Codex SDK Logo

Codex SDK for Elixir

CIElixirOTPHex.pmDocumentationLicense

An idiomatic Elixir SDK for embedding OpenAI's Codex agent in your workflows and applications. This SDK wraps the codex-rs executable, providing a complete, production-ready interface with streaming support and comprehensive event handling.

Documentation Menu

Features

Installation

Add codex_sdk to your list of dependencies in mix.exs:

def deps do
  [
    {:codex_sdk, "~> 0.16.0"}
  ]
end

Prerequisites

You must have the codex CLI installed. Install it via npm or Homebrew:

# Using npm
npm install -g @openai/codex

# Using Homebrew
brew install codex

The SDK does not vendor codex-rs; it shells out to the codex executable on your system. Path resolution follows this order:

  1. codex_path_override supplied in Codex.Options.new/1
  2. CODEX_PATH environment variable
  3. System.find_executable("codex")

Make sure the binary at the resolved location is executable and kept up to date. When the resolved path is a version-manager shim (for example asdf/mise), the SDK stabilizes it to the underlying installed executable when possible so child subprocess launches do not depend on the child working directory.

For authentication, sign in with your ChatGPT account (this stores credentials for the CLI):

codex
# Select "Sign in with ChatGPT"

Alternatively, set CODEX_API_KEY before starting your BEAM node. For normal CLI-backed SDK execution, auth resolution is:

  1. CODEX_API_KEY
  2. auth.jsonOPENAI_API_KEY
  3. ChatGPT OAuth tokens stored under CODEX_HOME (default ~/.codex/auth.json, with legacy credential file support)

The SDK now also exposes native OAuth login via Codex.OAuth:

{:ok, result} =
  Codex.OAuth.login(
    storage: :file,
    interactive?: true
  )

Persistent Codex.OAuth login writes upstream-compatible auth.json and respects upstream auth_mode. Memory-only OAuth is also available for host-managed and app-server external auth flows. openai_base_url does not change the OAuth issuer; use auth_issuer only when you need to override the login authority itself.

Environment-aware OAuth behavior matches current native-app guidance:

ChatGPT plan types are normalized before they surface through SDK auth/status structs or app-server external-auth forwarding. In particular, hc and enterprise normalize to "enterprise", while education and edu normalize to "edu".

If cli_auth_credentials_store = "keyring" is set in config and keyring support is unavailable, the SDK logs a warning and skips file-based tokens (remote model fetch falls back to bundled models). When cli_auth_credentials_store = "auto" and keyring is unavailable, the SDK falls back to file-based auth.

When an API key is supplied, the SDK forwards it as both CODEX_API_KEY and OPENAI_API_KEY to the codex subprocess to align with provider expectations.

Base URL precedence is: explicit :base_url in Codex.Options.new/1, then layered openai_base_url from config.toml, then OPENAI_BASE_URL, then the OpenAI default (https://api.openai.com/v1). User-defined [model_providers] entries extend the built-in provider set, but reserved built-ins such as openai, ollama, and lmstudio cannot be redefined.

Custom trust roots use CODEX_CA_CERTIFICATE first and SSL_CERT_FILE second. Blank values are ignored. The same PEM bundle is applied consistently to Codex CLI subprocesses, direct HTTP clients, remote model fetches, MCP HTTP/OAuth, realtime websockets, and voice HTTP requests.

Centralized Model Selection

codex_sdk no longer owns the active model catalog, fallback rules, or default selection policy. That authority now lives in cli_subprocess_core.

The authoritative path is:

Codex.Options.new/1 now delegates mixed-input normalization to CliSubprocessCore.ModelInput.normalize/3, then projects the current model and reasoning_effort from the authoritative shared model_payload. Codex.Models is now a read-only projection of the shared core catalog. It no longer owns a separate catalog or a separate fallback/defaulting path.

If a caller supplies an explicit model_payload, that payload stays authoritative. Repo-local env defaults such as CODEX_MODEL, CODEX_PROVIDER_BACKEND, and CODEX_OLLAMA_BASE_URL are fallback inputs only when the payload was not supplied explicitly.

Operationally, that means:

Local Ollama Through The Shared Core Contract

codex_sdk now consumes the core-owned Codex OSS payload for local Ollama.

Use:

{:ok, opts} =
  Codex.Options.new(%{
    model: "llama3.2",
    provider_backend: :oss,
    oss_provider: "ollama"
  })

That causes the shared core registry to:

The SDK does not infer those flags on its own.

If ollama_base_url: is supplied, that endpoint is carried inside the payload-owned env overrides as CODEX_OSS_BASE_URL. Raw Ollama roots are normalized to the OpenAI-compatible /v1 base before those env overrides are emitted. The exec and app-server transports both consume that payload data directly instead of keeping a second raw base-url path alive downstream.

When the chosen local model is outside Codex's built-in model metadata catalog, the upstream CLI may warn that it is using fallback metadata. That is an upstream degraded-mode distinction, not a hard model rejection in codex_sdk.

For the stateful app-server transport, the same resolved payload is rendered into supported codex app-server --config ... startup overrides plus thread/startmodelProvider selection. The SDK does not pass unsupported exec-only OSS flags to codex app-server.

./examples/run_all.sh --ollama uses that same route. It runs the CLI-backed example suite against local Ollama and skips the direct OpenAI realtime/voice examples, which are a separate subsystem and are not Ollama-backed.

