claude-codex-orchestrator
3 starsAn orchestrator and executor split across two models. Claude plans, decomposes and verifies; Codex writes the code once the spec is frozen, and makes no design calls of its own.
I build systems where AI agents do real engineering work.
Deterministic code intelligence, a fleet-wide ontology, and MCP tooling for spec-driven SDLC. Platform engineering at Genea.
Agents fail on missing context, not missing intelligence.
So I stopped asking models to guess at a codebase and started handing them ground truth: a queryable graph of what the system actually is, answered in milliseconds with no model in the loop. Retrieval that is deterministic and verifiable beats retrieval that is merely plausible.
What I build
Infrastructure that makes autonomous engineering work repeatable rather than lucky.
112k+ nodes
A fleet-wide code ontology
One graph unifying services, events, routes, models, database columns, frontend components and mobile screens. Cross-service questions resolve by graph traversal in under 100 ms, with no model in the loop and no answer that cannot be traced back to source.
22 MCP tools
Tooling an agent can hold
A thin surface consolidated down from more than a hundred tools. Fewer, higher-leverage capabilities cut the attention an agent burns on choosing, and every error comes back as a structured envelope it can correct against.
48 commands
Spec-driven SDLC
Spec to plan to scope to tasks to implementation to review, with an engineering constitution loaded on every run so standards get enforced rather than remembered.
Multi-model orchestration
Each task routed to the cheapest tier that can do it well, with judgment reserved for the moments where a mistake compounds. Read-only lanes advise and review; only write lanes touch code, and every diff gets verified. A task that fails twice escalates carrying its failure log, rather than retrying blind.
Production platform engineering
The day job underneath the agent work: event-driven microservices in Python, Node and Ruby over Postgres, Redis and Elasticsearch, running on Kubernetes across AWS and GCP.
Agent harness and hooks
Lifecycle hooks, session context injection and structure-only telemetry, so the harness itself can be measured and tuned instead of guessed at.
Open source
Tools pulled out of daily agent work and published under MIT.
claude-session-replay
A video player for what your agent actually did. It turns a session transcript into one self-contained HTML file you can scrub, step through and hand to a teammate. No dependencies to view, and nothing leaves your machine until you share the file.
View on GitHubclaude-router
2 starsRoute each task to the cheapest model tier that can do it well, through model-pinned subagents. Cheap models for volume, premium ones only where a mistake compounds.
Day to day
PythonNode.jsRuby
PostgreSQLMongoDBRedis
ElasticsearchDockerKubernetes
AWSGCP
Earlier work from a Hadoop and data-engineering chapter still lives on GitHub. The most-used piece is a Java payroll management system that picked up 24 stars.