Pass prompt via stdin in /codex skill to fix Windows argv limit (also closes #971, #1034)

Summary

The /codex skill in codex/SKILL.md invokes codex exec with the prompt as a positional argv argument in both Challenge mode (Step 2B) and Consult mode (Step 2C). When the constructed prompt exceeds roughly 32KB - which happens easily in Consult mode because Step 2C prepends the entire plan-file content to the prompt, and in Challenge mode when Claude packs a large git diff into the prompt - the invocation fails on Windows before the Codex binary ever starts.

Observed failure on gstack v0.16+ / codex-cli 0.128.0 / Windows 10 / Git Bash (msys2) / Node v22.22.1, with a ~57KB prompt:

/c/Users/<user>/AppData/Roaming/npm/codex: line 15: /c/Program Files/nodejs/node: Argument list too long
/c/Users/<user>/AppData/Roaming/npm/codex: line 15: /c/Program Files/nodejs/node: No error

Root cause is the Win32 CreateProcess ~32,767-character command-line limit (KB830473). The npm-installed codex shim launches node via Git Bash / cmd.exe and inherits that limit, so any argv larger than ~32KB never reaches the Codex Rust binary. The skill silently breaks for Windows users above that threshold, with a cryptic error from the npm shim that doesn't mention gstack, Codex, or the skill at all.

Why this is worth fixing now (and why it's not a duplicate)

This is closely related to two existing issues that complain about the same codex exec "<prompt>" pattern in skill bash blocks, from different angles:

• #971 --- codex exec hangs in skill bash blocks because stdin isn't closed (missing </dev/null).
• #1034 --- same hang in scripts/resolvers/review.ts and scripts/resolvers/design.ts, same proposed </dev/null patch.

Both of those propose adding </dev/null to keep stdin closed while still passing the prompt as argv. That fixes the hang but doesn't fix this Windows argv-limit case - the prompt still has to fit in CreateProcess.

There is a single change that resolves all three (#971, #1034, and this one) cleanly and follows OpenAI's documented pattern for large prompts: pipe the prompt through stdin instead of passing it as argv, using codex exec -. From the Codex CLI reference (https://developers.openai.com/codex/cli/reference): "Initial instruction for the task. Use - to pipe the prompt from stdin."

Steps to reproduce

On Windows in Git Bash, against any repo with a moderately large plan file (say >40KB):

1. Have ~/.claude/plans/<project>-*.md exist with a long plan (40-80KB is enough).
2. In a Claude Code session inside the project, run /codex with no arguments.
3. The skill auto-detects Consult mode and prepends the plan content to the prompt (Step 2C).
4. When the assembled prompt exceeds ~32KB, the Bash invocation fails:

Bash(... PROMPT=$(cat /tmp/codex/prompt.txt); codex exec "$PROMPT" -C "$_REPO" -s read-only ...)
  ⤷ ---STDERR---
  /c/Users/<user>/AppData/Roaming/npm/codex: line 15: /c/Program Files/nodejs/node: Argument list too long
  /c/Users/<user>/AppData/Roaming/npm/codex: line 15: /c/Program Files/nodejs/node: No error

A minimal manual repro (no skill involved):

python -c "print('x' * 50000)" > /tmp/prompt.txt
PROMPT="$(cat /tmp/prompt.txt)"
codex exec "$PROMPT" -s read-only      # fails with Argument list too long on Windows
codex exec - -s read-only < /tmp/prompt.txt  # works on Windows

Proposed fix

In codex/SKILL.md, switch both codex exec invocations from argv to stdin:

Step 2B (Challenge mode) - current:

codex exec "<prompt>" -s read-only -c 'model_reasoning_effort="xhigh"' --enable web_search_cached --json 2>/dev/null | python3 -c "..."

Step 2B - proposed:

printf '%s' "<prompt>" \
  | codex exec - -s read-only -c 'model_reasoning_effort="xhigh"' --enable web_search_cached --json 2>/dev/null \
  | python3 -c "..."

Step 2C (Consult mode) - same transformation, both for the new-session and resumed-session forms. The <<< "<prompt>" here-string is an alternative but is bash-specific; printf '%s' "$PROMPT" | codex exec - works in any POSIX shell.

This change:

1. Sidesteps the Win32 CreateProcess argv limit entirely - stdin has no such ceiling on any platform.
2. Naturally closes stdin (pipe reaches EOF after the prompt streams in), which also resolves #971 and #1034 without needing </dev/null patches.
3. Follows OpenAI's documented pattern for large prompts in Codex CLI.
4. Is a pure markdown / skill-template change - no binary or installer changes.

Happy to open a PR with the change if maintainers agree on the approach. I'd also suggest a follow-up audit of any other gstack skill that shells out to codex exec or similar long-prompt CLIs (autoplan, plan-eng-review, plan-ceo-review, design-review were mentioned in #1034) for the same pattern.

Environment

• gstack: latest from main (verified pattern still present in codex/SKILL.md Step 2B / 2C)
• codex-cli: 0.128.0
• OS: Windows 10 x64
• Shell: Git Bash (msys2)
• Node: v22.22.1
• Claude Code: latest

Related

• #971 --- codex exec stdin hang (same skill bash pattern, different symptom)
• #1034 --- codex exec stdin hang in review/design resolvers (proposes </dev/null; this issue argues the stronger fix is stdin-as-prompt)