/codex:rescue subagent silently fails to delegate — ${CLAUDE_PLUGIN_ROOT} is empty in subagent Bash

[namzzz]

Summary

When /codex:rescue routes to the codex:codex-rescue subagent, the subagent's forward command fails because ${CLAUDE_PLUGIN_ROOT} resolves to an empty string in the subagent's Bash environment. The command becomes node "/scripts/codex-companion.mjs" task … → Cannot find module '/scripts/codex-companion.mjs'. The subagent then silently falls back to doing its own grep-based investigation and returns a non-Codex answer, so no Codex job is ever created and the user is unaware Codex was never invoked.

Environment

• Plugin codex v1.0.4 (marketplace openai/codex-plugin-cc)
• Claude Code 2.1.186 (Agent SDK 0.3.186), VS Code extension
• codex-cli 0.133.0, Node v20.20.2, Linux (WSL2)
• ChatGPT auth, plugin enabled, project trusted, review-gate off

Repro

1. Install/enable the plugin; confirm /codex:setup reports ready.
2. Run a substantial /codex:rescue <multi-file investigation prompt>.
3. Observe the subagent returns a plausible answer fairly quickly.
4. Run /codex:status → no job exists (running: [], recent: [], latestFinished: null). The returned answer was authored by the subagent, not Codex.

Root cause

${CLAUDE_PLUGIN_ROOT} is a harness template token that Claude Code substitutes when rendering plugin commands and hooks (textually, per-plugin). It is not exported as a real environment variable into the Bash environment of a spawned subagent.

• plugins/codex/agents/codex-rescue.md instructs the subagent to run node "${CLAUDE_PLUGIN_ROOT}/scripts/codex-companion.mjs" task … via Bash.
• In the subagent's shell, ${CLAUDE_PLUGIN_ROOT} expands to empty → module-not-found.

Verified:

• In a subagent Bash, echo "[${CLAUDE_PLUGIN_ROOT}]" prints [] (empty) by default.
• Running the exact forward command with an empty var reproduces Cannot find module '/scripts/codex-companion.mjs'.
• Exporting CLAUDE_PLUGIN_ROOT=<plugin-root> (e.g. via a project env setting) makes the var available in subagent Bash, and /codex:rescue then forwards correctly and creates a real Codex job.

Impact

Silent, dangerous failure mode: the user believes Codex ran when it did not. This directly violates the plugin's own codex-result-handling rule ("if Codex was never successfully invoked, do not generate a substitute answer"), because the substitute is produced by the subagent before any result-handling guidance applies.

Suggested fixes

• Have the subagent resolve the script path robustly instead of relying on ${CLAUDE_PLUGIN_ROOT} being present in the shell env — e.g. pass the absolute companion path into the subagent prompt at command-render time (the slash-command layer already receives the expanded path), or derive it without the env var.
• Alternatively, make the subagent hard-fail (return nothing) when the forward command errors, rather than silently self-investigating — at minimum this surfaces the problem instead of masking it.