Use Codex.Models.default_model/0, Codex.Models.list_visible/1, and Codex.Models.default_reasoning_effort/1 as convenience readers over that shared contract.

See the OpenAI Codex documentation for more authentication options.

Quick Start

Basic Usage

# Start a new conversation
{:ok, thread} = Codex.start_thread()

# Run a turn and get results
{:ok, result} = Codex.Thread.run(thread, "Explain the purpose of GenServers in Elixir")

# Access the final response
IO.puts(result.final_response)

# Inspect all items (messages, reasoning, commands, file changes, etc.)
IO.inspect(result.items)

# Continue the conversation
{:ok, next_result} = Codex.Thread.run(thread, "Give me an example")

App-server Transport (Optional)

The SDK defaults to the :exec compatibility selector for the core-backed exec JSONL lane. To use the stateful app-server transport:

{:ok, codex_opts} = Codex.Options.new(%{api_key: System.fetch_env!("CODEX_API_KEY")})
{:ok, conn} = Codex.AppServer.connect(codex_opts)

{:ok, thread} =
  Codex.start_thread(codex_opts, %{
    transport: {:app_server, conn},
    working_directory: "/project"
  })

{:ok, result} = Codex.Thread.run(thread, "List the available skills for this repo")

{:ok, %{"data" => skills}} = Codex.AppServer.skills_list(conn, cwds: ["/project"])

When you need experimental app-server fields such as approvals_reviewer or granular approval policies, create the connection with Codex.AppServer.connect(codex_opts, experimental_api: true).

When the managed codex app-server child should run against an isolated repo or temporary Codex home, pass launch overrides to connect/2 itself:

tmp_home = Path.join(System.tmp_dir!(), "codex-sdk-app-server-home")

{:ok, conn} =
  Codex.AppServer.connect(codex_opts,
    cwd: "/project",
    process_env: %{
      "CODEX_HOME" => tmp_home,
      "HOME" => Path.dirname(tmp_home),
      "USERPROFILE" => Path.dirname(tmp_home)
    }
  )

cwd and process_env apply to the app-server child process. Per-thread working directories still belong on working_directory / cwd thread params.

For a managed remote app-server websocket instead of a local codex app-server child, use Codex.AppServer.connect_remote/2:

{:ok, conn} =
  Codex.AppServer.connect_remote(
    "wss://app-server.example/ws",
    auth_token_env: "CODEX_REMOTE_AUTH_TOKEN",
    client_name: "my_app",
    experimental_api: true
  )

connect_remote/2 keeps the same pid contract as connect/2, so the existing Codex.AppServer.* request helpers, disconnect/1, alive?/1, subscribe/2, unsubscribe/1, and respond/3 work unchanged. Bearer auth headers are only attached for wss:// or loopback ws:// endpoints; plain non-loopback ws:// plus auth_token or auth_token_env is rejected. Remote OAuth only supports oauth: [storage: :memory]; persistent child-login preflight (:file / :auto) is not available because remote mode does not spawn a local child or child CODEX_HOME.

connect/2 also supports OAuth-aware child auth bootstrapping:

{:ok, conn} =
  Codex.AppServer.connect(codex_opts,
    experimental_api: true,
    process_env: %{"CODEX_HOME" => tmp_home},
    oauth: [
      mode: :auto,
      storage: :memory,
      auto_refresh: true
    ]
  )

For oauth: [storage: :file | :auto], the SDK resolves auth against the effective child CODEX_HOME before launching the child. For oauth: [storage: :memory], it starts the child, logs in with external chatgptAuthTokens, and attaches a connection-owned refresh responder. Set auto_refresh: false when you want to handle account/chatgptAuthTokens/refresh requests yourself.

Multi-modal input is supported on app-server transport:

input = [
  %{type: :text, text: "Explain this screenshot"},
  %{type: :local_image, path: "/tmp/screenshot.png"}
]

{:ok, result} = Codex.Thread.run(thread, input)

Note: the :exec compatibility lane still accepts text input only; list inputs return {:error, {:unsupported_input, :exec}}.

App-server-only APIs include:

On app-server transport, thread options now forward current upstream routing fields such as ephemeral, service_name, and service_tier; turn options can override service_tier per Codex.Thread.run/3. Raw plugin response maps still preserve newer upstream auth metadata such as needsAuth, while the typed plugin API projects those payloads into Codex.Protocol.Plugin.* structs and preserves unknown upstream fields in extra maps. Subscriptions adapt mcpServer/startupStatus/updated into typed Codex.Events structs.

Runnable app-server demos now include examples/live_app_server_filesystem.exs for fs/* and examples/live_app_server_plugins.exs for plugin/list + plugin/read using a disposable repo-local marketplace fixture plus an isolated temporary CODEX_HOME, rather than your real plugin config; that example now uses the typed plugin wrappers and prints derived needs_auth state from the typed app summaries while the raw wrappers remain available. examples/live_app_server_approvals.exs uses the same child-process isolation pattern to enable the under-development approval features only inside a temporary CODEX_HOME, so it can exercise live command/file/permissions approval flows without mutating your real Codex settings or writing inside this repository.

