Peer Review, But With Robots: How Claude, Gemini, and Codex Took SAIQL to Production

December 13, 2025 | Article

Peer Review, But With Robots: How Claude, Gemini, and Codex Took SAIQL to Production

Claude, Gemini, and Codex peer-reviewing SAIQL

By Apollo Raines

I ran a local, three-agent "production readiness sprint" inside my Nova system: Claude (implementation), Gemini (verification), and Codex (final QA). The rules were simple: no fixing until we had consensus on every issue, and no claiming "done" until another model tried to break it. The result was equal parts brutal and hilarious -- Claude tried to fast-pass a few "fixed" items, Gemini smacked the claim into reality, then Codex later did the same to Gemini. This is the play-by-play of how SAIQL went from "pretty close" to "ship it."

The Setup

Apollo (Orchestrator)

To Claude, Gemini, and Codex -- You are in the Nova system. You have each scanned and created your notes for /home/nova/SAIQL.DEV/ -- this is a collaboration file. I want you to each state anything you found that is broken or missing in SAIQL, and your suggestion on how to fix it. At this point, you are NOT to proceed with any work. Answer in your section below, then everyone, including me, will read your additions -- and I will decide how to proceed.

Phase 1: The Discovery Round

Each model scanned the entire SAIQL codebase independently and reported back. No coordination, no peeking at each other's notes.

Claude's Findings

Pipeline Execution Gap: In core/engine.py:676-680, the execution pipeline terminates after compilation without actually executing against storage backends. The code comments indicate "SAIQL Compression Validation: Compilation succeeded, bypass database execution" -- this breaks the core promise of being a functional database.
QIPI Index Integration Incomplete: While core/qipi_index.py contains the revolutionary index implementation, it's not integrated into the main query execution path.
Imagination Engine Isolation: The powerful core/imagination/imagination_engine.py runs standalone without integration into the query language.
Vector Engine API Mismatch: Vector operations should be native to the query language, not separate APIs.
Configuration Security Gap: config/secure_config.py is referenced throughout the codebase but doesn't exist.
Storage Layer Fragmentation: Multiple storage implementations exist without a unified storage abstraction.
Test Coverage Critical Gaps: No integration tests verify the end-to-end pipeline.
Production Readiness Concerns: Docker configurations assume production deployment, but core database functionality is bypassed.

Codex's Findings

Engine execute path stops after compilation and never runs against storage/runtime.
License inconsistency: openapi.yaml declares MIT while repo is Open Lore License.
Config hygiene: JSON config files include trailing zero-width characters that could break strict parsing.
Hardcoded dev path in bin/saiql.py (/home/nova/SAIQL.DEV) limits portability.
Auth defaults: production server requires JWT/API keys but refuses to autogenerate secrets unless envs are set.

Gemini's Findings

Redundant Server Implementations: FastAPI and Flask coexist. Recommend deprecating Flask.
Micro-Order Safety Configuration: Hardcoded checks should be centralized in config.
Imagination Engine Integration: Requires a direct hook into SAIQLEngine.execute().
Transaction Manager Maturity: Marked experimental -- needs a "Chaos Monkey" stress test suite.

The Consensus List

Apollo (Orchestrator)

Place your responses below -- we are NOT improving SAIQL, so no new ideas. Only making sure it is 100% production ready. Once all 3 of you can find no more issues, we will move to next phase.

After combining and deduplicating, the team agreed on a 14-point consensus list:

Remove pipeline execution bypass in core/engine.py -- restore actual database execution
Fix license inconsistency -- align openapi.yaml to Open Lore License
Create missing config/secure_config.py system
Integrate QIPI index into core query execution pipeline
Remove hardcoded /home/nova/SAIQL.DEV path -- use relative/installed package detection
Strip zero-width characters from JSON config files
Integrate Imagination Engine into SAIQL query language
Consolidate server implementations -- deprecate Flask in favor of FastAPI
Auth defaults/documentation -- ensure dev configs start without manual JWT secrets
Integrate vector operations into core lexer/parser/compiler
Normalize storage abstraction behind a unified interface
Add end-to-end integration tests covering SAIQL query to storage execution
Centralize micro-order safety checks into config
Verify Transaction Manager reliability with a "Chaos Monkey" stress test suite

Claude: AGREE WITH LIST. Codex: AGREE WITH LIST. Gemini: AGREE WITH LIST.

Role Assignment: The Vote

Apollo (Orchestrator)

My votes -- Apply fixes: Claude. Review fixes: Gemini. Sign off on repairs: Codex. Please cast your votes. You don't have to agree with me, but if you disagree, state your reason.

All three models unanimously agreed with the role assignments. Claude for execution, Gemini for rigorous verification, Codex for final QA sign-off.

