pstack skills use Cursor names. In Codex, read this file once per session. Apply these mappings wherever a skill or its references use those names. <pstack> is the installed folder that contains skills/ and docs/. Use the tools and argument schemas exposed in the current session. Clients can expose different tools.
| Cursor | Codex |
|---|---|
Task |
The available spawn_agent tool. It may be in a collaboration namespace. |
AskQuestion |
Use the available structured question tool. request_user_input_async allows independent work while the answer is pending. Use request_user_input only in modes that expose it. Otherwise ask in your reply. |
| Open a todolist | Use update_plan when exposed. Otherwise keep the playbook checklist in the conversation and update it at each phase. A goal is not a todo list. |
Run another pstack skill (/name or "the name skill") |
Read <pstack>/skills/<name>/SKILL.md directly and follow it. Most pstack skills intentionally stay out of implicit discovery. When addressing the user, tell them to select the installed pstack skill in the composer or invoke it with $pstack:name, not a Cursor slash command. |
Read, Grep, Glob, Bash |
Use the session's file and shell tools. Use rg for searches. |
create-skill, deslop, control-cli, control-ui |
Use the corresponding installed skill or native tool when available. deslop, control-cli, and control-ui are bundled at <pstack>/skills/<name>/SKILL.md. create-skill maps to the installed skill-creator skill. Run CLI checks through the shell. Drive UI with the exposed browser or computer tools. Clean up code directly when there is no matching skill. Report any required live check that cannot run. |
The package generates agents/openai.yaml from each skill's disable-model-invocation flag. Explicit selection loads the skill. Read sibling skills by their packaged paths, even when they are absent from the automatic catalog. Explicit poteto-mode invocation applies to the current request. LStack also supports conversation persistence through the hooks described below. Cursor Custom Modes do not apply in Codex.
When creating or updating a skill, including automate-me's <handle>-mode, translate its invocation policy too. In that skill's agents/openai.yaml, set policy.allow_implicit_invocation to the inverse of disable-model-invocation. Preserve other existing fields. Frontmatter alone does not express Codex's documented invocation policy. For an explicit-only mode, write:
policy:
allow_implicit_invocation: false| Cursor | Codex |
|---|---|
subagent_type: generalPurpose |
Use the default general-purpose agent. Omit Cursor's subagent_type. |
poteto-agent |
Spawn a delegate whose brief includes the absolute paths to this adapter and <pstack>/skills/poteto-mode/SKILL.md, and tells it to read both before working. These plugin Markdown files do not register a Codex agent type. |
comment-sicko |
Spawn a reviewer with the scoped diff and the absolute paths to this adapter and <pstack>/agents/comment-sicko.md. |
readonly: true |
Give a read-only brief and verify the working tree afterward. A brief does not change permissions. Retain the session's sandbox and approval controls. |
readonly: false |
Use the normal delegate with the parent's permissions. |
environment: "cloud", environment: "local" |
Codex delegates share the filesystem. Give concurrent writers separate worktrees or disjoint output directories. Omit the Cursor fields. |
cloud_base_branch |
Put the exact starting branch or commit and the worktree path in the delegate's brief. |
run_in_background: true |
Omit. Spawned agents already run alongside the parent. |
resume |
Use followup_task or resume_agent when offered. Preserve the skill's restrictions on reusing a delegate. |
file_attachments |
Put absolute file paths in the brief. |
Use the returned agent identifier with the session's send, wait, and follow-up tools. wait_agent can report mailbox updates. Read the delivered messages to get the answer. A timeout is not completion. Use close_agent only when it exists. After a capacity error, wait for existing work. Do not retry with a different model. On hosts without a close tool, completion does not necessarily free a slot.
When selecting another model or effort requires a fresh context, use the spawn tool's fresh-context option (fork_turns: "none" on hosts that expose it) and supply the task, paths, and relevant constraints explicitly. Full-history forks may inherit model and effort and reject overrides.
Current Codex releases enable subagents by default. Local clients can disable them with agents.enabled = false. If no spawn tool is available, report that limitation. Independent tasks can run sequentially, but a required independent review remains unverified until another agent or reviewer performs it. Do not count the parent's own second pass as independent review.
