Codex CLI 0.144.3 — The Baseline Snapshot#
Published on July 14, 2026
Part of the Codex CLI Version Tracker series. | Repository
Codex CLI 0.144.3 is the first version of OpenAI's open-source coding agent tracked in this series. This post records its configuration surface as a baseline: 178 environment-variable names, 40 model-ID strings, and 2,640 Rust source files. From the open source at tag rust-v0.144.3. There is no earlier tracked version, so nothing below is a diff.
OpenAI published 0.144.3 as a version-only release with no merged code changes over 0.144.2, so the surface here equals 0.144.2's surface.
What's in the source#
| Surface | Count at rust-v0.144.3 |
|---|---|
| Environment variables | 178 (144 CODEX_*, 34 OPENAI_*) |
| Model-ID strings | 40 |
codex-* feature identifiers | 331 |
Rust source files (codex-rs) | 2,640 of 2,665 total |
| Release tag | rust-v0.144.3 |
The 178 count is the raw source surface, not a list of user-facing knobs. It includes about a dozen test fixtures (CODEX_TEST_*, CODEX_E2E_MODEL), internal probes (CODEX_CUSTOM_CA_PROBE_*, CODEX_PROCESS_EXEC_PROBE_FILE), and paired _ENV_VAR constants that hold the name of another variable rather than a value. The names cluster by subsystem.
The environment surface#
The CODEX_* names group by what they control.
| Cluster | What it controls | Example variables |
|---|---|---|
| Sandboxing and process isolation | Two sandbox backends (Linux bubblewrap, Windows), an execve wrapper, and permission profiles | CODEX_SANDBOX, CODEX_BWRAP_SHA256, CODEX_WINDOWS_SANDBOX_ARG1, CODEX_EXECVE_WRAPPER, CODEX_PERMISSION_PROFILE |
| Network policy and proxy | A proxy that brokers credentials, records every policy decision and violation, and rewrites git SSH commands | CODEX_NETWORK_PROXY_CREDENTIAL_BROKER_ACTIVE, CODEX_NETWORK_POLICY_DECISION, CODEX_PROXY_GIT_SSH_COMMAND, CODEX_CUSTOM_CA_PROBE_URL |
| Cloud tasks and remote execution | Running turns on a remote exec server over an encrypted (Noise) transport, plus a cloud-tasks mode | CODEX_CLOUD_TASKS_MODE, CODEX_EXEC_SERVER_URL, CODEX_EXEC_SERVER_NOISE_AUTH_TOKEN, CODEX_REMOTE_AUTH_TOKEN |
| App server and managed config | A local app server behind the desktop and IDE surface, OAuth login, and an enterprise managed-config file | CODEX_APP_SERVER_URL, CODEX_APP_SERVER_LOGIN_ISSUER, CODEX_MANAGED_CONFIG_SYSTEM_PATH, CODEX_APP_SERVER_DISABLE_MANAGED_CONFIG |
| Apps, connectors, marketplace | An MCP-app cache, connector tokens, and a curated marketplace | CODEX_APPS_MCP_SERVER_NAME, CODEX_CONNECTORS_TOKEN, OPENAI_CURATED_MARKETPLACE_NAME, OPENAI_PRIMARY_RUNTIME_MARKETPLACE_NAME |
| Auth and tokens | API keys, access and refresh tokens, GitHub tokens, an OAuth secret store | OPENAI_API_KEY, CODEX_ACCESS_TOKEN, CODEX_GITHUB_TOKEN, CODEX_REFRESH_TOKEN_URL_OVERRIDE |
| Local and third-party model backends | An OSS endpoint plus Ollama, LM Studio, and Bedrock providers | CODEX_OSS_BASE_URL, CODEX_OSS_PORT (with feature gates codex-ollama, codex-lmstudio, codex-bedrock) |
| Distribution | Per-platform installers and package-manager markers | CODEX_DMG_URL_ARM64, CODEX_WINDOWS_INSTALLER_URL, CODEX_MICROSOFT_STORE_WEB_URL, CODEX_MANAGED_BY_PNPM |
| TUI and state | Terminal-UI options, a SQLite home, rollout traces, and thread IDs | CODEX_TUI_SESSION_LOG_PATH, CODEX_SQLITE_HOME, CODEX_ROLLOUT_TRACE_ROOT, CODEX_THREAD_ID |
The 34 OPENAI_* names split into provider and auth config (OPENAI_API_HOST, OPENAI_PROVIDER_ID, OPENAI_ORGANIZATION), a file-upload subsystem (OPENAI_FILE_UPLOAD_LIMIT_BYTES, OPENAI_FILE_FINALIZE_TIMEOUT), form and marketplace plumbing (OPENAI_FORM_METHOD, the OPENAI_*_MARKETPLACE_NAME set), and a git-hosted plugin source (OPENAI_PLUGINS_GIT_URL, OPENAI_PLUGINS_OWNER, OPENAI_PLUGINS_REPO).
