Agent Skills: Codex Knowledge Patch

Codex Knowledge Patch

Use this skill for current Codex app, CLI, TUI, remote, cloud, plugin, MCP, sandbox, configuration, and collaboration behavior.

Reference index

| Reference | Topics | | --- | --- | | App and collaboration | Desktop availability, shared state, worktrees, skills, Automations, imports, multi-folder projects, browser, images, voice, and iOS | | CLI execution and cloud | Resume, working roots, image generation, scripted execution, cloud tasks, launcher, login, package detection, and sandbox runner | | Remote, MCP, and connectors | Remote TUI, app-server authentication, MCP setup, hosted connectors, WebSockets, plugins, marketplaces, and V2 testing | | Security and configuration | Approval modes, permission presets, web search, feature flags, sandbox roots, policy preflight, and diagnostics | | TUI and session controls | Personality, themes, review, steering, history, Fast and Plan modes, conversation controls, status lines, processes, and agents |

Breaking changes and deprecations

Retiring ChatGPT-authenticated models

On August 31, 2026, ChatGPT-authenticated sessions lose gpt-5.4 and gpt-5.4-mini. They remain available through the API and in API-key- authenticated sessions.

Migrate workspace defaults, saved model settings, managed configuration, custom agents, and scheduled tasks from gpt-5.4 to gpt-5.6-terra, and from gpt-5.4-mini to gpt-5.6-luna.

Deprecated approval policy

on-failure is deprecated. Approval choices are untrusted, on-request, and never.

--full-auto combines workspace-write with on-request; --yolo bypasses both safeguards.

Global flag placement

Place global flags inherited by subcommands after the subcommand name. Repeatable -c key=value overrides parse JSON values when possible.

codex exec --profile work -c 'features.unified_exec=true' --full-auto "Run the checks"

High-value CLI workflows

Resume the right session

Interactive resume defaults to sessions from the current working directory. Use --all to broaden the picker; --last or a session ID skips it.

codex resume --last
codex exec resume --last "Continue with the implementation"

Non-interactive resume retains the transcript, plan history, and approvals.

Produce scripted output

Pass - to codex exec to read a prompt from standard input. A scripted run can emit JSONL state-change events, constrain the final response with a JSON Schema, save only the final message, avoid rollout persistence with --ephemeral, or operate outside Git with --skip-git-repo-check.

printf '%s\n' "Inventory licenses" | codex exec - --json \
  --output-schema result.schema.json --output-last-message final.txt

Select writable project roots

--cd selects the working root without changing the shell directory. Repeat --add-dir to expose additional writable roots.

codex --cd apps/frontend --add-dir ../backend --add-dir ../shared

Submit and apply cloud work

codex cloud opens a task picker. codex cloud exec submits directly to a configured environment and accepts one to four independent attempts. Submission failures exit nonzero, and Ctrl+O selects an environment in the picker.

codex cloud exec --env ENV_ID --attempts 3 "Summarize open bugs"
codex cloud list --env ENV_ID --limit 20 --json
codex apply TASK_ID

Cloud listing supports environment filtering, cursor pagination, a 1–20 result limit, and JSON with a tasks array plus an optional cursor. Applying a task uses its latest diff, reports patched files, and exits nonzero on git apply conflicts.

Generate or edit images

Request image generation or editing in natural language or with the literal $imagegen. Attach an image when the task should transform or extend it.

codex --image existing.png '$imagegen Extend this asset'

Image-generation requests use the image-generation extension by default.

Permissions and configuration

Choose a permission preset

/permissions switches among Auto, Read-only, and Full Access.

• Auto permits reads, edits, and commands inside the working directory, but asks for outside or network access.
• Read-only withholds changes and commands pending plan approval.
• Full Access removes those prompts across the machine and network.

Control web search

Local CLI work uses cached search by default; full-access sandboxing defaults to live results. --search requests live results for one run.

Persist the choice with web_search = "live" or web_search = "disabled". JSON execution reports searches as web_search events.

Manage feature flags

Inspect and persist feature flags with codex features. Changes go to ~/.codex/config.toml, or to the selected profile when --profile is used.

codex features list
codex features enable unified_exec
codex features disable shell_snapshot

Preflight an execution policy

The preview codex execpolicy check evaluates a command against one or more rule files and emits JSON with the strictest decision and matching rules. Use --pretty to format it.

codex execpolicy check --rules ~/.codex/rules/default.rules --pretty -- git status

Interaction and collaboration

Steer or queue work

During an active turn, Enter injects instructions immediately. Tab queues a prompt, slash command, or ! shell command for the next turn. @ opens fuzzy workspace-file search, and !command runs locally under the active sandbox and approval policy.

Review without changing the tree

/review launches a dedicated reviewer without modifying the working tree. Choose a base-branch diff, all uncommitted changes, one commit, or custom instructions. Review uses the session model unless review_model overrides it in configuration.

Use explicit subagents

Subagents are spawned only when explicitly requested. Configure roles under [agents] in config.toml. Each subagent performs its own model and tool work, so parallel workflows consume additional tokens.

Use /agent to inspect or continue a spawned agent's thread.

Work across desktop projects

The desktop app reuses projects, session history, and configuration from the CLI and IDE extension. Built-in Git worktrees isolate parallel agent threads; review or comment on changes in-thread, open them in an editor, or check them out locally.

For multi-folder projects, the primary folder controls new chats, Git operations, and automatic discovery of AGENTS.md, skills, and config.toml. Secondary folders remain searchable, readable, and editable, and Review aggregates repositories and changed lines across the project.

Routing details

Read the app reference for desktop or mobile availability, Automations, portable or recorded skills, Computer History, voice, imports, browser context, multi-folder projects, or image review.

Read the CLI execution reference for resume behavior, structured output, cloud workflows, working roots, authentication, app launching, global package detection, image requests, or direct sandbox execution.

Read the remote reference for app-server hosting, remote authentication, MCP, hosted connectors, Responses WebSockets, plugins, marketplace sources, or the experimental V2 test client.

Read the security reference for approval semantics, sandbox scope, web-search modes, feature persistence, Windows roots, execution-policy checks, or configuration-layer diagnostics.

Read the TUI reference for interaction commands, session navigation, review, themes, status lines, terminal titles, background commands, personalities, Fast mode, Plan mode, copying, branching, or explicit agent threads.