Paths: File paths (shared/, references/) are relative to skills repo root. Locate this SKILL.md directory and go up one level for repo root.

Agent Installer

Type: L3 Worker Category: 0XX Shared

Installs or updates the two supported CLI agents: Codex CLI and Claude Code. Single pass per agent: install then immediately verify.

MANDATORY READ: Load shared/references/coordinator_summary_contract.md, shared/references/environment_worker_runtime_contract.md, and shared/references/worker_runtime_contract.md

---

Input / Output

Direction	Content
Input	OS info, disabled flags per agent, dry_run flag, optional runId, optional summaryArtifactPath
Output	Structured summary envelope with payload.status = completed / skipped / error, plus per-agent install outcomes in changes / detail

If summaryArtifactPath is provided, write the same summary JSON there. If not provided, return the summary inline and remain fully standalone. If runId is not provided, generate a standalone run_id before emitting the summary envelope.

Runtime

Runtime family: environment-worker-runtime

Phase profile:

1. PHASE_0_CONFIG
2. PHASE_1_INSTALL_VERIFY
3. PHASE_2_POST_CODEX_SANITY_CHECK
4. PHASE_3_WRITE_SUMMARY
5. PHASE_4_SELF_CHECK

Runtime rules:

• emit summary_kind=env-agent-install
• standalone runs generate their own run_id and write the default worker-family artifact path
• managed runs require both runId and summaryArtifactPath and must write the summary to the exact provided path
• always write the validated summary artifact before terminal outcome

Output Contract

Always build a structured env-agent-install summary envelope per:

• shared/references/coordinator_summary_contract.md
• shared/references/environment_worker_runtime_contract.md

Payload fields:

• targets
• agents_processed
• agents_skipped
• versions
• codex_drift_detected
• status

---

Agent Registry

Agent	Install Command	Health Check	Interactive
Codex	npm i -g @openai/codex	codex --version	codex or codex "prompt"
Claude	npm i -g @anthropic-ai/claude-code or claude update	claude --version	claude or claude "prompt"

Automation uses:

• Codex: codex exec ...
• Claude: claude -p ...

Note: codex / codex "prompt" are interactive TTY entrypoints. Non-interactive validation and review automation must use codex exec ..., because Codex rejects direct interactive mode when stdin is not a terminal.

---

Workflow

For each enabled agent: Install or Update -> Verify -> Record

Phase 1: Install & Verify

For each agent in registry, apply first matching rule:

Condition	Action	Report
disabled: true	SKIP	disabled by user
dry_run: true	Show planned command	dry run
Codex	npm install -g @openai/codex then codex --version	version or error
Claude installed	claude update then claude --version	version or error
Claude missing	npm install -g @anthropic-ai/claude-code then claude --version	version or error

Single pass: install and verify happen atomically per agent. No separate scan phase; the install result is the state.

Error handling:

Error	Detection	Response
npm not in PATH	npm --version fails	FAIL gracefully, report npm not found in PATH
Permission denied	stderr contains EACCES	FAIL, suggest npm install -g --prefix ~/.local {pkg}
Network error	stderr contains ETIMEDOUT or ENETUNREACH	FAIL, report network error
Unknown error	Any other non-zero exit	FAIL, include stderr

Output table:

Agent Installation:
| Agent  | Action    | Version  | Status |
|--------|-----------|----------|--------|
| Codex  | installed | 0.124.0  | ok     |
| Claude | updated   | 2.1.119  | ok     |

Phase 2: Post-Install Codex Sanity Check

After successful Codex install/update:

1. Check ~/.codex/config.toml top-level approval_policy and sandbox_mode.
2. Check configured Codex marketplaces for duplicate active paths to this repo.
3. If drift is found, report a WARN and defer remediation to ln-013-config-syncer.

This installer does not install marketplace plugins, rewrite Codex marketplace layout, or manage Codex execution defaults. It only reports drift so install success is not mistaken for a healthy two-agent environment.

---

Critical Rules

1. Never modify disabled flags. Respect them, never change them.
2. Fail gracefully. One agent failure does not block the other.
3. Global install only. CLI tools must be in PATH.
4. Report all changes. Include config observations in the final summary table.
5. Idempotent. Safe to run multiple times.
6. Claude and Codex only. Do not add unrelated agent branches.
7. Codex marketplace/plugin and execution-default remediation belongs to ln-013. This skill may detect drift, but must not rewrite marketplace/cache layout or Codex top-level permission defaults itself.

Anti-Patterns

DON'T	DO
Separate check/install/verify phases	Single pass: install then verify
Retry failed installs automatically	One attempt, report failure
Use sudo npm install	Suggest --prefix for permission issues
Install agents marked disabled	Skip with clear report
Configure MCP settings or marketplace plugins here	Route environment alignment to ln-013-config-syncer

---

Definition of Done

• Codex and Claude processed in single pass (install + verify)
• Disabled agents skipped with report
• Version verified immediately after each install
• Codex marketplace/plugin and execution-default sanity checked or explicitly reported for ln-013 follow-up
• Status table displayed
• Structured summary returned
• Summary artifact written to the managed or standalone runtime path

---

Version: 1.1.0 Last Updated: 2026-03-23