Local plugin authoring is a separate surface from those runtime wrappers:

{:ok, scaffold} =
  Codex.Plugins.scaffold(
    cwd: "/repo/root",
    plugin_name: "demo-plugin",
    with_marketplace: true,
    skill: [name: "hello-world", description: "Greets the user"]
  )

{:ok, marketplace} = Codex.Plugins.read_marketplace(scaffold.marketplace_path)

Use Codex.Plugins.* to create and update .codex-plugin/plugin.json, .agents/plugins/marketplace.json, and minimal local plugin trees with normal Elixir file IO. Use Codex.AppServer.plugin_* later if you want runtime verification against a running codex app-server. Normal authoring flows do not route through app-server fs/*. Phase-1 scaffold intentionally stops at the plugin tree, optional skill stub, and marketplace entry; it does not generate mix.exs, Dialyzer/PLT helpers, or build_support/*.

Raw versus typed plugin calls:

alias Codex.AppServer
alias Codex.Protocol.Plugin

{:ok, raw} =
  AppServer.plugin_read(conn, "/tmp/marketplace.json", "demo-plugin")

{:ok, %Plugin.ReadResponse{plugin: plugin}} =
  AppServer.request_typed(
    conn,
    "plugin/read",
    %Plugin.ReadParams{
      marketplace_path: "/tmp/marketplace.json",
      plugin_name: "demo-plugin"
    },
    Plugin.ReadResponse
  )

IO.inspect(raw["plugin"]["apps"], label: "raw apps")
IO.inspect(plugin.apps, label: "typed apps")

App-server v2 input blocks support text, image, localImage, skill, and mention. Legacy app-server v1 conversation flows are available via Codex.AppServer.V1.

Experimental feature enablement is forwarded without a stale local allowlist:

{:ok, %{"data" => features}} = Codex.AppServer.experimental_feature_list(conn)

{:ok, _} =
  Codex.AppServer.experimental_feature_enablement_set(conn,
    apps: true,
    plugins: false
  )

The SDK forwards the enablement map as given and lets the server validate the current supported keys.

Raw CLI Passthrough and Interactive Sessions

Use Codex.CLI.run/2 when you want literal command-surface parity with the upstream terminal client, and Codex.CLI.interactive/2 or Codex.CLI.start/2 when you need a long-running or PTY-backed session.

Under the hood, Codex.CLI.run/2 and the synchronous wrapper functions ride the shared CliSubprocessCore.Command lane. Codex.CLI.Session, Codex.AppServer, and Codex.MCP.Transport.Stdio preserve their public Codex entrypoints while mapping raw PTY, stdio transport, stdin, stderr, interrupt, and exit lifecycle onto CliSubprocessCore.RawSession.

The ownership line is now:

When codex_sdk is installed alongside agent_session_manager, ASM auto-detects the runtime kit and activates ASM.Extensions.ProviderSDK.Codex in ASM.Extensions.ProviderSDK.available_extensions/0 and ASM.Extensions.ProviderSDK.capability_report/0. That ASM seam is only a bridge into Codex-native helpers such as app-server entrypoints; the actual app-server, MCP, realtime, and voice APIs remain here.

{:ok, codex_opts} = Codex.Options.new(%{})

# Safe one-shot command wrappers
{:ok, completion} = Codex.CLI.completion("zsh", codex_opts: codex_opts)
IO.puts(completion.stdout)

{:ok, features} = Codex.CLI.features_list(codex_opts: codex_opts)
IO.puts(features.stdout)

# Arbitrary raw command surface
{:ok, result} =
  Codex.CLI.run(
    ["cloud", "list", "--json"],
    codex_opts: codex_opts
  )

IO.puts(result.stdout)

# Prompt-mode root codex session over a PTY
{:ok, session} =
  Codex.CLI.interactive(
    "Summarize this repository in three bullets.",
    codex_opts: codex_opts
  )

:ok = Codex.CLI.Session.close_input(session)
{:ok, session_result} = Codex.CLI.Session.collect(session)
IO.puts(session_result.stdout)

This layer is also the simplest way to reach CLI-only workflows such as codex completion, codex cloud, codex execpolicy, codex features, codex mcp-server, and the root interactive client without dropping down to System.cmd/3 yourself.

Current upstream parity helpers also include:

ws_auth atoms normalize to upstream CLI values such as :capability_token -> capability-token and :signed_bearer_token -> signed-bearer-token.

Streaming Responses

For real-time processing of events as they occur:

{:ok, thread} = Codex.start_thread()

{:ok, stream} = Codex.Thread.run_streamed(
  thread,
  "Analyze this codebase and suggest improvements"
)

# Process events as they arrive
for event <- stream do
  case event do
    %Codex.Events.ItemStarted{item: item} ->
      IO.puts("New item: #{item.type}")

    %Codex.Events.ItemCompleted{item: %{type: "agent_message", text: text}} ->
      IO.puts("Response: #{text}")

    %Codex.Events.TurnCompleted{usage: usage} ->
      IO.puts("Tokens used: #{usage.input_tokens + usage.output_tokens}")

    _ ->
      :ok
  end
end

Structured Output

Request JSON responses conforming to a specific schema:

schema = %{
  "type" => "object",
  "properties" => %{
    "summary" => %{"type" => "string"},
    "issues" => %{
      "type" => "array",
      "items" => %{
        "type" => "object",
        "properties" => %{
          "severity" => %{"type" => "string", "enum" => ["low", "medium", "high"]},
          "description" => %{"type" => "string"},
          "file" => %{"type" => "string"}
        },
        "required" => ["severity", "description"]
      }
    }
  },
  "required" => ["summary", "issues"]
}

{:ok, thread} = Codex.start_thread()

{:ok, result} = Codex.Thread.run(
  thread,
  "Analyze the code quality of this project",
  output_schema: schema
)

# Parse the JSON response
{:ok, data} = Jason.decode(result.final_response)
IO.inspect(data["issues"])

Runnable Examples

The repository ships with standalone scripts under examples/ that you can execute via mix run. Live scripts (prefixed live_) hit the real Codex CLI using your existing CLI login—no extra API key wiring needed. To run everything sequentially:

./examples/run_all.sh

Examples that start Codex turns prefer reasoning_effort: :low; the SDK will coerce that to a higher supported level when the selected model requires it.