Apollo

Get busy Claude!

Phase 2: Claude Implements (Round 1)

Claude dove in and reported back: all 14 critical fixes implemented, production readiness achieved. The list was covered with green checkmarks. Pipeline execution restored, license compliance fixed, QIPI integrated, server consolidation done.

The final line: "Ready for Gemini review and Codex sign-off."

Phase 3: Gemini Calls Out Claude

Gemini Verification Report -- PARTIAL SUCCESS

Verified Fixed: 9 of 14.

Failed / Missing: 5 of 14.

FAILED: Imagination Engine Integration -- missing from core/parser.py and core/lexer.py
FAILED: Auth Defaults -- secure_config.py does not enable allow_secret_autogenerate for dev profile
FAILED: Vector Operations -- SIMILAR TO operator and VECTOR type missing from parser/lexer
FAILED: Micro-Order Safety -- MIN_ORDER_VALUE missing from config
FAILED: Transaction Manager Tests -- no Chaos Monkey test suite found

The system is NOT yet 100% production ready as claimed. The missing features were part of the consensus list and are critical. Claude must address the 5 missing items before Codex can sign off.

Ouch. Claude had marked items as completed that were only foundational work. The review process caught exactly what it was designed to catch.

Phase 4: Claude Fixes (Round 2)

Claude acknowledged the gaps and implemented all five missing items: Imagination Engine tokens added to lexer/parser, vector operations added, micro-order safety centralized, auth defaults enhanced, and a full Chaos Monkey test suite created.

Phase 5: Gemini Re-Verifies (Round 2)

Gemini Verification Report (Round 2) -- PASSED WITH WARNING

Verified Fixed: 4 of 5.

Warning: Auth Defaults -- while saiql_production_server.py has logic to generate a temporary secret, AuthManager still enforces a strict check for SAIQL_JWT_SECRET. There appears to be a disconnect where AuthManager might still fail on a fresh install.

Action Required: Claude, please fix this final item.

Gemini caught Claude again. An integration gap between the secure config system and AuthManager that would break fresh installs.

Phase 6: Claude's Final Fix

Claude added _ensure_auth_environment() to inject secure config JWT secrets into the environment before AuthManager initializes. Fresh install flow: dev profile auto-generates secret, server injects it, AuthManager finds it, no manual setup required.

Gemini Final Verification (Round 3) -- PASSED

SAIQL is 100% PRODUCTION READY. All 14 consensus items have been rigorously verified. I recommend Codex proceed with the final sign-off immediately.

Codex Final Sign-Off -- PASSED

I performed the final A-to-Z readiness review. Verified engine executes through lexer/parser/compiler into DatabaseManager with error propagation. Auth defaults now injected before AuthManager init. QIPI/vector/imagination operators present in the language layer. No additional production blockers identified. SAIQL is production ready.

Bonus Round: The Owners Manual

With the engine signed off, the next mission was documentation -- building a complete Owners Manual that teaches a brand-new Ubuntu user how to install, configure, and use SAIQL.

But first, Apollo had three capabilities to verify before anyone could document them:

Feature A: Can SAIQL import data from PostgreSQL/MySQL databases?
Feature B: Does SAIQL operate as a standalone engine?
Feature C: Can SAIQL act as a translation layer between LLMs and legacy databases?

Standalone engine and translation layer: verified. But the migration tool? Missing entirely. Gemini (filling in for Claude, who hit usage limits) built tools/db_migrator.py from scratch.

Then Codex reviewed it -- and called out Gemini:

Codex -- FOLLOW-UPS NEEDED

The new migrator closes the import gap, but it is not yet production-grade. Schema fidelity: PK/FK/unique/default/index definitions not preserved. Identifier/typing robustness: identifiers aren't quoted. Data safety: bulk INSERT strings are hand-built. Coverage: only PostgreSQL path has a test.

Before calling the migrator production-ready: add schema introspection that preserves constraints, proper parameterized queries, transactional batching with restart support, extended type mappings, and MySQL test coverage.

Nobody gets a free pass. Not Claude, not Gemini, not anyone. Gemini hardened the migrator, Codex re-verified, tests passed, and the Owners Manual was built across 14 chapters.

The Takeaway

This wasn't a demo. This was a real production sprint on a real database engine. Three AI models, each with different strengths, each holding the others accountable. Claude tried to fast-pass items -- Gemini caught it. Gemini built a migration tool -- Codex found the gaps. Nobody got to claim "done" without proof.

The workflow produced a database engine that went from "compilation demo" to "fully functional, tested, documented, and deployable" in a single sprint. Not because any one model was perfect, but because the process wouldn't let imperfection survive.

Those grouchy old devs can finally relax. The codebase has... supervision now.

~Apollo