name: gemini-delegate description: Use when [large codebase >200 lines], [web search current info], [adversarial review/critique]. metadata: triggers: web search, adversarial review, code analysis, plan critique, large file analysis

Gemini Delegate (ACP)

Delegate tasks to Gemini via ACP bridge. Verification MANDATORY.

Bridge Mandates

• MANDATORY: scripts/gemini_bridge.py only. NEVER call gemini directly.
• Safe I/O: Use --prompt-stdin to prevent deadlocks.
• Persistence: Use --output-file AUTO for unique result files.
• Sessions: Capture/reuse SESSION_ID from JSON for follow-ups.
• Timeout: Total 600s (10m). First chunk 300s.

Grounding & Web Search

• Prefix: Queries MUST start with WebSearch: (e.g., WebSearch: API news 2026).
• Grounding: Prefix forces web grounding. No prefix = fabrication risk.
• Fallback: If bridge fails (timeout/crash), use built-in WebSearch. Report in one line.

Routing & Tasking

CLI Flags

Output Schema

{
  "success": true,
  "SESSION_ID": "...",
  "agent_messages": "...",
  "read_file_count": N,
  "fallback_occurred": boolean,
  "requested_model": "...",
  "actual_model_selection": "specified"|"automatic",
  "tool_calls": [...],
  "token_estimate": { "estimated_cost_usd": 0.00 }
}

STOP & VERIFY (Red Flags)

Violating the letter of these rules violates the spirit. STOP if:

• read_file_count is 0 for code analysis (Gemini is guessing).
• Response contains functions not present in source files.
• Web search response lacks specific source URLs.
• Requested Pro failed (fallback_occurred: true). Alert user immediately.
• Calling gemini directly. Revert and use bridge.
• Omitted WebSearch: prefix. Re-run with prefix.

Rationalization Table

Permissions & Logs

• Write: --approve-edits required for file writes.
• Feedback: Log outcome via --log-feedback after every task.
• Escalation: Rejection rate >50% -> escalate model tier.

References

• assets/prompt-template.md: Task templates.
• assets/reference.md: Logs, parallel runs, capsules.