Or run individual scripts:

# Basic blocking turn and item traversal
mix run examples/basic_usage.exs

# Streaming patterns (real-time, progressive, stateful)
mix run examples/streaming.exs progressive

# Live model defaults + compaction/usage handling (CLI login or CODEX_API_KEY)
mix run examples/live_usage_and_compaction.exs "summarize recent changes"

# Live exec controls (env injection, cancellation token, timeout)
mix run examples/live_exec_controls.exs "list files and print CODEX_DEMO_ENV"

# Structured output decoding and struct mapping
mix run examples/structured_output.exs struct

# Conversation/resume workflow helpers
mix run examples/conversation_and_resume.exs save-resume

# Concurrency + collaboration demos
mix run examples/concurrency_and_collaboration.exs parallel lib/codex/thread.ex lib/codex/exec.ex

# Local plugin authoring scaffold (writes to a disposable temp repo)
mix run examples/plugin_scaffold.exs

# Auto-run tool bridging (forwards outputs/failures to codex exec)
mix run examples/tool_bridging_auto_run.exs

# Live two-turn session using CLI login or CODEX_API_KEY
mix run examples/live_session_walkthrough.exs "your prompt here"

# Live tooling stream: shows shell + MCP events and falls back to last agent message
mix run examples/live_tooling_stream.exs "optional prompt"

# Live telemetry stream: prints thread/turn ids, source metadata, usage deltas, diffs, and compaction (low reasoning, fast prompt)
mix run examples/live_telemetry_stream.exs

# Live CLI demo (requires authenticated codex CLI or CODEX_API_KEY)
mix run examples/live_cli_demo.exs "What is the capital of France?"

# Live app-server plugin verification against a disposable local scaffold
mix run examples/live_app_server_plugins.exs

# Live Codex CLI passthrough helpers
mix run examples/live_cli_passthrough.exs completion zsh

# Live PTY-backed prompt-mode root session
mix run examples/live_cli_session.exs "Summarize this repository in three bullets."

Realtime Voice Interactions

For bidirectional voice interactions using the OpenAI Realtime API:

Codex.Realtime.Diagnostics.probe_text_turn/1 now uses a minimal schema-compatible probe and treats unknown_parameter-style schema drift as a protocol-incompatible skip reason instead of a hard failure. Codex.Realtime.Session also defers follow-up response.create calls until the active response reaches response.done, so overlapping user input and tool output no longer trigger premature create requests.

alias Codex.Realtime

# Create a realtime agent
agent = Realtime.agent(
  name: "VoiceAssistant",
  instructions: "You are a helpful voice assistant. Keep responses brief."
)

# Configure session options
config = %Codex.Realtime.Config.RunConfig{
  model_settings: %Codex.Realtime.Config.SessionModelSettings{
    voice: "alloy",
    turn_detection: %Codex.Realtime.Config.TurnDetectionConfig{
      type: :semantic_vad,
      eagerness: :medium
    }
  }
}

# Start a realtime session
{:ok, session} = Realtime.run(agent, config: config)

# Subscribe to events
Realtime.subscribe(session, self())

# Send audio and receive responses (commit on final chunk)
Realtime.send_audio(session, audio_data, commit: true)

Realtime.Session also traps linked WebSocket exits and keeps processing other session messages while tool calls are running.

Voice Pipeline (Non-Realtime)

For STT -> Workflow -> TTS processing:

alias Codex.Voice.{Pipeline, SimpleWorkflow, Config}

# Create a simple workflow
workflow = SimpleWorkflow.new(
  fn text -> ["You said: #{text}. How can I help?"] end,
  greeting: "Hello! I&#39;m ready to listen."
)

# Configure the pipeline
config = %Config{
  workflow_name: "VoiceDemo",
  tts_settings: %Config.TTSSettings{voice: :nova}
}

# Create and run the pipeline
{:ok, pipeline} = Pipeline.start_link(workflow: workflow, config: config)
{:ok, result} = Pipeline.run(pipeline, audio_input)

# Process streamed audio output
for event <- result do
  case event do
    %Codex.Voice.Events.VoiceStreamEventAudio{data: data} ->
      # Handle audio chunk
      play_audio(data)
    _ -> :ok
  end
end

See examples/realtime_*.exs and examples/voice_*.exs for comprehensive demos.

