| name | codex-bootstrap |
| description | Bootstrap a local Codex setup from this shared starter repo. Use when an engineer needs an interactive onboarding flow that asks for local paths, required MCP settings, and only the overrides that differ from the shared Enterprise ChatGPT defaults, then generates or repairs a real local Codex config and verifies that shared path placeholders were resolved. |
Codex Bootstrap
Use this skill when an engineer is setting up Codex from this shared starter repo or repairing a local config after the repo moved.
Setup model:
- Keep the shared repo as the versioned source of truth for
AGENTS.md, workflows/, agents/, skills/, and templates.
- Keep personal Codex runtime state in
~/.codex by default unless the engineer explicitly wants another Codex home.
- Generate a local user config that points at the shared repo’s real paths.
- Install a rendered global
AGENTS.md entrypoint in Codex home so the engineer’s Codex setup is complete even when Codex home is outside the shared repo.
Quick Start
- Prefer writing a local config outside the shared repo, usually
~/.codex/config.toml.
- Default Codex home to
~/.codex unless the engineer wants a different location.
- Ask only the necessary questions, one short question at a time:
- repo root for this shared starter
- whether they want a non-default Codex home
- output config path
- workspace parent
- whether they want to add the local Maven repository to writable roots for sandboxed builds, defaulting the path to
~/.m2/repository
- whether they want to override the shared Enterprise ChatGPT model defaults, and only then the optional provider and model
- whether to install the rendered global
AGENTS.md entrypoint in Codex home
- whether they want to opt out of the default-enabled
stlm-mcp, mcp-dope, chrome-devtools, sqlcl, mcp-atlassian-jira-sd, mcp-atlassian-jira-oci, centralconfluence, playwright, ots, mcp_shepherd, and lts-mcp integrations
- tell them the Devplat MCP Gateway plugin is configured by default through the
devplat-plugins marketplace; Bo Peep tools load with devplat_mcp_gateway__use target=mfo-bo-peep after local gateway setup
- for enabled
sqlcl, ask whether bootstrap should use SQLcl connection metadata from an On-Call Investigation team config and, if yes, collect the team-config path and optional team name
- whether they already have one shared
.env file for enabled helper-backed MCP servers; if not, collect the desired path, offer starter generation there, and collect the shared HOME override and log directory when they want to change the defaults
- whether they want to opt out of the default-enabled
codex-bootstrap, ots-ticket, jira-ticket, cm-review, create-module-knowledge-skills, internal-confluence-page, oncall-investigation, object-store, pr-description, repository-version-preflight, scm-pr, bitbucket-pr, release-check, and authZ-permissions-yaml-generator skills
- when
bitbucket-pr stays enabled, ask whether the engineer uses zsh or bash, then collect the shell init file path to use for sourcing that helper
- for enabled
bitbucket-pr, whether BASE_URL and BITBUCKET_TOKEN are already set up; if not, collect the Bitbucket host and create a local helper env file
- for enabled
jira-ticket, whether JIRA_URL and JIRA_PERSONAL_TOKEN are already set up in both Jira MCP env files (~/.env and ~/.env.jira-oci); if not, collect the Jira OCI host and offer to create the Jira OCI MCP env helper
- tell the engineer to source the generated Bitbucket helper from the selected shell init file so every shell session loads it
- if either enabled skill still lacks configured auth, stop bootstrap with a clear rerun instruction instead of silently completing
- Run the helper script in this skill to generate or update the config and install the global
AGENTS.md entrypoint.
- Verify the script linked enabled shared skills into
$CODEX_HOME/skills and reports no unresolved path placeholders.
- Summarize what was written, which skills and integrations were enabled, how
AGENTS.md and skill links were installed, any PR auth helper files that were created, and any remaining follow-up.
Undo
Use undo mode when the engineer wants to remove the local files that bootstrap generated without restoring older backups.
python3 skills/codex-bootstrap/scripts/bootstrap_codex.py \
--codex-home "~/.codex" \
--undo
Notes:
- Undo removes the generated local
config.toml, global AGENTS.md target, generated MCP env helper files, the default Bitbucket auth helper, and the generated Jira OCI MCP env helper when it still has the bootstrap header.
- Undo does not restore backups.
- Undo leaves any
.bak.* files in place so the engineer can inspect or restore them manually if needed.
