repo context: The 5 repos (stoa, stoa-docs, stoa-web, stoa-quickstart, stoactl) have no shared context. Working in stoa-docs gives Claude Code zero knowledge about the stoa monorepo architecture.
  8. ADR numbering conflict: ADR-027 in stoa/docs/architecture/adr/ ("Gateway Adapter Pattern") conflicts with ADR-027 in stoa-docs ("X.509 Header Authentication").

Industry references​

SourcePractice
Anthropic Claude Code Best PracticesHierarchical CLAUDE.md, .claude/rules/ for modular instructions, native skills, hooks
GitLab Handbook-FirstSingle source of truth, everything documented in the repo
Stripe RFC ProcessDesign docs (ADRs) before implementation, traced decisions
Vercel/TurborepoPer-package context in monorepos
AWS AI-DLCAdaptive workflows for AI-driven development lifecycle

Decision​

1. Hierarchical CLAUDE.md with modular rules​

Root CLAUDE.md (~3KB) serves as a compact entry point: project overview, component map, key URLs, session workflow, and @import for current sprint context. All detailed rules are extracted into .claude/rules/ files which Claude Code auto-loads based on path patterns.

8 rule files in .claude/rules/:

FileScopeContent
code-style-python.md**/*.pyruff, black, isort, mypy, line length
code-style-typescript.md**/*.{ts,tsx}eslint, prettier, path aliases
code-style-rust.mdstoa-gateway/**cargo fmt, clippy
testing.md**/tests/**pytest, vitest, playwright standards
security.mdterraform/**, charts/**, deploy/**Infrastructure protection, secrets
git-conventions.mdallCommit format, branch naming, PR template
deployment.mdcharts/**, deploy/**Helm, Terraform safety, Docker multi-arch
ai-workflow.mdallSession rules, review process, anti-drift

2. Component-level CLAUDE.md (lazy-loaded)​

Each of the 8 components gets a CLAUDE.md (~1-2KB) loaded only when Claude Code reads files in that directory. Contains: overview, tech stack, key commands, directory structure, patterns used, and dependency relationships.

Components: control-plane-api/, control-plane-ui/, portal/, mcp-gateway/, stoa-gateway/, cli/, e2e/, charts/stoa-platform/.

3. Native skills replacing template-based workflows​

Convert .stoa-ai/templates/PROMPT-ENTRY.md modes into 7 native Claude Code skills invocable via / command:

SkillReplacesPurpose
/fix-bug"Mode Fix Rapide"Reproduce, fix, verify, commit
/implement-featureStandard promptPlan, code, test, commit
/audit-component"Mode Audit"Read-only analysis
/refactor"Mode Refactor"Safe refactor with test guard
/create-adr(new)ADR creation from template
/review-pr(new)PR review checklist
/update-memory(new)Update memory.md with current state

4. Hooks for automation and guardrails​

Configure .claude/settings.json (committed) with:

  • PreToolUse hook on Edit/Write: Block modifications to protected infrastructure paths (terraform/environments/prod, .github/workflows)

5. Single memory.md with structured format​

One memory.md at root (delete .stoa-ai/memory.md). Structured with tables for session state, a decisions section, and known issues. Items older than 2 weeks archived to docs/archive/.

6. Cross-repo context via ~/.claude/CLAUDE.md​

A user-level ~/.claude/CLAUDE.md provides shared context across all STOA repos: repo table, cross-repo rules (API changes require docs update, quickstart must match Helm defaults), and shared standards.

7. Retire .stoa-ai/ entirely​

The AI Factory infrastructure (.stoa-ai/) predates Claude Code's native skills, hooks, and rules. All useful content is migrated:

  • Templates become skills in .claude/skills/
  • Session workflow becomes a rule in .claude/rules/ai-workflow.md
  • Scripts (slack-notify.sh, daily-digest.sh) move to scripts/ai-ops/
  • Phase plans archived to docs/archive/phases/

8. Clean permission model​

Reduce settings.local.json from 280 rules to ~30 using glob patterns (Bash(cargo:*) instead of individual cargo subcommands).

Consequences​

Positive​

  • No duplication — single source of truth for each context type
  • Faster context loading — 3KB root + lazy component files vs. 11KB monolith
  • Native integration — skills, hooks, rules work with Claude Code's built-in features
  • Cross-repo awareness — AI knows about all 5 repos when working in any one
  • Scalable — adding a new component = adding one CLAUDE.md file
  • Traceable — decisions logged, patterns documented, ADRs indexed
  • Maintainable — modular rules are easier to update than a single large file

Negative​

  • More files — ~35 new files across the repo (rules, skills, component docs)
  • Migration effort — one-time cost to restructure existing context
  • Learning curve — team must understand the hierarchical loading model

Mitigations​

  • All new files are small (1-2KB each) and focused on one concern
  • Migration is automated in a single session
  • The hierarchy is documented in this ADR and in the root CLAUDE.md itself

References​