elizaos--eliza/plugins/plugin-codex-cli/AGENTS.md

@elizaos/plugin-codex-cli

ChatGPT Codex model provider plugin for elizaOS — routes text generation through a user's ChatGPT Plus/Pro subscription using the OAuth token cache written by the official codex CLI.

Purpose / role

This plugin registers model handlers so Eliza agents can use ChatGPT Codex models (gpt-5, gpt-5.5, etc.) as their inference backend. It is not auto-enabled by an env var; it activates when an auth profile in the runtime config sets provider: "codex-cli", or when agents.defaults.subscriptionProvider is "openai-codex". It is node-only ("platforms": ["node"]).

Plugin surface

No actions, providers, evaluators, or routes are registered. The plugin registers model handlers only:

Model type registered	What it does
TEXT_SMALL	Codex-backed small text generation
TEXT_NANO	Codex-backed nano text generation
TEXT_MEDIUM	Codex-backed medium text generation
TEXT_LARGE	Codex-backed large text generation
TEXT_MEGA	Codex-backed mega text generation
RESPONSE_HANDLER	Codex-backed response handler model
ACTION_PLANNER	Codex-backed action planner model

All model types delegate to generateTextWithCodex() in index.ts, which calls CodexBackend.generate(). Streaming is supported via TextStreamResult.

Layout

plugins/plugin-codex-cli/
  index.ts                  Plugin entry point — registers model handlers, exports Plugin object
  index.node.ts             Node build re-export (re-exports index.ts default)
  index.browser.ts          Unsupported browser export; plugin is node-only
  auto-enable.ts            Auto-enable module (shouldEnable + shouldForce); referenced by package.json elizaos.plugin.autoEnableModule
  src/
    codex-backend.ts        CodexBackend class — HTTP client for ChatGPT Codex /responses SSE endpoint; FIFO queue + jitter
    codex-auth.ts           OAuth token load/save/refresh with file-level lock; isExpired() JWT check
    sse-parser.ts           Spec-compliant SSE AsyncGenerator parser for ReadableStream<Uint8Array>
    tool-format-openai.ts   toOpenAITool() / toOpenAITools() — maps elizaOS ToolDefinition to OpenAI function tool shape
  __tests__/
    codex-backend.test.ts   Unit tests for CodexBackend, message translation, and auth
  build.ts                  Bun.build script (node + browser bundles via Bun bundler; declarations via tsc)
  vitest.config.ts          Test config

Commands

All commands run from the plugin directory:

bun run --cwd plugins/plugin-codex-cli build        # compile (Bun.build + tsc for declarations)
bun run --cwd plugins/plugin-codex-cli dev          # watch build
bun run --cwd plugins/plugin-codex-cli test         # vitest run
bun run --cwd plugins/plugin-codex-cli lint         # biome check --write --unsafe
bun run --cwd plugins/plugin-codex-cli lint:check   # biome check (no write)
bun run --cwd plugins/plugin-codex-cli format       # biome format --write
bun run --cwd plugins/plugin-codex-cli typecheck    # tsgo --noEmit
bun run --cwd plugins/plugin-codex-cli clean        # rm -rf dist .turbo

Config / env vars

Var	Required	Default	Description
CODEX_AUTH_PATH	No	~/.codex/auth.json	Path to codex CLI OAuth cache file
CODEX_BASE_URL	No	https://chatgpt.com/backend-api/codex	Codex backend base URL; only chatgpt.com or localhost accepted
CODEX_MODEL	No	gpt-5.5	Model identifier; must be one of the supported models
CODEX_JITTER_MS_MAX	No	200	Max pre-request jitter in ms; set 0 to disable
CODEX_ORIGINATOR	No	codex_cli_rs	Originator header sent to the backend
CODEX_USER_AGENT	No	codex_cli_rs/0.124.0	User-Agent header (env only; not in agentConfig)

Settings are read via runtime.getSetting() first, falling back to process.env. None are required; the plugin will fail at request time if CODEX_AUTH_PATH is missing or the auth file is absent/expired with no valid refresh token.

No auto-enable env var trigger exists. Auto-enable logic lives in auto-enable.ts:

• shouldEnable: any auth profile has provider === "codex-cli".
• shouldForce: agents.defaults.subscriptionProvider === "openai-codex".

How to extend

Add a new model type handler:

1. Import the new ModelType constant from @elizaos/core in index.ts.
2. Add an entry to the codexModels object following the existing pattern: [ModelType.NEW_TYPE]: (runtime, params) => generateTextWithCodex(runtime, params, ModelType.NEW_TYPE).
3. Declare it in package.json under elizaos.plugin.capabilities if appropriate.

