Senior Engineer Guide

February 28, 2026 · View on GitHub

For engineers who want full control over how GAAI governs their project.

This guide covers: rules customization, memory architecture, CI integration, and governance enforcement.

What You Control

GAAI's behavior is entirely driven by files in .gaai/core/contexts/rules/. Every constraint, every allowed action, every escalation condition — all in text. You read and modify them directly.

The framework enforces what you write. Nothing more.

Rules Architecture

.gaai/core/contexts/rules/
├── README.rules.md           ← loading priority and index
├── orchestration.rules.md    ← agent authority and forbidden patterns
├── skills.rules.md           ← skill activation rules
├── artefacts.rules.md        ← artefact lifecycle and ownership
├── backlog.rules.md          ← backlog state transitions
├── memory.rules.md           ← memory loading and storage rules
└── context-discovery.rules.md ← when Discovery must run

Load order matters. orchestration.rules.md is always loaded first. It defines the authority hierarchy. Other rules are loaded by agents selectively based on the task.

Editing rules is safe. Agents read rules at runtime. No rebuild, no restart. Change a rule file → next agent activation picks it up.

Customizing for Your Project

1. Add project-specific conventions

Edit contexts/memory/patterns/conventions.md. This is where you capture conventions that should influence agent behavior:

## Code Style
- Use async/await — no raw Promise chains
- Error types: always typed (never `any`)
- Functions: max 40 lines

## Testing
- Unit tests required for all business logic
- Integration tests required for all API endpoints
- No mocking of database layer

## PR conventions
- One Story per PR
- PR title = Story title

The Delivery Agent loads memory/patterns/conventions.md when it builds execution context. Your conventions become implicit constraints on every implementation.

2. Add domain-specific rules

If your project has domain constraints (regulatory, security, architecture), add a custom rules file:

.gaai/core/contexts/rules/your-domain.rules.md

Reference it in README.rules.md so agents know to load it.

Example for a fintech project:

---
type: rules
id: RULES-FINTECH-001
scope: delivery
---
# Fintech Rules

## Payment Flows
- All payment state changes must be idempotent
- Every transaction must log: amount, currency, user_id, timestamp, result
- No payment logic in controllers — service layer only

## PII
- Mask card numbers in all logs (show last 4 only)
- User PII never in query parameters

3. Tighten the Discovery gate

context-discovery.rules.md defines when Discovery is mandatory before Delivery can proceed. By default it's advisory. To make it enforced:

## Enforcement
Discovery is MANDATORY before Delivery when:
- New user-facing feature
- API contract change
- Data schema change
- Cross-service dependency change

The Delivery Agent must refuse to proceed without a validated Story artefact for these categories.

Memory Architecture

GAAI uses three memory files by default. You can extend this.

.gaai/project/contexts/memory/
├── memory/project/context.md     ← always loaded (project fundamentals)
├── memory/decisions/_log.md   ← append-only decision log
├── memory/patterns/conventions.md    ← conventions and preferences
└── memory/_template.md   ← template for new entries

memory/project/context.md — loaded by every agent on every session. Keep it short and stable. It answers: what is this project, what constraints exist, what are the non-negotiables.

memory/decisions/_log.md — accumulates. Agents append decisions after extracting them. Review it periodically and archive old entries that are no longer relevant.

memory/patterns/conventions.md — your living style guide. Update it when conventions evolve. The rules-normalize skill can help normalize and deduplicate entries.

Adding domain-specific memory

For large projects, split memory by domain:

memory/
├── memory/project/context.md
├── memory/decisions/_log.md
├── memory/patterns/conventions.md
├── auth.memory.md        ← authentication decisions
├── api.memory.md         ← API design decisions
└── infra.memory.md       ← infrastructure constraints

Agents won't auto-load domain memory. They use memory-retrieve to load only what's relevant to the current task. Reference new files in README.memory.md so agents know they exist.

CI Integration

GAAI ships a validation script. Use it as a CI gate.

GitHub Actions

# .github/workflows/validate-structure.yml
name: GAAI Structure Validation
on: [push, pull_request]
jobs:
  validate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Validate .gaai/ integrity
        run: bash .gaai/core/scripts/health-check.sh

The health check validates:

  • All SKILL.md files have required name and description
  • All rule files are present
  • VERSION file exists and is valid
  • Backlog YAML is parseable
  • Cross-references between artefacts and backlog are consistent

Exit 0 = pass. Exit 1 = fail with report.

Pre-commit hook

# .git/hooks/pre-commit
#!/bin/bash
bash .gaai/core/scripts/health-check.sh

Blocks commits that break framework integrity.

Backlog Governance

The backlog in contexts/backlog/active.backlog.yaml is the sole authorization mechanism. Govern it.

Backlog item validation

The backlog.rules.md file defines what a valid backlog item looks like. Review and extend it for your project:

## Required Fields
All backlog items must have:
- id (format: BL-NNN)
- title
- status
- priority
- track (discovery|delivery)
- complexity (fibonacci)
- acceptance_criteria (min 2 criteria)

## Forbidden States
- No item may transition from `refined` to `in-progress` without artefact path set
- No item may be marked `done` without QA report

Backlog scheduler — governance views

The backlog-scheduler.sh script provides governance-focused views of the backlog beyond just "what's next":

# Priority conflicts: high-priority items blocked by lower-priority dependencies
.gaai/core/scripts/backlog-scheduler.sh --conflicts active.backlog.yaml

Example output:

Priority conflicts (2):

  ⚠️  BL-005 (high) is blocked by BL-002 (low)
      → Consider raising priority of BL-002 or lowering BL-005

  ⚠️  BL-007 (high) depends on BL-003 — dependency not found in backlog
      → BL-003 is listed as a dependency but not found in backlog

Run --conflicts during sprint planning or priority reviews to catch misalignments before they delay high-value work.

# Full dependency graph — which items are blocked and by what
.gaai/core/scripts/backlog-scheduler.sh --graph active.backlog.yaml

Add --conflicts to CI alongside health-check.sh if your team wants automated priority governance:

# .github/workflows/validate-structure.yml
- run: |
    bash .gaai/core/scripts/health-check.sh
    bash .gaai/core/scripts/backlog-scheduler.sh --conflicts \
      .gaai/project/contexts/backlog/active.backlog.yaml || true

Artefact traceability

Run artefact-sync.sh to validate that every backlog item with an artefact reference points to a real file:

bash .gaai/core/scripts/artefact-sync.sh

Use this in CI if traceability is required.

Memory Hygiene

Memory grows. Keep it useful.

Compact on context pressure: When an agent reports token pressure, run memory-compact — it distills entries into summaries, replacing raw notes with actionable knowledge.

Review quarterly:

  • memory/decisions/_log.md — archive decisions that are no longer relevant
  • memory/patterns/conventions.md — remove patterns that have changed
  • memory/project/context.md — update project fundamentals as the project evolves

Snapshot before risky work:

bash .gaai/core/scripts/memory-snapshot.sh

Creates a timestamped archive of the current memory state. Roll back to it if needed.

Team Setup

For team usage, see Team Setup Guide.

Key topics:

  • Shared memory in git
  • Branching and backlog conventions
  • Rule ownership and review process
  • Parallel Discovery sessions

→ Team Setup Guide → Design Decisions — why the framework is structured this way → Rules Reference