Codex

1schema_version = 42id = "codex"3label = "Codex"4description = "OpenAI's Codex CLI, with chat over ACP"5license = "Apache-2.0"6contributor = { name = "Tori", github = "gettori" }7icon = "codex"89# Measured against this pair with10# `node dev/acp-probe.mjs --agent codex-acp --per-model`, most recently11# 2026-09-20. Two versions because12# two binaries are involved: the PTY tab runs the `codex` a user installed, and13# the chat surface runs the ACP wrapper below, which drives that same `codex`14# underneath. `verified_against` is compared against `codex --version`, and15# health.rs reads the first semver-looking token, so the `codex-cli` version has16# to come first.17verified_against = "codex-cli 0.155.1 with @agentclientprotocol/codex-acp 1.12.0"18verified_on = "2026-09-20"1920[launch]21program = "codex"22# The TUI, for a PTY tab. Unrelated to the chat surface below, which launches23# the ACP wrapper instead.24base_args = []25# Codex's own "skip all confirmation prompts and execute without sandboxing".26# Its help calls it EXTREMELY DANGEROUS, the same standing as claude's27# `--dangerously-skip-permissions` and opencode's `--auto`.28yolo_args = ["--dangerously-bypass-approvals-and-sandbox"]29# Empty, which the UI reads as "this agent's sessions can't be resumed" (see30# ADAPTERS.md) rather than mis-resuming. Codex does have its own resume, but the31# id Tori holds for an ACP row is Tori's own: see opencode.toml for the full32# reasoning, which applies here unchanged.33resume_args = []3435# No [discovery], [parser] or [running]: a protocol-backed adapter reaches its36# sessions through ACP, never through files. See opencode.toml.3738[capabilities]39pty_quiet_ms = 200040# Not claimed. The quiet-PTY join is Claude-shaped and unmeasured against41# Codex's TUI, and a chat tab does not need it: over ACP the agent asks42# in-protocol and Tori renders the question.43needs_you = false44hooks = false4546# --- schema v3: accounts ---47#48# Measured 2026-08-14 against the `codex-cli` named in `verified_against`.49#50# `codex login status` puts the entire answer in its exit code: 0 signed in, 151# not, with the prose ("Logged in using ChatGPT" / "Not logged in") on **stderr**52# and stdout empty. So `whoami_kind = "exit_code"`, and Tori learns no account53# name from it - Codex does not print one, and inventing a parse of that prose54# would be a guess about a sentence nobody promised to keep.55#56# **No `home_env`, and `supports_isolation` is absent (so false).** `CODEX_HOME`57# does relocate the credential store: pointed at an empty directory, the probe58# reports not-logged-in while the real home stays signed in. What has *not* been59# measured is the thing isolation actually claims - two accounts holding60# simultaneous logins without one signing the other out - because that needs a61# second Codex account. Until somebody runs it, Codex offers no "add account"62# action, which is the honest rendering of "we do not know".63[accounts]64login_args = ["login"]65logout_args = ["logout"]66whoami_args = ["login", "status"]67whoami_kind = "exit_code"6869# --- schema v4: usage ---70#71# **No `sessions` rung, measured rather than assumed.** The `app-server`72# underneath does send `account/rateLimits/updated` around a turn, but73# `codex-acp` 1.2.0 files it into its own state and emits no session update74# (`acp_transport::tests::codex_forwards_no_rate_limits_over_acp`, 2026-09-06);75# the only way back out is `/status`, as markdown prose. So Codex's quota costs76# a read Tori schedules for itself: `usage_probe.rs` drives77# `codex app-server` for `account/rateLimits/read`, which answers the two78# windows, the plan and the credit balance in one exchange.79[usage]80sources = ["cli"]8182[chat]83transport = "acp"84# **Codex is the one bundled adapter whose chat binary is not its launch85# binary.