Skip to content

2. Require AI-assisted coding for platform development

Date: 2026-02-05

Status

Accepted

Context

The Vindicta Platform's 6-week roadmap (Feb 4 - Mar 17, 2026) has aggressive velocity targets:

  • 20 repositories requiring coordinated development
  • Week 1: Foundation scaffolds across 4 priority products (Agent-Auditor-SDK, WARScribe-Core, Primordia-AI, Vindicta-Portal)
  • Week 6: v1.0 production releases for 3 products
  • Daily schedule includes multiple PRs across repositories (e.g., Feb 4 had 4 PRs merged)

The platform architecture includes an agentic ecosystem with defined roles:

  • Senior Manager: Orchestrates platform health and cross-team coordination
  • Agile Delivery Lead: Executes /adl-standup and /adl-pr-review daily
  • Product Owner: Manages /po-sprint-planning and /po-roadmap-update

These agents require AI coding capabilities to execute their workflows autonomously.

Additionally, the Platform Constitution mandates:

  • Article II: Gemini as the AI engine (Gas Tank Model)
  • Article XVI: Async-First development for LLM orchestrations
  • Agent-Auditor-SDK: Implements rate limiting and priority queuing for AI requests

Question: Given the roadmap velocity requirements and agentic architecture, should AI-assisted coding be mandatory?

Decision

AI-assisted coding is REQUIRED for all Vindicta Platform development.

All development work MUST use approved AI coding tools:

  • Antigravity (primary AI coding assistant)
  • GitHub Copilot (code completion and reviews)
  • Agent-Auditor-SDK (agentic task orchestration)
  • Gemini AI Studio (direct API access for agents)

Enforcement Mechanisms

  1. Rate Limiting: Agent-Auditor-SDK RateLimiter class enforces Gemini free tier limits (60 RPM, 60K TPM)
  2. Gas Tank Model: AI requests MUST call acquire() before API calls, respecting quota limits
  3. Priority Queue: Human requests prioritized over agent background work
  4. Cost Controls: Operations stop immediately when quota exhausted (Constitution I compliance)

Consequences

Positive

  • ✅ Enables 6-week roadmap delivery: Manual development cannot achieve required velocity
  • ✅ Supports agent workflows: ADL/PO/SM agents can execute autonomously
  • ✅ Parallelizes development: Multiple PRs per day across repositories (demonstrated in Week 1 with 6 PRs merged)
  • ✅ Free tier compliant: Rate limiting prevents quota overages
  • ✅ Consistent with architecture: Leverages Agent-Auditor-SDK and Gemini API

Negative

  • ⚠️ API quota dependency: Platform velocity limited by Gemini free tier (mitigated by rate limiting)
  • ⚠️ Agent maturity risk: Requires Agent-Auditor-SDK stability (Week 1 PR #12 completed this)
  • ⚠️ Learning curve: Developers must adopt AI-first workflows (addressed by agent AGENT.md documentation)

Neutral

  • Agent workflows become the default interaction model
  • Human intervention required for architecture reviews and release approvals
  • Manual development remains possible but discouraged

Alternatives Considered

Alternative 1: No AI Assistance (Manual Only)

Rationale: Maximum human control, zero AI costs

Rejected because:

  • Cannot meet 6-week timeline
  • Single-threaded development (no parallelization)
  • Invalidates agentic ecosystem architecture
  • Contradicts Constitution Article II (Gemini mandate)

Alternative 2: Optional AI Assistance

Rationale: Flexibility for developers, can be enabled per-project

Rejected because:

  • Inconsistent velocity across repositories
  • Schedule uncertainty
  • Undermines agent autonomy
  • Doesn't leverage platform architecture

Compliance

Constitution Article How This Decision Complies
I. Economic Prime Directive Free tier enforced via RateLimiter (60 RPM/TPM limits)
II. Gas Tank Model acquire() method implements gas tank for Gemini API
III. Spec-Driven Methodology Specs remain human-authored; code AI-generated from specs
V. Agentic Rights & Responsibilities Formalizes AI assistants as first-class development participants
XVI. Async-First Mandate AI tools enable concurrent development across repositories

Implementation

Week 1 demonstrated successful AI-assisted development:

  • 6 PRs merged in 2 days (Feb 4-5)
  • Agent-Auditor-SDK PR #12: Rate limiting prevents 429 errors
  • WARScribe-Core PR #8: Edition abstraction with tests
  • Primordia-AI PR #6: DuckDB schema implementation
  • Vindicta-Portal PR #21, #22: Firebase config + design tokens
  • Platform-Docs PR #2: Branding updates

All implementations approved by architecture review with 100% Constitution compliance.

References