Human
Intent, policy, irreversible choices
Advisory governance for Codex
Engineering judgment,
with evidence.
Codex Cortex helps Codex form better contracts, choose the right workflow, catch material drift, and verify what can actually be proven—without becoming a second production writer.
- Form
- Skills-first local plugin
- Core
- 9 routed skills · 5 deterministic tools
- Pilot
- Codex Desktop + CLI on Windows 11
The governing issue
Codex can move fast.
The hard part is knowing where it should stop.
Ambiguous authority
Product intent, repository rules, runtime facts and model judgment can blur into one confident answer.
Verification inflation
A plausible explanation or a broad test suite gets mistaken for proof of the exact claim that matters.
Governance sprawl
More agents, gates and process accumulate—even when a smaller, well-bounded intervention would do.
CORTEX’S ANSWER
Put a thin advisory layer around Codex: one that separates authority, routes by intent, challenges only material drift, and mechanizes only the facts that benefit from deterministic guarantees.
A different category
Not more agents.
Better boundaries.
Cortex is not a prompt library, a worker swarm, or a workflow engine. It is a governance peer that helps Codex make better engineering decisions while leaving implementation ownership where it belongs.
Generic prompting
Helpful instructions
Improves a response, but rarely establishes durable authority or evidence boundaries.
Agent swarms
More parallel actors
Useful for independent work, but concurrency alone does not resolve who is allowed to decide what.
Heavy orchestration
More managed process
Can coordinate complex systems, but often adds persistent state and machinery the task never earned.
Codex Cortex
Bounded advisory governance
Routes the smallest useful intervention, keeps one production writer, and escalates only genuine authority disputes.
How it works
A small loop.
Clear ownership.
Users invoke Cortex through ordinary intent. Cortex selects the relevant workflow internally; you do not need to know the names of its skills.
- 01
CONTRACT
Make the task’s edges explicit.
Clarify the goal, authoritative inputs, allowed change shape, invariants, external effects, verification oracle and acceptance boundary.
Unspecified consequential behavior defaults to none. - 02
ROUTE
Choose the smallest useful workflow.
Build, review, migrate, simplify, transform, research or advise—selected from intent rather than exposed as a menu of modes.
No manufactured roles, hooks or telemetry. - 03
VERIFY
Prove mechanical claims mechanically.
Use tests, parsers, schemas, inventories, diffs and fingerprints where exact evidence is possible. Keep judgment in the reasoning layer.
Verification is distinct from human acceptance. - 04
RECONCILE
Correct once, then escalate cleanly.
Cortex issues a bounded objection. Codex corrects, diagnoses, reroutes or rebuts. Cortex reassesses once; only unresolved authority reaches the human.
The disputed surface stays narrow.
Core capabilities
Governance that stays
out of the way.
Authority separation
Distinguishes human intent, Cortex reasoning, Codex execution, deterministic proof, and repository/runtime truth.
Bounded correction
Gives Codex one focused opportunity to correct or rebut, preventing circular review loops and sprawling rewrites.
Intelligent routing
Selects the applicable workflow from ordinary language, including project formation, review, migration and checkpoint advice.
Deterministic verification
Uses five narrow local tools for inspection, rendering, validation, structural diffing and task-local intervention traces.
Practical reference injection
Adds a tiny, annotated reference packet only when authoritative external evidence materially improves the method or decision.
Human escalation
Returns genuine policy, destructive, irreversible or unresolved architectural choices to the person who owns them.
Design principles
Reliability without
governance theatre.
- 01
One production writer
Cortex advises. Codex implements. Review does not create a second source of code changes.
- 02
Evidence before confidence
Mechanically testable claims require deterministic evidence. Repository and runtime truth remain authoritative.
- 03
The smallest correct change
Fix the earliest owning layer and keep unspecified persistence, dependencies, I/O and compatibility behavior at none.
- 04
Machinery must earn its place
No persistent backend, continuous edit surveillance, generic Stop gate, or self-modifying policy engine in v1.
Example use cases
Ask in ordinary
engineering language.
Start with “Use Cortex to…” and describe the outcome. Cortex handles internal routing; explicit boundaries such as read-only, preserve behavior, or stop for acceptance remain authoritative.
BUILD
Use Cortex to build the smallest Codex-ready control plane for this new repository from these requirements.
REVIEW
Use Cortex to review this repository’s agent instructions, hooks and verifiers for contradictions and authority leaks. Read-only.
MIGRATE
Use Cortex to align this existing Codex setup while preserving supported behavior and repository authority.
SIMPLIFY
Use Cortex to retire stale or duplicated governance only where evidence supports removal. Preserve live product policy.
ADVISE
Use Cortex for a pre-closure review of this implementation. Raise only material objections and keep correction bounded.
Ready for a better boundary?
Give Codex a clearer
operating model.
Install the local plugin, start a fresh Codex Desktop task or CLI session, and invoke Cortex through ordinary intent. Windows 11 is the validated v1 pilot host.
codex plugin marketplace add .codex plugin add codex-cortex@codex-cortex-localNo remote backend. No account service. No dependency on Cortex after a generated package is handed off.