Repository instructions govern delegation. Without authorization, perform the same steps sequentially and report that independent review did not run. Do not bypass this through another Codex CLI or external model API.
| Cursor | Codex |
|---|---|
pstack-models.mdc, ~/.cursor/rules/pstack-models.mdc |
Read $CODEX_HOME/pstack-models.md (default ~/.codex/pstack-models.md) when a skill selects models. Codex does not automatically load this file. |
An unconfigured claude-*, grok-*, or gpt-* default |
Inherit the parent. A Cursor model slug is not evidence that Codex can run it. |
auto or inherit-parent |
Omit model and effort overrides. Each panel entry represents one independent agent. |
| A configured Codex model | Pass it only if the current spawn tool and available model inventory support it. |
<model>@<effort> |
Pass the model and supported reasoning_effort separately. |
Keep role names identical to setup-pstack. For a bare model name, preserve the session effort when the host exposes and accepts it. Otherwise explain that the host chooses the effort. An optional default effort row applies to real model entries without an @effort suffix. A value of session means the parent's effort. Keep configured Codex model names. If the host rejects a model, report the rejection. Use a confirmed available model only under the workflow's fallback rule. Otherwise inherit the parent and disclose the substitution.
Repeated models provide independent reviewers, but not model diversity. GPT models belong to the same family for Arena's cross-judge rule. When no different-family model is available, follow Arena's documented skip path and report the missing cross-judge. Another provider's CLI needs separate access and authorization.
The initial model sheet preserves LifeMash's current choices. When installing for LifeMash and the user has not created a sheet, copy it to $CODEX_HOME/pstack-models.md. Preserve an existing sheet. Read it when selecting delegates; this file does not change the main chat's running model or effort.
Astra is only for one specific judgment explicitly requested by the user. It is never an automatic fallback, panel member, or cross-judge, even if a sheet lists it. An authorized consultation defaults to one call at low, receives only necessary evidence, and does not delegate or retry automatically. Sol performs follow-up implementation. If the model is unavailable, inherit the parent and report the actual fallback; do not probe expensive alternatives.
Follow the skill's steps with these substitutions:
- Detect: inventory the models and effort levels the current spawn tool accepts. If no reliable inventory is exposed, offer
inherit-parentandauto. Ask for model choices only when needed. Do not persist guessed slugs. - Load: read the Codex sheet above, not a Cursor rule. Use its existing role values and budget as the starting point. Preserve its
default effortsetting when present. - Defaults: map each unconfigured role to
inherit-parent. Keep the number of entries in default panels. Explain that inherited entries follow the session's model and effort, so changing the budget does not override those entries. - Budget and confirm: show every role, model, effort, and panel size before writing. For a real model, store
<model>@<effort>when an effort is chosen. Offer only supported levels and keep an inheritance choice for effort. - Write: write the confirmed rows and budget to the Codex sheet, with no Cursor frontmatter. Change only that file. Do not put model preferences in
AGENTS.md,CLAUDE.md, orconfig.toml. - Confirm: the next model-selecting pstack skill reads the sheet. Continue the optional verification-skill offers using the Codex paths below.
| Cursor | Codex |
|---|---|
.cursor/skills/, ~/.cursor/skills/ |
.agents/skills/, ~/.agents/skills/ for project and personal skills. |
~/.cursor/plugins/, .cursor-plugin/plugin.json |
Use the installed <pstack> path supplied by the skill catalog. The generated package has plugin.json and a .codex-plugin/plugin.json compatibility manifest. |
.cursor/rules/, always-applied rules |
Project or personal AGENTS.md. The separate model sheet above is the exception. |
https://github.com/cursor/plugins/blob/main/pstack/<path> |
Read <pstack>/<path> when packaged. Give the public URL to the user when they need a link. |
~/.cursor/projects/<slug>/agent-transcripts/ |
Use a history tool scoped to the requested chat when available. For local rollouts, use the helper below to select this workspace's sessions before reading their contents. |
For compressed history, use Node 24.22 or a later 24.x release, or Node 26.10 or newer. Check node --version first. If the default runtime is older, run these helpers with npm exec --yes --package=node@24.22.1 -- node <script> <arguments> instead; this uses a cached runtime without changing the system Node installation. Older decoders can silently accept truncated Zstd data. The helpers reject compressed files on those runtimes.