Safety Rules
- Do not write personal absolute paths into tracked repo files unless the user explicitly asks.
- Prefer generating a user-local config from
config.example.toml instead of editing the template in place.
- Prefer installing a rendered
~/.codex/AGENTS.md copy that expands repo-relative routing paths to absolute repo paths so skill and action routing still works outside the shared repo workspace.
- Install enabled shared skills as symlinks under
$CODEX_HOME/skills that point back to the shared repo. Do not copy skill directories; copies drift when the repo skills are updated.
- Do not tell AIPack users to run bootstrap as a mandatory post-install step.
- If the target config already exists, back it up before overwriting.
- If a skill target already exists in
$CODEX_HOME/skills, back it up under $CODEX_HOME/skill-backups before replacing it with the repo symlink.
- In undo mode, remove only the generated local files and repo-backed skill symlinks; do not auto-restore backups.
- Default
stlm-mcp, mcp-dope, chrome-devtools, sqlcl, mcp-atlassian-jira-sd, mcp-atlassian-jira-oci, centralconfluence, playwright, ots, mcp_shepherd, lts-mcp, codex-bootstrap, ots-ticket, jira-ticket, cm-review, create-module-knowledge-skills, internal-confluence-page, oncall-investigation, object-store, pr-description, repository-version-preflight, scm-pr, bitbucket-pr, release-check, and authZ-permissions-yaml-generator to enabled unless the engineer explicitly opts out.
- When
stlm-mcp or mcp-dope stays enabled, collect one shared .env path for both MCPs and offer the shared HOME override and log directory when the engineer wants to change the defaults.
- Ask whether the engineer already has a shared env file for enabled helper-backed MCPs before offering generation.
- For
stlm-mcp, make sure the generated helper or existing dotenv contains OCI_CONFIG_FILE and OCI_PROFILE.
- If they do not, offer to generate a starter dotenv helper from the repo template at the chosen path.
- If
bitbucket-pr stays enabled, ask whether BASE_URL and BITBUCKET_TOKEN are already available and create a local helper env file when they are not.
- If
jira-ticket stays enabled, ask whether JIRA_URL and JIRA_PERSONAL_TOKEN are already available in both Jira MCP env files (~/.env and ~/.env.jira-oci) and offer to create the Jira OCI MCP env helper when Jira OCI is not configured.
- When a Bitbucket helper file is used, tell the engineer to source it from the selected shell init file.
- Jira Ticket reads the shared Jira MCP dotenv files directly and does not need shell startup sourcing.
- If
bitbucket-pr or jira-ticket stays enabled, require configured auth before bootstrap succeeds. A starter helper file alone is not enough; the engineer must populate it or otherwise export the required variables, then rerun bootstrap.
- Prefer the shared
skills/codex-bootstrap/scripts/refresh_auth.py helper for OP_TOKEN refreshes and OCI session validation/repair so shared docs and skills stay aligned.
- Do not run Devplat MCP Gateway credential refresh or configuration mutation from bootstrap. For Bo Peep, tell the engineer to run
mcpgw config BoPeep, mcpgw refresh-jwt, and start Docker/Colima locally when the gateway doctor reports those prerequisites.
