A
AGENTS.md
A markdown file at the repository or directory level that provides persistent custom instructions to Codex, scoped by proximity to the working directory with closer files taking precedence.
Approval Policy
A configuration setting (e.g., "on-request" or "never") that determines when Codex must pause and ask the user for permission before executing commands that cross the sandbox boundary.
C
Cloud Task
An asynchronous Codex job that runs hosted against a pushed GitHub repository rather than a local checkout, suitable for work that does not depend on untracked local artifacts.
Codex Exec
The non-interactive CLI command (codex exec) that runs a prompt as a scripted or CI-integrated job, defaulting to a read-only sandbox and supporting structured JSON output.
Complementary Workflow
An operating model where multiple AI agents fill distinct roles (e.g., one implements while another reviews) rather than competing on the same task, creating separation of duties.
Computer Use
A Codex capability that allows the agent to operate native desktop applications through visual interaction, used when browser-based or structured tool access is insufficient.
G
Goal Mode
A Codex execution mode that keeps a measurable objective active across a long multi-step session, serving as both the starting instruction and the completion test.
Guidance Layer
One of several scoped levels (prompt, AGENTS.md, repository config, global config, skill) where Codex instructions can live, with narrower scope taking precedence over broader scope.
H
Handoff
The process of moving a Codex thread and its associated changes from an isolated worktree back to the local checkout without manually copying files or creating branch conflicts.
I
Instruction Debt
The accumulation of vague or unverifiable rules (e.g., "write clean code") in agent guidance files that consume context without producing measurable behavior change.
M
MCP (Model Context Protocol)
An integration mechanism that connects Codex to external data sources or action systems outside the repository, enabling live queries and operations through standardized tool interfaces.
P
Plan Mode
A Codex execution mode used to interview the user, expose assumptions, and shape an ambiguous task into a clear objective before any code edits begin.
Plugin
An installable bundle that packages one or more Codex skills, tools, or configuration into a distributable unit that can be shared across users or repositories.
Promotion Ladder
The progressive escalation path from manual interactive execution to scripted automation, where a workflow must succeed repeatedly before being scheduled for unattended runs.
S
Sandbox Mode
A security configuration (read-only, workspace-write, or danger-full-access) that defines the technical boundary of what filesystem and network resources Codex can access during execution.
Separation of Duties
The practice of assigning implementation and review to different agents so that the reviewer does not inherit the implementer's assumptions or context bias.
Side Chat
A conversation mechanism that allows asking an agent for explanation or clarification mid-task without redirecting or interrupting the active work thread.
Skill
A reusable named procedure with metadata, instructions, and optional helper scripts that Codex loads on demand when its description matches the current task, using progressive disclosure.
Subagent
A parallel worker spawned by Codex to perform a bounded, typically read-only investigation or specialist review, returning distilled findings to the main thread for synthesis.
Surface Map
A mental model for choosing which Codex interface (CLI, IDE, app, web/cloud, or exec) best fits a given task based on requirements like interactivity, isolation, or GitHub integration.
T
Thread Automation
A scheduled Codex automation that continues an existing conversation thread on a recurring basis, maintaining accumulated context across runs.
V
Visual Feedback Loop
A verification pattern where the agent reproduces a UI state, captures multi-signal evidence (screenshot, DOM, console, network), makes a scoped change, and repeats the identical flow to confirm the fix.
W
Worktree
A Git-based isolated filesystem managed by the Codex app that gives a background task its own checkout separate from the user's local working directory, preventing concurrent-edit conflicts.
Writable Root
A narrowly scoped filesystem exception added to Codex's sandbox configuration that grants write access to one specific directory without broadening permissions globally.
No terms match your search.