Resuming Threads

Threads are persisted under $CODEX_HOME/sessions (default ~/.codex/sessions). Resume previous conversations:

thread_id = "thread_abc123"
{:ok, thread} = Codex.resume_thread(thread_id)

{:ok, result} = Codex.Thread.run(thread, "Continue from where we left off")

Resume the most recent session (equivalent to codex exec resume --last):

{:ok, thread} = Codex.resume_thread(:last)
{:ok, result} = Codex.Thread.run(thread, "Continue from where we left off")

Session Helpers

The CLI writes session logs under $CODEX_HOME/sessions (default ~/.codex/sessions). The SDK can list them and apply or undo diffs locally:

{:ok, sessions} = Codex.Sessions.list_sessions()

{:ok, result} = Codex.Sessions.apply(diff, cwd: "/path/to/repo")
{:ok, _undo} = Codex.Sessions.undo(ghost_snapshot, cwd: "/path/to/repo")

Configuration Options

# Codex-level options
{:ok, codex_options} =
  Codex.Options.new(
    api_key: System.fetch_env!("CODEX_API_KEY"),
    codex_path_override: "/custom/path/to/codex",
    telemetry_prefix: [:codex, :sdk],
    model: "o1",
    reasoning_effort: :high,  # :none | :minimal | :low | :medium | :high | :xhigh
    model_personality: :friendly,
    review_model: Codex.Models.default_model(),
    tool_output_token_limit: 512,
    history: %{persistence: "local", max_bytes: 1_000_000},
    config: %{"model_reasoning_summary" => "concise"}  # global --config baseline
  )

# Thread-level options
{:ok, thread_options} =
  Codex.Thread.Options.new(
    metadata: %{project: "codex_sdk"},
    labels: %{environment: "dev"},
    auto_run: true,
    sandbox: :strict,
    approval_timeout_ms: 45_000,
    ephemeral: true,          # app-server thread/fork lifecycle hint
    service_name: "my_app",   # app-server routing hint
    service_tier: :flex,      # :auto | :default | :flex | :priority
    web_search_mode: :cached,  # :disabled | :cached | :live (explicit :disabled forces disable override)
    personality: :pragmatic,   # :friendly | :pragmatic | :none (works consistently on exec/app-server)
    collaboration_mode: :plan  # :plan | :pair_programming | :code | :default | :execute | :custom (app-server)
  )

{:ok, thread} = Codex.start_thread(codex_options, thread_options)

# Run-level options (validated by Codex.RunConfig.new/1)
run_options = %{
  run_config: %{
    auto_previous_response_id: true
  }
}

{:ok, result} = Codex.Thread.run(thread, "Your prompt", run_options)
IO.inspect(result.last_response_id)
# Note: last_response_id remains nil until codex exec emits response_id fields.

# Turn-level options
turn_options = %{output_schema: my_json_schema, personality: :friendly, service_tier: :priority}

{:ok, result} = Codex.Thread.run(thread, "Your prompt", turn_options)

# Exec controls: inject env, set cancellation token/timeout/idle timeout (forwarded to codex exec)
turn_options = %{
  env: %{"CODEX_DEMO_ENV" => "from-sdk"},
  cancellation_token: "demo-token-123",
  timeout_ms: 120_000,
  stream_idle_timeout_ms: 300_000
}

# The SDK also sets CODEX_INTERNAL_ORIGINATOR_OVERRIDE=codex_sdk_elixir
# unless you provide your own value in `env`.

{:ok, stream} =
  Codex.Thread.run_streamed(thread, "List three files and echo $CODEX_DEMO_ENV", turn_options)

# Opt-in retry and rate limit handling
{:ok, thread_opts} =
  Codex.Thread.Options.new(
    retry: true,
    retry_opts: [max_attempts: 3],
    rate_limit: true,
    rate_limit_opts: [max_attempts: 3]
  )

Thread Option Boundary

codex_sdk is the layer that owns real Codex thread and execution options.

That includes:

If you are coming from a higher-level runner or orchestration layer, keep the distinction clear:

In other words:

Example:

{:ok, thread_options} =
  Codex.Thread.Options.new(
    sandbox: :workspace_write,
    ask_for_approval: :never,
    reasoning_effort: :xhigh
  )

Config Overrides

Options-level, thread-level, and turn-level config overrides are forwarded as --config key=value flags to the Codex CLI (exec transport). For app-server transport, typed derived settings plus options-level config overrides are merged into the structured config payload when unset. Four layers of precedence apply for exec — later wins:

  1. Options-level global — Codex.Options.new(config: ...)
  2. Derived — automatically generated from typed Codex.Options and Codex.Thread.Options fields
  3. Thread-level — Codex.Thread.Options.config_overrides
  4. Turn-level — config_overrides in turn opts passed to Thread.run/3

Nested maps are auto-flattened to dotted-path keys:

# These two are equivalent:
config_overrides: %{"features" => %{"web_search_request" => true}}
config_overrides: [{"features.web_search_request", true}]

Override values are validated at runtime and must be TOML-compatible primitives: strings, booleans, integers/floats, arrays, and nested maps. Unsupported values (nil, tuples, PIDs, functions, etc.) return an error before the CLI is invoked.

When you explicitly disable web search (web_search_enabled: false or web_search_mode: :disabled), the SDK emits web_search="disabled" so that thread-level intent overrides existing CLI config. When you leave defaults untouched, the SDK now mirrors current Codex CLI behavior: cached web search for normal local runs, and live web search when you opt into full-access sandboxing (:danger_full_access, :permissive, or dangerously_bypass_approvals_and_sandbox: true).