Workflow
1. Collect the minimum required inputs
Ask one short question at a time until you have:
- repo root
- whether Codex home should stay at the default
~/.codex
- output config path
- workspace parent
- whether the engineer wants the local Maven repository writable root for sandboxed builds
- whether the engineer wants to override the shared Enterprise ChatGPT defaults
- provider and model only when they opt into an override
- whether to install the rendered global
AGENTS.md entrypoint
- enabled skills, defaulting to all shared skills on
- whether
bitbucket-pr auth is already set up and, if not, the Bitbucket host plus helper env path
- whether
jira-ticket auth is already set up and, if not, the Jira OCI host plus whether to generate the Jira OCI MCP env helper
- whether the engineer uses
zsh or bash for shell startup and which init file should source the Bitbucket auth helper when Bitbucket stays enabled
- enabled MCP servers, defaulting to the full packaged set of
stlm-mcp, mcp-dope, chrome-devtools, sqlcl, mcp-atlassian-jira-sd, mcp-atlassian-jira-oci, centralconfluence, playwright, ots, mcp_shepherd, and lts-mcp
- Devplat MCP Gateway plugin wiring, which is configured through the plugin marketplace rather than the packaged MCP server list; Bo Peep uses
devplat_mcp_gateway__use target=mfo-bo-peep
- whether the engineer already has one shared
.env file for enabled helper-backed MCPs; if not, the desired path, whether to generate a starter env helper there, plus the shared HOME override and log directory when they want to change the defaults
- when SQLcl connection metadata should come from an On-Call Investigation team config, the config path and, if needed, the specific team name whose
team.sqlcl block should be used
Good defaults are:
- repo root: current repo root
- Codex home:
~/.codex
- global
AGENTS.md target: <codex home>/AGENTS.md
- workspace parent: parent directory of the current repo
- Maven repository:
~/.m2/repository when sandboxed builds need it
- log directory:
<codex home>/log
- Bitbucket helper env file:
<codex home>/bitbucket-pr.env
- Jira-SD MCP/Jira Ticket env file:
<home>/.env
- Jira OCI MCP/Jira Ticket env file:
<home>/.env.jira-oci
- zsh init file for Bitbucket auth helper:
~/.zshenv
- bash init file for Bitbucket auth helper:
~/.bashrc
- Jira-SD base URL:
https://jira-sd.mc1.oracleiaas.com
- Jira OCI base URL:
https://jira.oci.oraclecorp.com
- MCP helper env file for
stlm-mcp and mcp-dope: <codex home>/mcp.env
- default
stlm-mcp MCP env OCI settings: OCI_CONFIG_FILE=<home>/.oci/config, OCI_PROFILE=oc1, and OCI session auto-refresh settings
- shared
HOME for packaged MCPs: current user home directory
- Atlassian MCP dotenv files:
<home>/.env for Jira-SD and <home>/.env.jira-oci for Jira OCI
- README-derived starter keys for the shared
mcp-dope template helper used by both stlm-mcp and mcp-dope: OP_TOKEN and SSH_AUTH_SOCK
- model provider: blank by default
- provider default model:
gpt-5.5
2. Generate the config
Use the bundled script. Prefer explicit arguments when you already collected the values in chat.
If helper-backed MCPs stay enabled and you are not using --prompt, provide one shared --env-file plus --home-dir and --log-directory explicitly. stlm-mcp and mcp-dope must point at the same dotenv file. stlm-mcp also needs HOME, OCI_CONFIG_FILE, OCI_PROFILE, OCI session auto-refresh settings, and LOG_DIRECTORY in the MCP env block.
All packaged MCPs are enabled by default in non-prompt mode. Disable any the engineer does not use with --disable-mcp <name>.
When you want bootstrap to create starter MCP env helpers in non-prompt mode, pass --generate-mcp-env <server-name> once per enabled MCP whose file should be generated. If multiple enabled MCPs share the same env path, bootstrap writes one file from mcp-dope.env.template; stlm-mcp gets its OCI config/profile settings from the MCP env block.
Pass --m2-repository "<path>" when you want sandboxed builds to reuse a writable local Maven cache.
When bitbucket-pr still needs local auth/bootstrap help, pass the matching flags so the script can generate the helper file:
--bitbucket-auth-state needs-setup --bitbucket-base-url "<host>" --bitbucket-env-file "<path>"
When jira-ticket still needs local auth/bootstrap help, pass the matching flags so the script can generate the helper file:
--jira-auth-state needs-setup --jira-oci-base-url "<jira-oci-host>" --generate-jira-oci-mcp-env
When the engineer wants bootstrap to use SQLcl connection metadata from an On-Call Investigation team config, pass:
--sqlcl-team-config "<path-to-team-config>" --sqlcl-team-name "<team-name>"
If that team.sqlcl block includes connect_string, username, and password_env_var, bootstrap reads the password from the named variable in the current environment first, then in the configured password_env_file (default ~/.env), and creates or refreshes the saved SQLcl alias. If the block includes only connection_name, bootstrap treats it as an existing saved SQLcl connection and just records the alias plus any tunnel notes.