Swap the backend or auth mechanism:

• CodexBackend accepts loadAuth, refreshAuth, fetchImpl, and toolTranslator overrides in its constructor config — use these in tests and when integrating alternate auth flows.

Add a provider or action:

• This plugin intentionally has no actions or providers. Adding one follows the standard elizaOS pattern: define it, add it to the Plugin object's actions or providers array in index.ts.

Conventions / gotchas

• Node-only. The browser export (index.browser.ts) exists to satisfy the build; the plugin will not function in a browser runtime. The src/codex-auth.ts module uses node:fs, node:crypto, and node:os.
• Single in-flight FIFO queue. CodexBackend chains requests via this.tail promise — all calls on the same backend instance serialize. One CodexBackend instance per runtime is maintained via backendByRuntime WeakMap in index.ts.
• Base URL validation. CODEX_BASE_URL must target https://chatgpt.com or localhost; any other host throws at backend construction to prevent OAuth token exfiltration.
• 401 auto-refresh. On a 401 from the Responses endpoint, the backend refreshes the OAuth token (with a file lock) and retries exactly once.
• Tool calls return an object, not a string. When params.tools is non-empty, params.messages is provided, or the backend returns tool calls, generateTextWithCodex returns TextResultWithNativeTools ({ text, toolCalls, finishReason, usage }) rather than a plain string.
• responseSchema support. When callers pass a responseSchema and no explicit responseFormat, the backend wraps it in { type: "json_schema", schema } for the OpenAI Responses API structured output format.
• Per-call model override. Text handlers honor params.model before CODEX_MODEL. Workflow generation uses this for isolated Codex model tests without changing every Codex CLI text call.
• Auth file lock. codex-auth.ts uses a .lock file alongside auth.json (30 s stale timeout, 30 retries at 100 ms). Stale locks are cleaned up automatically.
• Supported models: gpt-5, gpt-5-codex, gpt-5.4, gpt-5.5, gpt-5.5-pro. Setting CODEX_MODEL to anything else will be sent as-is; the backend may reject it.
• See the root AGENTS.md for repo-wide architecture rules, logger conventions, and ESM requirements.

⛔ NON-NEGOTIABLE — evidence, trajectories & real end-to-end tests

The binding, repo-wide standard is AGENTS.md. Read it. Nothing in this package is done until it is proven done — a reviewer must confirm it works without reading the code, from the artifacts you attach. This applies to every feature, fix, refactor, and chore here. "Tests pass" is not proof; "CI is green" is not proof.

• Record AND read model trajectories. Capture the actual inputs and outputs of the model from a live LLM — not the deterministic proxy, not a mock: the prompt, the providers/context, the raw model output, every tool/action call, and the result. Then open the trajectory and review it by hand. A captured-but-unread trajectory is not evidence (packages/scenario-runner/bin/eliza-scenarios run <scenario> --report <out>).
• Real, full-featured E2E — no larp. Every feature ships detailed end-to-end tests that drive the real path end to end. Not the happy "front door" only: cover error paths, edge/empty/invalid input, concurrency, roles/permissions, and adversarial input. A test that asserts against a mock/stub/fixture standing in for the thing under test does not count. If the real model/device/chain/connector/account is hard to reach, make it reachable — that is the work, not an excuse to mock. If the existing tests here are shallow or mocked, fixing them is part of your change.
• Screenshots + logs at every phase, plus a complete walkthrough video/run-through of the entire feature or view, start to finish (bun run test:e2e:record).
• Manually review every artifact the change touches — never just the green check: client logs (console + network), server logs ([ClassName] …), the model trajectories in and out, before/after full-page screenshots, and the domain artifacts listed below for this package.
• No residuals. No shortcuts. The goal is not "done" — it is everything done. Clear every blocker by the hard path: build the real architecture, stand up the real model/device/service, actually test it. Never leave a TODO, a stub, a stepping-stone, or a "follow-up." When unsure, research thoroughly, weigh the options, and ship the best, highest-effort, production-ready version. Keep going until every possibility is exhausted.

Artifacts → attached inline in the PR (MP4 video, JPG screenshots, logs in <details>); attach each evidence type or explicitly mark it N/A with a reason — never leave it blank. If develop moved and changed behavior, re-capture evidence; stale proof is worse than none.

Capture & manually review for this package — CLI / tooling:

• The real command/flow invocation transcript (args in, stdout/stderr, exit code) and the artifacts it generated (files, scaffolds, manifests, screenshots/recordings).
• Failure paths: bad args, missing deps, partial state, permission/network errors.
• A recording/log of the actual run end to end — not a unit test of one helper.
• Any model interaction captured as a live trajectory and reviewed.