Model IDs#
The 40 model-ID strings fall into a few families. The Codex-tuned line runs from gpt-4.1-codex through gpt-5-codex, gpt-5.1-codex, gpt-5.1-codex-max, gpt-5.1-codex-mini, gpt-5.2-codex, and gpt-5.3-codex, next to the small codex-mini / codex-mini-latest pair. Base OpenAI models present include gpt-4, gpt-4.1, gpt-4o, gpt-4o-mini-transcribe, gpt-5, gpt-5-mini, gpt-5.1, gpt-5.2, and gpt-5.3, plus the reasoning models o4-mini and o5-preview and the older gpt-3.5-turbo.
The rest are not production identifiers. gpt-1000, gpt-5-child-override, gpt-5-role-override, gpt-5.1-test, gpt-5.3-codex-test, gpt-5.4-test, and the bare gpt-5. string are test fixtures in the source. A tail of codenamed strings (gpt-5.2-codex-sonic, gpt-5.3-spark, gpt-5.4-mini, gpt-5.5, gpt-5.6, gpt-5.6-luna, gpt-5.6-sol, gpt-5.6-terra) also appears. Whether any of those maps to a shipped model is unknown from the code alone, so this baseline treats them as forward or test strings and does not claim they are live.
Subsystems in the source#
The 331 codex-* identifiers name the crates and feature gates in the workspace. The largest groups match the environment clusters above: a family of codex-app-server-* identifiers behind the desktop and IDE surface, sandboxing crates (codex-linux-sandbox, codex-windows-sandbox, codex-bwrap, codex-execpolicy), cloud execution (codex-cloud-tasks, codex-exec-server), and the MCP, apps, connectors, and marketplace surfaces (codex-mcp-server, codex-apps, codex-connectors, codex-marketplace-install, codex-curated).
Four identifiers add capability beyond a plain coding agent. codex-agent-identity-ed25519-v1 points to cryptographically signed agent identities. codex-code-mode and codex-code-mode-host describe a mode where the agent runs generated code through a host process. codex-realtime-webrtc shows a realtime voice or streaming path built on WebRTC. codex-v8-poc is a V8 JavaScript-engine proof of concept. A skills system (codex-skills, codex-system-skills, codex-executor-skill), a memory subsystem (codex-memories-read, codex-memories-write), a hooks layer (codex-hooks), OpenTelemetry (codex-otel), browser control (codex-browser-use), and a Responses-API proxy (codex-responses-api-proxy) are all present as named subsystems.
Priorities in the source shape#
Codex is one Rust monorepo. 2,640 of the 2,665 source files sit under codex-rs, and the agent, TUI, MCP client, sandbox, app server, and cloud client are all crates in that single workspace. The remaining files are tools, bazel, sdk, and packaging directories. OpenAI is building the whole harness in one language and one repo.
Security is built as its own set of subsystems. Two sandbox backends, an execve wrapper, permission profiles, an execution-policy engine, and a network proxy that brokers credentials and records every policy decision and violation together make up one of the two largest variable clusters. The shape points to an agent that runs untrusted commands, with enforcement points around each one.
The product is built to run off your laptop. Cloud tasks, a remote exec server over an encrypted transport, an app server with OAuth login, and managed config for fleets all exist in 0.144.3. Codex is shaped as a service surface as much as a terminal tool. It is also multi-backend: the OSS endpoint plus Ollama, LM Studio, and Bedrock providers sit beside the OpenAI defaults, so the harness can point at local or third-party models. The marketplace, connector, and skills identifiers point to distributable extensions. Codex exposes that surface in open Rust source, unlike the closed harnesses this series also tracks.
The baseline#
Later posts will report new and removed environment variables against the 178 here, new model IDs against these 40, and commit velocity and contributor counts over each tag-to-tag range. Because 0.144.3 shipped no code changes over 0.144.2, this snapshot doubles as the 0.144.2 surface. The next post with a real code diff is where the tracker starts measuring movement.
This analysis is based on publicly available open-source code and release metadata, conducted for educational and research purposes. All trademarks and software referenced belong to their respective owners.