For recall, reflect, automate-me, and session pickup, list matching local sessions with:
node <pstack>/docs/harness/codex/sessions.mjs --cwd <absolute-workspace-path>The helper reads metadata from .jsonl and compressed .jsonl.zst files under sessions/ and archived_sessions/. Both directories are under $CODEX_HOME, which defaults to ~/.codex. It lists matching top-level chats newest first. --cwd requires an exact workspace match. It does not include chats started in subdirectories or other workspaces.
The helper reads the metadata line of every rollout under both directories, including rollouts from other workspaces. One empty or malformed rollout fails the whole listing. The error names the file. With the user's approval, move or repair that file. Otherwise pass --home <codex-home> to read a copy of the history that excludes it. Then run the command again. The helper does not skip the file. A skipped rollout can hide history.
Use --include-subagents when the task needs delegate history. The result labels those sessions with isSubagent and parentId. Use --id <session-id> to select a known chat, including a subagent. To read the selected rollout as plain JSONL, run:
node <pstack>/docs/harness/codex/sessions.mjs --cwd <absolute-workspace-path> --id <session-id> --readThis command also decodes compressed rollouts. It reads message contents only after selection. If a conversation is ephemeral or unavailable locally, use the current conversation or the session digest that the skill permits. Do not substitute another workspace's history.
The bundled worktree-audit.sh reads Cursor history only. For a Codex worktree cleanup, use the Codex wrapper instead:
node <pstack>/docs/harness/codex/worktree-audit.mjs [-b <base-ref>] <repo-path>The wrapper retains the shared audit's worktree and PR checks. It counts sessions started in each worktree or a subdirectory. For recent sessions linked to the repository, including delegates, it also checks stored messages for the worktree path. This catches work done outside a session's starting directory.
Both audits treat activity less than five elapsed days old as recent and display dates in local time. When the wrapper finds recent activity, it keeps that worktree out of the safe bucket. Ephemeral sessions and work absent from stored rollouts remain unverified. Follow the cleanup playbook's inspection and approval steps. An audit error leaves cleanup unverified. The audit does not delete worktrees.
Cursor Automations, routine webhooks, and Cursor cloud-agent dashboards have no direct Codex equivalent. Mark the affected step Skipped in Codex: <step>. <reason>. and continue independent steps. A step that depends on the missing result remains blocked. Use Codex automations only when the user requests one and the host exposes a suitable tool.
Bugbot findings on a GitHub PR are normal review threads. Read them through the forge CLI. If no Bugbot check exists, omit steps that start or await it and report that omission. Retain the workflow's own review and verification requirements. For a built-in PR tool, use the forge command the workflow gives, subject to the user's authorization.
Codex keeps its own permission and approval rules. A missing driver, unsupported feature, or declined operation is not permission to weaken a required check or claim it passed.
The mappings follow the current Codex subagent documentation, skill documentation, and plugin packaging documentation. The active tool schema controls when a client's capabilities differ.
Trusted local Codex hooks recognize poteto 모드 켜줘, poteto 모드 꺼줘, and poteto 모드 상태 on the first line of a prompt. English equivalents are poteto mode on, poteto mode off, and poteto mode status. A task may follow on later lines. A control-only message acknowledges the state without starting a playbook.
The hook keeps an enabled marker per conversation under PLUGIN_DATA and provides a short skill reference each turn and after compaction. A resumed conversation keeps its state; a new conversation, fork, or clear starts off. Ordinary $pstack:poteto-mode <task> remains a one-turn invocation. Hook context reporting ON is evidence of persistence; without it, do not claim the mode is saved. The mode does not change model settings or permissions.
Plugin hooks require native trust through /hooks. Local plugin hooks do not run with cloud orchestration. When unavailable, use explicit per-request invocation and report that persistence is unavailable.