Approval Hooks

Codex ships with approval policies and hooks so you can review potentially destructive actions before the agent executes them. Policies are provided per-thread:

policy = Codex.Approvals.StaticPolicy.deny(reason: "manual review required")

{:ok, thread_opts} =
  Codex.Thread.Options.new(
    sandbox: :strict,
    approval_policy: policy,
    approval_timeout_ms: 60_000
  )

{:ok, thread} = Codex.start_thread(%Codex.Options{}, thread_opts)

To integrate with external workflow tools, implement the Codex.Approvals.Hook behaviour and set it as the approval_hook:

defmodule MyApp.ApprovalHook do
  @behaviour Codex.Approvals.Hook

  def review_tool(event, context, _opts) do
    # Route to Slack/Jira/etc. and await a decision
    if MyApp.RiskEngine.requires_manual_review?(event, context) do
      {:deny, "pending review"}
    else
      :allow
    end
  end
end

{:ok, thread_opts} = Codex.Thread.Options.new(approval_hook: MyApp.ApprovalHook)
{:ok, thread} = Codex.start_thread(%Codex.Options{}, thread_opts)

Hooks can be synchronous or async (see Codex.Approvals.Hook for callback semantics), and all decisions emit telemetry so you can audit approvals externally.

Approval, Sandbox, And Auto-Run Cheatsheet

These settings are related but distinct:

Use the explicit knobs when you need precise behavior. Treat the convenience flags as shortcuts, not as the clearest onboarding path.

Codex respects upstream safe-command markers: tool events flagged with requires_approval: false bypass approval gating automatically, keeping low-risk workspace actions fast while still blocking requests that require review.

For app-server file-change approvals, hooks can return {:allow, grant_root: "/path"} to accept the proposed root for the current session.

App-server permission approvals use structured grant payloads rather than string decisions. Hooks can implement review_permissions/3 and return :allow, {:allow, permissions: ..., scope: :turn | :session}, or {:deny, reason}. App-server streams now also surface %Codex.Events.GuardianApprovalReviewStarted{}, %Codex.Events.GuardianApprovalReviewCompleted{}, and %Codex.Events.ServerRequestResolved{} when the connected Codex build emits guardian review and request-resolution notifications. Use approvals_reviewer: :user | :guardian_subagent on thread options to control upstream review routing. The SDK also emits %Codex.Events.CommandApprovalRequested{} and %Codex.Events.FileApprovalRequested{} for app-server request approvals, preserving upstream fields such as approval_id, command_actions, network_approval_context, additional_permissions, available_decisions, and grant_root in normalized snake_case. These app-server approval fields are experimental upstream, so connect with experimental_api: true before using them. For live request-permissions flows, use a granular approval policy with request_permissions: true, but note that upstream keeps request_permissions_tool, exec_permission_approvals, and guardian_approval disabled by default on stock CLI installs. The SDK accepts both the local inline shape (%{type: :granular, request_permissions: true}) and the upstream external-tagged shape (%{granular: %{request_permissions: true}}); malformed granular maps now fail fast instead of being silently dropped.

Tool-call events can also arrive pre-approved via approved_by_policy (or approved) from the CLI; the SDK mirrors that bypass and skips hooks while still emitting telemetry. Sandbox warnings are normalized so Windows paths dedupe cleanly (e.g., C:/Temp and C:\\Temp coalesce). See examples/sandbox_warnings_and_approval_bypass.exs for a runnable walkthrough.

File Attachments & Registries

Stage attachments once and reuse them across turns or threads with the built-in registry:

{:ok, attachment} = Codex.Files.stage("reports/summary.md", ttl_ms: :infinity)

thread_opts =
  %Codex.Thread.Options{}
  |> Codex.Files.attach(attachment)

{:ok, thread} = Codex.start_thread(%Codex.Options{}, thread_opts)

Query Codex.Files.metrics/0 for staging stats and force cleanup with Codex.Files.force_cleanup/0. Codex.Files.force_cleanup/0, Codex.Files.reset!/0, and Codex.Files.metrics/0 return {:error, reason} if the registry is unavailable. Use Codex.Files.list_staged_result/0 for explicit {:ok, list} | {:error, reason} responses; Codex.Files.list_staged/0 remains available as a compatibility helper that falls back to [] on startup errors. Staged files are runtime-scoped; the registry clears the staging directory on startup, so re-stage attachments after restarts.

MCP Tool Discovery

The SDK provides MCP client helpers for discovering and invoking tools from MCP servers:

# Connect to a stdio MCP server
{:ok, transport} =
  Codex.MCP.Transport.Stdio.start_link(
    command: "npx",
    args: ["-y", "mcp-server"]
  )

{:ok, client} =
  Codex.MCP.Client.initialize(
    {Codex.MCP.Transport.Stdio, transport},
    client: "codex-elixir",
    version: "0.1.0",
    server_name: "my_server"
  )

# List tools with filtering
{:ok, tools, client} = Codex.MCP.Client.list_tools(client,
  allow: ["read_file", "write_file"],
  deny: ["dangerous_tool"]
)

# List tools with qualified names (mcp__server__tool format)
{:ok, tools, client} = Codex.MCP.Client.list_tools(client, qualify?: true)

# Each tool includes:
# - "name" - original tool name
# - "qualified_name" - fully qualified name (e.g., "mcp__my_server__read_file")
# - "server_name" - server identifier

Codex.MCP.Transport.StreamableHTTP provides JSON-RPC over HTTP with bearer/OAuth auth support for remote MCP servers. Transport failures are normalized to {:error, reason} tuples.

