The coding agent, modeled.
603,461 lines of Codex as a product.
GitMir interpreted OpenAI's Codex CLI into a verified Object Context: how the agent core, tools, sandboxing and approvals, execution policy, MCP and model providers actually work — as 12,442 linked objects, not 2,307 files.
84/84 TASKS COMPLETED · ALL TEN DIMENSIONS FILLED · ZERO INTEGRITY FAILURES
Explore the drawing.
Open Sandboxing and approvals to see how escalation actually works; follow the journeys of a tool call; read the 33 lifecycles the agent's state machines walk.
THE DRAWING IS LIVE: OPEN AN AREA TO SEE WHAT IS INSIDE IT, SWITCH TO JOURNEYS, DATA AND LIFECYCLES — EVERY OBJECT IS SERVED FROM /api/models/codex.
Ten dimensions.
12,442 objects.
SCOPE EXCLUDES TESTS (965,801 LINES), GENERATED CODE AND CI — MEASURED BY THE INGEST CENSUS, NOT ESTIMATED. DOMAINS INCLUDE: AGENT CORE · TOOLS · COMMAND EXECUTION · SANDBOXING AND APPROVALS · EXECUTION POLICY · MCP · MODEL PROVIDERS · TERMINAL UI.
An agent, explained to agents.
The most asked-about codebase of the year
Everyone extends, wraps or competes with coding agents. The Codex model answers the questions its README cannot: which state machines govern a session, what the approval policy really permits, and which journeys a tool call walks.
A stress test for the method
Dense Rust internals, 624 events, 33 lifecycles — a different shape from Supabase Studio's product surface. The same pipeline filled all ten dimensions with zero integrity failures.
The measurements above come from the run ledger and the model index. The founder-attested measurement record for this run is in preparation — until it is signed, treat the figures as verified-by-tooling, not attested.