When Bitbucket or Jira auth is missing for an enabled skill, bootstrap writes the requested helper file and exits non-zero until the required variables are populated and available. Bitbucket helpers should be sourced from the selected shell init file; Jira MCP env files are read directly by the Jira MCPs and the Jira Ticket script.
python3 skills/codex-bootstrap/scripts/bootstrap_codex.py \
--repo-root "<repo-root>" \
--codex-home "<codex-home>" \
--output "<config-path>" \
--workspace-parent "<workspace-parent>" \
--m2-repository "<path-to-your-m2-repository>" \
--model-provider "<provider>" \
--model "<model>" \
--agents-mode copy \
--bitbucket-auth-state needs-setup \
--bitbucket-base-url "<bitbucket-host>" \
--bitbucket-env-file "<path-to-bitbucket-env-helper>" \
--jira-auth-state needs-setup \
--jira-oci-base-url "<jira-oci-host>" \
--generate-jira-oci-mcp-env \
--env-file "<path-to-shared-mcp-env>" \
--generate-mcp-env stlm-mcp \
--generate-mcp-env mcp-dope \
--sqlcl-team-config "<path-to-team-config>" \
--sqlcl-team-name "<team-name>" \
--home-dir "<home-dir>" \
--log-directory "<log-directory>"
Disable anything the engineer does not use:
python3 skills/codex-bootstrap/scripts/bootstrap_codex.py \
--output "~/.codex/config.toml" \
--agents-mode copy \
--disable-mcp sqlcl \
--disable-mcp mcp-atlassian-jira-sd \
--disable-mcp mcp-atlassian-jira-oci \
--disable-mcp lts-mcp \
--disable-mcp mcp_shepherd \
--disable-skill bitbucket-pr
Trim the setup down to only the browser MCP:
python3 skills/codex-bootstrap/scripts/bootstrap_codex.py \
--output "~/.codex/config.toml" \
--agents-mode copy \
--disable-mcp stlm-mcp \
--disable-mcp mcp-dope \
--disable-mcp sqlcl \
--disable-mcp mcp-atlassian-jira-sd \
--disable-mcp mcp-atlassian-jira-oci \
--disable-mcp centralconfluence \
--disable-mcp playwright \
--disable-mcp ots \
--disable-mcp lts-mcp \
--disable-mcp mcp_shepherd
If the engineer wants a terminal-driven interview instead of chat-driven collection, use prompt mode:
python3 skills/codex-bootstrap/scripts/bootstrap_codex.py --prompt
Use --allow-repo-write only when the engineer explicitly wants a machine-specific config inside the shared repo.
3. Verify placeholder cleanup
- Read the script output.
- If it reports unresolved path placeholders, stop and fix them before finishing.
- Call out any remaining non-path template placeholders only when the engineer chose to leave those values blank.
- When validating the shared bootstrap workflow itself, prefer
skills/codex-bootstrap/scripts/verify_bootstrap.sh; it uses disposable temp homes and cleans them up automatically.
4. Explain the result
Report:
- config path written
- backup path if one was created
- rendered global
AGENTS.md target
- enabled skills
- shared skill links installed under
$CODEX_HOME/skills
- enabled MCP servers
- Devplat MCP Gateway plugin status and Bo Peep local prerequisites
- any prerequisite notes for
chrome-devtools, sqlcl, mcp-atlassian-jira-sd, mcp-atlassian-jira-oci, centralconfluence, ots, mcp_shepherd, or lts-mcp
- any SQLcl saved connection that bootstrap configured from team metadata
- any generated Bitbucket PR helper files
- any generated Jira OCI MCP env helper files
- the shell init sourcing lines the engineer should add for those helper files
- any generated MCP env helper files
- any disabled optional integrations
- any remaining manual steps
Notes
- The helper script resolves shared repo references for:
- rendered global
AGENTS.md installation into Codex home
skills/*/SKILL.md- generated local
agents/*.toml derived from shared agents/*.md
- writable roots
- optional local Maven repository access for sandboxed builds
- shared helper-backed MCP
.env files plus shared HOME and log directories
- packaged Chrome DevTools MCP settings
- starter MCP env helper generation when requested
- optional Bitbucket PR env helper generation
- optional Jira OCI MCP env helper generation
- trusted repo example paths
- The shared starter expects local absolute paths inside the generated user config. Keep those paths user-local, not committed.
- The intended steady state is:
- shared repo: versioned source of truth
- Codex home: personal config, logs, auth, and global
AGENTS.md entrypoint
