| name | install |
| description | Install or refresh the bundled SpecKit Pro Codex custom subagents. Copies the plugin's TOML agent templates into ~/.codex/agents/ by default, verifies the installed files, and tells the user to restart Codex so the new subagents load.
|
SpecKit Install
Scope
Use this skill when the user wants to install, refresh, repair, or
verify the Codex custom subagents that ship with this plugin.
This skill exists only for the Codex runtime. It does not manage
Claude Code agents, Claude marketplace metadata, or Claude commands.
The bundled source of truth lives in the plugin at:
The default user-scope Codex destination is:
If the user explicitly asks for project scope, use .codex/agents/ in the
current project. Arbitrary destinations are rejected. Otherwise install to the
default user-scope path above.
What This Skill Installs
This skill installs the bundled TOML subagent templates that the
Codex autopilot expects to exist as real custom subagents:
autopilot-fast-helper.toml
phase-executor.toml
clarify-executor.toml
checklist-executor.toml
analyze-executor.toml
implement-executor.toml
codebase-analyst.toml
spec-context-analyst.toml
domain-researcher.toml
uat-runbook-author.toml
autopilot-fast-helper.toml is optional at runtime. The main
autopilot may use it for tiny advisory text-only prep work when
gpt-5.3-codex-spark is available, but autopilot must continue
without it if that model is unavailable in the current Codex
environment.
These files follow the official Codex subagent format: one standalone
TOML file per custom agent, with required name, description, and
developer_instructions fields plus Codex config such as model,
model_reasoning_effort, and sandbox_mode.
Operator note: on Codex, sandbox_mode = "read-only" does not sandbox
MCP server processes. The agent TOML cannot restrict tools, so to keep a
read-only agent provably unable to cause writes via MCP, the operator must
curate write-capable MCP servers OUT at the profile/config level (enabled = false, or enabled_tools/disabled_tools).
The bundled model policy runs every execution and consensus agent on
gpt-5.5. Reasoning effort remains exactly as declared by each bundled TOML.
autopilot-fast-helper is the only model exception: it stays on
gpt-5.3-codex-spark for tiny advisory text-only prep, never for SDD reasoning.
If gpt-5.5 is not available in the current Codex environment, set the
helper's model input to gpt-5.4 (or set
SPECKIT_CODEX_MODEL=gpt-5.4); the installer rewrites only destination
copies.
Plugin Upgrade and Agent Refresh Boundary
Codex plugin marketplace updates refresh the plugin cache but do not update
the separately registered files under ~/.codex/agents/. Run this skill after
every SpecKit Pro upgrade, then restart Codex:
codex plugin marketplace upgrade racecraft-plugins-public
@SpecKit Pro -> install (or `$install`)
# Restart Codex
The runner resolves codex-agents/ from the currently executing installed
plugin. It does not mutate plugin caches or marketplace directories. Its
dry_run is content-aware: same-named destination files whose rendered content
differs are planned for refresh, while current files are reported as no-ops.
Hard Constraints
- Never touch
.claude/agents/, .claude-plugin/, commands/, or any
Claude marketplace file.
- Never move or rename the bundled files in
codex-agents/. They are the
plugin's packaged templates and must stay on Codex-only paths.
- Copy only
*.toml files from codex-agents/.
- Do not delete unrelated user subagents already present in the target
directory.
- Overwrite only same-named SpecKit Pro agent files in the target directory.
- If the source bundle is missing or incomplete, STOP and report the exact
missing files.
- After an applied change, always finish by telling the user to restart Codex.
A no-op verification does not require another restart.
Procedure
1. Resolve explicit Codex-only paths
Resolve all paths before mutating anything:
- Determine the plugin root from the current skill location.
- Resolve the source directory at
../../codex-agents/ relative to this
skill.
- Resolve the destination directory:
- default:
~/.codex/agents/
- explicit project scope:
.codex/agents/ in the current project
- Resolve the executor/consensus model:
- default:
gpt-5.5
- fallback:
gpt-5.4 via the helper's model input or
SPECKIT_CODEX_MODEL=gpt-5.4
Do not infer a Claude path from a vague request. If the user says only
"install the agents", use ~/.codex/agents/.
2. Validate the bundled source set
Before copying:
- Verify the source directory exists.
- Verify all expected
*.toml files are present.
- Verify there are no legacy
.md Codex agent files in the source bundle.
If any required file is missing, stop immediately. Do not partially install.
3. Run the bundled installer helper
Use the deterministic installed-runtime helper for install-codex-agents
with the selected destination. Run it first in dry_run mode, then in
apply mode after the plan matches the requested destination and model
override. Use gpt-5.4 only when fallback mode was requested.
The structured request inputs are:
destination: omit for ~/.codex/agents/, or set to .codex/agents/ for
current-project scope
model: gpt-5.5 or gpt-5.4
The helper must be the only mechanism used for copying files. Do not
re-implement the copy loop inline unless the helper itself is broken and
you have already reported that failure.
When fallback mode is requested, verify every destination copy whose bundled
source model is gpt-5.5 uses model = "gpt-5.4". The Spark helper remains
unchanged, and all bundled source templates remain byte-identical.
4. Verify the installed destination
After the helper completes:
- Verify the destination directory exists.
- Verify every expected TOML file now exists in the destination.
- Verify the copied files are the same filenames as the bundled source set.
- For
gpt-5.5, verify every destination file is byte-identical to its
bundled source. For gpt-5.4, verify only the supported model assignment
was rewritten in destination copies.
- Preserve any unrelated user files in the destination.
- Require the helper's verification status to be
verified before reporting
success.
If verification fails, report the mismatch clearly and stop.
5. Report restart requirement
When the helper applied changes, your closing output must explicitly tell the
user:
- where the files were installed
- which files were copied or refreshed
- the effective executor/consensus model
- that they must restart Codex now
When the helper reports no_op, report that verification succeeded and no
additional restart is required.
Do not continue into autopilot setup or workflow execution in the same skill.
Installation ends once the files are copied, verified, and the user has been
told to restart Codex.
Output
Return a concise installation report like:
## SpecKit Codex Subagents Installed
**Source:** /absolute/path/to/plugin/codex-agents
**Destination:** <HOME>/.codex/agents
**Installed files:**
- autopilot-fast-helper.toml
- phase-executor.toml
- clarify-executor.toml
- checklist-executor.toml
- analyze-executor.toml
- implement-executor.toml
- codebase-analyst.toml
- spec-context-analyst.toml
- domain-researcher.toml
- uat-runbook-author.toml
**Next step:** Restart Codex now so the custom subagents are loaded.
Failure Handling
Stop instead of improvising when:
- the bundled
codex-agents/ directory is missing
- any required TOML file is missing
- the installer helper fails or rolls back
- the destination cannot be created
- post-copy verification does not match the bundled source set
If the install partially succeeded, report exactly what copied and what still
needs repair. Do not silently continue.