Tool name qualification now sanitizes each server/tool component to ASCII alphanumerics plus _ and - before joining them for OpenAI-facing tool names. Original MCP server/tool names are preserved for actual MCP calls. Names exceeding 64 characters are truncated with a SHA1 hash suffix for disambiguation:

Codex.MCP.Client.qualify_tool_name("server1", "tool_a")
#=> "mcp__server1__tool_a"

Codex.MCP.Client.qualify_tool_name("server.one", "tool.two-three")
#=> "mcp__server_one__tool_two-three"

# Long names are truncated with SHA1 suffix
Codex.MCP.Client.qualify_tool_name("srv", String.duplicate("a", 80))
#=> 64-character string with SHA1 hash suffix

Results are cached by default; bypass with cache?: false. See Codex.MCP.Client for full documentation and examples/live_mcp_and_sessions.exs for a runnable demo.

Shell Hosted Tool

The SDK provides a fully-featured shell command execution tool with approval integration, timeout handling, and output truncation:

alias Codex.Tools
alias Codex.Tools.ShellTool

# Register with default settings (60s timeout, 10KB max output)
{:ok, _} = Tools.register(ShellTool)

# Execute a simple command
{:ok, result} = Tools.invoke("shell", %{"command" => ["ls", "-la"]}, %{})
# => %{"output" => "...", "exit_code" => 0, "success" => true}

# With working directory
{:ok, result} = Tools.invoke("shell", %{"command" => ["pwd"], "workdir" => "/tmp"}, %{})

# With custom timeout and output limits
{:ok, _} = Tools.register(ShellTool,
  timeout_ms: 30_000,
  max_output_bytes: 5000
)

# With approval callback for sensitive commands
approval = fn cmd, _ctx ->
  if String.contains?(cmd, "rm"), do: {:deny, "rm not allowed"}, else: :ok
end

{:ok, _} = Tools.register(ShellTool, approval: approval)
{:error, {:approval_denied, "rm not allowed"}} =
  Tools.invoke("shell", %{"command" => ["rm", "file"]}, %{})

For custom execution, provide a custom executor:

custom_executor = fn %{"command" => cmd}, _ctx, _meta ->
  formatted = if is_list(cmd), do: Enum.join(cmd, " "), else: cmd
  {:ok, %{"output" => "custom: #{formatted}", "exit_code" => 0}}
end

{:ok, _} = Tools.register(ShellTool, executor: custom_executor)

For string shell scripts, use the shell_command tool:

alias Codex.Tools.ShellCommandTool

{:ok, _} = Tools.register(ShellCommandTool)
{:ok, result} = Tools.invoke("shell_command", %{"command" => "ls -la", "workdir" => "/tmp"}, %{})

Additional hosted tools include write_stdin (unified exec sessions via app-server) and view_image (local image attachments gated by features.view_image_tool or Thread.Options.view_image_tool_enabled).

See examples/shell_tool.exs for a complete demonstration.

FileSearch Hosted Tool

The SDK provides a local filesystem search tool with glob pattern matching and content search capabilities:

alias Codex.Tools
alias Codex.Tools.FileSearchTool

# Register with default settings
{:ok, _} = Tools.register(FileSearchTool)

# Find all Elixir files recursively
{:ok, result} = Tools.invoke("file_search", %{"pattern" => "lib/**/*.ex"}, %{})
# => %{"count" => 42, "files" => [%{"path" => "lib/foo.ex"}, ...]}

# Search file content with regex
{:ok, result} = Tools.invoke("file_search", %{
  "pattern" => "**/*.ex",
  "content" => "defmodule"
}, %{})
# => %{"count" => 10, "files" => [%{"path" => "lib/foo.ex", "matches" => [...]}]}

# Case-insensitive content search
{:ok, result} = Tools.invoke("file_search", %{
  "pattern" => "**/*.ex",
  "content" => "ERROR",
  "case_sensitive" => false
}, %{})

# Limit results
{:ok, result} = Tools.invoke("file_search", %{
  "pattern" => "**/*",
  "max_results" => 20
}, %{})

# Custom base path
{:ok, _} = Tools.register(FileSearchTool, base_path: "/project")

Supported glob patterns:

See examples/file_search_tool.exs for more examples.

MCP Tool Invocation

Invoke tools on MCP servers with built-in retry logic, approval callbacks, and telemetry:

# Basic invocation with default retries (3) and exponential backoff
{:ok, result} = Codex.MCP.Client.call_tool(client, "echo", %{"text" => "hello"})

# Custom retry and timeout settings
{:ok, result} = Codex.MCP.Client.call_tool(client, "fetch", %{"url" => url},
  retries: 5,
  timeout_ms: 30_000,
  backoff: fn attempt -> Process.sleep(attempt * 200) end
)

# With approval callback (for sensitive operations)
{:ok, result} = Codex.MCP.Client.call_tool(client, "write_file", args,
  approval: fn tool, args, context ->
    if authorized?(context.user, tool), do: :ok, else: {:deny, "unauthorized"}
  end,
  context: %{user: current_user}
)

Telemetry events are emitted for observability:

Custom Prompts and Skills

List and expand custom prompts from $CODEX_HOME/prompts, and load skills when features.skills is enabled:

{:ok, prompts} = Codex.Prompts.list()
{:ok, expanded} = Codex.Prompts.expand(Enum.at(prompts, 0), "FILE=lib/app.ex")