** `codex app-server` exists and other clients drive it, but `codex` has86# no `acp` subcommand, so the first-party wrapper is how Codex reaches an ACP87# client. Measured 2026-08-14 against both on one live turn: the wrapper carries88# the permission prompt, the exact before-and-after diff, the same four models89# and the same token-only usage that the native app-server does, for one TOML90# file instead of a second typed transport over `codex app-server generate-ts`.91# That surface is also moving: it went from 598 files / 6,230 lines on 0.144.5 to92# 642 / 6,923 on 0.147.0, three patch versions apart and inside one afternoon.93# What the wrapper gives up is `thread/archive`, `thread/rollback`,94# `thread/fork` and `turn/steer`, none of which Tori drives over ACP anyway.95program = "npx"96# Pinned to the version measured, matching how the ACP catalog pins its own97# rows. `npx -y @agentclientprotocol/codex-acp` unpinned would silently move the98# agent underneath a `verified_against` that names 1.12.0.99base_args = ["-y", "@agentclientprotocol/codex-acp@1.12.0"]100101# All empty because ACP has no command line for any of it, not because Codex102# lacks the feature. Measured on the handshake instead:103#104#   * ids are minted by the agent inside `session/new`,105#   * resume is `session/load`; `sessionCapabilities` advertises106#     `resume, list, close, delete, additionalDirectories`,107#   * a model switch is `session/set_config_option`,108#   * modes arrive as config options: `mode` (read-only, agent,109#     agent-full-access) and `collaboration_mode` (default, plan),110#   * reasoning effort is a `thought_level` select with six levels111#     (low, medium, high, xhigh, max, ultra).112#113# Re-measured 2026-08-17 with the catalogue probe (`catalog_probe.rs`, live test114# `the_real_codex_answers_a_catalogue_from_one_session`), which reads the same115# handshake without submitting a turn. Four config options, and two findings116# worth keeping:117#118#   * the effort selector's **id is `reasoning_effort`** while its **category is119#     `thought_level`**. Tori selects by category, which is the only reason it120#     finds this at all: opencode's ids happen to match their categories, and121#     matching on the id here would have found nothing.122#   * `collaboration_mode` is categorized as *itself*, so no bespoke control123#     claims it. It is the first live option to reach the chat's generic mirror,124#     and before that mirror existed it was read off the wire and dropped.125#126# `session/list` is advertised and answered, **scoped to the asking session's127# cwd**: a chat in a project directory is told about that directory's sessions128# and nothing else. So a probe session in Tori's probe directory is out of a129# user's history twice over, the agent's scoping and `adopt`'s own filter.130session_id_args = []131resume_args = []132fork_args = []133model_args = []134effort_args = []135mode_args = []136add_dir_args = []137138# No [[chat.models]].139#140# Measured 2026-09-20: the handshake advertises five (`gpt-5.6-sol`,141# `gpt-6-astra`, `gpt-5.6-terra`, `gpt-5.6-luna`, `gpt-5.5`). There is no142# shortfall to publish,143# and the list is per account: signed *out*, the same call returned seven, which144# is a reminder that a catalogue measured without an account is not this user's.145# A bundled table would be a guess about somebody else's entitlement, and the146# picker takes the live list whenever there is one.147148# No [[chat.modes]] or [[chat.effort]] either: all three selectors above arrive149# on the handshake, which is what the ACP transport reads them from.150151[chat.acp]152# Measured doing its own file I/O, so the client's fs stays declined.153serve_client_fs = false154# Measured with dev/mcp-probe.mjs --acp on 1.12.0: it starts a server handed in155# `mcpServers`, passes the `env` pairs through, and the model calls the tool156# with no `session/request_permission`.157send_mcp_servers = true158159[install]160# npm names the package that ships the launch binary, and all three verbs were161# run on macOS. The binary lands in npm's global bin, which has to be on the162# login shell's PATH for Tori to find it.163program = "npm"164args = ["install", "-g", "@openai/codex"]165update_args = ["install", "-g", "@openai/codex"]166uninstall_args = ["uninstall", "-g", "@openai/codex"]