fix: pass prompt via stdin to avoid "Argument list too long" on large PRs

[jbiskur]

Problem

run_opencode and run_gemini pass the entire prompt as a single argv argument:

opencode run -m "$full_model" "$(cat "$prompt_file")"      # line 453
gemini -y -m "$GEMINI_MODEL" --prompt "$(cat "$prompt_file")"  # line 365

On Linux the per-argument cap is MAX_ARG_STRLEN = 128 KB (PAGE_SIZE × 32). A PR with a large diff produces a 150 KB+ prompt, so the kernel rejects the exec with Argument list too long (exit 126) and validation hard-fails before the model is ever invoked.

Observed in the wild on a large PR (179 KB prompt):

/home/runner/.opencode/bin/opencode: Argument list too long
❌ OpenCode CLI failed (exit code: 126)
::error::Non-retryable error occurred.

Fix

Feed the prompt through stdin instead. A piped stream has no per-argument size limit. Verified both CLIs read a non-TTY stdin as the prompt directly from their source:

• opencode — packages/opencode/src/cli/cmd/run.ts: const piped = process.stdin.isTTY ? undefined : await Bun.stdin.text() → used as the message when no positional arg is given.
• gemini — packages/cli/src/gemini.tsx: if (!process.stdin.isTTY) stdinData = await readStdin() → pushed as --prompt. (--prompt is also deprecated upstream, so this drops a deprecation too.)

set -o pipefail is already enabled, so $? still reflects the CLI's exit code through the | tee pipe — retry detection and exit-code handling are unchanged.

Testing

• bash -n scripts/validate.sh — syntax clean.
• Confirmed a 179 KB argument triggers Argument list too long at exec on Linux, while the same payload over < file stdin has no limit.
• No behavioral change beyond prompt delivery: model, flags, output capture (| tee), retry loop, and exit-code semantics are identical.

🤖 Generated with Claude Code