{:ok, conn} = Codex.AppServer.connect(codex_opts)
{:ok, %{"data" => skills}} = Codex.Skills.list(conn, skills_enabled: true)
{:ok, content} = Codex.Skills.load(hd(hd(skills)["skills"]), skills_enabled: true)

Retry Logic

The SDK provides comprehensive retry utilities via Codex.Retry for handling transient failures:

alias Codex.Retry

# Basic retry with defaults (4 attempts, exponential backoff, 200ms base delay)
{:ok, result} = Retry.with_retry(fn -> make_api_call() end)

# Custom configuration
{:ok, result} = Retry.with_retry(
  fn -> risky_operation() end,
  max_attempts: 5,
  base_delay_ms: 100,
  max_delay_ms: 5_000,
  strategy: :exponential,
  jitter: true,
  on_retry: fn attempt, error ->
    Logger.warning("Retry #{attempt}: #{inspect(error)}")
  end
)

# Different backoff strategies
Retry.with_retry(fun, strategy: :linear)      # 100, 200, 300, 400ms...
Retry.with_retry(fun, strategy: :constant)    # 100, 100, 100, 100ms...
Retry.with_retry(fun, strategy: :exponential) # 100, 200, 400, 800ms... (default)

# Custom backoff function
Retry.with_retry(fun, strategy: fn attempt -> attempt * 50 end)

# Custom retry predicate
Retry.with_retry(fun, retry_if: fn
  :my_transient_error -> true
  _ -> false
end)

# Stream retry (retries entire stream creation on failure)
stream = Retry.with_stream_retry(fn -> make_streaming_request() end)
Enum.each(stream, &process_item/1)

Default retryable errors include: :timeout, :econnrefused, :econnreset, :closed, :nxdomain, 5xx HTTP errors, 429 rate limits, stream errors, and Codex.TransportError with retryable?: true. See examples/retry_example.exs for more patterns.

Telemetry & OTLP Exporting

OpenTelemetry exporting is disabled by default. To ship traces/metrics to a collector, set CODEX_OTLP_ENABLE=1 along with the endpoint (and optional headers) before starting your application:

export CODEX_OTLP_ENABLE=1
export CODEX_OTLP_ENDPOINT="https://otel.example.com:4318"
export CODEX_OTLP_HEADERS="authorization=Bearer abc123"

mix run examples/basic_usage.exs

When the flag is not set (default), the SDK runs without booting the OTLP exporter—avoiding tls_certificate_check warnings on systems without the helper installed.

The Codex CLI (codex-rs) has its own OpenTelemetry log exporter, configured separately via $CODEX_HOME/config.toml (default ~/.codex/config.toml) under [otel]. This is independent of the Elixir SDK exporter above.

[otel]
environment = "staging"
exporter = "otlp-grpc"
log_user_prompt = false

[otel.exporter."otlp-grpc"]
endpoint = "https://otel.example.com:4317"

See codex/docs/config.md for the full upstream reference. To point Codex at an isolated config directory from the SDK, pass env: %{"CODEX_HOME" => "/path/to/codex_home"} in turn options for exec transport, or Codex.AppServer.connect(codex_opts, process_env: %{"CODEX_HOME" => ...}) for a managed app-server child.

Architecture

The SDK follows a layered architecture built on OTP principles:

Process Model

┌─────────────┐
│   Client    │
└──────┬──────┘
       │
       ▼
┌─────────────────┐
│ Codex.Thread    │  (manages turn state)
└────────┬────────┘
         │
         ▼
┌──────────────────┐
│  Codex.Exec      │  (public exec API)
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│ Codex.Runtime.   │  (runtime kit over core session API)
│ Exec             │
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│ CliSubprocess    │  (shared session + raw transport core)
│ Core.Session     │
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│   codex-rs       │  (OpenAI's Codex CLI)
└──────────────────┘

Event Types

The SDK provides structured events for all Codex operations:

Thread Events

Session and Control Events

App-server protocol payload structs such as RequestUserInput.Question, CollaborationMode, and RateLimit.Snapshot are now schema-backed. The SDK normalizes known fields, preserves forward-compatible unknown wire fields in extra, and merges them back into to_map/1.

Item Events

Item Types

Examples

See the examples/ directory for comprehensive demonstrations. A quick index:

Run examples with:

mix run examples/basic_usage.exs

# Live CLI example (requires authenticated codex CLI)
mix run examples/live_cli_demo.exs "What is the capital of France?"

# Run all live examples in sequence
./examples/run_all.sh

Documentation

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

Related Projects

Made with ❤️ and Elixir

Model Selection Contract

/home/home/p/g/n/codex_sdk no longer owns active model-selection policy. The only authoritative resolver/defaulting/validation path is /home/home/p/g/n/cli_subprocess_core through CliSubprocessCore.ModelRegistry.resolve/3, CliSubprocessCore.ModelRegistry.validate/2, and CliSubprocessCore.ModelRegistry.default_model/2.

Codex.Options and the runtime execution path now consume the resolved payload returned by core and only render transport arguments from that payload. Any older references in this document to local bundled catalogs such as priv/models.json should be treated as historical packaging details, not policy authority.

Session History And Recovery

The runtime-facing Codex lane now publishes explicit session-control capabilities instead of making orchestration layers infer them from ad hoc behavior.

The intended recovery posture is to prefer exact thread resumption when a concrete provider session id is available, and only fall back to looser “latest session” continuation when the caller explicitly chooses that behavior.