Gemini Worker

Run Gemini CLI headlessly as a background worker. Gemini handles its own OAuth via

~/.gemini/oauth_creds.json — authenticate once interactively, then all headless runs

use cached credentials.

Key insight: gemini --acp requires a TTY and hangs headlessly. Use gemini -p "..." --yolo

instead — it takes a prompt as argument, runs to completion, and exits. See

references/acp-vs-headless.md for the full technical explanation.

Quick Start

# Simple prompt
gemini -p "Your prompt here" --yolo

# With workspace access (required for file reads/writes outside ~/.gemini)
gemini \
  --include-directories /path/to/dir1,/path/to/dir2 \
  --yolo \
  -p "Your prompt here"

# Long prompts via file (recommended for >1KB prompts)
gemini \
  --include-directories /tmp/task-dir \
  --yolo \
  -p "$(cat /tmp/task-dir/prompt.txt)"

Key Flags

Critical: Gemini CLI restricts file access to its workspace (~/.gemini/tmp/workspace

and --cwd). Use --include-directories to grant access to your project directories.

/tmp is outside the default workspace — include it explicitly if needed.

Anthropic Fallback Pattern

When Anthropic returns 529 Overloaded, Gemini is a capable drop-in worker:

cat > /tmp/fallback-task.txt << 'TASK'
You are a senior engineer. Read the spec at /path/to/spec.md.
Implement the feature. Write code to /path/to/output.ts.
Write a summary to /tmp/summary.md.
TASK

timeout 600 gemini \
  --include-directories /path/to/project,/tmp \
  --yolo \
  -p "$(cat /tmp/fallback-task.txt)"

Parallel Workers Pattern

# Spawn 3 parallel Gemini workers
gemini --include-directories /path/to/task1 --yolo -p "$(cat /tmp/task1.txt)" > /tmp/out1.txt &
gemini --include-directories /path/to/task2 --yolo -p "$(cat /tmp/task2.txt)" > /tmp/out2.txt &
gemini --include-directories /path/to/task3 --yolo -p "$(cat /tmp/task3.txt)" > /tmp/out3.txt &
wait  # Wait for all 3

Pre-fetch Pattern (WebFetchTool Workaround)

Gemini CLI's WebFetchTool is unreliable in headless mode. Pre-fetch with curl:

# Pre-fetch a URL to a file Gemini can read
curl -sL https://example.com/doc | python3 -c "
from html.parser import HTMLParser; import sys
class S(HTMLParser):
    def __init__(self): super().__init__(); self.t=[]
    def handle_data(self, d):
        if d.strip(): self.t.append(d.strip())
p=S(); p.feed(sys.stdin.read()); print('\n'.join(p.t[:500]))
" > /tmp/fetched-doc.txt

# Then run Gemini with access to /tmp
gemini --include-directories /tmp --yolo -p "Read /tmp/fetched-doc.txt and ..."

Wrapper Script

Use scripts/gemini-run.sh (Linux/macOS) or scripts/gemini-run.ps1 (Windows/PowerShell)

for timeout handling, logging, and output capture:

Linux/macOS (bash)

scripts/gemini-run.sh /tmp/prompt.txt "/path/to/project,/tmp" /tmp/output.txt

Windows (PowerShell)

.\scripts\gemini-run.ps1 C:\tmp\prompt.txt "C:\path\to\project,C:\tmp" C:\tmp\output.txt

Windows (Command Prompt)

scripts\gemini-run.cmd C:\tmp\prompt.txt "C:\path\to\project,C:\tmp" C:\tmp\output.txt

Known Limitations

• No access to OpenClaw tools (web_search, memory_search, cron, etc.)
• WebFetchTool unreliable headlessly — pre-fetch with curl instead
• /tmp outside workspace — add --include-directories /tmp/your-dir
• Large prompts (>4KB): write to file, reference by path in prompt
• Auth required once: run gemini interactively to complete OAuth, then headless works
• Output truncation: ask Gemini to write to file instead of relying on stdout for large outputs

References

• references/acp-vs-headless.md — Why ACP fails, how -p --yolo works, decision matrix
• references/prompt-patterns.md — Reusable prompt templates
• references/troubleshooting.md — Common errors and fixes
• scripts/gemini-run.sh — Wrapper with timeout and error handling